text
stringlengths 1
2.56M
| id
stringlengths 40
40
| metadata
dict |
|---|---|---|
\subsection{#1}}
\setlength{\textfloatsep}{3pt}
\algdef{SE}[SUBALG]{Indent}{EndIndent}{}{\algorithmicend\ }%
\algtext*{Indent}
\algtext*{EndIndent}
\title{\LARGE \bf
Centralized and Decentralized Techniques for Analysis and Synthesis of Non-Linear Networked Systems
\vspace{-15pt}
}
\author{Shirantha Welikala, Hai Lin and Panos J. Antsaklis
\thanks{The support of the National Science Foundation (Grant No. IIS-1724070, CNS-1830335, IIS-2007949) is gratefully acknowledged.}
\thanks{The authors are with the Department of Electrical Engineering, College of Engineering, University of Notre Dame, IN 46556, \texttt{{\small \{wwelikal,hlin1,pantsakl\}@nd.edu}}.}}
\begin{document}
\maketitle
\thispagestyle{empty}
\pagestyle{empty}
\begin{abstract}
In this paper, we develop centralized and decentralized techniques for analyzing and synthesizing networked systems comprised of interconnected sets of non-linear subsystems - only using the subsystem dissipativity properties. In particular, this paper extends a recent work that proposed dissipativity based centralized analysis techniques for non-linear networked systems as linear matrix inequality (LMI) problems. First, we consider four networked system configurations (NSCs) of interest and provide centralized stability/dissipativity \emph{analysis} techniques for them as LMI problems. Second, we show that the centralized interconnection topology \emph{synthesis} techniques for these NSCs can also be developed as LMI problems. This enables synthesizing the interconnection topology among the subsystems so that specific stability/dissipativity measures of the networked system are optimized. Next, we show that the proposed analysis and synthesis techniques can be implemented in a decentralized and compositional manner (i.e., subsystems can be added/removed conveniently). Finally, we include several numerical results to demonstrate our contributions.
\end{abstract}
\section{Introduction}\label{Sec:Introduction}
Large-scale networked systems have gained a renewed attention over the past several years due to their broad applications in infrastructure networks \cite{Agarwal2021,Arcak2022,Tang2021}, biological networks \cite{Chaves2018,Rufino2012,Arcak2008}, vehicle platooning \cite{Song2022Ax,Karafyllis2021,Antonelli2013}, electronic circuits \cite{Jafarian2019,Jeltsema2005}, mechanical networks \cite{Zhu2015,Papageorgiou2004} and so on. Related to such networked systems, the main research thrusts have been focused on addressing problems such as analysis/verification \cite{Ebihara2017,Arcak2016,Cremean2003}, optimization \cite{Welikala2020P7,Oltafi-Saber2004,Welikala2020J2}, abstraction \cite{Lavaei2020,Zamani2018,Lavaei2020b}, controller synthesis \cite{WelikalaP32022,Nejati2022,Morrison2021} and network topology synthesis \cite{Ghosh2021,Ebihara2017,Rufino2018,Ghanbari2016}. Moreover, the literature on large-scale networked systems can also be categorized based on the nature of the constituent subsystems as: discrete \cite{Welikala2022Ax2}, continuous \cite{Agarwal2021}, linear \cite{WelikalaP32022}, piece-wise linear \cite{Welikala2020J2}, non-linear \cite{Arcak2022}, stochastic \cite{Lavaei2020}, switched \cite{Lavaei2022} and hybrid \cite{Nejati2022}.
With respect to the literature above, this paper aims to address the research problems: analysis/verification and network topology synthesis of large-scale networked systems comprised of non-linear subsystems (henceforth referred to as \emph{analysis} and \emph{synthesis} of non-linear networked systems). In particular, we are interested in \emph{analyzing} the stability/dissipativity properties of a given non-linear networked system and, when necessary, \emph{synthesizing} the interconnection topology of the non-linear networked system to guarantee/optimize such stability/dissipativity properties.
To achieve these goals, we only assume the knowledge of the involved subsystem dissipativity properties such as their passivity, passivity indices, $L_2$-stability and $L_2$-gain values. Note that identifying such dissipativity properties is far more convenient, efficient and reliable than having to identify entire dynamic models \cite{Koch2021,WelikalaP42022,Arcak2022,Xia2014,Zakeri2021}. Further, we formulate the interested non-linear networked system analysis and synthesis problems as linear matrix inequality (LMI) problems. Thus, they can be conveniently implemented and efficiently solved \cite{Boyd1994,Lofberg2004}. Furthermore, we formulate these LMI problems so that they can be implemented in a decentralized and compositional manner. Consequently, new subsystems can be added/removed to/from a networked system conveniently without having to re-design the existing interconnections \cite{Agarwal2021,WelikalaP32022}. Due to these unique qualities, the proposed solutions in this paper can be used for higher-level design of large-scale networked systems. To summarize, we provide dissipativity-based centralized and decentralized LMI techniques to analyze and synthesize non-linear networked systems.
This paper was inspired by the well-known work \cite{Arcak2016,Arcak2022} where dissipativity based centralized LMI techniques have been developed to analyze non-linear networked systems. Compared to \cite{Arcak2016,Arcak2022}, we consider two special non-linear networked system configurations (NSCs) with multiple applications. Moreover, as mentioned earlier, we not only provide techniques for analysis but also address the corresponding synthesis and decentralization problems.
It should be noted that even though the analysis techniques proposed in \cite{Arcak2016,Arcak2022} have been derived in a compositional manner, they can only be evaluated in a centralized setting. To address this challenge, inspired by Sylvester's criterion \cite{Antsaklis2006}, the work in \cite{Agarwal2021,WelikalaP32022} propose a decentralized and compositional iterative process to analyze linear networked systems. In this paper, we adopt this idea to analyze non-linear networked systems in a decentralized and compositional manner.
A critical difference between this work and \cite{Agarwal2021,WelikalaP32022} is that we only use dissipativity properties of the subsystems, whereas \cite{Agarwal2021,WelikalaP32022} use subsystem dynamic models. Note, however, that this additional information and the linear dynamics assumed in \cite{Agarwal2021,WelikalaP32022} also enable the synthesis of standard feedback controllers and observers for linear networked systems. In contrast, due to the said complexities that we consider, here we do not aim to synthesize controllers. Instead, we aim to synthesize appropriate interconnection topologies (in a centralized or decentralized and compositional manner) for non-linear networked systems when the analysis fails.
Nevertheless, there is a wealth of literature that focuses on the synthesis of controllers for networked systems. Some recent examples are as follows. Considering a linear networked system, a decentralized observer based controller is synthesized in \cite{Elmahdi2015}. For stochastic hybrid networked systems, to enforce a particular class of global specifications, a compositional framework is proposed in \cite{Nejati2022} to construct local control barrier functions. This solution is extended in \cite{Jahanshahi2022} for discrete-time partially observable stochastic networked systems. For non-linear dynamical systems, a feed-forward control procedure is proposed in \cite{Morrison2021} using model reduction and bifurcation theory. A decentralized controller synthesis approach is proposed in \cite{Liu2019b} for discrete-time linear networked systems exploiting approximations of robust controlled invariant sets \cite{Rungger2017}. Compositionally constructed abstractions of networked systems are used to synthesize controllers for discrete-time stochastic \cite{Lavaei2019,Lavaei2022,Lavaei2017,Lavaei2020}, continuous-time hybrid \cite{Nejati2021,Awan2020} and nonlinear \cite{Zamani2018} networked systems.
As mentioned earlier, these controller synthesis techniques require additional information and assumptions regarding the involved subsystems and their interconnection topology. In contrast, in this paper, we only use subsystem dissipativity properties and focus on synthesizing the interconnection topology - which describes how each subsystem output is connected to the other subsystem inputs in the considered networked system (characterized by a block matrix called the ``interconnection matrix''). Note that this approach is more reasonable when there are ready-made standard controllers (perhaps with tunable parameters that may change their dissipativity properties). For example, see the scenarios considered in \cite{Cremean2003,Ghanbari2016,Ebihara2017,Rufino2018,Arcak2022,Xia2014,Zakeri2019}.
However, literature focusing directly on synthesizing this interconnection matrix of networked systems are few and far between (especially for non-linear networked systems). In fact, the leading paper that motivated us to address this particular synthesis problem is \cite{Xia2014} (see also \cite{Xia2018}). In particular, \cite{Xia2014,Xia2018} consider a networked system comprised only of one or two subsystems and propose a set of non-linear inequalities to guide the synthesis of the corresponding interconnection matrix. In contrast, we consider a networked system with an arbitrary number of subsystems and provide LMI conditions to efficiently and conveniently synthesize the interconnection matrix (centrally or decentrally). Therefore, the proposed synthesis results in this paper can be seen as a generalization of \cite{Xia2014,Xia2018}. Note also that the techniques proposed in \cite{Xia2014,Xia2018} have influenced several recent works on problems such as designing switched controllers \cite{Ghanbari2017}, networked controllers \cite{Rahnama2018} and adaptive controllers \cite{Zakeri2019}. Therefore, it is reasonable to expect that the proposed techniques in this paper may also be able to generalize the techniques proposed in \cite{Ghanbari2017,Rahnama2018,Zakeri2019}.
In addition, the work in \cite{Ebihara2017} has considered the interconnection topology synthesis problem limited to linear and positive networked systems. However, it requires the explicit knowledge of the involved linear positive subsystems. In \cite{Rufino2018}, symmetries in the interconnection matrix are exploited to reduce the computational complexity of the networked system analysis. However, this approach only allows the interconnection matrix to be diagonally scaled from a predefined value (to recover symmetries). For stability analysis of large-scale interconnected systems, similar symmetry based techniques have been used in \cite{Ghanbari2016,Goodwine2013}. However, they assume the interconnection matrix (hence the topology) as a given - rather than treating it as a decision variable. Taking a graph theoretic approach, necessary conditions for the stability of a class of non-linear networked systems are given in terms of the interconnection matrix in \cite{Cremean2003}. While this solution leads to identify several types of interconnection topologies that guarantee stability, it is only applicable to a small class of non-linear systems with known dynamics.
\paragraph{\textbf{Contributions}}
With respect to the literature mentioned above, our contributions can be summarized as follows:
(1) We consider several networked system configurations (NSCs) that can be used to model several exciting and widely used non-linear networked system configurations;
(2) For the centralized analysis (stability/dissipativity) of such NSCs, we propose linear matrix inequality (LMI) based techniques;
(3) For the centralized synthesis (interconnection topology) of such NSCs, we propose LMI based approaches;
(4) We propose decentralized and compositional (i.e., resilient to subsystem additions and removals) counterparts for the proposed centralized analysis and synthesis approaches; and
(5) We provide several numerical results to support our theoretical results.
\paragraph{\textbf{Organization}}
This paper is organized as follows. In Section \ref{Sec:Preliminaries}, we provide several preliminary concepts and results related to dissipativity, network matrices and their positive definiteness analysis. Different networked system configurations and applicable centralized techniques to analyze their stability and dissipativity are discussed in Sec. \ref{Sec:NetworkedSystems}. In Sec. \ref{Sec:CentralizedSynthesis}, we propose centralized techniques to synthesize the interconnection matrices involved in the considered networked systems. Section \ref{Sec:DecentralizedAnalysisAndSynthesis} provide the decentralized counterparts of the proposed centralized analysis and centralized synthesis techniques. Finally, Sec. \ref{Sec:NumericalResults} discusses several numerical examples before concluding the paper in Sec. \ref{Sec:Conclusion}.
\paragraph{\textbf{Notation}}
The sets of real and natural numbers are denoted by $\mathbb{R}$ and $\mathbb{N}$, respectively. We define $\mathbb{N}_N\triangleq\{1,2,\ldots,N\}$ where $N\in\mathbb{N}$.
An $n\times m$ block matrix $A$ can be represented as $A=[A_{ij}]_{i\in\mathbb{N}_n, j\in\mathbb{N}_m}$ where $A_{ij}$ is the $(i,j)$\tsup{th} block of $A$. $[A_{ij}]_{j\in \mathbb{N}_m}$ and $\text{diag}(A_{ii}:i\in\mathbb{N}_n)$ represent a block row matrix and a block diagonal matrix, respectively. We also define $\{A_i\} \triangleq \{A_{ii}\}\cup\{A_{ij},j\in\mathbb{N}_{i-1}\}\cup\{A_{ji}:j\in\mathbb{N}_i\}$. If $\Psi\triangleq[\Psi^{kl}]_{k,l \in \mathbb{N}_m}$, its block element-wise form \cite{WelikalaP32022} is denoted as $\mbox{BEW}_n(\Psi) \triangleq [[\Psi^{kl}_{ij}]_{k,l\in\mathbb{N}_m}]_{i,j\in\mathbb{N}_n}$ (note that, when indexing, subscripts and superscripts are used interchangeably, e.g., $A^{ij} \equiv A_{ij}$). The transpose of a matrix $A$ is denoted by $A^\top$ and $(A^\top)^{-1} = A^{-\top}$. The zero and identity matrices are denoted by $\mathbf{0}$ and $\mathbf{I}$, respectively (dimensions will be clear form the context). A symmetric positive definite (semi-definite) matrix $A\in\mathbb{R}^{n\times n}$ is represented as $A=A^\top>0$ ($A=A^\top \geq 0$). Unless stated otherwise, $A>0 \iff A=A^\top>0$. The symbol $\star$ is used to represent redundant conjugate matrices (e.g.,
$\scriptsize \bm{A & B \\ \star & C} \equiv \bm{A & B \\ B^\top & C}$, $\scriptsize \bm{A_{ij} & B_{ij} \\ \star & C_{ij}} \equiv \bm{A_{ij} & B_{ij} \\ B_{ji}^\top & C_{ij}}$ and $A^\top B\, \star \equiv A^\top B A$).
The symmetric part of a matrix $A$ is defined as $\H(A) \triangleq A+A^\top$ and $\H(A_{ij}) \triangleq A_{ij}+A_{ji}^\top$. $\mb{1}_{\{\cdot\}}$ represents the indicator function and $e_{ij} \triangleq \mathbf{I} \cdot \mb{1}_{\{i=j\}}$.
\section{Preliminaries}\label{Sec:Preliminaries}
\subsection{Equilibrium Independent Dissipativity}
Consider the non-linear dynamical system
\begin{equation}\label{Eq:GeneralSystem}
\begin{aligned}
\dot{x}(t) = f(x(t),u(t)),\\
y(t) = h(x(t),u(t)),
\end{aligned}
\end{equation}
where $x(t)\in\mathbb{R}^{n}$, $u(t)\in \mathbb{R}^{q}$, $y(t)\in\mathbb{R}^{m}$, and $f:\mathbb{R}^{n}\times \mathbb{R}^{q} \rightarrow \mathbb{R}^{n}$ and $h:\mathbb{R}^{n}\times \mathbb{R}^{q}\rightarrow \mathbb{R}^{m}$ are continuously differentiable. Suppose there exists a set $\mathcal{X}\subset \mathbb{R}^{n}$ where for every $\tilde{x}\in\mathcal{X}$ there is a unique $\tilde{u}\in\mathbb{R}^{q}$ satisfying $f(\tilde{x},\tilde{u})=\mathbf{0}$, and both $\tilde{u}$ and $\tilde{y}\triangleq h(\tilde{x},\tilde{u})$ are implicit functions of $\tilde{x}$.
Note that when $\mathcal{X}=\{x^*\}$ (a singleton), $x^*$ is the unique equilibrium of \eqref{Eq:GeneralSystem} (and $x^*=\mathbf{0}$ is a conventional assumption that we avoid making here). The \emph{equilibrium-independent dissipativity} (EID) property \cite{Arcak2022} defined next (which includes the conventional dissipativity property \cite{Willems1972a}) enables stability and dissipativity analysis of \eqref{Eq:GeneralSystem} without having explicit knowledge regarding the equilibrium point(s) of \eqref{Eq:GeneralSystem}.
\begin{definition}
The system \eqref{Eq:GeneralSystem} is equilibrium-independent dissipative (EID) with supply rate $s:\mathbb{R}^{q}\times\mathbb{R}^{m}\rightarrow \mathbb{R}$ if there exists a continuously differentiable storage function $V:\mathbb{R}^{n}\times \mathcal{X} \rightarrow \mathbb{R}$ satisfying
$\forall (x,\tilde{x},u)\in\mathbb{R}^{n}\times \mathcal{X} \times \mathbb{R}^{q}$:
$V(x,\tilde{x})>0$, $V(\tilde{x},\tilde{x})=0$ and
\begin{equation}
\dot{V}(x,\tilde{x}) = \nabla_x V(x,\tilde{x})f(x,u)
\leq s(u-\tilde{u},y-\tilde{y}).
\end{equation}
\end{definition}
Note that the above EID concept can be specialized based on the form of the used storage function $s(\cdot,\cdot)$.
For example, the well-known $(Q,S,R)$-dissipativity property \cite{WelikalaP32022} is defined with respect to a quadratic supply function. An equivalent dissipativity property (defined next) is used in the remainder of this paper for notational convenience.
\begin{definition}\label{Def:X-EID}
The system \eqref{Eq:GeneralSystem} is $X$-equilibrium independent dissipative ($X$-EID) if it is EID with respect to the quadratic supply rate
\begin{equation}
s(u-\tilde{u},y-\tilde{y}) \triangleq
\bm{u-\tilde{u}\\y-\tilde{y}}^\top
\underbrace{\bm{X^{11} & X^{12}\\mathcal{X}^{21} & X^{22}}}_{\triangleq X}
\bm{u-\tilde{u}\\y-\tilde{y}}.
\end{equation}
\end{definition}
If $\mathcal{X} = \{x^*\}$ where $x^*=\mathbf{0}$ and corresponding equilibrium control and output values are $u^* = \mathbf{0}$ and $y^* = \mathbf{0}$, respectively, $X$-EID and $(Q,S,R)$-dissipativity are equivalent and
\begin{equation}
X \triangleq \bm{X^{11} & X^{12}\\ X^{21} & X^{22}} \equiv
\bm{R & S^\top\\ S & Q}.
\end{equation}
Similar to the $(Q,S,R)$-dissipativity, $X$-EID covers several properties of interest based on the choice of $X$.
\begin{remark}\label{Rm:X-DissipativityVersions}
If the system \eqref{Eq:GeneralSystem} is $X$-EID with:
\begin{enumerate}
\item $X = \bm{\mathbf{0} & \frac{1}{2}\mathbf{I} \\ \frac{1}{2}\mathbf{I} & \mathbf{0}}$, then it is \emph{passive};
\item $X = \bm{-\nu\mathbf{I} & \frac{1}{2}\mathbf{I} \\ \frac{1}{2}\mathbf{I} & -\rho\mathbf{I}}$, then it is \emph{strictly passive} ($\nu$ and $\rho$ are input and output passivity indices, respectively \cite{WelikalaP42022});
\item $X = \bm{\gamma^2\mathbf{I} & \mathbf{0} \\ \mathbf{0} & -\mathbf{I}}$, then it is finite-gain \emph{$L_2$-stable} ($\gamma$ is the $L_2$-gain);
\end{enumerate}
in an equilibrium-independent manner.
\end{remark}
\subsection{Network Matrices}
Consider a directed network $\mathcal{G}_n=(\mathcal{V},\mathcal{E})$ where $\mathcal{V} \triangleq \{\Sigma_i:i\in\mathbb{N}_n\}$ is the set of subsystems (nodes), $\mathcal{E} \subset \mathcal{V}\times \mathcal{V}$ is the set of inter-subsystem interconnections (edges), and $n\in\mathbb{N}$. We next define a class of matrices named ``network matrices'' \cite{WelikalaP32022} that corresponds to a such network topology $\mathcal{G}_n$.
\begin{definition}\cite{WelikalaP32022}\label{Def:NetworkMatrices}
Given a network $\mathcal{G}_n=(\mathcal{V},\mathcal{E})$, any $n\times n$ block matrix $\Theta = \bm{\Theta_{ij}}_{i,j\in\mathbb{N}_n}$ is a \emph{network matrix} if:
(1) any information specific to the subsystem $i$ is embedded only in its $i$\tsup{th} block row or block column, and
(2) $(\Sigma_i,\Sigma_j) \not\in \mathcal{E}$ and $(\Sigma_j,\Sigma_i)\not\in \mathcal{E}$ implies $\Theta_{ij}=\Theta_{ji}=\mathbf{0}$ for all $i,j\in\mathbb{N}_n$.
\end{definition}
Based on this definition, any $n \times n$ block matrix $\Theta=[\Theta_{ij}]_{i,j\in\mathbb{N}_n}$ is a network matrix of $\mathcal{G}_n$ if $\Theta_{ij}$ is a coupling weight matrix corresponding to the edge $(\Sigma_i,\Sigma_j)\in\mathcal{V}$. Note also that any $n \times n$ block diagonal matrix $\Theta=\text{diag}(\Theta_{ii}:i\in \mathbb{N}_n)$ will be a network matrix of any network with $n\in\mathbb{N}$ subsystems if $\Theta_{ii}$ is specific only to the subsystem $i$.
The following lemma provides several useful properties of such network matrices - first established in \cite{WelikalaP32022}.
\begin{lemma}
\label{Lm:NetworkMatrixProperties}
\cite{WelikalaP32022}
Given a network $\mathcal{G}_n$, a few corresponding block network matrices $\Theta,\Phi,\{\Psi^{kl}:k,l\in\mathbb{N}_m\}$, and some arbitrary block-block matrix $\Psi\triangleq[\Psi^{kl}]_{k,l \in \mathbb{N}_m}$:
\begin{enumerate}
\item $\Theta^\top$, \ $\alpha \Theta + \beta \Phi$ are network matrices for any $\alpha,\beta \in \mathbb{R}$.
\item $\Phi \Theta$, $\Theta\Phi$ are network matrices whenever $\Phi$ is a block diagonal network matrix.
\item $\mbox{BEW}(\Psi)\triangleq [[\Psi^{kl}_{ij}]_{k,l\in\mathbb{N}_m}]_{i,j\in\mathbb{N}_n}$ is a network matrix.
\end{enumerate}
\end{lemma}
The above lemma enables claiming custom block matrices as ``network matrices'' by enforcing additional conditions. For example, if $M,P$ are block network matrices and $P$ is block diagonal, then: (1) $M^\top P, P M$ and $M^T P + PM$ are network matrices, and (2) if $\scriptsize \Psi\triangleq\bm{P & M^\top P\\ PM & P}$ is some block-block matrix, its \emph{block element-wise} (BEW) form $\mbox{BEW}(\Psi)\triangleq$ $\scriptsize \bm{\bm{P_{ii}e_{ij} & M_{ji}^\top P_{jj} \\ P_{ii}M_{ij} & P_{ii}e_{ij}}}_{i,j\in\mathbb{N}_N}$ is a network matrix.
\subsection{Positive Definiteness}
We next provide several useful lemmas on the positive (or negative) definiteness of a few matrix expressions of interest.
\begin{lemma}\label{Lm:AlternativeLMI_Schur}
Let $W \triangleq \Phi^\top \Theta \Phi - \Gamma$ where $\Phi,\Theta$ and $\Gamma$ are arbitrary matrices such that $\Theta > 0$ and $\Gamma = \Gamma^\top$.
Then:
$$W < 0 \iff
\bm{\Theta & \Theta \Phi \\ \Phi^\top \Theta & \Gamma} > 0.$$
\end{lemma}
\begin{proof}
Using the Schur's complement theory and the congruence principle \cite{Bernstein2009}, we can directly obtain:\\
$
W<0 \iff \bm{\Theta^{-1} & \Phi \\ \Phi^\top & \Gamma} > 0 \iff
\bm{\Theta & \Theta \Phi \\ \Phi^\top \Theta & \Gamma} > 0.
$
\end{proof}
\begin{lemma}\label{Lm:AlternativeLMI_LowerBound}
Let $W \triangleq \Phi^\top \Theta \Phi - \Gamma$ where $\Phi,\Theta$ and $\Gamma$ are arbitrary matrices such that $\Theta < 0$ and $\Gamma = \Gamma^\top$.
Then:
$$W < 0 \impliedby
- \Phi^\top \Theta \Psi - \Psi^\top \Theta \Phi + \Psi^\top \Theta \Psi + \Gamma > 0,$$
for any matrix $\Psi$ with appropriate dimensions.
\end{lemma}
\begin{proof}
Note that $\Theta<0 \implies (\Phi-\Psi)^\top \Theta (\Phi-\Psi)<0
\iff \Phi^\top \Theta \Phi - \Gamma < \Phi^\top \Theta \Psi + \Psi^\top \Theta \Phi - \Psi^\top \Theta \Psi -\Gamma$.
Thus, $W<0 \impliedby -\Phi^\top \Theta \Psi - \Psi^\top \Theta \Phi + \Psi^\top \Theta \Psi + \Gamma>0$.
\end{proof}
\begin{remark}
Note that, the use of $\Psi=\mathbf{0}$ in the above lemma yields $W <0 \impliedby \Gamma > 0$. However, the use of $\Psi=\alpha \mathbf{I}$ (assuming $\Phi$ is a square matrix) yields
\begin{equation}
W<0 \impliedby -\alpha\left( \Phi^\top \Theta + \Theta \Phi \right) + \alpha^2 \Theta + \Gamma > 0.
\end{equation}
Hence the latter choice for $\Psi$ (i.e., $\Psi = \alpha \mathbf{I}$) is more effective whenever
$-\alpha\left( \Phi^\top \Theta + \Theta \Phi \right) + \alpha^2 \Theta > 0$. Since $\alpha^2 \Theta < 0$, this condition holds whenever $\Phi^\top \Theta + \Theta \Phi>0$ with $\alpha < 0$ or $\Phi^\top \Theta + \Theta \Phi<0$ with $\alpha>0$.
\end{remark}
\begin{lemma}\label{Lm:AlternativeLMI_BEW}
\cite{WelikalaP32022}
Let $\Psi = [\Psi^{kl}]_{k,l\in\mathbb{N}_m}$ be an $m\times m$ block-block matrix where $\Psi^{kl}, \forall k,l\in\mathbb{N}_m$ are $n \times n$ block matrices. Then, $\Psi > 0 \iff \text{BEW}(\Psi) \triangleq [[\Psi^{kl}_{ij}]_{k,l\in\mathbb{N}_m}]_{i,j\in\mathbb{N}_n}>0$.
\end{lemma}
\begin{remark}
Lemmas \ref{Lm:AlternativeLMI_Schur} and \ref{Lm:AlternativeLMI_LowerBound} can be used to replace a particular form of a tri-linear matrix inequality condition with a bi-linear matrix inequality condition. This will be useful later on when deriving LMI conditions. Moreover, if $\Phi,\Theta,\Psi$ and $\Gamma$ used in Lemmas \ref{Lm:AlternativeLMI_Schur} and \ref{Lm:AlternativeLMI_LowerBound} are block network matrices of some network and $\Phi$ and $\Psi$ are also block diagonal, then, with the help of Lemmas \ref{Lm:NetworkMatrixProperties} and \ref{Lm:AlternativeLMI_BEW}, it is easy to see that the aforementioned bi-linear matrix inequality conditions can also be replaced with a positive-definiteness condition of a network matrix. This will be useful later on when decentralizing different LMIs of interest.
\end{remark}
Inspired by the Sylvester's criterion \cite{Antsaklis2006}, the following lemma provides a decentralized and compositional testing criterion to evaluate the positive definiteness of an $N\times N$ block matrix $W=[W_{ij}]_{i,j\in\mathbb{N}_N}$ (for more details, see \cite{Welikala2022Ax2}).
\begin{lemma}\label{Lm:MainLemma}
\cite{WelikalaP32022}
A symmetric $N \times N$ block matrix $W = [W_{ij}]_{i,j\in\mathbb{N}_N} > 0$ if and only if
\begin{equation}\label{Eq:Lm:MainLemma1}
\tilde{W}_{ii} \triangleq W_{ii} - \tilde{W}_i \mathcal{D}_i \tilde{W}_i^\top > 0,\ \ \ \ \forall i\in\mathbb{N}_N,
\end{equation}
where
\begin{equation}\label{Eq:Lm:MainLemma2}
\begin{aligned}
\tilde{W}_i \triangleq&\ [\tilde{W}_{ij}]_{j\in\mathbb{N}_{i-1}} \triangleq W_i(\mathcal{D}_i\mathcal{A}_i^\top)^{-1},\\
W_i \triangleq&\ [W_{ij}]_{j\in\mathbb{N}_{i-1}}, \ \ \
\mathcal{D}_i \triangleq \text{diag}(\tilde{W}_{jj}^{-1}:j\in\mathbb{N}_{i-1}),\\
\mathcal{A}_i \triangleq&\
\bm{
\tilde{W}_{11} & \mathbf{0} & \cdots & \mathbf{0} \\
\tilde{W}_{21} & \tilde{W}_{22} & \cdots & \mathbf{0}\\
\vdots & \vdots & \vdots & \vdots \\
\tilde{W}_{i-1,1} & \tilde{W}_{i-1,2} & \cdots & \tilde{W}_{i-1,i-1}
}.
\end{aligned}
\end{equation}
\end{lemma}
\begin{remark}\label{Rm:Lm:MainLemma}
If $W$ is a \emph{symmetric block network matrix} corresponding to some network $\mathcal{G}_N$, the above lemma can be used test $W>0$ in a decentralized and compositional manner (i.e., by sequentially testing $\tilde{W}_{ii}>0$ at each subsystem $\Sigma_i, i\in\mathbb{N}_N$).
\end{remark}
\begin{algorithm}[!t]
\caption{Testing/Enforcing $W>0$ in a Network Setting.}
\label{Alg:DistributedPositiveDefiniteness}
\begin{algorithmic}[1]
\State \textbf{Input: } $W = [W_{ij}]_{i,j\in\mathbb{N}_N}$
\State \textbf{At each subsystem $\Sigma_i, i \in \mathbb{N}_N$ execute:}
\Indent
\If{$i=1$}
\State Test/Enforce: $W_{11}>0$
\State Store: $\tilde{W}_1 \triangleq [W_{11}]$ \Comment{To be sent to others.}
\Else
\State \textbf{From each subsystem $\Sigma_j, j\in\mathbb{N}_{i-1}$:}
\Indent
\State Receive: $\tilde{W}_j \triangleq [\tilde{W}_{j1},\tilde{W}_{j2},\ldots,\tilde{W}_{jj}]$
\State Receive: Required info. to compute $W_{ij}$
\EndIndent
\State \textbf{End receiving}
\State Construct: $\mathcal{A}_i, \mathcal{D}_i$ and $W_i$.
\Comment{Using: \eqref{Eq:Lm:MainLemma2}.}
\State Compute: $\tilde{W}_i \triangleq W_i (\mathcal{D}_i\mathcal{A}_i^\top)^{-1}$
\Comment{From \eqref{Eq:Lm:MainLemma2}.}
\State Compute: $\tilde{W}_{ii} \triangleq W_{ii} - \tilde{W}_{i}\mathcal{D}_i\tilde{W}_i^\top$
\Comment{From \eqref{Eq:Lm:MainLemma1}.}
\State Test/Enforce: $\tilde{W}_{ii} > 0$
\State Store: $\tilde{W}_i \triangleq [\tilde{W}_i, \tilde{W}_{ii}]$
\Comment{To be sent to others}
\EndIf
\EndIndent
\State \textbf{End execution}
\end{algorithmic}
\end{algorithm}
The above lemma shows that testing positive definiteness of an $N\times N$ block matrix $W=[W_{ij}]_{i,j\in\mathbb{N}_N}$ can be broken down to $N$ separate smaller sequence of tests (iterations).
In a network setting (i.e., where $W$ is a block network matrix corresponding to a network $\mathcal{G}_N$), at the $i$\tsup{th} iteration (i.e., at the subsystem $\Sigma_i$), we now only need to test whether $\tilde{W}_{ii}>0$, where $\tilde{W}_{ii}$ can be computed using:
(1) $\{W_{ij}:j \in \mathbb{N}_i\}$ (related blocks to the subsystem $\Sigma_i$ extracted from $W$),
(2) $\{\tilde{W}_{ij}:j\in \mathbb{N}_{i-1}\}$ (computed using \eqref{Eq:Lm:MainLemma2} at subsystem $\Sigma_i$), and
(3) $\{\{\tilde{W}_{jk}:k\in\mathbb{N}_j\}:j\in\mathbb{N}_{i-1}\}$ (blocks computed using \eqref{Eq:Lm:MainLemma2} at previous subsystems $\{\Sigma_j:j\in\mathbb{N}_{i-1}\}$).
Therefore, it is clear that Lm. \ref{Lm:MainLemma} can be used to test/enforce the positive definiteness of a network matrix in a decentralized manner. Further, as shown in \cite{Welikala2022Ax2}, for some network topologies, this process is fully distributed (i.e., no communications are required between non-neighboring subsystems). Furthermore, it is established in \cite{WelikalaP32022} that this process is compositional (i.e., resilient to subsystem removals/additions from/to the network). This decentralized and compositional approach to test/enforce the positive-definiteness of a network matrix $W$ is summarized in the Alg. \ref{Alg:DistributedPositiveDefiniteness}.
\section{Networked Systems}\label{Sec:NetworkedSystems}
\subsection{Networked System Configurations}
\begin{figure}[!t]
\centering
\begin{subfigure}[t]{0.23\textwidth}
\centering
\captionsetup{justification=centering}
\includegraphics[width=1.0in]{Figures/Interconnection1.png}
\caption{NSC 1}
\label{Fig:Interconnection1}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.23\textwidth}
\centering
\captionsetup{justification=centering}
\includegraphics[width=1.4in]{Figures/Interconnection2.png}
\caption{NSC 2}
\label{Fig:Interconnection2}
\end{subfigure}
\begin{subfigure}[t]{0.23\textwidth}
\centering
\captionsetup{justification=centering}
\includegraphics[width=1.0in]{Figures/Interconnection3.png}
\caption{NSC 3}
\label{Fig:Interconnection3}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.23\textwidth}
\centering
\captionsetup{justification=centering}
\includegraphics[width=1.4in]{Figures/Interconnection4.png}
\caption{NSC 4}
\label{Fig:Interconnection4}
\end{subfigure}
\caption{Considered four networked (interconnected) system configurations (abbreviated as NSCs 1-4).}
\label{Fig:InterconnectedSystems}
\end{figure}
In this paper, we consider the four networked (interconnected) system configurations (NSCs) shown in Fig. \ref{Fig:InterconnectedSystems}. Each of these NSC is comprised of independent dynamic subsystems $\Sigma_i,i\in\mathbb{N}_N, \bar{\Sigma}_i, i\in\mathbb{N}_{\bar{N}}$ and a static $M$ matrix that defines how the subsystems, an exogenous input signal $w(t) \in \mathbb{R}^{r}$ (e.g., disturbance) and interested output signal $z(t) \in\mathbb{R}^{l}$ (e.g., performance) are interconnected, over $t \geq 0$.
Analogous to \eqref{Eq:GeneralSystem}, dynamics of subsystem $\Sigma_i, i\in\mathbb{N}_N$ are
\begin{equation}
\begin{aligned}
\dot{x}_i(t) = f_i(x_i(t),u_i(t)),\\
y_i(t) = h_i(x_i(t),u_i(t)),
\end{aligned}
\end{equation}
where $x_i(t)\in\mathbb{R}^{n_i}, u_i(t)\in\mathbb{R}^{q_i}, y_i(t)\in\mathbb{R}^{m_i}$ and $u \triangleq \bm{u_i^\top}_{i\in\mathbb{N}_N}^\top, y \triangleq \bm{y_i^\top}_{i\in\mathbb{N}_N}^\top$. Further, each subsystem $\Sigma_i, i\in\mathbb{N}_N$ is assumed to have a set $\mathcal{X}_i \subset \mathbb{R}^{n_i}$ where for every $\tilde{x}_i \in \mathcal{X}_i$, there is a unique $\tilde{u}_i \in \mathbb{R}^{q_i}$ satisfying $f_i(\tilde{x}_i,\tilde{u}_i)=0$, and both $\tilde{u}_i$ and $\tilde{y}_i\triangleq h_i(\tilde{x}_i,\tilde{u}_i)$ are implicit functions of $\tilde{x}_i$. Furthermore, subsystem $\Sigma_i, i\in\mathbb{N}_N$ is assumed to be $X_i$-EID, i.e., there is a storage function $V_i:\mathbb{R}^{n_i}\times \mathcal{X}_i\rightarrow \mathbb{R}$ satisfying $\forall(x_i,\tilde{x}_i,u_i)\in \mathbb{R}^{n_i}\times \mathcal{X}_i \times \mathbb{R}^{q_i}:$ $V_i(x_i,\tilde{x}_i)>0$, $V_i(\tilde{x}_i,\tilde{x}_i)=0$, and
\begin{equation*}
\dot{V}_i(x_i,\tilde{x}_i)
\leq \bm{u_i-\tilde{u}_i\\y_i-\tilde{y}_i}^\top
\underbrace{\bm{X_i^{11} & X_i^{12}\\ X_i^{21} & X_i^{22}}}_{X_i} \bm{u_i-\tilde{u}_i\\y_i-\tilde{y}_i}.
\end{equation*}
Subsystems $\bar{\Sigma}_i,i\in\mathbb{N}_{\bar{N}}$ are defined similarly but involves quantities denoted with a bar symbol (e.g., subsystem $\bar{\Sigma}_i,i\in\mathbb{N}_{\bar{N}}$ is $\bar{X}_i$-EID with a storage function $\bar{V}_i:\mathbb{R}^{\bar{n}_i} \times \bar{\mathcal{X}}_i \rightarrow \mathbb{R}$).
On the other hand, the format (block structure) of the static $M$ matrix involved in different NSC differs according to the quantities that it interconnects. In particular, for the considered four NSCs shown in Fig. \ref{Fig:InterconnectedSystems}, the interconnection relationships (that involves $M$) are respectively as follows:
\begin{align}
\label{Eq:NSC1Interconnection}\mbox{NSC 1}:&\ \ \ u \ \ = M y \equiv M_{uy} y,\\
\label{Eq:NSC2Interconnection}\mbox{NSC 2}:&\ \bm{u\\z} \, = M \bm{y\\w} \equiv
\bm{M_{uy} & M_{uw} \\ M_{zy} & M_{zw}}\bm{y\\w},\\
\label{Eq:NSC3Interconnection}\mbox{NSC 3}:&\ \bm{u \\ \bar{u}} \ = M \bm{y\\ \bar{y}} \equiv
\bm{M_{uy} & M_{u\bar{y}} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}}}\bm{y\\ \bar{y}},\\
\label{Eq:NSC4Interconnection}\mbox{NSC 4}:&\ \bm{u \\ \bar{u} \\ z} \, = M \bm{y\\ \bar{y} \\ w} \equiv
\bm{M_{uy} & M_{u\bar{y}} & M_{uw} \\
M_{\bar{u}y} & M_{\bar{u}\bar{y}} & M_{\bar{u}w}\\
M_{zy} & M_{z\bar{y}} & M_{zw}}
\bm{y\\ \bar{y} \\ w}.
\end{align}
\subsection{Connection to Related Work on NSCs}
The work in \cite{Arcak2016,Arcak2022} (and references therein) have already studied the NSCs 1-2 inspired by their applicability in different application domains such as in resource allocation in communications networks, multi-agent motion coordination and biochemical reaction networks. In particular, \cite{Arcak2016,Arcak2022} provide linear matrix inequality (LMI) conditions to analyze the EI-stability of NSC 1 and $X$-EID of NSC 2. However, these LMIs are only applicable for analysis (i.e., when the interconnection matrix $M$ is given) but not for synthesis (i.e., when $M$ needs to be designed), of NSCs. Moreover, even though \cite{Arcak2016,Arcak2022} derive the said LMI conditions in a compositional manner, they can only be evaluated in a centralized setting. These limitations of \cite{Arcak2016,Arcak2022} regarding NSCs 1-2 motivated this paper to focus on these synthesis and decentralization aspects.
Compared to \cite{Arcak2016,Arcak2022}, the earlier work in \cite{Xia2014,Zakeri2022} has actually considered the problem of synthesizing the interconnection matrix $M$. However, the proposed synthesizing strategy in \cite{Xia2014,Zakeri2022} is limited to the NSC 2 with a single subsystem (i.e., $N=1$). Moreover, this strategy \cite[Th. 3]{Xia2014} involves non-linear inequality conditions, and thus, it is not straightforward. In this paper, we will address these limitations while generalizing the result in \cite[Th. 3]{Xia2014}.
Another interesting contribution of \cite{Xia2014} is that it not only considers the problem of synthesizing $M$ for the NSC 2 (albeit with $N=1$), it also provide conditions under which a synthesized NSC 2 can be used as its controller for another system. This motivated the study of NSCs 3-4 as they can be seen as instants where a NSC 2 has been used as a controller to govern another networked system (comprised of subsystems $\bar{\Sigma}_i,i\in\bar{B}$). In this paper, we will generalize the result in \cite[Th. 4]{Xia2014} by providing LMI conditions for EI-stability of NSC 3 and $X$-EID of NSC 4, in both centralized and decentralized settings. Moreover, as shown in Tab. \ref{Tab:Configurations}, the NSC 4 (similarly, the NSC 3) can be used to represent several standard control system configurations.
Finally, we also point out that our recent work in \cite{WelikalaP32022} has addressed the decentralization of several standard LMI based control solutions associated with networked systems comprised of linear subsystems. Therefore, this paper can also be viewed as a extension of \cite{WelikalaP32022} for networked systems comprised of non-linear systems.
\begin{table}[!h]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|c|c|}\hline
Control System Configuration & \begin{tabular}{@{}c@{}} Format of $M$ \eqref{Eq:NSC4Interconnection}
\\ in the NSC 4
\end{tabular}\\ \hline
\begin{tabular}{@{}c@{}} \raisebox{-0.5\totalheight}{\includegraphics[width=2.6in]{Figures/Series.png}}
\\ (Series Configuration)
\end{tabular}
& $\scriptsize \bm{\mathbf{0} & -M_{\tilde{u}\bar{y}} & \mathbf{I} \\ M_{\bar{u}y} & \mathbf{0} & \mathbf{0} \\ \mathbf{0} & M_{z\bar{y}} & \mathbf{0}}$ \\ \hline
\begin{tabular}{@{}c@{}}
\raisebox{-0.5\totalheight}{\includegraphics[width=2.2in]{Figures/Parallel.png}}
\\ (Parallel Configuration)
\end{tabular}
& $\scriptsize \bm{\mathbf{0} & \mathbf{0} & \mathbf{I} \\ \mathbf{0} & \mathbf{0} & M_{\bar{u}w} \\ \mathbf{I} & M_{\tilde{z}y} & \mathbf{0}}$ \\ \hline
\begin{tabular}{@{}c@{}}
\raisebox{-0.5\totalheight}{\includegraphics[width=2.1in]{Figures/Feedback.png}}
\\ (Feedback Configuration)
\end{tabular}
& $\scriptsize \bm{\mathbf{0} & -M_{\tilde{u}\bar{y}} & \mathbf{I} \\ M_{\bar{u}y} & \mathbf{0} & \mathbf{0} \\ \mathbf{I} & \mathbf{0} & \mathbf{0}}$ \\ \hline
\begin{tabular}{@{}c@{}}
\raisebox{-0.5\totalheight}{\includegraphics[width=1.5in]{Figures/Reconfig.png}}
\\ (Feedback Reconfiguration Configuration \cite{Xia2014})
\end{tabular}
& $\scriptsize \bm{-M_{\tilde{u}y} & -M_{\tilde{u}\bar{y}} & \mathbf{I} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}} & \mathbf{0} \\ \mathbf{I} & \mathbf{0} & \mathbf{0}}$ \\ \hline
\begin{tabular}{@{}c@{}}
\raisebox{-0.5\totalheight}{\includegraphics[width=1.8in]{Figures/Approx.png}}
\\ (Approximate Simulation Configuration \cite{Song2022})
\end{tabular}
& $\scriptsize \bm{\mathbf{0} & \mathbf{0} & \mathbf{I} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}} & M_{\bar{u}w} \\ \mathbf{I} & -\mathbf{I} & \mathbf{0}}$ \\ \hline
\end{tabular}}
\caption{Some standard control system configurations and the corresponding format of $M$ \eqref{Eq:NSC4Interconnection} in the NSC 4.}
\label{Tab:Configurations}
\end{table}
\subsection{Centralized Stability and Dissipativity Analysis of NSCs}
The following four propositions provide centralized LMI conditions for stability and $\textbf{Y}$-dissipativity analysis of NSCs 1-4 (using the EID properties of the subsystems). It is noteworthy that these analysis techniques are respectively independent of the equilibrium points of the NSCs 1-4.
\begin{proposition}\label{Pr:NSC1Stability}\hspace{-2mm}\cite{Arcak2006}
Let $x^* \triangleq [x_i^{*\top}]_{i\in\mathbb{N}_N}^\top$ be an equilibrium point of NSC 1 with $x_i^* \in\mathcal{X}_i, \forall i\in\mathbb{N}_N$. Then, $x^*$ is stable for NSC 1 if there exist scalars $p_i>0, \forall i \in \mathbb{N}_N$ such that
\begin{equation}\label{Eq:Pr:NSC1Stability1}
\bm{M_{uy}\\ \mathbf{I}}^\top \textbf{X}_p \bm{M_{uy}\\ \mathbf{I}} \leq 0,
\end{equation}
where components of $M$ are from \eqref{Eq:NSC3Interconnection} and
\begin{equation}\label{Eq:Pr:NSC1Stability2}
\textbf{X}_p \triangleq \scriptsize
\bm{
\text{diag}(p_i X_i^{11}:i\in\mathbb{N}_N) & \text{diag}(p_i X_i^{12}:i\in\mathbb{N}_N)\\
\text{diag}(p_i X_i^{21}:i\in\mathbb{N}_N) & \text{diag}(p_i X_i^{22}:i\in\mathbb{N}_N)\\
}.\normalsize
\end{equation}
\end{proposition}
\begin{proposition}\label{Pr:NSC2Dissipativity}\cite{Arcak2016}
Let $x^* \triangleq [x_i^{*\top}]_{i\in\mathbb{N}_N}^\top$ be an equilibrium point of NSC 2 (under $w(t)=\mathbf{0}$) with $x_i^* \in\mathcal{X}_i, \forall i\in\mathbb{N}_N$ and $z^*$ be the corresponding output. Then, the NSC 2 is $\textbf{Y}$-dissipative, i.e., dissipative with respect to the supply rate:
\begin{equation}\label{Eq:Pr:NSC2Dissipativity0}
\bm{w\\z-z^*}^\top \textbf{Y} \bm{w\\z-z^*},
\end{equation}
if there exist scalars $p_i \geq 0, \forall i \in \mathbb{N}_N$ such that
\begin{equation}\label{Eq:Pr:NSC2Dissipativity1}
\bm{M_{uy} & M_{uw} \\ \mathbf{I} & \mathbf{0} \\
\mathbf{0} & \mathbf{I} \\ M_{zy} & M_{zw}}^\top
\bm{\textbf{X}_p & \mathbf{0} \\ \mathbf{0} & -\textbf{Y}}
\bm{M_{uy} & M_{uw} \\ \mathbf{I} & \mathbf{0} \\
\mathbf{0} & \mathbf{I} \\ M_{zy} & M_{zw}} \leq 0,
\end{equation}
where components of $M$ are from \eqref{Eq:NSC2Interconnection} and $\textbf{X}_p$ is as in \eqref{Eq:Pr:NSC1Stability2}.
\end{proposition}
\begin{proposition}\label{Pr:NSC3Stability}
Let $\hat{x}^* \triangleq \left[[x_i^{*\top}]_{i\in\mathbb{N}_N}, \ [\bar{x}_i^{*\top}]_{i\in\mathbb{N}_{\bar{N}}}\right]^\top$ be an equilibrium point of NSC 3 with $x_i^* \in \mathcal{X}_i, \forall i\in\mathbb{N}_N$ and $\bar{x}_i^* \in \bar{\mathcal{X}}_i, \forall i\in\mathbb{N}_{\bar{N}}$. Then, $\hat{x}^*$ is stable for the NSC 3 if there exist scalars $p_i > 0, \forall i \in \mathbb{N}_N$ and $\bar{p}_i > 0, \forall i \in \mathbb{N}_{\bar{N}}$ such that
\begin{equation}\label{Eq:Pr:NSC3Stability1}
\bm{M_{uy} & M_{u\bar{y}} \\ \mathbf{I} & \mathbf{0} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}} \\ \mathbf{0} & \mathbf{I}}^\top
\bm{\textbf{X}_p & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}}}
\bm{M_{uy} & M_{u\bar{y}} \\ \mathbf{I} & \mathbf{0} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}} \\ \mathbf{0} & \mathbf{I}} \leq 0,
\end{equation}
where components of $M$ are from \eqref{Eq:NSC3Interconnection}, $\textbf{X}_p$ is as in \eqref{Eq:Pr:NSC1Stability2} and
\begin{equation}\label{Eq:Pr:NSC3Stability2}
\bar{\textbf{X}}_{\bar{p}} \triangleq \scriptsize
\bm{
\text{diag}(\bar{p}_i\bar{X}_i^{11}:i\in\mathbb{N}_N) & \text{diag}(\bar{p}_i\bar{X}_i^{12}:i\in\mathbb{N}_N)\\
\text{diag}(\bar{p}_i\bar{X}_i^{21}:i\in\mathbb{N}_N) & \text{diag}(\bar{p}_i\bar{X}_i^{22}:i\in\mathbb{N}_N)
}.\normalsize
\end{equation}
\end{proposition}
\begin{proof}
Consider the Lyapunov function
\begin{equation}\label{Eq:Pr:NSC3StabilityStep1}
V(\hat{x}) \triangleq \sum_{i\in\mathbb{N}_N} p_i V_i(x_i,x_i^*) + \sum_{i\in\mathbb{N}_{\bar{N}}} \bar{p}_i \bar{V}_i(\bar{x}_i,\bar{x}_i^*)
\end{equation}
constructed from the storage functions of subsystems $\Sigma_i, i\in\mathbb{N}_N$ and $\bar{\Sigma}_i, i\in\mathbb{N}_{\bar{N}}$. Due to the EID of subsystems, we get:
\begin{align}
&\dot{V}(\hat{x}) \leq
\sum_{i\in\mathbb{N}_N} p_i \bm{u_i-u_i^*\\y_i-y_i^*}^\top X_i \bm{u_i-u_i^*\\y_i-y_i^*}
\nonumber \\
&+ \sum_{i\in\mathbb{N}_{\bar{N}}} \bar{p}_i \bm{\bar{u}_i-\bar{u}_i^*\\\bar{y}_i-\bar{y}_i^*}^\top \bar{X}_i \bm{\bar{u}_i-\bar{u}_i^*\\\bar{y}_i-\bar{y}_i^*} \nonumber\\
&= \bm{u-u^*\\y-y^*}^\top \hspace{-1mm}\textbf{X}_p\hspace{-1mm} \bm{u-u^*\\y-y^*}
+ \bm{\bar{u}-\bar{u}^*\\\bar{y}-\bar{y}^*}^\top \hspace{-1mm}\bar{\textbf{X}}_{\bar{p}}\hspace{-1mm} \bm{\bar{u}-\bar{u}^*\\\bar{y}-\bar{y}^*}. \label{Eq:Pr:NSC3StabilityStep2}
\end{align}
Note that the quantities with a superscript ``$*$'' correspond to the assumed equilibrium point $\hat{x}^*$. Using \eqref{Eq:NSC3Interconnection}, we can write
\begin{equation}\label{Eq:Pr:NSC3StabilityStep3}
\begin{aligned}
\bm{u-u^*\\y-y^*} =
\bm{M_{uy} & M_{u\bar{y}} \\ \mathbf{I} & \mathbf{0}}
\bm{y-y^*\\ \bar{y}-\bar{y}^*},\\
\bm{\bar{u}-\bar{u}^*\\\bar{y}-\bar{y}^*} =
\bm{M_{\bar{u}y} & M_{\bar{u}\bar{y}} \\ \mathbf{0} & \mathbf{I}}
\bm{y-y^*\\ \bar{y}-\bar{y}^*}.
\end{aligned}
\end{equation}
Applying \eqref{Eq:Pr:NSC3StabilityStep3} in \eqref{Eq:Pr:NSC3StabilityStep2} and combining the two terms, we get:
\begin{equation*}
\begin{aligned}
\dot{V}(\hat{x}) \leq
\bm{y-y^*\\ \bar{y}-\bar{y}^*}^\top
\bm{M_{uy} & M_{u\bar{y}} \\ \mathbf{I} & \mathbf{0} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}} \\ \mathbf{0} & \mathbf{I}}^\top
\bm{\textbf{X}_p & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}}} \, \star.
\end{aligned}
\end{equation*}
Thus, \eqref{Eq:Pr:NSC3Stability1} implies $\dot{V}(\hat{x}) \leq 0$. Further, by definition \eqref{Eq:Pr:NSC3StabilityStep1}, $\dot{V}(\hat{x}^*) = 0$. Hence $\hat{x}^*$ for the NSC 3 is Lyapunov stable.
\end{proof}
\begin{proposition}\label{Pr:NSC4Dissipativity}
Let $\hat{x}^* \triangleq \left[[x_i^{*\top}]_{i\in\mathbb{N}_N}, [\bar{x}_i^{*\top}]_{i\in\mathbb{N}_{\bar{N}}}\right]^\top$ be an equilibrium point of NSC 4 (under $w(t)=0$) with $x_i^* \in \mathcal{X}_i, \forall i\in\mathbb{N}_N$, $\bar{x}_i^* \in \bar{\mathcal{X}}_i, \forall i\in\mathbb{N}_{\bar{N}}$ and $z^*$ be the corresponding output. Then, the NSC 4 is $\textbf{Y}$-dissipative, i.e., dissipative with respect to the supply rate \eqref{Eq:Pr:NSC2Dissipativity0},
if there exist scalars $p_i \geq 0, \forall i \in \mathbb{N}_N$ and $\bar{p}_i, \forall i\in\mathbb{N}_{\bar{N}}$ such that
\begin{equation}\label{Eq:Pr:NSC4Dissipativity}
\bm{
M_{uy} & M_{u\bar{y}} & M_{uw} \\ \mathbf{I} & \mathbf{0} & \mathbf{0} \\
M_{\bar{u}y} & M_{\bar{u}\bar{y}} & M_{\bar{u}w} \\ \mathbf{0} & \mathbf{I} & \mathbf{0} \\
\mathbf{0} & \mathbf{0} & \mathbf{I}\\ M_{zy} & M_{z\bar{y}} & M_{zw}
}^\top
\bm{
\textbf{X}_p & \mathbf{0} & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} & -\textbf{Y}
}
\, \star \leq 0,
\end{equation}
where components of $M$ are from \eqref{Eq:NSC4Interconnection}, and $\textbf{X}_p$ and $\bar{\textbf{X}}_{\bar{p}}$ are as defined in \eqref{Eq:Pr:NSC1Stability2} and \eqref{Eq:Pr:NSC3Stability2}, respectively.
\end{proposition}
\begin{proof}
Consider $V(\hat{x})$ in \eqref{Eq:Pr:NSC3StabilityStep1} now as a storage function. Using the EID of subsystems, we can obtain an upper-bound for $\dot{V}(\hat{x})$ as in \eqref{Eq:Pr:NSC3StabilityStep2}. Thus, the NSC 4 is $\textbf{Y}$-dissipative if
\begin{equation}\label{Pr:NSC4DissipativityStep1}
\bm{u-u^*\\y-y^*}^\top \textbf{X}_p \,\star
+ \bm{\bar{u}-\bar{u}^*\\\bar{y}-\bar{y}^*}^\top \bar{\textbf{X}}_{\bar{p}} \,\star
\leq
\bm{w\\z-z^*}^\top \textbf{Y} \,\star.
\end{equation}
On the other hand, using \eqref{Eq:NSC4Interconnection}, we can write
\begin{equation}\label{Pr:NSC4DissipativityStep2}
\begin{aligned}
\bm{u-u^*\\y-y^*} =
\bm{M_{uy} & M_{u\bar{y}} & M_{uw} \\ \mathbf{I} & \mathbf{0} & \mathbf{0}}
\bm{y-y^*\\ \bar{y}-\bar{y}^* \\ w},\\
\bm{\bar{u}-\bar{u}^*\\\bar{y}-\bar{y}^*} =
\bm{M_{\bar{u}y} & M_{\bar{u}\bar{y}} & M_{\bar{u}w} \\ \mathbf{0} & \mathbf{I} & \mathbf{0} }
\bm{y-y^*\\ \bar{y}-\bar{y}^* \\ w},\\
\bm{w\\z-z^*} =
\bm{\mathbf{0} & \mathbf{0} & \mathbf{I} \\ M_{zy} & M_{z\bar{y}} & M_{zw}}
\bm{y-y^*\\ \bar{y}-\bar{y}^* \\ w}.
\end{aligned}
\end{equation}
Applying \eqref{Pr:NSC4DissipativityStep2} in \eqref{Pr:NSC4DissipativityStep1} and doing some simplifications, we can obtain the condition for NSC 4 to be $\textbf{Y}$-dissipative as:
\begin{equation*}
\bm{y-y^* \\ \bar{y}-\bar{y}^* \\ w}^\top
\hspace{-1mm}
\bm{
M_{uy} & M_{u\bar{y}} & M_{uw} \\ \mathbf{I} & \mathbf{0} & \mathbf{0} \\
M_{\bar{u}y} & M_{\bar{u}\bar{y}} & M_{\bar{u}w} \\ \mathbf{0} & \mathbf{I} & \mathbf{0} \\
\mathbf{0} & \mathbf{0} & \mathbf{I}\\ M_{zy} & M_{z\bar{y}} & M_{zw}
}^\top
\hspace{-1mm}
\bm{
\textbf{X}_p & \mathbf{0} & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} & -\textbf{Y}
} \star \leq 0.
\end{equation*}
Therefore, \eqref{Eq:Pr:NSC4Dissipativity} implies that $\dot{V}(\hat{x})$ is bounded by the storage function \eqref{Eq:Pr:NSC2Dissipativity0}, in other words, NSC 4 is $\textbf{Y}$-dissipative.
\end{proof}
\begin{remark}\label{Rm:StrictNegativeDefiniteness}
For the analysis of NSCs 1-4, Props. \ref{Pr:NSC1Stability}-\ref{Pr:NSC4Dissipativity} provide LMI conditions of the form $W \leq 0$. In the sequel, we address rather practical problems such as the synthesis of $M$ and the decentralized analysis/synthesis of NSCs 1-4. Thus, henceforth, we consider the LMI conditions given in Props. \ref{Pr:NSC1Stability}-\ref{Pr:NSC4Dissipativity} to be of the form $W<0$ (stricter than before).
\end{remark}
\section{Centralized Synthesis of Networked Systems} \label{Sec:CentralizedSynthesis}
In this section, we consider the problem of synthesizing the interconnection matrix $M$ \eqref{Eq:NSC1Interconnection}-\eqref{Eq:NSC4Interconnection} associated with the NSCs 1-4 so as to enforce stability or $\textbf{Y}$-dissipativity. To this end, we first introduce some additional notations as follows.
Let us denote the components of:
$\textbf{X}_p$ in \eqref{Eq:Pr:NSC1Stability2} as $\textbf{X}_p \equiv [\textbf{X}_p^{kl}]_{k,l\in\mathbb{N}_2}$ (i.e., $\textbf{X}_p^{kl} \triangleq \text{diag}(p_i X_i^{kl}:i\in\mathbb{N}_N),\forall k,l \in\mathbb{N}_2$),
$\bar{\textbf{X}}_{\bar{p}}$ in \eqref{Eq:Pr:NSC3Stability2} as $\bar{\textbf{X}}_{\bar{p}} \equiv [\bar{\textbf{X}}_{\bar{p}}^{kl}]_{k,l\in\mathbb{N}_2}$ (i.e., $\bar{\textbf{X}}_{\bar{p}}^{kl} \triangleq \text{diag}(\bar{p}_i \bar{X}_i^{kl}:i\in\mathbb{N}_N), \forall k,l \in\mathbb{N}_2$), and
$\textbf{Y}$ in \eqref{Eq:Pr:NSC2Dissipativity0} as $\textbf{Y} \equiv [\textbf{Y}^{kl}]_{k,l\in\mathbb{N}_2}$. Let us also denote:
$\textbf{X}^{12} \triangleq \text{diag}((X_i^{11})^{-1}X_i^{12}:i\in\mathbb{N}_N)$,
$\textbf{X}^{21} \triangleq \text{diag}(X_i^{21}(X_i^{11})^{-1}:i\in\mathbb{N}_N)$,
$\bar{\textbf{X}}^{12} \triangleq \text{diag}((\bar{X}_i^{11})^{-1}\bar{X}_i^{12}:i\in\mathbb{N}_{\bar{N}})$, and
$\bar{\textbf{X}}^{21} \triangleq \text{diag}(\bar{X}_i^{21}(\bar{X}_i^{11})^{-1}:i\in\mathbb{N}_{\bar{N}})$.
Further, we require the following assumption regarding the dissipativity properties of the subsystems $\{\Sigma_i: i\in\mathbb{N}_N\}$.
\begin{assumption}\label{As:PositiveDissipativity}
Each subsystem $\Sigma_i, i\in\mathbb{N}_N$ is $X_i$-EID with $X_i^{11}>0$. Also, each subsystem $\bar{\Sigma}_i, i\in\mathbb{N}_{\bar{N}}$ is $\bar{X}_i$-EID with $\bar{X}_i^{11}>0$.
\end{assumption}
\begin{remark}\label{Rm:As:PositiveDissipativity}
According to Rm. \ref{Rm:X-DissipativityVersions}, the above assumption directly holds if each subsystem (say $\Sigma_i, i\in\mathbb{N}_N$) is: (1) L2G($\gamma_i$) (as $X_i^{11} = \gamma_i^2 \mathbf{I} > 0$), or (2) IFP($\nu_i$) with $\nu_i<0$ (i.e., non-passive, as $X_i^{11}=-\nu_i\mathbf{I} > 0$). Therefore, As. \ref{As:PositiveDissipativity} holds in many cases of interest.
Note also that, even if a subsystem (say $\Sigma_i, i\in\mathbb{N}_N$) is IFP($\nu_i$) with $\nu_i \geq 0$ (i.e., passive), using the fact that IFP($\nu_i$) $\implies$ IFP($\nu_i-\epsilon_i$) for any $\epsilon_i>0$ \cite[Rm. 1]{Welikala2022Ax1}, As. \ref{As:PositiveDissipativity} can still be upheld by re-assigning the passivity indices. Another less conservative approach to handle this ``passive'' case is given later on in Rm. \ref{Rm:As:PositiveDissipativityFailure}.
\end{remark}
Finally, we also require the following assumption on $\textbf{Y}$.
\begin{assumption}\label{As:NegativeDissipativity}
The dissipativity specification for the NSCs 2 and 4 (i.e., $\textbf{Y}$-dissipativity) is such that $\textbf{Y}^{22}<0$.
\end{assumption}
\begin{remark}\label{Rm:As:NegativeDissipativity}
According to Rm. \ref{Rm:X-DissipativityVersions}, the above assumption holds if we expect the NSCs 2 and 4 to be: (1) L2G($\gamma$) (as $\textbf{Y}^{22} = -\mathbf{I} < 0$), or (2) OFP($\rho$) with $\rho>0$ (i.e., passive, as $\textbf{Y}^{22} = -\rho \mathbf{I} < 0$). Note that it is always desirable to make a system $L_2$-stable or passive. Therefore, As. \ref{As:NegativeDissipativity} is mild.
\end{remark}
\subsection{Centralized Synthesis of NSCs}
The following four propositions provide centralized LMI problems that can be used to synthesize the interconnection matrix $M$ \eqref{Eq:NSC1Interconnection}-\eqref{Eq:NSC4Interconnection} for the NSCs 1-4, respectively.
\begin{proposition}\label{Pr:NSC1Synthesis}
Under As. \ref{As:PositiveDissipativity}, a stabilizing interconnection matrix $M_{uy}$ \eqref{Eq:NSC1Interconnection} for the NSC 1 can be found via the LMI:
\begin{equation}\label{Eq:Pr:NSC1Synthesis}
\begin{aligned}
\mbox{Find: }& L_{uy}, \{p_i: i\in\mathbb{N}_N\}\\
\mbox{ Sub to: }& p_i > 0, \forall i\in\mathbb{N}_N,\\
& \bm{ \textbf{X}_p^{11} & L_{uy} \\ L_{uy}^\top & -(L_{uy}^\top \textbf{X}^{12} + \textbf{X}^{21}L_{uy} + \textbf{X}_p^{22})
}>0,
\end{aligned}
\end{equation}
as $M_{uy} \triangleq (\textbf{X}_p^{11})^{-1} L_{uy}$.
\end{proposition}
\begin{proof}
Using the new notations, we can restate the stability condition for the NSC 1 \eqref{Eq:Pr:NSC1Stability1} as (see also Rm. \ref{Rm:StrictNegativeDefiniteness}):
\begin{equation}\label{Pr:NSC1SynthesisStep1}
M_{uy}^\top \textbf{X}_p^{11} M_{uy} + M^\top_{uy} \textbf{X}_p^{12} + \textbf{X}_p^{21} M_{uy} + \textbf{X}_p^{22} < 0.
\end{equation}
Under As. \ref{As:PositiveDissipativity}, $\textbf{X}_p^{11} > 0$, and thus, Lm. \ref{Lm:AlternativeLMI_Schur} can be used to write an equivalent condition to \eqref{Pr:NSC1SynthesisStep1} as:
\begin{equation}\label{Pr:NSC1SynthesisStep2}
\bm{ \textbf{X}_p^{11} & \textbf{X}_p^{11} M_{uy} \\ M_{uy}^\top \textbf{X}_p^{11} & -(M^\top_{uy} \textbf{X}_p^{12} + \textbf{X}_p^{21} M_{uy} + \textbf{X}_p^{22})}>0.
\end{equation}
Since the objective is to design $M_{uy}$, we need the change of variables
$L_{uy} \triangleq \textbf{X}_p^{11}M_{uy}$. Consequently, $\textbf{X}_p^{21}M_{uy} = \textbf{X}_p^{21}(\textbf{X}_p^{11})^{-1} L_{uy} = \textbf{X}^{21}L_{uy}$ and
$M^\top_{uy} \textbf{X}_p^{12} = L_{uy}^\top (\textbf{X}_p^{11})^{-1}\textbf{X}_p^{12} = L_{uy}^\top\textbf{X}^{12}$. Thus, \eqref{Pr:NSC1SynthesisStep2} can be formulated as the LMI (in $\{p_i:i\in\mathbb{N}_N\}$ and $L_{uy}$) given in \eqref{Eq:Pr:NSC1Synthesis}.
\end{proof}
\begin{proposition}\label{Pr:NSC2Synthesis}
Under As. \ref{As:PositiveDissipativity} and \ref{As:NegativeDissipativity}, the NSC 2 can be made $\textbf{Y}$-dissipative by synthesizing the interconnection matrix $M \equiv \scriptsize \bm{M_{uy} & M_{uw}\\ M_{zy} & M_{zw}}$ \eqref{Eq:NSC2Interconnection} via the LMI:
\begin{equation}\label{Eq:Pr:NSC2Synthesis}
\begin{aligned}
\mbox{Find: }& L_{uy}, L_{uw}, M_{zy}, M_{zw}, \{p_i: i\in\mathbb{N}_N\}\\
\mbox{ Sub. to: }& p_i > 0, \forall i\in\mathbb{N}_N, \mbox{ and \eqref{Eq:Pr:NSC2Synthesis2}},
\end{aligned}
\end{equation}
with $M_{uy} \triangleq (\textbf{X}_p^{11})^{-1} L_{uy}$ and $M_{uw} \triangleq (\textbf{X}_p^{11})^{-1} L_{uw}$.
\end{proposition}
\begin{proof}
Using the introduced notations, we can restate the condition for $\textbf{Y}$-dissipativity of the NSC 2 (i.e., \eqref{Eq:Pr:NSC2Dissipativity1}) as
\begin{equation}\label{Eq:Pr:NSC2SynthesisStep1}
\begin{aligned}
\bm{M_{uy} & M_{uw}\\ M_{zy} & M_{zw}}^\top
\bm{\textbf{X}_p^{11} & \mathbf{0} \\ \mathbf{0} & -\textbf{Y}^{22}}
\bm{M_{uy} & M_{uw}\\ M_{zy} & M_{zw}}\\
+ \bm{ M_{uy}^\top \textbf{X}_p^{12} + \textbf{X}_p^{21}M_{uy} + \textbf{X}_p^{22} & \textbf{X}_p^{21}M_{uw} \\ M_{uw}^\top \textbf{X}_p^{12} & \mathbf{0}}\\
- \bm{\mathbf{0} & M_{zy}^\top \textbf{Y}^{21} \\ \textbf{Y}^{12}M_{zy} & M_{zw}^\top \textbf{Y}^{21} + \textbf{Y}^{12}M_{zw} + \textbf{Y}^{11}}
< 0.
\end{aligned}
\end{equation}
Under As. \ref{As:PositiveDissipativity} and \ref{As:NegativeDissipativity}, $\scriptsize \bm{\textbf{X}_p^{11} & \mathbf{0} \\ \mathbf{0} & -\textbf{Y}^{22}}>0$, and thus, Lm. \ref{Lm:AlternativeLMI_Schur} can be used to write an equivalent condition to \eqref{Eq:Pr:NSC2SynthesisStep1}. This equivalent condition, under the change of variables $L_{uy} \triangleq \textbf{X}_p^{11}M_{uy}$ and $L_{uw} \triangleq \textbf{X}_p^{11}M_{uw}$, can be re-formulated as the LMI (in $\{p_i:i\in\mathbb{N}_N\},L_{uy},L_{uw},M_{zy}$ and $M_{zw}$) given in \eqref{Eq:Pr:NSC2Synthesis}.
\end{proof}
\begin{proposition}\label{Pr:NSC3Synthesis}
Under As. \ref{As:PositiveDissipativity}, a stabilizing interconnection matrix $M \equiv \scriptsize \bm{M_{uy} & M_{u\bar{y}} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}}}$ \eqref{Eq:NSC3Interconnection} for the NSC 3 can be found via the LMI:
\begin{equation}\label{Eq:Pr:NSC3Synthesis1}
\begin{aligned}
\mbox{Find: } &L_{uy},L_{u\bar{y}},L_{\bar{u}y},L_{\bar{u}\bar{y}},\{p_i:i\in\mathbb{N}_N\},\{\bar{p}_i:i\in\mathbb{N}_{\bar{N}}\}\\
\mbox{Sub. to: } &p_i>0, \forall i\in\mathbb{N}_N, \ \ \bar{p}_i>0, \forall i\in\mathbb{N}_{\bar{N}}, \mbox{ and } \eqref{Eq:Pr:NSC3Synthesis2},
\end{aligned}
\end{equation}
as
$
\scriptsize \bm{M_{uy} & M_{u\bar{y}} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}}}
=
\bm{\textbf{X}_p^{11} & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}}^{11}}^{-1}
\bm{L_{uy} & L_{u\bar{y}} \\ L_{\bar{u}y} & L_{\bar{u}\bar{y}}}
$.
\end{proposition}
\begin{proof}
Using the introduced notations, we can re-state the stability condition for the NSC 3 \eqref{Eq:Pr:NSC3Stability1} as
\begin{equation}\label{Eq:Pr:NSC3SynthesisStep1}
\begin{aligned}
\bm{M_{uy} & M_{u\bar{y}} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}}}^\top
\bm{\textbf{X}_p^{11} & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}}^{11}}
\bm{M_{uy} & M_{u\bar{y}} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}}}\\
+
\bm{M_{uy}^\top \textbf{X}_p^{12}+\textbf{X}_p^{21} M_{uy}+\textbf{X}_p^{22} & \textbf{X}_p^{21} M_{u\bar{y}} \\ M_{u\bar{y}}^\top \textbf{X}_p^{12} & \mathbf{0}} \\
+
\bm{\mathbf{0} & M_{\bar{u}y}^\top \bar{\textbf{X}}_{\bar{p}}^{12} \\ \bar{\textbf{X}}_{\bar{p}}^{21} M_{\bar{u}y} &
M_{\bar{u}\bar{y}}^\top \bar{\textbf{X}}_{\bar{p}}^{12} + \bar{\textbf{X}}_{\bar{p}}^{21} M_{\bar{u}\bar{y}} + \bar{\textbf{X}}_{\bar{p}}^{22}} < 0.
\end{aligned}
\end{equation}
Under As. \ref{As:PositiveDissipativity}, $\scriptsize \bm{\textbf{X}_p^{11} & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}}^{11}}>0$, and thus, Lm. \ref{Lm:AlternativeLMI_Schur} is applicable here to write an equivalent condition - which, under the change of variables:
$
\scriptsize
\bm{L_{uy} & L_{u\bar{y}} \\ L_{\bar{u}y} & L_{\bar{u}\bar{y}}} =
\bm{\textbf{X}_p^{11} & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}}^{11}}
\bm{M_{uy} & M_{u\bar{y}} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}}}
$,
can be re-formulated as the LMI given in \eqref{Eq:Pr:NSC3Synthesis1}.
\end{proof}
\begin{proposition}\label{Pr:NSC4Synthesis}
Under As. \ref{As:PositiveDissipativity} and As. \ref{As:NegativeDissipativity}, the NSC 4 can be made $\textbf{Y}$-dissipative by synthesizing the interconnection matrix $M$ \eqref{Eq:NSC4Interconnection} via the LMI:
\begin{equation}\label{Eq:Pr:NSC4Synthesis1}
\begin{aligned}
\mbox{Find: }
&L_{uy}, L_{u\bar{y}}, L_{uw}, L_{\bar{u}y}, L_{\bar{u}\bar{y}}, L_{\bar{u}w}, M_{zy}, M_{z\bar{y}}, \\
&M_{zw}, \{p_i:i\in\mathbb{N}_N\}, \{\bar{p}_i:i\in\mathbb{N}_{\bar{N}}\} \\
\mbox{Sub. to: } &p_i \geq 0, \forall i\in\mathbb{N}_N, \ \
\bar{p}_i \geq 0, \forall i\in\mathbb{N}_{\bar{N}},\ \ \eqref{Eq:Pr:NSC4Synthesis2},
\end{aligned}
\end{equation}
with
$
\scriptsize
\bm{M_{uy} & M_{u\bar{y}} & M_{uw} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}} & M_{\bar{u}w}} =
\bm{\textbf{X}_p^{11} & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}}^{11}}^{-1} \hspace{-1mm} \bm{L_{uy} & L_{u\bar{y}} & L_{uw} \\ L_{\bar{u}y} & L_{\bar{u}\bar{y}} & L_{\bar{u}w}}
$.
\end{proposition}
\begin{proof}
Using the introduced notations, we can re-state the condition for $\textbf{Y}$-dissipativity of the NSC 4 (i.e., \eqref{Eq:Pr:NSC4Dissipativity}) as
\begin{equation}\label{Eq:Pr:NSC4SynthesisStep1}
\begin{aligned}
\bm{M_{uy} & M_{u\bar{y}} & M_{uw} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}} & M_{\bar{u}w} \\ M_{zy} & M_{z\bar{y}} & M_{zw}}^\top
\bm{\textbf{X}_p^{11} & \mathbf{0} & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}}^{11} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} & -\textbf{Y}^{22}}\star \\
+\bm{M_{uy}^\top \textbf{X}_p^{12}+\textbf{X}_p^{21}M_{uy}+\textbf{X}_p^{22} & \textbf{X}_p^{21}M_{u\bar{y}} & \textbf{X}_p^{21}M_{uw}\\ M_{u\bar{y}}^\top \textbf{X}_p^{12} & \mathbf{0} & \mathbf{0} \\ M_{uw}^\top \textbf{X}_p^{12} & \mathbf{0} & \mathbf{0}}\\
+\bm{\mathbf{0} & M_{\bar{u}y}^\top \bar{\textbf{X}}_{\bar{p}}^{12} & \mathbf{0} \\ \bar{\textbf{X}}_{\bar{p}}^{21}M_{\bar{u}y} & M_{\bar{u}\bar{y}}^\top \bar{\textbf{X}}_{\bar{p}}^{12} + \bar{\textbf{X}}_{\bar{p}}^{12}M_{\bar{u}\bar{y}}+\bar{\textbf{X}}_{\bar{p}}^{22} & \bar{\textbf{X}}_{\bar{p}}^{21} M_{\bar{u}w}\\ \mathbf{0} & M_{\bar{u}w}^\top\bar{\textbf{X}}_{\bar{p}}^{12} & \mathbf{0}}\\
-\bm{\mathbf{0} & \mathbf{0} & M_{zy}^\top \textbf{Y}^{21}\\ \mathbf{0} & \mathbf{0} & M_{z\bar{y}}^\top \textbf{Y}^{21}\\ \textbf{Y}^{12}M_{zy} & \textbf{Y}^{12}M_{z\bar{y}} & M_{zw}^\top\textbf{Y}^{21} + \textbf{Y}^{12}M_{zw} + \textbf{Y}^{11} } < 0.
\end{aligned}
\end{equation}
Similar to the proof of Prop. \ref{Pr:NSC2Synthesis}, under As. \ref{As:PositiveDissipativity} and As. \ref{As:NegativeDissipativity}, Lm. \ref{Lm:AlternativeLMI_Schur} is applicable here to write an equivalent condition for \eqref{Eq:Pr:NSC4SynthesisStep1}. This condition, under the change of variables
$
\scriptsize
\bm{L_{uy} & L_{u\bar{y}} & L_{uw} \\ L_{\bar{u}y} & L_{\bar{u}\bar{y}} & L_{\bar{u}w}} =
\bm{\textbf{X}_p^{11} & \mathbf{0} \\ \mathbf{0} & \bar{\textbf{X}}_{\bar{p}}^{11}}
\bm{M_{uy} & M_{u\bar{y}} & M_{uw} \\ M_{\bar{u}y} & M_{\bar{u}\bar{y}} & M_{\bar{u}w}}
$, can be re-formulated as the LMI \eqref{Eq:Pr:NSC4Synthesis1}.
\end{proof}
\begin{figure*}[!hb]
\centering
\hrulefill
\begin{equation}\label{Eq:Pr:NSC2Synthesis2}
\bm{
\textbf{X}_p^{11} & \mathbf{0} & L_{uy} & L_{uw} \\
\mathbf{0} & -\textbf{Y}^{22} & -\textbf{Y}^{22}M_{zy} & -\textbf{Y}^{22} M_{zw}\\
L_{uy}^\top & -M_{zy}^\top\textbf{Y}^{22} & -L_{uy}^\top \textbf{X}^{12}-\textbf{X}^{21}L_{uy}-\textbf{X}_p^{22} & -\textbf{X}^{21}L_{uw}+M_{zy}^\top\textbf{Y}^{21} \\
L_{uw}^\top & -M_{zw}^\top\textbf{Y}^{22} & -L_{uw}^\top \textbf{X}^{12}+\textbf{Y}^{12} M_{zy} & M_{zw}^\top\textbf{Y}^{21} + \textbf{Y}^{12}M_{zw} + \textbf{Y}^{11}
}>0
\end{equation}
\begin{equation}\label{Eq:Pr:NSC3Synthesis2}
\bm{
\textbf{X}_p^{11} & \mathbf{0} & L_{uy} & L_{u\bar{y}}\\
\mathbf{0} & \bar{\textbf{X}}_{\bar{p}}^{11} & L_{\bar{u}y} & L_{\bar{u}\bar{y}}\\
L_{uy}^\top & L_{\bar{u}y}^\top & -L_{uy}^\top \textbf{X}^{12} - \textbf{X}^{21} L_{uy} -\textbf{X}_p^{22} & -\textbf{X}^{21}L_{u\bar{y}}-L_{\bar{u}y}^\top \bar{\textbf{X}}^{12} \\
L_{u\bar{y}}^\top & L_{\bar{u}\bar{y}}^\top & -L_{u\bar{y}}^\top\textbf{X}^{12}-\bar{\textbf{X}}^{21} L_{\bar{u}y} & -L_{\bar{u}\bar{y}}^\top \bar{\textbf{X}}^{12} - \bar{\textbf{X}}^{21} L_{\bar{u}\bar{y}} - \bar{\textbf{X}}_{\bar{p}}^{22}
}>0,
\end{equation}
\begin{equation}\label{Eq:Pr:NSC4Synthesis2}
\bm{
\textbf{X}_p^{11} & \mathbf{0} & \mathbf{0} & L_{uy} & L_{u\bar{y}} & L_{uw} \\
\mathbf{0} & \bar{\textbf{X}}_{\bar{p}}^{11} & \mathbf{0} & L_{\bar{u}y} & L_{\bar{u}\bar{y}} & L_{\bar{u}w}\\
\mathbf{0} & \mathbf{0} & -\textbf{Y}^{22} & -\textbf{Y}^{22} M_{zy} & -\textbf{Y}^{22} M_{z\bar{y}} & \textbf{Y}^{22} M_{zw}\\
L_{uy}^\top & L_{\bar{u}y}^\top & - M_{zy}^\top\textbf{Y}^{22} & -L_{uy}^\top\textbf{X}^{12}-\textbf{X}^{21}L_{uy}-\textbf{X}_p^{22} & -\textbf{X}^{21}L_{u\bar{y}}-L_{\bar{u}y}^\top \bar{\textbf{X}}^{12} & -\textbf{X}^{21}L_{uw} + M_{zy}^\top \textbf{Y}^{21} \\
L_{u\bar{y}}^\top & L_{\bar{u}\bar{y}}^\top & - M_{z\bar{y}}^\top\textbf{Y}^{22} & -L_{u\bar{y}}^\top\textbf{X}^{12}-\bar{\textbf{X}}^{21}L_{\bar{u}y} & -(L_{\bar{u}\bar{y}}^\top \bar{\textbf{X}}^{12} + \bar{\textbf{X}}^{21}L_{\bar{u}\bar{y}}+\bar{\textbf{X}}_{\bar{p}}^{22}) & -\bar{\textbf{X}}^{21} L_{\bar{u}w} + M_{z\bar{y}}^\top \textbf{Y}^{21} \\
L_{uw}^\top & L_{\bar{u}w}^\top & -M_{zw}^\top \textbf{Y}^{22}& -L_{uw}^\top\textbf{X}^{12}+\textbf{Y}^{12}M_{zy} & -L_{\bar{u}w}^\top\bar{\textbf{X}}^{12}+ \textbf{Y}^{12} M_{z\bar{y}} & M_{zw}^\top\textbf{Y}^{21} + \textbf{Y}^{12}M_{zw} + \textbf{Y}^{11}
}>0
\end{equation}
\end{figure*}
\subsection{\textbf{Some General Remarks}}
The following remarks can be made on the synthesis techniques proposed for NSCs 1-4 respectively in Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis}.
\begin{figure*}[!hb]
\centering
\hrulefill
\begin{equation}\label{Eq:Pr:NSC2SynthesisOptimal2}
\bm{
\textbf{X}_p^{11} & \mathbf{0} & L_{uy} & L_{uw} \\
\mathbf{0} & \bar{\rho}\mathbf{I} & M_{zy} & M_{zw} \\
L_{uy}^\top & M_{zy}^\top & -\textbf{X}^{21}L_{uy} - L_{uy}^\top\textbf{X}^{12} -\textbf{X}_p^{22} & -\textbf{X}^{21}L_{uw}+\frac{1}{2}M_{zy}^\top \\
L_{uw}^\top & M_{zw}^\top & -L_{uw}^\top \textbf{X}^{12} + \frac{1}{2} M_{zy} & \frac{1}{2}M_{zw}^\top + \frac{1}{2}M_{zw} - \nu \mathbf{I}
}>0,
\end{equation}
\begin{equation}\label{Eq:Pr:NSC2Synthesis2Alternative}
\begin{aligned}
\bm{
-\textbf{Y}^{22} & -\textbf{Y}^{22}M_{zy} & -\textbf{Y}^{22}M_{zw} \\
-M_{zy}^\top\textbf{Y}^{22} & -L_{uy}^\top \textbf{X}^{12} - \textbf{X}^{21}L_{uy} - \textbf{X}_p^{22} + \Theta^{11} & -\textbf{X}^{21}L_{uw} + M_{zy}^\top \textbf{Y}^{21} + \Theta^{12} \\
-M_{zw}^\top\textbf{Y}^{22} & -L_{uw}^\top \textbf{X}^{12} + \textbf{Y}^{12}M_{zy} + \Theta^{21} & M_{zw}^\top \textbf{Y}^{21} + \textbf{Y}^{12}M_{zw} + \textbf{Y}^{11} + \Theta^{22}
}>0\\
\mbox{ with }
\bm{\Theta^{11} & \Theta^{12} \\ \Theta^{21} & \Theta^{22}} \equiv
\alpha^2 \bm{\textbf{X}_p^{11} & \textbf{X}_p^{11}\\\textbf{X}_p^{11} & \textbf{X}_p^{11}}
-\alpha \bm{
L_{uy}^\top + L_{uy} &
L_{uy}^\top + L_{uw} \\
L_{uw}^\top + L_{uy} &
L_{uw}^\top + L_{uw}
}, \mbox{ and }\alpha \in \mathbb{R}.
\end{aligned}
\end{equation}
\end{figure*}
\begin{remark}\textit{(Partial Topology Design)}\label{Rm:PartialTopologyDesign}
The interconnection matrix synthesizing techniques proposed in Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis}) can be used even when the interconnection matrix is partially known - as it will only reduce the number of variables in the corresponding LMI problems. In other words, the proposed techniques can be used not only to design an interconnection topology from scratch but also to fine-tune an existing interconnection topology (e.g., to determine coupling weights).
\end{remark}
\begin{remark}\textit{(Optimal Synthesis)}\label{Rm:EstimationOFPassivityIndices}
The LMI problem \eqref{Eq:Pr:NSC2Synthesis} (and \eqref{Eq:Pr:NSC4Synthesis1}) proposed for the NSC 2 (and 4) in Props. \ref{Pr:NSC2Synthesis} (and \ref{Pr:NSC4Synthesis}) can be modified to address the problems of synthesizing the \emph{optimal} interconnection matrix $M$ that: minimizes the system stability index L2G($\gamma$) or maximizes passivity indices IFP($\nu$) and OFP($\rho$). For example, for the latter case, the modified LMI problem \eqref{Eq:Pr:NSC2Synthesis} takes the form:
\begin{equation}\label{Eq:Pr:NSC2SynthesisOptimal1}
\begin{aligned}
\max_{\substack {L_{uy},L_{uw},M_{zy},M_{zw},\\ \nu, \rho, \{p_i: i\in\mathbb{N}_N\}}} & c_1 \nu - c_2 \bar{\rho} \\
\mbox{Sub. To: }& \nu > 0, \bar{\rho} > 0, p_i > 0, \forall i\in\mathbb{N}_N, \eqref{Eq:Pr:NSC2SynthesisOptimal2},\\
\end{aligned}
\end{equation}
where $c_1,c_2>0$ are some pre-selected cost coefficients, and $M_{uy} \triangleq (\textbf{X}_p^{11})^{-1} L_{uy}$, $M_{uw} \triangleq (\textbf{X}_p^{11})^{-1} L_{uw}$ and $\rho = 1/\bar{\rho}$. Note that \eqref{Eq:Pr:NSC2SynthesisOptimal2} above has been obtained from \eqref{Eq:Pr:NSC2Synthesis2} via applying $\textbf{Y} = \scriptsize \bm{-\nu \mathbf{I} & \frac{1}{2}\mathbf{I} \\ \frac{1}{2} \mathbf{I} & -\rho \mathbf{I}}$, the congruence principle \cite{Bernstein2009} and the change of variables $\bar{\rho}=1/\rho$.
\end{remark}
\begin{remark}\textit{(Estimating Stability/Passivity Indices)}
When the interconnection matrix is pre-defined, still, solving an optimal synthesis problem like \eqref{Eq:Pr:NSC2SynthesisOptimal1} (with some minor modifications) will reveal the stability/passivity indices of the networked system. The importance of such a knowledge (a dissipativity property) is evident from this paper itself.
\end{remark}
\begin{remark}(Analysis via Synthesis)\label{Rm:AnalysisViaSynthesis}
When the interconnection matrix is predefined, all the synthesis problems formulated in Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis} (with some minor modifications) can still be used to analyze the corresponding networked systems. Note that the aforementioned ``minor modifications'' basically refer to having to reverse the used change of variables (e.g., replacing $L_{uy}$ in \eqref{Eq:Pr:NSC1Synthesis} with $\textbf{X}_p^{11}M_{uy}$). In essence, for the analysis of the NSCs, instead of Props. \ref{Pr:NSC1Stability}-\ref{Pr:NSC4Dissipativity} we can use Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis}. Note also that LMI conditions in Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis} are less convoluted than those in Props. \ref{Pr:NSC1Stability}-\ref{Pr:NSC4Dissipativity}. We will use this to our advantage later on when decentralizing the analysis results proposed in Props. \ref{Pr:NSC1Stability}-\ref{Pr:NSC4Dissipativity}.
\end{remark}
\begin{remark}\textit{(Single Subsystem Case: NSC 1)}\label{Rm:SingleSubsystemCaseNSC1}
Consider the NSC 1 with $N=1$ and a scalar interconnection matrix $M_{uy}=m_{uy}\mathbf{I}$ \eqref{Eq:NSC1Interconnection} where $m_{uy}\in\mathbb{R}$. Now, if the subsystem $\Sigma_1$ is L2G($\gamma_1$), As. \ref{As:PositiveDissipativity} holds and $\textbf{X}_p^{11}=p_1\gamma_1^2\mathbf{I}$, $\textbf{X}_p^{22}=-p_1\mathbf{I}$, $\textbf{X}^{12} = \textbf{X}^{12} = \mathbf{0}$. Thus, the LMI condition \eqref{Eq:Pr:NSC1Synthesis} in Prop. \ref{Pr:NSC1Synthesis} can be written as
$
\scriptsize \bm{ p_1\gamma_1^2\mathbf{I} & l_{uy}\mathbf{I} \\ l_{uy}^\top \mathbf{I}& p_1\mathbf{I}} \normalsize >0 \iff
\scriptsize \bm{ p_1\gamma_1^2 & l_{uy} \\ l_{uy} & p_1} \normalsize >0 \iff
\{p_1 > 0, p_1^2 \gamma_1^2 - l_{uy}^2 > 0 \} \impliedby \{p_1 = k l_{uy}/\gamma_1, l_{uy}>0, k>1\}
$. Now, using the change of variables relationships, we get $m_{uy}=l_{uy}/(p_1\gamma_1^2)=1/k\gamma_1 \iff m_{uy}\gamma_1 = 1/k < 1$, which is the well-known small-gain condition for stability.
\end{remark}
\begin{remark}\textit{(Single Subsystem Case: NSC 2)}\label{Rm:SingleSubsystemCaseNSC2}
Consider the NSC 2 with $N=1$ and a scalar interconnection matrix $M = \scriptsize \bm{m_{uy}\mathbf{I} & m_{uw}\mathbf{I} \\ m_{zy}\mathbf{I} & m_{zw}\mathbf{I}}$ \eqref{Eq:NSC2Interconnection} where $m_{uy},m_{uw},m_{zy},m_{zw}\in\mathbb{R}$. Now, if the subsystem $\Sigma_1$ is L2G($\gamma_1$), As. \ref{As:PositiveDissipativity} holds and $\textbf{X}_p^{11}=p_1\gamma_1^2\mathbf{I}$, $\textbf{X}_p^{22}=-p_1\mathbf{I}$, $\textbf{X}^{12}=\textbf{X}^{21}=\mathbf{0}$. Let the dissipativity specification for the NSC 2 be $\textbf{Y}=\scriptsize\bm{-\nu_1\mathbf{I} & \frac{1}{2}\mathbf{I} \\ \frac{1}{2}\mathbf{I} & -\rho_1\mathbf{I}}$ with some $\nu_1, \rho_1 > 0$. Hence the As. \ref{As:NegativeDissipativity} also holds and the LMI problem \eqref{Eq:Pr:NSC2Synthesis} in Prop. \ref{Pr:NSC2Synthesis} now can be written as
\begin{equation}\label{Eq:Pr:NSC2SynthesisSingleSubsystemCase}
\begin{aligned}
\mbox{Find: }& l_{uy}, l_{uw}, m_{zy}, m_{zw}, \{p_i: i\in\mathbb{N}_N\}\\
\mbox{ Sub. to: }& p_i > 0, \forall i\in\mathbb{N}_N,\\
&\hspace{-2mm}\bm{
p_1\gamma_1^2 & 0 & l_{uy} & l_{uw} \\
0 & \rho_1 & \rho_1 m_{zy} & \rho_1 m_{zw}\\
l_{uy} & \rho_1 m_{zy} & p_1 & \frac{1}{2}m_{zy} \\
l_{uw} & \rho_1 m_{zw} & \frac{1}{2} m_{zy} & m_{zw} - \nu_1
}>0,
\end{aligned}
\end{equation}
where $m_{uy}=l_{uy}/(p_1\gamma_1^2)$ and $m_{uw}=l_{uw}/(p_1\gamma_1^2)$. It can be shown that the above LMI problem is a streamlined version of the interconnection matrix synthesis approach proposed for a single subsystem in \cite[Th. 3]{Xia2014}.
\end{remark}
\begin{remark}(Assumption \ref{As:PositiveDissipativity}) \label{Rm:As:PositiveDissipativityFailure}
For situations where As. \ref{As:PositiveDissipativity} does not hold, a conservative solution (a fix) was proposed in Rm. \ref{Rm:As:PositiveDissipativity}. However, there is an alternative (and less conservative) solution to this issue based on Lm. \ref{Lm:AlternativeLMI_LowerBound}. To illustrate this solution, consider the interconnection matrix synthesis problem for the NSC 2 (originally addressed in Prop. \ref{Pr:NSC2Synthesis}). Now, without loss of generality, assume each subsystem $\Sigma_i, i \in \mathbb{N}_N$ to be $X_i$-EID with $X_i^{11}<0$. Consequently, $X_p^{11} < 0$ and the corresponding LMI condition \eqref{Eq:Pr:NSC2Synthesis2} in Prop. \ref{Pr:NSC2Synthesis} clearly does not hold. This can be resolved soon-after the step \eqref{Eq:Pr:NSC2SynthesisStep1} in the proof of Prop. \ref{Pr:NSC2Synthesis} by using both Lm. \ref{Lm:AlternativeLMI_LowerBound} (to handle $X_p^{11} < 0$) and Lm. \ref{Lm:AlternativeLMI_Schur} (to handle $-\textbf{Y}^{22}>0$) - which leads to the LMI condition \eqref{Eq:Pr:NSC2Synthesis2Alternative}. In all, when As. \ref{As:PositiveDissipativity} does not hold, the LMI condition \eqref{Eq:Pr:NSC2Synthesis2} in Prop. \ref{Pr:NSC2Synthesis} should be replaced by \eqref{Eq:Pr:NSC2Synthesis2Alternative}. Using the same steps, for situations where As. \ref{As:PositiveDissipativity} does not hold, similar solutions can be developed to the other proposed synthesis techniques in Props. \ref{Pr:NSC1Synthesis}, \ref{Pr:NSC3Synthesis} and \ref{Pr:NSC4Synthesis} as well.
\end{remark}
\begin{remark}
(Equilibrium Independence)
Similar to the proposed analysis techniques in Props. \ref{Pr:NSC1Stability}-\ref{Pr:NSC4Dissipativity}, the proposed synthesis techniques in Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis} are also independent of the equilibrium points of the respective NSCs 1-4.
\end{remark}
\section{Decentralized Analysis and Synthesis of Networked Systems}\label{Sec:DecentralizedAnalysisAndSynthesis}
\begin{figure*}[!hb]
\centering
\hrulefill
\begin{equation}\label{Eq:Th:NSC1Stability2}
W_{ij} = \bm{ p_iX_i^{11}e_{ij} & p_iX_i^{11} M_{uy}^{ij} \\
p_jM_{uy}^{ji\top}X_j^{11} &
-p_jM^{ji\top}_{uy}X_j^{12} - p_iX_i^{21}M^{ij}_{uy} - p_iX_i^{22}e_{ij}}
\end{equation}
\begin{equation}\label{Eq:Th:NSC1Synthesis2}
W_{ij} = \bm{ p_iX_i^{11}e_{ij} & L_{uy}^{ij} \\ L_{uy}^{ji\top} & -(p_jL_{uy}^{ji^\top}X_j^{12} + p_iX_i^{21}L_{uy}^{ij} + p_iX_i^{22}e_{ij})}
\end{equation}
\end{figure*}
In this section, we briefly show that the each centralized analysis and synthesis technique proposed earlier in Props. \ref{Pr:NSC1Stability}-\ref{Pr:NSC4Synthesis} for the NSCs 1-4 can be implemented in a decentralized and compositional manner. For this purpose, we will use the Lemmas \ref{Lm:NetworkMatrixProperties}, \ref{Lm:AlternativeLMI_BEW} and \ref{Lm:MainLemma}. First, to simplify the decentralization, we make the following assumption regarding the subsystems.
\begin{assumption}\label{As:SymmetryForDecentralization}
In the NSC 2, the input signal $w \equiv [w_i^\top]^\top_{i\in\mathbb{N}_N}$ where $w_i\in \mathbb{R}^{r_i}$ and the output signal $z \equiv [z_i^\top]^\top_{i\in\mathbb{N}_N}$ where $z_i\in\mathbb{R}^{l_i}$.
Further, in the NSCs 3 and 4, $\bar{N}=N$ and each subsystem $\Sigma_i,i\in\mathbb{N}_N$ has a corresponding (virtual twin) subsystem $\bar{\Sigma}_i,i\in\mathbb{N}_{\bar{N}}$. Furthermore, in the NSC 4, the input signal $w \equiv [w_i^\top,\bar{w}_i^\top]^\top_{i\in\mathbb{N}_N}$ where $\bar{w}_i\in \mathbb{R}^{\bar{r}_i}$ and the output signal $z \equiv [z_i^\top, \bar{z}_i^\top]^\top_{i\in\mathbb{N}_N}$ where $\bar{z}_i\in\mathbb{R}^{\bar{l}_i}$.
\end{assumption}
Note that under As. \ref{As:SymmetryForDecentralization}, each block element of any interconnection matrix $M$ in \eqref{Eq:NSC1Interconnection}-\eqref{Eq:NSC4Interconnection} is a block network matrix. For example, the block element $M_{uy} \equiv [M_{uy}^{ij}]_{i,j\in\mathbb{N}_N}$ of $M$ (appearing in \eqref{Eq:NSC1Interconnection}-\eqref{Eq:NSC4Interconnection}) is a block network matrix. Note that the matrix $M_{uy}^{ij}$ represents how the input $u_i$ of the subsystem $\Sigma_i$ is affected by the output $y_j$ of the subsystem $\Sigma_j$.
Moreover, under As. \ref{As:SymmetryForDecentralization}, each block element of the matrices $\textbf{X}_p$ \eqref{Eq:Pr:NSC1Stability2}, $\bar{\textbf{X}}_{\bar{p}}$ \eqref{Eq:Pr:NSC3Stability2} and $\textbf{Y}$ \eqref{Eq:Pr:NSC2Dissipativity0} are block network matrices. For example, the $(k,l)$\tsup{th} block elements of $\textbf{X}_p$, $\textbf{Y}$ and $\bar{X}_{\bar{p}}, \forall k,l\in\mathbb{N}_2$ are the block network matrices $\textbf{X}_p^{kl}\equiv \text{diag}(p_i X_i^{kl}:i\in\mathbb{N}_N)$, $\bar{\textbf{X}}_{\bar{p}}^{kl}=\text{diag}(\bar{p}_i \bar{X}_i^{kl}:i\in\mathbb{N}_N)$ and $\textbf{Y}^{kl}$, respectively. The matrices $\textbf{X}^{12},\textbf{X}^{21},\bar{\textbf{X}}^{12}$ and $\bar{\textbf{X}}^{21}$ are also block network matrices (recall: $\textbf{X}^{12} \triangleq \text{diag}((X_i^{11})^{-1}X_i^{12}:i\in\mathbb{N}_N)$).
It is worth noting that all the network matrices identified above are also block diagonal except for the block elements of $M$ and $\textbf{Y}$ - which are by default non-block diagonal (i,e., general) network matrices unless specified otherwise.
As pointed out in Rm. \ref{Rm:Lm:MainLemma}, in a network setting, Lm. \ref{Lm:MainLemma} can be used to decentrally analyze (test/enforce) an LMI condition of the form $W>0$ when $W$ is a network matrix. However, if we consider the main LMI conditions derived in Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis} (e.g., see \eqref{Eq:Pr:NSC2Synthesis2}), while they take the form $\Psi>0$, $\Psi$, in any of these cases, is not a network matrix. In fact, in each of these four cases, $\Psi$ is a block-block matrix. Moreover, upon close examination using Lm. \ref{Lm:NetworkMatrixProperties}, it can be seen that each block element of each $\Psi$ is a network matrix (assuming $\textbf{Y}^{12},\textbf{Y}^{21}$ and $\textbf{Y}^{22}$ to be block diagonal network matrices). Therefore, using Lm. \ref{Lm:AlternativeLMI_BEW}, we can replace the condition $\Psi>0$ with an equivalent condition $W \triangleq \text{BEW}(\Psi)>0$. Now, based on Lm. \ref{Lm:NetworkMatrixProperties}, $W$ is a network matrix. In all, using Lemmas \ref{Lm:NetworkMatrixProperties} and \ref{Lm:AlternativeLMI_BEW}, the main LMI conditions derived in Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis} can be transformed to the form $W>0$ where $W$ is a network matrix. Consequently, Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis} (centralized synthesis techniques) can be decentrally implemented exploiting Lm. \ref{Lm:MainLemma}.
Note also that, the same argument is valid for the scenarios where As. \ref{As:PositiveDissipativity} is violated in Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis} (due to Rm. \ref{Rm:As:PositiveDissipativityFailure}). Moreover, as pointed out in Rm. \ref{Rm:AnalysisViaSynthesis}, Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis} can also be used to analyze the corresponding networked systems. Therefore, by extension, such centralized analysis techniques can also be decentrally implemented exploiting Lm. \ref{Lm:MainLemma}.
Based on the above discussion/proof, we now can formally state the decentralized versions of all the centralized analysis and synthesis techniques proposed in Props. \ref{Pr:NSC1Stability}-\ref{Pr:NSC4Synthesis} (as a set of eight theorems). However, due to space constraints, here we limit only to the NSC 1 and provide the decentralized versions of its centralized analysis and synthesis techniques proposed in Pr. \ref{Pr:NSC1Stability} and \ref{Pr:NSC1Synthesis} in the following two theorems.
\begin{theorem}\label{Th:NSC1Stability}
Under As. \ref{As:PositiveDissipativity} and \ref{As:SymmetryForDecentralization}, the NSC 1 is stable if at each subsystem $\Sigma_i, i\in\mathbb{N}_N$, the LMI problem
\begin{equation}\label{Eq:Th:NSC1Stability1}
\begin{aligned}
\mbox{Find: } &p_i \\
\mbox{ such that } &p_i>0, \ \tilde{W}_{ii}>0,
\end{aligned}
\end{equation}
is feasible, where $\tilde{W}_{ii}$ is computed from Lm. \ref{Lm:MainLemma} when analyzing $W=[W_{ij}]_{i,j\in\mathbb{N}_N}>0$ with $W_{ij}$ as in \eqref{Eq:Th:NSC1Stability2}.
\end{theorem}
\begin{theorem}\label{Th:NSC1Synthesis}
Under As. \ref{As:PositiveDissipativity} and \ref{As:SymmetryForDecentralization}, the NSC 1 is stable if at each subsystem $\Sigma_i, i\in\mathbb{N}_N$, the problem
\begin{equation}\label{Eq:Th:NSC1Synthesis1}
\begin{aligned}
\mbox{Find } &p_i, \{L_{uy}^i\} \\
\mbox{ such that } &p_i > 0, \tilde{W}_{ii}>0,
\end{aligned}
\end{equation}
is feasible, where $\tilde{W}_{ii}$ is computed from Lm. \ref{Lm:MainLemma} when analyzing $W=[W_{ij}]_{i,j\in\mathbb{N}_N}>0$ with $W_{ij}$ as in \eqref{Eq:Th:NSC1Synthesis2}. The newly synthesized interconnection matrix elements at subsystem $\Sigma_i$, i.e. $\{M_{uy}^i\}$ can be found using the relationships:
$M_{uy}^{ij} = (p_iX_i^{11})^{-1}L_{uy}^{ij}, \forall j\in\mathbb{N}_{i-1}$,
$M_{uy}^{ji} = (p_jX_j^{11})^{-1}L_{uy}^{ji}, \forall j\in\mathbb{N}_{i-1}$, and
$M_{uy}^{ii} = (p_iX_i^{11})^{-1}L_{uy}^{ii}$.
\end{theorem}
\section{Numerical Results}\label{Sec:NumericalResults}
In this section, we provide a simple numerical example to illustrate the theoretical results. We point out that this numerical example can also be seen as a ``networked'' version of the numerical example considered in \cite{Xia2014}.
\subsection{Subsystems}
As the set of subsystems $\{\Sigma_i:i\in\mathbb{N}_5\}$ (in the NSCs 1-4), we consider a set of stable, non-passive, single-input-single-output (SISO) \emph{controllers} described by the transfer functions: \begin{equation}\label{Eq:NumResSubsystems}
\Sigma_i: G_i(s) = \frac{a_is + b_i}{s+c_i}\exp(-d_i s)
\end{equation}
where
$[a_i]_{i\in\mathbb{N}_5} \equiv [-2,-0.3,-0.2,-1.3,-1.2]$,
$[b_i]_{i\in\mathbb{N}_5} \equiv [1,16,0.2,0.2,1]$,
$[c_i]_{i\in\mathbb{N}_5} \equiv [1,9,1,1,3]$ and
$[d_i]_{i\in\mathbb{N}_5} \equiv [1,1.2,0.8,0.9,1.1]$. Despite the involved delay component in \eqref{Eq:NumResSubsystems}, the $L_2$-gain values of these subsystems can be evaluated as
$[\gamma^2_i]_{i\in\mathbb{N}_5}=[4.00, 3.16, 0.04, 1.69, 1.44]$.
As the set of subsystems $\{\bar{\Sigma}_i:i\in\mathbb{N}_5\}$ (in the NSCs 3-4), we consider a set of unstable, non-passive, SISO \textbf{plants} described by the transfer functions:
\begin{equation}\label{Eq:NumResSubsystemsBar}
\bar{\Sigma}_i: H_i(s) = \frac{a_i s + b_i}{s + c_i}
\end{equation}
where
$[a_i]_{i\in\mathbb{N}_5} \equiv [1,1.5,2,0.5,3]$,
$[b_i]_{i\in\mathbb{N}_5} \equiv [2,3,2,1,1.5]]$ and
$[c_i]_{i\in\mathbb{N}_5} \equiv [-1,-3,-5,-2,-4]$. Due to the linear nature of \eqref{Eq:NumResSubsystemsBar}, output feedback passivity indices of these subsystems can be evaluated directly as $[\bar{\rho}_i]_{i\in\mathbb{N}_5}=[-0.50, -1.00, -2.50, -2.00, -2.67]$.
It is worth noting that even if the subsystems are unknown and non-linear, there are many on-line as well as off-line techniques to directly obtain their dissipativity properties such as the $L_2$-gain values and the passivity indices (e.g., see \cite{Xia2014,Zakeri2019,WelikalaP42022,Arcak2022}).
\subsection{Network Topology}
Note that the known subsystem dissipativity properties (evaluated L2G and OFP indices given above) satisfy the As. \ref{As:PositiveDissipativity}, to synthesize the interconnection matrix $M$ in the NSCs 1-4, we can respectively use the LMIs given in Props. \ref{Pr:NSC1Synthesis}-\ref{Pr:NSC4Synthesis}.
In particular, to guide the synthesis of the interconnection matrix $M$, we assume an underlying undirected \emph{graph topology} described respectively by the adjacency matrix $A=[A_{ij}]_{i,j\in\mathbb{N}_5}$ and a corresponding cost matrix $C=[C_{ij}]_{i,j\in\mathbb{N}_5}$:
\begin{equation}\label{Eq:NumResGraphTopology}
A = \scriptsize \bm{0 & 1 & 1 & 0 & 0 \\ 1 & 0 & 1 & 0 & 0 \\ 1 & 1 & 0 & 1 & 1 \\ 0 & 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 & 0},\ \ \normalsize
C= \scriptsize \bm{100 & 1 & 1 & 10 & 10 \\ 1 & 100 & 1 & 10 & 10 \\ 1 & 1 & 100 & 1 & 1 \\ 10 & 10 & 1 & 100 & 10 \\ 10 & 10 & 1 & 10 & 100}\normalsize.
\end{equation}
Note that $A_{ij}, i,j\in\mathbb{N}_5$ in \eqref{Eq:NumResGraphTopology} indicates whether an interconnection between subsystems $\Sigma_i$ and $\Sigma_j$ is allowed, while $C_{ij}$ in \eqref{Eq:NumResGraphTopology} represents the cost of such an interconnection.
Now, when synthesizing the interconnection matrix $M$ (i.e., solving the LMIs given in Props. \ref{Pr:NSC1Synthesis}, \ref{Pr:NSC2Synthesis}, \ref{Pr:NSC3Synthesis} and \ref{Pr:NSC4Synthesis}), we have two options: (1) use the adjacency matrix $A$ \eqref{Eq:NumResGraphTopology} to impose a \emph{hard graph constraint} that completely restricts the usage of certain elements in $M$, or (2) use the cost matrix $C$ \eqref{Eq:NumResGraphTopology} to impose a \emph{soft graph constraint} that penalizes the usage of certain elements in $M$. Note that these hard and soft graph constraints can simply be included in the interested LMIs problems by: (1) introducing constraints of the form $M_{uy}^{ij}=0$ $\forall i,j\in\mathbb{N}_5$ such that $A_{ij}=0$, or (2) using a cost function of the form $\sum_{i,j\in\mathbb{N}_5}\vert C_{ij}M_{uy}^{ij}\vert^2$, respectively.
\subsection{Observations}
\subsubsection{\textbf{NSC 1}}
First, to validate the Prop. \ref{Pr:NSC1Synthesis} and to illustrate the difference between the aforementioned hard and soft graph constraints, let us consider an example networked system of the form NSC 1 comprised of the subsystems in \eqref{Eq:NumResSubsystems} interconnected via an interconnection matrix $M_{uy}$ \eqref{Eq:NSC1Interconnection}. Figure \ref{Fig:NSC1Simulink} shows the Simulink implementation of this networked system. Note that, to perturb the initial state of the subsystems, we use a step signal as shown in Fig. \ref{Fig:NSC1w}. As shown in Fig. \ref{Fig:NSC1Case1y}, when $M_{uy}=\mathbf{I}$ is used (i.e., unit self loops), the networked system becomes unstable. This motivates the need to synthesize a stabilizing $M_{uy}$ using Prop. \ref{Pr:NSC1Synthesis}. Figures \ref{Fig:NSC1Case3y} and \ref{Fig:NSC1Case4y} show the observed output signals when the optimal $M_{uy}$ values:
\begin{equation*}
M_{uy} = 10^{-1} \times \scriptsize \bm{0 & 0.497 & 1.226 & 0 & 0 \\ 0.408 & 0 & 1.059 & 0 & 0 \\ -0.229 & -0.099 & 0 & 0.318 & 0.508 \\ 0 & 0 & 0.902 & 0 & 0 \\ 0 & 0 & 0.924 & 0 & 0}
\end{equation*}
and
\begin{equation*}
M_{uy} = \scriptsize \bm{0 & 0.099 & 0.043 & 0 & 0 \\ 0.110 & 0 & 0.048 & 0 & 0 \\ 0.967 & 0.966 & 0 & 1.510 & 1.510 \\ 0 & 0 & 0.094 & 0 & 0.011 \\ 0 & 0 & 0.104 & 0.012 & 0}
\end{equation*}
obtained under the hard and soft graph constraint are used, respectively. It is worth noting that when the soft graph constraint is used, an extra interconnection between the subsystems $\Sigma_5$ and $\Sigma_4$ is being used. In both cases, the synthesized $M_{uy}$ successfully stabilizes the networked system.
Note that, in the remaining numerical examples, we have used hard graph constraints to restrict new interconnections and soft graph constraints to penalize self connections.
\begin{figure}[!hb]
\centering
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=0.6\linewidth]{Results/NSC1.png}
\caption{NSC 1: Simulink implementation using the subsystems \eqref{Eq:NumResSubsystems}.}
\label{Fig:NSC1Simulink}
\end{subfigure}\\
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC1w.png}
\caption{Initial excitation signal.}
\label{Fig:NSC1w}
\end{subfigure}
\hfill
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC1Case1y.png}
\caption{Output $y(t)$ with $M=\mathbf{I}$.}
\label{Fig:NSC1Case1y}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC1Case3y.png}
\caption{Output $y(t)$ with optimal $M$ under a hard graph constraint.}
\label{Fig:NSC1Case3y}
\end{subfigure}
\hfill
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC1Case4y.png}
\caption{Output $y(t)$ with optimal $M$ under a soft graph constraint.}
\label{Fig:NSC1Case4y}
\end{subfigure}
\caption{\textbf{NSC 1:}
\textbf{(a)} An example NSC 1 implemented in Simulink by interconnecting the controllers (subsystems) $\{\Sigma_i:i\in\mathbb{N}_5\}$ in \eqref{Eq:NumResSubsystems} via an interconnection matrix $M\in\mathbb{R}^{5\times 5}$.
\textbf{(b)} The excitation signal used to perturb the subsystem initial conditions.
The observed output signals under: \textbf{(c)} $M=\mathbf{I}$, \textbf{(d)} optimally synthesized $M$ subject to a hard graph constraint, and \textbf{(e)} optimally synthesized $M$ subject to soft graph constraint.}
\label{Fig:NSC1}
\end{figure}
\subsubsection{\textbf{NSC 2}}
To validate the Prop. \ref{Pr:NSC2Synthesis}, we next use the previous networked system (shown in Fig. \ref{Fig:NSC1Simulink}) with an added input port ($w$) and an output port ($z$) - making a networked system of the form NSC 2 shown in Fig. \ref{Fig:NSC2Simulink}. This networked system, as shown in Fig. \ref{Fig:NSC2Case1w}, is unstable under the interconnection matrix choice $M= \scriptsize \bm{\mathbf{I} & \mathbf{I} \\ \mathbf{I} & \mathbf{I}}$ \eqref{Eq:NSC2Interconnection}. Note also that it is desirable to make this networked system passive as it then can be used to control another non-passive networked system. This motivates the need to synthesize $M$ such that this networked system (shown in Fig. \ref{Fig:NSC2Simulink}) is maximally passive (from $w$ to $z$). For this purpose, we used Prop. \ref{Pr:NSC2Synthesis} (see also Rm. \ref{Rm:EstimationOFPassivityIndices}) and synthesized $M$ that optimized the input feedforward and output feedback passivity indices (i.e., $\nu$ and $\rho$, respectively) of the networked system. The obtained optimal passivity indices are as follows:
\begin{equation}
\nu^* = 0 \ \ \mbox{ and } \ \ \rho^*=4.78.
\end{equation}
The output trajectories of the passivated networked system are shown in Figs. \ref{Fig:NSC2Case2y} and \ref{Fig:NSC2Case2z}.
To further verify the obtained passivity measures of the networked system, we next consider the simple example scenario shown in Fig. \ref{Fig:NSC2ApplicationSimulink}. In there, the passivated networked system (obtained above) is connected in feedback with a memoryless system $K_{sys}\mathbf{I}$ where $K_{sys}\in\mathbb{R}$. According to \cite[Th. 4]{Xia2014}, the composite system is passive if $K_{sys}+\rho^*>0$, or non-passive otherwise. Figures \ref{Fig:NSC2Case3z} and \ref{Fig:NSC2Case4z} show output trajectories observed when $K_{sys}$ is selected such that $K_{sys}+\rho^* = 1$ (hence passive) and $K_{sys}+\rho^* = -1$ (hence non-passive), respectively. Their opposite nature implies the change of the passivity property of the composite system.
\begin{figure}[!t]
\centering
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=0.75\linewidth]{Results/NSC2.png}
\caption{NSC 2: Simulink implementation using the subsystems \eqref{Eq:NumResSubsystems}.}
\label{Fig:NSC2Simulink}
\end{subfigure}\\
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC2Case1w.png}
\caption{Exogenous input $w(t)$.}
\label{Fig:NSC2Case1w}
\end{subfigure}
\hfill
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC2Case1z.png}
\caption{Output with $M= \scriptsize \bm{\mathbf{I} & \mathbf{I} \\ \mathbf{I} & \mathbf{I}}$.}
\label{Fig:NSC2Case1z}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC2Case2y.png}
\caption{Subsystem outputs with optimal $M$.}
\label{Fig:NSC2Case2y}
\end{subfigure}
\hfill
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC2Case2z.png}
\caption{Output with optimal $M$.}
\label{Fig:NSC2Case2z}
\end{subfigure}
\caption{\textbf{NSC 2:} \textbf{(a)} An example NSC 2 implemented in Simulink using the controllers \eqref{Eq:NumResSubsystems} and an interconnection matrix $M\in\mathbb{R}^{10 \times 10}$. \textbf{(b)} Used exogenous input signal (in addition to the initial excitation signal shown in Fig. \ref{Fig:NSC1w}). \textbf{(c)} Observed output under $M= \scriptsize \bm{\mathbf{I} & \mathbf{I} \\ \mathbf{I} & \mathbf{I}}$. Observed: \textbf{(d)} subsystem outputs and \textbf{(e)} the networked system output under optimally synthesized $M$.}
\label{Fig:NSC2}
\end{figure}
\begin{figure}[!t]
\centering
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=0.75\linewidth]{Results/NSC2Application.png}
\caption{NSC 2 in Fig. \ref{Fig:NSC2} with a feedback connection.}
\label{Fig:NSC2ApplicationSimulink}
\end{subfigure}\\
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC2Case3z.png}
\caption{Output $z(t)$ with $K_{sys} = - \rho^* + 1$ (passive).}
\label{Fig:NSC2Case3z}
\end{subfigure}
\hfill
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC2Case4z.png}
\caption{Output $z(t)$ with $K_{sys} = - \rho^* - 1$ (non-passive).}
\label{Fig:NSC2Case4z}
\end{subfigure}
\caption{\textbf{NSC 2 Application:} A verification of the obtained passivity properties of the networked system.}
\label{Fig:NSC2Application}
\end{figure}
\begin{figure}[!t]
\centering
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=0.75\linewidth]{Results/NSC3.png}
\caption{NSC 3: Simulink implementation using the subsystems \eqref{Eq:NumResSubsystems} (as controllers) and \ref{Eq:NumResSubsystemsBar} (as plants).}
\label{Fig:NSC3Simulink}
\end{subfigure}\\
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC3w.png}
\caption{Initial excitation signal.}
\label{Fig:NSC3w}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC3Case1y.png}
\caption{Output with $M=\scriptsize\bm{\mathbf{0} & \mathbf{I} \\ \mathbf{I} & \mathbf{0}}$.}
\label{Fig:NSC3Case1y}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC3Case2y.png}
\caption{Output $y$ with optimal $M$.}
\label{Fig:NSC3Case2y}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC3Case2yBar.png}
\caption{Output $\bar{y}$ with optimal $M$.}
\label{Fig:NSC3Case2yBar}
\end{subfigure}
\caption{\textbf{NSC 3:} \textbf{(a)} An example NSC 3 implemented in Simulink by interconnecting the controllers \eqref{Eq:NumResSubsystems} and the plants \eqref{Eq:NumResSubsystemsBar} via an interconnection matrix $M\in\mathbb{R}^{10 \times 10}$ \eqref{Eq:NSC3Interconnection}. \textbf{(b)} The excitation signal used to purturb the subsystem initial conditions. \textbf{(c)} Observed controller output signal $y(t)$ under $M=\scriptsize\bm{\mathbf{0} & \mathbf{I} \\ \mathbf{I} & \mathbf{0}}$. Observed: \textbf{(d)} controller and \textbf{(e)} plant output signals under optimally synthesized $M$.
}
\label{Fig:NSC3}
\end{figure}
\begin{figure}[!t]
\centering
\begin{subfigure}{0.48\textwidth}
\centering
\includegraphics[width=0.75\linewidth]{Results/NSC4.png}
\caption{NSC 4: Simulink implementation using the subsystems \eqref{Eq:NumResSubsystems} (as controllers) and \eqref{Eq:NumResSubsystemsBar} (as plants).}
\label{Fig:NSC4Simulink}
\end{subfigure}\\
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC4w.png}
\caption{Exogenous input $w(t)$.}
\label{Fig:NSC4w}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC4Case1z.png}
\caption{Output with $M=\scriptsize \bm{\mathbf{0} & \mathbf{0} & \mathbf{I} \\ \mathbf{I} & \mathbf{0} & \mathbf{0} \\ \mathbf{I} & -\mathbf{I} & \mathbf{0}}$.}
\label{Fig:NSC4Case1z}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC4Case2y.png}
\caption{Subsystem outputs with optimal $M$.}
\label{Fig:NSC4Case2y}
\end{subfigure}
\begin{subfigure}{0.23\textwidth}
\includegraphics[width=\linewidth]{Results/NSC4Case2z.png}
\caption{Outputs with optimal $M$.}
\label{Fig:NSC4Case2z}
\end{subfigure}
\caption{\textbf{NSC 4:} \textbf{(a)} An example NSC 4 implemented in matlab using the controllers \eqref{Eq:NumResSubsystems}, plants \eqref{Eq:NumResSubsystemsBar} and an interconnection matrix $M\in\mathbb{R}^{15\times15}$. \textbf{(b)} Used exogenous input signal (in addition to the initial excitation signal shown in Fig. \ref{Fig:NSC3w}). \textbf{(c)} Observed output under $M=\scriptsize \bm{\mathbf{0} & \mathbf{0} & \mathbf{I} \\ \mathbf{I} & \mathbf{0} & \mathbf{0} \\ \mathbf{I} & -\mathbf{I} & \mathbf{0}}$.
Observed: \textbf{(d)} subsystem outputs $y(t)$ and \textbf{(e)} networked system output $z(t)$, under optimally synthesized $M$.}
\label{Fig:NSC4}
\end{figure}
\subsubsection{\textbf{NSC 3}}
Next, to validate the Prop. \ref{Pr:NSC3Synthesis}, we use the previous networked system (shown in Fig. \ref{Fig:NSC2Simulink}) and connect it to the collection of subsystems (i.e., plants \eqref{Eq:NumResSubsystemsBar}) - making a networked system of the form NSC 3 as shown in Fig. \ref{Fig:NSC3Simulink}. As shown in Fig. \ref{Fig:NSC3Case1y}, this networked system is unstable under the interconnection matrix choice $M= \scriptsize \bm{\mathbf{I} & \mathbf{I} \\ \mathbf{I} & \mathbf{I}}$ \eqref{Eq:NSC3Interconnection}. This motivates the need to synthesize $M$ such that it stabilizes the networked system (i.e., stabilizes the group of unstable plants \eqref{Eq:NumResSubsystemsBar} via the group of controllers \eqref{Eq:NumResSubsystems}). Figures \ref{Fig:NSC3Case2y} and \ref{Fig:NSC3Case2yBar} show the observed output signals when the synthesized $M$ provided by the Prop. \ref{Pr:NSC3Synthesis} is used to interconnect the involved subsystems (i.e., plants and controllers).
\subsubsection{\textbf{NSC 4}}
Finally, to validate the Prop. \ref{Pr:NSC4Synthesis}, we use the previous networked system (shown in Fig. \ref{Fig:NSC3Simulink}) with an added input port ($w$) and an output port ($z$) - making a networked system of the form NSC 4 as shown in Fig. \ref{Fig:NSC4Simulink}. When the interconnection matrix is chosen as $M=\scriptsize \bm{\mathbf{0} & \mathbf{0} & \mathbf{I} \\ \mathbf{I} & \mathbf{0} & \mathbf{0} \\ \mathbf{I} & -\mathbf{I} & \mathbf{0}}$, as shown in Fig. \ref{Fig:NSC4Case1z}, this networked system amplifies the input signal eventually leading to an unstable behavior. To counter this, we need to synthesize $M$ such that the networked system become finite-gain $L_2$ stable with a gain $\gamma<1$. For this purpose, we used Prop. \eqref{Pr:NSC4Synthesis} and synthesized $M$ that minimizes the $L_2$-gain value of the networked system. Note that, to simplify the problem, we used the ``approximate Simulation'' $M$ matrix format given in Tab. \ref{Tab:Configurations}. The obtained optimal $L_2$-gain value is
\begin{equation}
\gamma^* = 0.3351.
\end{equation}
The output trajectories of the resulting finite-gain $L_2$ stable networked system are shown in Figs. \ref{Fig:NSC4Case2y} and \ref{Fig:NSC4Case2z}.
\section{Conclusion}\label{Sec:Conclusion}
In this paper, we considered several widely occurring networked system configurations - each comprised of an interconnected set of non-linear subsystems with known dissipativity properties. For these NSCs, exploiting the known subsystem dissipativity properties, centralized network analysis and network topology synthesis techniques were developed as LMI problems. Then, for these centralized techniques, decentralized counterparts were developed so that they can be executed in a decentralized and compositional manner (i.e., in a way that new/existing subsystems can be added/removed conveniently). Since each proposed technique only uses subsystem dissipativity properties and takes the form of an LMI problem, they can be efficiently implemented and solved. Consequently, the proposed techniques in this paper are ideal for higher-level large-scale networked system design. Finally, we discussed several illustrative numerical examples. Future work aims to extend the proposed techniques for switched and hybrid networked systems.
\bibliographystyle{IEEEtran}
| 7234eab24de1eaa6381dcc7f013c6f86771eaf22 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
About $13$ million people in Bangladesh are suffering from different degrees of hearing loss, of which $3$ million have hearing disability~\cite{alauddin2004deafness}.
There are around $1$ million using Bangladeshi Sign Language (BdSL) in their everyday life\cite{c2}. While communicating with a signer
, there are two major tasks for a non-signer: \textit{(i)} understanding the signs and \textit{(ii)} expressing the signs. Researchers made impressive contributions to task~\textit{(i)} by developing sign letters\footnote{signs that represent letters only.} recognition techniques from images (Fig.~\ref{fig:teaser}~\protect\includegraphics[scale=.18]{crc1.png}). Several works has been proposed for BdSL letters classification via machine learning techniques~\cite{islam2022improving, rahim2022soft, miah2022bensignnet, hasan2021shongket, khatun2021systematic, talukder2021okkhornama, Hoque_2020_ACCV, BdSLiciet}.
Task \textit{(ii)} still has less research attention since it is a difficult process for non-signers. A naive and tiresome approach to expressing signs is to use flashcards with signs and symbols. Being inspired by that,~\cite{shishir2020esharagan} proposed a system that generates symbols of signs using generative adversarial networks (GANs)~\cite{goodfellow2014generative}. However, their work only produces symbols of signs---which may raise questions regarding the necessity of such a system. Another way is by using animated avatars of signs, i.e.~\cite{KippAvat}; but there is no such system for BdSL. Moreover, an avatar-based system does not provide a realistic environment for communication.
All the above scenarios inspired us to make task~\textit{(ii)} more realistic yet effortless. Hence, we introduced \textit{\textbf{PerSign}: {Pers}onalized Bangladeshi {Sign} Letters Synthesis} which converts the image of a user into an image showing signs while keeping the person's profile unchanged. Fig.~\ref{fig:teaser}~\protect\includegraphics[scale=.19]{crc2.png} explains the working pipeline of our prototype. A user first uploads their profile photo ($I_P$) only once to our system. The $I_P$ must follow a specific rule of showing hand and palm (as shown in~\protect\includegraphics[scale=.19]{crc2.png}{$^{a}$}). After that, the user inserts the desired letter ($L$) to be expressed (e.g. \protect\includegraphics[scale=.17]{Ga.png} in \protect\includegraphics[scale=.19]{crc2.png}{$^{b}$}). Our system converts $I_P$ into $I_L$ by considering $L$. This can be seen as $I_L \leftarrow I_P+L$, where $I_L$ contains the same person in $I_P$ with unchanged face, skin tone, attire, and background (\protect\includegraphics[scale=.19]{crc2.png}{$^{c}$}). In that case, the person does not need any expertise in sign language. We believe, a signer will feel a natural environment if $I_L$ is shown, thus, making the communication more realistic and affectionate.
\textbf{\textit{\color{blue}Do we really need such a system?}}---in order to address this question, we performed a survey on a group of $6$ {guardians} and $11$ {teachers} of deaf children---who were also signers---regarding the necessity of our system.
We let participants upload profile images to \textit{PerSign} and asked them to rate the results on a scale of \textcircled{\small 1} to \textcircled{\small 5} according to \textit{Likert} rating method~\cite{likert1932technique}, with \textcircled{\small 1} being \textit{not necessary at all} and \textcircled{\small 5} being \textit{very necessary}. Out of total $17$ participants, $13$ and $3$ rated \textit{PerSign} with \textcircled{\small 5} and \textcircled{\small 4} respectively with an average rating of $4.705$. Most of the sign language teachers commented that \textit{personalized} signs are very helpful for general people to get closer to signers, especially when it comes to children.
\section{Implementation}
We employed the \textit{Generative Adversarial Network} (Fig.~\ref{fig:teaser}~\protect\includegraphics[scale=.18]{crc3.png}), an unsupervised deep learning technique that automatically learns the patterns from datasets in order for the model to produce a new output~\cite{goodfellow2014generative}. Our problem lies under a sub-domain of GAN which \textit{image-to-image translation}~\cite{isola2017image} and we adopted \textit{GestureGAN}~\cite{tang2018gesturegan}---a gesture-to-gesture translation method---to implement a prototype system. For this purpose, we built a dataset of images with hand gestures of arbitrary poses, sizes, and backgrounds.
Since we needed a paired dataset $\{I_P, I_L\}$ we could not reuse any of the existing unpaired ones. For localizing gestures and appearances, we exploited \textit{OpenPose}~\cite{8765346} to make the skeleton of the hands and face of the images and store the pair of input and output images. We then trained and tested the \textit{GestureGAN} model with our dataset, and constructed our system's working prototype. Fig.~\ref{fig:res} presents more results of \textit{PerSign}.
\begin{figure}[htb]
\includegraphics[width=0.49\textwidth]{res2.png}
\caption{Result analysis. (a) input profile image. (b) generated image with signs. (c) zoomed view of faces from input \textit{(left)} and output \textit{(right)}. We can see, that the face is retained.}
\label{fig:res}
\Description {A picture of the result obtained exclusively showing the face and hand-gestures.}
\end{figure}
\section{Conclusion and Future work }
In this poster, we proposed a framework---\textit{PerSign}---for synthesizing Bangladeshi Sign Letters that can be \textit{personalized}. Through our method, anyone will be able to communicate as a signer without having any expertise in BdSL. We built our own dataset and exploited \textit{GestureGAN} method to accomplish the task. Our work is still in progress and we gathered comments from the participants during the survey to find areas for improvement. Most of the users recommended applying the technique for a sequence of images to render video for a stream of input letters. We can achieve better results by increasing the number of diverse examples in our dataset. Though our dataset has samples for all BdSL letters, some specific letters need to be treated carefully because of their similarities in patterns.
Our final aim is to merge task (\textit{i})
and (\textit{ii})
into a single system to provide an \textit{one-stop solution} for two-way communication. Some of the users also suggested extending our work for full gestures, rather than letters only. We are also intent on implementing better \textit{GUI} with voice input. Last but not the least, we plan to conduct a thorough evaluation from experts and signers. We believe, this poster opens new avenues in sign language for further research.
\newpage
\newpage
\bibliographystyle{ACM-Reference-Format}
| ddd94a2507f4dbc4d448f22d1fff4855e81a4d64 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The aim of this paper is to highlight the combinatorial structure of a problem coming from symplectic geometry, that is, finding almost-complex structures on a given symplectic manifold satisfying a geometric and differential property. We will restrict to a particular class of symplectic manifolds that are called \emph{adjoint orbits}, that are diffeomorphic to a quotient $G/V$ where $G$ is a semisimple Lie group and $V$ a compact subgroup. On these manifolds one can choose a symplectic structure and an almost-complex structure that are \emph{invariant} by the group action of $G$ on $G/V$, meaning that these structures become algebraic objects on the Lie algebra $\mathfrak{g}$ of $G$. By the assumption of $G$ being semisimple, the symplectic and almost-complex structures can be studied in a purely Lie-theoretical way. In turn, also the differential properties of interest concerning the almost-complex structure can be formulated in a Lie theory framework. This is the setting studied in \cite{DellaVedovaGatti2022}, where the authors formulated the condition in terms of the root system of the underlying Lie algebra $\mathfrak{g}$. In this work, we take a step further and we rephrase the condition solely in terms of the combinatorics of the Vogan diagram of $\mathfrak{g}$.
The problem is explained with more details in the following. Let $(M,\omega)$ be a symplectic manifold, and let $J$ be a compatible almost-complex structure on it. This means that $J$ is an endomorphism of the tangent bundle of $M$, $TM$, such that $J^2=$id and satisfying $\omega(JX,JY)=\omega(X,Y)$ and $\omega(JX,X)>0$, for all vector fields $X,Y\in TM$. Manifolds equipped with these structures are called \emph{almost-K\"ahler manifolds}. When, in addition, $J$ is integrable, such manifolds turn out to be complex manifolds and are called \emph{K\"ahler manifolds}. One can define a closed two-form $\rho$ on $M$ associated with $J$ in the following way. Let $\nabla$ denote the Chern connection of $J$. The curvature $R$ of $\nabla$ is a two-form with values in $\End(TM)$, and $\rho$ is defined by the identity $\rho(X,Y) = \mathrm{tr}(JR(X,Y))$, for $X,Y$ vector fields on $M$. In particular, an open problem is to find when the form $\rho$ is a multiple of the symplectic form $\omega$, i.e.,
\begin{equation}\label{eq:rholambdaomega}
\rho=\lambda\omega,\quad \lambda\in\mathbb{R}.
\end{equation}
This condition has been extensively studied in several works, especially on K\"ahler manifolds \cite{Calabi1957,Yau1977,Yau1978,DonaldsonChenSun2015I,DonaldsonChenSun2015II,DonaldsonChenSun2015III,Tian1997,Tian1990,Tian2015}, just to name a few references. In this setting, when the condition \eqref{eq:rholambdaomega} is satisfied for some $\lambda\in\mathbb{R}$, the manifold is said to be \emph{K\"ahler-Einstein}. On the other hand, very little is known about equation \eqref{eq:rholambdaomega} for almost-K\"ahler manifolds. Hence it turns out to be important to look for examples of symplectic non-complex manifolds satisfying this equation. Almost-K\"ahler manifolds with this property will be called \emph{special}.
In this work, we focus on manifolds $M$ that are \emph{homogeneous spaces}, i.e., diffeomorphic to a quotient $G/V$ of a Lie group $G$ by a subgroup $V$. On such manifolds one can choose a symplectic and an almost-complex structure that are invariant by the action of $G$. In this case the condition $\rho=\lambda\omega$ becomes an algebraic condition on the Lie algebra $\mathfrak{g}$ of the group $G$, which greatly simplify the computations. Moreover, when $G$ is assumed to be a real non-compact semisimple Lie group and $V$ a compact subgroup, it turns out that the form $\rho$ can be expressed entirely in terms of the root system of the real semisimple Lie algebra $\mathfrak{g}$. It is possible to list all classical adjoint orbits satisfying \eqref{eq:rholambdaomega} having Lie algebra with rank up to $4$, and all the exceptional ones \cite{DellaVedovaGatti2022}. In particular, adjoint orbits $G/V$ with $G$ an exceptional simple Lie group satisfying \eqref{eq:rholambdaomega} are fully classified. A key ingredient in proving these results are Vogan diagrams \cite[Ch.~VI,Sec.~8]{Knapp1996}, which are combinatorial tools akin to Dynkin diagrams for non-compact real semisimple Lie algebras. Through Vogan diagrams it is possible express the equation \eqref{eq:rholambdaomega} on adjoint orbits of semisimple Lie groups in terms of the root system associated to the real non-compact semisimple Lie algebra $\mathfrak{g}$. In this work we write the equation only in terms of the indices of simple non-compact roots of the Vogan diagram. The results are quite technical, so we briefly summarize the content below.
\begin{thm}
Given a classical Vogan diagram of rank $\ell$ with $S$ the set of indices of non-compact simple roots, we provide explicit formulae to determine when equation $\rho=\lambda\omega$ is satisfied. These formulae depend only on the combinatorics of the set $S$.
\end{thm}
The results that build the above theorem also shed some light on the combinatorics of Vogan diagrams and how to count certain compact roots in real non-compact semisimple Lie algebras. In addition, they allow to implement a faster algorithm to verify when equation \eqref{eq:rholambdaomega} is satisfied on these orbits for some $\lambda\in\mathbb{R}$.
We will prove the results for classical simple Lie algebras only, since these can be trivially adapted to semisimple Lie algebras. In general it would be very interesting to find a unifying way of writing the combinatorial formulae, since the current results require to consider each family of classical Lie algebras $A$, $B$, $C$ and $D$ separately. For example, one can define an operator acting on Vogan diagrams or particular subgraphs and analyzing its properties, like its spectrum. We leave this study for future work.
The paper is organized as follows. In section \ref{sec:specialAKadjointorbits} we recall the relevant background material presented in \cite{DellaVedovaGatti2022}, omitting all the proofs. In section \ref{sec:specialCVD} we prove the results concerning equation \eqref{eq:rholambdaomega} and the combinatorics of Vogan diagrams. More precisely, we give explicit formulae to check if the equation \eqref{eq:rholambdaomega} holds with purely combinatorial arguments, for each family of classical Lie algebras $A$ \ref{sec:anCoefficients}, $B$ \ref{sec:bnCoefficients}, $C$ \ref{sec:cnCoefficients}, $D$ \ref{sec:dnCoefficients}.
\section{Special almost-K\"ahler adjoint orbits}\label{sec:specialAKadjointorbits}
In this section we recall definitions and results needed for the computations in section \ref{sec::spacialitycondition}. In particular, some facts about adjoint orbits of semisimple Lie groups, the structure theory of real semisimple Lie algebras and the algebraic equation \eqref{eq:rholambdaomega} for adjoint orbits of semisimple Lie groups. We will omit the proofs and the details of all the results, as they are already contained in \cite{DellaVedovaGatti2022}. Additional material on the general theory can be found in \cite{Knapp1996,Helgason1978,GriffithsSchmid1969,DellaVedova2019}.
\subsection{Adjoint orbits}\label{sec:adjointOrbits}
Let $G$ be a Lie group and denote by $\mathfrak g$ its Lie algebra. Then $G$ acts on $\mathfrak{g}$ by the \emph{adjoint action of} $G$ \emph{on} $\mathfrak{g}$
\begin{equation}
\text{Ad}:G\to \text{Aut}(\mathfrak{g}),\quad g\mapsto \text{Ad}_g,
\end{equation}
where $\text{Ad}_g$ is the differential at the identity $e\in G$ of the conjugation $h\to ghg^{-1}$, $\forall h\in G$. By differentiating the above map, one gets
\begin{equation}\label{eq:adjointaction}
\text{ad}:\mathfrak{g}\to\text{Der}(\mathfrak{g}),\quad x\mapsto\text{ad}_x,
\end{equation}
and $\text{ad}_x(y)=[x,y]$, where $[,]$ denotes the commutator and $\text{Der}(\mathfrak{g})$ is the Lie algebra of $\text{Aut}(\mathfrak{g})$. Also in this case we call the map \eqref{eq:adjointaction} \emph{adjoint action of} $\mathfrak{g}$ \emph{on itself}.
Assume that $G$ is, in addition, non-compact, real and semisimple. Let $v \in \mathfrak g$ be a chosen element such that its stabilizer $V$ is a compact subgroup $V \subset G$. Note that the orbit of $v$ under the adjoint action, also called \emph{adjoint orbit} of $v$, is diffeomorphic to $G/V$, by the orbit-stabilizer theorem. By definition of adjoint action, the Lie algebra of the stabilizer $V$ is
\begin{equation}
\mathfrak v = \{x \in \mathfrak g \ \vert \ [v,x]=0 \}.
\end{equation}
The \emph{Killing form} $B(x,y) = \mathrm{tr}(\mathrm{ad}(x) \mathrm{ad}(y))$, $x,y \in \mathfrak g$, is a bilinear symmetric form defined on a Lie algebra $\mathfrak{g}$. In particular, since $V$ is compact, it holds that $B$ restricts to a negative definite scalar product on the Lie algebra $\mathfrak v$ of $V$ \cite[Ch.~VI, Sec.~1]{Knapp1996}, while the orthogonal complement
\begin{equation}
\mathfrak m = \{x \in \mathfrak g \ \vert \ B(x,y)=0 \mbox{ for all } y \in \mathfrak v\}
\end{equation}
is canonically isomorphic to the tangent space at the identity coset $e$ of the adjoint orbit $G/V$ of $v$.
Recall that $\mathfrak{g}$ is semisimple, thus the Killing form is non-degenerate, by Cartan criterion of semisemplicity \cite[Ch.~1, Sec.~7]{Knapp1996}. So $B$ induces a canonical isomorphism between $\mathfrak g$ and its dual $\mathfrak g^*$.
As a consequence, $G/V$ turns out to be equipped with the so called \emph{Kirillov-Kostant-Souriau symplectic form} $\omega$ \cite[Sec.~1.2]{Kirillov2004}, which is $G$-invariant and, at the identity coset $e$, corresponds to the symplectic form (i.e., antisymmetric and non-degenerate) $\sigma$ on $\mathfrak m$ defined by
\begin{equation}
\sigma(x,y) = B(v,[x,y]) \quad x,y \in \mathfrak m.
\end{equation}
The relationship between the symplectic form $\omega$ on $G/V$ and $\sigma$ is analyzed in detail in \cite[Section 3]{DellaVedova2019}.
\subsection{Structure of the Lie algebra $\mathfrak g$}\label{sec::structureofg}
Given a complex Lie algebra $\mathfrak g_c$ obtained by complexification of $\mathfrak g$, there exists a unique complex conjugation $\tau$ on $\mathfrak{g}_c$ that fixes $\mathfrak g \subset \mathfrak g_c$. Let $\mathfrak k \subset \mathfrak g$ be a maximal compact subalgebra such that $\mathfrak v \subset \mathfrak k$. Then, considering the complexification $\mathfrak k_c \subset \mathfrak g_c$ of $\mathfrak k$ and its $\mathrm{ad}(\mathfrak k_c)$-invariant complement $\mathfrak p_c$ yields a decomposition $\mathfrak g = \mathfrak k \oplus \mathfrak p$, where $\mathfrak p = \mathfrak p_c \cap \mathfrak g$. This decomposition is called \emph{Cartan decomposition of} $\mathfrak{g}$.
Choose a maximal abelian subalgebra $\mathfrak{h}_0\subset\mathfrak{k}$ that contains $v$ and call $\mathfrak{h}_c\subset\mathfrak{g}_c$ its complexification. Then the adjoint representation of $\mathfrak h_c$ on $\mathfrak g_c$ produces a \emph{root space decomposition}
\begin{equation}
\mathfrak g_c = \mathfrak h_c \oplus \sum_{\alpha \in \Delta} \mathfrak g^\alpha,
\end{equation}
where the set of roots $\Delta$ is a finite subset of the dual space of $\mathfrak h_c$, and each root space,
\begin{equation}\label{eq::root-space}
\mathfrak g^\alpha = \left\{ x \in \mathfrak g_c \ \vert \ [h,x] = \alpha(h)x \mbox{ for all } h \in \mathfrak h_c \right\}
\end{equation}
has dimension one. For more details about the properties of root spaces and root space decompositions see \cite[Sec.~9.2]{Humphreys1972}. In particular, any root space $\mathfrak g^\alpha$ is contained either in $\mathfrak k_c$ or in $\mathfrak p_c$, and the root $\alpha$ is called \emph{compact} if its root space is contained in $\mathfrak k_c$ and \emph{non-compact} if it is contained in $\mathfrak p_c$.
Define the coefficients $\varepsilon_\alpha = -1$ if $\alpha$ is compact and $\varepsilon_\alpha = 1$ otherwise, for $\alpha\in \Delta$. In our case, it turns out to be convenient to choose always \emph{positive root systems}, where a positive root system is a subset $\Delta_+ \subset \Delta$ such that
\begin{enumerate}\renewcommand{\labelenumi}{\alph{enumi})}
\item for all $\alpha \in \Delta$, either $\alpha$ or $-\alpha$ belongs to $\Delta_+$,
\item if $\alpha, \beta \in \Delta_+$ and $\alpha+\beta \in \Delta$, then $\alpha + \beta \in \Delta_+$.
\end{enumerate}
A positive root is called simple if it cannot be written as a sum $\alpha + \beta$ where $\alpha, \beta \in \Delta_+$.
Once a positive root system $\Delta_+$ is fixed, the set of simple roots $\Sigma^+ \subset \Delta_+$ turns out to be a basis for $\mathfrak h_{\mathbf R}^*=(i\mathfrak{h}_0)^*$, since one can consider $\Delta$ as a subspace of $\mathfrak h_{\mathbf R}^*$. In addition, if $\alpha\in \Delta_+$ is a root, then it can be written as $\alpha = \sum_{\gamma \in \Sigma^+} n_\gamma \gamma$, with the coefficients $n_{\gamma}$ all positive integers.
Denote by $\Sigma^+_c = \{ \gamma \in \Sigma^+ \, | \, \varepsilon_\gamma = -1 \}$ the set of simple compact roots and by $\Sigma^+_n = \{ \gamma \in \Sigma^+ \, | \, \varepsilon_\gamma = 1 \}$ the set of simple non-compact roots. Then the set of simple roots can be decomposed as $\Sigma^+ = \Sigma^+_c \cup \Sigma^+_n$. It is possible to determine whether a root is compact or not by looking at the compactness of the simple roots $\gamma\in\Sigma^+$ and the coefficients $n_{\gamma}$'s, as the next result shows.
\begin{lem}\label{cor::epsilonalphaformula}
If a positive root $\alpha\in\Delta_+$ has the form $\alpha = \sum_{\gamma \in \Sigma^+} n_\gamma \gamma$, then
\begin{equation}
\varepsilon_\alpha = (-1)^{1+\sum_{\gamma \in \Sigma^+_n}n_\gamma}.
\end{equation}
\end{lem}
Remember that $\mathfrak{h}_0\subset \mathfrak{k}$, with $\mathfrak{k}$ compact, hence $B$ restricts to a positive scalar product on $\mathfrak h_{\mathbf R}=i\mathfrak{h}_0$.
As a consequence there is an isomorphism between $\mathfrak h_{\mathbf R}^*$ and $\mathfrak h_{\mathbf R}$ which takes $\psi\in\mathfrak{h}_{\mathbf{R}}^*$ to the unique $h_\psi\in\mathfrak{h}_{\mathbf{R}}$ such that $\psi(h) = B(h_\psi,h)$, for all $h \in \mathfrak h_{\mathbf R}$.
Thus, it is possible to define a positive scalar product on $\mathfrak h_{\mathbf R}^*$ as $(\psi,\psi') = B(h_{\psi},h_{\psi'})$.
This scalar product defined on $\mathfrak h_{\mathbf R}^*$ allows to define particular subspaces of $\mathfrak h_{\mathbf R}^*$, called \emph{Weyl chambers}, in the following way. Consider the set of hyperplanes $P_\alpha = \{ \psi \in \mathfrak h_{\mathbf R}^* \,|\, (\psi,\alpha)=0\}$, with $\alpha \in \Delta$. These hyperplanes divide $\mathfrak h_{\mathbf R}^*$ into a finite number of closed convex cones, named Weyl chambers. In particular, each positive root system $\Delta_+$ corresponds bijectively to a \emph{dominant Weyl chamber} defined by
\begin{equation}
C = \left\{ \psi \in \mathfrak h_{\mathbf R}^* \,|\, (\psi,\alpha) \geq 0 \mbox{ for all } \alpha \in \Delta_+ \right\}.
\end{equation}
Recall that our vector $v$ has been chosen to belong to $\mathfrak h_0$, hence $iv\in\mathfrak h_{\mathbf R}$. So, by the isomorphism between $\mathfrak h_{\mathbf R}$ and $\mathfrak h_{\mathbf R}^*$ recalled above, there exists a unique a co-vector $\varphi \in \mathfrak h_{\mathbf R}^*$ such that $h_\varphi = -iv$. In particular, one can always choose a positive root system $\Delta_+$ such that $\varphi$ belongs to the fundamental Weyl chamber $C$, and in the rest of the paper we assume such a choice of positive root system $\Delta_+$ has been made.
\subsection{Fundamental dominant weights}\label{ssec:fundamentalDominantWeights}
There exists a convenient basis of $C$ by means of fundamental dominant weights, which we now recall. More details are contained in \cite[Sec.~13.1]{Humphreys1972}.
Let $\ell$ be the rank of $\mathfrak g$, that is the dimension of $\mathfrak h_0$, so that we can label the simple roots form 1 to $\ell$, $\Sigma^+ = \{\gamma_1,\dots,\gamma_\ell\}$. Let $A = (A_{ij})$ be the \emph{Cartan matrix} associated to the Lie algebra $\mathfrak{g}_c$, which is defined as
\begin{equation}\label{eq:cartanMatrix}
A_{ij} = \frac{2 (\gamma_i,\gamma_j)}{(\gamma_i,\gamma_i)}.
\end{equation}
Then the \emph{fundamental dominant weights} $\varphi_1,\dots, \varphi_\ell$ are the elements of $\mathfrak h_{\mathbf R}$ defined by $\varphi_j = \sum_{i=1}^\ell (A^{-1})^{ij}\gamma_i$. In particular they form a basis of $\mathfrak h_{\mathbf R}^*$ suited for the computations we are going to perform.
\begin{lem}\label{lem::vassumoffunddomweights}
Let $\psi \in \mathfrak h_{\mathbf R}^*$ and write $\psi = \sum_{j=1}^\ell w^j \varphi_j$ for some reals $w^1, \dots, w^\ell$.
Then one has $(\psi,\alpha) \geq 0$ for each positive root $\alpha$ if and only if all $w^i$'s are non-negative.
Moreover, denoted by $\Delta_+ \setminus \psi^\perp$ the subset of positive roots which are not orthogonal to $\psi$, one has $\Delta_+ \setminus \psi^\perp = \mathrm{span} \left\{ \gamma_j \ |\ w^j \neq 0\right\} \cap \Delta_+$.
\end{lem}
The above lemma implies that the fundamental Weyl chamber $C$ is the (closed) convex cone spanned by the fundamental dominant weights $\varphi_1,\dots,\varphi_\ell$. Therefore we can write
\begin{equation}
\varphi = \sum_{i=1}^\ell v^i \varphi_i \quad \mbox{ for some } \quad v^1,\dots,v^\ell \geq 0,
\end{equation}
since we chose $\varphi\in C$.
We also recall an element of the root lattice that will come up frequently in our computation, which is denoted by $\delta$ \cite[Sec.~10.2]{Humphreys1972}. It is defined equivalently in terms of the roots or the fundamental dominant weights by
\begin{equation}\label{eq: deltavector}
\delta=\frac{1}{2}\sum_{\alpha\in\Delta_+}\alpha=\sum_{i=1}^{\ell}\varphi_i.
\end{equation}
\subsection{Canonical almost-complex structure}\label{sec::definitionJ}
In this section we recall a canonically defined homogeneous almost-complex structure on $M=G/V$, which turns out to be compatible with the Kirillov-Kostant-Souriau symplectic form $\omega$ \cite{DellaVedova2019,AlekseevskyPodesta2018}. An \emph{almost-complex structure} on $M$ is an element $J\in \text{End}(TM)$ such that $J^2=\text{id}$, and it is \emph{compatible} with $\omega$ if
\begin{equation}
\omega(JX,JY)=\omega(X,Y),\quad \omega(JX,X)>0,\quad\forall X,Y\in TM.
\end{equation}
Consider a root $\alpha \in \Delta$ and define $\lambda_\alpha = s_{\alpha}(\alpha,\varphi)\in \mathbb{R}$, where $s_{\alpha}=1$ if $\alpha\in\Delta_+$ and $s_{\alpha}=-1$ otherwise. Note that, by the assumption $\varphi\in C$, it holds that $\lambda_\alpha \geq 0$, and $\lambda_\alpha=0$ when $\alpha$ is orthogonal to $\varphi$. Then, for each root $\alpha \in \Delta_+$ define the vectors
\begin{equation}\label{eq::uandvintermsofe}
u_\alpha = \frac{i^{(1-\varepsilon_\alpha)/2}}{\sqrt 2} (e_\alpha + e_{-\alpha}), \qquad
v_\alpha = \frac{i^{(3-\varepsilon_\alpha)/2}}{\sqrt 2} s_\alpha (e_\alpha - e_{-\alpha}),
\end{equation}
where $e_{\alpha}\in\mathfrak{g}^{\alpha}$ is an element chosen such as $[e_{\alpha},e_{-\alpha}]=h_{\alpha}$ and satisfying other additional properties \cite[Pg.~265]{GriffithsSchmid1969}. Observe that one has $u_\alpha = u_{-\alpha}$ and similarly $v_\alpha = v_{-\alpha}$. As a consequence of the choices of the $e_{\alpha}$'s one has the following statement.
\begin{lem}\label{lem::advonualphaandvalpha}
For all roots $\alpha, \beta \in \Delta$ one has
\begin{enumerate}
\item $B(u_\alpha,u_\beta) = B(v_\alpha,v_\beta) = (\delta_{\alpha,\beta} + \delta_{\alpha,-\beta})\varepsilon_\alpha$, \label{item::Bualphaubeta}
\item $B(u_\alpha,v_\beta)=0$, \label{item::Bualphavbeta}
\item $u_\alpha, v_\alpha \in \mathfrak g$, \label{item::alphavalphaing}
\item $[v,u_\alpha] = \lambda_\alpha v_\alpha$ and $[v,v_\alpha] = -\lambda_\alpha u_\alpha$, \label{item::[v,ualpha]}
\end{enumerate}
where $\delta_{a,b}=1$ if $a=b$ and $0$ otherwise.
\end{lem}
As a consequence of the above lemma, we have the $B$-orthogonal decomposition
\begin{equation}\label{eq:decompositiong}
\mathfrak g = \mathfrak h_0 \oplus \sum_{\alpha \in \Delta_+} \Span\{u_\alpha ,v_\alpha\}=\mathfrak h_0\oplus\sum_{\alpha \in \Delta_+\cap \varphi^\perp} \Span\{u_\alpha ,v_\alpha\}\oplus\sum_{\alpha \in \Delta_+\setminus \varphi^\perp} \Span\{u_\alpha ,v_\alpha\},
\end{equation}
where $\Delta_+ \cap \varphi^\perp$ denotes the subset of positive roots which are orthogonal to $\varphi$. Note that, as a consequence of point \ref{item::Bualphavbeta} of lemma \ref{lem::advonualphaandvalpha}, a root $\alpha$ is in $\Delta_+ \cap \varphi^\perp$ if $\lambda_\alpha=0$ and in its complement if $\lambda_\alpha>0$. Thus, item \ref{item::[v,ualpha]} of Lemma \ref{lem::advonualphaandvalpha} shows that the Lie algebra $\mathfrak v$ of the stabilizer of $v$ can be decomposed as
\begin{equation}\label{eq:decompositionv}
\mathfrak{v}=\mathfrak h_0\oplus\sum_{\alpha \in \Delta_+\cap \varphi^\perp} \Span\{u_\alpha ,v_
\alpha\}.
\end{equation}
In particular, all roots belonging to $\Delta_+ \cap \varphi^\perp$ are compact, by compactness of $V$. On the other hand, putting together the decompositions \eqref{eq:decompositionv} and \eqref{eq:decompositiong}, one has
\begin{equation}
\mathfrak{m}=\sum_{\alpha \in \Delta_+\setminus \varphi^\perp} \Span\{u_\alpha ,v_\alpha\}.
\end{equation}
Finally, let $H$ be the complex structure on $\mathfrak h_0^\perp = \Span_{\alpha \in \Delta_+} \{u_\alpha,v_\alpha\}$ defined by
\begin{equation}\label{eq::defH}
Hu_\alpha = \varepsilon_\alpha v_\alpha, \qquad Hv_\alpha = - \varepsilon_\alpha u_\alpha \qquad \mbox{for all } \alpha \in \Delta.
\end{equation}
Recall that $H:\mathfrak h_0^\perp\to \mathfrak h_0^\perp$ and $H^2=\text{id}_{\mathfrak h_0^\perp}$. We denote by $J\in\text{End}(T(G/V))$ the canonical homogeneous almost-complex structure on the orbit $G/V$ induced by $H$ on $\mathfrak m$. A manifold equipped with a symplectic form and a compatible almost-complex structure is called \emph{almost-K\"ahler}, hence $(G/V,\omega,J)$ turns out to be an almost-K\"ahler manifold.
\subsection{The condition $\rho = \lambda \omega$}\label{sec::spacialitycondition}
In this section we recall the necessary basic notions of symplectic geometry and the differential equation of interest. Note that, in our setting, the equation $\rho = \lambda \omega$ will be written as an algebraic equation, and we will deal only with that.
Let $(M,\omega)$ be a symplectic manifold, and let $J$ be a compatible almost-complex structure on it. We can define a closed two-form $\rho$ on $M$, the \emph{Chern-Ricci form} of $J$, in the following way. Let $\nabla$ be the \emph{Chern connection} on $M$, i.e., the unique affine connection on $M$ such that $\nabla \omega = 0$, $\nabla J = 0$ and its torsion is exactly the Nijenhuis tensor of $J$. Its curvature $R$ is a two-form with values in $\text{End}(TM)$, and the form $\rho$ is defined by $\rho(X,Y) = \mathrm{tr}(JR(X,Y))$.
As explained in the introduction, a common question to ask is if the equation $\rho = \lambda \omega$ is satisfied for some constant $\lambda\in\mathbb{R}$. If this last equation is satisfied and $J$ is integrable (i.e., $M$ is a complex manifold), then $(M,\omega,J)$ is called \emph{K\"ahler-Einstein} manifold. On the other hand, a non-complex almost-K\"ahler manifold $(M,\omega,J)$ which satisfies $\rho = \lambda \omega$ is sometimes called \emph{Hermitian-Einstein}, \emph{special} \cite{DellaVedova2019}, or \emph{Chern-Einstein} \cite{AlekseevskyPodesta2018}. The nomenclature is not standard in this case. Below we are going to consider the condition $\rho = \lambda \omega$ on adjoint orbits \ref{sec:adjointOrbits} equipped with the almost-complex structure induced by $H$ \ref{sec::definitionJ}.
In our setting, the Chern-Ricci form $\rho$ of $J$ is determined by the two form $B(v',[\cdot,\cdot])$, with $ v'= 2 \sum_{\alpha \in \Delta_+ \setminus \varphi^\perp} [u_\alpha,v_\alpha]$ \cite[Sec.~4.2]{DellaVedova2019}.
By using the definition of $u_\alpha,v_\alpha$ \eqref{eq::uandvintermsofe}, one can write
\begin{equation}\label{eq::v'sumofroots}
v' = -2i \sum_{\alpha \in \Delta_+ \setminus \varphi^\perp} \varepsilon_\alpha h_\alpha.
\end{equation}
Hence, one can consider the element of the root lattice
\begin{equation}\label{eq::definitionvarphi'}
\varphi' = - 2 \sum_{\alpha \in \Delta_+ \setminus \varphi^\perp} \varepsilon_\alpha \alpha \in \mathfrak h_{\mathbf R}^*
\end{equation}
satisfying $h_{\varphi'} = -iv'$. At this point, the condition involving the Chern-Ricci form $\rho = \lambda \omega$ turns out to be equivalent to $v' = \lambda v$ and $\varphi' = \lambda \varphi$.
We introduce the element of the root lattice
\begin{equation}\label{eq:eta}
\eta = -2 \sum_{\alpha \in \Delta_+} \varepsilon_\alpha \alpha \in \mathfrak h_{\mathbf R}^*,
\end{equation}
that allows to write the condition in a more tractable way. This element can be thought as analogous to the $\delta$-vector \eqref{eq: deltavector} for real simple Lie algebras. Observe that $\eta$ depends on the semisimple Lie algebra $\mathfrak g$ and on the chosen set of positive roots. One can then write the condition $\rho = \lambda \omega$ as
\begin{equation}\label{eq::rho=lambdaomegaintermsoftau}
\eta - 2 \sum_{\alpha \in \Delta_+ \cap \varphi^\perp} \alpha = \lambda \varphi,
\end{equation}
where we used the fact that no non-compact roots can be orthogonal to $\varphi$. This is the equation we are going to study in section \ref{sec:specialCVD}, in terms of the combinatorics of Vogan diagrams. In particular, one has the following result, which is the building block to determine speciality of a given adjoint orbit.
\begin{thm}\label{cor::specialvALL}
For any $\varphi \in \mathfrak h_{\mathbf R}^*$ and any real $\lambda \in\{-1,0,1\}$ the following are equivalent:
\begin{itemize}
\item $\varphi$ belongs to the dominant Weyl chamber $C$, the stabilizer of $v=ih_\varphi$ is compact, and one has $\varphi'=\lambda \varphi$;
\item there exists $S \subset \{ 1,\dots,\ell \}$ such that $i \in S$ whenever $\gamma_i$ is a non-compact simple root, and $\varphi_S = \eta - 2 \sum_{\alpha \in \Span\{ \gamma_i | i \in S^c\} \cap \Delta_+} \alpha$ satisfies $(\varphi_S,\gamma_i) = \lambda |(\varphi_S,\gamma_i)|$ for all $i \in S$. Moreover
\begin{equation}
\varphi= \left\{
\begin{array}{ll}
\lambda \varphi_S & \mbox{if } \lambda = \pm 1 \\
\sum_{i \in S} v^i \varphi_i \mbox{ for some }v^i>0 & \mbox{if } \lambda = 0.
\end{array}
\right.
\end{equation}
\end{itemize}
\end{thm}
What the above theorem says is that we can choose a Vogan diagram and algorithmically check the signs of $(\varphi_S,\gamma_i)$ for each simple non-compact root $\gamma_i$. If the signs are all the same, then the associated orbit admits special canonical almost-complex structure.
\subsection{Vogan diagrams}\label{sec::Vogandiagrams}
Vogan diagrams are combinatorial objects used to classify real semisimple Lie algebras \cite[Chapter VI]{Knapp1996}. In our setting, they will play an important role in studying the equation \eqref{eq::rho=lambdaomegaintermsoftau} on adjoint orbits of simple Lie groups, as explained in the previous sections.
A \emph{Vogan diagram with trivial automorphism} is a Dynkin diagram with some (including no one or all) painted vertices, where painted vertices correspond to simple non-compact roots. In the following, by Vogan diagram we will always intend a Vogan diagram with trivial automorphism. The reason why these diagrams are important in our setting is contained in the following lemma.
\begin{lem}\label{lem::correspVogandiagramwithv}
Let $G$ be a real semisimple Lie group with Lie algebra $\mathfrak g$, and let $\ell$ be the rank of $\mathfrak g$.
To any $v \in \mathfrak g$ with compact stabilizer, one can associate a Vogan diagram
and a vector $(v^1,\dots,v^\ell) \in \mathbf R^\ell$ with $v^i \geq 0$.
Moreover $v^i>0$ if the $i$-th node of the Vogan diagram is painted.
\end{lem}
Since a Cartan subalgebra $\mathfrak h_0$ containing $v$ and a Weyl chamber $C$ containing $\varphi$ cannot be chosen in a canonical way, one has that it is possible to associate different Vogan diagrams to the same element $v$. However, once the Vogan diagram is chosen and the simple roots are labelled, the vector $(v^1,\dots,v^\ell)$ is uniquely determined.
The correspondence established by Lemma \ref{lem::correspVogandiagramwithv} can be reversed. Indeed, given a connected Vogan diagram one can find a positive root system $\Delta_+$, and determine which roots in $\Delta_+$ are compact by using the formula in lemma \ref{cor::epsilonalphaformula}. At this point, the Weyl Chamber associated to $\Delta_+$ is determined, hence one can choose $\varphi \in C$, $\varphi = \sum_{i=1}^\ell v^i \varphi_i$, with $v^i \geq 0$.
The important fact is that, given a non-compact real simple Lie group $G$, all adjoint orbits $(G/V,\omega,J)$ such that $\rho=\lambda\omega$, for some constant $\lambda\in\mathbb{R}$, can be listed in an algorithmic way (up to isomorphism and scaling). Indeed, by what we said above, this is equivalent to list (up to scaling) all special $\varphi$'s for all possible connected Vogan diagrams.
In \cite{DellaVedovaGatti2022} the authors classified all such vectors for classical Vogan diagrams up to rank $\ell=4$ and all the exceptional ones. This bound on the rank was mainly due to the fact that the algorithm to compute them becomes very slow with high ranks, since it requires to compute the full root system of the simple Lie algebra, together with its compact and non-compact roots. In the next section we are going to write the condition of special vector purely in terms of the indices of simple non-compact roots of a Vogan diagram. This makes the problem combinatorial and allows to scale the computation to much bigger Vogan diagrams.
\section{The condition $\rho=\lambda\omega$ through the combinatorics of Vogan diagrams}\label{sec:specialCVD}
In this section we are going to study the condition $\rho=\lambda\omega$ in terms of the combinatorics of Vogan diagrams. In particular, we will improve the results in \cite{DellaVedovaGatti2022}, providing a purely combinatorial condition for a special vector. Note that we will not consider adjoint orbits of exceptional type, since they were fully classified in \cite{DellaVedovaGatti2022}. The code that we used to test the formulae is available at the GitHub repository \href{https://github.com/alga-hopf/special-vogan-diagrams}{https://github.com/alga-hopf/special-vogan-diagrams}.
Consider the Vogan diagram of a non-compact real semisimple Lie algebra $\mathfrak{g}$ with $S=\lrbr{i_1,\ldots,i_m}$ the set of indices of simple non-compact roots. The vector
\begin{equation}
\varphi=\eta-2\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha
\end{equation}
can be written as
\begin{align}
\varphi =&\eta-2\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\\
=& -2\sum_{\alpha\in\Delta_+^{nc}}\alpha-2\sum_{\alpha\in\Delta_+^c}\alpha +4\sum_{\alpha\in\Delta_+^c}-2\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha \label{eq:phisp1}\\
=& -4\delta+4\sum_{\alpha\in\Delta_+^c}\alpha-2\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha \label{eq:phisp4},
\end{align}
where in \eqref{eq:phisp1} we used the definition of $\eta$ \eqref{eq:eta}, in \eqref{eq:phisp4} we used the definition of $\delta$ \eqref{eq: deltavector} and $\Delta_+^{nc}$, $\Delta_+^{c}$ denote the sets of positive non-compact and compact roots respectively. To understand whether the vector $\varphi$ is special, we have to compute the signs of its $i$th coefficients, $i\in S$, by theorem \ref{cor::specialvALL}. As explained in section \ref{ssec:fundamentalDominantWeights}, the coefficients $\xi_i$ of $\varphi$ in the basis of the fundamental dominant weights $\lrbr{\omega_1,\ldots,\omega_{\ell}}$ are given by
\begin{equation}
\xi_i=(A\varphi)_i\quad i\in S,
\end{equation}
where $A$ is the Cartan matrix of the underlying complex Lie algebra \eqref{eq:cartanMatrix}. Hence
\begin{equation}
\tilde{\varphi}=-4A\delta+4A\sum_{\alpha\in\Delta_+^c}\alpha-2A\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha,
\end{equation}
where $\tilde{\varphi}$ denotes $\varphi$ in the basis of fundamental dominant weights. Since, by definition, $\delta_i=1$ in the basis of the fundamental dominant weights, we have the following expression for the $\xi_i$'s.
\begin{lem}\label{eq:xiExpression}
In the basis of the fundamental dominant weights, the coefficients $\xi_i$'s can be expressed as
\begin{equation}
\xi_i=-4+4\left(A\sum_{\alpha\in\Delta_+^c}\alpha\right)_i-2\left(A\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i\vert i\in S^c}}\alpha\right)_i,
\end{equation}
where $A$ is the Cartan matrix of the underlying Dynkin diagram.
\end{lem}
Then, by theorem \ref{cor::specialvALL}, the considered Vogan diagram is special if all the $\xi_i$'s, $i\in S$, have the same sign. In order to check the signs of the $\xi_i$'s we need to compute explicitly the coefficients $\xi_i$, $i\in S$, which is the goal of the subsequent sections.
We can now start diving into the study of the coefficients $\xi_{i}$'s for Vogan diagrams of type $A_{\ell}$.
\subsection{$A_{\ell}$ family}\label{sec:anCoefficients}
In this section we provide a closed formula for the coefficients $\xi_i$ defined above. The arguments used in this section will be used also for other families of diagrams, thus we provide the details here without repeating them in other cases, if not necessary.
Consider a diagram of type $A_{\ell}$ with $S=\lrbr{i_1,\ldots,i_m}$ the set of indices of simple non-compact roots, and let $\{\gamma_1,\ldots,\gamma_{\ell}\}$ denote the simple roots. Note that, by definition,
\begin{equation}\label{eq: coeffAn}
\left(\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_i=0\quad \forall i\in S.
\end{equation}
Combining equation \eqref{eq: coeffAn} together with the definition of the Cartan matrix $A$ for Lie algebras of type $A_{\ell}$ \cite[Sec.~11.4, Tab.~1]{Humphreys1972} we get for $i_k\in S$, $1\leq k\leq m$,
\begin{equation}
\left( A\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_{i_k}=-\mathrm{dim}(\mathfrak{h}_{i_{k}})-\mathrm{dim}(\mathfrak{h}_{i_{k+1}}),
\end{equation}
where $\mathfrak{h}_{i_k}$ is the Cartan subalgebra of the Lie subalgebra included between the nodes $i_{k-1}+1$ and $i_{k}-1$. This follows from the fact that the roots are unbroken strings \cite[Sec.~9.4]{Humphreys1972} and that the roots of the smaller subalgebra of type $A$ included between $i_{k-1}+1$ and $i_{k}-1$ having $\alpha_{i_k-1}\neq 0$ are exactly $\sum_{i=n_1}^{i_k-1}\gamma_i$, $i_{k-1}<n_1<i_k$. Their number corresponds to the dimension of the corresponding Cartan subalgebra $\mathrm{dim}(\mathfrak{h}_{i_{k}})$. A similar argument holds for $\mathrm{dim}(\mathfrak{h}_{i_{k+1}})$. Moreover, we can write
\begin{equation}\label{eq:compactRootsSplit}
\sum_{\alpha\in\Delta_+^c}\alpha=\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i\vert i\in S^c}}\alpha+\sum_{\alpha\in\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha,
\end{equation}
and we have
\begin{align}
\left(A\sum_{\alpha\in\Delta_+^c}\alpha\right)_{i_k}=& \left(A\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_{i_k}+\left(A\sum_{\alpha\in\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_{i_k}\\
=&-\mathrm{dim}(\mathfrak{h}_{i_{k}})-\mathrm{dim}(\mathfrak{h}_{i_{k+1}})+\left(A\sum_{\alpha\in\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_{i_k}\\
=& -\mathrm{dim}(\mathfrak{h}_{i_{k}})-\mathrm{dim}(\mathfrak{h}_{i_{k+1}})+\left(A\delta^c\right)_{i_k}, \label{eq:acompactroots}
\end{align}
where $\delta^c=\sum_{\alpha\in\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i\vert i\in S^c}}\alpha$.
Putting together equation \eqref{eq:acompactroots} and lemma \ref{eq:xiExpression} we get
\begin{align}\label{eq: phiCoeff}
\xi_{i_k}=&2\left(-2-\mathrm{dim}(\mathfrak{h}_{i_{k}})-\mathrm{dim}(\mathfrak{h}_{i_{k+1}})+2\left(A\delta^c\right)_{i_k}\right)\\
=&2\left(-2-(i_{k}-i_{k-1}-1)-(i_{k+1}-i_k-1)+2(-\delta^c_{i_k-1}+2\delta^c_{i_k}-\delta^c_{i_k+1})\right) \label{eq: equality3an}\\
=&2\left(i_{k-1}-i_{k+1}+2(-\delta^c_{i_k-1}+2\delta^c_{i_k}-\delta^c_{i_k+1})\right) \label{eq: equality4an},
\end{align}
where in equality \eqref{eq: equality3an} we used the fact that $\mathrm{dim}(\mathfrak{h}_{i_{l}})=i_l-i_{l-1}-1$, with the assumptions that $i_0=0$ and $i_{m+1}=\ell+1$, and the structure of $A$.
At this point, in order to determine the coefficient $\xi_{i_k}$, we need to make the term $-\delta^c_{i_k-1}+2\delta^c_{i_k}-\delta^c_{i_k+1}$, $i_k\in S$, explicit. This requires some knowledge about the roots systems of Lie algebras of type $A_{\ell}$. In particular, the positive roots of Lie algebras type $A_{\ell}$ are
\begin{equation}
\Delta_+=\left\{ \sum_{i=n_1}^{n_2}\gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell\right\},
\end{equation}
while the positive compact roots are of the form
\begin{equation}
\Delta^c_+=\left\{ \sum_{i=n_1}^{n_2}\gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell,\ \lvert S\cap\{n_1,\ldots,n_2\}\rvert\in 2\mathbb{Z}\right\}.
\end{equation}
Before stating the theorem about the coefficients $\xi_i$ we introduce the following quantities that will be used extensively through the paper:
\begin{align}
S_1(j):=&\sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1},\quad S_2(j):=\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k} \\
T_1(j):=&\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1},\quad T_2(j):=\sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}\\
S_+(j):=&\sum_{k=j+1}^m 2(-1)^{k-j}i_k +i_{j+1}+(-1)^{m-j+1}(\ell+1)=S_1(j)-T_2(j) \label{eq: s+}\\
S_-(j):=&\sum_{k=1}^{j-1} 2(-1)^{k-j}i_k+i_{j-1}=S_2(j)-T_1(j). \label{eq: s-}
\end{align}
This notation is going to greatly simplify the formulae we are going to present.
\begin{thm}\label{thm: ancoeff}
Given a Vogan diagram of type $A_{\ell}$ with $S=\lrbr{i_1,\ldots,i_m}$ the set of indices of simple non-compact roots, for $i_j\in S$ one has
\begin{equation}\label{eq:xiAn}
\xi_{i_j}=2\left(i_{j-1}-i_{j+1}+2\left(S_+(j)-S_-(j)\right)\right).
\end{equation}
\end{thm}
\begin{proof}
First, assume that $i_0=0$ and $i_{m+1}=\ell+1$. We are going to prove that
\begin{equation}\label{eq: deltacThm}
-\delta_{i_j-1}^c+2\delta^c_{i_j}-\delta_{i_j+1}^c= S_1(j)+S_2(j) -T_1(j)-T_2(j).
\end{equation}
This, together with equality \eqref{eq: equality4an} and quantities \ref{eq: s+}, \eqref{eq: s-}, proves identity \eqref{eq:xiAn}.
Note that we can express
\begin{equation}
\delta^c_{i_j-1}=\delta^c_{i_j}-s_1+t_1,\quad \delta^c_{i_j+1}=\delta^c_{i_j}-s_2+t_2,
\end{equation}
where
\begin{itemize}
\item $s_1$ is the number of compact roots $\sum_{i=1}^{n}\alpha_i\gamma_i$ in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that \\$\alpha_{i_j}-\alpha_{i_j-1}=1$,
\item $s_2$ is the number of compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j}-\alpha_{i_j+1}=1$,
\item $t_1$ is the number of compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j-1}-\alpha_{i_j}=1$,
\item $t_2$ is the number of compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j+1}-\alpha_{i_j}=1$.
\end{itemize}
Hence
\begin{equation}
-\delta_{i_j-1}^c+2\delta^c_{i_j}-\delta_{i_j+1}^c=s_1-t_1+s_2-t_2.
\end{equation}
So it suffices to determine the numbers $s_1,s_2,t_1,t_2$ for each $i_j$. We start with the coefficient $s_1$ and we need to count the compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j}=1$ and $\alpha_{i_j-1}=0$. These are of the form $\sum_{l=i_j}^{n_1}\gamma_l$, with $S\cap|\lrbr{i_j,\ldots,n_1}|\in 2\mathbb{Z}$. These roots are
\begin{equation}
s_1=\sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}=S_1(j).
\end{equation}
A similar argument can be used to compute $s_2$, and we need to compute the the compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j}=1$ and $\alpha_{i_j+1}=0$, which are of the form $\sum_{l=n_1}^{i_j}\gamma_l$, with $S\cap|\lrbr{n_1,\ldots,i_j}|\in 2\mathbb{Z}$.
These are
\begin{equation}
s_2=\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}=S_2(j).
\end{equation}
For the coefficient $t_1$, the compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j}=0$ and $\alpha_{i_j-1}=1$ are of the form $\sum_{l=n_1}^{i_j-1}\gamma_l$, with $S\cap|\lrbr{n_1,\ldots,i_j-1}|\in 2\mathbb{Z}$.
This reads
\begin{equation}
t_1=\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}=T_1(j).
\end{equation}
Finally, for $t_2$ the compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j}=0$ and $\alpha_{i_j+1}=1$ are of the form $\sum_{l=i_j+1}^{n_2}\gamma_l$, with $S\cap|\lrbr{i_j+1,\ldots,n_1}|\in 2\mathbb{Z}$. So
\begin{equation}
t_2=\sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}=S_2(j).
\end{equation}
Thus,
\begin{equation}
s_1-t_1+s_2-t_2=S_1(j)-T_1(j)+S_2(j)-T_2(j).
\end{equation}
This proves equality \eqref{eq: deltacThm}
\end{proof}
Theorem \ref{thm: ancoeff} above gives a formula to compute the coefficients $\xi_i$ that depends only on the set $S$ and its combinatorics. In the next sections we are going to state similar results to the one for $A_{\ell}$. Even if the root systems are more intricate, the fundamental ideas are the ones developed in the current section.
\subsection{$B_{\ell}$ family}\label{sec:bnCoefficients}
In this section we are going to study the coefficients $\xi_i$ for Vogan diagrams of type $B_{\ell}$. Consider a Vogan diagram of type $B_{\ell}$ with $S=\lrbr{i_1,\ldots,i_m}$ the set of indices of simple non-compact roots, and denote $\{\gamma_1,\ldots,\gamma_{\ell}\}$ the set of simple roots. As we did in section \ref{sec:anCoefficients}, taking into account the form of the Cartan matrix $A$ of type $B_{\ell}$ \cite[Sec.~11.4,Tab.~1]{Humphreys1972}, one has
\begin{equation}
\begin{split}
\Gamma_{i_k}=&\left( A\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_{i_k} \\
=&\left\{ \begin{array}{lcl}
-\mathrm{dim}(\mathfrak{h}_{i_{k}})-\mathrm{dim}(\mathfrak{h}_{i_{k+1}}) & \text{if} &k\neq m\\
-\mathrm{dim}(\mathfrak{h}_{i_{m}})-(2\mathrm{dim}(\mathfrak{h}_{i_{m+1}})-1)& \text{if} &k= m, i_k\neq \ell \\
-2\mathrm{dim}(\mathfrak{h}_{\ell}) & \text{if} & $k=m$, i_k= \ell \end{array}\right. \\
=&\left\{ \begin{array}{lcl}
i_{k-1}-i_{k+1}+2 & \text{if} &k\neq m\\
2+i_m+i_{m-1}-2n & \text{if} &k= m, i_k\neq \ell \\
-2n+2i_{m-1}+2 & \text{if} & k=m, i_k= \ell \end{array}\right.,
\end{split}
\end{equation}
where we assume $i_0=0$ and $i_{m+1}=\ell+1$. The case $k= m, i_k\neq \ell$ comes from the fact that the roots in $\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ that have non-zero $i_m+1$ component are the roots of the Lie algebra of type $B$ included between $i_m+1$ and $\ell$, which are exactly $2\mathrm{dim}(\mathfrak{h}_{i_{k+1}})-1$. The case $k=m, i_k= \ell$ comes from the structure of the Cartan matrix $A$ and the fact that the Lie algebra included between $i_{m-1}+1$ and $i_{m}-1$ is of type $A_{n}$.
By equation \eqref{eq:compactRootsSplit}, similarly as for the $A_{\ell}$ case, we have
\begin{equation}
\begin{split}
\left(A\sum_{\alpha\in\Delta_+^c}\alpha\right)_{i_k}=& \left(A\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_{i_k} + \left(A\sum_{\alpha\in\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_{i_k}\\
=&\Gamma_{i_k}+\tau_{i_k},\\
\end{split}
\end{equation}
where $\delta^c=\sum_{\alpha\in\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i\vert i\in S^c}}\alpha$ and we put $\tau_{i_k}:=\left(A\delta^c\right)_{i_k}$. In particular,
\begin{equation}
\tau_k=\left\{\begin{array}{lcl}
-\delta^c_{i_k-1}+2\delta^c_{i_k}-\delta^c_{i_k+1} & \text{if} & i_k\neq \ell \\
-2\delta^c_{i_k-1}+2\delta^c_{i_k} & \text{if} & i_k=\ell
\end{array}\right.,
\end{equation}
by the structure of the Cartan matrix $A$.
Finally, as for the $A_{\ell}$ case,
\begin{equation}\label{eq: comformulabn}
\xi_{i_k}=-4-2\Gamma_{i_k}+4\left( A\sum_{\alpha\in\Delta_+^c}\alpha\right)_{i_k}=2\left(-2+\Gamma_{i_k}+2\tau_{i_k}\right).
\end{equation}
In order to write explicitly the coefficients $\tau_{i_k}$, we need to recall that the structure of the positive roots of type $B_{\ell}$,
\begin{equation}
\Delta_+=\left\{ \sum_{i=n_1}^{n_2}\gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell\right\}\cup \left\{ \sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell}2 \gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell-1\right\},
\end{equation}
and the positive compact roots
\begin{equation}
\begin{split}
\Delta^c_+=&\left\{ \sum_{i=n_1}^{n_2}\gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell,\ \lvert S\cap\{n_1,\ldots,n_2\}\rvert \in 2\mathbb{Z}\right\}\cup \\ & \cup\left\{ \sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell}2 \gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell-1,\ \lvert S\cap\{n_1,\ldots,n_2\}\rvert\in 2\mathbb{Z}\right\}.
\end{split}
\end{equation}
In the following theorem we compute the coefficients $\xi_i$.
\begin{thm}
Given a Vogan diagram of type $B_{\ell}$ with $S=\lrbr{i_1,\ldots,i_m}$ the set of indices of simple non-compact roots, for $i_j\in S$ one has
\begin{itemize}
\item If $j\neq m$
\begin{equation}\label{eq:bCoeff1}
\xi_{i_j}=2(-i_{j-1}+4i_j-3i_{j+1}+4S_+(j)+2(-1)^{j+m}).
\end{equation}
\item If $j= m$ and $i_m\neq \ell$
\begin{equation}\label{eq:bCoeff2}
\xi_{i_m}= 2(3i_m-i_{m-1}-2\ell-2).
\end{equation}
\item If $i_j=\ell$
\begin{equation}\label{eq:bCoeff3}
\xi_{\ell}= 2(2\ell-2i_{m-1}).
\end{equation}
\end{itemize}
\end{thm}
\begin{proof}
Assume that $i_0=0$ and $i_{m+1}=\ell+1$.
\begin{itemize}
\item Case $j\neq m$. We are going to prove that
\begin{equation}\label{eq: deltac_Bn1}
\tau_{i_j}= -i_{j-1}+2i_j-i_{j+1}+ 2S_1(j)-2T_2(j)+(-1)^{j+m}.
\end{equation}
This, together with equality \eqref{eq: comformulabn} proves equality \eqref{eq:bCoeff1}. In this case, $\tau_{i_j}$ can be written as
\begin{equation}
\tau_{i_j}=-\delta^c_{i_k-1}+2\delta^c_{i_k}-\delta^c_{i_k+1}=s_1-t_1+s_2-t_2,
\end{equation}
as for the $A_{\ell}$ case. We start with the computation of $s_1$. In order to count the compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j}-\alpha_{i_j-1}=1$, one can first consider the roots such that $\alpha_{i_j}=1$ and $\alpha_{i_j-1}=0$, which are of two types. The roots $\sum_{l=i_j}^{n_2}\gamma_l$, with $|S\cap\lrbr{i_j,\ldots,n_2}|\in 2\mathbb{Z}$, and $\sum_{l=i_j}^{n_1}\gamma_l+\sum_{l=n_1+1}^{\ell}2\gamma_l$, with $\lvert S\cap\{i_j,\ldots,n_1\}\rvert\in 2\mathbb{Z}$, that are exactly
\begin{equation}
\sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}+\sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}-\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}=2S_1(j)-\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}.
\end{equation}
Finally, we have to consider all the roots for which $\alpha_{i_j}=2$ and $\alpha_{i_j-1}=1$, that are of the form $\sum_{l=n_1}^{i_{j-1}}\gamma_l+\sum_{l=i_j}^{\ell}2\gamma_l$, with $\lvert S\cap\{n_1,\ldots,i_{j-1}\}\rvert\in 2\mathbb{Z}$. These are
\begin{equation}
\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}+(i_j-i_{j-1}-1)=T_1(j)+i_j-i_{j-1}-1.
\end{equation}
So
\begin{equation}
s_1=2S_1(j)+T_1(j)+i_j-i_{j-1}-1-\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}.
\end{equation}
For the coefficients $s_2$ and $t_1$ we can follow the same steps made for the $A_{\ell}$ case, so
\begin{align}
s_2=&\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}=S_2(j),\\
t_1=&\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}=T_1(j).
\end{align}
Finally, for $t_2$ one has that the roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j+1}-\alpha_{i_j}=1$ are the following. First we have the roots such that $\alpha_{i_j+1}=1$ and $\alpha_{i_j}=0$, that are of two types: the roots $\sum_{l=i_j+1}^{n_1}\gamma_l$, with $|S\cap\lrbr{i_j+1,\ldots,n_1}|\in 2\mathbb{Z}$, and $\sum_{l=i_j+1}^{n_1}\gamma_l+\sum_{l=n_1+1}^{\ell}2\gamma_l$, with $\lvert S\cap\{i_j+1,\ldots,n_1\}\rvert\in 2\mathbb{Z}$. These are
\begin{multline}
\sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}+\sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}+(i_{j+1}-i_j-1)+ \\-\delta_{\ell+1,i_{j+2\lceil\frac{m -(j+1)}{2} \rceil+1}}
=2T_2(j)+i_{j+1}-i_j-1-\delta_{\ell+1,i_{j+2\lceil\frac{m -(j+1)}{2} \rceil+1}}.
\end{multline}
Then there are the compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j+1}=2$ and $\alpha_{i_j}=1$, that are of the form $\sum_{l=n_1}^{i_j}\gamma_l+\sum_{l=i_j+1}^{\ell}2\gamma_l$, with $\lvert S\cap\{n_1,\ldots,i_j\}\rvert\in 2\mathbb{Z}$. These are
\begin{equation}
\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}=S_2(j).
\end{equation}
Thus,
\begin{equation}
t_2=2T_2(j)+S_2(j)+i_{j+1}-i_j-1-\delta_{\ell+1,i_{2\lceil\frac{m -(j+1)}{2} \rceil+1}}.
\end{equation}
Putting everything together we get
\begin{equation}
s_1+s_2-t_1-t_2= -i_{j-1}+2i_j-i_{j+1}+2S_1+(j)-2T_2(j)+(-1)^{j+m},
\end{equation}
where we used the fact that $-\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}+\delta_{\ell+1,i_{j+2\lceil\frac{m -(j+1)}{2} \rceil+1}}=(-1)^{j+m}$.
This proves equality \eqref{eq: deltac_Bn1}.
\item Case $j=m$, $i_m\neq \ell$. We will prove that
\begin{equation}\label{eq: deltac_Bn2}
\tau_{i_m}= i_j-i_{j-1}-1,
\end{equation}
that, together with equality \eqref{eq: comformulabn}, proves equality \eqref{eq:bCoeff2}.
Again, we have
\begin{equation}
\tau_{i_m}=-\delta^c_{i_m-1}+2\delta^c_{i_m}-\delta^c_{i_m+1}=s_1-t_1+s_2-t_2,
\end{equation}
so we only need to compute the coefficients $s_1,t_1,s_2,t_2$. We start with the coefficient $s_1$. Notice that the compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_m}-\alpha_{i_m-1}=1$, are only the roots such that $\alpha_{i_m}=2$ and $\alpha_{i_m-1}=1$, that are of type $\sum_{l=n_1}^{i_{m-1}}\gamma_l+\sum_{l=i_m}^{\ell}2\gamma_l$, with $S\cap\lvert\{n_1,\ldots,i_{m-1}\}\rvert\in 2\mathbb{Z}$. So
\begin{equation}
s_1=\sum_{k=1}^{\lfloor\frac{m+1}{2} \rfloor-1} i_{m-2k}-i_{m-2k-1}+(i_m-i_{m-1}-1)=T_1(m)+i_m-i_{m-1}-1.
\end{equation}
As above, for $s_2$ and $t_1$ we have the equalities
\begin{equation}
s_2=\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{m-2k+1}-i_{m-2k}=S_2(m),\quad t_1=\sum_{k=1}^{\lfloor\frac{m+1}{2} \rfloor-1} i_{m-2k}-i_{m-2k-1}=T_1(m).
\end{equation}
Finally, for $t_2$ one has that the roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_m+1}-\alpha_{i_m}=1$ are the ones for which $\alpha_{i_m+1}=2$ and $\alpha_{i_m}=1$. These are $\sum_{l=n_1}^{i_m}\gamma_l+\sum_{l=i_m+1}^{\ell}2\gamma_l$, with $\lvert\{n_1,\ldots,i_m\}\cap S\rvert\in 2\mathbb{Z}$, so
\begin{equation}
t_2=\sum_{k=1}^{\lfloor\frac{m}{2} \rfloor} i_{m-2k+1}-i_{m-2k}=S_2(m).
\end{equation}
Summing everything up we get
\begin{equation}
s_1+s_2-t_1-t_2= i_m-i_{m-1}-1,
\end{equation}
which proves equality \eqref{eq: deltac_Bn2}.
\item Case $j=m$, $i_m=\ell$. We will prove equality
\begin{equation}\label{eq: deltac_Bn3}
\tau_{\ell}= 2\left(\ell-i_{m-1}-1\right)
\end{equation}
that, together with equality \eqref{eq: comformulabn}, proves \eqref{eq:bCoeff3}. Remember that
\begin{equation}
\tau_{\ell}=-2\delta^c_{\ell-1}+2\delta^c_{\ell}=s_1-t_1.
\end{equation}
For $s_1$, the compact roots in $\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{\ell}-\alpha_{\ell-1}=1$ are the ones for which $\alpha_{\ell}=2$ and $\alpha_{\ell-1}=1$, that are of type $\sum_{l=n_1}^{\ell-1}\gamma_l+2\gamma_{\ell}$, with $\lvert S\cap\{n_1,\ldots,\ell-1\}\rvert\in 2\mathbb{Z}$. These are
\begin{equation}
s_1=\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}+(\ell-i_{m-1}-1)=T_1(j)+\ell-i_{m-1}-1.
\end{equation}
The coefficient $t_1$ is equal to $T_1(m)$, as in the previous cases, hence
\begin{equation}
s_1-t_1= \ell-i_{m-1}-1,
\end{equation}
which proves equality \eqref{eq: deltac_Bn3} and concludes the proof of the theorem.
\end{itemize}
\end{proof}
The above theorem allows us to compute the coefficients of Vogan diagrams of type $B_{\ell}$ and check their signs. The next two sections are dedicated to Vogan diagrams of type $C_{\ell}$ and $D_{\ell}$. The proofs in these cases are going to be significantly more intricate than the ones we have seen so far. Thus we will try to keep them short and refer to the previous proofs in case of similar computations.
\subsection{$C_{\ell}$ family}\label{sec:cnCoefficients}
Consider a Vogan diagram of type $C_{\ell}$, with $S=\{i_1,\ldots,i_m\}$ the set of indices of simple non-compact roots and denote by $\{\gamma_1,\ldots,\gamma_{\ell}\}$ the set of simple roots. As for the other cases, using the structure of the Cartan matrix $A$ for Lie algebras of type $C_{\ell}$ \cite[Sec.~11.4, Tab.~1]{Humphreys1972}, one can write
\begin{equation}\label{eq:gammaC}
\begin{split}
\Gamma_{i_k}=&\left( A\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_{i_k}\\
&=\left\{ \begin{array}{lcl}-\mathrm{dim}(\mathfrak{h}_{i_{k}})-\mathrm{dim}(\mathfrak{h}_{i_{k+1}}) & \text{if} &k\neq m,\ i_k\neq \ell-1\\
-\mathrm{dim}(\mathfrak{h}_{i_{m}})-2\mathrm{dim}(\mathfrak{h}_{i_{m+1}})& \text{if} &k= m, i_k\neq \ell-1 \\
-\mathrm{dim}(\mathfrak{h}_{\ell-1})-2 & \text{if} & k=m, i_k= \ell-1 \\
-\mathrm{dim}(\mathfrak{h}_{\ell-1}) & \text{if} & k\neq m, i_k= \ell-1 \\
\end{array}\right.\\
=& \left\{ \begin{array}{lcl}
i_{k-1}-i_{k+1}+2 & \text{if} &k\neq m,\ i_k\neq \ell-1\\
i_m+i_{m-1}+1-2\ell & \text{if} &k= m, i_k\neq \ell-1 \\
-\ell+i_{m-1} & \text{if} & k=m,i_k= \ell-1 \\
-\ell+i_{k-1}+2 & \text{if} & k\neq m,i_k= \ell-1
\end{array}\right.
\end{split}
\end{equation}
where, for $k= m, i_k\neq \ell-1$, the second summand comes from the fact that the roots in $\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ that have non-zero $i_m+1$ component are the roots of the Lie algebra of type $C$ included between $i_m+1$ and $\ell$, which are exactly $2\mathrm{dim}(\mathfrak{h}_{i_{m+1}})$. For the last two cases, the second summand is equal to $0$ if $i_m\neq \ell-1$, and equal to $-2$ if $i_m= \ell-1$, as a consequence of the fact that $A_{\ell-1,\ell}=-2$, and there is only one root between $\ell-1$ and $\ell$.
Similarly for the $A_{\ell}$ and $B_{\ell}$ cases, one has
\begin{equation}\label{eq: formulacncoefficients}
\xi_{i_k}=-4-2\Gamma_{i_k}+4\left( A\sum_{\alpha\in\Delta_+^c}\alpha\right)_{i_k}=2\left(-2+\Gamma_{i_k}+2\tau_{i_k}\right),
\end{equation}
where $\delta^c=\sum_{\alpha\in\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i\vert i\in S^c}}\alpha$ and we put $\tau_{i_k}:=\left(A\delta^c\right)_{i_k}$. Note that
\begin{equation}
\tau_{i_k}=\left\{\begin{array}{lcl}
-\delta^c_{i_k-1}+2\delta^c_{i_k}-\delta^c_{i_k+1} & \text{if} & i_k\neq \ell-1 \\
-\delta^c_{\ell-1}+2\delta^c_{\ell} & \text{if} & i_k=\ell\\
-\delta^c_{\ell-2}+2\delta^c_{\ell-1}-2\delta^c_{\ell} & \text{if} & i_k=\ell-1\\
\end{array}\right.,
\end{equation}
again as a consequence of the structure of $A$.
The coefficients $\tau_{i_k}$ depend on the structure of the roots of type $C_{\ell}$, which we recall below. The positive roots of type $C_{\ell}$ are
\begin{equation}
\begin{split}
\Delta_+=&\left\{ \sum_{i=n_1}^{n_2}\gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell\right\}\cup \cup \left\{\sum_{i=n_1}^{\ell-1}2 \gamma_i+\gamma_{\ell}\ \vert\ 1\leq n_1\leq \ell-1\right\}\cup \\ & \left\{ \sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell-1}2 \gamma_i+\gamma_{\ell}\ \vert\ 1\leq n_1\leq n_2\leq \ell-2\right\},
\end{split}
\end{equation}
and the positive compact roots are
\begin{equation}
\begin{split}\label{eq:rootscn}
\Delta^c_+=&\left\{ \sum_{i=n_1}^{n_2}\gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell,\ \lvert S\cap\{n_1,\ldots,n_2\}\rvert\in 2\mathbb{Z}\right\}\cup \\
& \cup\left\{ \sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell-1}2 \gamma_i+\gamma_{\ell}\ \vert\ 1\leq n_1\leq n_2\leq \ell-2,\ \lvert S\cap \{n_1,\ldots,n_2,\ell\}\rvert\in 2\mathbb{Z}\right\}\cup\\
& \cup\left\{ \sum_{i=n_1}^{\ell-1}2 \gamma_i+\gamma_{\ell}\ \vert\ 1\leq n_1\leq n-1,\ \ell\notin S\right\}
\end{split}
\end{equation}
The following theorem shows the structure of the coefficients $\xi_i$.
\begin{thm}\label{thm: xiCnall}
Given a Vogan diagram of type $C_{\ell}$ with $S=\lrbr{i_1,\ldots,i_m}$ the set of indices of simple non-compact roots, for $i_j\in S$ one has
\begin{itemize}
\item If $i_j\neq \ell-1$
\begin{itemize}
\item If $j\neq m$ and $\ell\notin S$
\begin{equation}\label{eq: xi_Cn1}
\xi_{i_j}=2\left(-i_{j-1}+4i_j-3i_{j+1} +4S_+(j)+4(-1)^{j+m}\right).
\end{equation}
\item If $j\neq m$ and $\ell\in S$
\begin{equation}\label{eq: xi_Cn2}
\xi_{i_j}= 2\left(3i_{j-1}-4i_j+i_{j+1}+4S_-(j)\right).
\end{equation}
\item If $j=m$ and $i_j=\ell$
\begin{equation}\label{eq: xi_Cn3}
\xi_{\ell}= 2\left(\ell-i_{m-1}-3+2S_-(j) \right).
\end{equation}
\item If $j=m$ and $i_j\neq \ell$
\begin{equation}\label{eq: xi_Cn4}
\xi_{i_m}= 2\left(3i_m-i_{m-1}-2\ell+1\right).
\end{equation}
\end{itemize}
\item If $i_j=\ell-1$
\begin{itemize}
\item If $\ell\in S$
\begin{equation}\label{eq: xi_Cn5}
\xi_{\ell-1}= 2(\ell+3i_{j-1}+6+ 4S_-(j)).
\end{equation}
\item If $\ell\notin S$
\begin{equation}\label{eq: xi_Cn6}
\xi_{\ell-1}= 2(\ell-i_{j-1}).
\end{equation}
\end{itemize}
\end{itemize}
\end{thm}
\begin{proof}
Assume that $i_0=0$ and $i_{m+1}=\ell+1$.
\begin{itemize}
\item Case $i_j\neq \ell-1$.
\begin{itemize}
\item Case $j\neq m$, $\ell\notin S$. We will prove that
\begin{equation}\label{eq: deltac_Cn1}
\tau_{i_j}=-i_{j-1}+2i_j-i_{j+1} +2(S_1(j)- T_2(j))+2(-1)^{j+m},
\end{equation}
that, together with equalities \eqref{eq:gammaC} and \eqref{eq: formulacncoefficients} proves \eqref{eq: xi_Cn1}. We need to compute
\begin{equation}
\tau_{i_j}=-\delta_{i_j-1}^c+2\delta^c_{i_j}-\delta_{i_j+1}^c=s_1-t_1+s_2-t_2.
\end{equation}
We start with the coefficient $s_1$. The compact roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j}-\alpha_{i_j-1}=1,2$ are the ones for which $\alpha_{i_j}=1$ and $\alpha_{i_j-1}=0$, or $\alpha_{i_j}=2$ and $\alpha_{i_j-1}=1$, or $\alpha_{i_j}=2$ and $\alpha_{i_j-1}=0$. Following similar arguments to the ones used in the proof of case $B_{\ell}$, and looking and the structure of the roots \eqref{eq:rootscn}, one has
\begin{equation}
\begin{split}
s_1=& 2\sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}+\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}+i_{j-2k-1}+(i_j-i_{j-1}-1)+2+ \\&-2\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}} \\
=& 2S_1(j)+T_1(j)+i_j-i_{j-1}+1-2\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}},
\end{split}
\end{equation}
where the term $\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}$ comes from the fact that $\alpha_{\ell}$ is fixed and equal to $1$ in the roots $\sum_{l=i_j}^{n_1}\gamma_l+\sum_{l=n_1+1}^{\ell-1}2 \gamma_l+\gamma_{\ell}$ and $\sum_{l=n_1}^{i_j-1}\gamma_l+\sum_{l=i_j}^{\ell-1}2 \gamma_i+\gamma_{\ell}$.
As above, the coefficients $s_2$ and $t_1$ are precisely $s_2=S_2(j)$ and $t_1=T_1(j)$.
Similarly to $s_1$, for $t_2$ one has that the roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j+1}-\alpha_{i_j}=1,2$ are the ones for which $\alpha_{i_j+1}=1$ and $\alpha_{i_j}=0$, or $\alpha_{i_j+1}=2$ and $\alpha_{i_j}=1$, or $\alpha_{i_j+1}=2$ and $\alpha_{i_j}=0$, hence
\begin{equation}
\begin{split}
t_2=& 2\sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}+(i_{j+1}-i_j-1)+\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}+2+\\&-2\delta_{\ell+1,i_{j+2\lceil\frac{m -(j+1)}{2} \rceil+1}} \\
=& i_{j+1}-i_j+1+2T_2(j)+S_1(j)-2\delta_{\ell+1,i_{j+2\lceil\frac{m -(j+1)}{2} \rceil+1}}
\end{split}
\end{equation}
Finally
\begin{equation}
s_1+s_2-t_1-t_2=-i_{j-1}+2i_j-i_{j+1}+2(S_1(j)- T_2(j))+2(-1)^{j+m}.
\end{equation}
This proves equality \eqref{eq: deltac_Cn1}.
\item Case $j\neq m$, $\ell\in S$. We will prove that
\begin{equation}\label{eq: deltac_Cn2}
\tau_{i_j}= i_{j-1}-2i_j+i_{j+1}+2(S_2(j) -T_1(j)).
\end{equation}
This, with equalities \eqref{eq:gammaC} and \eqref{eq: formulacncoefficients} proves \eqref{eq: xi_Cn2}. Recall that
\begin{equation}
\tau_{i_j}=-\delta_{i_j-1}^c+2\delta^c_{i_j}-\delta_{i_j+1}^c=s_1-t_1+s_2-t_2.
\end{equation}
The computations are similar as the previous case, except that the fact that $\gamma_{\ell}\in S$ changes the compactness of the roots with respect to the above point. Hence we have
\begin{equation}
\begin{split}
s_1=& \sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}+
\sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}+ (i_{j+1}-i_j)+\\
& + \sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}+i_{j-2k}-2\delta_{\ell+1,j+2\lceil\frac{m -j}{2} \rceil} \\
=& i_{j+1}-i_j+S_1(j)+T_2(j)+S_2(j)-2\delta_{\ell+1,j+2\lceil\frac{m -j}{2} \rceil}\\
t_2=&\sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}+\sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}+ \\
&\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}+(i_{j}-i_{j-1})-2\delta_{\ell+1,i_{j+2\lceil\frac{m -(j+1)}{2} \rceil+1}} \\
=& i_{j}-i_{j-1}+T_2(j)+S_1(j)+T_1(j)-2\delta_{\ell+1,i_{j+2\lceil\frac{m -(j+1)}{2} \rceil+1}},
\end{split}
\end{equation}
and, again $s_2=S_2(j)$, $t_1=T_1(j)$. Finally
\begin{equation}
s_1+s_2-t_1-t_2=i_{j-1}-2i_j+i_{j+1}+2(S_2(j) -T_1(j))+2(-1)^{j+m}.
\end{equation}
This proves equality \eqref{eq: deltac_Cn2}.
\item Case $j=m$ and $i_j=\ell$. We will prove that
\begin{equation}\label{eq: deltac_Cn3}
\tau_{\ell}= S_2(m)-T_1(m).
\end{equation}
This, with equalities \eqref{eq:gammaC} and \eqref{eq: formulacncoefficients} proves \eqref{eq: xi_Cn3}. In this case, $\tau_{\ell}=-\delta^c_{\ell-1}+2\delta^c_{\ell}$. Note that the compact roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{\ell}\neq 0$ are
\begin{align}
& \sum_{k=n_1}^{\ell} \gamma_k,\quad \lvert S\cap\lrbr{n_1,\ldots,\ell}\rvert\in 2\mathbb{Z}\\
& \sum_{k=n_1}^{n_2} \gamma_k+\sum_{k=n_2+1}^{\ell-1} 2\gamma_k+\gamma_{\ell},\quad \lvert S\cap\lrbr{n_1,\ldots,n_2}\rvert\in 1+2\mathbb{Z},
\end{align}
with coefficient $\alpha_{\ell}=1$. On the other hand, the compact roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{\ell-1}\neq 0$ are
\begin{align}
& \sum_{k=n_1}^{\ell-1} \gamma_k,\quad \lvert S\cap\lrbr{n_1,\ldots,\ell-1}\rvert\in 2\mathbb{Z}\\
& \sum_{k=n_1}^{\ell-1} \gamma_k+\gamma_{\ell},\quad \lvert S\cap \lrbr{n_1,\ldots,\ell-1}\rvert\in 1+2\mathbb{Z}\\
& \sum_{k=n_1}^{n_2} \gamma_k+\sum_{k=n_2+1}^{\ell-1} 2\gamma_k+\gamma_{\ell},\quad \lvert S\cap\lrbr{n_1,\ldots,n_2}\rvert\in 1+2\mathbb{Z},
\end{align}
with coefficients $\alpha_{\ell-1}=1,1,2$ respectively. Not all the roots contribute to the summation. In particular, the ones that do are
\begin{itemize}
\item $\sum_{k=n_1}^{\ell-1} \gamma_k,\ \lvert S\cap \lrbr{n_1,\ldots,\ell-1}\rvert\in 2\mathbb{Z}$, which contribute by $-\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}+i_{j-2k}$.
\item $\sum_{k=n_1}^{\ell-1} \gamma_k+\gamma_{\ell},\quad \lvert S\cap \lrbr{n_1,\ldots,\ell-1}\rvert\in 1+2\mathbb{Z}$, which contribute by $\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}$.
\end{itemize}
By summing all the contributes we get
\begin{equation}
\tau_{\ell}= -\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}+ \sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}+i_{j-2k}=-T_1(j)+S_2(j).
\end{equation}
This proves equality \eqref{eq: deltac_Cn3}.
\item Case $j=m$ and $i_j\neq \ell$. We are going to prove that
\begin{equation}\label{eq: deltac_Cn4}
\tau_{i_m}= i_m-i_{m-1}+1.
\end{equation}
This, with equalities \eqref{eq:gammaC} and \eqref{eq: formulacncoefficients} proves \eqref{eq: xi_Cn4}. This case is similar to the case $j\neq m$, $\ell\notin S$, except that we do not have to take into account the roots with $\alpha_{m+1}\neq 0,\ldots,\alpha_{\ell}\neq 0$, since these are not included in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$. Thus
\begin{equation}
\begin{split}
s_1=&\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{m-2k}-i_{m-2k-1}+(i_m-i_{m-1}-1)+2=T_1(m)+i_m-i_{m-1}+1. \\
t_2=&\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{m-2k+1}-i_{m-2k}=S_2(m),
\end{split}
\end{equation}
and $s_2=S_2(m)$, $t_1=T_1(m)$. This reads
\begin{equation}
s_1+s_2-t_1-t_2= i_m-i_{m-1}+1,
\end{equation}
which proves equality \eqref{eq: deltac_Cn4}.
\end{itemize}
\item Case $i_j=\ell-1$. In this case we have to compute
\begin{equation}
\tau_{\ell-1}=-\delta^c_{\ell-2}+2\delta^c_{\ell-1}-2\delta^c_{\ell},
\end{equation}
and the explicit computations, together with equalities \eqref{eq:gammaC} and \eqref{eq: formulacncoefficients}, prove \eqref{eq: xi_Cn5} and \eqref{eq: xi_Cn6}.
We have to consider two cases.
\begin{itemize}
\item Case $\ell\in S$. We will prove that
\begin{equation}\label{eq: deltac_Cn5}
\tau_{\ell-1}= 3-\ell+i_{j-1}+ 2(S_2(j)-T_1(j)).
\end{equation}
The roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{\ell-1}\neq 0$ are
\begin{align}
& \sum_{k=n_1}^{\ell-1}\gamma_k,\quad \lvert S\cap\lrbr{n_1,\ldots,\ell-1}\rvert \in 2\mathbb{Z} \label{eq: n-1An}\\
& \sum_{k=n_1}^{\ell-1}\gamma_k+\gamma_{\ell},\quad \lvert S\cap\lrbr{n_1,\ldots,\ell}\rvert \in 2\mathbb{Z} \label{eq: n-1An2} \\
& \sum_{k=n_1}^{n_2}\gamma_k+\sum_{k=n_2+1}^{\ell-1}2\gamma_k+\gamma_{\ell},\quad \lvert S\cap\lrbr{n_1,\ldots,n_2}\rvert \in 1+2\mathbb{Z}, \label{eq: n-1Bn}
\end{align}
with coefficient $\alpha_{\ell-1}=1$, $\alpha_{\ell-1}=1$ and $\alpha_{\ell-1}=2$ respectively. Call $L=\sum_{i=1}^{\ell} \lambda_i\gamma_i$ the sum of all roots of type \eqref{eq: n-1An}, $M=\sum_{i=1}^{\ell} \mu_i\gamma_i$ the sum of roots of type \eqref{eq: n-1An2} and $N=\sum_{i=1}^{\ell} \nu_i\gamma_i$ the sum of the roots of type \eqref{eq: n-1Bn}. Note that, $\delta^c_{\ell-1}=\lambda_{\ell-1}+\mu_{\ell-1}+\nu_{\ell-1}$. On the other hand, the roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{\ell}\neq 0$ are \eqref{eq: n-1An2} and \eqref{eq: n-1Bn}, both with coefficient $\alpha_{\ell}=1$. In particular, $\delta^c_{\ell}=\mu_{\ell-1}+\frac{1}{2}\nu_{\ell-1}$. So
\begin{equation}
\begin{split}
2\delta^c_{\ell-1}-2\delta^c_{\ell}=& 2\lambda_{\ell-1}+2\mu_{\ell-1}+2\nu_{\ell-1}-2\mu_{\ell-1}-\nu_{\ell-1}\\
=& \lambda_{\ell-1}+\mu_{\ell-1}+\nu_{\ell-1}+\lambda_{\ell-1}-\mu_{\ell-1}\\
=& \delta^c_{\ell-1}+\lambda_{\ell-1}-\mu_{\ell-1},
\end{split}
\end{equation}
which reads
\begin{equation}
\tau_{\ell-1}= s_1-t_1+\lambda_{\ell-1}-\mu_{\ell-1}.
\end{equation}
So it suffices to compute the coefficients $s_1, t_1, \lambda_{\ell-1},\mu_{\ell-1}$. We start with $\lambda_{\ell-1}$. This coefficient is given by the $(\ell-1)$th coefficient of the sum of the roots of type \eqref{eq: n-1An}, that corresponds to
\begin{equation}
\lambda_{\ell-1}=\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}=S_2(j).
\end{equation}
Similarly, the coefficient $\mu_{\ell-1}$ corresponds to
\begin{equation}
\mu_{\ell-1}=\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}+(i_j-i_{j-1}-1)=i_j-i_{j-1}-1+T_1(j).
\end{equation}
Then, similar computations as in the other sections reads $s_1=S_2(j)+1$, since we need to take into account also the root $\gamma_{\ell-1}+\gamma_{\ell}$, and and $t_1=T_1(j)$. Finally
\begin{equation}
s_1-t_1+\lambda_{\ell-1}-\mu_{\ell-1}= 3-\ell+i_{j-1}+ 2(S_2(j)-T_1(j)).
\end{equation}
This proves equality \eqref{eq: deltac_Cn5}.
\item Case $\ell\notin S$. We will prove that
\begin{equation}\label{eq: deltac_Cn6}
\tau_{\ell-1}= \ell-i_{j-1}
\end{equation}
The roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{\ell-1}\neq 0$ are
\begin{align}
& \sum_{k=n_1}^{\ell-1}\gamma_k,\quad \lvert S\cap\lrbr{n_1,\ldots,\ell-1}\rvert \in 2\mathbb{Z} \label{eq: n-1Anf}\\
& \sum_{k=n_1}^{\ell-1}\gamma_k+\gamma_{\ell},\quad \lvert S\cap\lrbr{n_1,\ldots,\ell-1}\rvert \in 2\mathbb{Z} \label{eq: n-1An2f} \\
& \sum_{k=n_1}^{n_2}\gamma_k+\sum_{k=n_2+1}^{\ell-1}2\gamma_k+\gamma_{\ell},\quad \lvert S\cap\lrbr{n_1,\ldots,n_2}\rvert \in 2\mathbb{Z}, \label{eq: n-1Bnf}
\end{align}
with coefficients $\alpha_{\ell-1}=1$, $\alpha_{\ell-1}=1$ and $\alpha_{\ell-1}=2$ respectively. Similarly as above, it turns out that $2\delta^c_{\ell}=\delta^c_{\ell-1}$, hence $2\delta^c_{\ell-1}-2\delta^c_{\ell}=\delta^c_{\ell-1}$. This, in turns, gives
\begin{equation}
\tau_{\ell-1}= -\delta^c_{\ell-1}+s_1-t_1+\delta^c_{\ell-1}=s_1-t_1.
\end{equation}
So it is enough to compute the coefficients $s_1, t_1$, and this can be done with the techniques used for the previous points and proofs. Hence
\begin{equation}
\begin{split}
s_1=&\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}+(i_j-i_{j-1}-1)+2=T_1(j)+i_j-i_{j-1}+1\\
t_1=&\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}=T_1(j),
\end{split}
\end{equation}
and this implies
\begin{equation}
s_1-t_1 = \ell-i_{j-1}.
\end{equation}
This proves equality \eqref{eq: deltac_Cn6} and concludes the proof of the theorem.
\end{itemize}
\end{itemize}
\end{proof}
The next section is the last one and concludes the computation of the paper.
\subsection{$D_{\ell}$ family}\label{sec:dnCoefficients}
In this final section we are going to study the coefficients $\xi_i$ for Vogan diagrams of type $D_{\ell}$. Consider a Vogan diagram of type $D_{\ell}$ with $S=\lrbr{i_1,\ldots,i_m}$ the set of indices of simple non-compact roots and denote by $\{\gamma_1,\ldots,\gamma_{\ell}\}$ the set of simple roots. As in the previous cases, with the definition of the Cartan matrix $A$ for Lie algebras of type $D_{\ell}$ \cite[Sec.~11.4,Tab.~1]{Humphreys1972}, we get
\begin{equation}\label{eq: gammadn}
\begin{split}
\Gamma_{i_k}=&\left( A\sum_{\alpha\in\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}}\alpha\right)_{i_k}\\
=&\left\{ \begin{array}{lcl}
-\mathrm{dim}(\mathfrak{h}_{i_{k}})-\mathrm{dim}(\mathfrak{h}_{i_{k+1}})-\delta_{\nu\in S}\delta_{\nu'\notin S}\delta_{k=m-1} & \text{if} &k\neq m,\ i_k\neq \ell-2\\
-\mathrm{dim}(\mathfrak{h}_{i_{m}})-2\mathrm{dim}(\mathfrak{h}_{i_{m+1}})+2& \text{if} &k= m, i_k\neq \ell-2 \\
-\mathrm{dim}(\mathfrak{h}_{i_{k}})-(\delta_{\ell-1\notin S}+\delta_{\ell\notin S}) & \text{if} &i_k= \ell-2\\
-2\mathrm{dim}(\mathfrak{h}_{i_{k}}) & \text{if} & i_k=\ell-1,\ell
\end{array}\right.\\
=&\left\{ \begin{array}{lcl}
i_{k-1}-i_{k+1}+2-\delta_{\nu\in S}\delta_{\nu'\notin S}\delta_{k=m-1} & \text{if} & k\neq m,\ i_k\neq \ell-2\\
-2\ell+i_m+ i_{m-1}+3 & \text{if} & k= m, i_k\neq \ell-2 \\
-\ell+i_{k-1}+3-\delta_{\ell-1\notin S}-\delta_{\ell\notin S} & \text{if} & i_k= \ell-2\\
-2\ell+2i_{k-1}+4 & \text{if} & i_k=\ell-1,\ell
\end{array}\right.,
\end{split}
\end{equation}
with $\nu$ and $\nu'$ equal to $\ell-1$, $\ell$ or $\ell$, $\ell-1$. The $\delta$ factor in the first case comes from the fact that if $k=m-1$ and $i_m=\nu$, then there is another root that contributes to the coefficient, that is $\sum_{i=i_k+1}^{\ell-1}\gamma_i+\gamma_{\nu'}$. The same holds for $\nu'$. For the third case the second summand comes from the fact that the roots of the Lie subalgebra with Dynkin diagram having vertices with indices greater than $\ell-2$ are only: $\gamma_{\ell-1}$ if $\gamma_{\ell}\in S$, $\gamma_{\ell}$ if $\gamma_{\ell-1}\in S$, $\gamma_{\ell-1}+\gamma_{\ell}$ if $\gamma_{\ell-1},\gamma_{\ell}\notin S$, none if $\gamma_{\ell-1},\gamma_{\ell}\in S$. The other terms can be computed as in previous families. Moreover
\begin{equation}\label{eq: xiCoefficientsDnPre}
\xi_{i_k}=-4-2\Gamma_{i_k}+4\left( A\sum_{\alpha\in\Delta_+^c}\alpha\right)_{i_k}=2\left(-2+\Gamma_{i_k}+2\tau_{i_k}\right),
\end{equation}
where $\delta^c=\sum_{\alpha\in\Delta_+^c\setminus\mathrm{span}\lrbr{\gamma_i\vert i\in S^c}}\alpha$ and $\tau_{i_k}:=\left(A\delta^c\right)_{i_k}$. As a consequence of the structure of the Cartan matrix $A$, the coefficients $\tau_{i_k}$ can be written explicitly as
\begin{equation}
\tau_{i_k}=\left\{\begin{array}{lcl}
-\delta^c_{i_k-1}+2\delta^c_{i_k}-\delta^c_{i_k+1} & \text{if} & i_k\neq \ell-2,\ell-1,\ell \\
-\delta^c_{\ell-3}+2\delta^c_{\ell-2}-\delta^c_{\ell-1}-\delta^c_{\ell} & \text{if} & i_k=\ell-2\\
-\delta^c_{\ell-2}+2\delta^c_{\ell-1} & \text{if} & i_j=\ell-1,\ell\\
\end{array}\right..
\end{equation}
Note that the coefficients $\tau_{\ell-1}$ and $\tau_{\ell}$ are the same, by symmetry of Dynkin diagrams of type $D_{\ell}$.
Again, the goal is to compute explicitly the coefficients $\tau_{i_k}$ in terms of the indices of the simple non-compact roots of the Vogan diagram. Since the computations depend on the structure of the roots of type $D_{\ell}$, we recall them here below. The roots for Lie algebra of type $D_{\ell}$ are
\begin{equation}
\begin{split}
\Delta_+=&\left\{ \sum_{i=n_1}^{n_2}\gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell-2\right\}\cup \left\{ \sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{j}\ \vert\ 1\leq n_1\leq \ell-2,\ j=\ell-1,\ell\right\} \cup\\
& \cup \left\{ \sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell-1}+\gamma_{\ell}\ \vert\ 1\leq n_1\leq \ell-2\right\}\cup \\ & \cup \left\{ \sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell-2}2 \gamma_i+\gamma_{\ell-1}+\gamma_{\ell}\ \vert\ 1\leq n_1\leq n_2\leq \ell-3\right\}
\end{split}
\end{equation}
and the positive compact roots are
\begin{equation}\label{eq: compactRootsDn}
\begin{split}
\Delta^c_+=&\left\{ \sum_{i=n_1}^{n_2}\gamma_i\ \vert\ 1\leq n_1\leq n_2\leq \ell-2,\ \lvert S\cap\{n_1,\ldots,n_2\}\rvert\in 2\mathbb{Z}\right\}\cup \\
& \cup \left\{ \sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{j}\ \vert\ 1\leq n_1\leq \ell-1,\ j=\ell-1,\ell,\ \lvert S\cap\{n_1,\ldots,\ell-2,j\}\rvert\in 2\mathbb{Z}\right\}\cup \\
& \cup \left\{ \sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell-1}+\gamma_{\ell}\ \vert\ 1\leq n_1\leq \ell-2,\ \lvert S\cap\{n_1,\ldots,\ell-2,\ell-1,\ell\}\rvert\in 2\mathbb{Z}\right\}\cup\\
& \cup\left\{ \sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell-2}2 \gamma_i+\gamma_{\ell-1}+\gamma_{\ell}\ \vert\ 1\leq n_1\leq n_2\leq \ell-3,\ \lvert \right.\\ &\phantom{\cup}\lvert S\cap\{n_1,\ldots,n_2,\ell-1,\ell\}\rvert\in 2\mathbb{Z}\Bigg{\}}\\
\end{split}
\end{equation}
At this point we can write explicitly the coefficients $\xi_i$ in term of the indices in $S$.
\begin{thm}
Given a Vogan diagram of type $D_{\ell}$ with $S=\lrbr{i_1,\ldots,i_m}$ the set of indices of simple non-compact roots, for $i_j\in S$ one has
\begin{itemize}
\item If $i_j\notin\lrbr{\ell-2,\ell-1,\ell}$
\begin{itemize}
\item If $j\neq m$ and $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=0,2$
\begin{equation}\label{eq: xi_Dn1}
\xi_{i_j}= 2(-i_{j-1}+4 i_j-3i_{j+1}+4(S_+(j)+(-1)^{m+j})
\end{equation}
\item If $j\neq m$ and $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=1$
\begin{equation}\label{eq: xi_Dn2}
\xi_{i_j}= 2(3i_{j-1}-4 i_j+i_{j+1}+4S_-(j) +\delta_{j,m-1})
\end{equation}
\item If $j= m$
\begin{equation}\label{eq: xi_Dn3}
\xi_{i_m}= 2( -2\ell+3i_m-i_{m-1}-1)
\end{equation}
\end{itemize}
\item If $i_j=\ell-2$
\begin{itemize}
\item If $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=0,2$
\begin{equation}\label{eq: xi_Dn4}
\xi_{\ell-2}=2(\ell-i_{j-1}-5+2(2\delta_{i_{m-1},\ell-1}\delta_{i_m,\ell}-\delta_{\ell-1\notin S}\delta_{\ell\notin S}))
\end{equation}
\item If $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=1$
\begin{equation}\label{eq: xi_Dn5}
\xi_{\ell-2}= 2(-3\ell+3i_{j-1}+8+4S_-(j))
\end{equation}
\end{itemize}
\item If $i_j=\ell-1,\ell$
\begin{itemize}
\item If $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=1$
\begin{equation}\label{eq: xi_Dn6}
\xi_{i_j}=2(-2\ell+2i_{j-1}+2+4S_-(j))
\end{equation}
\item If $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=2$
\begin{equation}\label{eq: xi_Dn7}
\xi_{i_j}=2(\ell-4-i_{m-2})
\end{equation}
\end{itemize}
\end{itemize}
\end{thm}
\begin{proof}
As for the other proofs assume that $i_0=0$ and $i_{m+1}=\ell+1$.
\begin{itemize}
\item Case $i_j\notin\lrbr{\ell-2,\ell-1,\ell}$.
\begin{itemize}
\item Case $j\neq m$ and $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=0,2$. We will prove that
\begin{equation}\label{eq: dnformula1}
\tau_{i_j}= -i_{j-1}+2 i_j-i_{j+1}+2(S_1(j)-T_2(j))+2(-1)^{m+j},
\end{equation}
that, together with the coefficients \eqref{eq: gammadn} and \eqref{eq: xiCoefficientsDnPre} will prove identity \eqref{eq: xi_Dn1}. Recall that
\begin{equation}
\tau_{i_j}=-\delta_{i_j-1}^c+2\delta^c_{i_j}-\delta_{i_j+1}^c=s_1-t_1+s_2-t_2.
\end{equation}
Similarly as before, by looking at the compact roots listed in \eqref{eq: compactRootsDn} one has
\begin{equation}
\begin{split}
s_1=& \sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}+\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}+(i_j-i_{j-1}-1)+ \\ &+\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1} + \sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}-3\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}\\
=& i_j-i_{j-1}-1+2S_1(j)+T_1(j)+\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}-2\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}.
\end{split}
\end{equation}
The coefficient $\delta_{\ell+1,i_{j+2\lceil\frac{m -j}{2} \rceil}}$ appears since the root $\sum_{l=i_j}^{\ell-2}\gamma_l+\gamma_{\ell-1}+\gamma_{\ell}$ gives contribution only if $|\lrbr{i_{j+1},\ldots,i_m}|$ is even. Similarly, $3\delta_{\ell+1,i_{j+2\lceil\frac{m -j-1}{2} \rceil+1}}$ appears to avoid overcounting the roots of type $\sum_{l=i_j}^{n_2}\gamma_l+\sum_{l=n_2+1}^{\ell-2}2\gamma_l+\gamma_{\ell-1}+\gamma_{\ell}$ with $|S\cap\lrbr{i_j,\ldots,\ell-2}|\in 2\mathbb{Z}$.
For $s_2$ and $t_1$ we have $s_2=S_2(j)$, $t_2=T_2(j)$. Finally, for $t_2$, similarly as for $s_1$ we have
\begin{equation}
\begin{split}
t_2= & \sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}+\delta_{\ell+1,i_{j+2\lceil\frac{m -j-1}{2} \rceil+1}}+(i_{j+1}-i_j-1)+ \\ & +\sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k} -3\delta_{\ell+1,i_{j+2\lceil\frac{m -j-1}{2} \rceil+1}} + \sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}\\
=& i_{j+1}-i_j-1 +2T_2(j)+S_2(j)-2\delta_{\ell+1,i_{j+2\lceil\frac{m -j-1}{2} \rceil+1}}.
\end{split}
\end{equation}
Hence
\begin{equation}
\begin{split}
s_1-t_1+s_2-t_2=& -i_{j-1}+2i_j-i_{j-1} +2(S_1(j)-T_2(j))+2(-1)^{m+j},
\end{split}
\end{equation}
This proves \eqref{eq: dnformula1}.
\item Case $j\neq m$ and $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=1$. We have to prove that
\begin{equation}\label{eq: dnformula2}
\tau_{i_j}= i_{j-1}-2 i_j+i_{j+1}+2(S_2(j)-T_1(j)) +\delta_{j,m-1}.
\end{equation}
This, together with the coefficients \eqref{eq: gammadn} and \eqref{eq: xiCoefficientsDnPre} will prove identity \eqref{eq: xi_Dn2}. Without loss of generality we can assume $\ell-1\in S$, since the same result holds for $\ell\in S$, by symmetry. We need to compute
\begin{equation}
\tau_{i_j}=-\delta_{i_j-1}^c+2\delta^c_{i_j}-\delta_{i_j+1}^c=s_1-t_1+s_2-t_2.
\end{equation}
The computations for $s_1$, $s_2$ and $t_1$ are similar as above, hence
\begin{equation}
\begin{split}
s_1=& \sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}+ \delta_{\ell+1,i_{j+2\lceil \frac{m-j-1}{2}\rceil+1}} +\sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}+ \\
& -\delta_{\ell+1,i_{j+2\lceil \frac{m-j}{2}\rceil }} -2\delta_{\ell+1,i_{j+2\lceil \frac{m-j-1}{2}\rceil +1}} + (i_{j+1}-i_j)+ \sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k} \\
=& i_{j+1}-i_j+S_1(j)+T_2(j)+S_2(j)-\delta_{\ell+1,i_{j+2\lceil \frac{m-j}{2}\rceil} } -\delta_{\ell+1,i_{j+2\lceil \frac{m-j-1}{2}\rceil+1}},
\end{split}
\end{equation}
and $s_2=S_2(j)$, $t_1=T_1(j)$. For $t_2$ we need to count the compact roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{i_j+1}-\alpha_{i_j}=1$. In order to do that we can proceed as for the other cases, obtaining
\begin{equation}
\begin{split}
t_2=& \sum_{k=1}^{\lceil\frac{m -j-1}{2} \rceil}i_{j+2k+1}-i_{j+2k}+\delta_{\ell+1,i_{j+2\lceil \frac{m-j}{2}\rceil}}+\sum_{k=1}^{\lceil\frac{m -j}{2} \rceil}i_{j+2k}-i_{j+2k-1}+ \\ &-\delta_{j,m-1} -2\delta_{\ell+1,i_{j+2\lceil \frac{m-j}{2}\rceil}}-\delta_{\ell+1,i_{j+2\lceil \frac{m-j-1}{2}\rceil+1}}+(i_j-i_{j-1})+ \\ &+\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}\\
=& i_j-i_{j-1}+T_2(j)+S_1(j)+T_1(j) -\delta_{\ell+1,i_{j+2\lceil \frac{m-j}{2}\rceil}} +\\ &-\delta_{\ell+1,i_{j+2\lceil \frac{m-j-1}{2}\rceil+1}}-\delta_{j,m-1}.
\end{split}
\end{equation}
In this case we see a $\delta_{j,m-1}$ as last summand. It comes from the fact that, when $j=m-2$, the contribute from $T_2(j)+\delta_{\ell+1,i_{j+2\lceil \frac{m-j}{2}\rceil}}$ should be $0$, since $\gamma_{m-2}+\ldots+\gamma_{\ell-2}+\gamma_{\ell}\notin \Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$. However, this contribute is $1$, since $\delta_{\ell+1,i_{j+2\lceil \frac{m-j}{2}\rceil}}=1$, so we need to subtract $1$ in order to get the correct number of roots. Summing all the contributes we get
\begin{equation}
s_1-t_1+s_2-t_2=i_{j-1}-2i_j+i_{j+1}+2(S_2(j)-T_1(j))+\delta_{j,m-1}.
\end{equation}
This proves formula \eqref{eq: dnformula2}.
\item Case $j=m$. We prove
\begin{equation}\label{eq: dnformula3}
\tau_{i_m}= i_m-i_{m-1}-1,
\end{equation}
that implies identity \eqref{eq: xi_Dn2}, together with the coefficients \eqref{eq: gammadn} and \eqref{eq: xiCoefficientsDnPre}. In this case we have
\begin{equation}
\tau_{i_j}=-\delta_{i_j-1}^c+2\delta^c_{i_j}-\delta_{i_j+1}^c=s_1-t_1+s_2-t_2,
\end{equation}
and the computations are as before. Hence
\begin{equation}
s_1=(i_j-i_{j-1}-1)+\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}=i_j-i_{j-1}-1+T_1(j),
\end{equation}
$s_2=S_2(j)$, $t_1=T_1(j)$ and
\begin{equation}
t_2=\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}=S_2(j).
\end{equation}
Thus
\begin{equation}
s_1-t_1+s_2-t_2=i_m-i_{m-1}-1,
\end{equation}
and this proves \eqref{eq: dnformula3}.
\end{itemize}
\item Case $i_j=\ell-2$. In this case we need to compute
\begin{equation}
\tau_{i_j}=-\delta^c_{\ell-3}+2\delta^c_{\ell-2}-\delta^c_{\ell-1}-\delta^c_{\ell}.
\end{equation}
There are two cases to consider.
\begin{itemize}
\item Case $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=0,2$. We will prove that
\begin{equation}\label{eq: dnformula4}
\tau_{\ell-2}=\ell-3-i_{j-1}+2\delta_{i_{m-1},\ell-1}\delta_{i_m,\ell},
\end{equation}
that implies identity \eqref{eq: xi_Dn4}, together with the coefficients \eqref{eq: gammadn} and \eqref{eq: xiCoefficientsDnPre}. This point is a little technical, so we will illustrate most of the computations, which will also be used in the subsequent points. The compact roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{\ell-2}\neq 0$ are
\begin{align}
&\sum_{i=n_1}^{\ell-2}\gamma_i,\quad |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z} \\
&\sum_{i=n_1}^{\ell-2}\gamma_i+\nu,\ \nu=\gamma_{\ell-1},\gamma_{\ell},\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 1+2\mathbb{Z}, \text{if}\ \gamma_{\ell-1},\gamma_{\ell}\in S \\
&\sum_{i=n_1}^{\ell-2}\gamma_i+\nu,\ \nu=\gamma_{\ell-1},\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z}, \text{if}\ \gamma_{\ell-1},\gamma_{\ell}\notin S \\
&\sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z} \\
&\sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell-2}2\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,n_2}|\in 2\mathbb{Z},
\end{align}
with coefficients $\alpha_{\ell-2}=1,1,1,1,2$ respectively. On the other hand, the compact roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{\ell-1}\neq 0$ or $\alpha_{\ell}\neq 0$ are
\begin{align}
&\sum_{i=n_1}^{\ell-2}\gamma_i+\nu,\ \nu=\gamma_{\ell-1},\gamma_{\ell},\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 1+2\mathbb{Z}, \text{if}\ \gamma_{\ell-1},\gamma_{\ell}\in S \\
&\sum_{i=n_1}^{\ell-2}\gamma_i+\nu,\ \nu=\gamma_{\ell-1},\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z}, \text{if}\ \gamma_{\ell-1},\gamma_{\ell}\notin S \\
&\sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z} \\
&\sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell-2}2\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,n_2}|\in 2\mathbb{Z},
\end{align}
with coefficients $0$ or $1$. Moreover, the compact roots in $\Delta^c_+\setminus\mathrm{span}\lrbr{\gamma_i \vert i\in S^c}$ such that $\alpha_{\ell-3}\neq 0$ are
\begin{align}
&\sum_{i=n_1}^{\ell-3}\gamma_i,\quad |S\cap\lrbr{n_1,\ldots,\ell-3}|\in 2\mathbb{Z},\ n_1\leq \ell-3 \\
&\sum_{i=n_1}^{\ell-2}\gamma_i,\quad |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z},\ n_1\leq \ell-3 \\
&\sum_{i=n_1}^{\ell-2}\gamma_i+\nu,\ \nu=\gamma_{\ell-1},\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 1+2\mathbb{Z}, \\ & \phantom{spazio}\text{if}\ \gamma_{\ell-1},\gamma_{\ell}\in S,\ n_1\leq \ell-3 \\
&\sum_{i=n_1}^{\ell-2}\gamma_i+\nu,\ \nu=\gamma_{\ell-1},\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z},\\ & \phantom{spazio}\text{if}\ \gamma_{\ell-1},\gamma_{\ell}\notin S,\ n_1\leq \ell-3 \\
&\sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z},\ n_1\leq \ell-3 \\
&\sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell-2}2\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,n_2}|\in 2\mathbb{Z},\ n_2\leq \ell-4 \\
&\sum_{i=n_1}^{\ell-3}\gamma_i+2\gamma_{\ell-2}+\gamma_{\ell-1}+\gamma_{\ell},\quad |S\cap\lrbr{n_1,\ldots,\ell-3}|\in 2\mathbb{Z},\ n_1\leq \ell-3.
\end{align}
Not all the roots contribute to the summation. In particular, the ones that do are
\begin{itemize}
\item $\sum_{i=n_1}^{\ell-3}\gamma_i,\ |S\cap\lrbr{n_1,\ldots,\ell-3}|\in 2\mathbb{Z},\ n_1\leq \ell-3$. These contribute to $-\delta^c_{\ell-3}+2\delta^c_{\ell-2}-\delta^c_{\ell-1}-\delta^c_{\ell}$ only for the $-\delta^c_{\ell-3}$ part, since the other $\delta^c_{\ell-2},\delta^c_{\ell-1},\delta^c_{\ell}$ give contribution $0$. So the given contribute is $-\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}$.
\item $\sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z},\ n_1\leq \ell-3$. These contribute to $-\delta^c_{\ell-3}+2\delta^c_{\ell-2}-\delta^c_{\ell-1}-\delta^c_{\ell}$ only for the $-\delta^c_{\ell-3}$ part, since $\delta^c_{\ell-2},\delta^c_{\ell-1},\delta^c_{\ell}$ factors cancel out. So the given contribute is $-\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}$.
\item $\sum_{i=n_1}^{\ell-2}\gamma_i,\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z},\ n_1\leq \ell-3$. These contribute to $-\delta^c_{\ell-3}+2\delta^c_{\ell-2}-\delta^c_{\ell-1}-\delta^c_{\ell}$ only for the $-\delta^c_{\ell-3}+2\delta^c_{\ell-2}$ part, since $\delta^c_{\ell-1},\delta^c_{\ell}$ give contribution $0$. So the given contribute is $\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}$.
\item $\sum_{i=n_1}^{n_2}\gamma_i+\sum_{i=n_2+1}^{\ell-2}2\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\ |S\cap\lrbr{n_1,\ldots,n_2}|\in 2\mathbb{Z},\ n_2\leq \ell-3$. These contribute to the whole coefficient $-\delta^c_{\ell-3}+2\delta^c_{\ell-2}-\delta^c_{\ell-1}-\delta^c_{\ell}$, by $+(\ell-2-i_{j-1}-1)+\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}$.
\item The roots $\gamma_{\ell-2}+\gamma_{\ell-1}$, $\gamma_{\ell-2}+\gamma_{\ell}$, if $\gamma_{\ell-1},\gamma_{\ell}\in S$. These contribute to $-\delta^c_{\ell-3}+2\delta^c_{\ell-2}-\delta^c_{\ell-1}-\delta^c_{\ell}$ only for the $2\delta^c_{\ell-2}-\delta^c_{\ell-1}-\delta^c_{\ell}$ part, with a total contribute of $2$.
\end{itemize}
So
\begin{equation}
\tau_{\ell-2}= \ell-3-i_{j-1}+2\delta_{i_{m-1},\ell-1}\delta_{i_m,\ell},
\end{equation}
amd this proves equality \eqref{eq: dnformula4}.
\item Case $\lvert S\cap \lrbr{\ell-1,\ell}\rvert=1$. We prove that
\begin{equation}\label{eq: dnformula5}
\tau_{\ell-2}= -i_j+i_{j-1}+2+2(S_2(j)-T_1(j))
\end{equation}
that implies identity \eqref{eq: xi_Dn5}, together with the coefficients \eqref{eq: gammadn} and \eqref{eq: xiCoefficientsDnPre}. Without loss of generality we can assume that $\gamma_m=\gamma_{\ell-1}$, and the proof in this case is similar to the one of the previous point. In particular, the roots that contribute in the summation are
\begin{itemize}
\item $\sum_{i=n_1}^{\ell-3}\gamma_i,\ |S\cap\lrbr{n_1,\ldots,\ell-3}|\in 2\mathbb{Z},\ n_1\leq \ell-3$, which contribute by $-\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}$.
\item $\sum_{i=n_1}^{\ell-2}\gamma_i,\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z},\ n_1\leq \ell-3$, which contribute by $\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}$.
\item $\gamma_{\ell-2}+\gamma_{\ell-1}$, which contributes by a factor $1$.
\item $\sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 1+2\mathbb{Z},\ n_1\leq \ell-3$, which contribute by $-(\ell-1-i_{j-1}-1)-\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}$.
\item $\sum_{i=n_1}^{\ell-3}\gamma_i+2\gamma_{\ell-2}+\gamma_{\ell-1}+\gamma_{\ell},\ |S\cap\lrbr{n_1,\ldots,\ell-3}|\in 1+2\mathbb{Z},\ n_1\leq \ell-3$, which contribute by $\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}$.
\end{itemize}
By summing all the contributes we get
\begin{equation}
\begin{split}
\tau_{\ell-1}= -\ell+3+i_{j-1}+2(S_2(j)-T_1(j)),
\end{split}
\end{equation}
which proves \eqref{eq: dnformula5}.
\end{itemize}
\item Case $i_j=\ell-1,\ell$. Without loss of generality we can assume $i_j=\ell-1$. For this case we have to compute
\begin{equation}
\tau_{i_j}=-\delta^c_{\ell-2}+2\delta^c_{\ell-1}.
\end{equation}
\begin{itemize}
\item Case $|S\cap \lrbr{\gamma_{\ell-1},\gamma_{\ell}}|=1$. We compute
\begin{equation}\label{eq: dnformula6}
\tau_{i_j}=2(S_2(j) -T_1(j)),
\end{equation}
that, together with the coefficients \eqref{eq: gammadn} and \eqref{eq: xiCoefficientsDnPre}, implies identity \eqref{eq: xi_Dn6}. Similarly to the previous point, the roots that contribute in the summation are
\begin{itemize}
\item $\sum_{i=n_1}^{\ell-2}\gamma_i,\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z}$, which contribute by $-\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}$.
\item $\sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell-1},\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 1+2\mathbb{Z}$, which contribute by $\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}$.
\item $\sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell},\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 2\mathbb{Z}$, which contribute by $-\sum_{k=1}^{\lfloor\frac{j+1}{2} \rfloor-1} i_{j-2k}-i_{j-2k-1}$.
\item $\sum_{i=n_1}^{\ell-2}\gamma_i+\gamma_{\ell-1}+\gamma_{\ell},\ |S\cap\lrbr{n_1,\ldots,\ell-2}|\in 1+2\mathbb{Z}$ which contribute by $\sum_{k=1}^{\lfloor\frac{j}{2} \rfloor} i_{j-2k+1}-i_{j-2k}$.
\end{itemize}
Thus one gets
\begin{equation}
\tau_{i_j}=2(S_2(j)-T_1(j)),
\end{equation}
which proves \eqref{eq: dnformula6}
\item Case $|S\cap \lrbr{\gamma_{\ell-1},\gamma_{\ell}}|=2$. We prove
\begin{equation}\label{eq: dnformula7}
\tau_{\ell-1}=\ell-2-i_{m-3},
\end{equation}
that, together with the coefficients \eqref{eq: gammadn} and \eqref{eq: xiCoefficientsDnPre}, implies identity \eqref{eq: xi_Dn7}. Similarly as above one can compute the roots that contribute to $\tau_{i_j}$, except that one has to take into account that the compactness of the roots changes. Then
\begin{equation}
\tau_{\ell-1}=\ell-2-i_{m-3},
\end{equation}
which proves identity \eqref{eq: dnformula7} and concludes the proof.
\end{itemize}
\end{itemize}
\end{proof}
\paragraph{Aknowledgments} The author is grateful to Alberto Della Vedova for discussions, ideas and thoughts in the initial stage of the work, and to G. Bruno De Luca for many corrections, suggestions and for the support in writing this paper.
\bibliographystyle{siam}
| f728da1ff55c93e0ddcbfb497cd62c729d3c9053 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Appendix}
\label{s:App}
In this section, we include additional comparisons to related work, additional definitions, proofs to the theorems in the main text, and additional experimental details.
The code to reproduce the figures and experiments is available here: \url{https://github.com/cavalab/proportional-multicalibration}.
\subsection{Related Work}
\label{s:App:related}
\paragraph{Definitions of Fairness}
There are myriad ways to measure fairness that are covered in more detail in other works~\citep{barocasFairnessMachineLearning2019,chouldechovaFrontiersFairnessMachine2018,castelnovoZooFairnessMetrics2021}.
We briefly review three notions here.
The first, \textit{demographic parity}, requires the model's predictions to be independent of patient demographics ($A$).
Although a model satisfying demographic parity can be desirable when the outcome should be unrelated to sensitive attributes~\citep{fouldsAreParityBasedNotions2020}, it can be unfair if important risk factors for the outcome are associated with those attributes~\citep{hardtEqualityOpportunitySupervised2016a}.
For example, it may be more fair to admit socially marginalized patients to a hospital at a higher rate if they are assessed less able to manage their care at home.
Furthermore, if the underlying rates of illness vary demographically, requiring demographic parity can result in a healthier patients from one group being admitted more often than patients who urgently need care.
When the base rates of admission are expected to differ demographically, we can instead ask that the model's errors be balanced across groups.
One such notion is \textit{equalized odds}, which states that for a given $Y$, the model's predictions should be independent of $A$.
Satisfying equalized odds is equivalent to having equal FPR and FNR for every group in $A$.
When the model is used for patient risk stratification, as in the target use case in this paper, it is important to consider a model's calibration for each demographic group in the data.
Because risk prediction models influence who is prioritized for care, an unfairly calibrated model can systematically under-predict risk for certain demographic groups and result in under-allocation of patient care to those groups.
Thus, guaranteeing group-wise calibration via an approach such as multicalibration also guarantees fair patient prioritization for health care provision.
In some contexts, risk predictions are not directly interpreted, but only used to \textit{rank} patients, which in some contexts is sufficient for resource allocation.
Authors have proposed various ways of measuring the fairness of model rankings, for example by comparing AUROC between groups~\citep{kallusAssessingAlgorithmicFairness2020}.
\paragraph{Approaches to Fairness}
Many approaches to achieving fairness guarantees according to demographic parity, equalized odds and its relaxations have been proposed~\citep{ dworkFairnessAwareness2012,hardtEqualityOpportunitySupervised2016a, berkConvexFrameworkFair2017,jiangIdentifyingCorrectingLabel2019a,kearnsPreventingFairnessGerrymandering2018}.
When choosing an approach, is important to carefully weigh the relative impact of false positives, false negatives, and miscalibration on patient outcomes, which differ by use case.
When group base rates differ (i.e., group-specific positivity rates), \emph{equalized odds and calibration by group cannot both be satisfied}~\citep{kleinbergInherentTradeoffsFair2016}.
Instead, one can often equalized multicalibration while satisfying relaxations of equalized odds such as \emph{equalized accuracy}, where $Accuracy = \mu TPR+(1-\mu)(1-FPR)$ for a group with base rate $\mu$.
However, to do so requires denigrating the performance of the model on specific groups~\citep{chouldechovaFairPredictionDisparate2017,pleissFairnessCalibration2017}, which is unethical in our context.
As mentioned in the introduction, we are also motivated to utilize approaches to fairness that 1) dovetail well with intersectionality theory, and 2) provide privacy guarantees.
Most work in the computer science/ machine learning space does not engage with the broader literature on socio-cultural concepts like intersectionality, which we see as a gap that makes adoption in real-world settings difficult~\citep{hanna2020towards}.
One exception to this statement is differential fairness~\citep{fouldsIntersectionalDefinitionFairness2019}, a measure designed with intersectionality in mind.
In addition to being a definition of fairness that provides equal protection to groups defined by intersections of protected attributes, models satisfying $\epsilon$-differential fairness also satisfy $\epsilon$-pufferfish privacy.
This privacy guarantee is very desirable in risk prediction contexts, because it limits the extent to which the model reveals sensitive information to a decision maker that has the potential to influence their interpretation of the model's recommendation.
However, prior work on differential fairness has been limited to using it to control for demographic parity, which is not an appropriate fairness measure for our use case~\citep{fouldsAreParityBasedNotions2020}.
Multicalibration has inspired several extensions, including relaxations such as multiaccuracy~\citep{kimMultiaccuracyBlackboxPostprocessing2019}, low-degree multicalibration~\citep{gopalanLowDegreeMulticalibration2022}, and extensions to conformal prediction and online learning~\citep{jungMomentMulticalibrationUncertainty2021,guptaOnlineMultivalidLearning2021}.
Noting that multicalibration is a guarantee over mean predictions on a collection of groups $\mathcal{C}$, \cite{jungMomentMulticalibrationUncertainty2021} propose to extend multicalibration to higher-order moments (e.g., variances), which allows one to estimate a confidence interval for the calibration error for each category.
\cite{guptaOnlineMultivalidLearning2021} extend this idea and generalize it to the online learning context, in which an adversary chooses a sequence of examples for which one wishes to quantify the uncertainty of different statistics of the predictions.
Recent work has also utilized higher order moments to ``interpolate" between the guarantees provided by multiaccuracy, which only requires accuracy in expectation for groups in $\mathcal{C}$, and multicalibration, which requires accuracy in expectation at each prediction interval~\citep{kimMultiaccuracyBlackboxPostprocessing2019}.
Like proportional multicalibration (\cref{def:PMC}), definitions of multicalibration for higher order moments provide additional criteria for quantifying model performance over many groups; in general, however, much of the focus in other work is on statistics for uncertainty estimation.
Like these works, one may view our proposal for proportional multicalibration as alternative definition of what it means to be multicalibrated.
The key difference is that proportional multicalibration measures the degree to which multicalibration depends on differences in outcome prevalence between groups, and in doing so provides guarantees of pufferfish privacy and differential calibration.
\cite{dworkLearningOutcomesEvidenceBased2019} study the relation of fair rankings to multicalibration, and, in a similar vein to differential fairness measures, formulate a fairness measure for group rankings using the relations between pairs of groups.
However, these definitions are specific to the ranking relation between the groups, whereas differential calibration cares only about the outcome differential (conditioned on model predictions) between pairs of groups.
\subsubsection{Differential Fairness}
\label{s:App:DF}
DF was explicitly defined to be consistent with the social theoretical framework of \emph{intersectionality}.
This framework dates back as early as the social movements of the '60s and '70s \citep{collins_intersectionality_2020} and was brought into the academic mainstream by pioneering work from legal scholar Kimberlé Crenshaw~\citep{crenshawDemarginalizingIntersectionRace1989,crenshaw_mapping_1991} and sociologist Patricia Hill Collins~\citep{collins_black_1990}.
Central to intersectionality is that hierarchies of power and oppression are structural elements that are fundamental to our society.
Through an intersectional lens, these power structures are viewed as interacting and co-constituted, inextricably related to one another.
To capture this viewpoint, DF~\citep{fouldsIntersectionalDefinitionFairness2019} constrains the differential of a general data mechanism among all pairs of groups, where groups are explicitly defined as the intersections of protected attributes in $\mathcal{A}$.
\begin{definition}[$\epsilon$-differential fairness~\citep{fouldsIntersectionalDefinitionFairness2019}]
\label{def:DF}
Let $\Theta$ denote a set of distributions and let $x \sim \theta$ for $\theta \in \Theta$.
A mechanism $M(x)$ is $\varepsilon$-differentially fair with respect to ($\mathcal{C}$,$\Theta$)
for all $\theta \in \Theta$ with $x \sim \theta$, and $m \in Range(M)$ if,
for all $(S_i,S_j) \in \mathcal{C} \times \mathcal{C}$ where $P(S_i|\theta)>0$, $P(S_j|\theta)>0$,
\begin{equation}\label{eq:eDF}
e^{-\varepsilon}\leq\frac{P_{M,\theta}(M(x)=m|S_i,\theta)}{P_{M,\theta}(M(x)=m|S_j,\theta)} \leq e^{\varepsilon}
\end{equation}
\end{definition}
\begin{definition}[Pufferfish Privacy]\label{def:puff}
Let the collection of subsets $\mathcal{C}$ represent sets of secrets.
A mechanism $M({x})$ is $\epsilon$-\emph{pufferfish private} \citep{kiferPufferfishFrameworkMathematical2014} with respect to $(\mathcal{C}, \Theta)$ if for all $\theta \in \Theta$ with ${x} \sim \theta$, for all secret pairs $(S_i,S_j) \in \mathcal{C} \times \mathcal{C}$ and $y \in \mbox{Range}(M)$,
\begin{equation}
e^{-\epsilon} \leq \frac{P_{M, \theta}(M(x) = y| S_i, \theta)}{P_{M, \theta}(M(x) = y|S_j, \theta)}\leq e^\epsilon \mbox{ ,} \label{def:pufferfish}
\end{equation}
when $S_i$ and $S_j$ are such that $P(S_i|\theta) > 0$, $P(S_j|\theta) > 0$.
\end{definition}
\paragraph{Note on pufferfish and differential privacy}
Although \cref{eq:eDF} is notable in its similarity to differential privacy~\citep{dwork2009differential}, they differ in important ways.
Differential privacy aims to limit the amount of information learned about any one individual in a database by computations performed on the data (e.g. $M(x)$).
Pufferfish privacy only limits information learned about the group membership of individuals as defined by $\mathcal{C}$.
\cite{kiferPufferfishFrameworkMathematical2014} describe in detail the conditions under which these privacy frameworks are equivalent.
\paragraph{Efficiency Property}
\label{s:df-efficient}
\cite{fouldsIntersectionalDefinitionFairness2019} also define an interesting property of $\varepsilon$-differential fairness that allows guarantees of higher order (i.e., marginal) groups to be met for free; the property is given in \cref{s:App:def}.
\begin{definition}[Efficiency Property~\citep{fouldsIntersectionalDefinitionFairness2019}] \label{def:inter}
Let $M(x)$ be an $\varepsilon$-differentially fair mechanism with respect to $(\mathcal{C},\Theta)$.
Let the collection of subsets $\mathcal{C}$ group individuals according to the Cartesian product of attributes $A \subseteq \mathcal{A}$.
Let $\cal G$ be any collection of subsets that groups individuals by the Cartesian product of attributes in $A'$, where $A' \subset A$ and $A' \neq \emptyset$.
Then $M(x)$ is $\varepsilon$-differentially fair in $(\cal G,\Theta)$.
\end{definition}
The authors call this the "intersectionality property", yet its implication is the opposite: if a model satisfies $\epsilon$-DF for the low level (i.e. intersectional) groups in $\mathcal{C}$, then it satisfies $\epsilon$-DF for every higher-level (i.e. marginal) group.
For example, if a model is ($\epsilon$)-differentially fair for intersectional groupings of individuals by race and sex, then it is $\epsilon$-DF for the higher-level race and sex groupings as well.
Whereas the number of intersections grows exponentially as additional attributes are protected~\citep{kearnsPreventingFairnessGerrymandering2018}, the number of total possible subgroupings grows at a larger combinatorial rate: for $p$ protected attributes, we have $\sum_{k=1}^p{ \binom{p}{k} m_a^k}$ groups, where $m_a$ is the number of levels of attribute $a$.
\paragraph{Limitations}
To date, analysis of DF for predictive modeling has been limited to defining $R(x)$ as the mechanism, which is akin to asking for \emph{demographic parity}.
Under demographic parity, one requires that model predictions be independent from group membership entirely, and this limits the utility of it as a fairness notion.
Although a model satisfying demographic parity can be desirable when the outcome should be unrelated to $\mathcal{C}$~\citep{fouldsAreParityBasedNotions2020}, it can be unfair if important risk factors for the outcome are associated with demographics~\citep{hardtEqualityOpportunitySupervised2016a}.
For example, if the underlying rates of an illness vary demographically, requiring demographic parity can result in a healthier patients from one group being admitted more often than patients who urgently need care.
\subsection{Additional Definitions}
\label{s:App:def}
\begin{definition}[$\alpha$-calibration~\citep{hebert-johnsonCalibrationComputationallyIdentifiableMasses2018}]
\label{def:calibration}
Let $S \subseteq \mathcal{X}$.
For $\alpha \in [0,1]$, $R$ is $\alpha$-\emph{calibrated} with
respect to $S$ if there exists some $S' \subseteq S$
with $\card{S'} \ge (1-\alpha)\card{S}$
such that
for all $r \in [0,1]$,
$$
\card{ \E_D [ y | R(x) = r, x \in S' ] - r} \le \alpha.
$$
\end{definition}
\begin{definition}[$\alpha$-MC~\citep{hebert-johnsonCalibrationComputationallyIdentifiableMasses2018}]
\label{def:MC}
Let $\mathcal{C} \subseteq 2^{\mathcal{X}}$ be a collection of subsets of $\mathcal{X}$, $\alpha \in [0,1]$.
A predictor $R$ is $\alpha$-multicalibrated on $\mathcal{C}$ if for all $S \in \mathcal{C}$,
$R$ is $\alpha$-calibrated with respect to $S$.
\end{definition}
We note that, according to \cref{def:calibration}, a model need only be calibrated over a sufficiently large subset of each group ($S'$) in order to satisfy the definition.
This relaxation is used to maintain a satisfactory definition of MC when working with discretized predictions.
That is, with \cref{def:calibration}, \cite{hebert-johnsonCalibrationComputationallyIdentifiableMasses2018} show that $(\alpha, \lambda)$-multicalibrated models are at most $2\alpha$-multicalibrated.
\subsubsection{Loss functions}
The following loss functions are empirical analogs of the definitions of $MC$, $PMC$, and $DC$, and are used in the experiment section to measure performance.
\begin{definition}[MC loss]
\label{def:mcloss}
Let $\mathcal{D} = \set{(y,x)_i}_{i=0}^{N} \sim D$, and let $\alpha, \lambda, \gamma > 0$.
Define a collection of subsets $\mathcal{C} \in 2^{\mathcal{X}}$ such that for all $S \in \mathcal{C}, |S| \geq \gamma N$.
Let $S_I = \set{x: R(x) \in I, x \in S}$ for $(S,I) \in \mathcal{C} \times \Lambda_\lambda$.
Define the collection $\mathcal{S}$ containing all $S_I$ satisfying $S_I \geq \alpha \lambda N$.
The MC loss of a model $R(x)$ on $\mathcal{D}$ is
\[
\max_{S_I \in \mathcal{S}}{
\frac{1}{|S_I|}
\card{\sum_{i \in S_I}{ y_i } - \sum_{i \in S_I}{ R_i }}
}
\]
\end{definition}
\begin{definition}[PMC loss]
\label{def:pmcloss}
Let $\mathcal{D} = \set{(y,x)_i}_{i=0}^{N} \sim D$, and let $\alpha, \lambda, \gamma, \rho > 0$.
Define a collection of subsets $\mathcal{C} \in 2^{\mathcal{X}}$ such that for all $S \in \mathcal{C}, |S| \geq \gamma N$.
Let $S_I = \set{x: R(x) \in I, x \in S}$ for $(S,I) \in \mathcal{C} \times \Lambda_\lambda$.
Define the collection $\mathcal{S}$ containing all $S_I$ satisfying $S_I \geq \alpha \lambda N$.
Let $\frac{1}{|S_I|}\sum_{i \in S_I}{ y_i } \geq \rho$.
The PMC loss of a model $R(x)$ on $\mathcal{D}$ is
\[
\max_{S_I \in \mathcal{S}}{
\frac{
\card{\sum_{i \in S_I}{ y_i } - \sum_{i \in S_I}{ R_i }}
}
{
\sum_{i \in S_I}{ y_i }
}
}
\]
\end{definition}
\begin{definition}[DC loss]
\label{def:dcloss}
Let $\mathcal{D} = \set{(y,x)_i}_{i=0}^{N} \sim D$, and let $\alpha, \lambda, \gamma > 0$.
Define a collection of subsets $\mathcal{C} \in 2^{\mathcal{X}}$ such that for all $S \in \mathcal{C}, |S| \geq \gamma N$.
Given a risk model $R(x)$ and prediction intervals $I$,
Let $S_I = \set{x: R(x) \in I, x \in S}$ for $(S,I) \in \mathcal{C} \times \Lambda_\lambda$.
Define the collection $\mathcal{S}$ containing all $S_I$ satisfying $S_I \geq \alpha \lambda N$.
The DC loss of a model $R(x)$ on $\mathcal{D}$ is
\[
\max_{(S_I^a,S_I^b) \in \mathcal{S} \times \mathcal{S}}{
\log{\card{
\frac{1}{|S_I^a|} \sum_{i \in S_I^a}{ y_i } - \frac{1}{|S_I^b|}\sum_{j \in S_I^b}{ y_j }
}}
}
\]
\end{definition}
\subsection{Theorem Proofs}
\label{s:proof}
\paragraph{\cref{thm:alg}}\label{proof:alg}
\textit{
\Paste{thm:alg}
}
\begin{proof}
We show that \cref{alg:PMC} converges using a potential function argument~\citep{bansalPotentialfunctionProofsGradient2019}, similar to the proof techniques for the MC boosting algorithms in \cite{hebert-johnsonMulticalibrationCalibrationComputationallyIdentifiable2018,kimMultiaccuracyBlackboxPostprocessing2019}.
Let $p^*_i$ be the underlying risk, $R_i$ be our initial model, and $R'_i$ be our updated prediction model for individual $i \in S_r$, where $S_r = \{x | x \in S, R(x) \in I\}$ and $(S,I) \in \mathcal{C} \times \Lambda_{\lambda}$.
We use $p^*$, $R$, and $R'$ without subscipts to denote these values over $S_r$.
We cannot easily construct a potential argument using progress towards ($\alpha$,$\lambda$)-PMC, since its derivative is undefined at $\E_D [ y | R \in I, x \in S]$=0.
Instead, we analyze progress towards the difference in the $\ell_2$ norm at each step.
\begin{align}
||p^*-R|| - ||p^*-R'|| &= \sum_{i \in S_r}{ (p_i^* - R_i)^2 } - \sum_{i \in S_r}{ (p_i^* - \text{squash}(R_i+\Delta r))^2 } \nonumber \\
&\geq \sum_{i \in S_r}{\left( (p_i^* - R)^2 - (p_i^* - (R_i+\Delta r))^2 \right) } \nonumber\\
&= \sum_{i \in S_r}{\left( 2p_i^* \Delta r - 2R_i \Delta r - \Delta r^2 \right) } \nonumber\\
&= 2 \Delta r \sum_{i \in S_r}{\left( p_i^* - R_i \right)} - |S_r|\Delta r^2 \label{eq:del}
\end{align}
From \cref{alg:PMC} we have
$$
\Delta r = \frac{1}{|S_r|}\sum_{i \in S_r}{( p_i^* - R_i )}
$$
Substituting into~\cref{eq:del} gives
\begin{align*}
||p^*-R|| - ||p^*-R'|| &\geq |S_r|{\Delta r}^2 \\
\end{align*}
We know that $|S_r| \geq \alpha \lambda \gamma N$, and that the smallest update $\Delta r$ is $\alpha \rho$.
Thus,
\begin{align*}
||p^*-R|| - ||p^*-R'|| &\geq \alpha^3 \rho^2 \lambda \gamma N \\
\end{align*}
Since our initial loss, $|| p^* - R||$, is at most $N$, \cref{alg:PMC} converges in at most $O(\frac{1}{\lambda^3 \rho^2 \lambda \gamma})$ updates for category $S_r$.
To understand the total number of steps, including those without updates, we consider the worst case, in which only a single category $S_r$ is updated in a cycle of the for loop (if no updates are made, the algorithm exits).
Since each repeat consists of at most $|C|/\lambda$ loop iterations, this results in $O(\frac{|C|}{\alpha^3 \lambda^2 \rho^2 \gamma})$ total steps.
\end{proof}
\subsection{Additional Theorems}\label{s:App:thm}
\subsubsection{Differentially calibrated models with global calibration are multicalibrated}
Here we show that, under the assumption that a model is globally calibrated (satisfies $\delta$-calibration), models satisfying $\varepsilon$-DC are also multicalibrated.
\begin{theorem}\label{thm:DCtoMC}
Let R(x) be a model satisfying ($\varepsilon$,$\lambda$)-DC and $\delta$-calibration.
Then $R(x)$ is ($1-e^{-\varepsilon}+\delta$, $\lambda$)-multicalibrated.
\end{theorem}
\begin{proof}
From~\cref{eq:DC} we observe that $\varepsilon$ is bounded by the two groups with the largest and smallest group- and prediction- specific probabilities of the outcome.
Let $I_M$ be the risk stratum maximizing $(\varepsilon,\lambda)$-DC, and let $p_n = \max_{S \in \mathcal{C}} P_D(y|R \in I_M, x \in S)$ and $p_d = \min_{S \in \mathcal{C}} P_D(y|R \in I_M,x \in S)$.
These groups determine the upper and lower bounds of $\varepsilon$ as $e^{-\varepsilon} \leq p_d/p_n$ and $p_n/p_d \leq e^{\varepsilon}$.
We note that $p_d \leq P_D(y|R \in I_M) \leq p_n$, since $P(y| R \in I_M) = \frac{1}{N} \sum_{S \in \mathcal{C}} |S| P_D(y|R \in I_M, x \in S)$, and $p_n$ and $p_d$ are the extreme values of $P(y|R\in I_M,x \in S)$ among $S$.
So, $\alpha$-MC is bound by the group outcome that most deviates from the predicted value, which is either $p_n$ or $p_d$.
Let $r = P_D( R|R \in I_M )$.
There are then two scenarios to consider:
\begin{enumerate}
\item $ \alpha \leq | p_n - r | = p_n - r $ when $r \leq \frac{1}{2}(p_n + p_d)$; and
\item $ \alpha \leq | p_d - r | = r - p_d $ when $r \geq \frac{1}{2}(p_n + p_d)$.
\end{enumerate}
We will look at the first case.
Let $p^*_r = P_D(y|R \in I_M)$.
Due to $\delta$-calibration, $p^*_r - \delta \leq r \leq p^*_r + \delta$.
Then
\begin{align*}
\alpha &\leq p_n - r \\
&\leq p_n - (p^*_r - \delta) \\
&\leq p_n - p_d + \delta \\
&= p_n (1-e^{-\varepsilon}) + \delta\\
\alpha &\leq 1 - e^{-\varepsilon} + \delta.
\end{align*}
Above we have used the facts that $r \leq p^*_r - \delta$, $p^*_r \geq p_d$, $p_d \leq e^{-\varepsilon}p_n$, and $p_n \leq 1$.
The second scenario is complementary and produces the identical bound.
\end{proof}
\cref{thm:DCtoMC} formally describes how $\delta$-calibration controls the baseline calibration error contribution to $\alpha$-MC, while $\varepsilon$-DC limits the deviation around this value by constraining the (log) maximum and minimum risk within each category.
\subsection{Multicalibrated models satisfy intersectional guarantees}
In contrast to DF, MC \citep{hebert-johnsonMulticalibrationCalibrationComputationallyIdentifiable2018} was not designed to explicitly incorporate the principles of intersectionality.
However, we show that it provides an identical efficiency property to DF in the theorem below.
\begin{theorem}\label{thm:intersectionalmc}
Let the collection of subsets $\mathcal{C} \subseteq 2^\mathcal{X}$ define groups of individuals according to the Cartesian product of attributes $A \subseteq \mathcal{A}$.
Let $\cal G \in 2^\mathcal{X}$ be any collection of subsets that groups individuals by the Cartesian product of attributes in $A'$, where $A' \subset A$ and $A' \neq \emptyset$.
If $R(x)$ satisfies $\alpha$-MC on $\mathcal{C}$, then $R(x)$ is $\alpha$-multicalibrated on $\cal G$.
\end{theorem}
In proving \cref{thm:intersectionalmc}, we will make use of the following lemma.
\begin{lemma}\label{lemma:express}
The $\alpha$-MC criteria can be rewritten as: for a collection of subsets $\mathcal{C} \subseteq \mathcal{X}$, $\alpha \in [0,1]$, and $r \in [0,1]$,
$$
\max_{c\in\mathcal{C}}\E_D [ y | R(x) = r, x \in c ]\leq r+\alpha
$$
and
$$
\min_{c\in\mathcal{C}}\E_D [ y | R(x) = r, x \in c ] \ge r-\alpha
$$
\end{lemma}
\begin{proof}
The lemma follows from~\cref{def:MC}, and simply restates it as a constraint on the maximum and minimum expected risk among groups at each prediction level.
\end{proof}
\begin{proof}[Proof of \cref{thm:intersectionalmc}]
We use the same argument as \cite{fouldsIntersectionalDefinitionFairness2019} in proving this property for DF.
Define $Q$ as the Cartesian product of the protected attributes included in $\mathcal{A}$, but not $\mathcal{A}'$.
Then for any $(y,x) \sim D$,
\begin{align}
\max_{g\in \cal G}\E_D [ y | R(x) = r, x \in g]
=& \max_{g \in \cal G}\sum_{q\in Q} \E_D [ y | R(x) = r, x \in g \cap q ]P[x \in q | x \in g]\\
\leq& \max_{g \in \cal G}\sum_{q\in Q} \max_{q'\in Q}\E_D [ y | R(x) = r, x \in g \cap q' ]P[x \in q | x \in g]\\
=&\max_{g \in \cal G}\max_{q'\in Q}\E_D [ y | R(x) = r, x \in g \cap q' ]\\
=&\max_{c\in \mathcal{C}} \E_D [ y | R(x) = r, x \in c ]
.
\end{align}
Moving from (5) to (6) follows from substituting the maximum value of $\E_D [ y | R(x) = r, x]$ for observations in the intersection of subsets in $\mathcal{G}$ and $Q$ which is the upper limit of the expression in (5).
Moving from (6) to (7) follows from recognizing that the sum $P[x\in q|x \in g]$ for all subsets in $\mathcal{Q}$ is 1.
Finally, moving from (7) to (8) follows from recognizing that the intersections of subsets in $\mathcal{G}$ and $\mathcal{Q}$ that satisfy (7), must define a subset of $\mathcal{C}$.
Applying the same argument, we can show that
$$
\min_{g\in \cal G}\E_D[y|R(x)=r,x\in g] \ge\min_{c\in \mathcal{C}} \E_D [ y | R(x) = r, x \in c ]
.
$$
Substituting into \cref{lemma:express},
$$
\max_{g\in \cal G}\E_D [ y | R(x) = r, x \in g]\leq \alpha+r\\
$$
and
$$
\min_{g\in \cal G}\E_D [ y | R(x) = r, x \in g ] \ge r -\alpha
$$
or
$$
\card{\E_D [ y | R(x) = r, x \in g ]-r}\le{\alpha}
$$
for all $g\in \cal G$. Therefore $R(x)$ is $\alpha$-multicalibrated with respect to $\cal G$.
\end{proof}
As a concrete example, imagine we have the protected attributes $A = \set{ \text{race} \in \set{B,W}, \text{gender} \in \set{M,F}}$.
According to \cref{thm:intersectionalmc}, $\mathcal{C}$ would contain four sets: $\{(B,M),(B,F),(W,M),(W,F)\}$.
In contrast, there are eight possible sets in $\cal G$: $\set{ (B,M),(B,F),(W,M),(W,F),(B,*),(W,*),(*,M), (*,F)}$, where the wildcard indicates a match to either attribute.
As noted in \cref{s:df-efficient}, the efficiency property is useful because the number of possible sets in $\cal G$ grows at a large combinatorial rate, rate as additional attributes are added; meanwhile $\mathcal{C}$ grows at a slower, yet exponential, rate.
For an intuition for why this property holds, consider that the maximum calibration error of two subgroups is at least as large as the maximum expected error of those groups combined; e.g., the maximum calibration error in a higher order groups such as $(B,*)$ will be covered by the maximum calibration error in either $(B,M)$ or $(B,F)$.
\subsection{Additional Experiment Details}
Models were trained on a heterogenous computing cluster.
Each training instance was limited to a single core and 4 GB of RAM.
We conducted a full parameter sweep of the parameters specified in \cref{tbl:params}.
A single trial consisted of a method, a parameter setting from \cref{tbl:params}, and a random seed.
Over 100 random seeds, the data was shuffled and split 75\%/25\% into train/test sets.
Results in the manuscript are summarized over these test sets.
\paragraph{Code}
\label{s:code}
Code for the experiments is available here: \url{https://github.com/by1tTZ4IsQkAO80F/pmc}.
Code is licensed under GNU Public License v3.0.
\paragraph{Data}
We make use of data from the \href{https://physionet.org/content/mimic-iv-ed/1.0/}{MIMIC-IV-ED} repository, version 1.0, to train admission risk prediction models~\citep{johnsonalistairMIMICIVED2021}.
This resource contains more than 440,000 ED admissions from Beth Isreal Deaconness Medical Center between 2011 and 2019.
We preprocessed these data to construct an admission prediction task in which our model delivers a risk of admission estimate for each ED visitor after their first visit to triage, during which vitals are taken.
Additional historical data for the patient was also included (e.g., number of previous visits and admissions).
A list of features is given in \cref{tbl:features}.
\begin{table}
\caption{Features used in the hospital admission task.}
\label{tbl:features}
\begin{tabularx}{\textwidth}{XX}
\toprule
Description & Features \\
\midrule
Vitals &
temperature, heartrate, resprate, o2sat, systolic blood pressure, diastolic blood press,
\\
Triage Acuity &
Emergency Severity Index~\citep{tanabeReliabilityValidityScores2004}
\\
Check-in Data &
chief complaint, self-reported pain score
\\
Health Record Data &
no. previous visits, no. previous admissions
\\
Demographic Data &
ethnoracial group, gender, age, marital status, insurance, primary language
\\
\bottomrule
\end{tabularx}
\end{table}
\subsection{Additional Results}
\label{s:app:results}
\cref{tbl:params} lists a few parameters that may affect the performance of post-processing for both MC and PMC.
Of particular interest when comparing MC versus PMC post-processing is the parameter $\alpha$, which controls how stringent the calibration error must be across categories to terminate, and the group definition ($A$), which selects which features of the data will be used to asses and optimize fairness.
In comparing \cref{def:MC,def:PMC}, we note PMC's tolerance for error is more ``aggressive" for a given value of $\alpha$, since $\E_D [ y | R \in I, x \in S] \in [0,1]$.
Thus a natural question is whether MC can match the performance of PMC on different fairness measures simply by specifying a smaller $\alpha$.
We shed light on this question in three ways.
First, we quantify how often the use of each post-processing algorithm gives the best loss for each metric and trial in \cref{tbl:wins}.
Next, we look at the performance of MC and PMC postprocessing over values of $\alpha$ and group definitions in \cref{fig:auroc,fig:mcloss,fig:pmcloss}.
Finally, we empirically compare MC- and PMC-postprocessing by the number of steps required for each to reach their best performance in \cref{fig:updates,tbl:time}.
\cref{tbl:wins} quantifies the number of trials for which the baseline model and the two post-processing variants produce the best model according to a given metric, over all paramter configurations.
In pure head-to-head comparisons, we observe that PMC-postprocessing produces models with the lowest fairness loss according to all three metrics (DC loss, MC loss, PMC loss) the majority of the time.
This provides strong evidence that, over a large range of $\alpha$ values, PMC post-processing is beneficial compared to MC-postprocessing.
From \cref{fig:auroc}, it is clear that post-processing has a minimal effect on AUROC in all cases; note the differences dissapear if we round to two decimal places.
When post processing with RF, we do note a relationship between lower values of $\alpha$ and a very slight decrease in performance, particularly for MC-postprocessing.
\cref{fig:mcloss,fig:pmcloss} show performance between methods on MC loss and PMC loss, respectively.
In terms of MC loss, PMC-postprocessing tends to produce models with the lowest loss, at $\alpha$ values greater than 0.01.
Lower values of $\alpha$ do not help MC-postprocessing in most cases, suggesting that these smaller updates may be overfitting to the post-processing data.
In terms of PMC loss (\cref{fig:pmcloss}), we observe that performance by MC-postprocessing is highly sensitive to the value of $\alpha$.
For smaller values of $\alpha$, MC-postprocessing is able to achieve decent performance by these metrics, although in all cases, PMC-postprocessing generates a model with a better median loss value at some configuration of $\alpha$.
The ability of MC-postprocessing to perform well in terms of PMC and DC loss for certain values of $\alpha$ makes intuitive sense.
If $\alpha$ can be made small enough, the calibration error $\card{\E_D [ R | R \in I, x \in S] - \E_D [ y | R \in I, x \in S]}$ on all categories will be small compared to the outcome prevalence, $\E_D [ y | R \in I, x \in S]$.
However, to achieve this performance by MC-postprocessing may require a large number of unnecessary updates for high risk intervals, since the DC and PMC of multicalibrated models are limited by low-risk groups (\cref{thm:MCtoDC}).
Furthermore, the number of steps in MC-postprocessing (and PMC-postprocessing) scales as an inverse high-order polynomial of $\alpha$ (cf. Thm. 2~\citep{hebert-johnsonMulticalibrationCalibrationComputationallyIdentifiable2018}).
We assess how many steps/updates MC and PMC take for different values of $\alpha$ in \cref{fig:updates}, and summarize empirical measures of running time in \cref{tbl:time}.
On the figure, we annotate the point for which each post-processing algorithm achieves the lowest median value of PMC loss across trials.
\cref{fig:updates} validates that PMC-postprocessing is more efficient than MC-postprocessing at producing models with low PMC loss, on average requiring 4.0x fewer updates to achieve its lowest loss on test.
From \cref{tbl:time} we observe that PMC typically requires a larger number of updates to achieve its best performance on MC loss (about 2x wall clock time and number of updates), whereas MC-postprocessing requires a larger number of updates to achieves its best performance on PMC loss and DC loss, due to its dependence on very small values of $\alpha$.
We accompany these results with the caveat that they are based on performance on one real-world task, and wall clock time measurements are influenced by the heterogenous cluster environment; future work could focus on a larger empirical comparison.
\begin{table}
\centering
\footnotesize
\caption{Across 100 trials of dataset shuffles, we compare the post-processing configurations in terms of the number of times they achieve the best score for the metric shown on the left.
PMC post-processing (\cref{alg:PMC}) achieves the best fairness the highest percent of the time, according to DC loss (63\%), MC loss (70\%), and PMC loss (72\%), while MC-postprocessed models achieve the best AUROC in 88\% of cases.
}
\input{tbls/winning_configs.tex}
\label{tbl:wins}
\end{table}
\begin{figure}
\includegraphics[width=\textwidth]{figs/catpoint_AUROC_vs_alpha_row-ML_col-groups_hue-postprocessing_annot-none.pdf}
\caption{
AUROC test performance versus $\alpha$ across experiment settings.
Rows are different ML base models, and columns are different attributes used to define $\mathcal{C}$.
The color denotes the post-processing method.
}
\label{fig:auroc}
\end{figure}
\begin{figure}
\includegraphics[width=\textwidth]{figs/catpoint_MC-loss_vs_alpha_row-ML_col-groups_hue-postprocessing_annot-none.pdf}
\caption{
MC loss test performance versus $\alpha$ across experiment settings.
Rows are different ML base models, and columns are different attributes used to define $\mathcal{C}$.
The color denotes the post-processing method.
}
\label{fig:mcloss}
\end{figure}
\begin{figure}
\includegraphics[width=\textwidth]{figs/catpoint_PMC-loss_vs_alpha_row-ML_col-groups_hue-postprocessing_annot-none.pdf}
\caption{
PMC loss test performance versus $\alpha$ across experiment settings.
Rows are different ML base models, and columns are different attributes used to define $\mathcal{C}$.
The color denotes the post-processing method.
}
\label{fig:pmcloss}
\end{figure}
\begin{figure}
\includegraphics[width=\textwidth]{figs/catpoint_n-of-Updates_vs_alpha_row-ML_col-groups_hue-postprocessing_annot-PMC-loss.pdf}
\caption{
Number of post-processing updates by MC and PMC versus $\alpha$ across experiment settings.
Rows are different ML base models, and columns are different attributes used to define $\mathcal{C}$.
The color denotes the post-processing method.
Each result is annotated with the median PMC loss for that method and parameter combination.
}
\label{fig:updates}
\end{figure}
\begin{table}
\centering
\footnotesize
\caption{For MC- and PMC-postprocessing, we compare the median number of updates and median wall clock time (s) taken to train for the configuration ($\alpha$,$A$) that achieved the best performance on each metric.
}
\footnotesize
\input{tbls/best_cfg_time.tex}
\label{tbl:time}
\end{table}
\section{Introduction}
Today, machine learning (ML) models have an impact on outcome disparities across sectors (health, lending, criminal justice) due to their wide-spread use in decision-making.
When applied in clinical decision-making, ML models help care providers decide whom to prioritize to receive finite and time-sensitive resources among a population of potentially very ill patients.
These resources include hospital beds~\citep{barak-correnPredictionPatientDisposition2021,dinhOvercrowdingKillsHow2021}, organ transplants~\citep{schnellinger2021mitigating}, specialty treatment programs~\citep{henryTargetedRealtimeEarly2015,obermeyerDissectingRacialBias2019}, and, recently, ventilator and other breathing support tools to manage the COVID-19 pandemic~\citep{rivielloAssessmentCrisisStandards2022}.
In scenarios like these, decision makers typically rely on risk prediction models to be \emph{calibrated}.
Calibration measures the extent to which a model's risk scores, $R$, match the observed probability of the event, $y$.
Perfect calibration implies that $P(y|R=r) = r$, for all values of $r$.
Calibration allows the risk scores to be used to rank patients in order of priority and informs care providers about the urgency of treatment.
However, models that are not equally calibrated among subgroups defined by different sensitive attributes (race, ethnicity, gender, income, etc.) may lead to systematic denial of resources to marginalized groups (e.g.~\citep{obermeyerDissectingRacialBias2019,ashana2021equitably,roberts_fatal_2011,zelnick2021association,ku2021racial}).
Just this scenario was observed by~\citet{obermeyerDissectingRacialBias2019} analyzed a large health system algorithm used to enroll high-risk patients into care management programs and showed that, at a given risk score, Black patients exhibited significantly poorer health than white patients.
To address equity in calibration, \citet{hebert-johnsonMulticalibrationCalibrationComputationallyIdentifiable2018} proposed a fairness measure called \textit{multicalibration} (MC), which asks that calibration be satisifed simultaneously over many flexibly-defined subgroups.
Remarkably, MC can be satisfied efficiently by post-processing risk scores without negatively impacting the generalization error of a model, unlike other fairness concepts like demographic parity~\citep{fouldsAreParityBasedNotions2020} and equalized odds~\citep{hardtEqualityOpportunitySupervised2016a}.
This has motivated the use of MC in practical settings (e.g.~\citet{bardaAddressingBiasPrediction2021a})
and has spurred several extensions~\citep{kimMultiaccuracyBlackboxPostprocessing2019,jungMomentMulticalibrationUncertainty2021,guptaOnlineMultivalidLearning2021,gopalanLowDegreeMulticalibration2022}.
If we bin our risk predictions, the MC criteria specifies that, for every group within each bin, the absolute difference between the mean observed outcome and the mean of the predictions should be small.
As~\citet{barocasFairnessMachineLearning2019} note, equity in calibration embeds the fairness notion called \emph{sufficiency}, which states: for a given risk prediction, the expected outcome should be independent of group membership.
Starting from this notion, we can assess the conditions under which MC satisfies sufficiency.
In this work, we derive a fairness criteria directly from sufficiency dubbed \emph{differential calibration} for its relation to differential fairness~\citep{foulds_intersectional_2019}.
We show that satisfying differential calibration can ensure that a model is equally ``trustworthy" among groups in the data.
By equally ``trustworthy'', we mean that a decision maker cannot reasonably come to distrust the model's risk predictions for specific groups, which may help prevent differences in decision-making between demographic groups, given the same risk prediction.
By relating sufficiency to MC, we describe a shortcoming of MC that can occur when the outcome probabilities are strongly tied to group membership.
Under this condition, the amount of calibration error \emph{relative to the expected outcome} can be unequal between groups.
This inequality hampers the ability of MC to (approximately) guarantee sufficiency, and thus guarantee equity in trustworthiness for the decision maker.
We propose a simple variant of MC called \textit{proportional multicalibration} (PMC) that ensures that the proportion of calibration error within each bin and group is small.
We prove that PMC bounds both multicalibration and differential calibration.
We show that PMC can be satisfied with an efficinet post-processing method, similarly to MC.
\looseness=-1
\subsection{Our Contributions}
In this manuscript, we formally analyze the connection of MC to the fairness notion of sufficiency.
To do so, we introduce differential calibration (DC), a sufficiency measure that constrains ratios of population risk between pairs of groups within prediction bins.
We describe how DC, like sufficiency, provides a sense of equal trustworthiness from the point of view of the decision maker.
With this definition, we prove the following.
First, models that are ($\alpha$,$\lambda$)-multicalibrated satisfy $(log \frac{r_{min}+\alpha}{r_{min}-\alpha}, \lambda)$-DC, where $r_{min}$ is the minimum expected risk prediction among categories defined by subgroups and prediction intervals.
We illustrate the meaning of this bound, which is that the proportion of calibration error in multicalibrated models may scale inversely with the outcome probability.
Based on these observations, we propose an alternate definition of MC, PMC, that controls the percentage error by group and risk strata (\cref{def:PMC}).
We show that models satisfying $(\alpha,\lambda)$-PMC are $(\frac{\alpha}{1-\alpha},\lambda)$-multicalibrated and $(\log \frac{1+\alpha}{1-\alpha})$-differentially calibrated.
Proportionally multicalibrated models thereby obtain robust fairness guarantees that are independent of population risk categories.
Furthermore, we define an efficient algorithm for learning predictors satisfying $\alpha$-PMC.
Finally, we investigate the application of these methods to predicting patient admissions in the emergency department, a real-world resource allocation task, and show that post-processing for PMC results in models that are accurate, multicalibrated, and differentially calibrated.
\section{Reconciling Multicalibration and Sufficiency}
\label{s:methods}
\subsection{Preliminaries}
\looseness=-1
We consider the task of training a risk prediction model for a population of individuals with outcomes, $y \in \set{0,1}$, and features, $x \in \mathcal{X}$.
Let $D$ be the joint distribution from which individual samples $(y, x)$ are drawn.
We assume the outcomes $y$ are random samples from underlying independent Bernoulli distributions, denoted as $p^*(x) \in [0,1]$.
Given an individual's attributes $x = (x_1,\;\dots,\;x_d)$, it will be useful to refer to subsets we wish to protect, e.g. demographic identifiers.
To do so, we define $\mathcal{A} = \set{A_1,\;\dots,\;A_p}$, $p \leq d$, such that $A_1 = \{x_{1i},\;\dots,\;x_{1k}\}$ is a finite set of values taken by attribute $x_1$.
Individuals can be further grouped into \emph{collections of subsets}, $\mathcal{C} \subseteq \text{2}^{\mathcal{X}}$, such that $S \in \mathcal{C}$ is the subset of individuals belonging to $S$, and $x \in S$ indicates that individual $x$ belongs to group $S$.
We denote our risk prediction model as $R(x): \mathcal{X}$ $\rightarrow [0,1]$.
In order to consider calibration in practice, the risk predictions are typically discretized and considered within intervals.
The coarseness of this interval is parameterized by a partitioning parameter, $\lambda \in (0, 1]$.
The \emph{$\lambda$-discretization} of $[0,1]$ is denoted by a set of intervals,
$\Lambda_{\lambda} = \set{ \set{I_j}_{j=0}^{1/\lambda -1}}$, where
$I_j = [ j\lambda, (j+1)\lambda ) $.
For brevity, most proofs in the following sections are given in~\cref{s:proof}.
\subsection{Multicalibration}
\looseness=-1
MC~\citep{hebert-johnsonCalibrationComputationallyIdentifiableMasses2018} guarantees that the calibration error for any group from a collection of subsets, $\mathcal{C}$ will not exceed a user-defined threshold, over the range of risk scores.
In order to work with bins of predictions, we will mostly concern ourselves with the discretized version of MC, defined below.
The non-discretized versions are given in~\cref{s:App:def}.
\begin{definition}[$(\alpha,\lambda)$-multicalibration]
\label{def:alMC}
Let $\mathcal{C} \subseteq \text{2}^{\mathcal{X}}$ be a collection of subsets of $\mathcal{X}$.
For any $\alpha, \lambda > 0$,
a predictor $R$ is \emph{$(\alpha,\lambda)$-multicalibrated} on $\mathcal{C}$
if,
for all $I \in \Lambda_{\lambda}$
and $S \in \mathcal{C}$ where $P_D(R \in I |x \in S) \geq \alpha \lambda $,
$$ \card{ \E_D [ y | R \in I, x \in S] - \E_D [ R | R \in I, x \in S]} \le \alpha .$$
\end{definition}
MC is one of few approaches to achieving fairness that does not require a significant trade-off to be made between a model's generalization error and the improvement in fairness it provides~\citep{hebert-johnsonCalibrationComputationallyIdentifiableMasses2018}.
As \cite{hebert-johnsonCalibrationComputationallyIdentifiableMasses2018} show, this is because achieving multicalibration is not at odds with achieving accuracy in expectation for the population as a whole.
This separates calibration fairness from other fairness constraints like demographic parity and equalized odds~\citep{hardtEqualityOpportunitySupervised2016a}, both of which may denigrate the performance of the model on specific groups~\citep{chouldechovaFairPredictionDisparate2017,pleissFairnessCalibration2017}.
In clinical settings, such trade-offs may be difficult or impossible to justify.
In addition to its alignment with accuracy in expectation, \citet{hebert-johnsonCalibrationComputationallyIdentifiableMasses2018} propose an efficient post-processing algorithm for MC similar on boosting.
We discuss additional extensions to MC in \cref{s:App:related}.
\subsection{Sufficiency and Differential Calibration}
MC provides a sense of fairness by approximating \emph{calibration by group}, which is perfectly satisfied when $P_D(y|R=r,x \in S)=r$ for all $S \in C$.
Calibration by group is closely related to the \emph{sufficiency} fairness criterion~\citep{barocasFairnessMachineLearning2019}.
Sufficiency is the condition where the outcome probability is independent from $\mathcal{C}$ conditioned on the risk score.
In the binary group setting ($\mathcal{C} = \{S_i, S_j\}$), sufficiency can be expressed as
$
P_D(y|R, x \in S_i) = P_D(y|R, x \in S_j)
$, or
\begin{equation}\label{eq:sufficiency}
\frac{ P_D(y|R, x \in S_i) }{ P_D(y|R, x \in S_j) } = 1.
\end{equation}
Unlike calibration by group, sufficiency does not stipulate that the risk scores be calibrated, yet from a fairness perspective, sufficiency and calibration-by-group are equivalent~\citep{barocasFairnessMachineLearning2019}.
Consider that one can easily transform a model satisfying sufficiency into one that is calibrated-by-group with a single function $f(R) \rightarrow [0,1]$, for example with Platt scaling~\citep{barocasFairnessMachineLearning2019}.
In both cases, the sense of \emph{fairness} stems from the desire for the risk scores, $R$ to capture everything about group membership that is relevant to predicting the outcome, $y$.
Under sufficiency, the risk score is equally informative of the outcome, regardless of group membership.
In this sense, a model satisfying sufficiency provides \emph{equally trustworthy} risk predictions to a decision maker, regardless of the groups to which an individual belongs.
Below, we define an approximate measure of sufficiency that constrains pairwise differentials between groups, and accomodates binned predictions:
\begin{definition}[Differential calibration]\label{def:DC}
Let $\mathcal{C} \subseteq \text{2}^{\mathcal{X}}$ be a collection of subsets of $\mathcal{X}$.
A model $R(x)$ is ($\varepsilon$,$\lambda$)-differentially calibrated with respect to $\mathcal{C}$ if, across prediction intervals $I \in \Lambda_{\lambda}$, for all pairs $(S_i,S_j) \in \mathcal{C} \times \mathcal{C}$ for which $P_D(S_i), P_D(S_j) >0$,
\begin{equation}\label{eq:DC}
e^{-\varepsilon} \leq
\frac{\E_D [ y | R \in I, x \in S_i]}
{\E_D [ y | R \in I, x \in S_j ]}
\leq e^{\varepsilon}
\end{equation}
\end{definition}
By inspection we see that $\epsilon$ in $(\epsilon,\lambda)$-DC measures the extent to which $R$ satisifies sufficiency.
That its, when $P(y|R \in I, x \in S_i) \approx P(y|R \in I, x \in S_j)$ for all pairs, $\varepsilon \approx 0$.
$(\varepsilon,\lambda)$-DC says that, within any bin of risk scores, the outcome $y$ is at most $e^{\varepsilon}$ times more likely among one group than another, and a minimum of $e^{-\varepsilon}$ less likely.
\cref{def:DC} fits into the general definition of a \emph{differential fairness} measure proposed by~\citet{fouldsIntersectionalDefinitionFairness2019}, although previously it was used to define demographic parity criteria~\citep{fouldsAreParityBasedNotions2020}.
We describe the relation in more detail in~\cref{s:App:DF}, including \cref{eq:DC}'s connection to differential privacy~\cite{dwork2009differential} and pufferfish privacy~\cite{kiferPufferfishFrameworkMathematical2014}.
\subsection{The differential calibration of multicalibrated models is limited by low-risk groups}
At a basic level, the form of MC and sufficiency differ: MC constrainins absolute differences between groups across prediction bins, whereas sufficiency constrains pairwise differentials between groups.
To reconcile MC and DC/sufficiency more formally, we pose the following question: if a model satisfies $\alpha$-MC, what, if anything does this imply about the $\varepsilon$-DC of the model?
(In \cref{s:App:thm},\cref{thm:DCtoMC}, we answer the inverse question).
We now show that multicalibrated models have a bounded DC, but that this bound is limited by small values of $R$.
\begin{theorem}
\label{thm:MCtoDC}
\Copy{thm:MCtoDC}{
Let $R(x)$ be a model satisfying ($\alpha$,$\lambda$)-MC on a collection of subsets $\mathcal{C} \in 2^{\mathcal{X}}$.
Let $r_{min} = \min_{(S, I) \in \mathcal{C} \times \Lambda_\lambda}{\E_D [ R | R \in I, x \in S]}$ be the minimum expected risk prediction among categories $(S, I) \in \mathcal{C} \times \Lambda_\lambda$.
Then R(x) is $(\log \frac{r_{min}+\alpha}{r_{min}-\alpha}, \lambda)$-differentially calibrated.
}
\end{theorem}
\begin{proof}
Let $r = \E_D [ R | R \in I, x \in S]$ and $p^* = \E_D [ y | R \in I, x \in S]$.
$(\alpha, \lambda)$-MC guarantees that $r - \alpha \leq p^* \leq r + \alpha$ for all groups $S \in \mathcal{C}$ and prediction intervals $r \in \Lambda_\lambda[0, 1]$.
Plugging these lower and upper bounds into \cref{eq:DC} yields
$
e^{\varepsilon} \geq {\frac{r+\alpha}{r-\alpha} }
$.
The maximum of this ratio, for a fixed $\alpha$, occurs at the smallest value of $r$; therefore
$
\varepsilon \geq \log \frac{r_{min} + \alpha}{r_{min} - \alpha} .
$
\end{proof}
\cref{thm:MCtoDC} illustrates the important point that, \emph{in terms of percentage error}, $MC$ does not provide equal protection to groups with different risk profiles.
Imagine a model satisfying (0.05,0.1)-MC for groups $S \in \mathcal{C}$.
Consider individuals receiving model predictions in the interval $(0.9,1]$.
MC guarantees that, for any category $\set{x : x \in S, R(x) \in I =(0.9,1]}$, the expected outcome prevalence ($\E_D[y|x \in S, R \in I]$) of at least $0.9 - \alpha = 0.85$.
This bounds the percent error among groups in the $(0.9, 1]$ prediction interval to 6\%.
In contrast, consider individuals for whom $R(x) \in (0.3-0.4]$; each group may have a true outcome prevalence as low as 0.25, which is an error of 20\% - about 3.4x higher than the percent error in the higher-risk group.
\section{Proportional Multicalibration}
We are motivated to define a measure that is efficiently learnable like MC (\cref{def:alMC}) but better aligned with the fundamental fairness notion of sufficiency, like DC (\cref{def:DC}).
To do so, we define PMC, a variant of MC that constrains the proportional calibration error of a model among subgroups and risk strata.
In this section, we show that bounding a model's PMC is enough to meaningfully bounds its DC and MC.
Furthermore, we provide an efficient algorithm for satisfying PMC based on a simple extension of MC/Multiaccuracy boosting~\citep{kimMultiaccuracyBlackboxPostprocessing2019}.
\begin{definition}[Proportional Multicalibration]\label{def:PMC}
A model $R(x)$ is $(\alpha,\lambda)$-proportionally multicalibrated with respect to a collection of subsets $\mathcal{C}$ if,
for all $S \in \mathcal{C}$ and $I \in \Lambda_\lambda$ satisfying $P_D(R(x) \in I | x \in S) \geq \alpha \lambda$,
\begin{equation}\label{eq:lPMC}
\frac{ \card{ \E_D [ y | R \in I, x \in S] - \E_D [ R | R \in I, x \in S] } }
{ \E_D [ y | R \in I, x \in S] }
\le \alpha.
\end{equation}
\end{definition}
Note that, in practice, we must ensure $\E_D [ y | R \in I, x \in S] \neq 0$ for \cref{def:PMC} to be defined.
We handle this by introducing a parameter $\rho > 0$ constraining the lowest expected outcome among categories $(S,I)$.
In the remainder of this section, we detail how PMC relates to suffiency/DC and MC.
We provide bounds on the values of $MC$ and $DC$ given a proportionally multicalibrated model, and we illustrate the relationship between these three metrics in \cref{fig:params}.
\paragraph{Comparison to Differential Calibration}
Rather than constraining the differentials of prediction- and group- specific outcomes among all pairs of subgroups in $\mathcal{C} \times \mathcal{C}$ as in DC (\cref{def:DC}), PMC constrains the relative error of each group in $\mathcal{C}$.
In practical terms, this makes it more efficient to calculate PMC by a factor of $O(\card{\mathcal{C}})$ steps compared to DC.
In addition, PMC does not require additional assumptions about the overall calibration of a model in order to imply guarantees of MC, since PMC directly constrains calibration rather than constraining sufficiency alone.
\begin{theorem}
\label{thm:PMCtoDC}
\Copy{thm:PMCtoDC}
{
Let R(x) be a model satisfying $(\alpha,\lambda)$-PMC on a collection $\mathcal{C}$.
Then $R(x)$ is $(\log \frac{1+\alpha}{1-\alpha}, \lambda )$-differentially calibrated.
}
\end{theorem}
\begin{proof}
Let $r = \E_D [ R | R \in I, x \in S]$ and $p^* = \E_D [ y | R \in I, x \in S]$.
If $R(x)$ satisfies $\alpha$-PMC (\cref{def:PMC}), then
$ r/(1 + \alpha) \leq p^* \leq r/(1 - \alpha)$.
Solving for the upper bound on $\varepsilon$-DC, we immediately have
$\varepsilon \leq \log \frac{r(1+\alpha)}{r(1-\alpha)} \leq \log \frac{1+\alpha}{1-\alpha} $.
\end{proof}
\cref{thm:PMCtoDC} demonstrates that $\alpha$-proportionally multicalibrated models satisfy a straightforward notion of differential fairness that depends monotonically only on $\alpha$.
The relationship between PMC and DC is contrasted with the relationship of MC and DC in \cref{fig:params}, left panel.
The figure illustrates how MC's sensitivity to small risk categories limits its DC.
\paragraph{Comparison to Multicalibration}
Rather than constraining the absolute difference between risk predictions and the outcome as in MC, PMC requires that the calibration error be a small fraction of the expected risk in each category $(S,I)$.
In this sense, it provides a stronger protection than MC by requiring calibration error to be a small fraction regardless of the risk group.
In many contexts, we would argue that this is also more aligned with the notion of fairness in risk prediction contexts.
Under MC, the underlying prevalence of an outcome within a group affects the fairness protection that is received (i.e., the percentage error that \cref{def:MC} allows).
Because underlying prevalences of many clinically relevant outcomes vary significantly among subpopulations, multicalibrated models may systematically permit higher percentage error to specific risk groups.
The difference in relative calibration error among populations with different risk profiles also translates in weaker sufficiency guarantees, as demonstrated in~\cref{thm:MCtoDC}.
In contrast, PMC provides a fairness guarantee that is independent of subpopulation risks.
In the following theorem, we show that MC is constrained when a model satisfies PMC.
\begin{theorem}
\label{thm:PMCtoMC}
\Copy{thm:PMCtoMC}{
Let $R(x)$ be a model satisfying \emph{$\alpha$-PMC} on a collection $\mathcal{C}$.
Then $R(x)$ is ($\frac{\alpha}{1-\alpha}$)-multicalibrated on $\mathcal{C}$.
}
\end{theorem}
\begin{proof}
To distinguish the parameters, let $R(x)$ be a model satisfying $\delta$-PMC.
Let $r = \E_D [ R | R \in I, x \in S]$ and $p^* = \E_D [ y | R \in I, x \in S]$.
Then $ r/(1 + \delta) \leq p^* \leq r/(1 - \delta) $.
We solve for the upper bound on $\alpha$-MC from \cref{def:MC} for the case when $p^* > r$.
This yields
\begin{align*}
\alpha &\leq p^* - r \\
&\leq \frac{r}{1-\delta} - r \\
&= r\frac{\delta}{1-\delta} \\
&\leq \frac{\delta}{1-\delta}
.
\end{align*}
\end{proof}
The right panel of \cref{fig:params} illustrates this relation in comparison to the DC-MC relationship described in \cref{s:App:thm}, \cref{thm:DCtoMC}.
At small values of $\epsilon$ and $\alpha$ and when the model is perfectly calibrated overall, $\alpha$-PMC and $\epsilon$-DC behave similarly.
However, given $\delta>0$, $\epsilon$-differentially calibrated models suffer from higher MC error than proportionally calibrated models when $\alpha$-PMC $< 0.3$.
The right graph also illustrates the feasible range of $\alpha$ for $\alpha$-PMC is $0 < \alpha < 0.5$, past which it does not provide meaningful $\alpha$-MC.
The steeper relation between $\alpha$-PMC and MC may have advantages or disadvantages, depending on context.
It suggests that, by optimizing for $\alpha$-PMC, small improvements to this measure can result in relatively large improvements to MC; conversely, $\epsilon$-DC models that are well calibrated may satisfy a lower value of $\alpha$-MC over a larger range of $\epsilon$.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{figs/parameter_comparison.pdf}
\caption{
A comparison of $\varepsilon$-DC, $\alpha$-MC, and $\alpha$-PMC in terms of their parameters $\alpha$ and $\epsilon$.
In both panes, the x value is a given value of one metric for a model, and the y axis is the implied value of the other metric, according to \cref{thm:DCtoMC}-\cref{thm:PMCtoMC}.
The left filled area denotes the dependence of the privacy/DC of $\alpha$-multicalibrated models on the minimum risk interval, $r_{min} \in [0.01, 1.0]$.
The right filled area denotes the dependence of the MC of $\epsilon$-differentially calibrated models on their overall calibration, $\delta \in [0.0, 0.5]$.
$\alpha$-PMC does not have these sensitivities.
}
\label{fig:params}
\end{figure}
\subsection{Learning proportionally multicalibrated predictors}
So far we have demonstrated that models satisfying PMC exhibit desirable guarantees relative to two previously defined measures of fair calibration, but have not considered whether PMC is easy to learn.
Here, we answer in the affirmative by proposing \cref{alg:PMC} to satisfy PMC and proving that it learns an ($\alpha$,$\lambda$)-PMC model in a polynomial number of steps.
\begin{theorem}
\label{thm:alg}
\Copy{thm:alg}{
Define $\alpha, \lambda, \gamma, \rho > 0$.
Let $\mathcal{C} \subseteq 2^{\mathcal{X}}$ be a collection of subsets of $\mathcal{X}$ such that, for all $S \in \mathcal{C}$, $P_D(S) > \gamma$.
Let $R(x)$ be a risk prediction model to be post-processed.
For all
$(S,I) \in \mathcal{C} \times \Lambda_{\lambda}$, let
$E[y|R\in I, x \in S] > \rho$.
There exists an algorithm that satisfies $(\alpha, \lambda)$-PMC with respect to $\mathcal{C}$
in $O(\frac{|C|}{\alpha^3\lambda^2\rho^2\gamma})$ steps.
}
\end{theorem}
We analyze \cref{alg:PMC} and show it satisfies \cref{thm:alg} in \cref{s:proof}.
\cref{alg:PMC} directly extends MCBoost\citep{pfistererMcboostMultiCalibrationBoosting2021}, but differs in that it does not terminate until $R(x)$ is within $\alpha \bar{y}$ for all categories, as opposed to simply within $\alpha$.
This more stringent threshold requires an additional $O(\frac{1}{\rho^2})$ steps, where $\rho>0$ is a lower bound on the expected outcome within a category $(S,I)$.
The parameter $\rho$ also serves to smooth empirical estimates of \cref{eq:lPMC} in our experiments.
\algblockdefx{MRepeat}{EndRepeat}{\textbf{repeat}}{}
\algnotext{EndRepeat}
\begin{algorithm}[t]
\caption{Proportional Multicalibration Post-processing}
\label{alg:PMC}
\begin{algorithmic}[1]
\footnotesize
\Require{ Predictor $R(x)$ \\
$\mathcal{C} \in 2^{\mathcal{X}}$ such that for all $S \in \mathcal{C}, P_D(S) \geq \gamma$ \\
$\alpha, \lambda, \gamma, \rho > 0$ \\
$\mathcal{D} = \set{(y,x)_i}_{i=0}^{N} \sim D$
}
\Statex
\Function{PMC}{$R$, $\mathcal{C}$, $\mathcal{D}$, $\alpha$, $\lambda$, $\gamma$, $\rho$}
\MRepeat \hspace{1em}
\Let{$\{(y,x)\}$}{sample $\mathcal{D}$}
\For{$S \in \mathcal{C}, I \in \Lambda_\lambda$
such that $P_D(R \in I , x \in S) \geq \alpha \lambda \gamma$ }
\Let{$S_r$}{$S \cap \set{x: R(x) \in I}$}
\Let{$\bar{r}$}{$\frac{1}{|S_r|}\sum_{x \in S_r}{R(x)}$} \Comment{average group prediction }
\Let{$\bar{y}$}{$\frac{1}{|S_r|}\sum_{x \in S_r}{y(x)}$} \Comment{average subgroup risk}
\If{$\bar{y} \leq \rho$}
\State continue
\EndIf
\Let{$\Delta r$}{$\bar{y} - \bar{r}$}
\If{$\card{\Delta r} \geq \alpha \bar{y}$}
\Let{$R(x)$}{$R(x) + \Delta r$ for all $x \in S_r$}
\Let{$R(x)$}{squash($R(x)$, $[0,1]$)} \Comment{squash updates to $[0,1]$}
\EndIf
\EndFor
\If{No Updates to R(x)}
\State break
\EndIf
\EndRepeat
\State \Return{$R$}
\EndFunction
\end{algorithmic}
\end{algorithm}
\section{Experiments}
\label{s:exp}
In our first set of experiments (\cref{s:exp}), we study MC and PMC in simulated population data to understand and validate the analysis in previous sections.
In the second section, we compare the performance of varied model treatments on a real world hospital admission task, using an implementation of \cref{alg:PMC}.
We make use of empirical versions of our fairness definitions which we refer to as \textit{MC loss} (\cref{def:mcloss}), \textit{PMC loss} (\cref{def:pmcloss}), and \textit{DC loss} (\cref{def:dcloss}), defined in \cref{s:App:def}.
\paragraph{Simulation study}
\label{s:exp:sim}
We simulate data from $\alpha$-multicalibrated models.
For simplicity, we specify a data structure with a one-to-one correspondence between subset and model estimated risk, such that for all $x$ in subset $S$, $R(x)=R(x|x\in S)=R(S)$.
Therefore all information for predicting the outcome based on the features in $x$ is contained in the attributes $\mathcal{A}$ that define subgroup $S$.
Prevalence is specified as
$p_i^*=P_D(y|x \in S_i)=0.2+0.01(i-1)$
and $i=1,\cdots,N_s$, where $N_s$ is the number of subsets $S$, defined by $\mathcal{A}$ and indexed by $i$ with increasing $p^*$.
For each group,
$R_i=R(S_i)=R(x|x \in S_i)=p_i^*-\Delta_i.$
We randomly select $\Delta_i$ for one group to be $\pm\alpha$ and for the remaining groups, $\Delta_i= \pm\delta$, where $\delta\sim \textrm{Uniform}(\min=0, \max=\alpha)$.
In all cases, the sign of $\Delta_i$ is determined by a random draw from a Bernoulli distribution. For these simulations we set $N_S=61$ and $\alpha=0.1$, such that $p^*_i\in[0.2,0.8]$ and $R_i\in[0.1,0.9]$. We generate $N_{sim}=1000$ simulated datasets, with $n=1000$ observations per group, and for each $S_i$, we calculate the ratio of the absolute mean error to $p^*_i$, i.e. the PMC loss function for this data generating mechanism.
We also simulate three specific scenarios where:
\begin{enumerate*}[label=\arabic*)]
\item $\card{\Delta_i}$ is equivalent for all groups (Fixed);
\item $\card{\Delta_i}$ increases with increasing $p_i^*$; and
\item $\card{\Delta_i}$ decreases with increasing $p_i^*$
\end{enumerate*}, with $\alpha=0.1$ in each case.
These scenarios compare when $\alpha$ is determined by all groups, the group with the lowest outcome prevalence, and the group with the highest outcome prevalence, respectively.
\paragraph{Hospital admission}
\label{s:exp:mimic}
Next, we test PMC alongside other methods in application to prediction of inpatient hospital admission for patients visiting the emergency department (ED).
The burden of overcrowding and long wait times in EDs is significantly higher among non-white, non-Hispanic patients and socio-economically marginalized patients~\citep{jamesAssociationRaceEthnicity2005,mcdonaldExaminingAssociationCommunityLevel2020a}.
Recent work has demonstrated risk prediction models that can expedite patient visits by predicting patient admission at an early stage of a visit with a high degree of certainty (AUC $\geq$ 0.9 across three large care centers)~\citep{barak-correnProgressivePredictionHospitalisation2017,barak-correnEarlyPredictionModel2017,barak-correnPredictionHealthcareSettings2021,barak-correnPredictionPatientDisposition2021}.
Our goal is to ensure no group of patients will be over- or under-prioritized over another by these models, which could exacerbate the treatment and outcome disparities that currently exist.
We construct a prediction task similar to previous studies but using a new data resource: the \href{https://physionet.org/content/mimic-iv-ed/1.0/}{MIMIC-IV-ED} repository~\citep{johnsonalistairMIMICIVED2021}.
The overall intersectional demographic statistics for these data are given in~\cref{tbl:mimic}.
In \cref{tbl:mimic} we observe stark differences in admission rates by demographic group and gender, suggesting that the use of a proportional measure of calibration could be appropriate for this task.
We trained and evaluated logistic regression (LR) and random forest (RF) models of patient admission, with and without post-processing for MC~\citep{pfistererMcboostMultiCalibrationBoosting2021} or PMC.
We tested a number of parameter settings given in~\cref{tbl:params}, running 100 trials with different shuffles of the data.
Comparisons are reported on a test set of 20\% of the data for each trial.
Additional experiment details are available in~\cref{s:app:results} and
code for the experiments is available here: \url{https://github.com/cavalab/proportional-multicalibration}.
The PMC-postprocessing method is available as a package as well: \url{https://github.com/cavalab/pmcboost}.
\section{Results}
\cref{fig:sim} shows the PMC loss of $\alpha$-multicalibrated models under the scenarios described in \cref{s:exp:sim}.
Proportional $\alpha$-MC constrains the ratio of the absolute mean error (AME) to the outcome prevalence, for groups defined by a risk interval $(R(x) \in I)$ and subset within a collection of subsets ($x \in S, S\in \mathcal{C})$. Without the proportionality factor $\card{ \E_D [ y | R \in I, x \in S] }^{-1}$ , $\alpha$-multicalibrated models allow a dependence between the group prevalence and the error or privacy loss permitted that is unfair for groups with lower outcome prevalence.
Results on the hospital admission prediction task are summarized in \cref{fig:mimic_results} and \cref{tbl:wins}.
PMC post-processing has a negligible effect on predictive performance ($<$0.1\% $\Delta$ AUROC, LR and RF) while reducing DC loss by 27\% for LR and RF models, and reducing PMC loss by 40\% and 79\%, respectively.
In the case of RF models, PMC post-processing reduces MC loss by 23\%, a significantly larger improvement than MC post-processing itself (19\%, $p$=9e-26).
\begin{table}
\centering
\scriptsize
\caption{
Admission prevalence (Admissions/Total (\%)) among patients in the MIMIC-IV-ED data repository, stratified by the intersection of ethnoracial group and gender.
}
\label{tbl:mimic}
\input{tbls/case_control_intersections.tex}
\end{table}
Due to normalization by outcome rates, the optimal value of $\alpha$ for PMC is likely to differ from the best value for MC (their relationship is shown in \cref{fig:params}).
For both methods, setting $\alpha$ too small may result in over-fitting.
To account for this, we quantified the number of trials for which a given method produced the best model according to a given metric, over all parameter configurations in \cref{tbl:params}.
PMC post-processing (\cref{alg:PMC}) achieves the best fairness the highest percent of the time, according to DC loss (63\%), MC loss (70\%), and PMC loss (72\%), while MC-postprocessed models achieve the best AUROC in 88\% of cases.
This provides strong evidence that, over a large range of $\alpha$ values, PMC post-processing is beneficial compared to MC-postprocessing.
We characterize the sensitivity of PMC and MC to $\alpha$ and provide a more detailed breakdown of these results in \cref{s:app:results}.
\begin{figure}
\begin{minipage}{.49\textwidth}
\includegraphics[width=\textwidth]{figs/sim_plot2.pdf}
\end{minipage}
\hspace{.01\textwidth}
\begin{minipage}{.49\textwidth}
\caption{
The relationship between MC, PMC, and outcome prevalence as illustrated via a simulation study in which the rates of the outcome are associated with group membership.
Gray points denote the PMC loss of a (0.1,0.1)-MC model on 1000 simulated datasets, and colored lines denote three specific scenarios in which each group's calibration error ($|\Delta|$) follows specific rules.
PMC loss is higher among groups with lower positivity rates in most scenarios unless the groupwise calibration error increases with positivity rate.
}
\label{fig:sim}
\end{minipage}
\end{figure}
\begin{table}[t]
\begin{minipage}[b]{.5\textwidth}
\centering
\footnotesize
\caption{Parameters for the hospital admission prediction experiment.}
\label{tbl:params}
\input{tbls/pmc_params.tex}
\end{minipage}
\hspace{.01\textwidth}
\begin{minipage}[b]{.44\textwidth}
\centering
\footnotesize
\caption{
The number of times each postprocessing method achieved the best score among all methods, out of 100 trials.
}
\input{tbls/winning_configs.tex}
\label{tbl:wins}
\end{minipage}
\end{table}
\begin{figure}
\includegraphics[width=\textwidth]{figs/box_AUROC_MC_PMC_DC.pdf}
\caption{
A comparison of LR and RF models, with and without MC and PMC post-processing, on the hospital admission task.
From left to right, trained models are compared in terms of test set AUROC, MC loss, PMC loss, and DC loss.
Points represent the median performance over 100 shuffled train/test splits with bootstrapped 99\% confidence intervals.
We test for significant differences between post-processing methods using two-sided Wilcoxon rank-sum tests with Bonferroni correction.
ns: $p <=$ 1;
**: 1e-03 $< p <=$ 1e-02;
***: 1e-04 $< p <=$ 1e-03;
****: $p <=$ 1e-04.
}
\label{fig:mimic_results}
\end{figure}
\section{Discussion and Conclusion}
\looseness=-1
In this paper we have analyzed multicalibration through the lens of suffiency and differential calibration to reveal the sensitivity of this metric to correlations between outcome rates and group membership.
We have proposed a measure, PMC, that alleviates this sensitivity and attempts to capture the ``best of both worlds" of MC and DC.
PMC provides equivalent percentage calibration protections to groups regardless of their risk profiles, and in so doing, bounds a model's differential calibration.
We provide an efficient algorithm for learning PMC predictors by postprocessing a given risk prediction model.
On a real-world and clinically relevant task (admission prediction), we have shown that post-processing LR and RF models with PMC leads to better performance across all three fairness metrics, with little to no impact on predictive performance.
\looseness=-1
Our preliminary analysis suggests PMC can be a valuable metric for training fair algorithms in resource allocation contexts.
Future work could extend this analysis on both the theoretical and practical side.
On the theoretical side, the generalization properties of the PMC measure should be established and its sample complexity quantified, as \citet{roseMachineLearningPrediction2018} did with MC.
Additional extensions of PMC could establish a bound on the accuracy of PMC-postprocessed models in a similar vein to work by \citet{kimMultiaccuracyBlackboxPostprocessing2019} and \citet{hebert-johnsonMulticalibrationCalibrationComputationallyIdentifiable}.
On the empirical side, future works should benchmark PMC on a larger set of real-world problems, and explore use cases in more depth.
| 705b9bb2fabb46b547024569cee6bfa6942b1a10 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The fundamental goal of machine learning algorithms is to identify the conditional distribution given any input and its label. In the training phase, it’s conventional to assume that the underlying classifier or function belongs to a certain class of functions. Therefore presuming that the approximation error is insignificant would be a necessary practice. This practice allows the training to emphasize on what is more practical to reduce the estimation error, which is the major error a classifier develops due to incomplete data training. The estimation error can be further decomposed into optimization and generalization errors, which are greatly complementary.
\par
Convexity, strong convexity, smoothness, and other features of the objective function (loss function) influence the optimization error. Furthermore, the convergence rate of the optimization problem relies on the algorithm used to solve it. For example, some algorithms have a linear convergence rate, and some have a sublinear or superlinear convergence rate. The computational complexity of an algorithm is a measure of how much computer resources the algorithm utilizes to solve the optimization problem. As a result, computational complexity can be quantified in units of storage, time, dimension, or all three simultaneously.
\par
A common methodology to quantify the computational complexity of optimization algorithms is by counting entire gradient evaluations required to obtain an optimal solution with a given accuracy $\epsilon$. The Gradient Descent algorithm is the most popular deterministic optimization algorithm with a linear convergence rate assuming $\mu$-strongly convex and $L$-smooth functions and a computational complexity of $\mathcal{O}( \frac{L}{\mu} N \log\frac{1}{\epsilon} )$ for ${N}$ data objective function. On the other hand, the Stochastic Gradient Descent is the most common algorithm that randomly picks a single function every iteration and thus has different computational complexity iteration $\mathcal{O}(\frac{1}{\epsilon} )$.
When $N$ is large, the preferred methods for solving the resulting optimization or sampling problem usually rely on stochastic estimates of the gradient of $f$. \par
Standard variance reduction techniques used for stochastic optimizations require additional storage or the computation of full gradients. Another approach for variance reduction is through adaptively increasing the sample size used to compute gradient approximations.
Some adaptive sampling optimization methods sizes have been studied in \cite{6-samplesize,Raghu-samplesize, daneshmand2016starting, Hashemi-samplesize, mokhtari2017first}. These methods have optimal complexity properties, making them useful for various applications. \cite{Hashemi-samplesize} uses variance-bias ratios to consider a test that is similar to the norm test, which is reinforced by a backup mechanism that ensures a geometric increase in the sample size. \cite{Raghu-samplesize} establishes terms for global linear convergence by investigating methods that sample the gradient and the Hessian. Other noise reduction methods like SVRG, SAG, and SAGA, either compute the full gradient at regular intervals or require storage of the component gradients, respectively \cite{Ayoub-samplesize} .
\section{Problem Definition}
The ultimate goal of most machine learning algorithms is to estimate the underlying distribution e.g.: $\mathcal{D}(\mathcal{X},\mathcal{Y})$ where $\mathcal{X}$ is the input feature space, and $\mathcal{Y}$ is the label space, in terms of some hypothesis function $h:\mathcal{X} \rightarrow \mathcal{Y}$, further on we assume that $h$ is determined by a parameter $w$ \cite{shalev2014understanding}.
Given any set $\mathcal{H}$ (that plays the role of hypotheses space) and domain $\mathbf{Z}:\mathcal{X\times Y}$: let $\ell$ be any function that maps from $\mathcal{H} \times \mathbf{Z}$ to the set of non-negative real numbers e.g. $\ell$ :
$\mathcal{H} \times \mathbf{Z} \rightarrow \mathbb{R}_+$ . The ability of the proposed hypothesis to estimate the underlying distribution is assessed by such loss functions.
The expected loss of a classifier, $h \in \mathcal{H}$, with regard to a probability distribution $\mathcal{D}$ over $Z$ is measured by the risk function \ref{eq:1}.
\begin{equation}
\label{eq:1}
L_{\mathcal{D}}(h) := \mathbb{E}_{z\sim \mathcal{D}} (\ell(h,z))
\end{equation}
Since this Expected Loss is built on the unknown distribution $\mathcal{D}$, the empirical risk over a given sample of the data $S = (z_1,\dots, z_m) \in Z_m$ is proven to be a good estimator of the expected loss, namely,
\begin{equation}
\label{eq:ERM}
L_s(h) := \frac{1}{m} \sum_{i=1}^{m}\ell(h,z_i)
\end{equation}
\label{sec:Definition}
\subsection{Computational complexity} The computational complexity is used to relate an excess error's upper bound (if one exists) to the available computational resources. Not only the convergence rate, but also the computational resources utilized to accomplish that convergence rate is important in order to have a superior algorithm. If we define a family $\mathcal{H}$ of candidates prediction functions, and let $h_s :=\argmin_{h\in\mathcal{H}}L_s(f)\implies$ ERM solution, $h^*:\argmin_{h}L_{\mathcal{D}}(h)\implies$ True solution (unknown)
$h_{\mathcal{H}} :=\argmin_{h\in\mathcal{H}}L_{\mathcal{D}}(h)\implies$ Best in class solution (unknown)
We can decompose the true loss (excess loss) as follow:
\[ \mathcal{E}(h_s,h^*) = \mathbb{E}[L(h^*_{\mathcal{H}})-L(h^*)]+\mathbb{E}[L(h_s)-H(h^*_{\mathcal{H}})] = \mathcal{E}_{app} + \mathcal{E}_{est}
\] Where the expectation w.r.t the samples.
\begin{itemize}
\item The approximation error $ \mathcal{E}_{app} $ measures how closely functions in $\mathcal{H}$
can approximate the optimal solution $h^*$.
\item The estimation error $\mathcal{E}_{est}$ measures the effect of minimizing the empirical risk $L_s(f )$ instead of the expected risk $L(f )$.
\item The estimation error is determined by the number of training
examples and the capacity of the family of functions.
\item The estimation error can be bounded using Rademacher
Complexity to measure the complexity of a family of functions.
\end{itemize}
Since the empirical risk $L_s (h )$ is already an approximation of the expected risk $L(h )$, it should not be necessary to carry out this
minimization with great accuracy. Assume that our algorithm returns an approximate solution $\hat{h}_s$ such that
\begin{equation}
\label{eq:rho}
L_s(\hat{h}_s) \leq L_s(h_s) +\delta_s
\end{equation}
where $\delta_s$ is a predefined positive tolerance. The new excess error can be decomposed as follows;
\begin{equation}
\mathcal{E}(\hat{h}_s,h^*) = \mathbb{E}[L(h^*_{\mathcal{H}})-L(h^*)]+\mathbb{E}[L(h_s)-L(h^*_{\mathcal{H}})]+\mathbb{E}[L(\hat{h}_s)-L(h_s)]= \mathcal{E}_{app}+\mathcal{E}_{est}+\mathcal{E}_{opt}
\end{equation}
Where the expectation w.r.t the samples. The additional term $\mathcal{E}_{opt}$ is optimization error. It reflects the impact of the approximate optimization on the generalization performance.
\subsection{General Model:}
The decomposition of the excess error leads to a trade-off minimization taking into account the number samples and allocating computation resources.
\begin{equation}
\begin{matrix}
\min_{\mathcal{H},\rho,m} & \mathcal{E}_{app} + \mathcal{E}_{est} + \mathcal{E}_{opt}\\
s.t. &m \leq m_{max}\\
&T(\mathcal{H},\rho,m) \leq T_{max}
\end{matrix}
\end{equation}
The variables are the size of the family of functions ${F}$, the optimizaton accuracy $\rho$ within the allotted training time $T_{max}$ , and the number of examples ${n}$ altered by using a subset of all available $n_max$ samples.\\
Typically when the size the class increases, the approximation error decreases, but the estimation error increases and nothing happens to optimization error because it's not related, but the computation time increase. When n increases the estimation error decreases and computation time increases, but no relation to approximation error or optimization error. When $\rho$ increases , the optimization error increases by definition, and computation time decreases.
\subsection{Statistical Error Minimization}In this paper we investigate the statistical error component of the excess error, which just comprises the difference of expected loss in some class and the empirical loss as shown below.
\[ \mathcal{E}_{stat} := \mathcal{E}(\hat{f}_n,f^*_{\mathcal{F}})= \mathbb{E}[E(\hat{f}_n)-E(f^*_{\mathcal{F}})]\]
Further we can add and substract some terms to have:
\begin{align*}
\mathcal{E}_{stat} := \mathcal{E}(\hat{f}_n,f^*_{\mathcal{F}})&= \mathbb{E}[E(\hat{f}_n)-E_n(\hat{f}_n)] +\underbrace{\mathbb{E}[E_n(\hat{f}_n)-E_n(f_n)]}_{\mathbb{E}[E_n(f_n)]=E(f_n)}\\
&+\underbrace{\mathbb{E}[E_n({f}_n)-E_n({f}^*_{\mathcal{F}})]}_{\leq0}+\underbrace{\mathbb{E}[E_n({f}^*_{\mathcal{F}})-E({f}^*_{\mathcal{F}})]}_{=0}\\
&\leq \underbrace{\mathbb{E}[E(\hat{f}_n)-E_n(\hat{f}_n)]}_{\mathcal{E}_{gen}} +\underbrace{\mathbb{E}[E_n(\hat{f}_n)-E_n(f_n)]}_{\mathcal{E}_{opt}}\\
\end{align*}
As a result, our primary goal is to minimize statistical error as follows:
\begin{equation}
\label{eq:min}
\begin{matrix}
\min_{\rho,m} & \mathcal{E}_{gen} + \mathcal{E}_{opt}\\
s.t. &m \leq m_{max}\\
&T(\rho,m) \leq T_{max}
\end{matrix}
\end{equation}
We list our assumptions below:
\begin{assumption}[Lipschits Continuity]\label{ass:Lipschitz continuous}
Assume for any $z\in\mathcal{Z}$, the loss function $l(\cdot;z)$ is G-Lipschitz continuous, i.e. $\forall w \in \mathcal{H}$,
\begin{align}
\nonumber |l(w;z)-l(w';z)|\leq G\left\|w-w'\right\|_2.
\end{align}
\end{assumption}
\begin{assumption}[Convexity]\label{ass:Convexity}
Assume for any $z\in\mathcal{Z}$ the loss function $l(\cdot;z)$ is convex function, i.e. $\forall w \in \mathcal{H}$,
\begin{align}
\nonumber l(w';z) \geq l(w;z) + \nabla l(w;z)^T (w' - w).
\end{align}
\end{assumption}
\begin{assumption}[L-Smooth]\label{ass:Smooth}
Assume for any $z\in\mathcal{Z}$ the loss gradient function $\nabla l(\cdot;z)$ is is L-Lipschitz continuous, i.e. $\forall w \in \mathcal{H}$,
\begin{equation}
\| \nabla \ell(w;z) - \ell(w';z)\| \leq L \| w - w' \|
\end{equation}
\end{assumption}
\section{Methodology}
\label{sec:others}
The contribution of this work is mainly deriving the computational complexity of sub-linear convergent algorithm with adaptive sample size training. In order to derive that we start by the generalized bound that haven been studied well in literature \cite{boucheron2005theory} :
\[ \mathbb{E} \left[\sup_{h\in\mathcal{H}} |L_s(h) - L(h)| \right] \leq V_n \approx \mathcal{O}\left(\frac{1}{n^\alpha}\right)
\]
where $\alpha \in [0 \; 0.5]$ depends on an algorithm used to solve ERM and other factors. The bound is found by \cite{vapnik1999nature} to be $\mathcal{O}(\sqrt{1/n \log{1/n}})\geq\mathcal{O}(\sqrt{1/n }) $, while in other references e.g. \cite{bartlett2006convexity} the bound is improved to $\mathcal{O}(1/n)$ under extra conditions in the regularizer. In any situation, the bound indicates that regardless of the ERM solution's optimization accuracy, there will always be a bound in the order $\mathcal{O}({1}/{n^\alpha})$, thus solving the ERM optimization problem with accuracy $\delta_s$ = 0 would not be beneficial to the final statistical error (estimation) minimization problem in \ref{eq:min} as illustrated in \cite{daneshmand2016starting}.
Thus, solving the ERM with a statistical accuracy of $\delta_s$ in \ref{eq:rho} equal to the $V_ n$ is sufficient to provide a uniformly stable result, namely hypothesis $h$. This solution is denoted in literature by calculating the ERM within its statistical accuracy.
\subsection{Adaptive Sample Size}
Following the work of \cite{mokhtari2017first}, an adaptive sample size scheme is employed to take advantage of the nature of ERM, namely, the finite sum of functions drawn identically and independently from the same distribution to achieve a higher convergence rate with lesser computational complexities. However, the research in \cite{mokhtari2017first} only focuses into linearly convergent algorithms (strongly convex loss function are implemented with the aid of L-2 norm regularizer). The adaptive sample size scheme starts with a small portion of the training samples and solves the correspoding ERM within its statistical accuracy, then expands to include new samples with the original one and solves the ERM with the initial solution found by the previous sample and repeats until all samples are finished.
\\
\par
In other words given training data samples $\mathcal{Z}$ with $|\mathcal{Z}|=s$, we initialize the training with small sample $S_m \subset \mathcal{Z}$ and solve ERM in \ref{eq:ERM} within its statistical accuracy namely $\mathcal{O}(1/m^\alpha)$ to find $\hat{h}_m $ defined by some weights $w_m$. Then expand the training sample to include new samples such that $S_m \subset S_n \subset \mathcal{Z}$ and solve the ERM with initial solution of $w_m$ to find the ERM solution $w_n$. Repeat this process until all data in ${s}$ are included .
\\
\par
The relationship between the consecutive solutions $w_n$ and $w_m$ with ${n = 2m}$ (the increase is discussed in section 4) is established by the following theorem. The bound is expressed in terms of the first sample statistical solution to indicate that solving the first ERM problem with zero accuracy is not required.
\begin{theorem}
\label{th:th1}
Given the solution $w_m$ that solves the ERM with tolerance $\delta_m$ on sample $S_m \subset \mathcal{Z}$ such that in expectation $\mathbb{E}[L_m(w_m) - L_m(w^*_m)] \leq \delta_m$. Assume the there exist an optimal solution $w_n$ on sample $S_n$ such that $S_m \subset \mathcal{Z}$ and its statistical accuracy $V_n$, then in expectation we have the empirical risk difference is bouneded in expctation between the $w_n^*$ and $w_m$ as:
\begin{equation}
\label{eq:th1}
\mathbb{E}[L_n(w_m) - L_n(w_n^*] \leq \delta_m + \frac{n-m}{n} (2V_{n-m} + V_m +V_n)
\end{equation}
\end{theorem}
\begin{proof}
Starting by rewrite difference between the empirical losses using two models and denote $S_{n-m}$ the set of samples in $S_n \cap (S_m \cap S_n)^c$ thus:
\begin{align}
\label{eq:pr1}
\mathbb{E}[L_n(w_m) - L_n(w_n^*)] &= \mathbb{E}[\underbrace{(L_n(w_m) - L_m(w_m))}_{1} + \underbrace{(L_m(w_m) - L_m(w_m^*))}_{2} \\
& + \underbrace{(L_m(w_m^*) - L_m(w_n^*))}_{3} + \underbrace{(L_m(w_n^*) - L_n(w_n^*))}_{4}]
\end{align}
The first difference is bounded by Lemma 5 in \cite{mokhtari2017first} as \[\mathbb{E}[|L_n(w_m) - L_m(w_m)|] \leq \frac{n-m}{n} (V_{n-m} +V_m) \]
The second difference us the optimization error which is assumed to be : \[ \mathbb{E}[L_m(w_m) - L_m(w_m^*)] \leq \delta_m \]
The third difference is bounde above by zero since $w_m^*$ is the minimizer of the empire risk $L_m$. \[\mathbb{E}[L_m(w_m^*) - L_m(w_n^*)] \leq 0 \]
The forth difference is bounded by Lemma 5 in \cite{mokhtari2017first} as \[\mathbb{E}[|L_m(w_n^*) - L_n(w_n^*)|] \leq \frac{n-m}{n} (V_{n-m} +V_n) \]
Putting all four bounds back in \ref{eq:pr1} to obtain the result in theorem \ref{eq:1}.
\end{proof}
Theorem 1 asserts that even with the most accurate $L_m$ ERM solution i.e. $\ delta_m =0$, the subsequent problem $L_n$ with $w_0 = w_m$ will always have an optimal solution that has a dependency on the $V_m$. Thus solving the $L_m$ should be only withing $\mathcal{O}(V_m)$ only to reduce the computational complexity. The results \ref{eq:th1} in theorem \ref{th:th1} can be simplified if we consider $V_n = 1/n^\alpha$ and $n = 2m$ in Lemma \ref{lemma:lemma1}.
\begin{lemma}
\label{lemma:lemma1}
\begin{align}
\mathbb{E}[L_n(w_m) - L_n(w_n^*] &\leq \delta_m + \frac{1}{2} \left(\frac{2}{(n-m)^\alpha} + \frac{1}{m^\alpha} +\frac{1}{n^\alpha}\right)\\
& = \delta_m + \frac{1}{2} \left(\frac{2}{m^\alpha} + \frac{1}{m^\alpha} +\frac{1}{(2m)^\alpha}\right)\\
& = \delta_m + \frac{1}{2 } \left(3 +\frac{1}{2^\alpha}\right)V_m
\end{align}
\end{lemma}
\subsection{Computational Complexity}
The computational complexity of an algorithm is a measure of the algorithm's recruitment of computer resources, and the less computing required to accomplish one iteration in an iterative process, the simpler the algorithm is. Typically, it's measured in cost units associated with the algorithm; for example, some algorithms are assessed in gradient evaluation or number of iterations, while others require counting the total number of computing activities performed by the machine. First the smooth loss function assumption is stated as below.
Theorem \ref{th:th2} provides the minimal number of iterations required to solve the ERM on subset $S_n$ within statistical accuracy, i.e. $\mathbb{E}[L_n(w_n)-L_n(w_n^*)] \leq V_n$ given that the optimization algorithm has a sublinear convergence rate.
\begin{theorem}
\label{th:th2}
Given the initial solution $w_m$ and assuming the optimal solution of the ERM on the subset $S_n \subset \mathcal{Z}$ to be $w^*_n$, the sublinear convergence optimization algorithm needs the following iteration to solve the ERM within its statistical accuracy:
\begin{equation}
T\geq \left[2^\alpha \left(\frac{5}{2 } +\frac{1}{2^{\alpha+1}}\right) \right]^\frac{1}{\zeta}
\end{equation}
Where T is the iteration number, and $\zeta$ is a positive constant determined by the optimization setting.
\end{theorem}
\begin{proof}
Given the initial solution $w_m$ and assuming the optimal solution of the ERM on the subset $S_n \subset \mathcal{Z}$ to be $w^*_n$, the sublinear convergence optimization bounds the difference in expectation as:
\begin{align}
\mathbb{E}[L_n(w_T) - L_n(W_n^*]& \leq \frac{1}{T^\zeta} (L_n(w_m) - L_n(w^*_n))\\
&\underset{a}{\leq}\frac{1}{T^\zeta}\left[V_m + \frac{1}{2 } \left(3 +\frac{1}{2^\alpha}\right)V_m\right]
\end{align}
Where the inequality (a) comes from the results in lemma \ref{lemma:lemma1} with $\delta_m = V_m$. In order to solve the ERM in $V_n$ accuracy we need to bound the RHS last equation by $V_n$ as follow:
\begin{align}
\frac{1}{T^\zeta}\left[V_m + \frac{1}{2 } \left(3 +\frac{1}{2^\alpha}\right)V_m\right] &\leq V_n\\
\frac{1}{T^\zeta} \left(\frac{5}{2 } +\frac{1}{2^{\alpha+1}}\right) &\leq \frac{1}{2^\alpha}\\
T\geq \left[2^\alpha \left(\frac{5}{2 } +\frac{1}{2^{\alpha+1}}\right) \right]^\frac{1}{\zeta}
\end{align}
\end{proof}
The number of iterations T in theorem \ref{th:th2} ensures that the solution of any phase (stage) meets the statistical accuracy utilizing this lower constraint based on the iterative optimization algorithm being used. With a batch or sample of data, the ERM problem in equation \ref{eq:ERM} is solved until the statistical accuracy of that batch is guaranteed, and the solution is then employed as an initial solution for the next batch. The requirement of statistical accuracy, on the other hand, necessitates access to the unknown minimizer $w^*_n$, thus Theorem \ref{th:th2} examines the minimum iterations needed such that an iterative method might utilize as a stopping criteria. Now the algorithm of solving the ERM problem in an adaptive way is illustrated in Algorithm 1.
\begin{algorithm}[!h]
\caption{Adaptive Sample Size iterative ERM solver Algorithm}\label{alg:cap}
\begin{algorithmic}
\label{alg:alg1}
\Require initial Sample size: $m_0$, Initial Solution $w^0$ such that $\mathbb{E}[L_{m_0}(w^0)-L_{m_0}(w^*)
\leq V_{m_0}]$, $\alpha$, $\zeta$
\Ensure $w_s \leq w_s^* +V_s$
\State $m \gets m_0$
\While{$n \leq |\mathcal{Z}|$}
\State $w_m \gets w^0$
\State $n \gets \min(2m,|\mathcal{Z}|)$
\State $T_{max} \gets \left[2^\alpha \left(\frac{5}{2 } +\frac{1}{2^{\alpha+1}}\right) \right]^\frac{1}{\zeta}\log(n)$
\State Solve ERM problem on $S_n\subset\mathcal{Z}$ with $T_{max}$ and initial solution $w_m$
\State $w^0 = w_n$
\EndWhile
\end{algorithmic}
\end{algorithm}
\section{Experiment}
The experiments in this part are carried out with first-order optimization algorithms that have a sub-linear convergence rate on an objective function that meets assumption 1, namely L-smooth and convex. The experiment's purpose is to assess the sub-optimality of these algorithms when adaptive sample size can be used against their fixed sample size counterpart. The gradient descent algorithm is selected from deterministic algorithms, whereas the ADAM algorithm is selected from stochastic algorithms. The logistic function with binary classification is the objective loss function to be minimized.\\
\par
We Refer to Gradient Descent with adaptive sample size as adaptive gradient descent (adaGD) and for ADAM with adaptive sample size as adaptive ADAM (adaADAM).
The characteristics and total samples attributes of the datasets considered are detailed in table \ref{tab:table1}. Only the digits zero and eight are represented in binary in the MNIST databases. The Gradient Descent algorithm is first ran on every data set for a large number of iteration to obtain the optimal value. Then the ADAM algorithm have fixed parameters as : 1st-order exponential decay $\beta_1 = 0.9$, 2nd-order exponential decay $\beta_2=0.999$, step size $\eta=0.01$ and a small value $\epsilon=1e-8$ to prevent zero-division. The batch size is chosen to be 5 in ADAM and in adaADAM. The gradient descent step size is chosen based on the L-smoothness parameter values: $\gamma =1/L$.\\
\par
\begin{table}[h]
\centering
\caption{Dataset Characteristics}
\begin{tabular}{c c c c}
\hline
Name: & MNIST & RCV1 & a1a \\
\hline
training size & 6000 & 20242 & 1,605 \\
testing size & 5774 & 677,399 & 30,956\\
features size & 784 & 123 & 47,236\\
\hline
\end{tabular}
\label{tab:table1}
\end{table}
\begin{figure}[!h]
\centering
\includegraphics[width=\textwidth]{adaGrad.png}
\caption{Comparison between Gradient descent and Adaptive sampling of Gradient descent}
\label{fig:fig1}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=\textwidth]{adaAdam.png}
\caption{Comparison between ADAM and Adaptive sampling of ADAM}
\label{fig:fig2}
\end{figure}
\section{Discussion}
This paper presents an adaptive sampling technique to simplify the ERM problem for first-order optimization algorithms with sublinear convergence under terms of convexity and L-smoothness. Based on the carried experiments, we can infer that adaptive sampling generally resulted in faster convergence for sublinear problems. The adaptive Gradient (adaGD) has reduced the computational complexity of minimizing the logistic loss on the three datasets MNIST, RCV1, and a1a, as shown in figure \ref{fig:fig1}. However, the MNIST dataset has the greatest reduction in complexity, which is characterized in gradient evaluations, while the a1a dataset has the least.
\\
\par
The reduction in computational complexity in adaptive ADAM (adaADAM) is not significant in some datasets, such as MNIST and RCV1, but it has proven to be significant in a1a as shown in figure \ref{fig:fig2}. The explanation for this could be that the ADAM algorithm stochastically shuffles the dataset after each epoch, which could result in the samples being repeated in different batches, which would employ the adaptive sampling technique implicitly. Future work related to exploring dataset characteristics that would limit convergence rate enhancement for sublinear problems through adaptive sampling is of interest.
\section*{Acknowledgments}
The work in this paper was a continuation of \cite{mokhtari2017first} and it was supported by the course instructor Dr. Bin Gu.
\bibliographystyle{apalike}
\section{Introduction}
\lipsum[2]
\lipsum[3]
\section{Headings: first level}
\label{sec:headings}
\lipsum[4] See Section \ref{sec:headings}.
\subsection{Headings: second level}
\lipsum[5]
\begin{equation}
\xi _{ij}(t)=P(x_{t}=i,x_{t+1}=j|y,v,w;\theta)= {\frac {\alpha _{i}(t)a^{w_t}_{ij}\beta _{j}(t+1)b^{v_{t+1}}_{j}(y_{t+1})}{\sum _{i=1}^{N} \sum _{j=1}^{N} \alpha _{i}(t)a^{w_t}_{ij}\beta _{j}(t+1)b^{v_{t+1}}_{j}(y_{t+1})}}
\end{equation}
\subsubsection{Headings: third level}
\lipsum[6]
\paragraph{Paragraph}
\lipsum[7]
\section{Examples of citations, figures, tables, references}
\label{sec:others}
\lipsum[8] \cite{kour2014real,kour2014fast} and see \cite{hadash2018estimate}.
The documentation for \verb+natbib+ may be found at
\begin{center}
\url{http://mirrors.ctan.org/macros/latex/contrib/natbib/natnotes.pdf}
\end{center}
Of note is the command \verb+\citet+, which produces citations
appropriate for use in inline text. For example,
\begin{verbatim}
\citet{hasselmo} investigated\dots
\end{verbatim}
produces
\begin{quote}
Hasselmo, et al.\ (1995) investigated\dots
\end{quote}
\begin{center}
\url{https://www.ctan.org/pkg/booktabs}
\end{center}
\subsection{Figures}
\lipsum[10]
See Figure \ref{fig:fig1}. Here is how you add footnotes. \footnote{Sample of the first footnote.}
\lipsum[11]
\begin{figure}
\centering
\fbox{\rule[-.5cm]{4cm}{4cm} \rule[-.5cm]{4cm}{0cm}}
\caption{Sample figure caption.}
\label{fig:fig1}
\end{figure}
\subsection{Tables}
\lipsum[12]
See awesome Table~\ref{tab:table}.
\begin{table}
\caption{Sample table title}
\centering
\begin{tabular}{lll}
\toprule
\multicolumn{2}{c}{Part} \\
\cmidrule(r){1-2}
Name & Description & Size ($\mu$m) \\
\midrule
Dendrite & Input terminal & $\sim$100 \\
Axon & Output terminal & $\sim$10 \\
Soma & Cell body & up to $10^6$ \\
\bottomrule
\end{tabular}
\label{tab:table}
\end{table}
\subsection{Lists}
\begin{itemize}
\item Lorem ipsum dolor sit amet
\item consectetur adipiscing elit.
\item Aliquam dignissim blandit est, in dictum tortor gravida eget. In ac rutrum magna.
\end{itemize}
\section{Conclusion}
Your conclusion here
\section*{Acknowledgments}
This was was supported in part by......
\bibliographystyle{unsrt}
| a28e39f7d97589abcb69be62077bbbff4847c7ea | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{SUPPLEMENTAL MATERIAL}
\subsection{Section 1: Photoelectron emission angle dependence of core-level spectra}
Figure 6 shows a comparison of core-level spectra measured at different photoelectron emission angles $\theta = 90^\circ$ and $60^\circ$ [see Fig. 6(a) for the definition of $\theta$].
Because of the finite escape depth of photoelectrons ($\lambda$) in the sample, the photoemission spectrum taken at a smaller $\theta$ is more surface sensitive; an effective escape depth $\lambda_{\rm eff}=\lambda\sin{\theta}$ is reduced by $\sim$15\% by changing $\theta$ from $90^\circ$ to $60^{\circ}$ [see Fig. 6(a)].
If the Sb-$4d$ core levels had surface and bulk components at higher and lower binding energies ($E_{\rm{B}}$’s), respectively, and the surface component became stronger in intensity by changing $\theta$ from $90^\circ$ to $60^\circ$ as observed for the Cs-$4d$ core levels (see below), the peak position of Sb-$4d$ core levels at $\theta = 60^\circ$ would be located at higher $E_{\rm{B}}$ than that at $\theta = 90^\circ$.
However, as shown in Fig. 6(b), such an energy shift is not seen within the present experimental uncertainty of 10 meV, which is much smaller than the 130-meV energy shift observed by the termination-dependent ARPES study.
This result is not consistent with the two-components scenario, but supports the doping scenario as the origin of the energy shift in the Sb-$4d$ core level [Fig. 1(d)].
In contrast, the Cs-$4d$ core levels show surface and bulk components.
As shown in Fig. 6(c), each spin-orbit satellite of Cs-$4d$ core levels ($d_{\rm{5/2}}$ or $d_{\rm{3/2}}$) consists of two peaks.
A peak at lower $E_{\rm B}$ originates from bulk Cs atoms, whereas that at higher $E_{\rm B}$ is from surface Cs atoms, as supported by the angular dependence of their intensity ratio; the shorter $\lambda_{\rm eff}$ at $\theta = 60^{\circ}$ leads to the stronger surface-derived (weaker bulk-derived) peak intensity of the Cs-$4d$ core levels.
\begin{figure}
\includegraphics[width=3.4in]{figS1.eps}
\caption{
(a) Schematic of experimental setup and the definition of photoelectron emission angle $\theta$, escape depth $\lambda$, and effective escape depth $\lambda_{\rm eff}$. (b) Comparison of EDCs in the Sb-$4d$ core-level region measured at $\theta = 90^{\circ}$ (black) and $60^{\circ}$ (red) for the Cs-terminated surface. (c) Same as (b) but for the Cs-$4d$ core-level region.
}
\end{figure}
\begin{figure}
\includegraphics[width=2in]{figS2.eps}
\caption{
Comparison of energy distribution curves (EDCs) in the Cs-$4d$ core-level region measured for Cs- and Sb-terminated surfaces (red and blue curves, respectively).
}
\end{figure}
\subsection{Section 2: Peak position of the Cs-$4d$ core levels}
The termination-dependent Cs-$4d$ core-level data in Fig. 7 shows that the surface components of the Cs-terminated surface shift toward lower $E_{\rm{B}}$ by $\sim$30 meV with respect to those of the Sb-terminated surface.
Since this direction is opposite from that observed in the Sb-$4d$ core levels, the observed Cs-$4d$ core-level shift cannot be understood in terms of the carrier doping, but probably reflects a difference in the density of Cs atoms at the surface.
It is known that the core level of alkaline-metal atoms shifts toward lower $E_{\rm{B}}$ with increasing the number of alkaline-metal atoms per area.
Such a characteristic energy shift is theoretically explained by a Born-Haber cycle \cite{PirugSurfSci1985} and provides a reasonable explanation on our observation.
Namely, the Cs-terminated surface has the higher density of Cs atoms compared to that of the Sb-terminated surface and hence the Cs-$4d$ core level appears at lower $E_{\rm{B}}$, whereas the Sb-$4d$ core levels shift toward higher $E_{\rm{B}}$ because of the doping-induced chemical potential shift.
A similar behavior was reported in K-dosed Fe(Se,Te) \cite{ZhangAPL2014} in which gradual K deposition on the surface leads to the shift of the Te-$4d$ core level toward higher $E_{\rm{B}}$ due to the electron doping to the Fe(Se,Te) layer, whereas the K-$3p$ core level shifts toward lower $E_{\rm{B}}$.
\begin{figure}
\includegraphics[width=3.4in]{figS3.eps}
\caption{
(a), (b) ARPES intensity measured along the KM cut at $T = 120$ K in the Cs- and Sb-terminated surfaces, respectively.
Green dashed curves are a guide for the eyes to trace the SP-band dispersions.
}
\end{figure}
\subsection{Section 3: Surface-termination dependence of the saddle-point position}
Figure 8 shows the ARPES data that supports the SP-band energy shift, in which a comparison of the ARPES intensity along the KMK cut between the Cs- and Sb-terminated surfaces is made.
One can directly recognize that the saddle point (SP) band (green dashed curve) is located at lower $E_{\rm{B}}$ in the Sb-terminated surface compared to that in the Cs-terminated surface, signifying the difference in the doping level.
\subsection{Section 4: Reproducibility of polar-surface formation}
We have performed ARPES measurements more than three times with different samples for each surface termination and confirmed the reproducibility, e.g. the termination-dependent band energy shifts, as seen by a comparison of the band dispersions measured with different samples in Fig. 9.
We also performed ARPES measurements with some different photon energies between 70 and 130 eV which have comparable probing depths, and obtained qualitatively the same results.
\subsection{Section 5: Band splitting observed in the Cs-terminated surface}
Figure 10 shows representative EDCs and momentum distribution curves (MDCs) at $T = 8$ K measured for the Cs-terminated surface.
One can clearly see a two-peaked structure in EDCs [Figs. 10(c) and 10(d)] and MDCs [Figs. 10(e) and 10(f)] which supports the band splitting associated with the CDW-induced band folding.
\begin{figure}
\includegraphics[width=3.4in]{figS4.eps}
\caption{
(a), (b) Comparison of the band dispersions at $T = 8$ K for the Sb-terminated surface between two different $\rm{CsV}_{3}\rm{Sb}_{5}$ samples (samples A and B). (c), (d) Same as (a) and (b) but for the Cs-terminated surface.}
\end{figure}
\subsection{Section 6: A signature of in-plane band folding}
Figures 11(a) and 11(b) show a comparison of the second-derivative ARPES intensity along the $\Gamma$KMK$\Gamma$ cut at $T = 8$ K and 120 K for the Cs-terminated surface.
We observed the discontinuity of the $\beta$ band dispersion along the $\Gamma$K line of the Cs-terminated surface at $T = 8$ K, as indicated by white arrows in Fig. 11(a).
This anomaly is of CDW origin because it disappears at $T = 120$ K [Fig. 11(b)] and is well reproduced by the density-functional theory that incorporates only the in-plane $2\times2$ periodicity [Fig. 11(c)].
Therefore, the Cs-terminated surface shows the unit-cell doubling not only along the $c$ axis but also in the $a-b$ plane.
On the other hand, for the Sb-terminated surface, the discontinuity of the $\beta$ band is absent even at 8 K, so both the c-axis and in-plane unit-cell doublings are suppressed.
\begin{figure}
\includegraphics[width=3.4in]{figS5.eps}
\caption{
(a), (b) ARPES intensity plots at $T = 8$ K measured along $k_{y}$ at $k_{y}=0$ ($\Gamma$KM cut) and $-\pi$ (KMK cut), respectively.
(c), (d) EDCs at representative energy slices indicated by red lines in (a).
(e), (f) MDCs at representative $\textbf{k}$ slices shown by red lines in (b).
Blue curves in (c)-(f) are the result of numerical fittings.
}
\end{figure}
\subsection{Section 7: Comparison of Fermi surface between Cs- and Sb-terminated surfaces}
Figures 12(a) and 12(b) show a comparison of the ARPES intensity map at the Fermi level ($E_{\rm{F}}$) around the K point between the Cs- and Sb-terminated surfaces.
The peak position of MDC at $E_{\rm{F}}$ is overlaid for the triangular-shaped electron Fermi surface (FS; black circles and red triangles, respectively).
Direct comparison of the MDC peak position in Fig. 12(c) shows the shrinkage of the triangular FS in the Sb-terminated surface, consistent with the hole-doped character.
\begin{figure}
\includegraphics[width=3.4in]{figS6.eps}
\caption{
(a), (b) Second-derivative ARPES-intensity plots for the Cs-terminated surface at $T = 8$ K and 120 K, respectively.
White arrows in (a) highlight the appearance of band hybridization by the in-plane band folding.
(c) Calculated band structure of $\rm{CsV}_{3}\rm{Sb}_{5}$ with the $2\times2\times1$ unit cell, reproduced from \cite{TanPRL2021}.
}
\end{figure}
\begin{figure}
\includegraphics[width=3.4in]{figS7.eps}
\caption{
(a), (b) ARPES intensity plots around the K point measured at $T = 8$ K for the Cs- and Sb-terminated surfaces, respectively.
Black and red triangles in (a) and (b) represent the Fermi wave vectors determined by numerical fittings to the MDCs at $E_{\rm{F}}$.
(c) Comparison of the Fermi wave vectors between the Cs- and Sb-terminated surfaces. The data points were symmetrized by assuming the $C_{3}$ symmetry centered at the K point.
}
\end{figure}
\begin{figure}
\includegraphics[width=3.4in]{figS8.eps}
\caption{
(a), (b) ARPES-intensity plots at $T = 120$ K and 8 K, respectively, measured along the KM cut.
(c), (d) Second-derivative ARPES-intensity plots of (a) and (b), respectively.
(e), (f) Calculated band structures with SOC for (e) $k_{z} = 0$ (red) and (f) $k_{z} = 0$ (red) and $\pi$ (blue) (reproduced from \cite{NakayamaPRB2021}).
The calculations were carried out using the full-potential linearized augmented plane-wave method implemented in the WIEN2K code with generalized gradient approximation (GGA) and the Perdew-Burke-Ernzerhof (PBE) type exchange-correlation potential.
Spin-orbit coupling was included self-consistently, while the lattice parameters were directly obtained from experiments.
The $k$-points mesh for the irreducible Brillouin zone is $17\times17\times10$.
The Muffin-tin radii are 2.50 a.u. for Cs and V and 2.60 a.u. for Sb, respectively.
The maximum modulus for the reciprocal vectors K$_{\rm{max}}$ was chosen to satisfy $R_{\rm{MT}}\times\rm{K}_{\rm{max}}=8.0$.
}
\end{figure}
\subsection{Section 8: Comparison between the experimental and calculated band structures}
To clarify the origin of band doubling in the charge-density-wave (CDW) phase for the Cs-terminated surface of $\rm{CsV}_{3}\rm{Sb}_{5}$, we compare the ARPES results with the first-principles band-structure calculations. Figures 13(a) and 13(b) show the ARPES intensity along the KM cut at $T = 120$ K (above $T_{\rm{CDW}}$) and 8 K (below $T_{\rm{CDW}}$), respectively, measured with $h\nu = 106$ eV photons corresponding to the out-of-plane wave vector of $k_{z} \sim 0$ \cite{NakayamaPRB2021}.
The second-derivative intensity is displayed in Figs. 13(c) and 13(d).
At $T = 120$ K, we observe five dispersive bands (labelled as $\beta, \gamma, \delta, \varepsilon,$ and $\eta$) near $E_{\rm{F}}$ [Figs. 13(a) and 13(c)], in agreement with the first-principles calculations for the KM cut ($k_{z} = 0$) [Fig. 13(e)].
On the other hand, at $T = 8$ K, these bands show doubling as marked by green arrows in Fig. 13(d) [doubling of the $\delta$ band is seen in the data obtained with higher energy resolution in Figs. 3(h) and 3(i)].
This band doubling is explained in terms of the band folding along the $k_{z}$ direction due to the three-dimensional (3D) CDW \cite{LiangPRX2021}.
Namely, the doubling of unit-cell along the c-axis causes the folding of energy bands from $k_{z} = \pi$ to 0, and vice versa.
The calculated band dispersions superimposed with those of $k_{z} = 0$ and $\pi$ [Fig. 13(f)] is in good agreement with the experimental results, supporting the 3D-CDW origin of the band doubling for the Cs-terminated surface.
It is noted that to obtain a better matching between the experiment and calculation, additional level splitting associated with the repulsion between the original and folded bands (especially for the $\varepsilon$ band) as well as the CDW-gap opening near $E_{\rm{F}}$ needs to be taken into account.
It is also noted that, besides the $2\times2\times2$ periodicity, the $2\times2\times4$ periodicity has been also reported in $\rm{CsV}_{3}\rm{Sb}_{5}$ \cite{OrtizPRX2021}.
We naively expect the quadruple band splitting when the charge order has the $2\times2\times4$ periodicity.
But our data show only doubling. In this respect, our results are more consistent with the $2\times2\times2$ periodicity.
\end{document} | 8bd6b2f8dad01920f4bd47a8922cb3e4df6daeb9 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{introduction}
Following the success of the use of neural networks across different fields of science~\cite{ching2018opportunities,topol2019high,hannun2019cardiologist,mehta2019high} for detecting patterns in data, proposals have been set forth in the quantum regime~\cite{fujii2017harnessing,qrpreview}.
In this direction, a particular quantum neural network architecture has emerged -- termed \emph{quantum reservoir processing}, in analogy to classical reservoir computing~\cite{montavon2012neural}. In such architecture, the quantum network serving as a processor is composed of randomly interacting quantum systems (the \emph{nodes}), not requiring precise control.
The function of this kind of network is learned in a training procedure that measures the system individually and fixes only a single output layer, which makes it experimentally friendly.
This architecture has been proposed for executing classical tasks~\cite{fujii2017harnessing,govia2021quantum,xu2021superpolynomial} (showing performance advantage over classical networks) and quantum tasks such as state characterisation~\cite{ghosh2019quantum,ghosh2020reconstructing}, quantum state preparation~\cite{ghosh2019quantum2,creating2020}, gate compression~\cite{ghosh2021realising}, and quantum metrology~\cite{krisnanda2022phase} (see Refs.~\cite{markovic2020quantum,qrpreview} for reviews).
Remarkably, for characterisation and metrological tasks, it is not necessary to perform correlation measurements, and it suffices to measure only local observables such as average occupation numbers or intensities of the network nodes.
The platform is versatile and it holds the potential to directly estimate important quantities such as quantum entanglement, which is the focus of our study.
Entanglement is a special type of correlation between two or more objects, the presence of which witnesses their quantum nature~\cite{horodecki2009quantum}.
In experiments involving objects that cannot be accessed directly, their quantum character could be revealed by using such inaccessible systems as mediators between two accessible probes.
The revelation of an entanglement gain between the probes then provides proof of another quantum signature -- known as quantum discord -- of the mediators~\cite{krisnanda2017revealing}.
This experimental scheme has been put forward as a proposal to probe quantum signatures of gravity through the observation of gravity-induced entanglement between masses~\cite{bose2017spin,marletto2017gravitationally,krisnanda2020observable} (see also Refs.~\cite{balushi2018,belenchia2018quantum,qvarfort2020,van2020quantum,rijavec2021decoherence,margalit2021realization,Pedernales2022} for recent developments and discussion).
This motivates the general framework presented in this paper, which is aimed at sensing (possibly weak) entanglement and its application to gravity-induced entanglement.
To date, there are essentially two main schemes proposed for the observation of gravity-induced entanglement, which suffer of different practical difficulties.
The Bose \emph{et al.}-Marletto-Vedral (BMV) scenario~\cite{bose2017spin,marletto2017gravitationally} requires preparation of a macroscopic superposition of each of two nearby massive bodies, whose later dynamics might showcase gravitational entanglement.
The challenging state-preparation stage is bypassed in the proposal of Ref.~\cite{krisnanda2020observable}, which resorts to continuous-variable (CV) entanglement between masses that begin in natural and easy-to-arrange Gaussian states. In this scheme, the entanglement detection remains as a demanding step.
We show that a relatively simple neural-network architecture is sufficient to achieve a two-order-of-magnitude improvement in the precision with which entanglement of massive systems can be estimated, when compared to state-of-the-art values~\cite{palomaki2013entangling}.
Specifically, we utilise a reservoir quantum network (QN) for precise entanglement sensing.
In particular, quantum objects whose entanglement we want to scrutinise (the input) are put in contact with a QN.
The observables from the QN -- which can be as simple as the mean excitation numbers of the nodes -- are post-processed through a single output layer.
This layer is trained so that the final output estimates quantum entanglement of the input objects.
Our general platform is particularly useful in situations where the input is not accessible for direct measurements, the latter are complicated (this is particularly the case for those that necessitate conditional or correlation measurements), or in cases where the input is less resilient to measurement errors than the QN.
First, we will introduce the general framework with generic quantum systems.
We show that a QN can learn from a random set of non-entangled input states and nevertheless is able to estimate the amount of entanglement at the testing stage.
For a more realistic scenario considering measurement errors, we show that the entanglement precision scales better than
$\Delta_E \propto 1/\sqrt{\mathcal{N}}$, where $\mathcal{N}$ is the number of measured observables.
We shall refer to $1/\sqrt{\mathcal{N}}$ scaling as the standard quantum limit (SQL).
Finally, we demonstrate an explicit application of our framework and its features to the recent endeavour whose goal is to reveal quantum features of gravity by measuring gravity-induced entanglement between masses.
In particular, we show that measurements on cavity modes, which have interacted with the masses, can be post-processed to estimate the gravity-induced entanglement. Importantly, our approach offers better sensitivity compared to direct measurements on the masses.
\section{The general framework}
Our thought platform is depicted in Fig.~\ref{Fig_framework}.
Consider that quantum objects, whose entanglement is to be estimated, serve as the input.
They come in contact with a processor, namely, a QN composed of quantum nodes.
Note that the QN nodes require minimal control, e.g., they can be randomly interacting with each other, the input, and environment.
We also allow that they are pumped by external coherent sources (e.g., in an optical system, lasers).
The purpose of the contact is a flow of information from the input to the QN.
By retrieving the observables from the QN and processing them via a single output layer, one obtains a final output signal.
We will show that by training a set of weights and biases in the single output layer, the output signal estimates entanglement of the input objects.
\begin{figure}[h]
\centering
\includegraphics[width=0.48\textwidth]{Fig1}
\caption{Illustration of a quantum neuromorphic platform for entanglement sensing.
It involves to-be-measured input objects that are connected to a quantum network composed of uncontrolled nodes having random interactions with the input $K_{nm}$ and between themselves $J_{mm^{\prime}}$.
Both the input and the nodes interact with their environments, denoted by $\gamma_{n}$
and $\kappa_m$, respectively.
The observables from the QN are processed by a trained output layer, producing a signal that estimates input entanglement.
}
\label{Fig_framework}
\end{figure}
Let us consider generic quantum systems and their dynamics with which we demonstrate the general framework described above.
In what follows, we consider continuous-variable systems (bosons).
Additionally, the platform also works for generic discrete systems (e.g., qubits), see \textcolor{blue}{Appendix~A} for details, as well as hybrid discrete-continuous systems, see \textcolor{blue}{Appendix~B}.
We begin by modelling the dynamics of the input $\rho_{\text{in}}$ and the QN $\rho_{\text{qn}}$ for a time $\tau$, after which the observables of the QN are recorded.
The coherent part of the dynamics is described by the following Hamiltonian, written in a frame rotating with the pump frequency
\begin{equation}
\begin{aligned}
H&=\sum_{n = 1}^2 \hbar \Delta_n \hat a_n^{\dagger}\hat a_n + \sum_{m = 1}^M \hbar (\Lambda_m \hat b_m^{\dagger}\hat b_m+P_m (\hat b_m+\hat b_m^{\dagger})) \\
&+\sum_{\llbracket n,m\rrbracket} \hbar K_{nm} \mathcal{F}(\hat a_n,\hat b_m)+ \sum_{\llbracket m,m^{\prime}\rrbracket}\hbar J_{mm^{\prime}}\mathcal{F}(\hat b_m, \hat b_{m^{\prime}}),
\label{EQ_gen_h}
\end{aligned}
\end{equation}
where $\hat a_n$ ($\hat b_m$) denotes the annihilation operator for the $n$th input object ($m$th QN node).
The detunings of all local frequencies $\{\omega_n,\Omega_m\}$ with respect to the frequency of the pump $\Theta_p$ are denoted by $\Delta_n=\omega_n-\Theta_p$ and $\Lambda_m=\Omega_m-\Theta_p$.
The contact between the input and QN is represented by the couplings $K_{nm}$, whereas the interactions within the QN are denoted by $J_{mm^{\prime}}$.
For simplicity, we take the operator function to represent interactions that are ample in nature, i.e., $\mathcal{F}(\hat X,\hat Y)\equiv \hat X \hat Y^{\dagger}+\hat Y \hat X^{\dagger}$.
Each QN node may be coherently driven with strength $P_m$.
The bracket $\llbracket \cdot,\cdot \rrbracket$ denotes a particular configuration of the couplings, e.g., all-to-all.
We note that the simulation of the system can be made efficient when dealing with Gaussian states~\cite{adesso2014continuous}.
These tools are applicable as the generic dynamics we consider here preserves Gaussianity, i.e., it involves a Hamiltonian that is at most quadratic in operators (Eq.~(\ref{EQ_gen_h})) and Gaussian dissipative processes (see below).
In this case, complete description of the system is contained in a covariance matrix (CM) $\bm V$ with elements $V_{ij}\equiv \langle u_i u_j +u_j u_i \rangle/2-\langle u_i \rangle \langle u_j \rangle$, where the vector ${\bm u}\equiv [\hat q_1,\hat r_1,\hat q_2,\hat r_2,\hat x_1,\hat p_1,\cdots,\hat x_M,\hat p_M]^T$ is composed of dimensionless position and momentum quadratures (of the input and QN nodes, respectively) that are expressed as $\hat q_n=(\hat a_n+\hat a_n^{\dagger})/\sqrt{2}$, $\hat r_n=(\hat a_n-\hat a_n^{\dagger})/(i\sqrt{2})$, $\hat x_m=(\hat b_m+\hat b_m^{\dagger})/\sqrt{2}$, and $\hat p_m=(\hat b_m-\hat b_m^{\dagger})/(i\sqrt{2})$.
One can obtain the dynamics of the quadratures in the Heisenberg picture from the Hamiltonian of Eq.~(\ref{EQ_gen_h}), which with added noise terms gives rise to a set of Langevin equations (LEs) that can be written in a matrix form: $\dot {\bm{u}}(t) = {\bm A} {\bm{u}}(t) + {\bm{h}}(t)$.
The drift matrix ${\bm A}$ contains the parameters $\{\Delta_n,\Lambda_m,K_{nm},J_{mm^{\prime}},\gamma_n,\kappa_m\}$ and the vector ${\bm h}(t)$ incorporates the pump and noise terms, see \textcolor{blue}{Appendix~C} for details.
The noise terms are of uncoloured Gaussian type, and written as $\sqrt{2\gamma_n}\hat a_n^{\text{in}}$ and $\sqrt{2\kappa_m}\hat b_m^{\text{in}}$, where $\langle \hat a_n^{\text{in}}(t)\hat a_{n^{\prime}}^{\text{in},\dagger}(t^{\prime})\rangle=\delta_{nn^{\prime}}\delta(t-t^{\prime})$ and $\langle \hat b_m^{\text{in}}(t)\hat b_{m^{\prime}}^{\text{in},\dagger}(t^{\prime})\rangle=\delta_{mm^{\prime}}\delta(t-t^{\prime})$~\cite{walls2007quantum}.
The solution of the LEs is given by
\begin{equation}
{\bm u}(t)={\bm W}_+(t){\bm u}(0)+{\bm W}_+(t)\int_0^t dt^{\prime} {\bm W}_-(t^{\prime}){\bm h}(t^{\prime}),
\end{equation}
where ${\bm W}_{\pm}(t)=\exp{(\pm {\bm A}t)}$.
This further gives the dynamical equation for the CM: $\dot {\bm V}={\bm A}{\bm V}(t)+{\bm V}(t){\bm A}^T+{\bm D}$, where ${\bm D}=\mbox{diag}[\gamma_1,\gamma_1,\gamma_2,\gamma_2,\kappa_1,\kappa_1,\cdots,\kappa_M,\kappa_M]$.
The observables of the QN $\langle \hat O_{mk}\rangle=\mbox{tr}(\rho(\tau) \hat O_{mk})$ ($k$ labels observables from the same $m$th node) at time $\tau$ can be obtained from ${\bm u}(\tau)$ and ${\bm V}(\tau)$.
For more detailed expressions, see \textcolor{blue}{Appendix~C}.
Here we consider local observables, for simplicity.
We will see that it is sufficient to work with average occupation numbers (intensities) as the observables, although any additional variables that can be measured can further help.
The observables define an output layer upon which a training procedure is used to find a linear combination of the observables that will define the final system output.
The training is performed with ridge regression using a random set of input CMs $\{\bm V_{\text{in},l}\}_{l=1}^{N_{\text{tr}}}$ as follows.
Each of the input CMs will be in contact with the QN and produce a set of $N_{\text{ob}}$ observables at time $\tau$, recorded as a vector $\bm v_l$.
The observables are used to first estimate the input state (its \emph{unique} elements), from which entanglement is calculated.
In the present case, each element of the CM (labelled $s$) is estimated linearly as $f_s=\bm \beta_s [1;\bm v_l]$, where $\bm \beta_s=[\beta_0,\beta_1,\cdots, \beta_{N_{\text{ob}}}]$ contains the coefficients to be obtained with ridge regression.
In particular, $\bm \beta_s=(\bm X^T\bm X+\lambda \mathbb{1})^{-1}\bm X^T\bm Y_s$, where $\bm X=[1,\bm v_1^T;1,\bm v_2^T;\cdots ;1,\bm v_{N_{\text{tr}}}^T]$ contains all the observables in the training set, $\bm Y_s$ contains the target $s$th element, and $\lambda$ is the ridge parameter.
This allows us to obtain an estimated input CM $\tilde {\bm V}_{\text{in}}$ from the trained output layer $\{\bm \beta_s\}$, given measured QN observables.
Consequently, the estimated entanglement is computed using the logarithmic negativity $E=L_e(\tilde {\bm V}_{\text{in}})$~\cite{negativity}.
In what follows, we define the entanglement estimation error as
\begin{equation}\label{EQ_def_error}
\Delta_E=\sqrt{\sum_{l^{\prime} = 1}^{N_{\text{te}}} \frac{(E_{\text{est},l^{\prime}}-E_{\text{in},l^{\prime}})^2}{N_{\text{te}}}},
\end{equation}
where $N_{\text{te}}$ is the number of random input CMs in the testing set.
\section{Entanglement estimation}\label{S_ee}
Here we present the performance of entanglement estimation.
In simulations, the parameters are taken as random $\{\Delta_n,\Lambda_m,K_{nm},J_{mm^{\prime}},P_m,10\gamma_n,10\kappa_m\}\in [0,1]\Gamma$, where $\Gamma$ is an overall strength in units of frequency, and evolution time $\tau=\pi/2\Gamma$.
One set of random parameters will be taken to define one particular QN.
When assessing the performance of the scheme, we will average over different parameter choices, to provide a general assessment of the architecture rather than any specific parameter choice.
Indeed, one advantage of our scheme is that the considered systems do not need precise control of their parameters.
Figure~\ref{Fig2}(a) shows the entanglement estimation error against the number of QN nodes.
The sudden shift shown by the arrow indicates $\Delta_E\sim 10^{-10}$ is obtained for QNs having at least $4$ nodes.
This can be understood as follows.
Recall that the number of independent parameters required to fully characterise an $N$-mode Gaussian state is $2N(2N+1)/2$.
This suggests that to faithfully estimate the state of a two-mode Gaussian input, one requires at least $10$ observables from the QN.
This is fulfilled by having at least $4$ QN nodes as each node is itself in a Gaussian state and hence requires three independent real parameters (e.g., we take two diagonal and one off-diagonal entries from the local CM) to be determined.
The inset shows the entanglement profile of the input CM ${\bm V}_{\text{in}}$ used in both training and testing, with $N_{\text{tr}}=50$ and $N_{\text{te}}=100$, respectively.
See \textcolor{blue}{Appendix~D} for the generation of random input CMs.
A closer look at the comparison between the estimated and input entanglement during testing is plotted in Figs.~\ref{Fig2}(b) and (c) for the case where the QN is composed of $3$ and $4$ nodes, respectively.
It can be seen that the latter offers minute errors.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{Fig2}
\caption{Performance of entanglement sensing for generic dynamics.
(a) Estimation error vs number of nodes used in the QN.
The inset shows the profile of entanglement of the randomly generated input CMs in both training and testing.
Panels (b) and (c) present explicit comparison between the estimated and input entanglement during testing where the QN is composed of $3$ and $4$ nodes, respectively.
}
\label{Fig2}
\end{figure}
We have also simulated the case where we record one observable (the mean excitation $\langle \hat b_m^{\dagger}\hat b_m \rangle$) from each QN node.
In this case, one requires either the addition of two-photon pump, i.e., $\sum_m P_m^{\prime}(\hat b^2_m+\hat b_m^{\dagger2})$ with random strengths (relatively weaker) $P_m^{\prime}\in [0,1]\Gamma/10$ or the presence of ultra-strong coupling $\mathcal{F}(\hat b_m,\hat b_{m^{\prime}})=(\hat b_m+\hat b_m^{\dagger})(\hat b_{m^{\prime}} +\hat b_{m^{\prime}}^{\dagger})$.
The reason for this is that simpler interactions or drives in Eq.~(\ref{EQ_gen_h}) are not sufficient for complex information transfer during the dynamics, which would allow the mean excitation of the QN nodes to completely recover information regarding the input objects.
We found that the shift to low estimation error requires at least 10 QN nodes, again consistent with the number of independent parameters of the input CM, see \textcolor{blue}{Appendix~E} for details.
As the scheme estimates the CM of the input objects before computing entanglement, it opens up the possibility to use a training set consisting of separable input CMs without affecting its entanglement-testing capabilities.
We used the same setup as in Fig.~\ref{Fig2}(a) based on 4 QN nodes and performed training using only non-entangled input CMs.
The testing was performed with entangled input CMs, finding a profile similar to the inset in Fig.~\ref{Fig2}(a).
Indeed, the comparison between the estimated and input entanglement is similar to the one in Fig.~\ref{Fig2}(c) (see \textcolor{blue}{Appendix~F} for details).
We note that although each input CM in the training set is not entangled, they are still correlated.
Similarly, learning from separable input objects is also possible for discrete systems (cf. \textcolor{blue}{Appendix~F}).
\section{Scaling beyond the SQL}
For a more realistic model, we incorporate measurement errors of the observables from the QN.
The observables now read $\langle \hat O_{mk}\rangle \rightarrow \langle \hat O_{mk}\rangle+\epsilon_{mk}$, where $\{\epsilon_{mk}\}$ are generated from a normal distribution with zero mean and standard deviation $\zeta/2$.
In what follows, we take $\zeta=10^{-3}$.
Similar to Fig.~\ref{Fig2}(a), we present the estimation errors in Fig.~\ref{Fig3}(a).
The dots indicate the scaling of the error with respect to the number of nodes $M$.
From Fig.~\ref{Fig3}(a), one can see the signature of the shift previously observed in Fig.~\ref{Fig2}(a).
In particular, the scaling of the estimation error becomes clearer for $M\ge 4$ in Fig.~\ref{Fig3}(a).
It can be seen that $\Delta_E$ can exhibit scaling beyond the dashed curve, i.e., beyond the SQL $\propto 1/\sqrt{\mathcal{N}} = 1 / \sqrt{3M} \propto 1/\sqrt{M}$.
Another alternative to obtain independent observables from the QN is through time-multiplexing.
For instance, we consider a single observable from each QN node, i.e., the mean excitation $\langle \hat b_m^{\dagger}\hat b_m \rangle$ and measure it at $\mathcal{T}$ different times.
This gives a total of $N_{\text{ob}}=M\mathcal{T}$ observables.
We demonstrate the case for $\mathcal{T}=3$, i.e., at $\tau=\{1,2,3\} \pi/2\Gamma$ in Fig.~\ref{Fig3}(b).
In this case, we have added random two-photon pump $P_m^{\prime}\in [0,1]\Gamma/10$ (see \textcolor{blue}{Appendix~E} for the case with ultra-strong coupling).
One can see similar scaling as in panel (a).
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{Fig3}
\caption{Scaling beyond the SQL.
(a) Estimation error plotted against the number of nodes $M$.
Panel (b), taking only mean excitation from each QN node as the observable with time-multiplexing performed $\mathcal{T}=3$ times.
The scaling for the SQL is given by the dashed curve in each panel.
}
\label{Fig3}
\end{figure}
\section{Gravity-induced entanglement}
We present an application of the entanglement sensing scheme to estimate gravity-induced entanglement (GIE) generated between masses.
Consider two identical spherical objects, each with mass $m$, trapped in a 1D harmonic potential.
This configuration has been theoretically predicted to generate entanglement between the masses through gravitational interactions~\cite{krisnanda2020observable}.
Here, each mass is probed by a cavity mode, see Fig.~\ref{Fig4}(a).
The probes are turned on by the pump on the respective cavities, $\mathcal{E}_a$ (left) and $\mathcal{E}_b$ (right).
This way, the observables from the cavity modes can be processed through an output layer, which then produces an estimate of the GIE.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{Fig4}
\caption{Sensing gravity-induced entanglement (GIE) between masses.
(a) The setup of two trapped masses interacting gravitationally.
After a time $\tau_0$, two cavity modes are turned on, whose observables are processed to estimate the GIE.
(b) The estimated GIE for different initial evolution of the masses $\tau_0$ vs testing instances. One observes that the standard deviation $\delta_E<10^{-4}$.
The parameters used in simulations are $m=1$~kg made of Osmium with density $22.59$~g/cm$^3$, $\omega=0.1$~Hz with $\omega/\gamma \gg 1$, $r_0=1.73$, $\tau=1$~$\mu$s, $L_{a(b)}=25$~mm, laser wavelength $1064$~nm, $P_{a(b)}=50$~mW, $\{\kappa_{a},\kappa_b,\Delta_{a},\Delta_{b}\}=2.36\times 10^5$~Hz, $L \simeq 2R$ with $R$ the radius of the mass, and $\zeta/2=10^{-2}$.
See \textcolor{blue}{Appendix~G} for references of the parameters.}
\label{Fig4}
\end{figure}
First, we consider the dynamics without the probes, in which the Hamiltonian reads
\begin{equation}
H_0=\frac{\hbar \omega}{2}(\hat p_A^2+\hat x_A^2+\hat p_B^2+\hat x_B^2)-\frac{\hbar Gm}{\omega L^3}(\hat x_A-\hat x_B)^2,\label{EQ_H_2m}
\end{equation}
where $\hat x_{A(B)}$ denotes the dimensionless displacement of mass $A(B)$, $\omega$ the frequency of the trapping potentials, and $L$ the equilibrium distance between the masses.
We have used $\hat x_{A(B)}= x_{A(B)}\sqrt{m\omega/\hbar}$ and $\hat p_{A(B)}=p_{A(B)}/\sqrt{\hbar m\omega}$, where $x_{A(B)}$ and $p_{A(B)}$ are the displacement and momentum operators, respectively.
The gravitational interaction is expanded from $-Gm^2/(L-(\hat x_A-\hat x_B)\sqrt{\hbar/m\omega})$ up to a quadratic term, $(\hat x_A-\hat x_B)^2$, which is necessary for entanglement generation as it contains non-local coupling $\propto \hat x_A\hat x_B$ acting on both masses.
We have neglected the constant and linear term $\propto(\hat x_A-\hat x_B)$ as the former is simply an energy offset and the latter a bi-local operator (cannot create entanglement) that constitutes to shifting the equilibrium position of the masses.
One can construct a set of LEs from Eq.~(\ref{EQ_H_2m}) with the addition of damping $\gamma$ and Brownian-like noises $\hat \xi_{A(B)}$ affecting the masses (see \textcolor{blue}{Appendix~G} for details).
As we deal with Gaussianity-preserving dynamics, we use the tools for CV systems.
This includes the description of the system within a CM and its evolution to $\bm V(\tau_0)$ from which properties of the system can be calculated (see \textcolor{blue}{Appendix~G}).
At time $\tau_0$, the probes are turned on, where the Hamiltonian (in a rotating frame with the frequency of the lasers) now reads
\begin{equation}
\begin{aligned}
H&=H_0+\hbar \Delta_{0a} \hat a^{\dagger}\hat a+\hbar \Delta_{0b} \hat b^{\dagger}\hat b+i\hbar \mathcal{E}_a(\hat a^{\dagger}-\hat a) \\
&+i\hbar \mathcal{E}_b(\hat b^{\dagger}-\hat b)-\hbar G_{0a}\hat a^{\dagger}\hat a \hat x_A+\hbar G_{0b}\hat b^{\dagger}\hat b \hat x_B,\label{EQ_H_tot}
\end{aligned}
\end{equation}
where $\hat j = \hat a, \hat b$ denotes the annihilation operator of the left and right cavity mode, $\Delta_{0j}=\omega_{j}-\omega_{lj}$ the cavity-laser detuning, $\mathcal{E}_{j}=\sqrt{2P_j\kappa_j/\hbar \omega_{lj}}$ the driving strength of the cavity, $P_j$ the laser power with frequency $\omega_{lj}$, $\kappa_j=\pi c/2F_jL_j$ the cavity decay rate with finesse $F_j$ and length $L_j$, $G_{0j}=(\omega_j/L_j)\sqrt{\hbar/m \omega}$ the optomechanical coupling strength.
From Eq.~(\ref{EQ_H_tot}), one can construct a set of linearised LEs, which are then used to evolve the CM $\bm V(\tau_0)$ to $\bm V(\tau_0+\tau)$ at which the observables from the cavity modes are recorded.
In what follows, we take into account the features shown previously for entanglement sensing using generic systems.
As the task is estimating entanglement of a two-mode CM (of the masses), at least 10 observables are required for recording.
This is taken from 10 independent CM elements of the joint cavity modes.
From the central limit theorem it follows that $\zeta/2 \propto 1/\sqrt{N_{\text {rep}}}$, where $N_{\text {rep}}$ is the number of repetitions that an element is measured.
To make a comparison with entanglement measurement in Ref.~\cite{palomaki2013entangling} whereby $N_{\text {rep}}=10^4$, we shall assume error statistics with $\zeta \sim 2/\sqrt{N_{\text {rep}}} = 2\times 10^{-2}$.
As the initial CM at $t=0$, we use squeezed (local) thermal state for the masses $\mbox{diag}[e^{2r_0},e^{-2r_0},e^{2r_0},e^{-2r_0}](1+2\bar n)/2$ with $r_0$ being the squeezing strength and $\bar n$ the mean thermal phonon number, and vacuum for the cavity modes.
The training is performed using random separable input states $\bm V_{\text{in}}(\tau_0)$, which are generated using random $\bar n>0$.
This is such that entanglement does not yet grow for initial thermal states within $\tau_0$.
On the other hand, testing is performed with $\bar n=0$.
For better precision, one can use time-multiplexing during the dynamics with the probes at $\{1,2,\cdots,\mathcal{T}\}\tau$.
We present the estimated GIE for different initial accumulation time $\tau_0$ in Fig.~\ref{Fig4}(b).
We have taken $\mathcal{T}=4$ (see \textcolor{blue}{Appendix~H} for the scaling of standard deviation $\delta_E$ against $\mathcal{T}$), $N_{\text{tr}}=50$, and $N_{\text{te}}=100$.
The standard deviations of the GIE in Fig.~\ref{Fig4}(b) follow $\delta_E<10^{-4}$, which is two orders of magnitude better than the experimentally achieved $\sim 10^{-2}$ in Ref.~\cite{palomaki2013entangling}.
We also computed the estimated GIE from direct measurements, which is done by adding measurement errors directly to the elements of $\bm V(\tau_0)$.
In this case, standard deviation $\delta_E\sim 10^{-4}$ is only possible if the system permits three orders of magnitude weaker measurement error strength $\zeta/2 =10^{-5}$.
This demonstrates the efficiency of our method, which requires less number of single-shot measurements $N_{\text{rep}}$ to obtain precision comparable to measurements directly on the masses, i.e., with noisy $\bm V(\tau_0)$.
Figure~\ref{Fig4}(b) shows that our method is able to estimate GIE efficiently for $\tau_0=0.5$~s.
We note that this is shorter than the coherence times resulting from thermal photons from environment and collisions with air molecules (both in the range of about $5$~s) if the experiments were conducted on Earth with liquid Helium in ultrahigh vacuum~\cite{krisnanda2020observable}.
\section{DISCUSSION}
We have shown that a simple neural network (quantum reservoir processor) can be used for efficient estimation of quantum entanglement. Our main motivation for development of such a method is provided by present efforts to design experiments capable of detection of gravity-induced entanglement.
The introduced method shows that the entanglement precision can be improved by two orders of magnitude from what was achieved in Ref.~\cite{palomaki2013entangling}.
The entanglement sensing step is crucial for masses initialised in natural Gaussian states
and any improvement on it relaxes other requirements of the setup.
The most direct one is the requirement on coherence times: since smaller values of entanglement become detectable, the system can be measured earlier.
With entanglement estimation accuracy on the order $10^{-4}$ detection of GIE could be performed within decoherence times available on Earth, whereas accuracy $10^{-2}$ would rather require an experiment in space~\cite{krisnanda2020observable}.
Moreover, in order to understand how other experimental parameters can be changed, let us recall that the \emph{figure of merit} for entanglement generated via gravity between trapped masses $m$ separated by a distance $L$ is given by $2 G m / \omega^2 L^3$, where $\omega$ characterises the trapping potential or spread of the initial wave function of each mass~\cite{krisnanda2020observable}.
Therefore, better entanglement precision also translates to smaller masses in the experiment that could be placed further apart.
The method presented in this paper also holds potential for other settings where one estimates entanglement of the easily accessed probes with precision advantage and reveals quantumness of a macroscopic mediating object.
In particular, this includes an extension of Refs.~\cite{lambert2013quantum,scholes2017using,collini2010coherently,panitchayangkoon2010long} towards showing quantum properties of photosynthetic bacteria~\cite{krisnanda2018probing} or that of a macroscopic mechanical membrane in the membrane-in-the-middle optomechanics setting~\cite{aspelmeyer2014cavity,paternostro2007creating,krisnanda2017revealing}.
Additionally, we note that our scheme can work not only for CV or discrete systems, but also hybrid configurations such as discrete systems as input and CV systems as the QN or vice versa (see \textcolor{blue}{Appendix~B}).
\begin{acknowledgements}
We thank Sanjib Ghosh, Kevin Dini, and Yvonne Gao for stimulating discussion.
T.K. and T.C.H.L. acknowledge the support by the Singapore Ministry of Education under its AcRF Tier 2 grant MOE2019-T2-1-004.
T.P. is supported by the Polish National Agency for Academic Exchange NAWA Project No. PPN/PPO/2018/1/00007/U/00001. MP acknowledges the support by the European Union's Horizon 2020 FET-Open project TEQ (766900), the Leverhulme Trust Research Project Grant UltraQuTe (grant RGP-2018-266), the Royal Society Wolfson Fellowship (RSWF/R3/183013), the UK EPSRC (EP/T028424/1), and the Department for the Economy Northern Ireland under the US-Ireland R\&D Partnership Programme (USI 175).
\end{acknowledgements}
\vspace{0.1cm}
\noindent{\bf Author contributions:} TK, TP, and TCHL conceived the initial project direction; TK carried out all the calculations and derivations under the supervision of TP, MP, and TCHL; MP and TP assisted in designing the setup for gravity-induced entanglement; TK wrote the paper with contributions from TP, MP, and TCHL. All authors discussed the results and revised the paper.
\vspace{0.1cm}
\noindent{\bf Competing interests:} The authors declare no competing interests.
\vspace{0.1cm}
\noindent {\bf Data availability:} All data needed to evaluate the conclusions in the paper are present in the paper and/or the \textcolor{blue}{Appendix}.
\clearpage
| 8861d526d9514eac46de8930bdf87d733b7dbac5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{introduction} | c18fe8c9635405da2e1c054a31b4bb79569f0fce | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:introduction}
The Weibull distribution is one of the most important probability distributions in analysis of lifetime data. The probability density function and cumulative density function of a Weibull random variable $T$ with shape parameter $k>0$ and scale parameter $\lambda>0$ are
\begin{align}
\label{eqn:y:complete}
p_T(t|k, \lambda) = \left(\frac{k}{\lambda^k}\right) t^{k-1} \exp\left(-\left(\frac{t}{\lambda}\right)^k\right), \quad F_T(t | k, \lambda) = 1 - \exp\left(-\left(\frac{t}{\lambda}\right)^k\right),
\end{align}
respectively. With lifetime data, we often do not observe complete data and instead have joint realisations of the random variables $(Y = y, \Delta = \delta)$ where
\begin{eqnarray}
\label{eqn:y:censored}
Y &=& \min (T, C) \\
\label{eqn:Delta}
\Delta_i &=& {\rm I}(T \leq C) =
\begin{cases}
1, & \text{if } T \leq C \; ({\rm observed\; survival})\\
0, & \text{if } T > C \; ({\rm observed\; censoring})
\end{cases}
\end{eqnarray}
where the random variables $T$ and $C$ are assumed to be independent and denote the survival time and the censoring time of an item, respectively. In words, we observe the survival time $T=t$ of an item if it is less than the corresponding censoring time $C = c$ (i.e., $t \leq c$) ; otherwise, we only know that the item survived past time $c$ (i.e., $t > c$). The censoring time may be a fixed constant (say, $c$) or a random variable that may depend on other factors. Given $n$ i.i.d. lifetime data points ${\bf y} = (y_1,\ldots,y_n)$ with or without censoring, an important problem is to estimate the unknown parameters $k$ and $\lambda$ and thus learn about the survival distribution of the items.
The most common approach to parameter estimation is the method of maximum likelihood, where the unknown parameters are set to values that maximise the (log-) likelihood of the observed data. Unfortunately, in the case of the Weibull shape parameter the corresponding maximum likelihood estimate is known to have large bias with both complete and censored data (see, for example, \cite{Ross94, MackisackStillman96, Hirose99}) and this is especially evident for small sample sizes and/or under large amounts of censoring. This manuscript introduces the Bayesian minimum message length (MML) approach to inductive inference and demonstrates how MML can be used to estimate Weibull parameters in both the complete and censored data setting. We show that with an appropriate choice of prior distributions the MML estimate of the shape parameter improves on the maximum likelihood estimate, given censored or complete data, and is competitive with alternative proposals that modify the maximum likelihood estimate to reduce bias. Furthermore, we demonstrate how the MML principle can be used to discriminate between the lognormal and Weibull distributions with censored data. Empirical experiments suggest that model selection with MML is an excellent alternative to commonly used information criteria such as the Bayesian information criterion.
\section{Minimum message length}
\label{sec:mml}
The minimum message length (MML) principle~\cite{WallaceBoulton68, WallaceFreeman87, WallaceDowe99a, Wallace05} is a Bayesian information-theoretic framework for inductive inference that provides a new, unified approach to parameter estimation and model selection. Given data, the key step in applying MML is the computation of the length of a message that describes (encodes) the data, with the assumption that the message comprises two components:
\begin{enumerate}
\item the \emph{assertion}, encoding of the structure of the model, including all model parameters $\bm{\theta} \in \bm{\Theta} \in \mathbb{R}^p$; and
\item the \emph{detail}, encoding the data $D$ using the model $p(D | \bm{\theta})$ from the assertion.
\end{enumerate}
The length of the assertion measures the complexity of the model, with simpler models having a shorter assertion compared to more complex models. The length of the detail measures how well the model named in the assertion fits the data; more complex models will have shorter detail lengths compared to simpler models. The length of the combined two-part message, $I(D, \bm{\theta})$, is
\begin{equation}
\label{eqn:mml:codelength}
I(D, \bm{\theta}) = \underbrace{ I(\bm{\theta}) }_{\rm assertion} + \underbrace{I(D | \bm{\theta})}_{\rm detail}
\end{equation}
i.e., the sum of the length of the assertion, $I(\bm{\theta})$, and the length of detail, $I(D|\bm{\theta})$. Inference in the MML framework proceeds by finding the model
\begin{equation}
\hat{\bm{\theta}}(D) = \argmin_{\bm{\theta} \in \bm{\Theta}} \left\{ I(D, \bm{\theta}) \right\}
\end{equation}
that minimises the length of the two-part message message. Minimising the MML codelength requires balancing complexity of a model (assertion) with the corresponding fit to the data (detail) with the preferred model being the simplest model that fits the data sufficiently well. A key advantage of MML is that the unit of measurement, the codelength (generally measured in $\log_e$ digits, called nits or nats), is universal in the sense that allows inference and comparison of models with different model structures (e.g., linear regression vs. decision tree) and parameters within a single, unified framework.
Precise computation of codelengths is known to be a NP-hard proble in general. As such, there exist many MML approximations to the codelength (\ref{eqn:mml:codelength})~\cite{WallaceBoulton75,Wallace05}, with the MML87 approximation~\cite{WallaceFreeman87,Wallace05} being the most widely applied due to it's relative computational simplicity. Under suitable regularity conditions, the MML87 codelength approximates (\ref{eqn:mml:codelength}) by
\begin{equation}
\label{eqn:mml87:codelength}
I_{87}(D, \bm{\theta}) = \underbrace{-\log \pi(\bm{\theta}) + \frac{1}{2} \log \abs{J_{\bm{\theta}}(\bm{\theta})} + \frac{p}{2} \log \kappa_p}_{\rm assertion} + \underbrace{\frac{p}{2} - \log p(D|\bm{\theta})}_{\rm detail}
\end{equation}
where $\pi_{\bm{\theta}}(\bm{\theta})$ is the prior distribution of the parameters $\bm{\theta}$, $\abs{J_{\bm{\theta}}(\bm{\theta})}$ is the determinant of the expected Fisher information matrix, $p(D|\bm{\theta})$ is the likelihood function of the model and $\kappa_p$ is a quantization constant~\cite{ConwaySloane98,AgrellEriksson98}; for small $p$ we have
\begin{equation}
\kappa_1 = \frac{1}{12}, \quad \kappa_2 = \frac{5}{36 \sqrt{3}}, \quad \kappa_3 = \frac{19}{192 \times 2^{1/3}},
\end{equation}
while, for moderate to large $p$, $\kappa_p$ is well-approximated by~\cite{Wallace05}:
\begin{equation}
\frac{p}{2} (\log \kappa_p + 1) \approx -\frac{p}{2} \log 2\pi + \frac{1}{2} \log p \pi - \gamma,
\end{equation}
where $\gamma \approx 0.5772$ is the Euler--Mascheroni constant. The MML87 approximation is invariant under smooth one-to-one reparametarizations of the likelihood function and is asymptotically equivalent to the Bayesian information criterion (BIC)~\cite{Schwarz78} as $n \to \infty$ with $p>0$ fixed; that is,
\begin{equation}
\label{eqn:mml87:bic}
I_{87}(D, \bm{\theta}) = - \log p(D|\bm{\theta}) + \frac{p}{2} + O(1)
\end{equation}
where the $O(1)$ term depends on the prior distribution, the Fisher information and the number of parameters $p$.
There exist many successful applications of the MML principle in statistics and machine learning literature, including factor analysis~\cite{WallaceFreeman92}, time series~\cite{Schmidt08a,SchmidtMakalic16a}, linear causal models~\cite{WallaceKorb99} and mixture models~\cite{WallaceDowe00,SchmidtMakalic12}), among others.
\section{Complete data}
\subsection{Maximum likelihood estimates}
\label{sec:complete:mle}
Consider first the setting of complete data with no censoring. The negative log-likelihood of data ${\bf y} = (y_1,\ldots,y_n)$ is
\begin{equation}
\label{eqn:complete:nll}
- \log p_T({\bf y} | k, \lambda) = n \log \left(\frac{\lambda^k}{k}\right) - (k-1) \left(\sum_{i=1}^n \log y_i\right) + \sum_{i=1}^n \left(\frac{y_i}{\lambda}\right)^k
\end{equation}
The maximum likelihood (ML) estimates of $k,\lambda$ are
\begin{equation}
\hat{\lambda}^{k}({\bf y}) = \frac{1}{n} \sum_{i=1}^n y_i^{k},
\end{equation}
where $\hat{k}({\bf y})$ is defined implicitly by
\begin{equation}
\label{eqn:mle:kscore}
\frac{n}{k} + \sum_{i=1}^n \log y_i - \frac{n \sum_i y_i^k \log y_i}{\sum_i y_i^k} = 0
\end{equation}
and must be solved for numerically. While the ML estimate of $\lambda$ is reasonable, the ML estimate of $k$ is known to exhibit large bias and perform poorly in terms of squared error risk, especially for small sample sizes~\cite{MackisackStillman96}.
Several attempts have been made to construct a modified ML estimate with improved performance. Ross~\cite{Ross94} derives the simple adjustment formula
\begin{equation}
\label{eqn:mle:kross}
\hat{k}_{\rm R}({\bf y}) = \left(\frac{n-2}{n-0.68}\right) \hat{k}_{\rm ML}({\bf y})
\end{equation}
for the ML estimate that reduces the bias to typically better than about $0.05$\%, though this adjustment applies to complete data only. Similarly, Hirose~\cite{Hirose99} derives tables with correction coefficients that can be used to obtain modified ML estimates of both $k$ and $\lambda$ with reduced bias. In a somewhat different approach, Yang and Xie~\cite{YangXie03} apply the modified profile likelihood proposed by Cox and Reid~\cite{CoxReid87,CoxReid92} to derive a penalized maximum likelihood estimate of $k$. Specifically, the Yang and Xie estimate of $\lambda$ is equivalent to the ML estimate while the new estimate of the shape parameter $k$ is obtained by numerically solving
\begin{equation}
\label{eqn:ml:yangxie}
\frac{n-2}{k} + \sum_{i=1}^n \log y_i - \frac{n \sum_i y_i^k \log y_i}{\sum_i y_i^k} = 0.
\end{equation}
which is similar to (\ref{eqn:mle:kscore}), the only difference being $(n-2)$ in the numerator of the first term. Yang and Xie empirically show that their estimate of $k$ is less biased than the ML estimate and is more efficient than the simple modification (\ref{eqn:mle:kross}) proposed by Ross. In the next section, we show how to derive an MML estimate of the Weibull distribution parameters and demonstrate that the Yang and Xie modified maximum likelihood estimate is an MML87 estimate for a particular prior distribution.
\subsection{Minimum message length estimates}
\label{sec:complete:mml}
To derive the MML87 codelength~(\ref{eqn:mml87:codelength}) we require the determinant of the expected Fisher information matrix
\begin{equation}
| J(k, \lambda) | = \frac{n^2 \pi^2}{6 \lambda ^2},
\end{equation}
and prior distributions for both parameters. Assuming that $k$ and $\lambda$ are independent a priori, we opt for the half-Cauchy distributions
\begin{equation}
\label{eqn:mmlprior:complete}
\pi(k, \lambda) = \pi(k) \pi(\lambda), \quad \pi(k) = \frac{2}{\pi (1 + k^2)}, \quad \pi(\lambda) = \frac{2}{\pi (1 + \lambda^2)}.
\end{equation}
As $\lambda$ is a scale parameter, a heavy tailed distribution like the half-Cauchy is appropriate and recommended in, for example, \cite{PolsonScott12}. Additionally, the half-Cauchy distribution is suitable for the shape parameter $k$ as $k=1$ denotes a fixed (constant) failure rate and decreasing ($k < 1$) and increasing ($k > 1$) failure rate are assumed equally likely a priori; that is,
\begin{equation}
\int_0^1 \pi(k) dk = \int_1^\infty \pi(k) dk = \frac{1}{2} .
\end{equation}
The complete MML87 codelength for the Weibull distribution is
\begin{equation}
\label{eqn:mml87:weibull:codelength}
I_{87}(D, k, \lambda) = -\log \left(\frac{4}{\pi^2 (1+k^2)(1+\lambda^2)} \right) + \frac{1}{2} \log \left( \frac{n^2 \pi^2}{6 \lambda ^2} \right) - \log p_T({\bf y} | k, \lambda) + 1 + \log \kappa_2
\end{equation}
where the negative log-likelihood function $- \log p_T({\bf y} | k, \lambda)$ is given in (\ref{eqn:complete:nll}) and $\kappa_2 = 5/(36 \sqrt{3})$ (see Section~\ref{sec:mml}). Unfortunately, with this selection of prior distributions, the MML87 estimates of $k$ and $\lambda$ must be obtained by numerically minimising (\ref{eqn:mml87:weibull:codelength}).
It is straightforward to see that the modified maximum likelihood estimate of Yang and Xie (\ref{eqn:ml:yangxie}) is the MML87 estimate obtained under the prior distribution
\begin{equation}
\pi(k, \lambda) = \pi(k) \pi(\lambda), \quad \pi(k) \propto \frac{1}{k^2}, \quad \pi(\lambda) \propto \frac{1}{\lambda}.
\end{equation}
which is improper unless lower and upper bound limits are imposed on both the shape and scale parameters. The implied prior distribution for $\lambda$ is the usual scale invariant distribution often used to model a scale parameter while the prior distribution for the shape parameter $k$ is heavy tailed and Cauchy-like asymptotically. As the aforementioned implied prior distributions are similar to (\ref{eqn:mmlprior:complete}) in their behaviour, it is expected that both the Yang and Xie modified maximum likelihood estimate and the MML87 estimate proposed in this manuscript will yield similar parameter estimates with virtually identical properties.
\section{Censored data}
\label{sec:censored}
We now examine inference of the Weibull distribution in the presence of Type I fixed as well as random censoring. Consider first the fixed censoring setup where observations are censored after some period of time $c > 0$. In particular, we observe the lifetime of an item only if $T_i \leq c$, otherwise we observe the censoring time $c$. The likelihood function of $n$ observed data points $D = \{(y_1, \delta_1), \ldots, (y_n, \delta_n)\}$ is
\begin{equation}
p(D) = \prod_{i=1}^n p_{T}(y_i)^{\delta_i} (1 - F_{T}(y_i))^{1 -\delta_i}
\end{equation}
where $\delta_i = 1$ if the survival time is observed, and $\delta_i = 0$ if the censoring time is observed (see Section~\ref{sec:introduction}).
In contrast, under random censoring, both the lifetime $T_i$ and the censoring time $C_i$ are assumed to be mutually independent random variables. Here, the likelihood function of $n$ observed data points $D = \{(y_1, \delta_1), \ldots, (y_n, \delta_n)\}$ can be written as
\begin{equation*}
p(D) = \left(\prod_{i=1}^n p_{T}(y_i)^{\delta_i} (1 - F_{T}(y_i))^{1 -\delta_i}\right) \left( \prod_{i=1}^n p_C(y_i)^{1-\delta_i} (1 - F_{C}(y_i))^{\delta_i} \right)
\end{equation*}
where $p_T(t|\theta)$ and $F_T(t|\theta)$ denote the probability density and the cumulative density function of the random variable $T$, respectively. We assume the random censoring setup examined in \cite{DanishAslam12}, where both $T_i$ and $C_i$ are Weibull random variables
\begin{equation}
\label{eqn:exponential}
T_i \sim {\rm Weibull}(\theta, \beta), \quad C_i \sim {\rm Weibull}(\theta, \alpha), \quad i = 1,\ldots,n,
\end{equation}
where $\alpha,\beta>0$ are the scale parameters and $\theta > 0$ is the common shape parameter.
The joint probability density function of $Y_i = {\rm min}(T_i, C_i)$ and $\Delta_i = I(T_i < C_i)$ is
\begin{equation}
p_{Y,\Delta}(y, \delta | \alpha, \beta, \theta) = \left( \frac{\theta}{\alpha^\theta} \right) \left( \frac{\alpha}{\beta} \right)^{\delta_i \theta} y^{\theta-1} \exp\left(- \left(\frac{1}{\alpha ^{\theta }}+\frac{1}{\beta ^{\theta }}\right) y^{\theta }\right) .
\end{equation}
Next we derive maximum likelihood estimates for the Weibull distribution under type I and random censoring.
\subsection{Maximum likelihood estimates}
\label{sec:censored:mle}
Consider the type I censoring setup as described in Section~\ref{sec:censored}. The likelihood of $n$ data points $D = \{(y_1, \delta_1), \ldots, (y_n, \delta_n)\}$ is
\begin{align}
p(D)
&= \left(\frac{k}{\lambda^k}\right)^d \exp\left(-\frac{1}{\lambda^k} \sum_{i=1}^n y_i^k \right) \prod_{i=1}^n y_i^{\delta_i (k-1)}
\label{eqn:censored:nll}
\end{align}
The maximum likelihood (ML) estimates of $k,\lambda$ are
\begin{equation}
\hat{\lambda}^{k}({\bf y}) = \frac{1}{d} \sum_{i=1}^n y_i^{k},
\end{equation}
where $d = \sum_{i=1}^n \delta_i$ and $\hat{k}({\bf y})$ is given implicitly by
\begin{equation}
\label{eqn:mle:censored:kscore}
\frac{d}{k} + \sum_{i=1}^n \delta_i \log y_i - \frac{d \sum_i y_i^k \log y_i}{\sum_i y_i^k} = 0 \, .
\end{equation}
The maximum likelihood estimate of $k$ is known to exhibit large bias in small samples and when the proportion of censoring is high. Sirvanci and Yang~\cite{SirvanciYang84} propose the alternative estimate
\begin{equation}
\hat{k}^{-1}(D) = \frac{1}{d g(d / n)} \sum_{i=1}^n \delta_i (\log c - \log y_i) ,
\end{equation}
where the function $g(\cdot)$ given by
\begin{equation}
g(p) = \log \log (1 - p)^{-1} - \frac{1}{p} \int_0^p \log \log (1-t)^{-1} \, dt .
\end{equation}
is a bias correction factor for the bias in estimating $1/k$. Sirvanci and Yang derive finite sample properties of this estimate and show that it has high relative efficiency in estimating $1/k$ over a range of censoring levels (10\% -- 90\% censoring) provided $0 < d < n$. Using the same strategy as in the complete data case (see Section~\ref{sec:complete:mle}), Yang and Xie~\cite{YangXie03} propose a new modified maximum likelihood estimate of the shape parameter $k$ that is obtained by solving
\begin{equation}
\label{eqn:ml:yangxie:censored}
\frac{d-1}{k} + \sum_{i=1}^n \delta_i \log y_i - \frac{d \sum_i y_i^k \log y_i}{\sum_i y_i^k} = 0.
\end{equation}
However, this modified profile score function requires $d>1$ to yield a positive estimate for $k$.
Next, we examine the random censoring setup described in Section~\ref{sec:censored}. The likelihood of the data under the random censoring model is
\begin{equation}
p_D(D | \alpha, \beta, \theta) = \left(\frac{\theta }{\alpha^{\theta }}\right)^n \left(\frac{\alpha }{\beta }\right)^{d \theta } \exp\left(- \left(\frac{1}{\alpha ^{\theta }}+\frac{1}{\beta ^{\theta }}\right) \sum_{i=1}^n y_i^{\theta }\right) \prod_{i=1}^n y_i^{\theta - 1}
\end{equation}
where, as before, $d = \sum_{i=1}^n \delta_i$. From this, the maximum likelihood estimates of $(\alpha,\beta)$ are
\begin{equation}
\hat{\alpha}_{\rm ML} = \left(\frac{\sum_{i=1}^n y_i^\theta}{n-d}\right)^{1/\theta }, \quad \hat{\beta}_{\rm ML} = \left(\frac{\sum_{i=1}^n y_i^\theta}{d}\right)^{1/\theta }
\end{equation}
while the maximum likelihood estimate of $\theta$ must be obtained by numerical optimisation.
Clearly, the maximum likelihood estimates $(\hat{\alpha}_{\rm ML}, \hat{\beta}_{\rm ML})$ exist only if $d \in (0, n)$. Alternatively, maximum likelihood estimates may be obtained by noting the following.
\begin{thm}
\label{thm:jointpdf}
The joint probability density function of $(Y_i, \Delta_i)$ can be written as
\begin{equation}
p_{Y,\Delta}(y, \delta | \alpha, \beta, \theta) = p_{\Delta}(\delta | \phi) \, p_{Y}(y | k, \lambda),
\end{equation}
where $\Delta \sim {\rm binom}(n, \phi)$ and $Y \sim {\rm Weibull} (k, \lambda)$ and
\begin{equation}
\phi = P(T \leq C) = \frac{\alpha^\theta}{\alpha^\theta+\beta^\theta}, \quad k = \theta, \quad \lambda = \frac{\beta}{(1 + (\beta/\alpha)^\theta)^{1/\theta}}.
\end{equation}
\end{thm}
The proof is straightforward and is omitted. By Lemma~\ref{thm:jointpdf} and invariance of the maximum likelihood estimate, the maximum likelihood estimates of $\alpha,\beta$ and $\theta$ can also be obtained from the usual maximum likelihood estimates for the binomial and Weibull distributions
\begin{equation}
\hat{\phi}_{\rm ML} = \frac{1}{n} \sum_{i=1}^n \delta_i, \quad \hat{\lambda}^{\hat{k}}_{\rm ML} = \frac{1}{n} \sum_{i=1}^n y_i^{\hat{k}},
\end{equation}
where $\hat{k}$ is given implicitly by
\begin{equation}
\frac{1}{n} \sum_{i=1}^n \log y_i + \frac{1}{k} - \frac{\sum_i y_i^k \log y_i}{\sum_i y_i^k} = 0
\end{equation}
and by noting that
\begin{equation}
\label{eqn:param:transform}
\theta = k, \quad \alpha = \lambda (1 - \phi)^{-1/\theta}, \quad \beta = \lambda \phi^{-1/\theta}.
\end{equation}
These estimates exist only if $\phi_{\rm ML} \in (0, 1)$ or, equivalently, $d \in (0, n)$.
\subsection{Minimum message length estimates}
\label{sec:censored:mml}
We consider first MML inference under the type I censoring setup described in Section~\ref{sec:censored}. Let
\begin{equation}
p = F_T(c | k, \lambda), \quad q = -\log(1 - p), \quad E_1(z) = \int_1^\infty \exp(-t z) / t \, dt,
\end{equation}
where $E_1(\cdot)$ is the usual exponential integral function. As with the complete data setting, we assume independent half-Cauchy prior distributions (see (\ref{eqn:mmlprior:complete})) for both the shape and the scale parameters. The expected Fisher information matrix with type I censoring is
\begin{equation*}
J(k,\lambda) = n
\left(
\begin{array}{cc}
J_{k,k} & J_{k,\lambda} \\
J_{k,\lambda} & J_{\lambda,\lambda}
\end{array}
\right)
\end{equation*}
\begin{align*}
J_{\lambda,\lambda} &= p \left(\frac{k}{\lambda}\right)^2, \quad
J_{k,\lambda} = \frac{E_1(q)-p-(p-1) \log(q)+\gamma }{\lambda } \\
J_{k,k}
&= \frac{(\log (q)+1) (p+(p-2) \log (q)-2 E_1(q)-2 \gamma ) + 2 q \, _3F_3(1,1,1;2,2,2;-q)}{k^2} ,
\end{align*}
where $_3F_3(\cdot;\cdot;\cdot)$ is the generalized hypergeometric function and $\gamma \approx 0.5772$ is the Euler--Mascheroni constant. The determinant of the expected Fisher information matrix
\begin{equation}
| J(k,\lambda) | = \left(\frac{n}{\lambda}\right)^2 \left[ 2 p q \, _3F_3(1,1,1;2,2,2;-q)-(\gamma+\log (q)-\text{li}(1-p))^2 \right],
\end{equation}
where $\text{li}(\cdot)$ is the logarithmic integral function, is clearly a complicated function of the probability of no censoring, $p$. The second term approaches
\begin{eqnarray}
\lim_{p \to 1} \left[ 2 p q \, _3F_3(1,1,1;2,2,2;-q)-(\gamma+\log (q)-\text{li}(1-p))^2 \right] = \frac{\pi^2}{6}
\end{eqnarray}
as $p$ gets closer to 1, and is otherwise always less than ($\pi^2/6$). A simple approximation to the log-determinant using a rational function of $p$ is
\begin{equation}
\log | J(k,\lambda) | \approx \frac{-89.8213 p^3+194.713 p^2-87.8331 p-10.7758}{-24.0356 p^3+7.79886 p^2+28.0711 p+1} + \log \left(\frac{n}{\lambda}\right)^2
\end{equation}
with the absolute approximation error of the order $0.005$ for all $0.1 \leq p \leq 0.9$; a higher order rational approximation can be used if smaller approximation error is desired over the entire range of $p \in (0,1)$. The MML87 codelength for the Weibull distribution with type I censoring is
\begin{equation}
\label{eqn:mml87:censored:codelength}
I_{87}(D, \bm{\theta}) = -\log \left(\frac{4}{\pi^2 (1+k^2)(1+\lambda^2)} \right) + \frac{1}{2} \log | J(k,\lambda) | - \log p_T({\bf y} | k, \lambda) + 1 + \log \kappa_2
\end{equation}
where the negative log-likelihood function $- \log p_T({\bf y} | k, \lambda)$ is given in (\ref{eqn:censored:nll}) and $\kappa_2 = 5/(36 \sqrt{3})$ (see Section~\ref{sec:mml}). As with the complete data case the MML87 estimates of $k$ and $\lambda$ must be obtained by numerically minimising (\ref{eqn:mml87:censored:codelength}).
Consider next the random censoring setup described in Section~\ref{sec:censored} where the lifetime $T_i$ and the censoring time $C_i$ are mutually independent Weibull random variables with a common shape parameter. From Lemma~\ref{thm:jointpdf}, the joint density of $(Y_i, \Delta_i)$ can be written as a product of a binomial distribution $\Delta \sim (n, \phi)$ and Weibull distribution $Y | \Delta \sim \text{Weibull}(k, \lambda)$. This implies that an MML code for the data $D$ could comprise two messages with the first message encoding the binary censoring indicators $\bm{\delta} = (\delta_1,\ldots,\delta_n)$, followed by another message that encodes the lifetimes ${\bf y} = (y_1,\ldots,y_n)$ given the censoring data $\bm{\delta}$. With this encoding, the total MML codelength for the data $D = \{(y_1, \delta_1), \ldots, (y_n, \delta_n)\}$ is
\begin{equation}
\label{eqn:cond:codelength}
I_{87}(D, \alpha, \beta, \theta) = I_{87}(\bm{\delta}, \phi) + I_{87}({\bf y}, k, \lambda |\bm{\delta}),
\end{equation}
where $\phi$ is the probability of observing an uncensored datum. As with maximum likelihood, MML87 is invariant under one-to-one parameter transformations implying that MML87 estimates of $(\alpha, \beta, \theta)$ can be obtained from MML87 estimates of $(\phi, k, \lambda)$ using the relations (\ref{eqn:param:transform}).
The MML87 codelength of the binomial distribution was derived in, for example, \cite{Wallace05,WallaceDowe00} and, for a uniform prior distribution on $\phi$, is given by
\begin{equation}
\label{eqn:Idelta}
I_{87}(\bm{\delta}, \phi) = -\left(k + \frac{1}{2}\right) \log \phi - \left(n + \frac{1}{2} - k\right) \log (1-\phi) + \frac{1}{2}(1 + \log (n/12) )
\end{equation}
where, as before, $k = (\sum_i \delta_i)$. The minimum of the codelength is at the MML87 estimate
\begin{equation}
\hat{\phi}_{87}(\bm{\delta}) = \frac{k + 1/2}{n + 1} .
\end{equation}
The conditional codelength of the surivival times ${\bf y}$ given the censoring indicators $\bm{\delta}$, $I_{87}({\bf y}, k, \lambda | \bm{\delta})$ is simply the MML87 codelength for the Weibull distribution discussed in Section~\ref{sec:complete:mml}. Note that it is of course possible to derive the MML87 joint codelength and construct a single message for the data $D$, similar to the complete data case discussed in Section~\ref{sec:complete:mml}. Due to the invariance of the MML87 codelength, both approaches will yield exactly the same inferences.
\section{Experiments}
\label{sec:experiments}
Numerical experiments were performed to measure the performance of the newly proposed MML87 estimates compared to the maximum likelihood estimate and the modified maximum likelihood estimate of Yang and Xie~\cite{YangXie03} with complete (see Section~\ref{sec:experiments:complete}) and type I censored data (see Section~\ref{sec:experiments:censored}).
\begin{table*}[tb]
\scriptsize
\begin{center}
\begin{tabular}{ccccccccc}
\toprule
$n$ & $k$ & \multicolumn{3}{c}{Bias} & & \multicolumn{3}{c}{Mean Squared Error} \\
& & MLE & MMLE & MML87 & ~ & MLE & MMLE & MML87 \\
\cmidrule{1-9}
\multirow{4}{*}{10} & 0.5 & 0.085 & {\bf 0.008} & 0.063 & & 0.038 & {\bf 0.023} & 0.029\\
& 1.0 & 0.168 & {\bf 0.015} & 0.085 & & 0.152 & {\bf 0.094} & 0.099\\
& 5.0 & 0.850 & {\bf 0.085} & 0.117 & & 3.836 & 2.352 & {\bf 2.336}\\
& 10.0 & 1.692 & {\bf 0.164} & 0.181 & & 14.973 & 9.143 & {\bf 9.124}\\
\vspace{-2mm} \\
\cmidrule{2-9}
\vspace{-2mm} \\
\multirow{4}{*}{20} & 0.5 & 0.038 & {\bf 0.004} & 0.030 & & 0.012 & {\bf 0.009} & 0.011\\
& 1.0 & 0.076 & {\bf 0.008} & 0.040 & & 0.048 & {\bf 0.037} & 0.038\\
& 5.0 & 0.371 & {\bf 0.031} & 0.045 & & 1.194 & 0.927 & {\bf 0.923}\\
& 10.0 & 0.774 & {\bf 0.093} & 0.100 & & 4.881 & 3.761 & {\bf 3.757}\\
\vspace{-2mm} \\
\cmidrule{2-9}
\vspace{-2mm} \\
\multirow{4}{*}{50} & 0.5 & 0.015 & {\bf 0.002} & 0.012 & & 0.004 & {\bf 0.003} & 0.003\\
& 1.0 & 0.029 & {\bf 0.004} & 0.016 & & 0.015 & {\bf 0.013} & 0.014\\
& 5.0 & 0.143 & {\bf 0.016} & 0.021 & & 0.366 & 0.329 & {\bf 0.328}\\
& 10.0 & 0.279 & {\bf 0.025} & 0.028 & & 1.456 & 1.311 & {\bf 1.310}\\
\vspace{-3mm} \\
\bottomrule
\vspace{+1mm}
\end{tabular}
\caption{Bias and mean squared error for maximum likelihood (MLE), modified maximum likelihood (MMLE) and MML87 estimates of $k$ computed over $10^5$ simulations runs with $\lambda = 1$.\label{tab:results:complete}}
\end{center}
\end{table*}
\subsection{Complete data}
\label{sec:experiments:complete}
The MML87 estimate of the shape parameter $k$ derived in Section~\ref{sec:complete:mml} is now compared to the maximum likelihood (MLE) estimate (\ref{eqn:mle:kscore}) and the modified maximum likelihood (MMLE) estimate (\ref{eqn:ml:yangxie}) using simulated data. In each simulation run, $n$ data points were generated from the model Weibull$(k, \lambda = 1)$ where $n = \{10, 20, 50\}$ and the shape parameter was set to $k \in \{0.5, 1, 5, 10\}$. Given the data, MLE, MMLE and MML87 estimates were computed and compared in terms of bias and mean squared error. For each value of $(k, n)$ $10^5$ simulations were performed and the average bias and mean squared error results are shown in Table~\ref{tab:results:complete} for each estimate.
It is clear that the MMLE and MML87 estimates improve significantly on the maximum likelihood estimate in terms of both bias and mean squared error for each tested value of $(n,k)$. We further note that the MMLE estimate of $k$ is slightly less biased than the proposed MML87 estimate, though the two estimates are virtually indistinguishable in terms of the average mean squared error. As discussed in Section~\ref{sec:complete:mml}, the MMLE estimate is a special case of the MML87 estimator for a particular choice of the prior distribution with complete data, and it is therefore expected that the two estimates will have similar behaviour.
\subsection{Censored data}
\label{sec:experiments:censored}
We also compared the MML87 estimate (see Section~\ref{sec:censored:mml}) to the maximum likelihood estimate (MLE) (\ref{eqn:mle:censored:kscore}) and the modified maximum likelihood estimate (MMLE) (\ref{eqn:ml:yangxie:censored}) under type I censored data. The experimental setup was identical to that for complete data with the following changes: (i) the proportion of uncensored observations was set to $p \in \{0.3, 0.5, 0.7, 0.9\}$, and (ii) $n \in \{20,30,40\}$ data points were generated during each simulation run. We restricted the experiments to exclude data sets where the number of uncensored observations $d (=\sum_i \delta_i) < 2$, as the MLE and MMLE estimates are not defined for small $d$. In addition to the bias and the mean squared error in estimating the shape parameter, we computed the Kullback--Leibler (KL) divergence~\cite{KullbackLeibler51} between the data generating model and each estimated model (see Appendix~A). The results averaged over $10^5$ simulations runs for each combination of $(n,p,k)$ are shown in Table~\ref{tab:results:censored}.
We again observe that the MLE estimate of $k$ is strongly biased particularly for small $k$ and $p$. While the MMLE is less biased than the proposed MML87 estimate, the MML87 estimate achieves smaller mean squared error and smaller KL divergence compared to the MMLE in all experiments. Additionally, we observe that the KL divergence for the MMLE model is similar to the MLE model, despite the significant reduction in bias of estimating the shape parameter $k$ achieved by the MMLE. Clearly the proposed MML87 estimate is an improvement over the MLE and highly competitive against estimators that are primarily designed to reduce bias in the MLE, such as the one proposed by Yang and Xie~\cite{YangXie03}.
\begin{table*}[tbph]
\scriptsize
\begin{center}
\begin{tabular}{cccccccccccccccccc}
\toprule
$n$ & $p$ & $k$ & \multicolumn{3}{c}{Bias} & & \multicolumn{3}{c}{Mean Squared Error} & & \multicolumn{3}{c}{KL Divergence} \\
& & & MLE & MMLE & MML87 & ~ & MLE & MMLE & MML87 & ~ & MLE & MMLE & MML87 \\
\cmidrule{1-14}
\multirow{16}{*}{20} & \multirow{4}{*}{ 0.3} & 0.5 & 0.114 & {\bf 0.002} & 0.070 & & 0.158 & 0.077 & {\bf 0.040} & & 0.069 & 0.060 & {\bf 0.042}\\
& & 1.0 & 0.055 & {\bf 0.005} & 0.038 & & 0.042 & 0.031 & {\bf 0.026} & & 0.060 & 0.056 & {\bf 0.043}\\
& & 5.0 & 0.037 & {\bf 0.006} & 0.021 & & 0.020 & 0.017 & {\bf 0.016} & & 0.057 & 0.054 & {\bf 0.044}\\
& & 10.0 & 0.019 & {\bf -0.001} & 0.006 & & 0.011 & 0.010 & {\bf 0.010} & & 0.048 & 0.046 & {\bf 0.039}\\
& \multirow{4}{*}{ 0.5} & 0.5 & 0.114 & {\bf 0.002} & 0.070 & & 0.158 & 0.077 & {\bf 0.040} & & 0.069 & 0.060 & {\bf 0.042}\\
& & 1.0 & 0.055 & {\bf 0.005} & 0.038 & & 0.042 & 0.031 & {\bf 0.026} & & 0.060 & 0.056 & {\bf 0.043}\\
& & 5.0 & 0.037 & {\bf 0.006} & 0.021 & & 0.020 & 0.017 & {\bf 0.016} & & 0.057 & 0.054 & {\bf 0.044}\\
& & 10.0 & 0.019 & {\bf -0.001} & 0.006 & & 0.011 & 0.010 & {\bf 0.010} & & 0.048 & 0.046 & {\bf 0.039}\\
& \multirow{4}{*}{ 0.7} & 0.5 & 0.114 & {\bf 0.002} & 0.070 & & 0.158 & 0.077 & {\bf 0.040} & & 0.069 & 0.060 & {\bf 0.042}\\
& & 1.0 & 0.055 & {\bf 0.005} & 0.038 & & 0.042 & 0.031 & {\bf 0.026} & & 0.060 & 0.056 & {\bf 0.043}\\
& & 5.0 & 0.037 & {\bf 0.006} & 0.021 & & 0.020 & 0.017 & {\bf 0.016} & & 0.057 & 0.054 & {\bf 0.044}\\
& & 10.0 & 0.019 & {\bf -0.001} & 0.006 & & 0.011 & 0.010 & {\bf 0.010} & & 0.048 & 0.046 & {\bf 0.039}\\
& \multirow{4}{*}{ 0.9} & 0.5 & 0.114 & {\bf 0.002} & 0.070 & & 0.158 & 0.077 & {\bf 0.040} & & 0.069 & 0.060 & {\bf 0.042}\\
& & 1.0 & 0.055 & {\bf 0.005} & 0.038 & & 0.042 & 0.031 & {\bf 0.026} & & 0.060 & 0.056 & {\bf 0.043}\\
& & 5.0 & 0.037 & {\bf 0.006} & 0.021 & & 0.020 & 0.017 & {\bf 0.016} & & 0.057 & 0.054 & {\bf 0.044}\\
& & 10.0 & 0.019 & {\bf -0.001} & 0.006 & & 0.011 & 0.010 & {\bf 0.010} & & 0.048 & 0.046 & {\bf 0.039}\\
\vspace{-2mm} \\
\cmidrule{2-14}
\vspace{-2mm} \\
\multirow{16}{*}{30} & \multirow{4}{*}{ 0.3} & 0.5 & 0.067 & {\bf 0.002} & 0.048 & & 0.059 & 0.039 & {\bf 0.026} & & 0.042 & 0.039 & {\bf 0.028}\\
& & 1.0 & 0.035 & {\bf 0.003} & 0.024 & & 0.020 & 0.017 & {\bf 0.016} & & 0.038 & 0.036 & {\bf 0.029}\\
& & 5.0 & 0.023 & {\bf 0.004} & 0.013 & & 0.012 & 0.010 & {\bf 0.010} & & 0.036 & 0.035 & {\bf 0.030}\\
& & 10.0 & 0.016 & {\bf 0.003} & 0.008 & & 0.007 & 0.007 & {\bf 0.007} & & 0.034 & 0.033 & {\bf 0.029}\\
& \multirow{4}{*}{ 0.5} & 0.5 & 0.067 & {\bf 0.002} & 0.048 & & 0.059 & 0.039 & {\bf 0.026} & & 0.042 & 0.039 & {\bf 0.028}\\
& & 1.0 & 0.035 & {\bf 0.003} & 0.024 & & 0.020 & 0.017 & {\bf 0.016} & & 0.038 & 0.036 & {\bf 0.029}\\
& & 5.0 & 0.023 & {\bf 0.004} & 0.013 & & 0.012 & 0.010 & {\bf 0.010} & & 0.036 & 0.035 & {\bf 0.030}\\
& & 10.0 & 0.016 & {\bf 0.003} & 0.008 & & 0.007 & 0.007 & {\bf 0.007} & & 0.034 & 0.033 & {\bf 0.029}\\
& \multirow{4}{*}{ 0.7} & 0.5 & 0.067 & {\bf 0.002} & 0.048 & & 0.059 & 0.039 & {\bf 0.026} & & 0.042 & 0.039 & {\bf 0.028}\\
& & 1.0 & 0.035 & {\bf 0.003} & 0.024 & & 0.020 & 0.017 & {\bf 0.016} & & 0.038 & 0.036 & {\bf 0.029}\\
& & 5.0 & 0.023 & {\bf 0.004} & 0.013 & & 0.012 & 0.010 & {\bf 0.010} & & 0.036 & 0.035 & {\bf 0.030}\\
& & 10.0 & 0.016 & {\bf 0.003} & 0.008 & & 0.007 & 0.007 & {\bf 0.007} & & 0.034 & 0.033 & {\bf 0.029}\\
& \multirow{4}{*}{ 0.9} & 0.5 & 0.067 & {\bf 0.002} & 0.048 & & 0.059 & 0.039 & {\bf 0.026} & & 0.042 & 0.039 & {\bf 0.028}\\
& & 1.0 & 0.035 & {\bf 0.003} & 0.024 & & 0.020 & 0.017 & {\bf 0.016} & & 0.038 & 0.036 & {\bf 0.029}\\
& & 5.0 & 0.023 & {\bf 0.004} & 0.013 & & 0.012 & 0.010 & {\bf 0.010} & & 0.036 & 0.035 & {\bf 0.030}\\
& & 10.0 & 0.016 & {\bf 0.003} & 0.008 & & 0.007 & 0.007 & {\bf 0.007} & & 0.034 & 0.033 & {\bf 0.029}\\
\vspace{-2mm} \\
\cmidrule{2-14}
\vspace{-2mm} \\
\multirow{16}{*}{40} & \multirow{4}{*}{ 0.3} & 0.5 & 0.047 & {\bf 0.002} & 0.035 & & 0.033 & 0.025 & {\bf 0.018} & & 0.029 & 0.028 & {\bf 0.021}\\
& & 1.0 & 0.025 & {\bf 0.002} & 0.017 & & 0.014 & 0.012 & {\bf 0.011} & & 0.027 & 0.026 & {\bf 0.022}\\
& & 5.0 & 0.017 & {\bf 0.003} & 0.009 & & 0.008 & 0.008 & {\bf 0.007} & & 0.027 & 0.026 & {\bf 0.023}\\
& & 10.0 & 0.014 & {\bf 0.004} & 0.008 & & 0.005 & 0.005 & {\bf 0.005} & & 0.026 & 0.026 & {\bf 0.023}\\
& \multirow{4}{*}{ 0.5} & 0.5 & 0.047 & {\bf 0.002} & 0.035 & & 0.033 & 0.025 & {\bf 0.018} & & 0.029 & 0.028 & {\bf 0.021}\\
& & 1.0 & 0.025 & {\bf 0.002} & 0.017 & & 0.014 & 0.012 & {\bf 0.011} & & 0.027 & 0.026 & {\bf 0.022}\\
& & 5.0 & 0.017 & {\bf 0.003} & 0.009 & & 0.008 & 0.008 & {\bf 0.007} & & 0.027 & 0.026 & {\bf 0.023}\\
& & 10.0 & 0.014 & {\bf 0.004} & 0.008 & & 0.005 & 0.005 & {\bf 0.005} & & 0.026 & 0.026 & {\bf 0.023}\\
& \multirow{4}{*}{ 0.7} & 0.5 & 0.047 & {\bf 0.002} & 0.035 & & 0.033 & 0.025 & {\bf 0.018} & & 0.029 & 0.028 & {\bf 0.021}\\
& & 1.0 & 0.025 & {\bf 0.002} & 0.017 & & 0.014 & 0.012 & {\bf 0.011} & & 0.027 & 0.026 & {\bf 0.022}\\
& & 5.0 & 0.017 & {\bf 0.003} & 0.009 & & 0.008 & 0.008 & {\bf 0.007} & & 0.027 & 0.026 & {\bf 0.023}\\
& & 10.0 & 0.014 & {\bf 0.004} & 0.008 & & 0.005 & 0.005 & {\bf 0.005} & & 0.026 & 0.026 & {\bf 0.023}\\
& \multirow{4}{*}{ 0.9} & 0.5 & 0.047 & {\bf 0.002} & 0.035 & & 0.033 & 0.025 & {\bf 0.018} & & 0.029 & 0.028 & {\bf 0.021}\\
& & 1.0 & 0.025 & {\bf 0.002} & 0.017 & & 0.014 & 0.012 & {\bf 0.011} & & 0.027 & 0.026 & {\bf 0.022}\\
& & 5.0 & 0.017 & {\bf 0.003} & 0.009 & & 0.008 & 0.008 & {\bf 0.007} & & 0.027 & 0.026 & {\bf 0.023}\\
& & 10.0 & 0.014 & {\bf 0.004} & 0.008 & & 0.005 & 0.005 & {\bf 0.005} & & 0.026 & 0.026 & {\bf 0.023}\\
\vspace{-3mm} \\
\bottomrule
\vspace{+1mm}
\end{tabular}
\caption{Bias, mean squared error and Kullback--Leibler (KL) divergence for maximum likelihood (MLE), modified maximum likelihood (MMLE) and MML87 estimates of $k$ computed over $10^5$ simulations runs with $\lambda = 1$; $p$ denotes the proportion of uncensored observations. \label{tab:results:censored}}
\end{center}
\end{table*}
\subsection{Model selection}
Recall that the minimum message length principle unifies parameter estimation and model selection within the same framework. In this section, we demonstrate how MML can be used to infer whether observed data was generated by a Weibull distribution or by a lognormal distribution~\cite{SiswadiQuesenberry82,UpadhyayPeshwani03,KimYum08}. To use MML to discriminate between competing models, we need to compute the codelength of each model. For complete data with no censoring, the probability density function of the lognormal distribution with mean $\mu \in \mathbb{R}$ and standard deviation $\sigma > 0$ is
\begin{equation}
p(y | \mu, \sigma) = \frac{1}{\sqrt{2 \pi } \sigma y} \exp\left(-\frac{(\log (y)-\mu )^2}{2 \sigma ^2}\right) .
\end{equation}
The negative log-likelihood for data ${\bf y}$ is
\begin{equation}
\label{eqn:nll:logn:complete}
-\log p_T({\bf y} | \mu, \sigma) = \frac{n}{2} \log (2\pi) + n \log \sigma + \sum_{i=1}^n \log y_i + \frac{1}{2\sigma^2} \sum_{i=1}^n (\log y_i - \mu)^2 .
\end{equation}
The determinant of the expected Fisher information for the lognormal model is well-known
\begin{equation}
\label{eqn:fisher:logn:complete}
|J(\mu,\sigma)| = \frac{2 n^2}{\sigma ^4} .
\end{equation}
Similar to Section~\ref{sec:complete:mml}, we select heavy-tailed prior distributions for both parameters
\begin{equation}
\label{eqn:priors:logn}
\pi(\mu,\sigma) = \pi(\mu) \pi(\sigma), \quad \pi(\mu) = \frac{1}{\pi (1 + \mu^2)}, \quad \pi(\sigma) = \frac{2}{\pi (1 + \sigma^2)}.
\end{equation}
Substituting (\ref{eqn:nll:logn:complete}), (\ref{eqn:fisher:logn:complete}) and (\ref{eqn:priors:logn}) into (\ref{eqn:mml87:codelength}) yields the MML87 codelength for the lognormal distribution. Due to this choice of prior distributions, the MML87 estimates of $\mu$ and $\sigma$ must be obtained numerically. To determine whether observed data follows the Weibull or the lognormal distribution, we compute the codelength of the data under each model and select the model with the smallest codelength.
An experiment was setup to compare the MML87 model selection performance against the commonly used Bayesian information criterion (BIC)~\cite{Schwarz78} and the scale transformation maximal invariant statistic (SI)~\cite{QuesenberryKent82}. A comparison of BIC and SI on discriminating between the Weibull and lognormal models was examined in~\cite{KimYum08}. Similar to the experimental setup in~\cite{KimYum08}, we generated $n \in (10,25,50,100,200)$ data points from either the Weibull(1,1) or the Lognormal(1,1) model, as both BIC and SI are invariant under scale and shape transformations. Each method was then asked to select the best fitting model for the observed data and the experiment was repeated for $10^5$ iterations. The performance of each method was measured in terms of probability of correct selection and the results are shown in Table~\ref{tab:results:mdl:complete}.
While all three methods tested performed similarly for medium to large sample sizes, the average accuracy of MML87 is significantly higher compared to BIC and SI under small sample sizes. Additionally, while the SI statistic tended to favour the lognormal distribution, no such preference was observed for MML or BIC.
\begin{table*}[tb]
\scriptsize
\begin{center}
\begin{tabular}{cccccccccccc}
\toprule
$n$ & \multicolumn{3}{c}{${\bf y} \sim$ Weibull} & & \multicolumn{3}{c}{${\bf y} \sim$ Lognormal} & & \multicolumn{3}{c}{Average accuracy}\\
& MML87 & BIC & SI & ~ & MML87 & BIC & SI & ~ & MML87 & BIC & SI \\
\cmidrule{1-12}
10 & {\bf 0.738} & 0.677 & 0.596 & & 0.714 & 0.663 & {\bf 0.742} & & {\bf 0.726} & 0.670 & 0.669\\
25 & {\bf 0.838} & 0.807 & 0.783 & & 0.826 & 0.803 & {\bf 0.828} & & {\bf 0.832} & 0.805 & 0.806\\
50 & {\bf 0.917} & 0.904 & 0.894 & & {\bf 0.913} & 0.904 & 0.913 & & {\bf 0.915} & 0.904 & 0.904\\
100 & {\bf 0.975} & 0.972 & 0.970 & & {\bf 0.973} & 0.971 & 0.973 & & {\bf 0.974} & 0.971 & 0.971\\
200 & {\bf 0.997} & 0.997 & 0.997 & & {\bf 0.998} & 0.997 & 0.998 & & {\bf 0.997} & 0.997 & 0.997\\
\vspace{-3mm} \\
\bottomrule
\vspace{+1mm}
\end{tabular}
\caption{Probability of correctly selecting the data generating model for MML87, BIC and SI computed over $10^5$ simulation runs with complete data only. \label{tab:results:mdl:complete}}
\end{center}
\end{table*}
We can of course use the MML principle to discriminate between the Weibull and lognormal distributions based on type I censored data. In this case, the negative log-likelihood function of the data is
\begin{eqnarray}
-\log p(D|\mu,\sigma) &=& d \log \sigma + \frac{d}{2} \log (2\pi) + \sum_{i=1}^n \delta_i \log(y_i) + \frac{1}{2\sigma^2} \sum_{i=1}^n \delta_i (\log(y_i) - \mu)^2 \nonumber \\
&&- (n-d)\log\left(1 - \Phi\left(\frac{\log(c) - \mu}{\sigma}\right)\right)~\label{eqn:nll:logn:typeI}
\end{eqnarray}
where $\Phi(\cdot)$ is cumulative density function of the standard normal distribution. Let
\begin{eqnarray}
z = \left(\frac{\log(c) - \mu}{\sigma}\right), \quad p = \Phi(z), \quad M = \Phi^{-1}(p),
\end{eqnarray}
where $z$ is the standardised censoring point and $\Phi^{-1}(\cdot)$ is the inverse cumulative density function of the standard normal distribution. The expected Fisher information matrix is
\begin{equation*}
J(\mu,\sigma) = \frac{n}{\sigma^2}
\left(
\begin{array}{cc}
\frac{e^{-M^2}}{2 \pi (1 - p)}-\frac{e^{-\frac{M^2}{2}} M}{\sqrt{2 \pi }}+p & \frac{e^{-M^2} M}{2 \pi(1 - p)}-\frac{e^{-\frac{M^2}{2}} \left(M^2+1\right)}{\sqrt{2 \pi }} \\
\frac{e^{-M^2} M}{2 \pi (1 - p)}-\frac{e^{-\frac{M^2}{2}} \left(M^2+1\right)}{\sqrt{2 \pi }} & \frac{e^{-M^2} M^2}{2 \pi(1 -p)}-\frac{e^{-\frac{M^2}{2}} \left(M^3+M\right)}{\sqrt{2 \pi }}+2 p \\
\end{array}
\right)
\end{equation*}
with determinant $|J(\mu,\sigma)|$ given by
\begin{equation}
\frac{2 n^2}{\sigma^4} \left[\frac{e^{-M^2} \left(M^2 (1-2 p)-3 p+1\right)}{4 \pi (p-1)}+\frac{e^{-\frac{1}{2} \left(3 M^2\right)} M}{4\sqrt{2} \pi ^{3/2} (1- p)}-\frac{e^{-\frac{M^2}{2}} M \left(M^2+3\right) p}{2\sqrt{2 \pi }}+ p^2\right].
\label{eqn:fisher:logn:typeI}
\end{equation}
which is equal to the determinant of the expected Fisher information matrix for complete data multiplied by a correction factor that takes into account the proportion of censoring. We use the same prior distributions for the parameters as in the case of complete data.
The MML87 codelength for the lognormal distribution with type I censoring is obtained by substituting (\ref{eqn:nll:logn:typeI}), (\ref{eqn:fisher:logn:typeI}) and (\ref{eqn:priors:logn}) into (\ref{eqn:mml87:codelength}) yields. As with the case of complete data, the MML87 estimates of $\mu$ and $\sigma$ must be obtained by numerical optimisation. We repeated the same model selection experiment as performed with complete data but this time varied the proportion of censoring from $10\%$ to $75\%$, similar to~\cite{KimYum08}. The performance of each method was measured in terms of probability of correct selection and the results are shown in Table~\ref{tab:results:mdl:typeI}.
As shown in~\cite{KimYum08}, BIC performs better than SI, with the latter always preferring the Weibull distribution for large amounts of data. In terms of model selection accuracy, it is clear that the proposed MML87 method is superior to both BIC and SI, especially with small sample sizes or large amounts of censoring. Lastly, we note that MML codelengths derived in this paper can also be used in more complex applications such as mixture models and decision trees; for example, we may use the Weibull distribution to model data in the terminal nodes of a tree or the attributes of a class in a finite mixture model.
\begin{table*}[btp]
\scriptsize
\begin{center}
\begin{tabular}{ccccccccccccccccc}
\toprule
$n$ & $p$ & \multicolumn{3}{c}{${\bf y} \sim \text{Weibull}$} & & \multicolumn{3}{c}{${\bf y} \sim \text{Lognormal}$} & & \multicolumn{3}{c}{Average accuracy} \\
& & MML87 & BIC & SI & ~ & MML87 & BIC & SI & ~ & MML87 & BIC & SI \\
\multirow{4}{*}{25} & 0.10 & {\bf 0.723} & 0.685 & 0.645 & & 0.853 & 0.828 & {\bf 0.854} & & {\bf 0.788} & 0.756 & 0.749\\
& 0.30 & {\bf 0.613} & 0.563 & 0.494 & & 0.842 & 0.808 & {\bf 0.856} & & {\bf 0.728} & 0.686 & 0.675\\
& 0.50 & {\bf 0.536} & 0.456 & 0.357 & & 0.833 & 0.795 & {\bf 0.866} & & {\bf 0.685} & 0.625 & 0.612\\
& 0.75 & {\bf 0.603} & 0.289 & 0.146 & & 0.768 & 0.818 & {\bf 0.928} & & {\bf 0.685} & 0.554 & 0.537\\
\vspace{-2mm} \\
\cmidrule{2-13}
\vspace{-2mm} \\
\multirow{4}{*}{50} & 0.10 & {\bf 0.830} & 0.810 & 0.790 & & {\bf 0.911} & 0.897 & 0.910 & & {\bf 0.870} & 0.853 & 0.850\\
& 0.30 & {\bf 0.727} & 0.696 & 0.655 & & 0.875 & 0.852 & {\bf 0.880} & & {\bf 0.801} & 0.774 & 0.768\\
& 0.50 & {\bf 0.629} & 0.582 & 0.532 & & {\bf 0.844} & 0.810 & 0.841 & & {\bf 0.736} & 0.696 & 0.686\\
& 0.75 & {\bf 0.589} & 0.409 & 0.419 & & {\bf 0.822} & 0.785 & 0.744 & & {\bf 0.706} & 0.597 & 0.582\\
\vspace{-2mm} \\
\cmidrule{2-13}
\vspace{-2mm} \\
\multirow{4}{*}{100} & 0.10 & {\bf 0.925} & 0.917 & 0.909 & & {\bf 0.963} & 0.958 & 0.963 & & {\bf 0.944} & 0.937 & 0.936\\
& 0.30 & {\bf 0.839} & 0.822 & 0.812 & & {\bf 0.919} & 0.907 & 0.914 & & {\bf 0.879} & 0.865 & 0.863\\
& 0.50 & 0.736 & 0.709 & {\bf 0.864} & & {\bf 0.876} & 0.854 & 0.521 & & {\bf 0.806} & 0.781 & 0.692\\
& 0.75 & 0.625 & 0.525 & {\bf 0.954} & & {\bf 0.837} & 0.787 & 0.061 & & {\bf 0.731} & 0.656 & 0.508\\
\vspace{-2mm} \\
\cmidrule{2-13}
\vspace{-2mm} \\
\multirow{4}{*}{200} & 0.10 & {\bf 0.983} & 0.981 & 0.979 & & {\bf 0.992} & 0.991 & 0.992 & & {\bf 0.987} & 0.986 & 0.986\\
& 0.30 & 0.935 & 0.929 & {\bf 0.964} & & {\bf 0.968} & 0.964 & 0.896 & & {\bf 0.951} & 0.946 & 0.930\\
& 0.50 & 0.847 & 0.834 & {\bf 0.998} & & {\bf 0.922} & 0.910 & 0.024 & & {\bf 0.884} & 0.872 & 0.511\\
& 0.75 & 0.695 & 0.642 & {\bf 1.000} & & {\bf 0.854} & 0.815 & 0.000 & & {\bf 0.775} & 0.729 & 0.500\\
\bottomrule
\vspace{+1mm}
\end{tabular}
\caption{Probability of correctly selecting the data generating model for MML87, BIC and SI computed over $10^5$ simulation runs with type I censored data. The probability of censoring ($p$) is given in the second column.\label{tab:results:mdl:typeI}}
\end{center}
\end{table*}
| 8364e9073497fb8c6637d3cc00863aafbca7fbcc | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Dense prediction refers to the process of predicting the label for each point in a point cloud. It is widely known that dense prediction plays a pivotal role in 3D robotic perception and autonomy, enabling an array of tasks such as semantic segmentation, depth completion, and scene flow estimation.
\begin{figure}[htbp]
\centering
\includegraphics[width=3.3in]{./figures/qua_scannet.pdf}
\caption{In a dense prediction task, i.e., 3D semantic segmentation, we note
the segmentation prediction (top), segmentation error (middle) and dense uncertainty map (bottom, estimated by \sysname+) of two scenes from ScanNet validation split. Incorrect predictions tend to have high uncertainties.}
\label{fig_qua_scannet}
\end{figure}
UNet \cite{ronneberger2015u} based network has been the de-facto choice of today's point neural network architecture design for cloud dense prediction \cite{choy2019fully, ao2021spinnet, thomas2019kpconv}. In a UNet-like network, one can observe that the input and the output of two correspondingly linked layers have the same number of points, e.g., if the input point cloud is denoted by a $N \times 3$ tensor, then the output of its correspondingly linked layer is a $N \times D$ tensor. In this regard, the output can also be viewed as an embedding map, and a dense prediction network can then be decomposed as an embedding learning network and a task-specific regressor (or classifier).
Therefore,
the heart of the dense prediction task is embedding learning.
Embedding learning aims to learn a discriminative embedding model that pulls samples of the same class closer and pushes those of different classes away from each other in the embedding space. Successful embedding learning empowers many downstream tasks, including image retrieval \cite{musgrave2020metric}, face recognition \cite{meng2021magface} and zero-shot learning \cite{bucher2016improving}. In addition to improving the embedding model's discriminative capability, quantifying its uncertainty is also attracting much attention.
For dense prediction tasks of point clouds, it is desirable that an uncertainty level could be provided in conjunction with the point-wise labels to make its downstream decision-making more information-aware. Consider a scenario where an autonomous vehicle is predicting semantic labels of each point on the road, a prediction with an estimated uncertainty level would be helpful for the computer to decide when to trust its prediction and moreover, utilize the uncertainty to optimize the vehicle's planning and control. Such promising benefits have stimulated the development of various uncertainty estimation methods for different dense prediction tasks. In 3D semantic segmentation tasks, for example, the
popular approaches
include (1) using the output of the logit layer to calculate softmax entropy~\cite{czolbe2021segmentation}, (2) building a two-head network to predict the mean and variance of an embedding separately \cite{kendall2017uncertainties}, and (3) resorting to a BNN model and approximating posterior weights with MCD \cite{qi2021neighborhood}.
However, two major issues remain in existing uncertainty estimation methods for dense prediction of 3D point clouds.
First, points can only interact in the limited
receptive field of convolution kernels, and they need a shared MLP to realize an implicit interaction among logits (see Fig. \ref{fig_pipeline2}.a).
Such under-treatment of cross-point dependencies, unfortunately, often results in sub-optimal uncertainty estimation as evidenced by~\cite{monteiro2020stochastic}.
Second,
a notable trait of the predominant dense prediction networks is that they are sequential compositions of embedding learning networks and task-specific regressors (or classifiers). While prior arts have shown that
enforcing embedding learning in regression or classification tasks
can yield better predictive performance \cite{li2021learning,wang2021exploring},
it is largely under-explored
if utilizing embedding learning
can also give rise to better-calibrated uncertainty.
In this paper, we propose a novel and generic uncertainty estimation pipeline, called \sysname\ in the paper for \textbf{C}ross-point embedding \textbf{U}ncertainty \textbf{E}stimation, to bridge the gap between the dense prediction of point clouds and its dense uncertainty quantification. \sysname\ involves building a probabilistic embedding model and enforcing metric alignments of massive points in the embedding space. In view of the aforementioned issues, \sysname\ identifies the importance of
embedding learning, and exploits this embedding space via a diagonal multivariate Gaussian model amenable to cross-point interactions. Moreover, we propose \sysname+ that further utilizes cross-point dependencies by a low-rank multivariate Gaussian model. Low-rank covariance matrix in \sysname+ explicitly expresses off-diagonal elements' dependencies while maintaining computational efficiency.
Specifically, our contributions are stated as follows:
\begin{itemize}
\item For the first time we propose a generic dense uncertainty estimation framework for dense prediction tasks of 3D point clouds.
\item We propose a novel method that fully explores cross-point information for dense uncertainty estimation.
\item We validate our proposed method on two representative dense prediction tasks, with the experimental results consistently showing that our method produces better-calibrated uncertainty than state-of-the-arts without losing any predictive performance.
\item Source code of both \sysname\ and \sysname+ is available at: \url{https://github.com/ramdrop/cue}.
\end{itemize}
\section{Related Work}
\begin{figure*}[htbp]
\centering
\includegraphics[width=6.6in]{./figures/pipeline2.pdf}
\caption{An overview of a) traditional probabilistic prediction pipeline (e.g., aleatoric uncertainty \cite{kendall2017uncertainties}) and b) the proposed \sysname\ and \sysname+. We take semantic segmentation for instance where there are $5$ points in the input point cloud and $2$ classes in the labels. A traditional probabilistic prediction pipeline treats logits as distributions where logits can only interact implicitly through a shared MLP (Dashed triangle). In contrast, \sysname\ explores cross-point embeddings by building a probabilistic embedding model (Red curves) and enforcing metric alignments (Blue arrows), and \sysname+ goes further by replacing the diagonal covariance matrix with a low-rank covariance matrix. }
\label{fig_pipeline2}
\end{figure*}
\subsection{Dense Prediction of 3D Point Cloud}
With the dense nature of the 3D point cloud, we focus on its dense prediction tasks, e.g., 3D geometric feature learning and 3D semantic segmentation.
\textit{3D Geometric Features Learning}: To find the correspondences in the absence of relative transformation information, a series of methods is to convert point clouds from the 3D Euclidean space to a feature space, where the correspondences are the nearest neighbors. Early work focus on hand-crafted features, such as SHOT \cite{salti2014shot} and FPFH \cite{rusu2009fast}, we kindly refer readers to \cite{guo2016comprehensive} for more details about hand-crafted features.
Recently deep learned geometric features are becoming popular, which are generally based on volumetric and point-wise operations on point clouds: (1) Volumetric: 3DMatch \cite{zeng20173dmatch} learns patch descriptors by applying a 3D convolutional neural network on volumetric input. FCGF \cite{choy2019fully} directly applies 3D CNN to volumetric point clouds with the hardest contrastive loss, generating dense point features. (2) Point-wise: PointNet \cite{qi2017pointnet} uses multiple parallel shared MLP to learn global or dense features. DGCNN \cite{wang2019dynamic} combines point-wise MLP with dynamic graph neural networks, obtaining flexible and effective feature extractors for unordered point clouds.
SpinNet \cite{ao2021spinnet} proposes a reference axis with a spherical voxelization to learn viewpoint-invariant point descriptors. Nevertheless, the above methods focus on improving predictive performance while ignoring the inherent uncertainty in massive points.
\textit{3D Semantic Segmentation}:
PointNet \cite{qi2017pointnet} is the very first work for 3D point cloud learning, and its shared-MLP architecture shows strong representation capability. However, the perturbation invariance of point clouds is obtained at the cost of ignorance of the local context. Following works propose different solutions to make for this limitation: PointNet++ \cite{qi2017pointnet++} adopts hierarchical sampling strategies, KPConv \cite{thomas2019kpconv} proposes a kernel-based MLP operation mimicking convolution, MinkowskiNet \cite{choy20194d} extends 2D convolution to 3D voxel and specifically design sparse operation python library for point clouds, and recently PointTransformer \cite{zhao2021point} shows the power of Transformer mechanisms in point cloud processing.
\subsection{Dense Uncertainty Estimation}
\textit{Embedding Learning Uncertainty}:
Kendall \cite{kendall2017uncertainties} categorizes uncertainties in deep learning as two types: aleatoric uncertainty and epistemic uncertainty. Aleatoric uncertainty stems from data noises, while epistemic uncertainty refers to model uncertainty, which can be reduced with sufficient training data. Embedding learning is usually applied to image recognition tasks, where most methods focus on estimating aleatoric uncertainty: PFE \cite{shi2019probabilistic} models face embeddings as Gaussian distributions and uses the proposed Mutual Likelihood Score to measure the likelihood of two embeddings belonging to the same class. DUL \cite{chang2020data} proposes to learn aleatoric uncertainty for both regression and classification face recognition tasks. BTL \cite{warburg2021bayesian} proposes a Bayesian loss to learn aleatoric uncertainty in place recognition. RUL \cite{zhang2021relative} uses relative uncertainty measurements to learn aleatoric uncertainty.
In the above image recognition tasks, a single feature is learned for a whole image. But in the dense prediction task of the point cloud, a single point cloud will involve learning thousands of features (i.e., equals to the number of points in the point cloud). Furthermore, image recognition is applied to regular-size images, while point clouds are totally unordered and varied-size. The massive features within a batch and irregular input size render it rather challenging to estimate dense uncertainty for a 3D point cloud.
\textit{Semantic Segmentation Uncertainty}:
Popular uncertainty estimation methods for semantic segmentation include
softmax entropy \cite{czolbe2021segmentation}
, Bayesian Neural Network (BNN) \cite{kendall2015bayesian},
learned aleatoric uncertainty \cite{kendall2017uncertainties}
, auxiliary network \cite{zheng2021rectifying}
and variance propagation based on Assumed Density Function (ADF) \cite{cortinhal2020salsanext}. Please refer to \cite{jungo2019assessing} for a thorough overview.
However, these dominant approaches for semantic segmentation usually treat pixels or points as independent of each other (see Fig. \ref{fig_pipeline2}.a). Such ignorance of cross-pixel or cross-point dependencies tends to result in noisy uncertainty estimation \cite{monteiro2020stochastic}.
Embedding learning has been explored in image segmentation: \cite{wang2021exploring} and \cite{tang2022contrastive} show contrastive learning optimizes embedding space and improve prediction performance in a semantic segmentation task, \cite{li2021learning} proves that optimized embeddings contribute to predictive performance. However, all the above methods exploit embedding learning for improving predictive performance,
rather than estimating dense uncertainty.
SSN \cite{monteiro2020stochastic} has used a low-rank multivariate Gaussian model to account for cross-pixel dependencies. But it is developed for logits, which does not involve embedding optimization.
Our \sysname\
is based on a probabilistic embedding model and enforces metric alignments in the embedding space by using bayesian triplet loss.
Bayesian triplet loss has been used in \cite{warburg2021bayesian}
in image recognition. The major differences are: (1) the image recognition \cite{warburg2021bayesian} requires a single embedding for an image (i.e., whole pixels), while massive point-wise embeddings are desired in \sysname. Thus, we design additional sophisticated sampling strategies and efficient networks for unordered point clouds; and (2) the probabilistic embedding model of \cite{warburg2021bayesian} ignores the cross-point dependencies. Thus, we propose \sysname+ to alleviate this issue by a low-rank multivariate Gaussian model.
\section{Method}
\subsection{Preliminary}
A dense prediction network maps a batch of points to a set of scalars. The process can be decomposed into a metric learning phase and a task-oriented regression or classification phase.
Formally, given a point cloud $\boldsymbol{\mathcal{P}} \in \mathbb{R}^{N \times 3}$, the network $f_{\theta}$ first maps it to a set of embeddings $\boldsymbol{\mathcal{X}} \in \mathbb{R}^{N \times D}$, where $N$ is the number of points, and $D$ is the embedding dimension:
\begin{equation}
\rm metric \ \ learning:\ \boldsymbol{\mathcal{X}} = f_{\theta}(\boldsymbol{\mathcal{P}})
\end{equation}
which is followed by a task-oriented regressor (or classifier) $f_{r}$ that generates predictions $\boldsymbol{\mathcal{Y}} \in \mathbb{R}^{N \times 1}$ (or $\boldsymbol{\mathcal{Y}} \in \mathbb{R}^{N \times C}$ where $C$ is the number of class )
for the set of embedding:
\begin{equation}
\rm regression \ or\ classification:\ \boldsymbol{\mathcal{Y}} = f_{r}(\boldsymbol{\mathcal{X}})
\end{equation}
In the above formulation, predictions are regarded as deterministic,
while the inherent noise from data is ignored. A probabilistic prediction model (e.g., probabilistic semantic segmentation~\cite{kendall2017uncertainties}) casts the prediction as a Gaussian distribution, which provides uncertainty level along with prediction (See Fig. \ref{fig_pipeline2}.a). But embeddings are still deterministic and equally weighted, which means each embedding will contribute equally to the regressor (or classifier).
Inspired by probabilistic contrastive learning in face recognition \cite{shi2019probabilistic,chang2020data}, we adopt a probabilistic embedding model for a point cloud, where embeddings are represented by a diagonal multivariate Gaussian distribution:
\begin{equation}
\label{eq_distribution}
\boldsymbol{\mathcal{X}} \sim \boldsymbol{\mathcal{N}}\left ( \boldsymbol{\mu}, \boldsymbol{\Lambda^2}\right )
\end{equation}
where $\boldsymbol{\mu}=f_\mu(\boldsymbol{\mathcal{P}})\in \mathbb{R}^{N \times D}$ and $\boldsymbol{\Lambda^2}=f_\sigma(\boldsymbol{\mathcal{P}}) \in \mathbb{R}^{(N \times D) \times (N \times D)}$
is a diagonal matrix. $f_\mu$ and $f_\sigma$ represents the mean branch and variance branch of the network $f_\theta$. We will later propose a full-covariance multivariate Gaussian model and show its superiority in Sec. \ref{lrmg}.
\subsection{Exploring Cross-point Embeddings}
After building the probabilistic embedding model, we now discuss how to optimize the embedding space and derive the uncertainty.
An overview of a traditional probabilistic prediction pipeline, the proposed \sysname\ and \sysname+ is presented in Fig. \ref{fig_pipeline2}. A traditional probabilistic prediction only allows logits to interact implicitly through a shared MLP, while \sysname\ explores cross-point embeddings by building a probabilistic embedding model and enforcing metric alignments, and \sysname+ goes further by replacing the diagonal covariance matrix with a low-rank covariance matrix. In what follows we will first describe \sysname\ that is based on the diagonal multivariate Gaussian model. Then an improved version \sysname+ will be introduced, which is based on the low-rank multivariate Gaussian model.
\subsubsection{\sysname}
Given a triplet $\{\boldsymbol{P_a},\boldsymbol{P_p}, \boldsymbol{P_n} \vert \boldsymbol{P_i} \in \mathbb{R}^{1 \times 3}, i=a, p, n\}$, their embeddings are obtained as $\{\boldsymbol{X_a},\boldsymbol{X_p}, \boldsymbol{X_n} \vert \boldsymbol{X_i} \in \boldsymbol{\mathcal{N}}(\boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i^2), \boldsymbol{\mu}_i \in \mathbb{R}^{1 \times D}, \boldsymbol{\Sigma^2}_i \in \mathbb{R}^{1 \times D}, i=a, p, n\}$, where the subscripts $a, p, n$ denote an anchor, positive and negative sample, respectively. In the probabilistic setting, we are interested in the probability of the positive embedding being closer than the negative to the anchor:
\begin{equation}
\label{eq_bayesian_triplet}
P(\Vert \boldsymbol{X}_a - \boldsymbol{X}_p \Vert - \Vert \boldsymbol{X}_a - \boldsymbol{X}_n \Vert + m < 0)
\end{equation}
Rewrite it as:
\begin{equation}
\label{eq_normal}
P(\tau<-m)
\end{equation}
where the new distribution $\tau= \sum_d^D \boldsymbol{T} ^d=\sum_d^D (\boldsymbol{X}_a^d - \boldsymbol{X}_p^d)^2 - (\boldsymbol{X}_a^d - \boldsymbol{X}_n^d)^2$, and $d$ means $d^{th}$ dimension . According to central limit theorem, $\tau$ will approximate a normal distribution when $D$ is large, i.e.,$\frac{\tau - \mu_{\tau}}{\sigma_{\tau}} \thicksim \mathcal{N}(0, 1)$, where $\mu_\tau$ and $\sigma^2_\tau$ are the mean and the variance of the distribution $\tau$. Then Eq. \ref{eq_normal} is solved as:
\begin{equation}
P(\tau<-m) = \Phi_{\mathcal{N}(0, 1)} (\frac{-m - \mu_{\tau}}{\sigma_{\tau}})
\end{equation}
where $\Phi$ is the Conditional Density Function (CDF). Now the task is converted to finding an analytical solution of $\mu_\tau$ and $\sigma_\tau$.
The mean $\mathbb{E}^\prime [\tau]$ and variance $\mathbb{D}^\prime[\tau]$ of a single dimension is given as follows (the superscript $d$ at right-hand side is omitted for brevity):
\begin{equation}
\begin{aligned}
\mathbb{E}[\boldsymbol{T}^d] &= \mu_{p}^{2}+\sigma_{p}^{2}-\mu_{n}^{2}-\sigma_{n}^{2}-2 \mu_{a}\left(\mu_{p}-\mu_{n}\right) \\
\mathbb{D}[\boldsymbol{T}^d] &= 2[\sigma_{p}^{4}+2 \mu_{p}^{2} \sigma_{p}^{2}+2\left(\sigma_{a}^{2}+\mu_{a}^{2}\right)\left(\sigma_{p}^{2}+\mu_{p}^{2}\right)- 2 \mu_{a}^{2} \mu_{p}^{2}\\&-4 \mu_{a} \mu_{p} \sigma_{p}^{2}] + 2[\sigma_{n}^{4}+2 \mu_{n}^{2} \sigma_{n}^{2}+2\left(\sigma_{a}^{2}+\mu_{a}^{2}\right)\left(\sigma_{n}^{2}+\mu_{n}^{2}\right)\\&-2 \mu_{a}^{2} \mu_{n}^{2}-4 \mu_{a} \mu_{n} \sigma_{n}^{2}] +4 \mu_{p} \mu_{n} \sigma_{a}^{2}
\end{aligned}
\end{equation}
Since the embedding model is assumed to be isotropic, we arrive at:
\begin{equation}
\mu_\tau = \sum_d^D \mathbb{E}[\boldsymbol{T} ^d], \quad \sigma _\tau^2 = \sum_d^D \mathbb{D}[\boldsymbol{T} ^d]
\end{equation}
In summary, after the network generates a set of embeddings for a point cloud, we calculate the probability of the positive embedding being closer than the negative to the anchor, and the goal of training is to minimize the metric loss derived from Eq. \ref{eq_bayesian_triplet}:
\begin{equation}
\label{eq_metric_loss}
L_{M} = -\frac{1}{T} \sum _{t=1}^T P(\Vert \boldsymbol{X}_{t,a} - \boldsymbol{X}_{t,p} \Vert - \Vert \boldsymbol{X}_{t,a} - \boldsymbol{X}_{t,n} \Vert + m < 0)
\end{equation}
where $T$ is the number of total triplets in a mini-batch.
\subsubsection{\sysname+}
\label{lrmg}
\begin{figure}[t]
\centering
\includegraphics[width=3.3in]{./figures/network.pdf}
\caption{The network architectures of \sysname\ and \sysname+: $\boldsymbol{\mathcal{P}}$ means a 3D point cloud, $\boldsymbol{\mu}$ embeddings' mean, $\boldsymbol{\Lambda}$ diagonal elements of embeddings' covariance matrix, $\boldsymbol{P}$ scale factor of embeddings' covariance matrix.}
\label{fig_network}
\end{figure}
Points usually show spatial correlation with their neighbors. For example, points at the boundaries of an object usually exhibit high uncertainty since the points around the boundary have varied semantic labels.
But \sysname\ fails to model point-wise dependencies because the diagonal covariance matrix of \sysname\ (see Eq. \ref{eq_distribution}) is based on the assumption that points are independent of each other. To solve this issue, we propose further capturing the point-wise dependencies by a full-covariance multivariate Gaussian model. Specifically, the diagonal covariance matrix in Eq. \ref{eq_distribution} is replaced with a full covariance matrix $\boldsymbol{\Sigma}^2 \in \mathbb{R}^{(N \times D) \times (N \times D)}$:
\begin{equation}
\boldsymbol{\mathcal{X}} \sim \boldsymbol{\mathcal{N}}\left ( \boldsymbol{\mu}, \boldsymbol{\Sigma^2}\right )
\end{equation}
where $\boldsymbol{\mu} \in \mathbb{R}^{N \times D}$. However, the computational complexity of the full covariance matrix $\boldsymbol{\Sigma}^2$ scales with the square of $N$, and a point cloud usually consists of tens of thousands of points, i.e., $N>10^4$. This makes training networks difficult. To alleviate this issue, we resort to a low-rank parameterization of the covariance matrix \cite{magdon2010approximating}:
\begin{equation}
\boldsymbol{\Sigma}^2 = \boldsymbol{P}\boldsymbol{P}^T+\boldsymbol{\Lambda}^2
\end{equation}
where the scale factor $\boldsymbol{P} \in \mathbb{R}^{(N\times D)\times K}$ and $K$ is the rank of the parameterization, $\boldsymbol{\Lambda}^2 \in \mathbb{R}^{(N\times D) \times (N\times D)}$ and $\boldsymbol{\Lambda}^2$ is a diagonal matrix. Therefore, we refer to the pipeline based on a low-rank covariance matrix as \sysname+. Compared with \sysname, \sysname+ learns parameters of additional elements other than diagonal elements of the covariance matrix. This makes the point-wise dependencies explicitly described by the learned variances.
For ease of application, we choose $K=1$. Then the equivalent of the embedding $\boldsymbol{\mathcal{X}}$ is obtained as
\begin{equation}
\begin{aligned}
\boldsymbol{\mathcal{X}}
=& \boldsymbol{\mu} + \left(\boldsymbol{P} + \boldsymbol{\Lambda} \right ) \cdot \boldsymbol{\mathcal{N}}\left (\boldsymbol{0}, \boldsymbol{I} \right ) \\
\end{aligned}
\end{equation}
$L_M$ is then used to train the network. By experimental results we show that \sysname+ generates better-calibrated uncertainty than \sysname\ (see Sec. \ref{experiments}).
The network architectures of the proposed \sysname\ and \sysname+ are shown in Fig. \ref{fig_network}, where $\boldsymbol{\mathcal{P}}$ means a 3D point cloud. The backbone encoder and decoder can be chosen from any UNet-like network. We add three branches to predict the mean $\boldsymbol{\mu}$, diagonal covariance matrix $\boldsymbol{\Lambda^2}$ and the scale factor $\boldsymbol{P}$. $\boldsymbol{\mu}$ branch ends with an L2-Normalization layer, while $\boldsymbol{\Lambda^2}$ and $\boldsymbol{P}$ branches with softplus layers.
\section{Experimental Results}
\label{experiments}
\subsection{3D Geometric Feature Learning}
While \sysname\ and \sysname+ are generic to dense prediction tasks, sampling strategies for triplets should be adapted
according
to different downstream tasks. Here, we present practical sampling strategies for two different
tasks:
3D geometric feature learning and 3D semantic segmentation.
\textit{3D geometric feature learning} aims to learn a discriminative mapping function represented by a deep neural network, such that raw points in the Euclidean space are mapped to the feature space. Ideally, points with similar geometric characteristics should be close to each other in the feature space. \cite{choy2019fully} studies different sampling strategies, including hardest-triplet sampling and random triplet sampling, where triplet loss is adopted. We follow their sampling methods but adapt their conventional triplet loss to our metric loss $L_M$. Specifically, given point clouds $\boldsymbol{\mathcal{P}}_i$ and $\boldsymbol{\mathcal{P}}_j$ and the relative transformation $\boldsymbol{\mathcal{T}}$, we first sample anchor embeddings $\boldsymbol{X}_{i,a}$ and $\boldsymbol{X}_{j,a}$. Then, we randomly choose its positives $\boldsymbol{X}_{i,p}$, $\boldsymbol{X}_{j,p}$ and negatives $\boldsymbol{X}_{i,n}$, $\boldsymbol{X}_{j,n}$. Finally, we calculate the metric loss $L_M$ of the triplets $\{\boldsymbol{X}_{i,a}, \boldsymbol{X}_{i,p}, \boldsymbol{X}_{i,n}\}$ and $\{ \boldsymbol{X}_{j,a}, \boldsymbol{X}_{j,p}, \boldsymbol{X}_{j,n}\}$ for training.
\begin{figure}[t]
\centering
\includegraphics[width=3in]{./figures/ece_fcgf.pdf}
\caption{Reliability diagram on the 3D Match Benchmark. \sysname\ and \sysname+ are closer to the ideally-calibrated line than others.}
\label{fig_ece_fcgf}
\end{figure}
\begin{table}
\centering
\label{table_fcgf}
\caption{Predictive performance and uncertainty quality on the 3D Match Benchmark.
}
\begin{tabular}{l|c|c}
\hline
Method & [email protected] ↑ & ECE ↓ \\
\hline
FPFH$^*$\cite{rusu2009fast} & 36.4 & \textbackslash{} \\
PerfectMatch$^*$\cite{gojcic2019perfect} & 94.9 & \textbackslash{} \\
FCGF$^*$\cite{choy2019fully} & 95.3 & \textbackslash{} \\
SpinNet\cite{ao2021spinnet} & 97.5 & \textbackslash{} \\
FCGF\cite{choy2019fully} & 97.5 & \textbackslash{} \\
\hline
FCGF+RG & 97.5 & 0.251 \\
FCGF+MCD & 94.1 & 0.344 \\
\rowcolor[rgb]{1,0.949,0.8} FCGF+\sysname & 97.5 & 0.142 \\
\rowcolor[rgb]{1,0.949,0.8} FCGF+\sysname+ & 97.6 & 0.135 \\
\hline
\end{tabular}
\begin{tablenotes}
\footnotesize
\item[1] $^*$ denotes predicting correspondences without a symmetric test \cite{horache20213d}.
\end{tablenotes}
\end{table}
\begin{figure}[htbp]
\centering
\includegraphics[width=3.3in]{./figures/qua_fcgf.pdf}
\caption{Matching results and dense uncertainty map (estimated by \sysname+) of a point cloud from 3D Match Benchmark. Incorrect correspondences ($1$ and $2$ areas) tend to have high uncertainties. }
\label{fig_qua_fcgf}
\end{figure}
\noindent \textbf{Datasets.} We use the 3D Match dataset \cite{zeng20173dmatch}, following the official training and evaluation splits.
\noindent \textbf{Model Architectures.}
FCGF is the first 3D convolutional network to integrate metric learning in a fully-convolutional setting. We choose FCGF \cite{choy2019fully} as our backbone because it holds state-of-the-art predictive performance with fast training and inferencing. To empower the deterministic FCGF to estimate the uncertainty of each point, we integrate it with our proposed \sysname\ and \sysname+ as is shown in Fig. \ref{fig_network}.
\noindent \textbf{Training Details.}
We train FCGF following the original paper \cite{choy2019fully}, i.e., Hardest-contrastive loss, $100$ epoches with SGD optimizer and batch size $4$, learning rate starts from $0.1$ with exponetial decay rate $0.99$, dada augmentation includes random scaling $\in [0.8, 1.2]$ and random rotation $\in [0 ^\circ , 360^\circ)$.
\noindent \textbf{Competing Methods.}
\begin{itemize}
\item RG: After training the FCGF, we randomly form ten bins of points and then calculate ECE.
\item MCD: We insert dropout layers with dropout rate $p=0.1$ after every convolutional layer. We take $N=40$ samples from the weights' posterior distribution at test time.
\item \sysname: To assure original predictive performance, we freeze the $\boldsymbol{\mu}$ branch and train $\boldsymbol{\Lambda^2}$ branches with the metric loss $L_M$.
\item \sysname+: We freeze the $\boldsymbol{\mu}$ branch, and train $\boldsymbol{\Lambda^2}$ and $\boldsymbol{P}$ branches with the metric loss $L_M$.
\end{itemize}
Note that MCD produces epistemic uncertainty, while our methods generate aleatoric uncertainty. We include it here for a comprehensive comparison.
\noindent \textbf{Evaluation Metrics.}
To evaluate the predictive performance, we use Feature Matching Recall with $0.1m$ inlier distance threshold and $0.05$ inlier recall threshold ([email protected]) \cite{choy2019fully}. We adopt the widely used Expected Calibrated Error (ECE) \cite{warburg2021bayesian} and the reliability diagram \cite{warburg2021bayesian} to evaluate uncertainty quality, where we calculate the Hit Ratio \cite{choy2019fully} of points in the same bin.
\noindent \textbf{Results.}
We evaluate the above methods on the 3D Match Benchmark \cite{zeng20173dmatch}. We establish correspondences by the nearest neighbor search in the embedding space, where each correspondence has an estimated uncertainty.\footnote[1]{We follow the covariance formulation in \cite{warburg2021bayesian} and use the sum of two points' uncertainty as the correspondence's uncertainty.} Table. \ref{table_fcgf} shows the predictive performance and uncertainty quality of different methods on the 3DMatch dataset. MCD shows degraded predictive performance due to the dropout layers significantly harming the network's representation ability. Since the $\boldsymbol{\mu}$ branch is inherited from the backbone network, \sysname\ and \sysname+ do not sacrifice any predictive accuracy. Compared with MCD, \sysname\ reduces ECE by $0.202$. Besides, \sysname+ outperforms \sysname\ with ECE $0.135$.
Fig. \ref{fig_ece_fcgf} illustrates the reliability diagram on the 3D Match Benchmark. The ideal line means points with higher uncertainty levels should have lower hit ratios. RG produces a horizontal line, while MCD fails to produce a sensible estimation. \sysname\ and \sysname+ present closer lines to the ideal line. Fig. \ref{fig_qua_fcgf} shows the matching results and dense uncertainty map estimated by \sysname+ of a point cloud. We can observe that incorrect correspondences ($1$ and $2$ areas) tend to have high uncertainties.
In summary, the proposed \sysname\ and \sysname+ provide well-calibrated uncertainty that can be used as an effective tool to filter incorrect correspondence.
\subsection{3D Semantic Segmentation}
\begin{figure}[t]
\centering
\includegraphics[width=3in]{./figures/ece_scannet.pdf}
\caption{Reliability diagram on the ScanNet validation split. \sysname\ and \sysname+ are closer to the ideal calibrated line than other methods.}
\label{fig_ece_scannet}
\end{figure}
\textit{3D semantic segmentaion} aims to learn a classification network that predicts class labels for each point in a point cloud. To optimize the embedding space by enforcing metric alignments, we first randomly sample anchors from a point cloud $\boldsymbol{\mathcal{P}}$, and then, within the neighbors of each anchor $\boldsymbol{X}_{a}$, choose embeddings with the same class label as positives $\boldsymbol{X}_{p}$, and those with different class label as negatives $\boldsymbol{X}_{n}$. Finally, we calculate the metric loss $L_M$ of the triplet $\{ \boldsymbol{X}_{a}, \boldsymbol{X}_{p}, \boldsymbol{X}_{n}\}$ for training.
\noindent \textbf{Datasets.}
Following \cite{park2022fast}, we use the ScanNet dataset \cite{dai2017scannet} and evaluate models on the ScanNet validation split.
\noindent \textbf{Model Architectures.}
Considering inference latency and accuracy, we choose MinkowskiNet42 (Mink) \cite{choy20194d, park2022fast} as our 3D semantic segmentation backbone. The semantic segmentation network is the same as that in Fig. \ref{fig_network}, except that we add a convolution layer as the segmentation classfier before the L2-Normalization layer of the $\boldsymbol{\mu}$ branch.
\begin{table}
\centering
\label{table_scannet}
\caption{Predictive performance and uncertainty quality on the ScanNet validation split.
}
\begin{tabular}{l|l|c|c}
\hline
& Method & mIOU ↑ & ECE ↓ \\
\hline
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}Training \\without \\uncertainty\end{tabular}} & PointNet\cite{qi2017pointnet} & 0.535 & \textbackslash{} \\
& PointConv\cite{wu2019pointconv} & 0.610 & \textbackslash{} \\
& KPConv
deform \cite{thomas2019kpconv} & 0.692 & \textbackslash{} \\
& SparseConvNet\cite{graham20183d} & 0.693 & \textbackslash{} \\
& Mink\cite{choy20194d} & 0.715 & \textbackslash{} \\ & Mink+SE\cite{jungo2019assessing} & 0.715 & 0.251 \\
\hline
\multirow{4}{*}{\begin{tabular}[c]{@{}c@{}}Training \\with \\uncertainty\end{tabular}} & Mink+AU\cite{kendall2017uncertainties} & 0.717 & 0.254 \\
& Mink+MCD(p=0.2) & 0.658 & 0.176 \\
& Mink+MCD(p=0.05) & 0.663 & 0.170 \\ & {\cellcolor[rgb]{1,0.949,0.8}}Mink+\sysname\ & {\cellcolor[rgb]{1,0.949,0.8}}0.721 & {\cellcolor[rgb]{1,0.949,0.8}}0.142 \\
& {\cellcolor[rgb]{1,0.949,0.8}}Mink+\sysname+ & {\cellcolor[rgb]{1,0.949,0.8}}0.727 & {\cellcolor[rgb]{1,0.949,0.8}}0.141 \\
\hline
\end{tabular}
\end{table}
\begin{figure}[htbp]
\centering
\includegraphics[width=3.3in]{./figures/qua_compare.pdf}
\caption{Segmentation errors (left column) and dense uncertainty maps (right column) on a scene from ScanNet validation split. \sysname\ and \sysname+ produce better-calibrated dense uncertainty maps than others. For correct predictions (rectangular area $1$), \sysname\ is under-confident while \sysname+ is more confident than \sysname.}
\label{fig_qua_compare}
\end{figure}
\noindent \textbf{Training Details.}
We train the model for $10^5$ steps with an SGD optimizer, learning rate starting from $0.1$ with a cosine annealing schedule and a linear warmup. We use a batch size of $8$. For more training details, we kindly refer readers to \cite{park2022fast}.
\noindent \textbf{Competing Methods.}
We compare \sysname\ and \sysname+ with the following popular uncertainty estimation methods from image segmentation:
\begin{itemize}
\item Softmax Entropy \cite{jungo2019assessing} (SE):
\begin{equation}
H = -\sum_c^Cp_c\log(p_c)/log(C) \in [0,1]
\end{equation}
where $C$ is the number of classes, $p_c$ is a probability by the softmax layer.
\item Aleatoric Uncertainty \cite{kendall2017uncertainties, jungo2019assessing} (AU): Logits are modeled as Gaussian distribution, whose mean and variance are predicted by two heads of the network. We use MC sampling ($n=10$ ) to draw samples from the logits distribution and optimize the network with Cross Entropy Loss.
\item MCD \cite{jungo2019assessing}: MCD estimates epistemic uncertainty because dropout at test time approximates random sampling of the model's weights. Test time inference is obtained by
$ p_c = \frac{1}{N}\sum_n^Np_{n,c} $, where $p_c$ denotes the output of the Softmax layer. We set the number of MC samples $N=40$ as suggested by \cite{kendall2016modelling}. We evaluate MCD with two dropout probability settings: $p=0.2$ and $p=0.05$. Aleatoric uncertainty and MCD uncertainty generates high-dimensional variance vectors, which are converted to uncertainty levels by $y(1-0.5q)+(1-y)(0.5q)$, where $q\in[0,1]$ is the normalized variance\cite{jungo2019assessing}.
\item \sysname\ / \sysname+: We train the \sysname\ / \sysname+ network from scratch with a weighted sum of cross entropy loss and the metric loss $L = L_{CR} + \lambda L_M$,
where we set $\lambda=1$ for all experiments.
\end{itemize}
\noindent \textbf{Evaluation Metrics.}
We use the mean IOU (mIOU) to evaluate the predictive performance. mIOU refers to the ratio of the intersection of ground-truth labels and predicted labels to their union, and a higher mIOU indicates better performance. The reliability diagram \cite{warburg2021bayesian} and ECE \cite{warburg2021bayesian} are adopted to evaluate uncertainty quality, where we calculate the precision of points in each bin.
\noindent \textbf{Results.}
Table. \ref{table_scannet} presents the predictive performance and uncertainty quality on the ScanNet validation split. In terms of predictive performance, we observe that AU, \sysname\ produce comparable results to Mink, while MCD shows degraded performance due to dropout layers decreasing representative power. \sysname+ promotes the Mink's predictive power with $0.13$ boost in mIOU. From the perspective of uncertainty quality, SE achieves a $0.251$ ECE, outperformed by MCD (p=0.05) with a $0.170$ ECE, while \sysname\ and \sysname+ provide significantly improved uncertainty with ECE $0.142$ and $0.141$, i.e., \sysname+ reduces ECE of SE by $43.8\%$.
Fig. \ref{fig_ece_scannet} shows the reliability diagram on the ScanNet validation split. It can be seen that \sysname\ is close to the ideal calibrated line, while \sysname+ improves \sysname\ in the low-uncertainty region. Fig. \ref{fig_qua_scannet} presents the qualitative results of \sysname+, where we can observe a significant correlation between segmentation prediction error and estimated uncertainty, i.e., Incorrect predictions tend to have high uncertainties.
Fig. \ref{fig_qua_compare} presents segmentation errors and dense uncertainty maps by different methods on the ScanNet validation split. For incorrect predictions (black points in the magnified area), we can observe that SE fails to detect them and shows high confidence, while \sysname\ and \sysname+ are uncertain about those incorrect predictions. AU is under-confident in most areas, while MCD (p=0.05) cannot produce sensible results.
For correct predictions (Rectangular area $1$),
\sysname\ is under-confident while \sysname+ is more confident than \sysname.
The above results indicate that \sysname\ and \sysname+ provide better-calibrated uncertainty than existing methods without compromising any predictive performance. \sysname+ outperforms \sysname\ in both predictive performance and uncertainty quality. This shows that embedding learning contributes to uncertainty estimation of dense prediction tasks and low-rank multivariate Gaussian model is more effective than a diagonal one.
\section{Conclusion}
\label{conclusion}
Observing the fact that dense prediction networks are sequential compositions of embedding learning networks and task-specific regressors (or classifiers), we propose \sysname\ that estimates dense uncertainty by building a probabilistic embedding model and enforcing metric alignments with a diagonal multivariate Gaussian model. Besides, we propose \sysname+ that further enhances cross-point interactions with a low-rank multivariate Gaussian model, which explicitly expresses off-diagonal elements' dependencies while maintaining computational efficiency. Experimental results on the 3D Match Benchmark and the ScanNet dataset prove that \sysname\ and \sysname+ are generic and effective tools for 3D dense uncertainty estimation.
| f53f2c7d2d14d650583ca03b14fbc0370b0e5d2a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Distributional reinforcement learning~\citep{bellemare2017distributional, dabney2017distributional, dabney2018implicit, yang2019fully, zhou2020non, nguyen2020distributional,luo2021distributional, sun2022distributional} characterizes the intrinsic randomness of returns within the framework of Reinforcement Learning~(RL). When the agent interacts with the environment, the intrinsic uncertainty of the environment seeps in the the stochasticity of rewards the agent receives and the inherently chaotic state and action dynamics of physical interaction, increasing the difficulty of the RL algorithm design. Distributional RL is aimed at representing the entire distribution of returns in order to capture more intrinsic uncertainty of the environment, and therefore to use these value distributions to evaluate and optimize the policy. This is in stark contrast to the classical RL that only focuses on the expectation of the return distributions, such as temporal-difference~(TD) learning~\citep{sutton2018reinforcement} and Q-learning~\citep{watkins1992q}.
As a promising branch of RL algorithms, distributional RL has demonstrated the state-of-the-art performance in a wide range of environments, e.g., Atari games, in which the representation of return distributions and the distribution divergence between the current and target return distributions within each Bellman update are pivotal to its empirical success~\citep{dabney2018implicit, sun2021interpreting, sun2022distributional}. Specifically, categorical distributional RL, e.g., C51~\citep{bellemare2017distributional, rowland2018analysis}, integrates a categorical distribution by approximating the density probabilities in pre-specified bins with a bounded range and Kullback-Leibler~(KL) divergence, serving as the first successful distributional RL family in recent years. Quantile Regression~(QR) distributional RL, e.g., QR-DQN~\citep{dabney2017distributional}, approximates Wasserstein distance by the quantile regression loss and leverages quantiles to represent the whole return distribution. Other variants of QR-DQN, including Implicit Quantile Networks~(IQN)~\citep{dabney2018implicit} and Fully parameterized Quantile Function~(FQF)~\citep{yang2019fully}, can even achieve significantly better performance across plenty of Atari games. Moment Matching distributional RL~\citep{nguyen2020distributional} learns deterministic samples to evaluate the distribution distance based on Maximum Mean Discrepancy, while a more recent work called Sinkhorn distributional RL~\citep{sun2022distributional} interpolates Maximum Mean Discrepancy and Wasserstein distance via Sinkhorn divergence~\citep{sinkhorn1967diagonal}. Meanwhile, distributional RL also inherits other benefits in risk-sensitive control~\citep{dabney2018implicit}, policy exploration settings~\citep{mavrin2019distributional, rowland2019statistics} and robustness~\citep{sun2021exploring}.
Despite the remarkable empirical success of distributional RL, the illumination on its theoretical advantages is still less studied. A distributional regularization effect~\citep{sun2021interpreting} stemming from the additional value distribution knowledge has been characterized to explain the superiority of distributional RL over classical RL, but the benefit of the proposed regularization on the optimization of algorithms has not been investigated as the optimization plays a key role in RL algorithms. In the literature of strategies that can help the learning in RL, a recent progress mainly focuses on the policy gradient methods~\citep{sutton2018reinforcement}. \cite{mei2020global} show that the policy gradient with a softmax parameterization converges at a $\mathcal{O}(1/t)$ rate, with constants depending on the problem and initialization, which significantly expands the existing asymptotic convergence results. Entropy regularization~\citep{haarnoja2017reinforcement,haarnoja2018soft} has gained increasing attention as it can significantly speed up the policy optimization with a faster linear convergence rate~\citep{mei2020global}. \cite{ahmed2019understanding} provide a fine-grained understanding on the impact of entropy on policy optimization, and emphasize that any strategy, such as entropy regularization, can only affect learning in one of two ways: either it reduces the noise in the gradient estimates or it changes the optimization landscape. These commonly-used strategies that accelerate RL learning inspire us to further investigate the optimization impact of distributional RL arising from the exploitation of return distributions.
In this paper, we study the theoretical superiority of distributional RL over classical RL from the optimization standpoint. We begin by analyzing the optimization impact of different strategies within the Neural Fitted Z-Iteration~(Neural FZI) framework and point out two crucial factors that contribute to the optimization of distributional RL, including the distribution divergence and the distribution parameterization error. The smoothness property of distributional RL loss function has also be revealed leveraging the categorical parameterization, yielding its stable optimization behavior. The uniform stability in the optimization process can thus be more easily achieved for distributional RL in contrast to classical RL. In addition to the optimization stability, we also elaborate the acceleration effect of distributional RL algorithms based on the value distribution decomposition technique proposed recently. It turns out that distributional RL can be shown to speed up the convergence and perform favorably if the value distribution is approximated appropriately, which is measured by the variance of gradient estimates. Empirical results collaborate that distributional RL indeed enjoys a stable gradient behavior by observing smaller gradient norms in terms of the observations the agent encounters in the learning process. Besides, the variance reduction of gradient estimates for distributional RL algorithms with respect to network parameters also provides strong evidence to demonstrate the smoothness property and acceleration effects of distributional RL. Our study opens up many exciting research pathways in this domain through the lens of optimization, paving the way for future investigations to reveal more advantages of distributional RL.
\section{Preliminary Knowledge}\label{sec:preliminary}
\paragraph{Classical RL.} In a standard RL setting, the interaction between an agent and the environment is modeled as a Markov Decision Process~(MDP) ($\mathcal{S}, \mathcal{A}, R, P, \gamma$), where $\mathcal{S}$ and $\mathcal{A}$ denote state and action spaces. $P$ is the transition kernel dynamics, $R$ is the reward measure and $\gamma \in (0,1)$ is the discount factor. For a fixed policy $\pi$, the return, $Z^{\pi}=\sum_{t=0}^{\infty} \gamma^t R_t$, is a random variable representing the sum of discounted rewards
observed along one trajectory of states while following the policy $\pi$. Classical RL focuses on the value function and action-value function, the expectation of returns $Z^{\pi}$. The action-value function $Q^\pi(s, a)$ is defined as $Q^{\pi}(s, a)=\mathbb{E}\left[Z^{\pi}(s, a)\right]=\mathbb{E}\left[\sum_{t=0}^{\infty} \gamma^t R\left(s_t, a_t\right) \right]$, where $s_0=s$, $a_0=a$, $s_{t+1}\sim P(\cdot|s_t, a_t)$, and $a_t \sim \pi(\cdot|s_t)$.
\paragraph{Distributional RL.} Distributional RL, on the other hand, focuses on the action-value distribution, the full distribution of $Z^{\pi}(s, a)$ rather than only its expectation, i.e., $Q^\pi(s, a)$. Leveraging knowledge on the entire value distribution can better capture the uncertainty of returns and thus can be advantageous to explore the intrinsic uncertainty of the environment~\citep{dabney2018implicit,mavrin2019distributional}. The scalar-based classical Bellman updated is therefore extended to distributional Bellman update, which allows a flurry of distributional RL algorithms, mainly including Categorical distributional RL~\citep{bellemare2017distributional} and Quantie Regression Distributional RL~\citep{dabney2017distributional,dabney2018implicit}.
\paragraph{Categorical Distributional RL.} As the first successful distributional RL family, Categorical distributional RL~\citep{bellemare2017distributional} approximates the action-value distribution $\eta$ by a categorical distribution $\hat{\eta}=\sum_{i=1}^{k}f_i \delta_{l_i}$ where $l_1, l_2, ..., l_k$ is a set of fixed supports and $\{f_i\}_{i=1}^k$ are learnable probabilities, normally parameterized by a neural network. A projection is also introduced to have the joint support with a newly distributed target probabilities, equipped by a KL divergence to compute the distribution distance between the current and target value distribution within each Bellman update. In practice, C51~\citep{bellemare2017distributional}, an instance of Categorical Distributional RL with $k=51$, performs favorably on a wide range of environments.
\paragraph{Quantile Regression~(QR) Distributional RL.} QR Distributional RL~\citep{dabney2017distributional,dabney2018implicit} approximates the value distribution $\eta$ by a mixture of Dirac $\hat{\eta}=\frac{1}{N}\sum_{i=1}^{N}\delta_{\tau_i}$, where $\tau_i=F^{-1}_\eta(\frac{2i-1}{2N})$ are the learnable quantile values at the fixed quantiles $\{\frac{2i-1}{2N}\}$ and $F^{-1}$ is the inverse cumulative distribution function of $\eta$. Since the quantile regression loss proposed in QR distributional RL can be used to approximate the Wasserstein distance, it gains favorable performance on Atari games. Moreover, the performance has been further improved by a series of variants based on quantile regression loss~\citep{dabney2018implicit,yang2019fully,zhou2020non}. For example, Implicit Quantile Network~(IQN)~\citep{dabney2018implicit} utilizes a continuous mapping for the quantile function $F^{-1}_\eta(\frac{2i-1}{2N})$ rather than a fixed set of quantiles, which expands the expressiveness power of function approximators to represent the value distribution.
\section{Optimization Effect of Distributional RL}\label{sec:optimization}
We consider the function approximation setting to analyze the optimization benefit of distributional RL. In Section~\ref{sec:neuralFZI}, we begin by showing the different roles of components in distributional RL on the entire optimization of RL algorithms within the Neural FZI framework. Further, in Section~\ref{sec:stability} we reveal the desirable smoothness properties of distributional RL loss function as opposed to classical RL, contributing to the stable optimization. Finally, the acceleration effect of distributional RL stemming from the additional value distribution is analyzed in Section~\ref{sec:acceleration}, which is characterized by the variance of gradient estimates.
\subsection{How to Optimize Neural Fitted Z-Iteration for Distributional RL?}\label{sec:neuralFZI}
In classical RL, \textit{Neural Fitted Q-Iteration}~(Neural FQI)~\citep{fan2020theoretical,riedmiller2005neural} provides a statistical interpretation of DQN~\citep{mnih2015human} while capturing its two key features, i.e., the leverage of target network and the experience replay:
\begin{equation}\begin{aligned}\label{eq:Neural_Q_fitting}
Q_\theta^{k+1}=\underset{Q_{\theta}}{\operatorname{argmin}} \frac{1}{n} \sum_{i=1}^{n}\left[y_{i}-Q_\theta^k\left(s_{i}, a_{i}\right)\right]^{2},
\end{aligned}\end{equation}
where the target $y_{i}=r(s_i, a_i)+\gamma \max _{a \in \mathcal{A}} Q^k_{\theta^*} \left(s_{i}^{\prime}, a\right)$ is fixed within every $T_{\text{target}}$ steps to update target network $Q_{\theta^*}$ by letting $\theta^*=\theta$. The experience buffer induces independent samples $\left\{\left(s_{i}, a_{i}, r_{i}, s_{i}^{\prime}\right)\right\}_{i \in[n]}$ and ideally without the optimization and TD approximation errors, Neural FQI is exactly the updating under Bellman optimality operator~\citep{fan2020theoretical}. Similarly, in distributional RL, \cite{sun2021interpreting,ma2021conservative} proposed \textit{Neural Fitted Z-Iteration}~(Neural FZI), a distributional version of Neural FQI based on the parameterization of $Z_\theta$:
\begin{equation}
\begin{aligned}\label{eq:Neural_Z_fitting}
Z_\theta^{k+1}=\underset{Z_{\theta}}{\operatorname{argmin}} \frac{1}{n} \sum_{i=1}^{n} d_p (Y_{i}, Z_\theta^k\left(s_{i}, a_{i}\right)),
\end{aligned}
\end{equation}
where the target $Y_{i}=R(s_i, a_i)+\gamma Z^k_{\theta^*} \left(s_{i}^{\prime}, \pi_Z(s_i^\prime)\right)$ is a random variable, whose distribution is also fixed within every $T_{\text{target}}$ steps. The target follows a greedy policy rule, where $\pi_Z(s^\prime_i)= \operatorname{argmax}_{a^\prime} \mathbb{E}\left[Z_{\theta^*}^{k}(s_i^\prime, a^\prime)\right]$ and $d_p$ is the choice of distribution distance. Within the Neural FZI process, we can easily perceive that there are mainly two crucial components that determine the comprehensive optimization of distributional RL algorithms.
\begin{itemize}
\item \textbf{The choice of $d_p$}. $d_p$ in fact has two-fold impacts on the optimization of the whole Neural FZI. Firstly, $d_p$ determines the convergence rate of distributional Bellman update. For instance, distributional Bellman operator under Crámer distance is $\gamma^{\frac{1}{2}}$-contractive, and is a $\gamma$-contraction when $d_p$ is Wasserstein distance. Apart from the impact on the distributional Bellman update speed, $d_p$ also largely affects the continuous optimization problem to estimate parameter $\theta$ in $Z_\theta$ within each iteration of Neural FZI, including the convergence speed and the bad or good local minima issues.
\item \textbf{The parameterization manner of $Z_\theta$}. The distribution representation way of $d_p$ plays an integral part of the optimization for deep RL algorithms. For example, with more expressiveness power on quantile functions, IQN outperforms QR-DQN on a wider range of Atari games, which is intuitive as a more informative representation way can approximate the true value distribution more reasonably. A smaller value distribution parameteriation error is also potential to help the optimization albeit in an indirect avenue.
\end{itemize}
Owing to the fact that convergence rates of distributional Bellman update under typical $d_p$ are basically known, our optimization analysis mainly focuses on the impact of $d_p$ and the paramterization error of $Z_\theta$ on the continuous optimization within Neural FZI of distributional RL by comparing Neural FQI of classical RL. In Sections~\ref{sec:stability} and \ref{sec:acceleration}, we attribute the optimization benefits of distributional RL to its distribution objective function, consisting of the aforementioned two factors, as opposed to the vanilla least squared loss in classical RL.
\subsection{Stable Optimization Analysis based on Categorical Parameterization}\label{sec:stability}
To allow for a theoretical analysis, we resort to the categorical parameterization equipped with KL divergence in categorical distributional RL~\citep{bellemare2017distributional}, e.g., C51, in order to investigate the stable optimization properties within each iteration in Neural FZI. Concretely, we assume $Z_\theta$ is absolutely continuous and the current and target value distributions under KL divergence within a bounded range have joint supports~\citep{arjovsky2017towards}, under which the KL divergence is well-defined. Note that this analysis strategy is slightly different from vanilla Categorical distributional RL, which also introduces a projection to redistribute probabilities of target value distribution by the neighboring smoothing without the joint support assumption. We slightly simplify Categorical distributional RL by assuming that the target distribution is still within the pre-specified support, which is still easy to satisfy in practice given a relative large bounded range $[l_0, l_k]$ in advance.
To approximate the categorical distribution, we leverage the \textit{histogram function} $f^{s, a}$ with $k$ uniform partitions on the support to parameterize the approximated probability density function of $Z(s, a)$. With KL divergence as $d_p$, we can eventually derive the distribution objective function to be optimized within each update in Neural FZI, which is similar to the histogram distributional loss proposed in \citep{imani2018improving}.
In particular, we denote $\mathbf{x}(s)$ as the state feature on each state $s$, and we let the support of $Z(s, a)$ be uniformly partitioned into $k$ bins. The output dimension of $f^{s,\cdot}$ can be $|\mathcal{A}| \times k$, where we use the index $a$ to focus on the function $f^{s, a}$. Hence, the function $f^{s, a}: \mathcal{X} \rightarrow[0,1]^{k}$ provides a k-dimensional vector $f^{s, a}(\mathbf{x}(s))$ of the coefficients, indicating the probability that the target is in this bin given the state feature $\mathbf{x}(s)$ and action $a$. Next, we use \textit{softmax} based on the linear approximation $\mathbf{x}(s)^{\top} \theta_{i}$ to express $f^{s, a}$, i.e., $f_{i}^{s, a, \theta}(\mathbf{x}(s))=\exp \left(\mathbf{x}(s)^{\top} \theta_{i}\right) / \sum_{j=1}^{k} \exp \left(\mathbf{x}(s)^{\top} \theta_{j}\right)$. For simplicity, we use $f_{i}^{\theta}(\mathbf{x}(s))$ to replace $f_{i}^{s, a, \theta}(\mathbf{x}(s))$. Note that the form of $f^{s, a}$ is similar to that in Softmax policy gradient optimization~\citep{mei2020global,sutton2018reinforcement}, but here we focus on the value-based RL rather than the policy gradient RL. Our prediction probability $f_i^{s, a}$ is redefined as the probability in the $i$-th bin over the support of $Z(s, a)$, thus eventually serving as a density function. While the linear approximator is clearly limited, this is the setting where so far the cleanest results have been achieved and understanding this setting is a necessary first step towards the bigger problem of understanding distributional RL algorithms. Under this categorical parameterization equipped with KL divergence, the resulting distributional objective function $\mathcal{L}_\theta(s, a)$ for the continuous optimization for each $s, a$ pair in each iteration of Neural FZI~(Eq.~\ref{eq:Neural_Z_fitting}) can be expressed as:
\begin{equation}\begin{aligned}\label{eq:histogram}
\mathcal{L}_\theta(s, a) = -\sum_{i=1}^{k} \int_{l_{i}}^{l_{i}+w_{i}} p^{s, a}(y) \log \frac{f_{i}^\theta(\mathbf{x}(s))}{w_{i}} d y \propto -\sum_{i=1}^{k} p^{s, a}_{i} \log f_{i}^\theta(\mathbf{x}(s)),
\end{aligned}\end{equation}
where $\theta=\{\theta_1, ..., \theta_{k}\}$ and $p^{s, a}_i$ is the probability in the $i$-th bin of the true density function $p^{s, a}(x)$ for $Z(s, a)$ defined in Eq.~\ref{eq:decomposition}. $w_i$ is the width for the $i$-th bin $(l_i, l_{i+1}]$. The derivation of the categorical distributional loss under the categorical parameterization is given in Appendix~\ref{appendix:histogram}. To attain the stable optimization property of distributional RL, we firstly derive the appealing properties of the new categorical distributional loss in Eq.~\ref{eq:histogram}, as shown in Proposition~\ref{prop:lipschitz}.
\begin{prop}\label{prop:lipschitz} (Properties of Categorical Distributional Loss)
Assume the state features $\Vert \mathbf{x}(s) \Vert \leq l$ for each state $s$, then $\mathcal{L}_\theta$ is $kl$-Lipschitz continuous, $kl^2$-smooth and convex w.r.t. the parameter $\theta$.
\end{prop}
Please refer to Appendix~\ref{appendix:lemma_lipschitz} for the proof. The derived smoothness properties of $d_p$ under the categorical distributional loss plays an integral role in the stable optimization for distributional RL. In stark contrast, classical RL optimizes a least squared loss function~\citep{sutton2018reinforcement} in Neural FQI. It is known that the least squared estimator has no bounded Lipschitz constant in general and is only $\lambda_\text{max}$-smooth, where $\lambda_\text{max}$ is the largest singular value of the design or data matrix. More specifically, for the categorical distributional loss in distributional RL, we have $\Vert \nabla_\theta \mathcal{L}_\theta \Vert \le kl$, while the gradient norm in classical RL is $|y_i - Q_\theta^k(s, a)|\Vert \mathbf{x}(s) \Vert$, where $Q_\theta^k(s, a)=\sum_{i=1}^{k} (l_i+l_{i+1})f_{i}^\theta(\mathbf{x}(s)) / 2 w_{i}$ under the same categorical parameterization for a fair comparison. Clearly, $Q_\theta^k(s, a)$ can be sufficiently large if the support $[l_0, l_k]$ is specified to be large, which is common in environments where the agent is able to attain a high level of expected returns~\citep{bellemare2017distributional}. As such, $|y_i - Q_\theta^k(s, a)|$ can vary significantly more than $k$ and therefore classical RL with the potentially larger upper bound of gradient norms is prone to the instability optimization issue.
After providing the intuitive comparison in terms of gradient norms above, we next show that distributional RL loss can induce an uniform stability property under the desirable smoothness properties analyzed in Proposition~\ref{prop:lipschitz}. We recap the definition of uniform stability for an algorithm while running \textit{Stochastic Gradient Descent}~(SGD) in Definition~\ref{def:stability}.
\begin{myDef}\label{def:stability}(Uniform Stability)~\citep{hardt2016train}
Consider a loss function $g_w(z)$ parameterized by $w$ encountered on the example $z$, a randomized algorithm $\mathcal{M}$ is uniformly stable if for all data sets $\mathcal{D}, \mathcal{D}^\prime$ such that $\mathcal{D}, \mathcal{D}^\prime$ differ in at most one example, we have
\begin{equation}\begin{aligned}\label{eq:uniform_stable}
\sup_{z} \mathbb{E}_{\mathcal{M}}\left[g_{\mathcal{M}(\mathcal{D}) }(z)-g_{\mathcal{M}\left(\mathcal{D}^{\prime}\right) }\left(z\right)\right] \leq \epsilon_{\text {stab }}.
\end{aligned}\end{equation}
\end{myDef}
In Theorem~\ref{theorem:lipschitz}, we show that while running SGD to solve the categorical distributional loss within each Neural FZI, the continuous optimization process in each iteration is $\epsilon_{\text{stab}}$-uniformly stable.
\begin{theorem}\label{theorem:lipschitz} (Stable Optimization for Distributional RL) Suppose that we run SGD under $\mathcal{L}_\theta$ in Eq.~\ref{eq:histogram} with step sizes $\lambda_t \le 2 / kl^2$ for $T$ steps. Assume $\Vert \mathbf{x}(s) \Vert \leq l$ for each state $s$ and action $a$, then we have $\mathcal{L}_\theta$ satisfies the uniform stability in Definition~\ref{def:stability} with $\epsilon_{\text {stab }} \leq \frac{4kT}{n}$, i.e.,
\begin{eqnarray}
\begin{aligned}
\mathbb{E}\left|\mathcal{L}_{\theta_T}(s, a) - \mathcal{L}_{\theta_T^\prime}(s, a)\right| \leq \frac{4kT}{n},
\end{aligned}
\end{eqnarray}
where $\theta_T$ and $\theta_T^\prime$ are the minimizers after $T$ steps under the dataset $\mathcal{D}$ and $\mathcal{D}^\prime$, respectively.
\end{theorem}
Please refer to the proof of Theorem~\ref{theorem:lipschitz} in Appendix~\ref{appendix:lipschitz}. The stable optimization has multiple advantages. In deep learning optimization literature~\citep{hardt2016train}, an uniform stability can guarantee $\epsilon_{\text {stab }}$-bounded generalization gap. In reinforcement learning, algorithms with more stability tend to achieve a better final performance~\citep{bjorck2021towards,li2021functional,ahmed2019understanding}.
In summary, under the categorical parameterization equipped with KL divergence, the continuous optimization objective function within each update of Neural FZI for distributional RL is uniformly stable with the stability errors shrinking at the rate of $O(n^{-1})$, and the immediately obtained bounded generalization gap also guarantees a desirable local minima. This advantage can be owing to the desirable smoothness property of categorical distributional loss with a potentially smaller upper bound of gradient norms compared with classical RL. Empirically, in Section~\ref{sec:experiments}, we validate the stable gradient behaviors of categorical distributional RL, and similar results are also observed in Quantile Regression distributional RL. By contrast, without these smooth properties, classical RL may not yield the stable optimization property directly. For example, $\lambda_\text{max}$-smooth may be of less help for the optimization given a bad conditional number of the design matrix where $\lambda_\text{max}$ could be sufficiently large. The potential optimization instability for classical RL can be used to explain its inferiority to distributional RL in most environments, although it may not explain why distributional RL could not perform favorably in certain games~\citep{ceron2021revisiting}. We leave the comprehensive explanation as future works.
\paragraph{Remark on Non-linear Categorical Parameterization.} Although the aforementioned stability optimization conclusions are established on the linear categorical parameterization on the value distribution of $Z^\pi$. Similar conclusions can be extended in the non-convex optimization case with a non-linear categorical parameterization by techniques proposed in \citep{hardt2016train}. We also empirically validate our theoretical conclusions in the experiments by directly applying practical neural network parameterized distributional RL algorithms.
\subsection{Acceleration Effect of distributional RL}\label{sec:acceleration}
To characterize the acceleration effect of distributional RL, we additionally leverage the recently proposed \textit{value distribution decomposition}~\citep{sun2021interpreting} to decompose the target $p^{s, a}$.
\paragraph{Value Distribution Decomposition.} In order to decompose the optimization impact of value distribution into its expectation and the remaining distribution part, we adopt the wisdom from robust statistics via a variant of \textit{gross error model}~\citep{huber2004robust}. Value distribution decomposition~\citep{sun2021interpreting} was successfully applied to derive the distributional regularization effect of distributional RL. We utilize $F^{s, a}$ to express the distribution function of $Z^\pi(s, a)$ and we consider the function class of $F^{s, a}$ that satisfies the following expectation decomposition:
\begin{equation}\begin{aligned}\label{eq:decomposition}
F^{s, a}(x)=(1-\epsilon)\mathbbm{1}_{\{x \ge \mathbb{E}\left[Z^\pi(s, a)\right]\}}(x) + \epsilon F^{s, a}_{\mu}(x),
\end{aligned}\end{equation}
where the distribution function $F^{s, a}_{\mu}$ is determined by $F^{s, a}$ and $\epsilon$ to measure the impact of remaining distribution \textit{independent of} its expectation $\mathbb{E}\left[Z^\pi(s, a)\right]$. $\epsilon$ controls the proportion of $F^{s, a}_\mu(x)$ and the indicator function $\mathbbm{1}_{\{x \ge \mathbb{E}\left[Z^\pi(s, a)\right]\}}=1$ if $x \ge \mathbb{E}\left[Z^\pi(s, a)\right]$, otherwise 0. Although the function class of $F^{s, a}$ is restricted to satisfy this decomposition equality, it is still rich with the rationale rigorously demonstrated in \citep{sun2021interpreting}. To reveal the speeding up effect of distributional RL loss, we consider the density function form of Eq.~\ref{eq:decomposition}, i.e., $p^{s, a}(x)=(1-\epsilon)\delta_{\{x=\mathbb{E}\left[Z^\pi(s, a)\right]\}}(x) + \epsilon \mu^{s, a}(x)$, where $\delta_{\{x=\mathbb{E}\left[Z^\pi(s, a)\right]\}}$ is a Dirac function centered at $\mathbb{E}\left[Z^\pi(s, a)\right]$ to characterize the expectation impact and $\mu^{s, a}$ is the density function of $F^{s, a}_\mu$ to measure the addition value distribution information. Within Neural FZI, our goal is to minimize $\frac{1}{n}\sum_{i=1}^{n}\mathcal{L}_\theta(s_i, a_i)$. We denote $G^k(\theta)$ as the expectation of $\mathcal{L}_\theta$, i.e., $G^k(\theta)=\mathbb{E}_{(s, a)\sim \rho^\pi}\left[\mathcal{L}_\theta(s, a)\right]$. Based on the categorical parameterization in Section~\ref{sec:stability}, the convex and smooth properties with respect to the parameter $\theta$ in $f_\theta$ as shown in Proposition~\ref{prop:lipschitz} still hold for $G^k(\theta)$. We use $G(\theta)$ for $G^k(\theta)$ for simplicity and rewrite $\mathcal{L}_\theta(s, a)$ as $\mathcal{L}_\theta(g^{s, a}, f^{s, a}_\theta)$, where the target density function $g^{s, a}$ can be $p^{s, a}$, $\mu^{s, a}$ or $\delta_{\{x=\mathbb{E}\left[Z^\pi(s, a)\right]\}}$, and $f^{s, a, \theta}$ is rewritten as $f^{s, a}_\theta$ for conciseness. As the KL divergence enjoys the property of unbiased gradient estimates, we let the variance of its stochastic gradient over \textit{the expectation} $\delta_{\{x=\mathbb{E}\left[Z^\pi(s, a)\right]\}}$ be bounded, i.e.,
\begin{equation}\begin{aligned}
\mathbb{E}_{(s, a)\sim \rho^\pi}\left[\|\nabla \mathcal{L}_\theta(\delta_{\{x=\mathbb{E}\left[Z^\pi(s, a)\right]\}}, f_\theta^{s, a}))-\nabla G(\theta)\|^{2}\right]=\sigma^{2}.
\end{aligned}\end{equation}
Next, following the similar label smoothing analysis in \citep{xu2020towards}, we further characterize the approximation degree of $f^{s, a}_\theta$ to the target value distribution $\mu^{s, a}$ by measuring its variance as $\kappa \sigma^2$:
\begin{equation}\begin{aligned}\label{eq:acceleration_kappa}
\mathbb{E}_{(s, a)\sim \rho^\pi}\left[\|\nabla \mathcal{L}_\theta(\mu^{s, a}, f_\theta^{s, a}))-\nabla G(\theta)\|^{2}\right]=\hat{\sigma}^2:=\kappa \sigma^{2}.
\end{aligned}\end{equation}
Notably, $\kappa$ can be used to measure the approximation error between $f_\theta^{s, a}$ and $\mu^{s, a}$ and we do not assume $\hat{\sigma}^2$ to be bounded as $\kappa$ can be arbitrarily large. This expression $\kappa \sigma^2$ for $\hat{\sigma}^2$ allows us to utilize $\kappa$ to characterize different acceleration effects for distributional RL given different $\kappa$. Concretely, a favorable approximation of $f_\theta^{s, a}$ to $\mu^{s, a}$ would lead to a small $\kappa$ that contributes to the acceleration effect of distributional RL as shown in Theorem~\ref{theorem:acceleration}. Based on Eq.~\ref{eq:acceleration_kappa}, we immediately have Proposition~\ref{prop:acceleration}.
\begin{prop}\label{prop:acceleration}
Based on the value distribution decomposition in Eq.~\ref{eq:decomposition}, and Eq.~\ref{eq:acceleration_kappa}, we have:
\begin{equation}\begin{aligned}\label{eq:acceleration_kappa_complete}
\mathbb{E}_{(s, a)\sim \rho^\pi}\left[\|\nabla \mathcal{L}_\theta(p^{s, a}, f_\theta^{s, a}))-\nabla G(\theta)\|^{2}\right] \le (1-\epsilon)^2\sigma^{2} + \epsilon^2 \kappa \sigma^{2}.
\end{aligned}\end{equation}
\end{prop}
Please refer to Appendix~\ref{appendix:acceleration_lemma} for the proof.
Before comparing the sample complexity in the optimization process of both classical and distributional RL, we provide the definition of the first-order $\tau$-stationary point, which is preferred in the optimization of deep learning rather than the a simple stationary point in order to guarantee the generalization.
\begin{myDef}\label{definition:acceleration} (First-order $\tau$-Stationary Point)
While solving $\min_\theta G(\theta)$, the updated parameters $\mathbb{\theta}_T$ after $T$ steps is a first-order $\tau$-stationary point if $\Vert \nabla G(\theta_T) \Vert \le \tau$, where the small $\tau$ is in $(0, 1)$.
\end{myDef}
Based on Definition~\ref{definition:acceleration}, we formally characterize the acceleration effects for distributional RL in Theorem~\ref{theorem:acceleration} that depends upon approximation errors between $\mu^{s, a}$ and $f^{s, a}_\theta$ measured by $\kappa$.
\begin{theorem}\label{theorem:acceleration} (Sample Complexity and Acceleration Effects of Distributional RL) While running SGD to minimize $\mathcal{L}_\theta$ in Eq.~\ref{eq:decomposition} within Neural FZI, we assume the step size $\lambda=1/kl^2$, $\epsilon=1/(1+\kappa)$ across (2) and (3), and the sample is uniformly drawn from $T$ samples, then:
(1) (\textbf{Classical RL}) When minimizing $\mathcal{L}_\theta(\delta_{\{x=\mathbb{E}\left[Z^\pi(s, a)\right]\}}, f_\theta^{s, a})$, $T = O(\frac{1}{\tau^4})$ such that $\mathcal{L}_\theta$ converges to a $\tau$-stationary point in expectation.
(2) (\textbf{Distributional RL with $\kappa \le \frac{\tau}{2\sigma}$}) When minimizing $\mathcal{L}_\theta(p^{s, a}, f_\theta^{s, a})$, let $T=\frac{4 G(\theta_0)}{\lambda \tau^2}=O(\frac{1}{\tau^2})$, $\mathcal{L}_\theta$ converges to a $\tau$-stationary point in expectation.
(3) (\textbf{Distributional RL with $\kappa > \frac{\tau}{2\sigma}$}) When minimizing $\mathcal{L}_\theta(p^{s, a}, f_\theta^{s, a})$, let $T=\frac{G(\theta_0)}{\lambda \kappa^2 \sigma^2}=O(\frac{1}{\tau^2})$, $\mathcal{L}_\theta$ does not converge to a $\tau$-stationary point, but can guarantee a $O(\kappa^2)$-stationary point.
\end{theorem}
The proof is provided in Appendix~\ref{appendix:acceleration_theorm}. Theorem~\ref{theorem:acceleration} is inspired by the intuitive connection between the value distribution in distributional RL and the label distribution in label smoothing technique~\citep{xu2020towards}. Importantly, Theorem~\ref{theorem:acceleration} demonstrates that solving categorical distributional loss of distributional RL can speed up the convergence if a distribution approximation error is favorable. Otherwise, the convergence point, albeit stationary, may not guarantee a desirable performance under an agnostic $\kappa$, which may be very large on certain environments.
\textit{The results in Theorem~\ref{theorem:acceleration} in fact incorporate the impact of both two key ingredients, including the choice of $d_p$ and parameterization error of $Z_\theta$, analyzed in Section~\ref{sec:neuralFZI} on the optimization of distributional RL.} \textbf{In the first scenario~((2) in Theorem~\ref{theorem:acceleration})}, there is only a small approximation or paramterization error between $f_\theta^{s, a}$ and $p^{s, a}$~(or $\mu^{s, a}$), corresponding to a small $\kappa$ with $\kappa\le \frac{\tau}{2\sigma}$. In this case, solving $\mathcal{L}_\theta$ based on the categorical parameterization can reduce the sample complexity from $O(\frac{1}{\tau^4})$ to $O(\frac{1}{\tau^2})$ compared with classical RL in (1) of Theorem~\ref{theorem:acceleration}, and meanwhile guarantees a $\tau$-stationary point. \textbf{In the second scenario~((3) in Theorem~\ref{theorem:acceleration})} especially for some challenging environments with much intrinsic uncertainty, we can also attain a relatively large approximation error or parameterization error of $Z_\theta$ with a large $\kappa > \frac{\tau}{2\sigma}$ as the distributional TD approximation error could be potentially large in practice. Under this circumstance, distributional RL algorithms may fail to speed up the convergence or achieve the superior performance compared with classical RL as $\mathcal{O}(\kappa^2)$ could be potentially large on some complex environments. If $\mathcal{O}(\kappa^2)$ is proper, distributional RL can still potentially perform reasonably due the to $\mathcal{O}(\kappa^2)$-stationary point guarantee.
Theses theoretical results also coincide with past empirical observations~\citep{dabney2017distributional, ceron2021revisiting}, where distributional RL algorithms outperform classical RL in most cases, but are inferior in certain environments. Based on our results in Theorem~\ref{theorem:acceleration}, we contend that these certain environments have much intrinsic uncertainty, the distribution parameterization error between $Z_\theta$ and the true value distribution under the distributional TD approximation is still too large~($\kappa > \frac{\tau}{2\sigma}$) to guarantee a favorable convergence point for distributional RL algorithms with different $d_p$, which is intuitive.
\section{Experiments}\label{sec:experiments}
\begin{figure*}[t!]
\centering
\includegraphics[width=1.0\textwidth,trim=0 0 0 0,clip]{0Figure_performance.pdf}
\caption{\textbf{Performance.} Learning curve of AC, DAC~(C51) and DAC~(IQN) over 5 seeds with smooth size 5 across eight MuJoCo games.}
\label{fig:performance}
\end{figure*}
We perform extensive experiments on eight continuous control MuJoCo games to validate the theoretical optimization advantage of distributional RL algorithms analyzed in Section~\ref{sec:optimization}, including the stable gradient behaviors of distributional RL to achieve the uniform stability as well as the acceleration effects determined by the distribution parameterization error.
\paragraph{Implementation.} Our implementation is based Soft Actor Critic~(SAC)~\citep{haarnoja2018soft} and distributional Soft Actor Critic~\citep{ma2020dsac}. We eliminate the optimization impact of entropy regularization in these algorithm implementations, and thus we denote the resulting algorithms as Actor Critic~(AC) and Distributional Actor Critic~(DAC) for the conciseness. For DAC, we firstly perform the C51 algorithm to the critic to extend the classical critic loss to the distributional version denoted by \textit{DAC~(C51)} as our theoretical analysis in Sections~\ref{sec:stability} and \ref{sec:acceleration} are mainly based on categorical parameterization. We further apply our empirical demonstration on Quantile Regression distributional RL heuristically, i.e., Implicit Quantile Network~(IQN), which is denoted as \textit{DAC~(IQN)}. Hyper-parameters and more implementation details are provided in Appendix~\ref{appendix:implementation}.
\begin{figure*}[b!]
\centering
\includegraphics[width=0.9\textwidth,trim=0 0 0 0,clip]{1Figure_gradient.pdf}
\caption{ \textbf{Stable Optimization.} The critic gradient norms in the logarithmic scale regarding \textit{the state} during the training of AC, DAC~(C51), DAC~(IQN) over 5 seeds on eight MuJoCo environments.}
\label{fig:optimization}
\end{figure*}
\subsection{Performance and Uniform Stability in distributional RL Optimization}
Figure~\ref{fig:performance} suggests that DAC~(IQN) in orange lines outperforms its classical version AC~(black lines) across all environments, while DAC~(C51) in red lines is inferior to AC on humanoid, walker2d and reacher. This could be explained by a more flexible parameterization of IQN over C51.
We then demonstrate the advantage of uniform optimization stability for distributional RL over classical RL. According to Theorem~\ref{theorem:lipschitz}, the stable optimization of distribution loss with Neural FZI is described as a bounded loss difference for a neighboring dataset in terms of each state $s$ and action $a$. In other words, the error bound holds by taking the supreme over each state the agent encounters. To measure this algorithm stability, while far from perfect, we consider to leverage \textit{the average gradient norms with respect to the state feature $\mathbf{x}(s)$} in the whole optimization process as the proxy due to the fact that the gradient could measure the sensitivity of loss function regarding each state the agent observes. From Figure~\ref{fig:optimization}, it turns out that both DAC~(C51) and DAC~(IQN) entail a much smaller gradient norm magnitude as opposed to their classical version AC~(black lines) across all eight MuJoCo environments, which corroborates the theoretical advantage of the uniform optimization stability for distributional RL analyzed in Theorem~\ref{theorem:lipschitz}.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.9\textwidth,trim=0 0 0 0,clip]{2Figure_acceleration.pdf}
\caption{ \textbf{Acceleration Effect.} The critic gradient norms in the logarithmic scale regarding \textit{network parameters} in the training of AC, DAC~(C51), DAC~(IQN) over 5 seeds on MuJoCo environments.}
\label{fig:acceleration}
\end{figure*}
\subsection{Smoothness Property and Acceleration Effect of Distributional RL}
Theorem~\ref{theorem:acceleration} demonstrates that distributional RL can speed up the convergence if the distribution parameterization is appropriate, characterized by the variance of the gradient estimates with a small $\kappa$ (case (2) in Theorem~\ref{theorem:acceleration}). To demonstrate it, we use the proxy by evaluating the $\ell_2$-norms of gradients \textit{with respect to network parameters} of the critic for AC and DAC. We mainly focus on a direct comparison between vanilla AC and DAC algorithm, although their network architectures are slightly different. Similar results under the same architecture and via the value distribution decomposition of Eq.~\ref{eq:decomposition} are provided in Appendix~\ref{appendix:exp_acceleration_kappa}.
Figure~\ref{fig:acceleration} showcases that both DAC~(C51) and DAC~(IQN) have smaller gradient norms in terms of network parameters $\theta$ compared with AC in the whole optimization process, which directly validates that distributional RL loss is more likely to enjoy smoothness properties in Proposition~\ref{prop:lipschitz}. In terms of acceleration effects, the property of stationary points, albeit being different, in cases (2) and (3) of Theorem~\ref{theorem:acceleration} guarantees bounded gradient norms, but the precise evaluation of $\kappa$ is tricky in order to discriminate either case (2) or (3) for each algorithm in a specific environment. Nevertheless, by considering the fact that DAC~(IQN) outperforms DAC~(C51) in most environments in Figure~\ref{fig:performance}, we hypothesize that the inferiority of DAC~(C51) on humanoid, walker2d and reacher could be owing to its larger parameterization errors $\kappa$ in these environments. This results in the worse performance of DAC~(C51) compared with DAC~(IQN) that is more likely to accord with the case (3) in Theorem~\ref{theorem:acceleration} due to its richer distribution expressiveness power than C51.
\section{Discussions and Conclusion}
Our optimization analysis of distributional RL is based on the categorical parameterization, and the alternative analysis on Wasserstein distance can be an integral complementary for our conclusions. Acceleration effects could be further investigated to explain whether a typical distributional RL algorithm can perform favorably in a specific environment. We leave them as future works.
In our paper, we answer the question: \textit{how does value distribution in distributional RL help the optimization} from two perspectives, including the stable optimization analysis based on the smoothness property of categorical distributional loss, as well as the acceleration effects determined by the variance of gradient estimates. We theoretically and empirically show that distributional RL embraces stable gradient behaviors and could speed up the convergence if the distribution approximation is desirable or the parameterization error is sufficiently small.
\paragraph{Ethics Statement.} Due to the fact that our study is about the theoretical properties of distributional RL algorithms, we do not think our research is involved with any ethics issues.
\paragraph{Reproducibility Statement.} As stated in Section~\ref{sec:experiments}, our implementation is based on the public code of SAC~\citep{haarnoja2018soft} and Distributional SAC~\citep{ma2020dsac}. We also provide implementation details in Appendix~\ref{appendix:implementation} for reproducibility. For the theoretical results, rigorous proof is also given in Appendix from \ref{appendix:histogram} to \ref{appendix:acceleration_theorm}.
| 5bb2baaf00bcf43326f04272db503b7ba252898a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Galactic scale flows (both inflowing and outflowing) are an essential ingredient in models of galaxy formation. In particular, flows are responsible for regulating the baryon cycle and are necessary to reproduce observations on both galactic and cosmic scales. A simple, but useful explanation is that flows work in equilibrium. For example, supernovae or AGN driven feedback is necessary to expel matter from the interstellar medium (ISM) to enrich the circumgalactic medium (CGM, \citealt{Tumlinson2017}) and intergalactic medium (IGM, \citealt{Oppenheimer2006}), and in doing so, actively suppresses star formation; consequently, accretion is then necessary to refuel star formation to - for example - recreate the observed time evolution of the star formation rate (e.g., \citealt{Erb2008,Mannucci2010,Papovich2011}). In this paradigm, knowing the physical properties of flows is essential for testing models of galaxy formation against observations. Indeed, flow parameters are often introduced into hydrodynamical simulations in an ad hoc fashion to incorporate the physics of star formation which occurs on unresolved scales (e.g., \citealt{Springel2005,Vogelsberger2014}). Furthermore, simulations which rely on self consistent models of star formation to derive flows (e.g., \citealt{Hopkins2014,Hopkins2018}) need to be checked as well to ensure their results agree with observations. For a review on the status of galaxy formation models which emphasizes the importance of observationally constraining flows see \cite{Somerville2015}.
Observationally, flows are identified by the asymmetric absorption and emission line broadening observed in their spectra where the width reflects the maximum velocity of the flow and the depth reflects the optical depth of the relevant transition \citep{Prochaska2011,Scarlata2015,Carr2018}. Over the past decade, absorption and emission line studies have made tremendous progress constraining the properties of flows across multiple wavelengths. For example, local UV studies relying on the Hubble Space Telescope's Cosmic Origins Spectrograph (COS) have successfully constrained the cool and warm phases of the CGM from low and high ionization state metal lines \citep{Stocke2013,Werk2014,Peeples2014,Heckman2017,Carr2021}, while sub-milimeter studies with the Herschel Space Observatory's PACS instrument have constrained the cold phase of the CGM via atomic and molecular lines \citep{Gonzalez-Alfonso2012,Falstad2015,Herrera-Camus2020}. Ground based studies in the optical with Keck and milimeter with ALMA have placed constraints on flows at even higher redshifts \citep{Martin2012,Rubin2014,Gallerani2018}. The hot phase of the CGM is the least well constrained due to the difficulty in observing the faint x-ray emission produced at such high temperatures ($\sim 10^7$ K); however, evidence for outflows has been observed in the x-ray emission of nearby galaxies \citep{Strickland1997,Laha2018}. Together these studies reveal that outflows are ubiquitous, multiphase, and energetic reaching speeds on the order of hundreds to a few thousand kilometers per second. In contrast, inflows are less energetic, typically reaching speeds less than 200 kilometers per second, and are observed far less frequently. For example, in a study of $\sim 200$ galaxies from $0.4 < z < 1.4$, \cite{Martin2012} found evidence for inflows in only $3-6\%$ of their sample compared to the roughly $20\%$ detection rate for outflows in the absorption lines of Fe II (i.e., partially ionized gas $\sim10^4$K) under similar restrictions on the net observed velocity. For a summary on the current state of observations of both inflows and outflows see \cite{Roberts-Borsani2019} and \cite{Veilleux2020}.
Given the important role accretion plays in models of galaxy formation, the overall lack of detections of inflows is surprising. Indeed, hydrodynamical simulations predict that a significant amount of the gas feeding star formation at low redshifts is metal enriched \citep{Ford2014,Angles-Alcazar2017}. For example, using the FIRE cosmological simulations, \cite{Angles-Alcazar2017} found that all galaxies in their simulations re-accrete $>50\%$ of the gas ejected in winds. A possible explanation for the lack of detections is that galactic inflows occur along streams with low covering fractions, and hence, whether or not inflowing gas is detected will be highly sensitive to the viewing angle \citep{Kimm2011}. Furthermore, the spectral features of the more energetic outflows may mask that of inflows \citep{Martin2012,Rubin2014}. For example, the symmetric emission component of an outflow may fill in the weaker redshifted absorption feature necessary to identify inflows in resonant lines.
The observational approach to study the CGM (i.e., background vs down the barrel spectroscopy) matters in terms of what information can be recovered from flows and how efficiently. In the case of background spectroscopy, where light from a distant quasar is used to probe the CGM of a single galaxy, the orientation of the flow cannot be determined from the absorption lines alone. When this is the case, the galaxy's morphology and orientation can be used to predict the direction of the flow; however, these analyses suffer from model degeneracies and favor certain orientations for constraining flow properties (e.g., \citealt{Ho2020}). Down-the-barrel spectroscopy, on the other hand, has the potential to unambiguously identify the orientation of a flow and has been used to study the spectra of outflows (e.g., \citealt{Rivera-Thorsen2015,Zhu2015,Chisholm2017a,Gazagnes2018,Carr2021}) and to estimate mass outflow rates (e.g., \citealt{Chisholm2016,Chisholm2017b,Xu2022}). There is, however, very little modeling available to study the spectra of inflows in this area. The lack of modeling has been recognized (see \citealt{Faucher2017}), and any progress made would be a step forward.
In this paper, we present a modified version of the semi-analytical line transfer (SALT) model of \cite{Scarlata2015} to predict the spectra of galactic inflows obtained via down the barrel spectroscopy. The model pertains only to the scattering of continuum radiation, and does not include emission from the inflows themselves. As such, it cannot be used to interpret - for example - radio observations, but can be used to interpret the spectra of cool partially ionized gas where collisional excitation is expected to be low (e.g., \citealt{Rubin2012}). We plan to consider an emission component in a future paper.
The rest of this paper is organized as follows. In section 2, we adapt the SALT model of \cite{Scarlata2015} to account for galactic inflows. In section 3, we explore the parameter space of the inflowing SALT model in various physically motivated scenarios including spherical inflows, inflows with covering fractions less than unity, and galactic fountains. Our discussion and conclusions follow in section 4.
\section{Modeling}
\label{Modeling}
This paper is the third in a series of theoretical papers aimed at calculating the spectra of galactic flows. In the first paper, \cite{Scarlata2015} presented a semi-analytical line transfer (SALT) model to predict the spectra of spherical galactic outflows. Their work was later generalized by \cite{Carr2018} to include spectra of bi-conical outflows. The aim of the current paper is to adapt SALT to predict the spectra of galactic inflows (both spherical and with covering fractions less than unity). This model is intended to interpret the inverse P Cygni profiles observed in the UV lines associated with the low ionization metals (e.g., Mg II, Si II, etc.) thought to trace cool inflows during metal recycling (e.g., \citealt{Ford2014}).
For consistency, we preserve what notation we can from previous works. All models are to assume a spherical source of isotropically emitted radiation of radius, $R_{SF}$, surrounded by an envelope of flowing material which extends to a terminal radius, $R_W$. A diagram has been provided in Figure~\ref{f1}. The $\xi$-axis runs perpendicular to the line of sight and is measured using normalized units, i.e., $\xi=r/R_{SF}$. The $s$-axis runs parallel with the line of sight and is measured using the same normalized units.
For convenience, we have provided a complete list of the SALT parameters for the outflowing and inflowing SALT model along with their definitions in Table~\ref{tab:pars}. More detailed definitions can be found in \cite{Carr2018,Carr2021}.
\subsection{SALT Model for Galactic Inflows}
\begin{figure}
\centering
\includegraphics[scale=.4]{f1.png}
\caption{Envelope of material (shown in blue) of radius $R_W$ surrounding a galaxy (shown in green) of radius $R_{SF}$. The $\xi$ and $s$-axes are written in normalized units $(r/R_{SF})$. The envelope represents a flow characterized by a density and velocity field.}
\label{f1}
\end{figure}
We begin by assuming the velocity field of the in-falling gas follows a power law of the form,
\begin{equation}
\begin{aligned}
v &= v_{\infty} - v_0\left(\frac{r}{R_{SF}}\right)^{\gamma}&&\text{for}\ r < R_{W} \\[1em]
v &= 0 &&\text{for} \ r \geq R_{W}, \\
\end{aligned}
\end{equation}
where $v = v_{\infty}-v_0$ represents the velocity at $R_{SF}$ and $\gamma$ takes positive values. To ensure an inflow, we set the orientation of the velocity field to be positive along the line of sight when positioned between the observer and the source (i.e., opposite the orientation of the outflowing case). This velocity field assumes that accretion begins from rest at some finite distance, $R_W$, away from the galaxy. Such a situation is suitable for metal enriched gas that has been thrown out of the galaxy, comes to a halt, and falls back in. Aside from this insight, the exact expression for the velocity field was chosen for analytical simplicity, and we direct the reader to \citealt{Fielding2022} for a discussion on the physics influencing the velocity fields of flows. We assume an arbitrary density field, $n(r) = n_0(r/R_{SF})^{-\delta}$, for the in-falling medium. The Sobolev approximation is still valid in this context \citep{Lamers1999}, and we assume photons can only interact locally with the inflow at a single point. Staying consistent with prior works, we will use the normalized units: $y = v/v_0$ and $x = v \cos{\theta}/v_0$ where $v\cos{\theta}$ is the projection of $v$ onto the line of sight.
Following the derivation of the outflowing SALT model in Carr et al. (2022, in prep), we seek to identify the surfaces of constant observed velocity, $\Omega_x$, in the inflow; however, due to radial symmetry, it will suffice to construct $\Gamma_x$ (i.e., the intersection of $\Omega_x$ with the $s\xi$-plane). Following the derivation of \cite{Carr2018}, we compute
\begin{eqnarray}
\Gamma_x(y) &&\resizebox{0.38\textwidth}{!}{$= \left([y_{\infty}-y]^{1/\gamma}\left(\frac{x}{y}\right),\left[(y_{\infty}-y)^{2/\gamma}-(y_{\infty}-y)^{2/\gamma}\left(\frac{x}{y}\right)^2\right]^{1/2}\right)$}\\
&&= (S,\Xi),
\end{eqnarray}
\noindent where $S(y)$ and $\Xi(y)$ refer to the parameterizations of the $s$ and $\xi$ coordinates of $\Gamma_x$, respectively. Several different examples of $\Gamma_x$ are shown in Figure~\ref{f2} for different values of $x$ and $\gamma$. An interesting difference between these curves and those of an outflowing velocity field\footnote{Note this claim is making the additional assumption that the outflow is accelerating. For decelerating velocity fields, $\Gamma_x$ displays the same behavior as inflowing velocity fields.} (see \citealt{Carr2018}), is that all curves for an inflowing velocity field trace back to the source (i.e., towards regions of high density). As we will demonstrate, this is one of the reasons why the line profile for an inflow gives the appearance of an 'inverse' P Cygni profile.
\begin{figure}[]
\centering
\includegraphics[scale=0.48]{f2.png}
\caption{Various curves of constant observed velocity for inflowing velocity fields. The curves drawn with the darkest, middle, and lightest shades correspond to velocity fields with $\gamma = 2.0,1.0$, and $0.5$, respectively. The different colors represent different observed velocities.}
\label{f2}
\end{figure}
Observe that rays emitted parallel to the line of sight can intersect a surface of constant observed velocity once, twice, or zero times. An example of a surface for which all three scenarios are possible is shown in Figure~\ref{f3}. This issue does not occur in an outflowing velocity field and requires special attention. Consider a continuum ray emitted at a height, $h$, above the $s$-axis. To determine if this ray will pass through the surface of constant observed velocity, $\Omega_x$, we must first find the maximum height, $\Xi(y_{max})$, reached by $\Gamma_x$ in the $s\xi$-plane. We have
\begin{eqnarray}
\resizebox{0.41\textwidth}{!}{$\frac{d\Xi}{dy} = \frac{\gamma x^2y^{-3}[y_{\infty}-y]^{2/\gamma}+(x^2y^{-2}-1)[y_{\infty}-y]^{(2-\gamma)/\gamma}}{-\Xi([y_{\infty}-y]^{(1-\gamma)/\gamma}xy^{-1}+\gamma[y_{\infty}-y]^{1/\gamma}xy^{-2})}$,}
\end{eqnarray}
and after minimizing we deduce that
\begin{eqnarray}
0 &=& y^3+(\gamma-1)x^2y-\gamma y_{\infty}x^2,
\end{eqnarray}
which has the real valued solution
\begin{eqnarray}
\resizebox{0.41\textwidth}{!}{$y_{\rm{max}} = \left(-\frac{q}{2}+\left(\frac{q^2}{4}+\frac{p^3}{27}\right)^{1/2}\right)^{1/3}+\left(-\frac{q}{2}-\left(\frac{q^2}{4}+\frac{p^3}{27}\right)^{1/2}\right)^{1/3}$,}
\end{eqnarray}
where $p = (\gamma-1)x^2$ and $q = -\gamma x^2y_{\infty}$. Thus, if $\Xi(y_{max}) < h$, then the continuum ray will miss $\Omega_x$. A shell with intrinsic velocity $y_{max}$ which intersects $\Omega_x$ at its maximum is shown in red in Figure~\ref{f3}.
\begin{figure}[]
\centering
\includegraphics[scale=0.34]{f3.png}
\caption{\emph{Left} Crosssectional view of a spherical inflow. Continuum rays are drawn in purple and are shown passing through a surface of constant observed velocity whose curve of constant observed velocity, $\Gamma_x$, is shown in blue. Note that the rays can either intersect $\Gamma_x$ once, twice, or zero times. For the continuum ray which intersects $\Gamma_x$ twice, the shells of intrinsic velocities $y_{h_1}$ and $y_{h_2}$ which pass through the intersections are shown in black. A shell of intrinsic velocity, $y_{max}$, intersecting the maximum of $\Gamma_x$, is shown in red. \emph{Right} A close up view of the intersections. The maximum of $\Gamma_x$ as well as the points of intersection with the continuum rays provide enough information to compute the absorption profile as described in the text. }
\label{f3}
\end{figure}
If $\Xi(y_{\infty}-1) < h < \Xi(y_{max})$, then the continuum ray will intersect $\Omega_x$ twice. Setting $\Xi = h$, we deduce the following relation,
\begin{eqnarray}
\frac{h^2}{(y_{\infty}-y)^{2/\gamma}}=1-\left(\frac{x}{y}\right)^2,
\label{intersections}
\end{eqnarray}
which has solutions $y_{h_1}$ and $y_{h_2}$ corresponding to the intrinsic velocities of the shells intersecting $\Gamma_x$ at height $h$. Shells with intrinsic velocities $y_{h_1}$ and $y_{h_2}$ for a fixed value of $h$ are shown in black in Figure~\ref{f3}. Finally, if $h < \Xi(y_{\infty}-1)$ the continuum ray will intersect $\Omega_x$ only once.
Putting it all together, we compute the absorption spectrum as follows. We define the integrand, $\rm{ABS}$, as
\begin{eqnarray}
\rm{ABS} &=& 2 \pi h (1-e^{-\bar{\tau}_S})|dh/dy|/\pi \\
&=&\resizebox{0.35\textwidth}{!}{$\frac{2}{\gamma}\left[\frac{x^2}{y^2}(y_{\infty}-y)^{\frac{2-\gamma}{\gamma}}+\frac{\gamma x^2}{y^3}(y_{\infty}-y)^{\frac{2}{\gamma}}-(y_{\infty}-y)^{\frac{2-\gamma}{\gamma}}\right]$}\nonumber\\
&\times& (1-e^{-\bar{\tau}_S(y)}),
\label{SF_abs1}
\end{eqnarray}
where $\bar{\tau}_S(y)$ is defined over the following cases.
\begin{figure*}[]
\centering
\includegraphics[scale=0.58]{f4.png}
\caption{\textbf{\emph{Left}} The absorption profile, \textbf{\emph{Middle}} emission profile, and \textbf{\emph{Right}} Inverse P Cygni profile for a spherical inflow. The asymmetry in the emission profile reflects the occultation effect where photons emitted from behind the source are blocked by the source from the observer's field of view. The parameters for this inflow are listed on the left most panel.}
\label{f4}
\end{figure*}
\begin{gather}
\bar{\tau}_S \equiv
\begin{cases}
\tau_S(y_{h_1})+\tau_S(y_{h_2})&\ \rm{if} \ \Xi(y_{\infty}-1)\leq \Xi(y)< \Xi(y_{\rm{max}}) \\[1em]
\tau_S(y) & \ \rm{otherwise}\\
\end{cases}
\label{SF_abs2}
\end{gather}
and $\tau_S$ is the Sobolev optical depth \citep{Lamers1999} defined as
\begin{eqnarray}
\tau_S &=& \frac{\pi e^2}{mc} f_{lu}\lambda_{lu} n_l(r)\left[1-\frac{n_ug_l}{n_lg_u}\right] \frac{r/v}{1+\sigma {\cos^2{\phi}}},
\label{Castor}
\end{eqnarray}
where $f_{ul}$ and $\lambda_{ul}$ are the oscillator strength and wavelength, respectively, for the $ul$ transition, $\sigma = \frac{d \ln(v)}{d\ln(r)} - 1$, and $\phi$ is the angle between the velocity and incoming photon ray \citep{Castor1970}. All other quantities take their usual definition. Writing Equation~\ref{Castor} in terms relevant to the SALT model, and by neglecting stimulated emission (i.e., $\left[1-\frac{n_ug_l}{n_lg_u}\right] =1$), we get
\begin{eqnarray}
\tau_S(y) = \tau_0 \frac{(y_{\infty}-y)^{(1-\delta)/\gamma}}{y+[\gamma(y_{\infty}-y)-y](x/y)^2]},
\label{inflow_tau}
\end{eqnarray}
where
\begin{eqnarray}
\tau_0 = \frac{\pi e^2}{mc} f_{lu}\lambda_{lu} n_0 \frac{R_{SF}}{v_0}.
\end{eqnarray}
Note that we are assuming that the population of the excited state is negligible ($n_u\sim0$), and are therefore excluding the possibility of emission from the inflow and are focusing exclusively on continuum scattering. We will consider an emission component in a future paper and extend the range of our model to include radio observations and the interpretation of the inverse P Cygni profiles frequently observed in the spectra of cool atomic and molecular inflows (e.g., \citealt{Gonzalez-Alfonso2012,Falstad2015,Herrera-Camus2020}).
Finally, if $\Xi(y_{max})\leq1$, then the normalized absorption profile becomes:
\begin{eqnarray}
\resizebox{0.4\textwidth}{!}{$
\frac{I_{\rm{abs,red}}}{I_0} = \frac{F(x)}{F_c(x)} - \frac{F(x)}{F_c(x)}\int_{x}^{\rm{min}(y_{max},y_{\infty}-1)}\rm{ABS}(y) \ dy,$}
\end{eqnarray}
where $F(x)/F_c(x)$ is the flux at $x$ normalized by the continuum flux. If $\Xi(y_{max})>1$, then the normalized absorption profile becomes
\begin{eqnarray}
\frac{I_{\rm{abs,red}}}{I_0}= \frac{F(x)}{F_c(x)} - \frac{F(x)}{F_c(x)} \int_{x}^{y_h}\rm{ABS}(y) \ dy,
\end{eqnarray}
where $y_h = \rm{min}(y_{h_2}(1),y_{h_1}(1))$, and $y_{h_2}(1)$ and $y_{h_1}(1)$ are the solutions of equation~\ref{intersections} for $h=1$.
We can compute the emission component of the inflowing SALT model using the same general procedure used in the outflowing case (see Carr et al. 2022, in prep). To this end, we think of $\Gamma_x(y)$ as a homotopy (i.e., a continuous map between surfaces of constant observed velocity), and compute the amount of energy absorbed by each surface that lies within the shell as we move in observed velocity space from $x = y\cos{\Theta_C}$ to $x = y$ where $\Theta_C = \arcsin{\left([y_{\infty}-y]^{-1/\gamma}\right)}$. Similar to the absorption case, we must be careful to identify which rays intersect $\Omega_x(y)$ before reaching the shell. With this in mind, we compute the total energy absorbed by a shell at velocity $y$ in the inflow as
\begin{eqnarray}
&&\hspace*{-1cm}L_{\rm{shell}} = \nonumber \\
&&\hspace*{-1cm}\resizebox{0.43\textwidth}{!}{$ \int_{y\cos{\Theta_C}}^{y}\frac{2L(x)}{\gamma}\left[\frac{x^2}{y^2}(y_{\infty}-y)^{\frac{2-\gamma}{\gamma}}+\frac{\gamma x^2}{y^3}(y_{\infty}-y)^{\frac{2}{\gamma}}-(y_{\infty}-y)^{\frac{2-\gamma}{\gamma}}\right] \rm{SF} dx,$}
\end{eqnarray}
where $L(x)$ is the total energy emitted by the shell at resonance with material moving at observed velocity, $x$. We refer to SF as the Sobolev factor and define it for the following cases. If $y \geq y_{max}$, then
\begin{eqnarray}
\rm{SF} \equiv -[1-e^{-\tau_S(y)}];
\label{neg_change}
\end{eqnarray}
otherwise,
\begin{gather}
\resizebox{0.48\textwidth}{!}{$
\rm{SF} \equiv
\begin{cases}
e^{-\tau_S(y_{h_2})}[1-e^{-\tau_S(y_{h_1})}]& \ \rm{if} \ \Xi(y_{\infty}-1)\leq \Xi(y)< \Xi(y_{\rm{max}}) \\[1em]
1-e^{-\tau_S(y)} & \ \rm{otherwise,}\\
\end{cases}
$}
\label{SF_em}
\end{gather}
where $\tau_S$ is given by Equation~\ref{inflow_tau}. The leading negative sign in Equation~\ref{neg_change} accounts for the sign change in $dh$ when $y>y_{max}$. Now that we know the total energy absorbed by the shell, the reemission of energy in terms of the observed velocities can be computed exactly as in the outflowing case, albeit for a change in orientation (see Carr et al. 2022, in prep). We compute the normalized red emission profile as
\begin{eqnarray}
&&\hspace*{-.9cm}I_{\rm{em,red}}/I_0 = \nonumber \\
&&\hspace*{-.9cm}\resizebox{0.43\textwidth}{!}{$\int_{x}^{y_{\infty}-1} \frac{F(x^{\prime})}{F(x)}\frac{dy}{2y} \int_{y\cos{\Theta_C}}^{y}\frac{2}{\gamma}\left[\frac{{x^{\prime}}^2}{y^2}(y_{\infty}-y)^{\frac{2-\gamma}{\gamma}}+\frac{\gamma {x^{\prime}}^2}{y^3}(y_{\infty}-y)^{\frac{2}{\gamma}}-(y_{\infty}-y)^{\frac{2-\gamma}{\gamma}}\right] \rm{SF} dx^{\prime},$}
\end{eqnarray}
where $SF$ is defined for the cases given above. Similar to the outflowing case, we account for the occultation effect on the blue emission profile by removing all photons reemitted from behind the source at heights $h = \Xi < 1$. To this end, we define the scale factor $\Theta_{\rm{blue}} \equiv \Theta(\Xi-1)$ such that
\begin{gather}
\Theta \equiv
\begin{cases}
1 & \rm{if \ } \ \Xi > 1 \\[1em]
0 & \rm otherwise.\\
\end{cases}
\end{gather}
The normalized blue emission profile becomes
\begin{eqnarray}
&&\hspace*{-.9cm}I_{\rm{em,blue}}/I_0 = \nonumber \\
&&\hspace*{-.9cm}\resizebox{0.43\textwidth}{!}{$\int_{x}^{y_{\infty}-1} \Theta_{\rm{blue}}\frac{F(x^{\prime})}{F(x)}\frac{dy}{2y} \int_{y\cos{\Theta_C}}^{y}\frac{2}{\gamma}\left[\frac{{x^{\prime}}^2}{y^2}(y_{\infty}-y)^{\frac{2-\gamma}{\gamma}}+\frac{\gamma {x^{\prime}}^2}{y^3}(y_{\infty}-y)^{\frac{2}{\gamma}}-(y_{\infty}-y)^{\frac{2-\gamma}{\gamma}}\right] \rm{SF} dx^{\prime}.$}
\end{eqnarray}
The technique developed by \cite{Carr2018} to account for the effects of a biconical outflow geometry still holds and can be applied to the line profile in the obvious way; however, rays which can interact with a surface of constant observed velocity more than once must be handled carefully since the geometry of the flow may be different at each interaction. When this is the case, we can approximate the expression for $SF$ in Equation~\ref{SF_em} as
\begin{eqnarray}
f_g(y_{h_1})[I(y_{h_2})-I(y_{h_2})e^{-\tau_S(y_{h_1})}],
\label{new_g}
\end{eqnarray}
where
\begin{eqnarray}
I = 1-f_g(y_{h_2})[1-e^{-\tau_S(y_{h_2})}],
\end{eqnarray}
and $f_g$ is the geometric factor as defined by \cite{Carr2018}. Likewise, the Sobolev factor defined in Equations~\ref{SF_abs1} \& \ref{SF_abs2} will also need to change to the expression in Equation~\ref{new_g}, but with the addition of the term, $f_g(y_{h_2})[1-e^{-\tau_S(y_{h_2})}]$. The techniques developed by \cite{Carr2018,Carr2021} to account for a dusty CGM, a limiting observing aperture, and holes in the outflow still hold and can be applied in the obvious way. Lastly, the multiple scattering procedure of \cite{Scarlata2015} used to account for resonant and fluorescent reemission still applies.
The absorption, emission, and full line profile or inverse P Cygni profile for a spherical inflow are shown in Figure~\ref{f4}. The most striking difference when compared to the traditional line profiles of outflows is that absorption now occurs to the right of systemic velocity or at positive observed velocities using our orientation. The emission profile is now more sharply peaked with a wider base compared to that of an outflow. This is because, in our model, the highest density material near the source is moving at the highest velocities - opposite of the outflowing case - thus the majority of energy is now spread over a larger observed velocity range; however, absorption (and reemission) can still occur in shells with low intrinsic velocities all the way to zero intrinsic velocity (outflows are cut off at the launch velocity) and accumulate emission into a sharp peak.
\section{Results}
In this section, we explore the parameter space of the inflowing SALT model by generating spectra for a variety of distinct physically motivated scenarios for the emergence of inflows in galactic systems. In other words, we explore the range of applicability for detecting galactic inflows using down-the-barrel spectroscopy - an area where modeling has been especially lacking (see \citealt{Faucher2017}). We consider the case of pure inflows - both spherical and with partial covering fractions - as well as galactic fountains. In our fiducial model, we assume $v_{\infty} = 200 \ \rm{km} \ {s}^{-1}$ and a column density ranging from $14 \leq \log{(N \ [\rm{cm^{-2}}])} \leq 15$. These values were chosen to be in agreement with \cite{Roberts-Borsani2019} and \cite{Martin2012}. For the remaining parameters, we assume $\gamma = 0.5$, $\delta = 1.5$, $\kappa = 0.0$ (i.e., dust free), $v_0 = 15 \ \rm{km} \ {s}^{-1}$, and $R_{AP} > R_W$ (i.e., flow is fully captured by the observing aperture) while $\alpha$, $\psi$, and $f_c$ vary. For the chosen velocity field, a value of $\gamma \leq 1$ ensures that the acceleration ($\sim vdv/dr$) of the inflow increases monotonically with decreasing radius, and a value of $\delta < 2$ ensures that the mass inflow rate ($\sim r^2nv$) will start to decrease with decreasing radius some finite distance from the source. The latter can be expected since material is likely to be stripped from the flow during infall, and the ionizing flux should increase as the material moves closer to the galaxy thereby increasing the ionization state.
\subsection{Pure Inflows}
By far the most convincing evidence for inflows in absorption line studies is the direct detection of inverse P Cygni profiles (e.g., \citealt{Rupke2021}). These line profiles are characterized by having an overall net blueshift in emission and an overall net redshift in absorption. The latter of which is a stark quality unique to the spectra of inflowing gas; however, to declare the detection of an inflow, the redshifted absorption must reach observed velocities exceeding the limitations of the spectral resolution of the study. Furthermore, the absorption must be asymmetric with respect to line center (e.g., \citealt{Martin2012}) to distinguish the inflow from possible ISM absorption or thermal/turbulent line broadening (Carr et al. 2022, in prep). Here we investigate which parameter ranges favor detections of inflows in absorption for cases of pure spherical inflows and inflows with covering fractions less than unity.
\begin{figure*}[]
\centering
\includegraphics[scale=0.5]{f5.png}
\caption{\textbf{\emph{Left Column}} The absorption profile (top), emission profile (middle), and Inverse P Cygni profile (bottom) of a spherical inflow for various values of $\delta$ with $\gamma = 0.5$. \textbf{\emph{Right Column}} The same as the left column except $\gamma$ varies and $\delta = 1.5$. The remaining parameters for these profiles are $\alpha = 90^{\circ}$, $\psi = 0^{\circ}$, $\log{(N \ [\rm{cm}^{-2}])}=15$, $v_0 = 15 \ \rm{km \ s^{-1}}$, $v_w = 200 \ \rm{km \ s^{-1}}$, $R_{AP} > R_W$, $\kappa = 0$, and $f_c = 1.0$. }
\label{f5}
\end{figure*}
\begin{figure*}[]
\centering
\includegraphics[scale=0.55]{f6.png}
\caption{\textbf{ \emph{Left Panel}} Line profiles for biconical inflows of various opening angles oriented along the line of sight (i.e., $\psi = 0^{\circ}$). \textbf{\emph{Right Panel}} Line profiles for biconical inflows with opening angle, $\alpha = 30^{\circ}$, at oriented at various angles with respect to the line of sight. The remaining parameters for these profiles are $\gamma = 0.5$, $\delta = 1.5$, $\log{(N \ [\rm{cm}^{-2}])}=15$, $v_0 = 15 \ \rm{km \ s^{-1}}$, $v_w = 200 \ \rm{km \ s^{-1}}$, $R_{AP} > R_W$, $\kappa = 0$, and $f_c = 1.0$. }
\label{f6}
\end{figure*}
We examine the effects of varying the power laws of both the density field (left column) and velocity field (right column) on the line profile of a spherical inflow in Figure~\ref{f5}. We have chosen to model the Fe II $2382$\AA \ resonant transition which was used in the study by \cite{Martin2012} to identify inflows and outflows in galaxies at redshifts $0.4<z<1.4$ as an example. All atomic data for the lines used in this paper is provided in Table~\ref{tab:atomicdata}. In each column, we show how the absorption (top panel), emission (middle panel), and full inverse P Cygni profile (bottom panel) vary as functions of $\delta$ (left column) and $\gamma$ (right column) for fixed column density, $\log{(N \ [\rm{cm}^{-2}])} = 15$. In regards to density, the amount of absorption at low observed velocities increases with decreasing $\delta$. This is because for steep density fields, most absorption is happening near the source at high observed velocities. When the density field becomes shallower, however, more absorption can occur at larger radii or at lower observed velocities. The spectrum varies significantly less over our chosen range for $\gamma$, but absorption does appear to increases at lower observed velocities for increasing $\gamma$. This occurs because the size of the inflow (i.e., $R_W = R_{SF}(v_{\infty}/v_0)^{1/\gamma}$) increases with decreasing $\gamma$. Therefore, the bulk of material at low velocity is moving further away from the galaxy at lower density. We can conclude that steeper density and shallower velocity fields yield better conditions for identifying inflows. Under these conditions, most absorption will occur near the maximum observed velocity of the flow creating a more asymmetric (with regards to systemic velocity) absorption profile.
In Figure~\ref{f6}, we break the assumption of spherical symmetry and consider biconical inflows. These are the same geometrical models used by \cite{Carr2018}, but are now inflowing. We mention that this is only an example, and note that the SALT model can just as easily be used in the case of a single cone geometry. In the left panel of Figure~\ref{f6}, we vary the opening angle, $\alpha$, of an inflow oriented along the line of sight. As is the case with outflows (see \citealt{Carr2018}), there is less reemission for smaller opening angles. This includes reducing the amount of red emission infilling\footnote{This is an analogous effect to the blue emission infilling which is typically associated with outflows.}, which enhances the increase of absorption for smaller opening angles. The opposite effect occurs when increasing the orientation angle, $\psi$, which is shown for an inflow with an opening angle $\alpha = 30^{\circ}$ in the right panel of Figure~\ref{f6}. As $\psi$ increases, the inflowing gas is moved away from the line of sight, which results in less absorption since there is less material between the observer and the source. Coincidently, more and more emission appears near systemic or zero observed velocity, reflecting the position of the inflow with respect to the observer. The importance of this example is to demonstrate that when spherical symmetry is broken either the absorption or emission equivalent width can dominate based on the orientation of the inflow with respect to the observer. In regards to detecting inflows, a red shifted absorption profile is still a dead give away for inflowing gas; however, a pure emission profile could be mistaken for outflowing gas. The only difference between emission profiles that arise from outflows and inflows, is that, in the case of inflows, they are much more narrow peaked with wider bases\footnote{Emission profiles with wider bases and narrow peaks are also characteristic of outflows with decelerating velocity fields.}. In this instance, fitting these lines with a single Gaussian - as is typically done in more traditional modeling techniques - may be difficult. More than likely fitting these emission lines would require two Gaussians (one for the narrow component near systemic velocity and the second for the wider base) and be misinterpreted as representing two kinematically distinct components. Thus the necessity to fit emission profiles with more than one Gaussian could be evidence for inflowing gas.
\subsection{Outflow and Inflows: Galactic Fountains}
\begin{figure*}[]
\centering
\includegraphics[scale=0.8]{f7.png}
\caption{Orientation effects on galactic systems with inflows and outflows. All diagrams are to be viewed from the right. Inflowing gas is shown in light blue while outflowing gas is shown in salmon. The would be spectrum for each flow is shaded in matching color. The composite spectrum is shown at a resolution of $50 \ \rm{km} \ \rm{s}^{-1}$ (black) and $120 \ \rm{km} \ \rm{s}^{-1}$ (dark grey). \textbf{\emph{Upper Left}} Looking down the line of site into an outflow with inflows moving perpendicular to the line of sight. \textbf{\emph{Upper Right}} Looking down the line of sight of an inflow with an outflow occurring perpendicular to the line of sight. \textbf{\emph{Bottom Middle}} Both inflow and outflow are observed edge-on. The remaining parameters for these profiles are as follows. Outflows: $\gamma = 1.0$, $\delta = 3.0$, $\log{(N \ [\rm{cm}^{-2}])}=15$, $v_0 = 25 \ \rm{km \ s^{-1}}$, $v_w = 500 \ \rm{km \ s^{-1}}$, $R_{AP} > R_W$, $\kappa = 0$, and $f_c = 1.0$. Inflows: $\gamma = 0.5$, $\delta = 1.5$, $\log{(N \ [\rm{cm}^{-2}])}=15$, $v_0 = 15 \ \rm{km \ s^{-1}}$, $v_w = 200 \ \rm{km \ s^{-1}}$, $R_{AP} > R_W$, $\kappa = 0$, and $f_c = 1.0$.}
\label{f7}
\end{figure*}
Realistically, both accretion and feedback occur simultaneously in galaxies - for example, during gas recycling when material expelled by the galaxy is returned via gravity. This is a fair assumption given the high detection rate of outflows and the necessity for inflows brought on by simulations. In fact, it has been suggested that the presence of outflows can mask inflows in spectral lines preventing the latter's detection (e.g., \citealt{Martin2012}). This can foreseeably occur if the symmetric, with respect to line center, emission profile of an outflow fills in the weaker absorption feature of an inflow. Since numerical simulations show that outflows typically emerge in the direction perpendicular to the face of the disc while inflows move in the plane of the disc, orientation should also play a factor.
In this section, we attempt to explore which conditions favor the detection of inflows in galactic fountains (i.e., galactic systems with both inflows and outflows). To prevent the masking scenario mentioned above, we have chosen to model the Fe II 2343\AA\ line. This line has two strong fluorescent channels at wavelengths 2365\AA\ and 2381\AA, and will therefore incur very little resonant emission (see \citealt{Scarlata2015} for a detailed description of the role fluorescence plays in resonant scattering). In addition, ignoring the emission component will also allow us to much more easily model the simultaneous outflowing and inflowing of a single ionic species in SALT. In fact, under these conditions we can compute the spectrum from each model (inflow and outflow) independently and take the final spectrum as the superposition; otherwise, more modeling would be necessary to account for photons which are scattered by both the outflow and inflow. We also note that this situation can be enforced observationally by limiting the portion of the outflow captured by the observing aperture. By limiting the field of view to exclude the outflow except along the line of sight, one can effectively eliminate the entire emission component of a line profile (see \citealt{Scarlata2015} for justification). This technique is feasible with IFU spectroscopy where one can obtain spectra from sub images.
We explore how the line profile of a galactic fountain changes with orientation in Figure~\ref{f7}. We show both a medium ($50 \ \rm{km} \ \rm{s}^{-1}$) and a low ($120 \ \rm{km} \ \rm{sec}^{-1}$) resolution spectra which has been smoothed with a Gaussian kernel. The latter is about the high resolution limit of the study by \cite{Martin2012}. In this setup, we are assuming that the outflows and inflows occupy separate regions of the CGM with inflows moving perpendicular to a biconical outflow. Specifically, the models are made by generating the spectrum of a spherical inflow, removing the contribution of a biconical inflow from that spectrum, and then replacing it with a spectrum produced by a biconical outflow refilling that space. The remaining volume filled by inflowing gas is roughly the shape of a torus. Leaving such a large volume for the inflows represents a best case scenario for observing them. As expected, the orientation (compare the top two profiles) strongly affects our ability to detect an inflow at all resolutions: If there is no gas moving along the line of sight then we cannot detect any redshifted absorption, and thus have no possible way to identify an inflow given the weak emission component for this line. In the bottom panel, the low resolution spectrum of both the outflow and inflow appears to have merged together making it impossible to distinguish the inflow from another absorption feature such as absorption occurring in the ISM. This appears to be another form of masking by the outflow since this would not be the case if the outflow were absent (see the blue profile). There does appear to be hope of identifying the inflow at $50 \ \rm{km} \ \rm{s}^{-1}$ resolution, however, since the asymmetric redshifted absorption feature becomes visible.
In the case of a true fountain, where warm ($5050\rm{K} < T < 2\times 10^4\rm{K}$) ejected gas doesn't have enough energy to escape the gravitational potential well, large regions of the CGM can have both an inflowing and outflowing component (see \citealt{Kim2018}). We attempt to model such situations in Figure~\ref{f8}, where we show various outflows with an inflowing gas component occupying the same general volume. For each scenario, we have chosen covering fractions for the outflows, $f_{c,o}$, and inflows, $f_{c,i}$, such that $f_{c,o} + f_{c,i} = 1$ and $f_{c,i} <f_{c,o}$ with the order representing the fact that inflows are typically expected to have smaller covering fractions than outflows \citep{Kimm2011}. Once again, we are not able to identify inflows at the $120 \ \rm{km} \ \rm{sec}^{-1}$ resolution level given our model constraints, but are able to at the $50 \ \rm{km} \ \rm{s}^{-1}$ level. This strongly suggests that the number of inflows detected in large low resolution surveys is undercounted, and can be improved by moving to higher, but still obtainable resolutions by active instruments.
\begin{figure*}[]
\centering
\includegraphics[scale=0.83]{f8.png}
\caption{Galactic fountains with inflows represented by streams observed in the Fe II 2343\AA \ line viewed at different orientations and geometries. The Fe II 2343\AA \ line has a strong fluorescent component (not shown) and therefore has reduced resonant emission. All diagrams are to be viewed from the right. Inflowing gas is shown in light blue, outflowing gas is shown in salmon, and regions of overlap are shown in violet. The resulting spectra for each type of flow is shaded in the matching color. The composite spectrum is shown at a resolution of $50 \ \rm{km} \ \rm{s}^{-1}$ (black) and $120 \ \rm{km} \ \rm{s}^{-1}$ (dark grey). \textbf{\emph{Top Left}} A spherical fountain. \textbf{\emph{Top Right}} A biconical fountain observed parallel to the line of sight. \textbf{\emph{Bottom Left}} A biconical fountain oriented perpendicular to the line of sight. \textbf{\emph{Bottom Right}} A biconical fountain observed edge on. The remaining parameters are as follows. Outflows: $v_0 = 25 \ [\rm{km} \ \rm{s}^{-1}]$, $v_{\infty} = 500 \ [\rm{km} \ \rm{s}^{-1}]$, $R_{AP} > R_W$, $\log{(N \ [\rm{cm}^{-2}])}=15$, $\kappa = 0$, $\delta = 3.0$, and $\gamma = 1.0$. Inflows: $v_0 = 15 \ [\rm{km} \ \rm{s}^{-1}]$, $v_{\infty} = 200 \ [\rm{km} \ \rm{s}^{-1}]$, $R_{AP} > R_W$, $\log{(N \ [\rm{cm}^{-2}])}=15$, $\kappa = 0$, $\delta = 1.5$, and $\gamma = 0.5$. }
\label{f8}
\end{figure*}
\section{Discussion and Conclusions}
While the detection rate for accretion or galactic inflows is much lower than that of outflows in absorption line studies (see \citealt{Martin2012} and \citealt{Rubin2014}), the important role inflows play in models of galaxy formation suggests that they should be just as common as outflows. This fact implies that the effects of inflows on spectra lines are either weak (i.e., inflows have low covering fractions and/or low column densities) or hidden (i.e., the effects of inflows are masked by outflows and/or are mistaken for the ISM).
In regards to the latter, given the available constraints on inflows, our models suggest that we should be able to resolve the properties of inflows with observations reaching spectral resolution of $\approx 50\ \rm{km} \ \rm{sec}^{-1}$ or $R=6000$. In addition to high spectral resolution, the detection of outflows/inflows requires spectra with high signal-to-noise ratio in the continuum\footnote{We do not explore the dependency on signal-to-noise here, but extensive discussion can be found in the literature.}. The COS spectrograph on board of the Hubble Space Telescope provides access to the required spectral resolution, but has no multiplexing capabilities, and is limited in the fluxes that it's able to observe. Observing large samples of objects at the required resolution and SNR is thus a challenging task. Future space-based missions, such as the large infrared/optical/ultraviolet telescope recommended by the Astro2020 Decadal Survey, will substantially improve the situation for low-redshift galaxies. For galaxies at $z>0.6$ some of the resonant spectral features are redshifted in the optical and can be observed from the ground. A number of spectrographs on large-aperture telescopes provide both the resolution and the mutiplexing capability that allow the study of statistical samples of galaxies in reasonable time frames. Examples include FLAMES, the multi-object, intermediate and high resolution spectrograph of the Very Large Telescope, and the IMACS multi-object spectrograph on the 6.5m Magellan telescope. Additionally, focusing on ionic transition lines with strong fluorescent channels which naturally have less resonant emission (such as the Fe~II~2343\AA \ transition studied in this work) will mitigate the effect outflows have on masking inflows.
The SALT model presented in this work will be invaluable in constraining the properties of gas inflows and outflows identified in high spectral-resolution data.
\acknowledgments
\section*{Acknowledgements}
The authors would like to thank Jacob Miller for graciously helping to create the figures in this work. The authors thank the anonymous referee for an in depth reading of the manuscript and for providing valuable insight for parameter constraints.
| 558813cd31c333a01241caaa6d325ebdf503ec47 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Introduction}
Gentle algebras arose as generalizations of iterated tilted algebras of type $A$ in \cite{assem1981generalized} and \cite{assem1987iterated}. In representation theory, gentle algebras have been of interest mainly due to the fact that they form a rich, but sufficiently well-behaved class of algebras. However, gentle algebras appear naturally in many other areas of mathematics as well, a preeminent example being the field of homological mirror symmetry (for example, see \cite{haiden2014flat}). Gentle algebras have been extensively studied, and their representation theory is well-understood.
Recently, bounded derived categories of gentle algebras have been a focus of substantial research activity that followed up on \cite{bekkert2003indecomposables}, which gave a characterization of their indecomposable objects: morphisms between these indecomposable were fully described in \cite{arnesen2016morphisms}, their mapping cones were calculated in \cite{ccanakcci2019mapping} and \cite{ccanakcci2021corrigendum}, and a geometric model for these derived categories has been established in \cite{opper2018geometric} and used, among other things, to characterize silting objects thereof in \cite{amiot2019complete} and give complete derived invariants in \cite{opper2019auto} and \cite{amiot2019complete}.
The geometric model has already proven to be a useful tool for finer study of bounded derived categories of gentle algebras as well: for example, silting objects therein and related notions are studied in \cite{chang2020geometric} and \cite{chang2022recollements} using specific ways to cut the marked surface underlying the geometric model.
A very similar model has been also developed to study categories of modules over gentle algebras~\cite{baur2021geometric}.
\smallskip
In this paper, we study semiorthogonal decompositions of bounded derived categories of gentle algebras. Semiorthogonal decompositions are a useful tool for understanding triangulated categories because they allow us to view an entire triangulated category as constructed from simpler subcategories; they are studied in algebraic geometry for the derived category of coherent sheaves on algebraic varieties (for example, see \cite{kuznetsov2014semiorthogonal}), algebraic topology (where they are often referred to as Bousfield localizations, \cite[\S3]{hovey1997stablehomotopy}) as well as representation theory.
We show that, for gentle algebras, semiorthogonal decompositions of their bounded derived categories can be understood as suitable cuts of the marked surfaces in their geometric models, which can be understood as a partial converse to the results in \cite{chang2020geometric} and \cite{chang2022recollements}.
Our main result is the following theorem:
\begin{theorem*}[Theorem \ref{TDecompositionsProperSequences}]
Let $\Lambda$ be a gentle algebra. There is a one-to-one correspondence between semiorthogonal decompositions of $\Db(\Lambda\textrm{-}\mathrm{mod})$ with arbitrarily many terms and proper sequences of good cuts of the marked surface associated to~$\Lambda$.
\end{theorem*}
The strategy of proving this theorem is rather straight-forward. Firstly, we simplify the situation by studying only two-term semiorthogonal decompositions, and we formulate necessary conditions for approximations of the indecomposable projectives of the gentle algebras. In the geometric model, these necessary conditions are then showed to imply that the indecomposable of approximations of projectives yield a bipartite admissible dissection of the marked surface.
On the other hand, we show that each bipartite admissible dissection of the marked surface gives rise to a semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$, but many such dissection give the same semiorthogonal decomposition. Therefore, we seek an invariant to distinguish between \textit{equivalent} bipartite admissible dissections, and we find it in a way how the surface can be cut using the dissection. Using this invariant, called \textit{good cut}, we establish that there is a one-to-one correspondence between semiorthogonal decompositions of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ with arbitrarily many terms and proper sequences of good cuts of the marked surface.
Subsequently, we prove that semiorthogonal decompositions of $\Db(\Lambda \textrm{-}\mathrm{mod})$ are uniquely determined extensions of semiorthogonal decompositions of $\Kb(\Lambda \textrm{-}\mathrm{proj})$.
\section{Preliminaries and conventions}
In this section, we define semiorthogonal decompositions and gentle algebras, and, subsequently, we review some results on the bounded derived category of a gentle algebra and its geometric model that will prove useful in the following sections. We also set some conventions used throughout this text as adopted from~\cite{arnesen2016morphisms}.
\subsection{Semiorthogonal decompositions}
In this subsection, we give a definition of a semiorthogonal decomposition of a triangulated category, and we review some facts about them.
\begin{definition}[Right and left admissible subcategory]\label{DAdmissibleSubcategory}
Let $\mathcal{T}$ be a triangulated category and $\mathcal{C} \subseteq \mathcal{T}$ be a full triangulated subcategory; we say that $\mathcal{C}$ is \emph{right admissible} if the inclusion functor of $\mathcal{C}$ into $\mathcal{T}$ admits a right adjoint. Similarly, $\mathcal{C}$ is \emph{left admissible} if the inclusion functor has a left adjoint.
\end{definition}
\begin{definition}[Semiorthogonal decomposition]\label{DSemiorthogonalDecomposition}
Let $\mathcal{T}$ be a triangulated category. Suppose that $\mathcal{C}_1, \dots, \mathcal{C}_m$ are full triangulated subcategories of $\mathcal{T}$ and that $\mathcal{T}_i$ is the smallest triangulated subcategory of $\mathcal{T}$ containing $\mathcal{C}_1, \dots, \mathcal{C}_i$. If $\mathcal{T}_m = \mathcal{T}$, $\mathcal{C}_i$ is right admissible subcategory of $\mathcal{T}_i$ for all $1 \leq i \leq m$, and $\Hom_\mathcal{T}(\mathcal{C}_j, \mathcal{C}_i) = 0$ for all $1 \leq i < j \leq n$, then we write $\langle \mathcal{C}_1, \dots, \mathcal{C}_m \rangle$, and we say that $\langle \mathcal{C}_1, \dots, \mathcal{C}_m \rangle$ is a \emph{semiorthogonal decomposition} of $\mathcal{T}$.
\end{definition}
Before we recall some useful properties of semiorthogonal decompositions, we need to introduce some notation. Suppose that $\mathcal{B}$ is a subcategory of a triangulated category $\mathcal{T}$, we denote $\mathcal{B}^\bot = \{D \in \mathcal{T} \, | \, \forall B \in \mathcal{B}: \Hom_\mathcal{T}(B, D) = 0 \}$ and, similarly, ${^\bot}\mathcal{B} = \{D \in \mathcal{T} \, | \, \forall B \in \mathcal{B}: \Hom_\mathcal{T}(D, B) = 0 \}$.
\begin{proposition}[Lemma 3.1 in \cite{bergh2000abstract}]\label{PSemiorthogonalDecomposition}
Let $\mathcal{T}$ be a triangulated category, and let $\mathcal{C}_1$ and $\mathcal{C}_2$ be full triangulated subcategories of $\mathcal{T}$ such that $\Hom_\mathcal{T}(\mathcal{C}_2, \mathcal{C}_1) = 0$. Then the following are equivalent:
\begin{enumerate}[(i)]
\item $\mathcal{C}_1$ and $\mathcal{C}_2$ generate $\mathcal{T}$ as a triangulated category.
\item For every $D \in \mathcal{T}$ there exists a distinguished triangle $C_2 \to D \to C_1 \to C_2[1]$ with $C_2 \in \mathcal{C}_2$ and $C_1 \in \mathcal{C}_1$.
\item The inclusion functor $i_1$ of $C_1$ into $\mathcal{T}$ has a left adjoint $j_1$, and $C_2 = {^\bot}\mathcal{C}_1$.
\item The inclusion functor $i_2$ of $C_2$ into $\mathcal{T}$ has a right adjoint $j_2$, and $C_1 = \mathcal{C}_2^\bot$.
\end{enumerate}
Moreover, $i_2 j_2 D \to D \to i_1 j_1 D \to i_2 j_2 D[1]$ is the distinguished triangle in (ii); these distinguished triangles are also functorial.
\end{proposition}
\begin{proof}
This is a variation of Lemma 3.1 in \cite{bondal1990representation}. Although, the ideas behind it can be traced back as far as \cite{beilinson1982faisceaux}.
\end{proof}
It can be observed that a semiorthogonal decomposition $\langle \mathcal{C}_1, \dots, \mathcal{C}_m \rangle$ can be viewed as a sequence of two term semiorthogonal decompositions $\langle \mathcal{T}_i, \mathcal{C}_{i+1} \rangle$ of $\mathcal{T}_{i+1}$, $1 \leq i \leq m-1$, where $\mathcal{T}_i$ is the smallest triangulated subcategory of $\mathcal{T}$ containing $\mathcal{C}_1, \dots, \mathcal{C}_i$.
For more facts on triangulated categories, we refer the reader to \cite{holm2010triangulated}, for example. The reader may find more information on semiorthogonal decompositions in Chapter 10 therein.
\subsection{Gentle algebras}
In this subsection, we define gentle algebras, and, following \cite{arnesen2016morphisms}, we establish some conventions for working with them.
\begin{definition}\label{DGentleAlgebras}
Let $Q = (Q_0, Q_1)$ be a finite connected quiver. In accordance with \cite{arnesen2016morphisms}, we read paths in $Q$ from right to left. A finite-dimensional bound path algebra $\Lambda \cong kQ/I$ is called \emph{gentle} if:
\begin{enumerate}
\item Each vertex $x \in Q_0$ is a source and a target of at most two arrows.
\item For any arrow $a \in Q_1$ there is at most one arrow $b \in Q_1$ such that $ab \notin I$ and $c \in Q_1$ such that $ac \in I$.
\item For any arrow $a \in Q_1$ there is at most one arrow $b \in Q_1$ such that $ba \notin I$ and $c \in Q_1$ such that $ca \in I$.
\item The ideal $I$ is generated by paths of length two.
\end{enumerate}
\end{definition}
For the reader's convenience, we also follow the Convention 1.2 of \cite{arnesen2016morphisms}. If we denote the indecomposable projective left $\Lambda$ corresponding to a vertex $x \in Q_0$ as $P(x)$, then we can observe that there is bijection between paths from $x$ to $y$ that do not lie in $I$ and canonical basis elements of $\Hom_\Lambda(P(y), P(x))$ (given a path $p$ from $x$ to $y$ in $Q$ and not in $I$, we have a map from $P(y) \to P(x)$ that maps $u$, a path from $y$ in $Q$, to $up$). Therefore, by abuse of notation, we identify a path $p$ from $x$ to $y$ in $Q$ and not in $I$ with its corresponding basis element of $\Hom_\Lambda(P(y), P(x))$.
Following \cite{arnesen2016morphisms}, we also assume that the ground field $k$ is algebraically closed. Finally, we identify $\Db(\Lambda \textrm{-}\mathrm{mod})$ with the equivalent triangulated category $K^{-,b}(\Lambda \textrm{-}\mathrm{proj})$ of complexes of finitely generated projective $\Lambda$-modules bounded from the right with bounded cohomology. In the following text, $\Lambda = kQ/I$ will be a gentle algebra over an algebraically closed field $k$.
\begin{example}[A running example from~\cite{arnesen2016morphisms}]\label{ERunningExample}
For future reference, in order to illustrate our subsequent definitions and results, we introduce here a particular example of a gentle algebra.
This algebra is taken from one of our main sources, \cite[p.~6]{arnesen2016morphisms}, and
it is given by the following quiver with the quadratic relations indicated by the dotted arrows.
$$
\begin{tikzpicture}
\node (V0) at ( 0.0, 0.0) {$0$};
\node (V1) at (-1.3, 1.3) {$1$};
\node (V2) at (-1.3,-1.3) {$2$};
\node (V3) at ( 1.3, 1.3) {$3$};
\node (V4) at ( 1.3,-1.3) {$4$};
\draw[dotted] (V0)+( 0.35,-0.35) arc (-45 : 45 : 0.5cm);
\draw[dotted] (V0)+(-0.35, 0.35) arc (135 : 225 : 0.5cm);
\draw[dotted] (V1)+( 0,-0.5 ) arc (270 : 315 : 0.5cm);
\draw[dotted] (V2)+( 0, 0.5 ) arc ( 90 : 45 : 0.5cm);
\draw[dotted] (V3)+( 0,-0.5 ) arc (270 : 225 : 0.5cm);
\draw[dotted] (V4)+( 0, 0.5 ) arc ( 90 : 135 : 0.5cm);
\path[commutative diagrams/.cd, every arrow, every label]
(V0) edge node[swap] {$a$} (V1)
(V1) edge node[swap] {$b$} (V2)
(V2) edge node[swap] {$c$} (V0)
(V0) edge node[swap] {$d$} (V4)
(V4) edge node[swap] {$e$} (V3)
(V3) edge node[swap] {$f$} (V0);
\end{tikzpicture}
$$
\end{example}
\subsection{Objects in the derived category of a gentle algebra}
In this subsection, we recall the combinatorial description of indecomposable objects in the derived category of a gentle algebra using homotopy strings and bands from Section~4 in \cite{bekkert2003indecomposables}. We also draw on discussions of this combinatorial description in Sections~2 in \cite{arnesen2016morphisms} and \cite{opper2018geometric}.
This subsection is structured in a straightforward way: at first, we define the combinatorial objects (homotopy strings, homotopy bands, and infinite homotopy strings) used for the characterization of objects in $\Db(\Lambda\textrm{-}\mathrm{proj})$, then we show how to construct corresponding complexes (string complexes, band complexes, and infinite string complexes), and, finally, we formulate the characterization theorem of \cite{bekkert2003indecomposables}.
At first, we consider formal inverses $\overline{a}$ to arrows in $a \in Q_1$ such that $s(\overline{a}) = t(a)$ and $t(\overline{a}) = s(a)$. Also, we define that $\overline{\overline{a}} = a$ for all $a \in Q_1$.
For a path $p = a_1 \dots a_m$ in $Q$ with $a_1, \dots, a_m \in Q_1$, the \emph{inverse path} $\overline{p}$ is defined as $\overline{a_m} \dots \overline{a_1}$.
Arrows in paths are composed from right to left as usual, that is $t(a_i)=s(a_{i-1})$ for all $1 < i \le m$.
\begin{definition}[{Walks and generalized walks; after discussion at the beginning of \cite[Section~4]{bekkert2003indecomposables}}]\label{DGeneralizedWalks}
A \textit{walk} $w$ is a sequence $w_n \dots w_1$ where each $w_i$ is an arrow in $Q_1$ or an inverse arrow in $\overline{Q_1}$ and where $t(w_{i-1}) = s(w_{i})$ for $1 < i \leq n$. Trivial walks corresponding to vertices $v \in Q_0$ are also allowed.
A \textit{generalized walk} $w$ is a sequence $w_n \dots w_1$ where each $w_i$ is a non-trivial path $p$ in $Q$ or an inverse thereof, $\overline{p}$ and where consecutive paths connect at endpoints, $t(w_{i-1}) = s(w_{i})$ for $1 < i \leq n$. The inverse $\overline{w}$ of $w$ is defined naturally.
\end{definition}
\begin{remark}\label{RConventionsForUnfoldedDiagrams}
Given a walk or a generalized walk $w = w_n \dots w_1$, we usually draw it, as in \cite{arnesen2016morphisms}, in the form of a diagram
$$\begin{tikzcd}
\bullet \arrow[dash]{r}{w_n} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{w_1} & \bullet
\end{tikzcd}$$
where the line segment labeled with $w_i$ stands for an arrow pointing to the right if $w_i$ is a path in $Q$, and to the left if $w_i$ is an inverse of a path in $Q$.
Note that the arrows in such a diagram have the \emph{opposite} orientation compared to the corresponding paths in the quiver $Q$.
For instance, referring to Example~\ref{ERunningExample},
the generalized walk $w=\overline{e}\overline{f}cbp$, where $p$ is the path $p=af$, is depicted by the diagram
$$\begin{tikzcd}
\bullet \arrow[r, leftarrow, "e"] &
\bullet \arrow[r, leftarrow, "f"] &
\bullet \arrow[r, "c"] &
\bullet \arrow[r, "b"] &
\bullet \arrow[r, "af"] &
\bullet
\end{tikzcd}
$$
The reason for this peculiarity is that, following the conventions of~\cite{arnesen2016morphisms}, our diagrams represent sequences of maps between indecomposable projective left $\Lambda$-modules corresponding to individual paths or their inverses in the walk, and such maps are given by right multiplication by the paths.
In our particular example, the diagram of the generalized walk thus stands for
$$\begin{tikzcd}
P(4) \arrow[r, leftarrow, "-\cdot e"] &
P(3) \arrow[r, leftarrow, "-\cdot f"] &
P(0) \arrow[r, "-\cdot c"] &
P(2) \arrow[r, "-\cdot b"] &
P(1) \arrow[r, "-\cdot af"] &
P(3).
\end{tikzcd}
$$
\end{remark}
\begin{definition}[{Homotopy letters and homotopy strings; \cite[Subsection~4.1]{bekkert2003indecomposables}, \cite[Subsection~2.1]{arnesen2016morphisms}, \cite[Definition~2.1]{opper2018geometric}}]\label{DHomotopyStrings}
A \emph{direct homotopy letter} is a non-trivial path $p$ in $Q$ which is not contained in $I$.
An \emph{inverse homotopy letter} is by definition the inverse $\overline{p}$ of a direct homotopy letter $p$.
A \emph{homotopy letter} is a common name for direct and inverse homotopy letters.
A (finite reduced) \emph{homotopy string} $w = w_n \dots w_1$ is a (possibly trivial) generalized walk such that:
\begin{enumerate}[(i)]
\item it consists of homotopy letters $w_i$, for $1 \leq i \leq n$;
\item if $w_i$ and $w_{i-1}$ are both direct or inverse, then $w_i w_{i-1} \in I$ or $\overline{w_i w_{i-1}} \in I$, respectively, for $1 < i \leq n$;
\item if $w_i$ is direct and $w_{i-1}$ is inverse or $w_i$ is inverse and $w_{i-1}$ is direct, then $w_i$ and $\overline{w_{i-1}}$ do not start with the same arrow or $\overline{w_i}$ and $w_{i-1}$ do not end with the same arrow, respectively, for $1 < i \leq n$.
\end{enumerate}
\end{definition}
\begin{definition}[{Gradings and homotopy bands; \cite[Subsection~4.1]{bekkert2003indecomposables}, \cite[Subsection~2.2]{arnesen2016morphisms}}]\label{DGrandingsHomotopyBands}
Let $w = w_n \dots w_1$ be a homotopy string. We define a \emph{grading} $\mu$ on $w$ to be a function $\mu\colon\{0, \dots, n\} \to \mathbb{Z}$ such that $\mu(i-1) = \mu(i) + 1$ if $w_i$ is a direct homotopy letter and $\mu(i-1) = \mu(i) - 1$ if $w_i$ is an inverse homotopy letter for all $1 \leq i \leq n$.
We also denote by $\overline{\mu}\colon\{0, \dots, n\} \to \mathbb{Z}$ the function given by $\overline{\mu}(i)=\mu(n-i)$, which is a grading on the inverse homotopy string $\overline{w}$.
A non-trivial homotopy string $w = w_n \dots w_1$ is called a \emph{homotopy band} if:
\begin{enumerate}
\item its endpoints coincide, $s(w_1) = t(w_n)$;
\item there exists a grading $\mu$ on $w$ such that $\mu(0) = \mu(n)$;
\item $w$ is not a proper power of another homotopy string;
\item one of $w_1, w_n$ is direct and the other is inverse.
\end{enumerate}
\end{definition}
\begin{remark} ~
\begin{enumerate}
\item Regarding homotopy strings and homotopy bands, we opted to for the terminology used in \cite{arnesen2016morphisms} and \cite{opper2018geometric}. In \cite{bekkert2003indecomposables}, they are referred to as generalized strings and generalized bands.
\item As usual, we will consider natural equivalence relations on the sets of homotopy strings and homotopy bands.
For homotopy strings, the equivalence $\sim_\mathsf{St}$ is given by identifying $w$ and $\overline{w}$.
For bands, we define $\sim_\mathsf{Ba}$ by considering two homotopy bands equivalent if they differ only by cyclic rotation and possibly taking inverse.
The equivalence relation extends in an obvious way to pairs $(w,\mu)$ where $w$ is a homotopy string or band and $\mu$ is a grading.
\end{enumerate}
\end{remark}
\begin{remark}\label{RGradedDiagrams}
If we have a homotopy string $w = w_n \dots w_1$ with a grading $\mu$,
we represent it using an \textit{unfolded diagram} as in \cite{arnesen2016morphisms}, which is just the diagram for the corresponding walk as explained in Remark~\ref{RConventionsForUnfoldedDiagrams} together with values of the function $\mu$ above the vertices of the diagram:
$$\begin{tikzcd}[row sep=0.1cm]
\mu(n) & \mu(n-1) & \mu(1) & \mu(0) \\
\bullet \arrow[dash]{r}{w_n} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{w_1} & \bullet
\end{tikzcd}$$
A particular example using the homotopy string from Remark~\ref{RConventionsForUnfoldedDiagrams} is
$$\begin{tikzcd}[row sep=0.1cm]
2 & 1 & 0 & 1 & 2 & 3
\\
\bullet \arrow[r, leftarrow, "e"] &
\bullet \arrow[r, leftarrow, "f"] &
\bullet \arrow[r, "c"] &
\bullet \arrow[r, "b"] &
\bullet \arrow[r, "af"] &
\bullet
\end{tikzcd}$$
If $w = w_n \dots w_1$ is a homotopy band with a grading $\mu$, its \emph{unfolded diagram} by definition infinitely repeats both to the left and to the right. We usually also decorate the homotopy letter $w_1$ with a fixed scalar $\lambda \in k^{\times}$, because this is an additional datum needed to define a corresponding one-dimensional band complex (see Definition~\ref{DOneDimBands} below):
$$\begin{tikzcd}[row sep=0.1cm, arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}]
& & \mu(0) & \mu(n-1) & \mu(1) & \mu(0) & \\
{} & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{\lambda w_1} & \bullet \arrow[dash]{r}{w_n} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{\lambda w_1} & \bullet \arrow[dash]{r}{w_n} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
In order to see a particular example, consider the band $w=\overline{d}\overline{e}\overline{f}cba$ for the gentle algebra from Example~\ref{ERunningExample} (see \cite[Example 2.3]{arnesen2016morphisms}). The corresponding unfolded diagram then reads
$$\begin{tikzcd}[row sep=0.1cm, arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}]
& & 3 & 2 & 1 & 0 & 1 & 2 & 3 & &
\\
{} &
\arrow[dash, decorate = true]{l}
\bullet \arrow[r, "\lambda a"] &
\bullet \arrow[r, leftarrow, "d"] &
\bullet \arrow[r, leftarrow, "e"] &
\bullet \arrow[r, leftarrow, "f"] &
\bullet \arrow[r, "c"] &
\bullet \arrow[r, "b"] &
\bullet \arrow[r, "\lambda a"] &
\bullet \arrow[r, leftarrow, "d"] &
\bullet \arrow[dash, decorate = true]{r} &
{}
\end{tikzcd}$$
\end{remark}
If the underlying algebra $\Lambda$ has infinite global dimension, we need to work with infinite homotopy strings in addition to homotopy strings and homotopy bands in order to describe all indecomposable objects in $\Db(\Lambda\textrm{-}\mathrm{mod})$. Infinite homotopy strings arise from oriented cycles $a_m \dots a_1$ in $Q$ that have full relations, meaning that $a_i a_{i-1} \in I$ for $1 < i \leq n$ and $a_1 a_m \in I$.
\begin{definition}[{Infinite homotopy strings; \cite[Subsection 4.3 above Lemma 5]{bekkert2003indecomposables}, \cite[Definition~2.6 and the following discussion]{arnesen2016morphisms}}]\label{DInfiniteHomotopyStrings}
Let $w = w_n \dots w_1$ be a homotopy string such that $w_n$ is direct (in particular $w$ is non-trivial).
We say that $w$ is \emph{left resolvable} if there exists a cycle $a_m \dots a_1$ in $Q$ with full relations and $a_1 w$ is a homotopy string.
Moreover, we call $w$ \emph{primitive left resolvable} if $w$ is left resolvable, but $w_{n-1} \dots w_1$ is not.
The notions of \emph{right resolvability} and \emph{primitive right resolvability} are defined dually so that $w$ is (primitive) right resolvable if and only if the inverse homotopy string $\overline{w}$ is (primitive) left resolvable.
For $w$ that is left resolvable
with a cycle $a_m \dots a_1$ in $Q$ with full relations, we form a \emph{left infinite homotopy} string ${^\infty w}$ by adding countably many copies of $a_m \dots a_1$ to the left of $w$ obtaining:
$$ {^\infty w} = \dots a_m a_{m-1} \dots a_1 \dots a_m a_{m-1} \dots a_1 w_1 \dots w_n.$$
For a right resolvable homotopy string, we form a \emph{right infinite homotopy} string $w^\infty$ in similar way.
If a homotopy string is both left and right resolvable, we can also form a \emph{two-sided infinite homotopy string} ${^\infty w^\infty}$ by combining the two constructions.
\end{definition}
\begin{remark} ~
\begin{enumerate}
\item Infinite homotopy strings are not specifically named in \cite{bekkert2003indecomposables}, so we opted to the term used in \cite{arnesen2016morphisms} and \cite{opper2018geometric}, which is compatible with naming of homotopy strings and homotopy bands.
\item Every left infinite homotopy string ${^\infty w}$ is determined by a unique primitive left resolvable homotopy string $w$. Analogously, every right infinite homotopy string $w^\infty$ is determined by a unique primitive right resolvable homotopy string $w$, and similarly for two-sided infinite homotopy strings.
\item If $w$ is left resolvable, any grading $\mu\colon\{0, \dots, n\} \to \mathbb{Z}$ uniquely extends to a grading ${^\infty\mu}\colon\{0, \dots, n, n+1, n+2, \dots\} \to \mathbb{Z}$ of the left infinite homotopy string ${^\infty w}$, and we can depict ${^\infty w}$ with ${^\infty\mu}$ in the form of an unfolded diagram analogous to that of ordinary homotopy strings as in Remark~\ref{RGradedDiagrams}.
Analogous comments apply to right and two-sided infinite homotopy strings.
For instance, if $w=a\overline{d}$ for the algebra from Example~\ref{ERunningExample} with the grading $\mu\colon\{0, \dots, n\} \to \mathbb{Z}$ such that $\mu(0)=0$, then we can extend $w$ to ${^\infty w^\infty}$ and $\mu$ to ${^\infty\mu^\infty}\colon\mathbb{Z}\to\mathbb{Z}$ and draw ${^\infty w^\infty}$ and ${^\infty \mu^\infty}$ in the form of an unfolded diagram
$$\begin{tikzcd}[row sep=0.1cm, column sep=0.7cm, arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}]
& & -2 & -1 & 0 & 1 & 0 & -1 & -2 &
\\
{} &
\arrow[dash, decorate = true]{l}
\bullet \arrow[r, "a"] &
\bullet \arrow[r, "c"] &
\bullet \arrow[r, "b"] &
\bullet \arrow[r, "a"] &
\bullet \arrow[r, leftarrow, "d"] &
\bullet \arrow[r, leftarrow, "e"] &
\bullet \arrow[r, leftarrow, "f"] &
\bullet \arrow[r, leftarrow, "d"] &
\bullet \arrow[dash, decorate = true]{r} &
{}
\end{tikzcd}$$
\item We can define formal inverses of infinite homotopy strings or infinite homotopy strings with a grading in a natural way. If $w$ is a primitive left resolvable homotopy string, we put $\overline{^\infty w} = \overline{w}^\infty$. Similarly, if $w$ is both primitive left and right resolvable, we put $\overline{^\infty w^\infty} = {^\infty\overline{w}^\infty}$.
That is, left infinite homotopy strings invert to right infinite homotopy strings and vice versa; whereas, inverting a two-sided infinite homotopy string produces yet another two-sided infinite homotopy string.
We again define an equivalence relation $\sim_{\infty\mathsf{St}}$ on the set of all infinite homotopy strings as well as on the set of all pairs consisting of an infinite homotopy string and a grading, which identifies a (graded) infinite homotopy string with its inverse.
\end{enumerate}
\end{remark}
\begin{convention}
From now on we will always assume that homotopy strings, infinite homotopy strings, and homotopy bands are equipped with some grading (which may remain only implicit in some arguments).
\end{convention}
\begin{definition}[{{}String and infinite string complexes; \cite[Definition~2]{bekkert2003indecomposables}, \cite[Subsection 2.1]{arnesen2016morphisms}}]\label{DStringComplexes}
To a homotopy string $w = w_n \dots w_1$ or an infinite homotopy string $w$ with grading $\mu$ we associate a complex of projectives $P_{(w, \mu)}$ as follows:
\begin{itemize}
\item the projective module in cohomological degree $j$ is a (necessarily finite) direct sum of indecomposable projectives:
$$\bigoplus_{\mu(i) = j} P(t(w_i))$$
where $t(w_0)$ stands for $s(w_1)$ by convention if $w$ is finite or only left infinite;
\item the differential is defined componentwise as follows: its only non-zero components are $P(t(w_i)) \overset{-\cdot w_i}{\longrightarrow} P(s(w_i))$ if $w_i$ is direct and $P(s(w_i)) \overset{-\cdot\overline{w_i}}{\longrightarrow} P(t(w_i))$ if $w_i$ is inverse, for all indices $i$.
\end{itemize}
If $w$ is finite, call the resulting complex $P_{(w, \mu)}$ a \emph{string complex}.
If $w$ is infinite, we call it an \emph{infinite string complex}.
\end{definition}
\begin{remark} ~
\begin{enumerate}
\item If $w$ is a homotopy string, then clearly $P_{(w, \mu)}\in\Kb(\Lambda\textrm{-}\mathrm{proj})$.
If $w$ is infinite, then $P_{(w, \mu)}$ has finitely generated projective components, it is bounded on the right and has a bounded cohomology by~\cite[Lemma~5(1)]{bekkert2003indecomposables}. Hence, up to quasi-isomorphism, we can consider $P_{(w, \mu)}$ as an object of $\Db(\Lambda\textrm{-}\mathrm{mod})$.
\item Two string complexes $P(w_1, \mu_1)$ and $P(w_2, \mu_2)$ are isomorphic if and only if $(w_1, \mu_1) \sim_\mathsf{St} (w_2, \mu_2)$.
Two infinite string complexes $P(w_1, \mu_1)$ and $P(w_2, \mu_2)$ are isomorphic if and only if $(w_1, \mu_1) \sim_{\infty\mathsf{St}} (w_2, \mu_2)$.
A string complex is never isomorphic to an infinite string complex.
We refer to \cite[Theorem 3]{bekkert2003indecomposables} for details.
\end{enumerate}
\end{remark}
\begin{definition}[{One-dimensional band complexes; \cite[Definition 3]{bekkert2003indecomposables}, \cite[Subsection~2.2]{arnesen2016morphisms}}]\label{DOneDimBands}
To a homotopy band $w = w_n \dots w_1$ with grading $\mu$ and a fixed scalar $\lambda \in k^{\times}$, we associate a complex of projectives $B_{(w, \mu), \lambda, 1}$, where we consider the indices $i$ of modulo $n$:
\begin{itemize}
\item the projective module in degree $j$ is a direct sum of indecomposable projectives:
$$\bigoplus_{0 \leq i < n: \mu(i) = j} P(t(w_i)),$$
in the same way as for string complexes;
\item the differential is defined similarly to string complexes if the length of the string is $n\ge 3$:
its only non-zero components are $P(t(w'_i)) \overset{-\cdot w'_i}{\longrightarrow} P(s(w'_i))$, where $w'_i = w_i$ if $w_i$ is direct and $w'_i = \overline{w_i}$ if $w_i$ is inverse, with the exception of $i=1$ where the component is $P(t(w'_1)) \overset{-\cdot \lambda w'_1}{\longrightarrow} P(s(w'_1))$;
\item in the somewhat degenerate case $n=2$, we necessarily have $s(w_1)=s(w_2)$, $t(w_1)=t(w_2)$ and the entire complex has two terms and is of the form
$$\begin{tikzcd}
P(t(w'_1)) \arrow[rr, "-\cdot(w'_2 + \lambda w'_1)"] &&
P(s(w'_1)).
\end{tikzcd}$$
\end{itemize}
We refer to such complexes as to \emph{one-dimensional band complexes}.
\end{definition}
\begin{remark}\label{RHigherDimensionalBands}
For each homotopy band $w = w_n \dots w_1$ with grading $\mu$, a fixed scalar $\lambda \in k^{\times}$ and $n>1$, there are also indecomposable \emph{higher-dimensional band complexes} $B_{(w, \mu), \lambda, n}\in\Kb(\Lambda\textrm{-}\mathrm{proj})$.
We refer to \cite[Definition~3]{bekkert2003indecomposables} and especially \cite[Section 5]{arnesen2016morphisms} for a detailed account.
As far as we are concerned, it is only important that for each $w$, $\mu$ and $\lambda$ and $n \ge 1$, we have a triangle in $\Kb(\Lambda\textrm{-}\mathrm{proj})$ of the form
$$\begin{tikzcd}
B_{(w, \mu), \lambda, n} \arrow{r} &
B_{(w, \mu), \lambda, n-1} \arrow{r} \oplus B_{(w, \mu), \lambda, n+1} \arrow{r} &
B_{(w, \mu), \lambda, n} \arrow{r} &
B_{(w, \mu), \lambda, n}[1],
\end{tikzcd}$$
where $B_{(w, \mu), \lambda, 0}=0$ by convention if $n=1$.
In particular, if $\mathcal{C}\subseteq\Kb(\Lambda\textrm{-}\mathrm{proj})$ is a full subcategory closed under extensions and summands, then $B_{(w, \mu), \lambda, n}\in\mathcal{C}$ for some $n\ge 1$ if and only if $B_{(w, \mu), \lambda, n}\in\mathcal{C}$ for \emph{all} $n\ge 1$.
\end{remark}
\begin{remark}
The only non-trivial isomorphisms between band complexes are $B_{(w_1, \mu_1), \lambda, n} \cong B_{(w_2,\mu_2), \lambda^\varepsilon, n}$, where $(w_1,\mu_1) \sim_\mathsf{Ba} (w_2,\mu_2)$ and $\varepsilon=\pm1$, where the sign depends on whether the passage from $w_1$ to $w_2$ can be achieved only by a cyclic rotation or whether we need to pass to an inverse homotopy band as well.
There are no isomorphisms between band complexes and homotopy string complexes or infinite homotopy string complexes.
\end{remark}
Finally, having defined all necessary notions, we can state the result of \cite{bekkert2003indecomposables} that characterizes indecomposable objects of $\Db(\Lambda \textrm{-}\mathrm{mod})$ as string, infinite string, and band complexes:
\begin{theorem}[{Characterization of indecomposables in $\Db(\Lambda\textrm{-}\mathrm{mod})$; \cite[Theorem~3]{bekkert2003indecomposables}}]\label{TIndecomposablesStringsBands}
There are one-to-one correspondences between:
\begin{enumerate}
\item isomorphism classes of indecomposable objects of $\Kb(\Lambda\textrm{-}\mathrm{proj})$ and isomorphism classes of string and band complexes;
\item isomorphism classes of indecomposable objects of $\Db(\Lambda\textrm{-}\mathrm{mod})$ which are not in $\Kb(\Lambda\textrm{-}\mathrm{proj})$ and isomorphism classes of infinite string complexes.
\end{enumerate}
\end{theorem}
\subsection{Morphisms in the derived category of a gentle algebra}
After having described the indecomposable objects in $\Db(\Lambda\textrm{-}\mathrm{mod})$, we focus on morphisms between them. In this subsection, we present a combinatorial description of morphisms, as given in \cite{arnesen2016morphisms}.
We structure this subsection similarly to the previous one. We begin by giving a combinatorial description of certain distinguished maps between string, infinite string, and one-dimensional complexes, and we conclude by stating the main result of \cite{arnesen2016morphisms} that gives a basis of maps between both in the category of complexes over $\Lambda\textrm{-}\mathrm{proj}$ and the bounded derived category $\Db(\Lambda\textrm{-}\mathrm{mod})$ in terms of those special maps.
Before we define any maps between string, infinite string, and one-dimensional complexes, we note that by virtue of their definition we can describe the maps using their unfolded diagrams (cf. discussion at the beginning of Section 3 in \cite{arnesen2016morphisms}). Suppose that $M$ and $N$ are string, infinite string, or one-dimensional band complexes and that $f\colon M \to N$ is a map of complexes between them. The map $f$ consists of maps $f^i: M^i \to N^i$; since we have defined $M^i$ and $N^i$ as direct sums of certain indecomposable projectives, $f^i$ can be thought of as a matrix of maps between them given by linear combination of paths between corresponding vertices in the unfolded diagrams of $M$ and $N$.
We note that there are examples of how maps between unfolded diagrams translate to maps between corresponding complexes throughout Section 3.1 in \cite{arnesen2016morphisms}.
In accordance with \cite{arnesen2016morphisms}, we omit the scalar $\lambda \in k^\times$ for one-dimensional complexes from the unfolded diagrams to allow for more clarity in the definitions. However, the reader is encouraged to keep in mind that it is still implicitly present.
\begin{definition}[Single and singleton single maps; Definition 3.1 and 3.7 in \cite{arnesen2016morphisms}]\label{DSingletonSingleMaps}
Suppose that $M$ and $N$ are string, infinite string, or one-dimensional band complexes and that $p$ is a non-stationary path in the quiver $Q$. Then the following configuration of unfolded diagrams:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[]{d}{p} \arrow[dash]{r}{m_i} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow[dash]{r}{n_i} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
gives rise to a \emph{single map} from $M$ to $N$ if the following conditions are met:
\begin{itemize}
\item[\textbf{(L1)}] if $m_{i+1}$ is direct, then $m_{i+1} p = 0$;
\item[\textbf{(R1)}] if $m_i$ is inverse, then $m_i p = 0$;
\item[\textbf{(L2)}] if $n_{i+1}$ is inverse, then $p n_{i+1} = 0$;
\item[\textbf{(R2)}] if $n_i$ is direct, then $p n_i = 0$.
\end{itemize}
The map is called \emph{singleton single map} if it arises from one of the following four configurations (up to inverting one of the homotopy strings) that satisfy the conditions above:
\begin{enumerate}[(i)]
\item $$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[]{d}{p} & \\
N: & & \bullet \arrow[dash]{r}{n_i} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
\item $$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[]{d}{p} \arrow{r}{m_i}[swap]{=p p_R} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & & \bullet \arrow[dash]{r}{n_i} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
\item $$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[]{d}{p} & \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_{i+1}}[swap]{=p_L p} & \bullet \arrow[dash]{r}{n_i} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
\item $$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[]{d}{p} \arrow{r}{m_i}[swap]{= p p_R} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_{i+1}}[swap]{= p_L p} & \bullet \arrow[dash]{r}{n_i} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
\end{enumerate}
The paths $p$, $p_L$ and $p_R$ are required to be non-stationary. Moreover, in cases (i) and (ii), if $m_{i+1}$ is inverse, $p$ is not an initial homotopy substring of $m_{i+1}$ and $m_{i+1}$ is not an initial homotopy substring of $p$; in cases (i) and (iii), if $n_i$ is inverse, $p$ is not a terminal homotopy substring of $n_i$ and $n_i$ is not $n_i$ is not a terminal homotopy substring of~$p$.
\end{definition}
\begin{remark}
In the unfolded diagrams above and also later in the text, we follow the convention that the arrows without specified direction, which are depicted as $\begin{tikzcd} \bullet \arrow[dash]{r} & \bullet \end{tikzcd}$, may stand also for a zero component.
On the other hand, arrows indicated by $\begin{tikzcd} \bullet \arrow{r} & \bullet \end{tikzcd}$ and $\begin{tikzcd} \bullet & \arrow{l} \bullet \end{tikzcd}$ must be non-zero.
The fact that the unfolded diagram possibly continues to the left, to the right, and in between is indicated by $\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}] {} & \arrow[dash, decorate = true]{l} \bullet \end{tikzcd}$, $\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}] \bullet \arrow[dash, decorate = true]{r} & {} \end{tikzcd}$, and $\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}] \bullet \arrow[dash, dotted]{r} & \bullet \end{tikzcd}$, respectively.
\end{remark}
\begin{remark}
The last two conditions in the definition above guarantee two things: due to their first parts, the fact that the configurations (i) to (iv) for singleton single maps are disjoint, and, due to their second parts, non-existence of non-trivial chain homotopies to other single or double maps, so the maps of complexes are indeed singleton in their homotopy class.
\end{remark}
\begin{definition}[Double and singleton double maps; Definition 3.3 and 3.8 in \cite{arnesen2016morphisms}]\label{DSingletonDoubleMaps}
Suppose that $M$ and $N$ are string, infinite string, or one-dimensional band complexes and that $f$ is a non-stationary path in the quiver $Q$. Then the following configuration of unfolded diagrams:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[']{d}{p_L} \arrow{r}{m_i} & \bullet \arrow[]{d}{p_R} \arrow[dash]{r}{m_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow{r}{n_i} & \bullet \arrow[dash]{r}{n_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
gives rise to a \emph{double map} from $M$ to $N$ if the following conditions are met:
\begin{itemize}
\item[\textbf{(C)}] $p_L n_i = m_i p_R$;
\item[\textbf{(L1)}] if $m_{i+1}$ is direct, then $m_{i+1} p_L = 0$;
\item[\textbf{(R1)}] if $m_{i-1}$ is inverse, then $m_{i-1} p_R = 0$;
\item[\textbf{(L2)}] if $n_{i+1}$ is inverse, then $p_L n_{i+1} = 0$;
\item[\textbf{(R2)}] if $n_{i-1}$ is direct, then $p_R n_{i-1} = 0$.
\end{itemize}
The map is called \emph{singleton double map} if it arises from the configuration that satisfies the conditions above and, moreover, there exists a non-stationary path $p'$ so that:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[']{d}{p_L} \arrow{r}{m_i}[swap]{=p_L p'} & \bullet \arrow[]{d}{p_R} \arrow[dash]{r}{m_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow{r}{n_i}[swap]{=p' p_R} & \bullet \arrow[dash]{r}{n_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
\end{definition}
\begin{definition}[Graph maps; Definition 3.9 in \cite{arnesen2016morphisms}]\label{DGraphMaps}
Suppose that $M$ and $N$ are string, infinite string, or one-dimensional band complexes. Consider a maximal overlap of the unfolded diagrams as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_L} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_p} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_{p-1}} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{\ell_2} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_1} & \bullet \arrow[equal]{d} \arrow[dash]{r}{m_R} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_L} & \bullet \arrow[dash]{r}{\ell_p} & \bullet \arrow[dash]{r}{\ell_{p-1}} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{\ell_2} & \bullet \arrow[dash]{r}{\ell_1} & \bullet \arrow[dash]{r}{n_R} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
It gives rise to a \emph{graph map} from $M$ to $N$ if one the following endpoint conditions is met
and one of dual endpoint conditions \textbf{(RG1)}, \textbf{(RG2)}, or \textbf{(RG$\infty$)} is met as well:
\begin{itemize}
\item[\textbf{(LG1)}] if $m_L$ and $n_L$ are both direct or both inverse, then there exists a non-stationary path $f_L$ such that $m_L = f_L n_L$ or $n_L = m_L f_L$, respectively;
\item[\textbf{(LG2)}] if $m_L$ and $n_L$ are neither both direct nor both inverse, then $m_L$ is zero or inverse and $n_L$ is zero or direct;
\item[\textbf{(LG$\infty$)}] both strings continue infinitely to the left: the diagram continues infinitely to the left with commuting squares in which its vertical maps are isomorphisms;
\end{itemize}
\end{definition}
\begin{definition}[Quasi-graph maps; Definition 3.11 in \cite{arnesen2016morphisms}]\label{DQuasiGraphMaps}
Suppose that $M$ and $N$ are string, infinite string, or one-dimensional band complexes. Then a maximal overlap of the unfolded diagrams as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_L} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_p} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_{p-1}} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{\ell_2} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_1} & \bullet \arrow[equal]{d} \arrow[dash]{r}{m_R} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_L} & \bullet \arrow[dash]{r}{\ell_p} & \bullet \arrow[dash]{r}{\ell_{p-1}} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{\ell_2} & \bullet \arrow[dash]{r}{\ell_1} & \bullet \arrow[dash]{r}{n_R} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
gives rise to a \emph{quasi-graph map} from $M$ to $N[1]$ provided that
\begin{itemize}
\item none of the endpoint conditions \textbf{(LG1)}, \textbf{(LG2)}, \textbf{(LG$\infty$)} \textbf{(RG1)}, \textbf{(RG2)}, or \textbf{(RG$\infty$)} holds,
\item if $p=0$ and both $m_L$ and $n_R$ are direct, then $m_L n_R=0$,
\item if $p=0$ and both $m_R$ and $n_L$ are inverse, then $m_R n_L=0$.
\end{itemize}
The quasi-graph map is represented by single maps from $M$ to $N[1]$ given by $\ell_p, \dots, \ell_1$ and possibly also double maps arising from $m_L, n_L$ and $m_R, n_R$ (cf. Definition 3.12 in \cite{arnesen2016morphisms}).
\end{definition}
\begin{remark}
The additional conditions in the last definition when $p=0$ do not seem to be treated in the literature, but they are necessary to define the corresponding single map $M\to N[1]$. Consider for example the quasi-graph map situation:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{m_L} & \bullet \arrow[equal]{d} & & {}\\
N: & & \bullet \arrow{r}{n_R} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
The corresponding map $M\to N[1]$ is then represented for example by the following unfolded diagram, and this makes sense only provided that $m_L n_R=0$:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{m_L} & \bullet \arrow{d}{n_R} & & {}\\
N: & \bullet \arrow{r}{n_R} & \bullet \arrow[dash]{r} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
\end{remark}
\begin{theorem}[Basis of maps between string, infinite string, and one-dimensional band complexes; Proposition 4.1 and Theorem 3.15 in \cite{arnesen2016morphisms}]\label{TBasisMaps}
Suppose that $M$ and $N$ are string, infinite string, or one-dimensional band complexes. Then the following statements hold:
\begin{enumerate}[(i)]
\item graph maps, single maps, and double maps form a $k$-linear basis of the space $\Hom_{C(\Lambda\textrm{-}\mathrm{proj})}(M, N)$;
\item graph maps, singleton single maps, singleton double maps, and quasi-graph maps form a $k$-linear basis of $\Hom_{\Db(\Lambda\textrm{-}\mathrm{mod})}(M, N)$;
\end{enumerate}
\end{theorem}
\begin{remark}
In light of the theorem above, we refer to graph maps, single maps, and double maps as to basis maps. A graph map is not homotopic to a scalar multiple of another basis map as proved in Subsection 4.4 in \cite{arnesen2016morphisms}. Proposition 4.8 in \cite{arnesen2016morphisms} illustrates that quasi-graph maps represent homotopy classes of homotopy non-trivial single or double maps that are homotopic to some other single or double maps (cf. Definition 3.12 \cite{arnesen2016morphisms}). The remaining homotopy non-trivial single and double maps are called singleton because they are not homotopic to any scalar multiple of any other basis map (cf. Remark 4.7 in \cite{arnesen2016morphisms})
\end{remark}
The study of morphisms including higher-dimensional band complexes can be essentially reduced to the one-dimensional case with some additional effort as shown in Section 5 in \cite{arnesen2016morphisms}.
\subsection{Geometric model of a derived category of a gentle algebra}
There is an elegant geometric model of a derived category of gentle algebra developed by \cite{opper2018geometric}; structure of the derived category is captured by curves on a marked surface and their intersections. In this subsection, we present the geometric model in a formalism of \cite{amiot2019complete}, which differs slightly from the one of \cite{opper2018geometric} (and also for the essentially equivalent model~\cite{baur2021geometric} which was developed to describe the module category). Successively, we focus on several aspects of the geometric model, namely: marked surfaces and how they model gentle algebras, (graded) curves on the marked surface as models of objects in the derived category, morphisms between objects represented by intersections of curves, and representation of mapping cones in the geometric model.
\subsubsection{Marked surfaces and their dissections}
\begin{definition}[Marked surface; Definition 1.7 in \cite{amiot2019complete}]\label{DMarkedSurfaces}
Let $S$ be the interior of a compact oriented open smooth surface with boundary $\partial S$; let $M = M_{\color{dark-green}{\circ}} \cup M_{\color{red} \bullet}$ be finite set of marked points on $S \cup \partial S$ such that each connected component of $\partial S$ contains at least one marked point and that $\color{dark-green}{\circ}$-marked points (elements of $M_{\color{dark-green}{\circ}}$) and $\color{red} \bullet$-marked points (elements of $M_{\color{red} \bullet}$) alternate on each connected component of $\partial S$, and let $P = P_{\color{dark-green}{\circ}} \cup P_{\color{red} \bullet}$ be a finite set of marked points in $S$ called punctures (similarly as for marked points, elements of $P_{\color{dark-green}{\circ}}$ and $P_{\color{red} \bullet}$ are called $\color{dark-green}{\circ}$-punctures and $\color{red} \bullet$-punctures, respectively). If $\partial S$ is empty, it is required that both $P_{\color{dark-green}{\circ}}$ and $P_{\color{red} \bullet}$ be non-empty.
We refer to elements of $M_{\color{dark-green}{\circ}} \cup P_{\color{dark-green}{\circ}}$ and $M_{\color{red} \bullet} \cup P_{\color{red} \bullet}$ as to $\color{dark-green}{\circ}$-points and $\color{red} \bullet$-points, respectively.
\end{definition}
\begin{definition}[$\color{dark-green}{\circ}$-arcs and $\color{red} \bullet$-arcs; Definition 1.8 in \cite{amiot2019complete}]\label{DMarkedPointArcs}
A \emph{$\color{dark-green}{\circ}$-arc} is smooth map $\gamma$ from $(0,1)$ to $S \, \backslash \, P$ such that its endpoints, $\lim_{x \to 0} \gamma(x)$ and $\lim_{x \to 1} \gamma(x)$, lie in $M_{\color{dark-green}{\circ}} \cup P_{\color{dark-green}{\circ}}$ and that $\gamma$ is not contractible to an element in $M_{\color{dark-green}{\circ}} \cup P_{\color{dark-green}{\circ}}$. The notion of a \emph{$\color{red} \bullet$-arcs} is defined similarly, with $\color{red} \bullet$-marked points and $\color{red} \bullet$-punctures as their endpoints.
\end{definition}
\begin{definition}[Admissible dissection; Definition 1.9 in \cite{amiot2019complete}]\label{DAdmissibleDissection}
A collection of pairwise non-intersecting and pairwise different \emph{$\color{dark-green}{\circ}$-arcs} $\{\gamma_1, \dots, \gamma_r\}$ on a marked surface $(S, M, P)$ is \emph{admissible} if its arcs do not enclose a subsurface containing no punctures of $P_{\color{red} \bullet}$ and with no boundary segment of $S$ on its boundary. A maximal admissible collection of $\color{dark-green}{\circ}$-arcs is called an \emph{admissible $\color{dark-green}{\circ}$-dissection}. These notions are used for $\color{red} \bullet$-arcs in an analogous manner.
\end{definition}
\begin{remark}
Throughout this text, the term admissible dissection stands for an admissible $\color{dark-green}{\circ}$-dissection.
\end{remark}
\begin{proposition}[Properties of admissible dissections; Propositions 1.11 to 1.13 in \cite{amiot2019complete}]\label{PAdmissibleDissectionProperties}
Let $(S, M, P)$ be a marked surface, and let $\Delta$ be an admissible dissection. Then the following statements hold:
\begin{enumerate}[(i)]
\item The number of $\color{dark-green}{\circ}$-arcs in $\Delta$ equals $|M_{\color{dark-green}{\circ}}| + |P| + b + 2g - 2$ arcs, where $g$ is the genus of $S$ and $b$ is the number of connected components of $\partial S$.
\item The connected components of the complement of $\Delta$ in $S \, \backslash \, P$ are homeomorphic to either an open disk with precisely one $\color{red} \bullet$-marked point on its boundary or to an open punctured open disk with no $\color{red} \bullet$-marked point on its boundary and precisely one $\color{red} \bullet$-puncture in the interior.
\item There exists, up to homotopy, a unique $\color{red} \bullet$-admissible dissection $\Delta^{\ast}$ such that each $\color{dark-green}{\circ}$-arc of $\Delta$ intersects exactly one $\color{red} \bullet$-arc of $\Delta^{\ast}$.
\end{enumerate}
\end{proposition}
\begin{definition}[Algebra associated to an admissible dissection; Definition 1.2 in \cite{amiot2019complete}]\label{DDissectionAssociatedAlgebra}
Let $(S, M, P)$ be a marked surface, and let $\Delta$ its admissible dissection. We set the algebra associated to the admissible dissection $\Delta$, $A(\Delta)$, to be the quotient of a path algebra of the quiver $Q(\Delta)$ over $k$ by the ideal $I(\Delta)$ defined as follows:
\begin{itemize}
\item The vertices of $Q(\Delta)$ are in bijection with $\color{dark-green}{\circ}$-arcs in $\Delta$.
\item There is an arrow $i \to j$ of $Q(\Delta)$ whenever $\color{dark-green}{\circ}$-arcs corresponding to $i$ and $j$ meet at the same $\color{dark-green}{\circ}$-marked point or $\color{dark-green}{\circ}$-puncture such that $j$ immediately follows $i$ the counter-clockwise order around the $\color{dark-green}{\circ}$-point (this means that there is no $\color{dark-green}{\circ}$-arc of $\Delta$ with the $\color{dark-green}{\circ}$-point as an endpoint between those corresponding to $i$ and $j$).
\item The ideal $I(\Delta)$ is generated by the following relations: $i \to j$ and $j \to k$ compose to zero if the $\color{dark-green}{\circ}$-arcs corresponding to $i$ and $j$ meet at the endpoint of the $\color{dark-green}{\circ}$-arc corresponding to $j$ other than where the $\color{dark-green}{\circ}$-arcs corresponding $\color{dark-green}{\circ}$-arcs to $j$ and $k$ meet.
\end{itemize}
\end{definition}
\begin{theorem}[Correspondence between marked surfaces and gentle algebras; Proposition 1.21 in \cite{opper2018geometric}]\label{TMarkedSurfaceGentle}
There is a bijection $[(S, M, P), \Delta] \to A(\Delta)$ between the set of homotopy classes of marked surfaces $(S, M, P)$ having no $\color{dark-green}{\circ}$-punctures ($P_{\color{dark-green}{\circ}} = \emptyset$) with an admissible dissection $\Delta$ to the set of isomorphism classes of gentle algebras over $k$.
\end{theorem}
\subsubsection{Models of objects in the derived category}
Henceforth, we assume that the marked surface $(S, M, P)$ has no $\color{dark-green}{\circ}$-punctures, in other words $P_{\color{dark-green}{\circ}} = \emptyset$.
\begin{definition}[$\color{dark-green}{\circ}$-infinite arc; after Definition 1.19 in \cite{opper2018geometric}]\label{DInfiniteArc}
Let $(S, M, P)$ be a marked surface. An \emph{infinite $\color{dark-green}{\circ}$-arc} is a smooth map $\gamma$ from $(0,1)$ to $S \, \backslash \, P$ such that
\begin{itemize}
\item $\lim_{x \to 0} \gamma(x) \in M_{\color{dark-green}{\circ}}$ and $\lim_{x \to 1} \gamma(x) \in P_{\color{red} \bullet}$ (i.e.\ it goes from a $\color{dark-green}{\circ}$-marked point to a $\color{red} \bullet$-puncture) or
\item $\lim_{x \to 0} \gamma(x) \in P_{\color{red} \bullet}$ and $\lim_{x \to 1} \gamma(x) \in P_{\color{red} \bullet}$ (i.e.\ it goes from one $\color{red} \bullet$-puncture to another).
\end{itemize}
\end{definition}
\begin{remark}
The arcs are called infinite because of the they can be for various purposes more appropriately viewed as not really ending in $\color{red} \bullet$-punctures, but rather infinitely wrapping around these $\color{red} \bullet$-punctures in the counter-clockwise direction.
\end{remark}
\begin{definition}[Graded arcs and closed curves; Definition 2.4 in \cite{amiot2019complete} and Definition 2.10 in \cite{opper2018geometric}]\label{DGradedCurves}
Let $(S, M, P)$ be a marked surface, and let $\Delta$ its admissible dissection. Suppose that $\gamma$ is a $\color{dark-green}{\circ}$-arc, $\color{dark-green}{\circ}$-infinite arc, or closed curve in $S \, \backslash \, P$. We assume that $\gamma$ intersects the arcs of $\Delta^{\ast}$ minimally and transversally. A \emph{grading} $f$ on $\gamma$ is a function $f\colon \gamma \cap \Delta^{\ast} \to \mathbb{Z}$, where $\gamma \cap \Delta^{\ast}$ is the totally (if $\gamma$ is an arc) or cyclically (if $\gamma$ is a closed curve) ordered set of intersections of $\gamma$ with the $\color{red} \bullet$-arcs in $\Delta^{\ast}$. The grading $f$ must satisfy the following condition: assume that $q$ succeeds $p$ in $\gamma \cap \Delta^{\ast}$, then $\gamma$ enters a disk enclosed by $\color{red} \bullet$-arcs in $\Delta^{\ast}$ and a segment of $\partial S$ at $p$ and leaves it at $q$; there is a single $\color{dark-green}{\circ}$-point (cf. Proposition \ref{PAdmissibleDissectionProperties}(ii) above for admissible $\color{red} \bullet$-dissections) in the disk or on its boundary, and $f(q) = f(p) + 1$ if the $\color{dark-green}{\circ}$ lies to the left of $\gamma$, or $f(q) = f(p) - 1$ otherwise.
\end{definition}
\begin{remark}
As it is remarked in \cite{opper2018geometric} (Remark 2.11), grading of $\color{dark-green}{\circ}$-arc, $\color{dark-green}{\circ}$-infinite arc, or closed curve, if it exists, is fully determined by its value at a single intersection, and $\color{dark-green}{\circ}$-arc and $\color{dark-green}{\circ}$-infinite arc can always be equipped with a grading.
\end{remark}
\begin{theorem}[Correspondence between objects of $\Db(A(\Delta)\textrm{-}\mathrm{mod})$ and graded arcs and closed curves; Theorem 2.12 in \cite{opper2018geometric}]\label{TObjectsGradedCurves}
Let $(S, M, P)$ be a marked surface, and let $\Delta$ its admissible dissection. Then the following holds:
\begin{itemize}
\item Graded $\color{dark-green}{\circ}$-arcs are in bijection with isomorphism classes of string complexes in $\Db(A(\Delta)\textrm{-}\mathrm{mod})$.
\item The pairs of graded closed curves together with isomorphism classes of indecomposable finite-dimensional $k[X^{\pm1}]$-modules are in bijection with isomorphism classes of band complexes in $\Db(A(\Delta)\textrm{-}\mathrm{mod})$.
\item Graded $\color{dark-green}{\circ}$-infinite arcs are in bijection with isomorphism classes of infinite string complexes in $\Db(A(\Delta)\textrm{-}\mathrm{mod})$.
\end{itemize}
\end{theorem}
\begin{remark}
The proof of \ref{TObjectsGradedCurves} in \cite{opper2018geometric} consists of assigning a homotopy string, homotopy band, or infinite homotopy string to $\gamma$, a $\color{dark-green}{\circ}$-arc, $\color{dark-green}{\circ}$-infinite arc, or closed curve, respectively. Vertices of unfolded diagram of the homotopy string, homotopy band, or infinite homotopy string correspond to successive intersections with the dual dissection $\Delta^{\ast}$ (grading may be assigned as in Definition~
\ref{DGradedCurves}). Arrows between the vertices are constructed as follows: if two successive vertices in the unfolded diagram correspond to $p$ and $q$ in $\gamma \cap \Delta^{\ast}$, then $\gamma$ enters a disk $\mathcal{D}$ enclosed by $\color{red} \bullet$-arcs in $\Delta^{\ast}$ and a segment of $\partial S$ at $p$ and leaves it at $q$. Inside the disk $\mathcal{D}$, $\gamma$ crosses $\color{dark-green}{\circ}$-arcs $\gamma_{n+1}, \dots, \gamma_1$ in that order, whose endpoint is the single $\color{dark-green}{\circ}$ on the segment of $\partial S$ of the boundary of the disk. Suppose that the single $\color{dark-green}{\circ}$ lies to the left of $\gamma$ in the disk $\mathcal{D}$. By Definition \ref{DDissectionAssociatedAlgebra} above, there are arrows $a_i: \gamma_{i+1} \to \gamma_i$ in $Q(\Delta)$ for $1 \leq i \leq n$ such that $a_n \dots a_i$ is a direct string for $Q(\Delta)$. Therefore, there is a direct string $a_n \dots a_1$ between the vertices corresponding to successive $p, q \in \gamma \cap \Delta^{\ast}$. If the single $\color{dark-green}{\circ}$ lies to the left of $\gamma$ in the disk $\mathcal{D}$, an inverse string that sits between the vertices corresponding to $p, q \in \gamma \cap \Delta^{\ast}$ is obtained similarly.
\end{remark}
\subsubsection{Models of morphisms in the derived category}
In Theorem 3.1 in \cite{opper2018geometric}, a correspondence between basis morphisms in $\Db(A(\Delta)\textrm{-}\mathrm{mod})$ between indecomposable objects, in the sense of \cite{arnesen2016morphisms} (graph maps, quasi-graph maps, singleton single maps, and singleton double maps; cf. Theorem \ref{TBasisMaps}), and \textit{oriented graded intersections} of the corresponding $\color{dark-green}{\circ}$-arcs, $\color{dark-green}{\circ}$-infinite arcs, and closed curves is established. For the purposes of this text, we do not need the full detail of the characterization in \cite{opper2018geometric}; it is enough to note that if $(\gamma_1, f_1)$ is an arc, infinite arc or a closed curve, and so is $(\gamma_2, f_2)$, the canonical maps from $P_{(\gamma_1, f_1)}$ to $P_{(\gamma_2, f_2)}$ correspond to intersections and common endpoints of $\gamma_1$ and $\gamma_2$ where their gradings agree locally.
An intersection $p \in \gamma_1 \cap \gamma_2$ or a common endpoint $p$ lies in a disk $\mathcal{D}$ enclosed by $\color{red} \bullet$-arcs in $\Delta^{\ast}$ and a segment of $\partial S$ or on the segment of $\partial S$ on the boundary of $\mathcal{D}$, respectively. Denote $J$ to be the set of $\gamma_1 \cap \Delta^{\ast}$ and $\gamma_2 \cap \Delta^{\ast}$ adjacent to $p$ on $\gamma_1$ and $\gamma_2$, respectively. We say that the gradings $f_1$ and $f_2$ agree locally if there are $q, r \in J$ such that $r$ immediately, among elements of $J$, follows $q$ in the counter-clockwise order around $p$, $q \in \gamma_1 \cap \Delta^{\ast}$, $r \in \gamma_2 \cap \Delta^{\ast}$, and $f_1(q) = f_2(r)$. This is summarised the following figure, which is an adaptation of Figure 9 in Remark 3.8 in \cite{opper2018geometric}:
\begin{figure}[H]
\centering
\begin{tikzpicture}[x = 1cm, y = 1cm]
\node[regular polygon, regular polygon sides = 8, minimum size = 4 cm] (Oct) at (0,0) {};
\foreach \x in {1,...,8}
{\node[coordinate, name = {OctNode\x}] at (Oct.corner \x) {};}
\draw[red, dotted] (OctNode1) -- (OctNode2);
\draw[red] (OctNode2) -- (OctNode3) node[coordinate, midway, name = Int2] {};
\draw[red, dotted] (OctNode3) -- (OctNode4);
\draw[red] (OctNode4) -- (OctNode5);
\draw[red] (OctNode6) -- (OctNode7);
\draw[red, dotted] (OctNode7) -- (OctNode8);
\draw[red] (OctNode8) -- (OctNode1) node[coordinate, midway, name = Int1] {};
\draw (OctNode5) to[out = 45, in = 135] node[coordinate, midway, name = Green] {} (OctNode6);
\fill[pattern = north east lines] (OctNode5) to[out = 45, in = 135] (OctNode6) -- cycle;
\draw[blue] (0,0) -- ($1.5*(Int1)$) node[at end, right, black] {$\gamma_1$};
\draw[blue] (0,0) -- ($1.5*(Int2)$) node[at end, left, black] {$\gamma_2$};
\foreach \x in {1,...,8}
{\draw[red, fill = red] (OctNode\x) circle (0.1);}
\draw[green, fill = white] (Green) circle (0.1);
\draw (Int1) circle (0.1);
\draw (Int2) circle (0.1);
\node[above = 0.1cm] at (Green) {$p$};
\node[right = 0.1cm] at (Int1) {$q$};
\node[left = 0.1cm] at (Int2) {$r$};
\end{tikzpicture}
\caption{The graded curves $(\gamma_1, f_1)$ and $(\gamma_2, f_2)$ intersect at a point which lies inside a connected component $\mathcal{D}$ of the complement of $\Gamma^*$ in $S \backslash P$, a disk, or they meet at the unique $\color{dark-green}{\circ}$-point $p$ which lies on its boundary. Their intersections, $r, q$, with $\color{red} \bullet$-arcs in $\Gamma^*$ which form the boundary have equal grading, $f_1(q) = f_2(r),$ and $r$ follows $q$ in the counter-clockwise direction as viewed from $p$.}
\label{FGradedIntersection}
\end{figure}
There are a two exceptions to this: if $P_{(\gamma, f)}$ is a one-dimensional band complex, the identity on $P_{(\gamma, f)}$ and the map $\xi: P_{(\gamma, f)} \to \tau P_{(\gamma, f)}[1] = P_{(\gamma, f)}[1]$ from the Auslander-Reiten triangle $\tau P_{(\gamma, f)} \to E \to P_{(\gamma, f)} \overset{\xi}{\to} \tau P_{(\gamma, f)}[1]$ are not represented by such an intersection.
Moreover, as an easy corollary formulated in Remark 3.8 in \cite{opper2018geometric}, we obtain that:
\begin{itemize}
\item If $\gamma_1$, $\gamma_2$ intersect at $p \in S \, \backslash \, P$ and $f_1$ is a grading on $\gamma_1$, then there exists a grading $f_2$ on $\gamma_2$ such that $p$ corresponds to a map from $P_{(\gamma_1, f_1)}$ to $P_{(\gamma_2, f_2)}$ and from $P_{(\gamma_2, f_2)}$ to $P_{(\gamma_1, f_1[1])}$.
\item If $\gamma_1$, $\gamma_2$ intersect at $p \in M_{\color{dark-green}{\circ}}$ and $f_1$ is a grading on $\gamma_1$, then $p$ corresponds to a map either from $P_{(\gamma_1, f_1)}$ to $P_{(\gamma_2, f_2)}$ or from $P_{(\gamma_2, f_2)}$ to $P_{(\gamma_1, f_1)}$ for some unique grading $f_2$ on $\gamma_2$. There is a morphism from $P_{(\gamma_1, f_1)}$ to $P_{(\gamma_2, f_2)}$ if $\gamma_2$ follows $\gamma_1$ in the counter-clockwise order around $p$, and vice versa.
\end{itemize}
Calculation mapping cones of the basis maps between the indecomposable objects can also be performed in the geometric model as illustrated in Theorem 4.1 in \cite{opper2018geometric}. This result on mapping cones is a geometric analogue of the characterization of mapping cones of the basis maps as given in \cite{ccanakcci2019mapping} and \cite{ccanakcci2021corrigendum}.
A basis map between two indecomposable objects is represented by an oriented graded intersection of corresponding graded curves in the geometric model; as an object the mapping cone of a basis map is given by resolving the intersection of the corresponding graded curves (the mapping cone may not be indecomposable). Specifically, if the basis map is represented by an oriented graded intersection at a common endpoint, its mapping cone is given by concatenation of the two graded curves (cf. Propositions 1.6, 1.20 and 3.7 in \cite{amiot2019complete}).
\section{Conditions on approximations of projectives}
The purpose of this section is to find necessary conditions for approximations of indecomposable projectives arising from a semiorthogonal decomposition of bounded derived category of $\Lambda$, a gentle algebra. Specifically, given $\langle \mathcal{L}, \mathcal{R} \rangle$, a semiorthogonal decomposition of $\Db(\Lambda \textrm{-}\mathrm{mod})$, there is a distinguished triangle $R_i \to P(i) \to L_i \to R_i[1]$ with $R_i \in \mathcal{R}$ and $L_i \in \mathcal{L}$ for every vertex $i$ of the underlying quiver, and we seek to understand what $R_i$ and $L_i$ may look like for all vertices $i$. In this section, we show that there are very restrictive conditions on such approximations, a fact that we exploit in the next section.
First, given $f\colon M \to N$, a graph map, singleton single map or singleton double map between string complexes, we study the situation when there is corresponding map $f'\colon N \to M[1]$ and what type this map is.
Our considerations are motivated by Remark 3.8 in \cite{opper2018geometric}, which implies that such a map $f$ has a corresponding map $f'$ if and only if it is represented by an oriented graded intersection of corresponding $\color{dark-green}{\circ}$-arcs lies in interior of the associated marked surface. This motivation may also help explain the choice of terminology in the next definition.
\begin{definition}[Boundary graph maps]\label{DBoundaryGrapMaps}
Suppose that $f\colon M \to N$ is graph map (as in Definition \ref{DGraphMaps} above) between string complexes in $\Db(\Lambda\textrm{-}\mathrm{mod})$; we say that it is \emph{boundary graph map} if it arises from a maximal overlap of the unfolded diagrams as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_L} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_p} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_{p-1}} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{\ell_2} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_1} & \bullet \arrow[equal]{d}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_L} & \bullet \arrow[dash]{r}{\ell_p} & \bullet \arrow[dash]{r}{\ell_{p-1}} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{\ell_2} & \bullet \arrow[dash]{r}{\ell_1} & \bullet
\end{tikzcd}$$
where $m_L$ and $n_L$ satisfy an endpoint condition (LG1) or $m_L, n_L$ satisfy (LG2) and $n_L m_L \neq 0$ if they are both non-zero.
\end{definition}
\begin{remark}
Note that situation that $n_L$ and $m_L$ are both non-zero and satisfy (LG2) with $n_L m_L = 0$ may not happen unless $p = 0$, meaning that the overlap between the unfolded diagram is trivial, formed by a single vertex.
\end{remark}
\begin{lemma}\label{LBoundaryGrapMaps}
Let $f\colon M \to N$ be a graph between string complexes in $\Db(\Lambda\textrm{-}\mathrm{mod})$ given by a maximal overlap of unfolded diagrams from $M$ to $N$ by Definition 3.9 in \cite{arnesen2016morphisms}. The same overlap of unfolded diagrams considered in the other direction, from $N$ to $M$, gives rise to a quasi-graph map $f'\colon N \to M[1]$ if and only if $f$ is not a boundary graph map.
\end{lemma}
\begin{proof}
At first, suppose that $p > 0$, in other words the maximal overlap between unfolded diagrams of $M$ and $N$ is given by a non trivial homotopy string. By Definition 3.9 in \cite{arnesen2016morphisms}, the overlap of unfolded diagrams of $M$ and $N$ is as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_L} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_p} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_{p-1}} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{\ell_2} & \bullet \arrow[equal]{d} \arrow[dash]{r}{\ell_1} & \bullet \arrow[equal]{d} \arrow[dash]{r}{m_R} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_L} & \bullet \arrow[dash]{r}{\ell_p} & \bullet \arrow[dash]{r}{\ell_{p-1}} & \bullet \arrow[dash, dotted]{r} & \bullet \arrow[dash]{r}{\ell_2} & \bullet \arrow[dash]{r}{\ell_1} & \bullet \arrow[dash]{r}{n_R} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
such that and one of (LG1), (LG2) and one of (RG1), (RG2) endpoint conditions are met. Suppose $f$ is non-boundary, so $m_L \neq n_L$ and $m_R \neq n_R$. We show that if we consider this overlap of unfolded diagrams from $N$ to $M$ instead, none of the endpoint conditions hold.
Suppose that (LG1) holds for the overlap from $M$ to $N$; arrows $m_L$ and $n_L$ are either both direct or inverse, and there exists a non-stationary path $p_L$ such that $m_L = p_L n_L$ if they are direct and $m_L p_L = n_L$ if they are inverse. This means that $n_L$ is a sub-path of $m_L$ if arrows $m_L$ and $n_L$ are direct and $m_L$ is a sub-path of $n_L$ if arrows $m_L$ and $n_L$ are inverse.
If (LG2) holds for the overlap from $M$ to $N$, then arrows $m_L$ and $n_L$ are neither both direct nor inverse, and $m_L$ being non-zero implies that it is inverse and $n_L$ being non-zero implies that it is direct.
Therefore, if (LG1) holds for the overlap from $M$ to $N$, it cannot hold for the overlap from $N$ to $N$ as this would imply that $n_L = m_L$, and (LG2) cannot hold here either since the arrows $n_L$ are $m_L$ are equally oriented; on the other hand, should (LG2) hold for the overlap from $M$ to $N$, (LG1) cannot hold due to different orientation of $n_L$ and $m_L$ and neither can (LG2), which would imply that $n_L = m_L = 0$.
The right endpoint conditions are discussed dually. Since none of the endpoint conditions are met for the overlap from $N$ to $M$, it yields a quasi-graph map from $N$ to $M[1]$ by Definition \ref{DQuasiGraphMaps} above.
Should $f$ be boundary with $m_R = n_R = 0$, the overlap considered from $N$ to $M$ does not yield a quasi-graph map since it satisfies (RG2).
Suppose next that $p=0$, i.e.\ the overlap is trivial.
The case analysis is more complicated then since the overlap from $N$ to $M$ can be depicted in two different ways and we must check whether \emph{either of them} yields a quasi-graph map:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_L} & \bullet \arrow[equal]{d} \arrow[dash]{r}{n_R} & \bullet \arrow[dash, decorate = true]{r} & {} &
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_L} & \bullet \arrow[equal]{d} \arrow[dash]{r}{n_R} & \bullet \arrow[dash, decorate = true]{r} & {} &
\\
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_L} & \bullet \arrow[dash]{r}{m_R} & \bullet \arrow[dash, decorate = true]{r} & {} &
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_R} & \bullet \arrow[dash]{r}{m_L} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
Should $m_R=0=n_R$ and (LG1) hold, it is easy to see that none of these yields a quasi-graph map, as the following diagrams show that when both $m_L$ and $n_L$ are direct, then (RG2) holds (the case with inverse homotopy letters is similar):
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_L} & \bullet \arrow[equal]{d} & {} &
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_L} & \bullet \arrow[equal]{d} & & {} &
\\
M: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{m_L} & \bullet & {} &
M: & & \bullet \arrow[<-]{r}{m_L} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
Similarly, the situation where $m_R=0=n_R$ and (LG2) hold with both $m_L$ and $n_L$ non-zero leads to unfolded diagrams:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_L} & \bullet \arrow[equal]{d} & {} &
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_L} & \bullet \arrow[equal]{d} & & {} &
\\
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[<-]{r}{m_L} & \bullet & {} &
M: & & \bullet \arrow{r}{m_L} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
The first of these never defines a quasi-graph map as (RG2) is satisfied, and the other defines a quasi-graph map precisely when $n_Lm_L=0$, that is when $f$ is not a boundary map. If one of $m_L$ and $n_L$ or both of them vanish, the arguments are similar.
Suppose finally that $p=0$ and one of $m_L$ and $n_L$ is non-zero, as is one of $m_R$ and $n_R$.
If (LG2) and (RG2) hold simultaneously for an overlap which gives the graph map, we assume without loss of generality that $n_L m_L \neq 0$ or one of them is zero and $n_R m_R \neq 0$ or one of them is zero (cf. Definition 5.1 in \cite{ccanakcci2019mapping}). This can be always arranged for by inverting one of the unfolded diagrams.
This implies that $n_Lm_R=0=n_Rm_L$ and the diagram define a quasi-graph map $N\to M[1]$.
Suppose we are in the situation where (LG1) and (RG2) simultaneously hold. If both $m_R$ and $n_R$ are non-zero, we have the following diagram
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r} & \bullet \arrow[dotted]{d}[']{p_L} \arrow{r}{m_L} & \bullet \arrow[equal]{d} \arrow{r}{m_R} & \bullet \ar[dash]{r} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r} & \bullet \arrow{r}{n_L} & \bullet \ar[<-]{r}{n_R} & \bullet \arrow[dash]{r} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
Then we can build the following well-defined single map from $N$ to $M[1]$ representing a quasi-graph map:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r} & \bullet \arrow{d}{n_L} \arrow{r}{n_L} & \bullet \arrow[dash]{r} & \bullet \arrow[dash, decorate = true]{r} & {} \\
M[1]: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{m_L} & \bullet \arrow[dash]{r} & \bullet \arrow[dash]{r} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
In fact, the overlap viewed from $N$ to $M$ violates (LG1) as well as (RG2) as long as at least one of $m_R$ and $n_R$ is non-zero. Hence, the quasi-graph map can be defined in the same way also in this case.
Finally, analogous arguments easily take care also of the case where $p=0$ and (LG1) and (RG1) are simultaneously satisfied for the overlap from $M$ to $N$.
\end{proof}
\begin{lemma}\label{LDualMorphisms}
Suppose that $f\colon M \to N$, between $M, N \in \Db(\Lambda\textrm{-}\mathrm{mod})$ string complexes is a singleton single map (as in Definition \ref{DSingletonSingleMaps}) or a singleton double map (as in Definition \ref{DSingletonDoubleMaps}). If $f$ is a singleton single map is of type $(ii)$, then there is a singleton single map $f'\colon N \to M[1]$ of type $(iii)$, and vice versa. If $f$ is a singleton double map, then there is a singleton single map $f'\colon N \to M[1]$ of type $(iv)$, and vice versa.
\end{lemma}
\begin{proof}
Suppose we have a singleton single map $f\colon M \to N$ in configuration $(ii)$ as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[]{d}{p} \arrow{r}{m_{i}}[swap]{=pp _R} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & & \bullet \arrow[dash]{r}{n_{i}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
such that $n_{i+1}$ if inverse and viewed as an path does not start with $p$. We claim that there exists the following single map $f'$ of type $(iii)$:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N: & & \bullet \arrow[]{d}{p_R} \arrow[dash]{r}{n_{i}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
M[1]: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{m_{i}}[swap]{=pp _R} & \bullet \arrow[dash]{r}{m_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
Suppose that $n_{i}$ is inverse, it corresponds to a path from its target to its source. Also, $p$ corresponds to such a path, and has the same starting vertex as $n_{i}$. We want to show that $p$ and $n_{i}$ do not start with the same arrow. If they did, due to $\Lambda$ being gentle, one would need to be a sub-path of the other. The situation of $p$ being a not necessarily proper sub-path of $n_{i}$ is excluded by definition. On the other hand, provided that $n_{i}$ is a proper sub-path of $p$, we show that the map $f$ is null-homotopic. Consider the following homotopy with $p'$ a non-stationary path for which $n_i p' = p$ as paths in $Q$:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[]{d}{p} \arrow[dotted]{dl}[']{p'} \arrow{r}{m_{i}}[swap]{=pp _R} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_{i}} & \bullet &
\end{tikzcd}$$
If $m_{i+1}$ is direct, we have that $0 = m_{i+1} m_{i} = m_{i+1} p p_R = m_{i+1} p' n_{i} p_R$, and, given that $p' n_{i} p_R \neq 0$, $m_{i+1} p'$ needs to be zero. So we have a homotopy between $f$ and the zero map.
Because $n_{i}$ and $p$ do not start with the same arrow, $p_R n_{i}$ as path in $Q$ needs to lie in $I$, and $n_{i} p_R = 0$ as maps between projectives over $\Lambda$.
Now, assume that $m_{i-1}$ is direct; we need to show that $p_R m_{i-1} = 0$. We already have that $p p_R m_{i-1} = 0$. Since $p p_R \neq 0$, $p_R m_{i-1}$ equals zero. The map $f'$ is therefore well-defined. The other direction follows dually.\\\\
Let $f\colon M \to N$ be a singleton double map. Such map gives rise to the following unfolded diagrams:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[']{d}{p_L} \arrow{r}{m_i}[swap]{=p_L p'} & \bullet \arrow[]{d}{p_R} \arrow[dash]{r}{m_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow{r}{n_i}[swap]{=p' p_R} & \bullet \arrow[dash]{r}{n_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
The non-stationary path $p'$ gives rises to the following well-defined singleton single map in the configuration $(iv)$:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow[]{d}{p'} \arrow{r}{n_i}[swap]{= p' p_R} & \bullet \arrow[dash, decorate = true]{r} & {}\\
M[1]: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{m_i}[swap]{= p_L p'} & \bullet \arrow[dash]{r}{m_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
The other direction follows dually.
\end{proof}
Now, in order to establish some constraints on approximations of indecomposable projectives in semiorthogonal decompositions, we study the following problem and its dual variant: given a map $g: N \to P$ with $N$ an indecomposable object in $\Db(\Lambda \textrm{-}\mathrm{mod})$ and $P$ corresponding to shift of an indecomposable projective, what maps $f\colon M \to N$ compose to zero with $g$? It turns out that we may find some suitable maps with the property for all types of indecomposables in $\Db$, which imposes strong restrictions on approximations of indecomposable projectives.
\begin{proposition}\label{PStringsProjectives}
Suppose that $f\colon M \to N$, between $M, N \in \Db(\Lambda\textrm{-}\mathrm{mod})$ string complexes.
\begin{enumerate}
\item If $f$ is non-zero and homotopically equivalent to a non-singleton single or double map or to a singleton single map in configurations $(iii)$ or $(iv)$ and there is a map $g: N \to P$ such that $P$ is projective and concentrated in single degree $i$, then $fg$ is null-homotopic.
\item Dually, if $f$ is non-zero and homotopically equivalent to a non-singleton single or double map or to a singleton single map in configurations $(ii)$ or $(iv)$ and there is a map $g: P \to M$ such that $P$ is projective and concentrated in single degree $i$, then $gf$ is null-homotopic.
\end{enumerate}
\end{proposition}
\begin{proof}
In order to prove this proposition, we consider $M, N$, and $P$ as complexes of projectives and the maps between them as maps of complexes. We discuss three possible cases:
\begin{enumerate}[(i)]
\item \textit{The map $f\colon M \to N$ is a single map that is not singleton and not null-homotopic.} The discussion in the proof of Proposition 4.8 in \cite{arnesen2016morphisms} yields that there are two possible configurations for such a map in terms of unfolded diagrams of $M$ and $N$:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet & \bullet \arrow[dash]{l}[']{m_{i+1}} \arrow[]{d}{p} \arrow{r}{m_{i} = p} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow[dash]{r}{n_{i}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
or
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet & \bullet \arrow[dash]{l}[']{m_{i+1}} \arrow[]{d}{p} \arrow[dash]{r}{m_{i}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_{i+1} = p} & \bullet \arrow[dash]{r}{n_{i}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
If the target of $p$ is of degree other than $i$, then $fg$ is zero. Assume, therefore, that the target of $P$ is of degree $i$. In terms of the unfolded diagram of $N$, $g$ can be thought of being given by maps of projectives in degree $i$ to the projective at degree $i$ in $P$. For discussing the composition with $g: N \to P$, it suffices to examine composition of $p$ with the map $q$ going from its target projective to the projective at degree $i$ in $P$.
In unfolded diagrams, this gives rise to the following two situations:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet & \bullet \arrow[dash]{l}[']{m_{i+1}} \arrow[]{d}{p} \arrow{r}{m_{i} = p} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow[]{d}{q} \arrow[dash]{r}{n_{i}} & \bullet \arrow[dash, decorate = true]{r} & {} \\
P: & & \bullet &
\end{tikzcd}$$
or
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet & \bullet \arrow[dash]{l}[']{m_{i+1}} \arrow[]{d}{p} \arrow[dash]{r}{m_{i}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_{i+1} = p} & \bullet \arrow[]{d}{q} \arrow[dash]{r}{n_{i}} & \bullet \arrow[dash, decorate = true]{r} & {} \\
P: & & \bullet &
\end{tikzcd}$$
In the second situation, $pq$ needs to zero so that $g$ is a well-defined map of complexes, and, thus, $fg = 0$. In the first situation, provided that $pq \neq 0$, we can construct a homotopy with a single non-zero component showing $fg$ is null-homotopic as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet & \bullet \arrow[dash]{l}[']{m_{i+1}} \arrow[]{d}{pq} \arrow{r}{m_{i} = p} & \bullet \arrow[dotted]{dl}{q} \, \arrow[dash, decorate = true]{r} & {} \\
P: & & \bullet &
\end{tikzcd}$$
Due to the fact that $pq \neq 0$, any potentially inverse $m_{i+2}$ has to compose to zero with $q$, so this is a homotopy with a zero map.
\item \textit{The map $f\colon M \to N$ is a double map that is not singleton and not null-homotopic.} By Lemma 4.13 in \cite{arnesen2016morphisms}, there is only one possible configuration of unfolded diagrams:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[']{d}{p_L} \arrow[]{r}{m_i = p_L} & \bullet \arrow[]{d}{p_R} \arrow[dash]{r}{m_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow[]{r}{n_i = p_R} & \bullet \arrow[dash]{r}{n_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
We proceed similarly as in the previous case. Target of $p_L$ or $p_R$ needs to be in degree $i$; otherwise, the composition $fg$ is trivial. Suppose that $q$ is the component of $g$ going from target of $p_L$ or $p_R$ to the projective at degree $i$ in $P$.
First, we deal with the situation that $p_L$ is in degree $i$. This yields the following unfolded diagram:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[']{d}{p_L} \arrow[]{r}{m_i = p_L} & \bullet \arrow[]{d}{p_R} \arrow[dash]{r}{m_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow[]{d}{q} \arrow[]{r}{n_i = p_R} & \bullet \arrow[dash]{r}{n_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
P: & & \bullet & &
\end{tikzcd}$$
As in the previous case, we can exhibit a homotopy between $fg$ and the zero map provided that $p_L q \neq 0$ as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[']{d}{p_L q} \arrow[]{r}{m_i = p_L} & \bullet \arrow[dotted]{dl}{q} \arrow[dash]{r}{m_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
P: & & \bullet & &
\end{tikzcd}$$
The other case, in which $p_R$ is degree $i$ is simpler. We have the following situation:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{m_{i+1}} & \bullet \arrow[']{d}{p_L} \arrow[]{r}{m_i = p_L} & \bullet \arrow[]{d}{p_R} \arrow[dash]{r}{m_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow[]{r}{n_i = p_R} & \bullet \arrow[]{d}{q} \arrow[dash]{r}{n_{i-1}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
P: & & & \bullet &
\end{tikzcd}$$
In order for $g$ to be a proper map of complexes, $p_R q$ needs to be zero, which is why $fg = 0$.
\item \textit{The map $f\colon M \to N$ is a singleton single map in configuration $(iii)$ or $(iv)$.} The corresponding unfolded diagram looks as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
M: & \arrow[dash, decorate = true]{l} \bullet & \bullet \arrow[dash]{l}[']{m_{i+1}} \arrow[]{d}{p} \arrow[dash]{r}{m_{i}} & \bullet \arrow[dash, decorate = true]{r} & {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{n_{i+1} = p_L p} & \bullet \arrow[dash]{r}{n_{i}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
Again, the target of $p$ needs to be equal to the source of $q$; the composition $fg$ is trivial otherwise. Denote $q$ the component of $g$ going from target of $p$ to the projective at degree $i$ in $P$. We have the following situation:
As $g$ is a map of complexes, we have that $p_L p q = 0$. By Definition \ref{DSingletonSingleMaps} above, $p$, $p_L$ are non-stationary paths in the quiver $Q$ such that no sub-path of $p p_L$ is in the ideal $I$. The map $q$ can be thought of as a linear combination of paths $\sum_j \alpha_j q_j$ from its target to its source.
Recall that $kQ$ has basis of paths in $Q$ as vector space over $k$ and that $I$ is generated as an ideal of $kQ$ by paths of length two as $\Lambda$ is gentle. We observe that $I$ as vector space over $k$ is generated by paths that lie in $I$. Suppose that $\sum_j \alpha_j q'_j \in I$ for paths $q'_j$ in $Q$ and scalars $\alpha_j \in k$. As $I$ is generated by paths in $I$, we can express $\sum_j \alpha_j q'_j$ as a linear combination of paths in $I$. However, paths form a basis $kQ$, so every $q_i$ needs to lie in $I$ already.
The fact that $p_L p q = 0$ means that $\sum_j \alpha_j q_j p p_L \in I$. By our discussion, every $q_j p p_L$ lies in $I$. Since no sub-path of $p p_L$ lies in $I$, a sub-path of $q_j p$ needs to lie in $I$. This yields that $p q_j = 0$ as maps between projectives, which yields that $p q = 0$. Hence, $fg = 0$.
\end{enumerate}
The other part of this proposition follows by dual argument.
\end{proof}
\begin{proposition}\label{PBandsProjectives}
Suppose $N \in \Kb(\Lambda\textrm{-}\mathrm{proj})$ is a band complex. Denote $f\colon N[-1] \to N$ the quasi-graph map corresponding to the identity on $N[-1]$. If there is a map $g$ from $N$ to $P$ or vice versa with $P$ being projective concentrated in a single degree $i$, then $fg$ and $gf[1]$, respectively, is homotopically trivial.
\end{proposition}
\begin{proof}
This follows directly from the fact that the map $f$ is a part of an almost split triangle and that $g$ is neither a split monomorphism nor a split epimorphism.
If $N$ is a one-dimensional band complex, we can also give a direct proof as follows.
Suppose that a component of the map $g: N \to P$ is given as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+1}} & \bullet \arrow[]{d}{q} \arrow[dash]{r}{n_{i}} & \bullet \arrow[dash, decorate = true]{r} & {} \\
P: & & \bullet &
\end{tikzcd}$$
In this case, $q$ may be any map between the projectives which are its source and target, respectively. Up to inverting the upper string, $n_{i+1}$ is inverse.
Consider the following representative of $f\colon N[-1] \to N$, which is homotopically non-trivial:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N[-1]: & \arrow[dash, decorate = true]{l} \bullet & \bullet \arrow{l}[']{n_{i+1}} \arrow{d}{n_{i+1}} \arrow[dash]{r}{n_{i}} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow[dash]{r}{n_{i+2}} & \bullet & \bullet \arrow{l}[']{n_{i+1}} \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
Clearly, $fg = 0$. The other part of this proposition follows by dual argument.
\end{proof}
\begin{proposition}\label{PInfiniteStringsProjectives}
Suppose $N \in \Db(\Lambda\textrm{-}\mathrm{mod})$ be an infinite string complex, possibly two-sided. If there is a map $g$ from $N$ to $P$ or vice versa with $P$ being concentrated in a single degree $i$, then there exist a map $f\colon N[-j] \to N$ or vice versa such that $fg$ and $gf[j]$, respectively, is homotopically trivial.
\end{proposition}
\begin{proof}
For the purposes of this proof, we consider $N$ and $P$ as complexes of projectives. We assume that the unfolded diagram of $N$ consists of a string $w = \begin{tikzcd} \bullet \arrow{r}{w_n} & \bullet \arrow[dotted, dash]{r} & \bullet \arrow[dash]{r}{w_1} & \bullet \end{tikzcd}$ which is preceded by infinitely many copies of a string $a = \begin{tikzcd} \bullet \arrow{r}{a_m} & \bullet \arrow[dotted, dash]{r} & \bullet \arrow{r}{a_1} & \bullet \end{tikzcd}$ formed from a repetition-free cycle $a_m \dots a_n$ with full relations in the underlying quiver, so it has the following form:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{a_m} & \bullet \arrow[dotted, dash]{r} & \bullet \arrow{r}{a_1} & \bullet \arrow{r}{w_n} & \bullet \arrow[dotted, dash]{r} & \bullet \arrow[dash]{r}{w_1} & \bullet
\end{tikzcd}$$
Note that, by Definition \ref{DInfiniteHomotopyStrings}, the degree of projectives right to the source of $w_n$ is at least equal to $d$, the degree of the source of $w_n$, and that $a_m \dots a_1$ is a repetition-free cyclic path in $Q$ with full relations. We also observe that $w_n = a_m w'$ for a non-trivial path $w'$ as $a_1 w_n = 0$ and $\Lambda$ is gentle.
For any $b \in \mathbb{N}$, we consider the following map from $N[-bm]$ to $N$ denoted $f_b$:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N[-bm]: & \arrow[dash, decorate = true]{l} \bullet \arrow[equals]{d} \arrow{r}{a_m} & \bullet \arrow[equals]{d} \arrow[dotted, dash]{r} & \bullet \arrow[equals]{d} \arrow{r}{a_1} & \bullet \arrow[equals]{d} \arrow{r}{a_m} & \bullet \arrow{d}{w'} \arrow{r}{a_1} & \bullet & \arrow[dash, decorate = true]{l} {}\\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{a_m} & \bullet \arrow[dotted, dash]{r} & \bullet \arrow{r}{a_1} & \bullet \arrow{r}{w_n} & \bullet \arrow[dash]{r}{w_{n-1}} & \bullet \arrow[dash, decorate = true]{r} & {}
\end{tikzcd}$$
This is a graph map as it satisfies endpoint conditions (LG$\infty$) and (RG1) in Definition \ref{DGraphMaps}. The former condition holds automatically; whereas, the latter condition translates to $w_n = a_m w'$, which is established above.
At first, consider a map $g: P \to N$. Because $P$ is in degree $i$, it is possible to find $b \in \mathbb{N}$ large enough that $f\colon N \to N[bm]$ as above is non-zero only in degrees strictly smaller that $i$; therefore, $gf$ equals zero.
Second, we deal with maps of type $g: N \to P$, where $P$ is concentrated in a single degree $i$. Suppose first that $i < d$. Because all letters to the left of the source of $w_1$ of degree $d$ are direct and because all projectives to the right of the source of $w_1$ have degree at least $d$, situation is then the following:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{a_j} & \bullet \arrow[]{d}{q} \arrow{r}{a_{j-1}} & \bullet \arrow[dotted, dash]{r} & \bullet \arrow{r}{a_1} & \bullet \arrow{r}{w_1} & \bullet \arrow[dotted, dash]{r} & \bullet \arrow[dash]{r}{w_n} & \bullet\\
P: & & \bullet & & & & & & &
\end{tikzcd}$$
In order for $g$ to be properly defined, we need to have $q = a_{j-1}q'$ as $a_{j-1}$ is an arrow and $\Lambda$ is gentle. The map $q'$ is a homotopy between $g$ and the zero map.
Suppose that $i > d+1$, now; then the map $f_1$ defined above is non-zero only in degrees at most $d+1$, which means that $f_1 g = 0$.
Now, we discuss the situation of $i = d$. We consider the composition $f_1 g$, and the situation is as follows:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N[-m]: & \arrow[dash, decorate = true]{l} \bullet \arrow[equals]{d} \arrow{r}{a_m} & \bullet \arrow[equals]{d} \arrow[dotted, dash]{r} & \bullet \arrow[equals]{d} \arrow{r}{a_1} & \bullet \arrow[equals]{d} \arrow{r}{a_m} & \bullet \arrow{d}{w'} \arrow{r}{a_1} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{a_m} & \bullet \arrow[dotted, dash]{r} & \bullet \arrow{r}{a_1} & \bullet \arrow[]{d}{q} \arrow{r}{w_n} & \bullet \arrow[dash]{r}{w_{n-1}} & \bullet \arrow[dash, decorate = true]{r} & {} \\
P: & & & & \bullet & & &
\end{tikzcd}$$
Although there may be some other components to $g$ other than $q$, their sources lie to the right from the source of $q$, which renders them irrelevant since they compose to zero with $f_1$. Because $a_1 q$ composes to zero, by a similar argument as above, we conclude that $q = a_m q'$ and that $q'$ is the null homotopy for $f_1 g$.
Finally, suppose that $i = d+1$. The following diagram describes the situation:
$$\begin{tikzcd}[arrows = {decorate = false, decoration={snake, segment length=2mm, amplitude=0.25mm}}, /tikz/column 1/.append style={anchor=base east}]
N[-m]: & \arrow[dash, decorate = true]{l} \bullet \arrow[equals]{d} \arrow{r}{a_m} & \bullet \arrow[equals]{d} \arrow[dotted, dash]{r} & \bullet \arrow[equals]{d} \arrow{r}{a_1} & \bullet \arrow[equals]{d} \arrow{r}{a_m} & \bullet \arrow{d}{w'} \arrow{r}{a_1} & \bullet \arrow[dash, decorate = true]{r} & {} \\
N: & \arrow[dash, decorate = true]{l} \bullet \arrow{r}{a_m} & \bullet \arrow[dotted, dash]{r} & \bullet \arrow{r}{a_1} & \bullet \arrow{r}{w_n} & \bullet \arrow[]{d}{q} \arrow[dash]{r}{w_{n-1}} & \bullet \arrow[dash, decorate = true]{r} & {} \\
P: & & & & & \bullet & &
\end{tikzcd}$$
Similarly as for $i = d$, we need to discuss about components of $g$ other than $q$. We know that $w_n = a_m w'$, and we need to have that $w_n q = 0$ for the map $g$ to be properly defined. The composition $f_1 g$ is therefore trivial.
The proof for two-sided infinite string complexes goes along the same lines. The arguments in the discussion above are applied for both infinite parts concurrently.
\end{proof}
The following theorem, which gives substantial restrictions on approximations of indecomposable projectives in semiorthogonal decompositions, is the main result of this section:
\begin{theorem}\label{TProjectivesDecomposition}
Let $\langle\mathcal{L}, \mathcal{R}\rangle$ be a semiorthogonal decomposition of $\Db(\Lambda\textrm{-}\mathrm{mod})$, and denote $P_1, \dots, P_n \in \Db(\Lambda\textrm{-}\mathrm{mod})$ the indecomposable direct summands of $\Lambda$. Moreover, let us have the following approximations:
$$\bigoplus_{t} R(i,t)^{m_R(i,t)} \overset{\lambda(i)}{\longrightarrow} P(i) \overset{\varrho(i)}{\longrightarrow} \bigoplus_{u} L(i,u)^{m_L(i,u)} \overset{\mu(i)}{\longrightarrow} \bigoplus_{t} R(i,t)^{m_R(i,t)}[1]$$
such that $R(i,t) \in \mathcal{R}$ for all $t$ and $L(i,u) \in \mathcal{L}$ for all $u$ and they are all indecomposable. We claim that:
\begin{enumerate}[(i)]
\item All $R(i,t)$ and $L(i',u')$ are string complexes.
\item All $\Hom_{\Db(\Lambda\textrm{-}\mathrm{mod})}(R(i,t)[j], R(i',t'))$, $\Hom_{\Db(\Lambda\textrm{-}\mathrm{mod})}(R(i,t)[j], L(i',u))$, and $\Hom_{\Db(\Lambda\textrm{-}\mathrm{mod})}(L(i,u)[j], L(i',u'))$ have a $k$-linear basis that comprises of boundary graph maps and singleton single maps of type $(i)$.
\end{enumerate}
\end{theorem}
\begin{proof}
At first, we show that no band or infinite string complexes may be direct summands of the approximations.
Suppose that $R(i,t)$ is a band complex for some $i$ and $t$. Denote $f\colon R(i,t)[-1] \to R(i,t)$ the quasi-graph map corresponding to the identity on $R(i,t)[-1]$ and $\iota\colon R(i,t) \to \bigoplus_{t} R(i,t)^{m_R(i,t)}$ one of the canonical injections. We have the following diagram:
$$\begin{tikzcd}
R(i,t)[-1] \arrow{d}{f \iota} \arrow{r} & 0 \arrow{d}{0} \arrow{r}{} & R(i,t) \arrow{r}{-\mathrm{id}_{R(i,t)[-1]}[1]} & R(i,t) \arrow{d}{(f\iota)[1]} \\
\bigoplus_{t} R(i,t)^{m_R(i,t)} \arrow{r}{\lambda(i)} & P_i \arrow{r}{\varrho(i)} & \bigoplus_{u} L(i,u)^{m_L(i,u)} \arrow{r}{\mu(i)} & \bigoplus_{t} R(i,t)^{m_R(i,t)}[1]
\end{tikzcd}$$
The square on the left commutes by Proposition \ref{PBandsProjectives}.
There exists a homotopically non-trivial map $g: R(i,t) \to \bigoplus_{u} L(i,u)^{m_L(i,u)}$ by the axioms of triangulated categories.
This a contradiction as $R(i,t) \in \mathcal{R}$, and there are no non-trivial morphisms from $\mathcal{R}$ to $\mathcal{L}$.
Provided that $L(i,u)$ is a band complex, we denote $f\colon L(i,u)[-1] \to L(i,u)$ the quasi-graph map corresponding to the identity on $L(i,u)[-1]$ and $\pi\colon L(i,u) \to \bigoplus_{u} L(i,u)^{m_L(i,u)}$ one of the canonical projections. We consider the following diagram:
$$\begin{tikzcd}
P_i \arrow{d}{0} \arrow{r}{\varrho(i)} & \bigoplus_{u} L(i,u)^{m_L(i,u)} \arrow{d}{\pi f[1]} \arrow{r}{\mu(i)} & \bigoplus_{t} R(i,t)^{m_R(i,t)}[1] \arrow{r}{\lambda(i)[1]} & P[1] \arrow{d}{0} \\
0 \arrow{r} & L(i,u)[1] \arrow{r}{\mathrm{id}_{L(i,u)[1]}} & L(i,u)[1] \arrow{r} & 0
\end{tikzcd}$$
Again, the square on the left commutes by Proposition \ref{PBandsProjectives}, and there exists a homotopically non-trivial map $g: \bigoplus_{t} R(i,t)^{m_R(i,t)} \to L(i,u)[1]$, which yields a contradiction.
In a similar manner, we discuss that neither any $R(i,t)$ nor any $L(i,u)$ may be an infinite string complex using \ref{PInfiniteStringsProjectives} instead of \ref{PBandsProjectives}. \\\\
Second, we prove that basis maps between indecomposable direct summands of the approximations may only be certain special graph maps and singleton single maps.
By Theorem \ref{TBasisMaps}, the vector space $\Hom_{\Db(\Lambda\textrm{-}\mathrm{mod})}(R(i',t')[j], R(i, t))$ over $k$ has a basis of graph maps, quasi-graph maps, and singleton single and double maps. Suppose there is a map $f\colon R(i',t')[j] \to R(i, t)$ in this basis that is not boundary graph map or a singleton single map of type $(i)$. Without loss of generality, $f$ is either a quasi-graph map or singleton single map of type $(iii)$ or $(iv)$. If $f$ is a graph map that is not boundary, there exists a quasi-graph map $f'\colon R(i, t) \to R(i',t')[j+1]$ by Lemma \ref{LBoundaryGrapMaps}, and we discuss maps from $R(i, t)$ to $R(i',t')[j+1]$ instead. By Lemma \ref{LDualMorphisms}, similar argument works for $f$ a singleton single map or singleton double map.
Denote $\iota$ a canonical inclusion of $R(i,t)$ into $\bigoplus_{t} R(i,t)^{m_R(i,t)}$, and consider the following diagram:
$$\begin{tikzcd}
R(i',t')[j] \arrow{d}{f \iota} \arrow{r} & 0 \arrow{d}{0} \arrow{r}{} & R(i',t')[j+1] \arrow{r}{-\mathrm{id}_{R(i',t')[j]}[1]} & R(i',t')[j+1] \arrow{d}{(f\iota)[1]} \\
\bigoplus_{t} R(i,t)^{m_R(i,t)} \arrow{r}{\lambda(i)} & P_i \arrow{r}{\varrho(i)} & \bigoplus_{u} L(i,u)^{m_L(i,u)} \arrow{r}{\mu(i)} & \bigoplus_{t} R(i,t)^{m_R(i,t)}[1]
\end{tikzcd}$$
The argument is similar as for the case bands above. The square on the left commutes by Proposition \ref{PBandsProjectives}. There exists a homotopically non-trivial map $g: R(i',t')[j+1] \to \bigoplus_{u} L(i,u)^{m_L(i,u)}$ yielding a contradiction.\\\\
For $\Hom_{\Db(\Lambda\textrm{-}\mathrm{mod})}(R(i',t)[j], L(i,u))$ and $\Hom_{\Db(\Lambda\textrm{-}\mathrm{mod})}(L(i',u')[j], L(i,u))$, the proof unfolds in a similar fashion as above.
\end{proof}
\begin{remark}
Equivalently, the theorem above can be proved using the fact that inclusions of $\mathcal{L}$ and $\mathcal{R}$ have left and right adjoints, respectively (cf. Proposition \ref{PSemiorthogonalDecomposition}). Suppose that we have an approximation $R \to P \to L \to R[1]$ with $R \in \mathcal{R}$ and $L \in \mathcal{L}$ for $P \in \Db\Lambda \textrm{-}\mathrm{mod})$. Suppose there is $0\ne f\colon R' \to R$ such that $R' \in \mathcal{R}$ and $f$ composes to zero with the approximation morphism $R \to P$. This is a contradiction as $\Hom_{\Db(\Lambda \textrm{-}\mathrm{mod})}(R',P) \cong \Hom_\mathcal{R}(R', R)$ via the approximation morphism $R \to P$. This argument can be made dually for $L$.
\end{remark}
\section{Semiorthogonal decompositions in the geometric model}
In this section, we give a one-to-one correspondence of between two-term semiorthogonal decompositions of $\Db(\Lambda \textrm{-}\mathrm{mod})$ and certain ways of cutting the underlying marked surface $(S, M, P)$ of the geometric model of $\Db(\Lambda \textrm{-}\mathrm{mod})$. We then extend this result to semiorthogonal decompositions with more than two terms.
Unless otherwise indicated, we only consider two-term semiorthogonal decompositions in this section.
\subsection{Bipartite admissible dissections}
We begin by translating the necessary conditions posed on a semiorthogonal decomposition $\langle \mathcal{L}, \mathcal{R} \rangle$ of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ by Theorem \ref{TProjectivesDecomposition} to the geometric model of $\Db(\Lambda \textrm{-}\mathrm{mod})$.
Taking models of indecomposable direct summands of approximations of indecomposable projectives in $\langle \mathcal{L}, \mathcal{R} \rangle$, we exhibit a system of $\color{dark-green}{\circ}-$arcs that may meet only at endpoints and that an a $\color{dark-green}{\circ}-$arcs from $\mathcal{R}$ need to always $\color{dark-green}{\circ}-$arcs in the counter-clockwise order around a common endpoint.
Furthermore, we prove a statement that can be viewed as a converse to Theorem \ref{TProjectivesDecomposition} by showing that any bipartite admissible dissection gives a semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$.
\begin{definition}[Bipartite dissection; after Definition 1.9 in \cite{amiot2019complete}]\label{DBipartiteAdmissibleDissection}
An admissible collection of $\color{dark-green}{\circ}$-arcs $\Gamma = \{\gamma_1, \dots, \gamma_r\}$ on the marked surface $(S, M, P)$ (see Definition \ref{DAdmissibleDissection} above) together with a function $p: \Gamma \to \{1, 2\}$ is called bipartite if $p(\gamma') \geq p(\gamma)$ for every two arcs $\gamma, \gamma' \in \Gamma$ such that $\gamma$ and $\gamma'$ have the same endpoint $e \in M_{\color{dark-green}{\circ}} \cup P_{\color{dark-green}{\circ}}$ and $\gamma'$ follows $\gamma$ in the counter-clockwise order around $e$. A maximal bipartite admissible collection is referred to as bipartite dissection.
\end{definition}
\begin{proposition}\label{PBipartiteDissectionCompletion}
Let $(\Gamma, p)$ be a bipartite admissible collection of $\color{dark-green}{\circ}$-arcs on the surface $(S, M, P)$ with $P_{\color{dark-green}{\circ}}$ empty, then there exists a bipartite admissible dissection $(\Gamma', p')$ such that $\Gamma \subseteq \Gamma'$ and $p'$ extends $p$.
\end{proposition}
\begin{proof}
Similarly as in Proposition 1.12 in \cite{amiot2019complete}, consider a connected component $\mathcal{P}$ of the complement of the $\color{dark-green}{\circ}$-arcs in $\Gamma$ in $S \, \backslash \, P$. Suppose that there is a $\color{red} \bullet$-puncture in $\mathcal{P}$ as well some some other $\color{red} \bullet$-vertices, $\color{red} \bullet$-punctures or $\color{red} \bullet$-marked points. In this case it is possible to add a $\color{dark-green}{\circ}$-arc $\gamma'$ whose two endpoints coincide and which the $\color{red} \bullet$-puncture and no other red vertices. If no arc $\gamma \in \Gamma$ with $p(\gamma) = 2$ follows $\gamma'$ in the counter-clock wise order around the sole endpoint of $\gamma'$, $\gamma$ is assigned $1$; otherwise, it is assigned $2$.
Now, suppose that there are two or more $\color{red} \bullet$-marked points in $\mathcal{P}$, that is no $\color{red} \bullet$-punctures. In this case, the boundary of $\mathcal{P}$ is made up of segments of $\partial S$ with a $\color{red} \bullet$-marked point each and of segments of successive $\color{dark-green}{\circ}$-arcs in , which alternate with one another. Because there are at least two $\color{red} \bullet$-marked points in $\mathcal{P}$, there at least two different $\color{dark-green}{\circ}$-arc segments $\Sigma_1$ and $\Sigma_2$ of the boundary of $\mathcal{P}$ separated by a single $\partial S$ segment $T$ with one $\color{red} \bullet$-marked point $t$.
Assume that $\Sigma_2$ follows $\Sigma_1$ in the counter-clockwise order around $t$. Specifically, $t$ divides $T$ into two parts $T_i$ whose boundary consists of $t$ and a $\color{dark-green}{\circ}$ marked point $s_i$ that lies on the boundary of $\Sigma_i$, for $i = 1, 2$, and we assume that $T_2$ follows $T_1$ in the counter-clockwise order around $t$. Moreover, we assume that $\Sigma_i$ are made up of successive $\color{dark-green}{\circ}$-arcs $\sigma^{(i)}_1, \dots, \sigma^{(i)}_{n_i}$, for $i = 1, 2$, such that one of the endpoints of $\sigma^{(1)}_{n_1}$ is $s_1$ and one of the endpoints of $\sigma^{(2)}_1$ is $s_2$. Note that one of the endpoints of $\sigma^{(i)}_1$ or $\sigma^{(i)}_{n_i}$ lies on a $\partial S$ segment of the boundary of $\mathcal{P}$; therefore, no $\color{dark-green}{\circ}$-arc $\gamma \in \Gamma$ may lie to the left of $\sigma^{(i)}_1$ or to the right of $\sigma^{(i)}_{n_i}$ in the counter-clockwise order around the endpoint lying on the $\partial S$ segment.
Denote $s'_1$ the $\color{dark-green}{\circ}$-marked point lying on the $\partial S$ segment of the boundary $\mathcal{P}$. Then, it is possible to add a $\color{dark-green}{\circ}$-arc $\gamma'$ from $s'_1$ to $s_2$ that lies inside $\mathcal{P}$ and is assigned $1$. All other $\color{dark-green}{\circ}$ arcs in $\Gamma$ from $s'_1$ or $s_2$ follow $\gamma'$ in the counter-clockwise orientation, and $\gamma'$ cuts off exactly one $\color{red} \bullet$-marked point from $\mathcal{P}$; hence we may extend $(\Gamma, p)$ by $\gamma'$ to a larger bipartite admissible collection.
By the inductive procedure described above, it is possible to extend $(\Gamma, p)$ to a bipartite admissible dissection $(\Gamma', p')$.
\end{proof}
\begin{proposition}\label{PBipartiteDissectionDecomposition}
Let $\Lambda$ be a gentle algebra, and let $(S, M, P)$ be its associated marked surface. A bipartite admissible dissection $(\Gamma, p)$ yields a semiorthogonal decomposition $\langle \mathcal{L}, \mathcal{R} \rangle$ of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ where $\mathcal{L}$ is generated by $\gamma \in \Gamma$ such that $p(\gamma) = 1$ and $\mathcal{R}$ is generated by $\gamma \in \Gamma$ such that $p(\gamma) = 2$.
\end{proposition}
\begin{proof}
To prove that $\langle \mathcal{L}, \mathcal{R} \rangle$ is semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$, it suffices to show that that there are no non-zero morphisms from $\mathcal{R}$ to $\mathcal{L}$ and that $\mathcal{L}$ and $\mathcal{R}$ generate the entire triangulated category $\Kb(\Lambda \textrm{-}\mathrm{proj})$ by Proposition \ref{PSemiorthogonalDecomposition}.
The morphisms between from $P_{(\gamma_1, f_1)}$ to $P_{(\gamma_2, f_2)}$ are up to shift given by oriented graded intersections $\gamma_1$ from $\gamma_2$; as $\color{dark-green}{\circ}$-arcs in $\Gamma$ may intersect only at endpoints, there are non-trivial morphisms from $P_{(\gamma_1, f_1)}$ to $P_{(\gamma_2, f_2)}$ only if $\gamma_2$ follows $\gamma_1$ in the counter-clockwise order around a common endpoint. In such case, it is required that $p(\gamma_2) \geq p(\gamma_1)$. So there are no non-trivial maps from $P_{(\gamma_1, f_1)}$ to $P_{(\gamma_2, f_2)}$ should $p(\gamma_1)$ be greater than $p(\gamma_2)$.
Since there are no non-trivial morphisms from the generators of $\mathcal{R}$ to the shifts generators of $\mathcal{L}$, a simple induction shows that there no non-trivial morphisms from $\mathcal{R}$ to $\mathcal{L}$.
The triangulated category $\Kb(\Lambda \textrm{-}\mathrm{proj})$ is generated by all string complexes, so it is enough to prove that they are generated by $\mathcal{L}$ and $\mathcal{R}$. Consider a $\color{dark-green}{\circ}$-arc $\sigma$ given that any string complex is of form $P_{(\sigma, h)}$ for some $\color{dark-green}{\circ}$-arc $\sigma$, and suppose that the arcs in $\Gamma$ and $\sigma$ are in a minimal position.
By Proposition 1.12 in \cite{amiot2019complete} the connected components of the complement of $\Gamma$ in $S \, \backslash \, P$ are homeomorphic either to an open disk with a single $\color{red} \bullet$-marked point on the boundary or to an open punctured disk with a $\color{red} \bullet$-puncture and no $\color{red} \bullet$-marked points on the boundary. The $\color{dark-green}{\circ}$-arc $\sigma$ crosses components $\mathcal{P}_1, \dots, \mathcal{P}_n$ of the complement of $\Gamma$ in $S \, \backslash \, P$ in this order. Inductively, we will construct a sequence: $$\gamma_1^{(1)}, \dots, \gamma_{m_1}^{(1)}, \gamma_1^{(2)}, \dots, \gamma_{m_2}^{(2)}, \dots, \gamma_1^{(n)}, \dots, \gamma_{m_n}^{(n)}$$ such that:
\begin{enumerate}[(i)]
\item one of the endpoints of $\sigma$ is an endpoint of $\gamma_1^{(1)}$,
\item the other endpoint of $\sigma$ is an endpoint of $\gamma_{m_n}^{(n)}$, and
\item successive curves in the sequence share an endpoint, and
\item the iterated concatenation of the arcs in the sequence is homotopic to $\sigma$.
\end{enumerate}
The subsequence $\gamma_1^{(i)}, \dots, \gamma_{m_i}^{(i)}$ is constructed in a straight-forward way. If the $\color{red} \bullet$-point pertaining to the component $\mathcal{P}_i$ lies on the boundary, the segment of $\sigma$ lying in $\mathcal{P}_i$ cuts in into two components: one containing the $\color{red} \bullet$-point pertaining to $\mathcal{P}_i$ and the other without a marked point. Denote $\delta_1^{(i)}, \dots, \delta_{m'_i}^{(i)}$ the $\color{dark-green}{\circ}$-arcs in $\Gamma$ that lie or partially lie on the boundary of the component of $\mathcal{P}_i$ cut by $\sigma$ without the $\color{red} \bullet$-point pertaining to $\mathcal{P}_i$ such that $\delta_j^{(i)}$ and $\delta_{j+1}^{(i)}$ share a common endpoint for $1 \leq j \leq m'-1$ and $\sigma$ meets $\delta_1^{(i)}$ before $\delta_{m'_i}^{(i)}$. Clearly, the segment of $\sigma$ in $\mathcal{P}_i$ is homotopic to the iterated concatenation of $\delta_1^{(i)}, \dots, \delta_{m'_i}^{(i)}$.
If the $\color{red} \bullet$-point pertaining to the component $\mathcal{P}_i$ is a puncture, $\sigma$ may wind around it and intersect itself. $\delta_1^{(i)}, \dots, \delta_{m'_i}^{(i)}$ the $\color{dark-green}{\circ}$-arcs in $\Gamma$ on the boundary of $\mathcal{P}_i$ such that $\sigma$ crosses their dual $\color{red} \bullet$-arcs in $\Gamma^*$ within $\mathcal{P}_i$ in that order together with the $\color{dark-green}{\circ}$-arcs in $\Gamma$ that $\sigma$ crosses as it goes through $\mathcal{P}_i$, which would form the first and the last element of the sequence, if applicable and not included already.
The sequence is defined so that the segment of $\sigma$ that lies in $\mathcal{P}_i$ is homotopic iterated concatenation of $\delta_1^{(i)}, \dots, \delta_{m'_i}^{(i)}$. We use the fact that the segment of $\sigma$ is homotopic to the iterated concatenation of its sub-segments that lie in connected components of the complement of $\Gamma*$ in $S \backslash P$. If the segment $\sigma$ in $\mathcal{P}_i$ crosses in the connected component $\mathcal{Q}$ of the complement of $\Gamma*$ in $S \backslash P$, it has to cross two adjacent $\color{red} \bullet$-arcs on its boundary; otherwise, it would cross a $\color{dark-green}{\circ}$-arc, which it does not. Therefore, the corresponding $\color{dark-green}{\circ}$-arcs in $\delta_j^{(i)}, \delta_{j+1}^{(i)} \in \Gamma$ have a common endpoint, and their concatenation is homotopic to the respective sub-segment of $\sigma$ in $\mathcal{Q}$. This can be schematically illustrated by the following figure:
\begin{figure}[H]
\centering
\begin{tikzpicture}[x = 1cm, y = 1cm]
\pgfmathsetmacro{\length}{15/sqrt(50+10*sqrt(5))}
\node[coordinate] (Red) at (0,-0.5*\length) {};
\node[coordinate] (Green) at (0,0.5*\length) {};
\draw[dark-green] (Green) -- ++(216:\length) node[coordinate, at end] (GreenLeft) {} node[near start, below, black] {$\delta^{(i)}_j$};
\draw[dark-green] (Green) -- ++(-36:\length) node[coordinate, at end] (GreenRight) {} node[near start, below, black] {$\delta^{(i)}_{j+1}$};
\draw[red] (Red) -- ++(54:\length) node[coordinate, at end] (RedAuxRight) {};
\draw[red] (Red) -- ++(126:\length) node[coordinate, at end] (RedAuxLeft) {};
\draw[red, dotted] (RedAuxRight) -- ++(54:0.5*\length) node[near start, below right, black] {$\delta^{(i)*}_{j+1}$};
\draw[red, dotted] (RedAuxLeft) -- ++(126:0.5*\length) node[near start, below left, black] {$\delta^{(i)*}_j$};
\path (GreenLeft) -- (Red) node[coordinate, midway] (SigmaLeft) {};
\path (GreenRight) -- (Red) node[coordinate, midway] (SigmaRight) {};
\draw[blue] (SigmaLeft) to[out = 45, in = 135] (SigmaRight);
\node[blue, below = 0*\length of SigmaLeft] {$\sigma$};
\draw[dark-green, fill = white] (Green) circle (0.1);
\draw[dark-green, fill = white] (GreenLeft) circle (0.1);
\draw[dark-green, fill = white] (GreenRight) circle (0.1);
\draw[red, fill = red] (Red) circle (0.1);
\end{tikzpicture}
\caption{The sub-segment of $\sigma$ in $Q$ and its representation by $\color{dark-green}{\circ}$-arcs in $\Gamma$}
\label{FBipartiteDissectionDecomposition1}
\end{figure}
We note that the same principle has been tacitly applied in the case above, when $\mathcal{P}_i$ had a $\color{red} \bullet$-marked point on the boundary. This principle also underpins the proof of how the indecomposable objects in $\Db(\Lambda \textrm{-}\mathrm{mod})$ are reflected in its geometric model in Theorem 2.12 in \cite{opper2018geometric}.
If $i = 1$, $\delta_1^{(1)}$ shares an endpoint with $\sigma$, we define $\gamma_i^{(1)} = \delta_i^{(1)}$ for $1 \leq i \leq m$ and we set $m = m'-1$ if the $\color{red} \bullet$-marked point lies to same side of $\sigma$ in $\mathcal{P}_2$ as in $\mathcal{P}_2$ and $m = m'$ otherwise, which includes the situation when $i = n$ and $\sigma$ shares an endpoint with $\delta_{m'_1}^{(1)}$. If $i > 1$, $\sigma$ crosses $\delta_1^{(i)}$ as it enters $\mathcal{P}_i$, we define $\gamma_j^{(i)} = \delta_{j+1}^{(i)}$ for $1 \leq j \leq m$ and we set $m = m'-1$ if the $\color{red} \bullet$-marked point lies to same side of $\sigma$ in $\mathcal{P}_{i+1}$ as in $\mathcal{P}_i$ and $m = m'$ otherwise, which includes the situation when $i = n > 1$ and $\sigma$ shares an endpoint with $\delta_{m'_i}^{(i)}$.
As above, we note that the last $\color{dark-green}{\circ}$-arc $\delta^{(i)}_{m'_i}$ is included in the final sequence precisely when $\sigma$ crosses the dual $\color{red} \bullet$-arc in $\Gamma^*$ in addition to crossing the $\color{dark-green}{\circ}$-arc separating $\mathcal{P}_{i}$ and $\mathcal{P}_{i+1}$.
By definition, our sequence satisfies (i) and (ii). As for (iii), it suffices to check it only for $\gamma_{m_i}^{(i)}$ and $\gamma_{1}^{(i+1)}$. Suppose that $\sigma$ lies to the same side of the $\color{red} \bullet$-marked point in both $\mathcal{P}_i$ and $\mathcal{P}_{i+1}$; schematically, we are in the following situation up to change in orientation of $\sigma$, which is assumed to go from left to right:
\begin{figure}[H]
\centering
\begin{tikzpicture}[x = 1cm, y = 1cm]
\def5{5}
\def3{3}
\pgfmathsetmacro{\penfactor}{0.5*cos(180/5)}
\pgfmathsetmacro{\penonecenter}{-3*\penfactor}
\pgfmathsetmacro{\pentwocenter}{3*\penfactor}
\node[regular polygon, regular polygon sides = 5, minimum size = 3 cm, shape border rotate = 18] (Pen1) at (\penonecenter,0) {};
\foreach \x in {1,...,5}
{\node[coordinate, name = {Pen1Node\x}] at (Pen1.corner \x) {};}
\draw[dark-green, dotted] (Pen1Node1) -- (Pen1Node2) node[coordinate, midway, name = {Sigma1}] {};
\draw (Pen1Node2) to[out = -9, in = 71] node[coordinate, midway, name = {Pol1Red}] {} (Pen1Node3);
\fill[pattern = north east lines] (Pen1Node2) to[out = -9, in = 71] (Pen1Node3) -- cycle; \draw[red, fill = red] (Pol1Red) circle (0.1);
\draw[dark-green, dotted] (Pen1Node3) -- (Pen1Node4);
\draw[dark-green] (Pen1Node4) -- (Pen1Node5) node[coordinate, midway, name = {Sigma2}] {};
\draw[dark-green] (Pen1Node5) -- (Pen1Node1) node[midway, below left, black] {$\gamma_{m_i}^{(i)}$};
\node[regular polygon, regular polygon sides = 5, minimum size = 3 cm, shape border rotate = -18] (Pen2) at (\pentwocenter,0) {};
\foreach \x in {1,...,5}
{\node[coordinate, name = {Pen2Node\x}] at (Pen2.corner \x) {};}
\draw[dark-green] (Pen2Node1) -- (Pen2Node2) node[midway, below right, black] {$\gamma_1^{(i+1)}$};
\draw[dark-green, dotted] (Pen2Node3) -- (Pen2Node4);
\draw (Pen2Node4) to[out = 99, in = 189] node[coordinate, midway, name = {Pol2Red}] {} (Pen2Node5);
\fill[pattern = north west lines] (Pen2Node4) to[out = 99, in = 189] (Pen2Node5) -- cycle; \draw[red, fill = red] (Pol2Red) circle (0.1);
\draw[dark-green, dotted] (Pen2Node5) -- (Pen2Node1) node[coordinate, midway, name = {Sigma3}] {};
\foreach \x in {1,...,5}
{\draw[dark-green, fill = white] ({Pen1Node\x}) circle (0.1);}
\foreach \x in {1,4,5}
{\draw[dark-green, fill = white] ({Pen2Node\x}) circle (0.1);}
\draw[blue] ($1.5*(Sigma1)$) to[out = 315, in = 180] (Sigma2) to[out = 0, in = 225] ($1.5*(Sigma3)$);
\node[blue, below left] (Sigma2) {$\sigma$};
\end{tikzpicture}
\caption{Constructing $\Gamma$-sequence for $\sigma$ going from $\mathcal{P}_i$ to $\mathcal{P}_{i+1}$ with respective $\color{red} \bullet$-points being to different sides of $\sigma$}
\label{FBipartiteDissectionDecomposition2}
\end{figure}
The $\color{dark-green}{\circ}$-arcs $\gamma_{m_i}^{(i)}$ and $\gamma_1^{(i+1)}$ have a shared endpoint. We know that $\delta_{m'_i}^{(i)} = \delta_1^{(i+1)}$ and that $\delta_{m'_i - 1}^{(i)} = \gamma_{m_i}^{(i)}$ and $\delta_{m'_i}^{(i)}$, and $\delta_1^{(i+1)}$ and $\delta_2^{(i+1)} = \gamma_1^{(i+1)}$ share an endpoint, respectively. However, the shared endpoint has to lie on the opposite side of the intersection of $\delta_{m'_i}^{(i)} = \delta_1^{(i+1)}$ and $\sigma$ as the $\color{red} \bullet$-point correspond to the respective component in both cases, so it has be the same endpoint.
Suppose that $\sigma$ lies to the same side of the $\color{red} \bullet$-marked point in both $\mathcal{P}_i$ and $\mathcal{P}_{i+1}$; schematically, we are in the following situation up to change in orientation of $\sigma$, which is assumed to go from left to right:
\begin{figure}[H]
\centering
\begin{tikzpicture}[x = 1cm,y = 1cm]
\def5{5}
\def3{3}
\pgfmathsetmacro{\penfactor}{0.5*cos(180/5)}
\pgfmathsetmacro{\penonecenter}{-3*\penfactor}
\pgfmathsetmacro{\pentwocenter}{3*\penfactor}
\node[regular polygon, regular polygon sides = 5, minimum size = 3 cm, shape border rotate = 18] (Pen1) at (\penonecenter,0) {};
\foreach \x in {1,...,5}
{\node[coordinate, name = {Pen1Node\x}] at (Pen1.corner \x) {};}
\draw[dark-green, dotted] (Pen1Node1) -- (Pen1Node2) node[coordinate, midway, name = {Sigma1}] {};
\draw (Pen1Node2) to[out = -9, in = 71] node[coordinate, midway, name = {Pol1Red}] {} (Pen1Node3);
\fill[pattern = north east lines] (Pen1Node2) to[out = -9, in = 71] (Pen1Node3) -- cycle;
\draw[red, fill = red] (Pol1Red) circle (0.1);
\draw[dark-green, dotted] (Pen1Node3) -- (Pen1Node4);
\draw[dark-green] (Pen1Node4) -- (Pen1Node5) node[coordinate, midway, name = {Sigma2}] {} node[midway, above right, black] {$\gamma_1^{(i+1)}$};
\draw[dark-green] (Pen1Node5) -- (Pen1Node1) node[midway, below left, black] {$\gamma_{m_i}^{(i)}$};
\node[regular polygon, regular polygon sides = 5, minimum size = 3 cm, shape border rotate = -18] (Pen2) at (\pentwocenter,0) {};
\foreach \x in {1,...,5}
{\node[coordinate, name = {Pen2Node\x}] at (Pen2.corner \x) {};}
\draw[dark-green, dotted] (Pen2Node3) -- (Pen2Node4) node[coordinate, midway, name = {Sigma3}] {};
\draw[dark-green, dotted] (Pen2Node4) -- (Pen2Node5);
\draw (Pen2Node5) to[out = 171, in = 261] node[coordinate, midway, name = {Pol2Red}] {} (Pen2Node1);
\fill[pattern = north east lines] (Pen2Node5) to[out = 171, in = 261] (Pen2Node1) -- cycle;
\draw[red, fill = red] (Pol2Red) circle (0.1);
\draw[dark-green, dotted] (Pen2Node1) -- (Pen2Node2);
\foreach \x in {1,...,5}
{\draw[dark-green, fill = white] ({Pen1Node\x}) circle (0.1);}
\foreach \x in {1,4,5}
{\draw[dark-green, fill = white] ({Pen2Node\x}) circle (0.1);}
\draw[blue] ($1.25*(Sigma1)$) -- (Sigma2) to[out = -18, in = 108] ($1.5*(Sigma3)$);
\node[blue, below left] (Sigma2) {$\sigma$};
\end{tikzpicture}
\caption{Constructing $\Gamma$-sequence for $\sigma$ going from $\mathcal{P}_i$ to $\mathcal{P}_{i+1}$ with respective $\color{red} \bullet$-points being to different sides of $\sigma$}
\label{FBipartiteDissectionDecomposition3}
\end{figure}
We can employ a similar argument as above to reason that the $\color{dark-green}{\circ}$-arcs $\gamma_{m_i}^{(i)} = \delta_{m'_i}^{(i)}$ and $\gamma_1^{(i+1)} = \delta_2^{(i+1)}$ have the same endpoint.
Now, we observe that $\sigma$ is homotopic to the iterated concatenation of $\gamma_1^{(1)}, \dots, \gamma_{m_1}^{(1)},$ $\gamma_1^{(2)}, \dots, \gamma_{m_2}^{(2)}, \dots,$ $\gamma_1^{(n)}, \dots, \gamma_{m_n}^{(n)}$. If we concatenate $\gamma_{m_i}^{(i)}$ and $\gamma_{1}^{(i+1)}$ and denote the resulting arc $\gamma_{i,i+1}$, then the iterated concatenation of $\gamma_{i-1,i},$ $\gamma_2^{(i)}, \dots, \gamma_{m_i - 1}^{(i)},$ and $\gamma_{i,i+1}$ is homotopic to the segment of $\sigma$ in $\mathcal{P}_i$. The fact $\sigma$ is homotopic to the iterated concatenation of its segments in $\mathcal{P}_1, \dots, \mathcal{P}_n$ completes the argument.
By Theorem 4.1 in \cite{opper2018geometric} as recalled above, concatenation of $\color{dark-green}{\circ}$-arcs $\varrho_1$ and $\varrho_2$ at their common endpoint around which $\varrho_2$ follows $\varrho_1$ in the counter-clockwise order corresponds to forming the mapping cone of the arising morphisms from $P_{(\varrho_1, g_1)}$ to $P_{(\varrho_2, g_2)}$ up to shift (cf. the proof of Proposition 3.7 in \cite{amiot2019complete}). This observation can be used to deduce that $P_{(\sigma, h)}$ lies in the triangulated subcategory of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ generated by $P_{\gamma_j^{(i)}}$.
Let $(\Gamma, p)$ be a bipartite admissible dissection of $(S, M, P)$. We recalled in the proof above that, by Proposition 1.12 in \cite{amiot2019complete}, the connected components of the complement of $\Gamma$ in $S \, \backslash \, P$ are homeomorphic either to an open disk with a single $\color{red} \bullet$-marked point on the boundary or to an open punctured disk with a $\color{red} \bullet$-puncture and no $\color{red} \bullet$-marked points on the boundary.
We consider a component $\mathcal{P}$ whose boundary contains arcs $\gamma_1, \gamma_2 \in \Gamma$ such that $p(\gamma_1) \neq p(\gamma_2)$. The component $\mathcal{P}$ needs to be of the first type. If its boundary did not contain a segment of $\partial S$, it would consist entirely of arcs in $\Gamma$, and there would invariably be curves $\gamma_3, \gamma_4 \in \Gamma$ in its boundary such that they share an endpoint, $\gamma_4$ follows $\gamma_3$ in the counterclockwise order around the endpoint, but $p(\gamma_4) < p(\gamma_3)$. In the counter-clockwise order, the boundary of $\mathcal{P}$ looks as follows: a segment of $\partial S$ adjacent, a sequence of arcs in $\Gamma$ assigned $1$ by $p$, and a sequence of arcs in $\Gamma$ assigned $2$ by $p$.
\end{proof}
It turns out that the procedure of representing a $\color{dark-green}{\circ}$-arc as an iterated concatenation of $\color{dark-green}{\circ}$-arcs in $(\Gamma, p)$ employed in the proof of the proposition above can be used to get approximation triangles for string objects in the induced semiorthogonal decomposition $\langle \mathcal{L}, \mathcal{R} \rangle$ of $\Kb(\Lambda \textrm{-}\mathrm{proj})$. Before that, however, we need to deal with some technicalities.
\begin{lemma}\label{LMappingConeReduction}
Let $A \nameto{f_i} B_i$, $i = 1, 2$, be morphisms in a triangulated category $\mathcal{T}$, then $B_1 \oplus \mathrm{cone}((f_1, f_2)^T)$ is isomorphic to $B_1 \oplus \mathrm{cone}(-f_2 \circ a_1[-1])$ such that $a_1$ comes from the distinguished triangle: $$A \nameto{f_1} B_1 \nameto{b_1} \mathrm{cone}(f_1) \nameto{a_1} A[1].$$ Furthermore, $\mathrm{cone}((f_1, f_2)^T)$ is isomorphic to $\mathrm{cone}(-f_2 \circ a_1[-1])$ provided that $\mathcal{T}$ is Krull-Schmidt.
Dually, let $A_i \to B$, $i = 1, 2$ be morphisms in a triangulated category $\mathcal{T}$, then $A_1[1] \oplus \mathrm{cone}(f_1, f_2))$ is isomorphic to $A_1[1] \oplus \mathrm{cone}(b_1 f_2)$ such that $b_1$ comes from the distinguished triangle: $$A_1 \nameto{f_1} B \nameto{b_1} \mathrm{cone}(f_1) \nameto{a_1} A_1[1].$$ Also, $\mathrm{cone}((f_1, f_2))$ is isomorphic to $\mathrm{cone}(b_1 f_2)$ provided that $\mathcal{T}$ is Krull-Schmidt.
\end{lemma}
\begin{proof}
We apply the octahedral axiom to the composition of morphisms: $$\mathrm{cone}(f_1)[-1] \nameto{-a_1[-1]} A \nameto{(f_1, f_2)^T} B_1 \oplus B_2$$
We note that $-a_1[-1]$ composes to zero with $f_1$.
We have the following distinguished triangles:
$$\mathrm{cone}(f_1)[-1] \nameto{-a_1[-1]} A \nameto{f_1} B_1 \nameto{b_1} \mathrm{cone}(f_1),$$
$$A \nameto{(f_1, f_2)^T} B_1 \oplus B_2 \nameto{b} \mathrm{cone}((f_1, f_2)^T) \nameto{a} A[1],$$
and two direct sum of two distinguished triangles:
$$\mathrm{cone}(f_1)[-1] \nameto{-f_2 \circ a_1[-1]} B_2 \nameto{g} \mathrm{cone}(-f_2 \circ a_1[-1]) \nameto{h} \mathrm{cone}(f_1)$$
and
$$0 \nameto{} B_1 \nameto{\mathrm{id}_{B_1}} B_1 \nameto{} 0,$$
which corresponds to the distinguished triangle for $-(f_1, f_2)^T \circ a_1[-1]$ because the first coordinate is zero.
Using the octahedral axiom, it is possible to construct:
$$B_1 \nameto{} B_1 \oplus \mathrm{cone}(-f_2 \circ a_1[-1]) \nameto{} \mathrm{cone}((f_1, f_2)^T) \nameto{} B_1[1]$$
The rightmost map in the distinguished triangle equals $f_1[-1] \circ a$, which is zero because $(f_1[-1], f_2[-1])^T \circ a$ is zero. Therefore, the distinguished triangle splits, and $B_1 \oplus \mathrm{cone}((f_1, f_2)^T) \cong B_1 \oplus \mathrm{cone}(-f_2 \circ a_1[-1])$.
Dually, we apply the octahedral axiom to the following composition:
$$A_1 \oplus A_2 \nameto{(f_1, f_2)} B \nameto{b_1} \mathrm{cone}(f_1)$$.
Combining distinguished triangles:
$$A_1 \oplus A_2 \nameto{(f_1, f_2)} B \nameto{b} \mathrm{cone}((f_1, f_2)) \nameto{a} (A_1 \oplus A_2)[1],$$
$$B \nameto{b_1} \mathrm{cone}(f_1) \nameto{a_1} A_1[1] \nameto{-f_1[1]} B[1],$$
and the direct sum of:
$$A_2 \nameto{b_1 f_2} \mathrm{cone}(f_1) \nameto{g} \mathrm{cone}(b_1 f_2) \nameto{h} A_2[1]$$
and
$$A_1 \nameto{\mathrm{id}_{A_1}} A_1 \nameto{} 0 \nameto{} A_1[1],$$
we obtain the following:
$$\mathrm{cone}((f_1, f_2)) \nameto{} A_1 \oplus \mathrm{cone}(b_1 f_2) \nameto{} A_1[1] \nameto{} \mathrm{cone}((f_1, f_2))[1]$$
This triangle splits, however, because the rightmost morphism is zero as the first coordinate of the composition $b \circ (f_1, f_2)$, which equals zero. We thus obtain that $A_1[1] \oplus \mathrm{cone}((f_1, f_2)) \cong A_1[1] \oplus \mathrm{cone}(b_1 f_2)$.
If $\mathcal{T}$ is Krull-Schmidt and $C \oplus D \cong C \oplus D'$, it is easy to see that the uniqueness of the direct decomposition into indecomposables yields that $D \cong D'$. The remaining claims of this lemma follow from this observation.
\end{proof}
\begin{proposition}\label{PBipartiteDissectionTriangles}
Let $(\Gamma, p)$ be a bipartite admissible dissection of the marked surface $(S, M, P)$, and let $\langle \mathcal{L}, \mathcal{R} \rangle$ be the associated semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$. For every string object $P_{(\sigma, h)}$, there exist string objects $P_{(\delta_1, g_1)}, \dots, P_{(\delta_n, g_n)}$ such that:
\begin{enumerate}[(i)]
\item $\sigma$ is homotopic to an iterated concatenation of $\delta_1, \dots, \delta_n$;
\item either $P_{(\delta_i, g_i)} \in \mathcal{L}$ for all odd $i$ and $P_{(\delta_j, g_j)} \in \mathcal{R}$ for all even $j$ or $P_{(\delta_i, g_i)} \in \mathcal{R}$ for all odd $i$ and $P_{(\delta_j, g_j)} \in \mathcal{L}$ for all even $j$;
\item $\delta_i$ and $\delta_{i+1}$ for $1 \leq i \leq n-1$ share an endpoint, and every shared endpoint gives rise to a map $f_{i, i+1}: P_{(\delta_i, g_i)} \to P_{(\delta_{i+1}, g_{i+1})}$ should $P_{(\delta_i, g_i)} \in \mathcal{L}$ or vice versa;
\item the mapping cone $\bigoplus_{i \, \mathrm{even}} P_{(\delta_i, g_i)} \nameto{f} \bigoplus_{i \, \mathrm{odd}} P_{(\delta_j, g_j)}$ where $f$ consists of maps $f_{i, j}$ arising from shared endpoints of adjacent $\delta_i$ and $\delta_j$ or vice versa is isomorphic to $P_{(\sigma, h)}$.
\end{enumerate}
\end{proposition}
\begin{proof}
Assume that $P_{\sigma, h}$ is a string object reprsented by a $\color{dark-green}{\circ}$-arc $\sigma$ with grading $h$. As in the proof of Proposition \ref{PBipartiteDissectionDecomposition}, we can construct a sequence: $$\gamma_1^{(1)}, \dots, \gamma_{m_1}^{(1)}, \gamma_1^{(2)}, \dots, \gamma_{m_2}^{(2)}, \dots, \gamma_1^{(n)}, \dots, \gamma_{m_n}^{(n)}$$ of $\color{dark-green}{\circ}$-arcs in $\Gamma$ such that $\sigma$ is homotopic to the iterated concatenation thereof.
From this sequence, we concatenate its maximal segments such that all $\color{dark-green}{\circ}$-arcs in the segment belong to the same part of $(\Gamma, p)$. We obtain a sequence of $\delta_1, \dots, \delta_n$, whose iterated concatenation remains homotopic to $\sigma$. Without loss of generality, we may assume that $\delta_i$ for $i$ odd are concatenations of $\gamma \in \Gamma$ such that $p(\gamma) = 1$ and conversely for $\delta_j$ with $j$ even. Consequently, for any choice of gradings, string objects associated to $\delta_i$, $i$ odd, belong $\mathcal{L}$; whereas, string objects associated to $\delta_j$, $j$ even, belong $\mathcal{R}$.
By construction, $\delta_i$ and $\delta_{i+1}$ share an endpoint for all $1 \leq i \leq n-1$. We may choose gradings $g_1, \dots, g_n$ such that the common endpoints induce maps from $f_{i, i+1}: P_{(\delta_i, g_i)} \to P_{(\delta_{i+1}, g_{i+1})}$, for odd $1 \leq i \leq n-1$, and $f_{i, i-1}: P_{(\delta_i, g_i)} \to P_{(\delta_{i-1}, g_{i-1})}$, for odd $3 \leq i \leq n$, and that the concatenation of $(\delta_1, g_1), \dots, (\delta_n, g_n)$ has the same grading as $(\sigma, h)$ (this simply entails choosing $g_1$ such that there is map $P_{(\sigma, h)} \to P_{(\delta_1, g_1[1])}$ at their shared endpoint).
Having established that $P_{(\delta_i, g_i)}, \dots, P_{(\delta_i, g_i)}$ satisfy conditions (i) to (iii), it remains to prove that $P_{\sigma, h}$ is the mapping cone of the following map: $$\bigoplus_{i \, \mathrm{odd}} P_{(\delta_i, g_i)} \nameto{f} \bigoplus_{i \, \mathrm{even}} P_{(\delta_i, g_i)}$$
where the only components of $f$ from $P_{(\delta_i, g_i)}$ are $f_{i, i-1}$ and $f_{i, i+1}$, if defined.
We note that the case of $n = 1$ is trivial and that the case of $n = 2$ is a direct application of Theorem 4.1 in \cite{opper2018geometric}, which yields that the mapping cone of a map arising from two $\color{dark-green}{\circ}$-arcs meeting at a $\color{dark-green}{\circ}$-endpoint corresponds to their concatenation at this common endpoint. Subsequently, we proceed in calculation of the mapping cone by induction. For the induction, we abstract from the condition (ii).
Noting that $\Kb(\Lambda \textrm{-}\mathrm{proj})$ is a Krull-Schmidt category, we apply Lemma \ref{LMappingConeReduction} to the following datum:
$$P_{(\delta_1, g_1)} \oplus \bigoplus_{i > 1, \, \mathrm{odd}} P_{(\delta_i, g_i)} \nameto{(f_{1,2}, f')} \bigoplus_{i \, \mathrm{even}} P_{(\delta_i, g_i)}$$
It yields that the mapping cone of $f$ isomorphic to the mapping cone of the composition:
$$\bigoplus_{i > 1, \, \mathrm{odd}} P_{(\delta_i, g_i)} \nameto{f'} P_{(\delta_2, g_2)} \oplus \bigoplus_{i > 2 \, \mathrm{even}} P_{(\delta_i, g_i)} \nameto{p \oplus \mathrm{id}} \mathrm{cone}(f_{1, 2}) \oplus \bigoplus_{i > 2 \, \mathrm{even}} P_{(\delta_i, g_i)}$$
where $\mathrm{cone}(f_{1,2})$ is isomorphic to $P_{(\delta_{1,2}, h_{1,2})}$, $\delta_{1,2}$ is the concatenation of $\delta_1$ and $\delta_2$ at their shared endpoint that gives rise to $f_{1,2}$ and the grading $h_{1,2}$ is such that there is a map $P_{(\delta_{1,2}, h_{1,2})} \to P_{(\delta_1, h_1[1])}$ at the other endpoint of $\delta_1$, and $p: P_{(\delta_2, g_2)} \to \mathrm{cone}(f_{1,2})$ is the map from the distinguished triangle pertaining to $f_{1,2}$. The map $p$ is represented by a common endpoint of $\delta_2$ and $\delta_{1,2}$, which is the other endpoint of $\delta_2$ shared with $\delta_3$. This means that $p \circ f_{3,2}$ is non-zero and it is represented by a common endpoint.
We reduced the number of $\color{dark-green}{\circ}$-arcs by one, and we have thus performed the induction step, and our proof is complete.
\end{proof}
\subsection{Good cuts of a marked surface and semiorthogonal decompositions}
In this subsection, we associate a specific way how to cut $(S, M, P)$, a \textit{good cut}, to every bipartite admissible dissections thereof, and we prove that two bipartite admissible dissections of $(S, M, P)$ yield the same semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ if and only if they cut the underlying marked surface along the same lines up to homotopy.
Also, we show that for any good cut of $(S, M, P)$ it is always possible to construct an associated bipartite admissible of $(S, M, P)$ such that they cut the underlying marked surface along the same lines up to homotopy. Hence, we obtain a one-to-one correspondence between semiorthogonal decompositions of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ and good cuts of $(S, M, P)$.
\begin{definition}\label{DDividingArcs}
Given $(\Gamma, p)$, a bipartite admissible dissection of $(S, M, P)$, we say that a connected component, $\mathcal{P}$, of the complement of $\Gamma$ in $S \, \backslash \, P$ is mixed if its boundary contains arcs in different parts of $(\Gamma, p)$. There is an arc in $\mathcal{P}$ from the unique $\color{dark-green}{\circ}$-point on the boundary at which arcs in different parts of $(\Gamma, p)$ meet to the $\color{red} \bullet$-point on the $\partial S$-segment of the boundary of $\mathcal{P}$. We call it a \textit{dividing arc} of $\mathcal{P}$.
\end{definition}
\begin{lemma}\label{LCrossingDividingArcs}
Let $(\Gamma, p)$ be a bipartite admissible dissection of $(S, M, P)$, and let $\sigma$ be $\color{dark-green}{\circ}$-arc in $S$. The arc $\sigma$ does not cross any dividing arc of a mixed component of the complement of $\Gamma$ in $S \, \backslash \, P$ if and only if $P_{(\sigma, h)}$ lies in $\mathcal{L}_{(\Gamma, p)}$ or $\mathcal{R}_{(\Gamma, p)}$.
\end{lemma}
\begin{proof}
We consider a sequence, $\gamma_1^{(1)}, \dots, \gamma_{m_1}^{(1)}, \gamma_1^{(2)}, \dots, \gamma_{m_2}^{(2)}, \dots, \gamma_1^{(n)}, \dots, \gamma_{m_n}^{(n)}$, of $\color{dark-green}{\circ}$-arcs in $\Gamma$ such that $\sigma$ is homotopic to the iterated concatenation thereof as constructed in the proof of Proposition \ref{PBipartiteDissectionDecomposition}. The function $p$ attains the same value for all arcs in the sequence because $\sigma$ does not cross any dividing arc of a mixed component of the complement of $\Gamma$ in $S \, \backslash \, P$.
The object $P_{(\sigma, h)}$ lies in the triangulated subcategory generated by $P_{\gamma_j^{(i)}}$. Because all $\gamma_j^{(i)}$ belong to the same part of $(\Gamma, p)$, $P_{(\sigma, h)}$ lies inside either $\mathcal{L}_{(\Gamma, p)}$ or $\mathcal{R}_{(\Gamma, p)}$.
Now, we suppose that $\sigma$ crosses a dividing arc of a mixed component $\mathcal{P}$ of the complement of $\Gamma$ in $S \, \backslash \, P$. This means that $\sigma$ has to cross $\gamma_1 \in \Gamma$ such that $p(\gamma_1) = 1$ or share an endpoint with $\gamma_1$ such that $\gamma_1$ follows $\sigma$ in the counter-clockwise order around it and that $\sigma$ has to cross $\gamma_2 \in \Gamma$ such that $p(\gamma_2) = 2$ or share an endpoint with $\gamma_2$ such that $\sigma$ follows $\gamma_2$ in the counter-clockwise order around it. Therefore, there are non-trivial morphisms $P_{(\sigma, h)} \to P_{(\gamma_1, c_1)}$ and $P_{(\gamma_2, c_2)} \to P_{(\sigma, h)}$ for suitable gradings. Because $P_{(\gamma_1, c_1)} \in \mathcal{L}_{(\Gamma, p)}$ and $P_{(\gamma_2, c_2)} \in \mathcal{R}_{(\Gamma, p)}$, we obtain that $P_{(\sigma, h)}$ has to be trivial.
\end{proof}
\begin{proposition}\label{PBipartiteDissectionsEquivalence}
Two bipartite admissible dissection of $(S, M, P)$ give rise to the same semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ if and only if dividing arcs of their mixed components are homotopic.
\end{proposition}
\begin{proof}
Let us have two bipartite admissible dissections $(\Gamma_1, p_1)$ and $(\Gamma_2, p_2)$. Without loss of generality, we suppose that there exists a dividing arc $\Psi$ of $(\Gamma_1, p_1)$ that is not, up to homotopy, a dividing arc of $(\Gamma_2, p_2)$. If $\Psi$ crosses an arc in $\Gamma_2$, it follows from Lemma \ref{LCrossingDividingArcs} above that $(\Gamma_1, p_1)$ and $(\Gamma_2, p_2)$ do not give the same semiorthogonal decompositions of $\Kb(\Lambda \textrm{-}\mathrm{proj})$.
We can assume that $\Psi$ lies in a connected component $\mathcal{P}$ of the complement of $\Gamma_1$ in $S \, \backslash \, P$. There are two $\color{dark-green}{\circ}$-arcs $\sigma_1, \sigma_2 \in \Gamma_2$ on the boundary of $P$ sharing a common endpoint with $\Psi$; the $\color{dark-green}{\circ}$-endpoint of $\Psi$ may not be adjacent to the $\color{red} \bullet$-endpoint on the same boundary component of $S$ since $\Psi$ would not be a dividing arc otherwise. Also, $\sigma_1$ and $\sigma_2$ need to belong to the same part of $(\Gamma_2, p_2)$ because $\Psi$ is not a dividing arc of $(\Gamma_2, p_2)$. Therefore, the object of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ corresponding, up to grading, to the concatenation of $\sigma_1$ and $\sigma_2$ lies in $\mathcal{L}_{(\Gamma_2, p_2)}$ or $\mathcal{R}_{(\Gamma_2, p_2)}$, but it belongs to neither $\mathcal{L}_{(\Gamma_1, p_1)}$ nor $\mathcal{R}_{(\Gamma_1, p_1)}$ by Lemma \ref{LCrossingDividingArcs} as it crosses a dividing arc $\Psi$ of $(\Gamma_2, p_2)$.\\
Suppose that $(\Gamma_1, p_1)$ and $(\Gamma_2, p_2)$, two bipartite admissible dissections of $(S, M, P)$, give rise to different semiorthogonal decompositions of $\Kb(\Lambda \textrm{-}\mathrm{proj})$. Provided that there exists a $\color{dark-green}{\circ}$-arc $\sigma \in \Sigma_1$ such that the corresponding object of $\Kb(\Lambda \textrm{-}\mathrm{proj})$, up to grading, does not lie in either $\mathcal{L}_{(\Gamma_2, p_2)}$ or $\mathcal{R}_{(\Gamma_2, p_2)}$, it follows from Lemma \ref{LCrossingDividingArcs} that $\sigma$ crosses a dividing arc of $(\Gamma_2, p_2)$. Hence, $(\Gamma_1, p_1)$ and $(\Gamma_2, p_2)$ do not have, up to homotopy, the same dividing arcs.
Suppose now that all objects of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ corresponding to $\color{dark-green}{\circ}$-arcs in $\Gamma_1$ given some grading lie either in $\mathcal{L}_{(\Gamma_2, p_2)}$ or in $\mathcal{R}_{(\Gamma_2, p_2)}$. Consider a function $p_1'$ that assigns an arc $\sigma \in \Gamma_1$ the value $1$ if the corresponding object $\Kb(\Lambda \textrm{-}\mathrm{proj})$ lies in $\mathcal{L}_{(\Gamma_2, p_2)}$ and the value $2$ if the corresponding object $\Kb(\Lambda \textrm{-}\mathrm{proj})$ lies in $\mathcal{R}_{(\Gamma_2, p_2)}$. In this way, we obtain a bipartite admissible dissection $(\Gamma_1, p_1')$. We have that $\mathcal{L}_{(\Gamma_1, p_1')} \subseteq \mathcal{L}_{(\Gamma_2, p_2)}$ and $\mathcal{R}_{(\Gamma_1, p_1')} \subseteq \mathcal{R}_{(\Gamma_2, p_2)}$; thus $(\Gamma_1, p_1')$ and $(\Gamma_2, p_2)$ give rise to the same semiorthogonal decompositions of $\Kb(\Lambda \textrm{-}\mathrm{proj})$.
From the proof of the other implication above, we get that $(\Gamma_1, p_1')$ and $(\Gamma_2, p_2)$ have the same dividing arcs. However, $(\Gamma_1, p_1)$ and $(\Gamma_1, p_1')$ cannot have the same dividing arcs because it would imply that $p_1 = p_1'$.
\end{proof}
\begin{definition}\label{DCuttingMarkedSurface}
Let $(S, M, P)$ be marked surface. A \textit{dividing arc} is a smooth map $\omega: (0,1) \to S \, \backslash \, P$ such that $g_\omega = \lim_{x \to 0} \omega(x) \in M_{\color{dark-green}{\circ}}$ and $r_\omega = \lim_{x \to 1} \omega(x) \in M_{\color{red} \bullet}$ such that $g_\omega$ and $r_\omega$ are not adjacent on a component of $\partial S$. We can obtain a new marked surface, the \textit{cut surface}, $(S_\omega, M_\omega, P_\omega)$ by cutting $(S, M, P)$ along $\omega$.
The marked points of $(S_\omega, M_\omega, P_\omega)$ are $P_\omega = P$, $M_\omega {}_{\color{dark-green}{\circ}} = (M_{\color{dark-green}{\circ}} \backslash \{g_\omega\}) \cup \{g_\omega^\ell, g_\omega^r\}$ and $M_\omega {}_{\color{red} \bullet} = (M_{\color{red} \bullet} \backslash \{r_\omega\}) \cup \{r_\omega^\ell, r_\omega^r\}$ such that $g_\omega^\ell$ to $r_\omega^\ell$ precede $g_\omega^r$ to $r_\omega^r$ in the counter-clockwise order on their shared component of the boundary of $S_\omega$. We refer to $g_\omega^\ell, r_\omega^\ell$ and $g_\omega^r, r_\omega^r$ as to \textit{left added marked points} and \textit{right added marked points}.
\begin{figure}[H]
\centering
\begin{tikzpicture}[x = 1cm,y = 1cm]
\node[coordinate, name = GOmega] at (-3, 1) {};
\node[coordinate, name = ROmega] at (-3, -1) {};
\draw[dashed] (GOmega) -- (ROmega) node[midway, right] {$\omega$};
\draw (GOmega) arc (270:250:5) node[coordinate, name = GOmegaAuxLeft] {};
\draw (GOmega) arc (270:290:5) node[coordinate, name = GOmegaAuxRight] {};
\fill[pattern = north east lines] (GOmegaAuxRight) -- (GOmegaAuxLeft) arc (250:290:5) -- cycle;
\draw (ROmega) arc (90:110:5) node[coordinate, name = ROmegaAuxLeft] {};
\draw (ROmega) arc (90:70:5) node[coordinate, name = ROmegaAuxRight] {};
\fill[pattern = north east lines] (ROmegaAuxRight) -- (ROmegaAuxLeft) arc (110:70:5) -- cycle;
\draw[dark-green, fill = white] (GOmega) circle (0.1) ;
\draw[red, fill = red] (GOmegaAuxLeft) circle (0.1) ;
\draw[red, fill = red] (GOmegaAuxRight) circle (0.1) ;
\draw[red, fill = red] (ROmega) circle (0.1) ;
\draw[dark-green, fill = white] (ROmegaAuxLeft) circle (0.1) ;
\draw[dark-green, fill = white] (ROmegaAuxRight) circle (0.1) ;
\node[below left] at (GOmega) {$g_\omega$};
\node[above left] at (ROmega) {$r_\omega$};
\draw[->, decorate, decoration = {snake, segment length = 2mm, amplitude = 0.25mm}] (-1,0) -- (1,0) node[] {};
\node[coordinate, name = GOmegaCut] at (3, 1) {};
\node[coordinate, name = ROmegaCut] at (3, -1) {};
\path (GOmegaCut) arc (270:265:5) node[coordinate, name = GOmegaCut1] {};
\path (GOmegaCut) arc (270:275:5) node[coordinate, name = GOmegaCut2] {};
\path (ROmegaCut) arc (90:95:5) node[coordinate, name = ROmegaCut1] {};
\path (ROmegaCut) arc (90:85:5) node[coordinate, name = ROmegaCut2] {};
\draw (GOmegaCut1) arc (265:250:5) node[coordinate, name = GOmegaCut1Aux] {};
\draw (GOmegaCut2) arc (275:290:5) node[coordinate, name = GOmegaCut2Aux] {};
\draw (ROmegaCut1) arc (95:110:5) node[coordinate, name = ROmegaCut1Aux] {};
\draw (ROmegaCut2) arc (85:70:5) node[coordinate, name = ROmegaCut2Aux] {};
\draw (GOmegaCut1) -- (ROmegaCut1);
\draw (GOmegaCut2) -- (ROmegaCut2);
\fill[pattern = north east lines] (GOmegaCut1Aux) arc (250:265:5) -- (ROmegaCut1) arc (95:110:5) -- (ROmegaCut2Aux) arc (70:85:5) -- (GOmegaCut2) arc (275:290:5) -- cycle;
\draw[dark-green, fill = white] (GOmegaCut1) circle (0.1) ;
\draw[red, fill = red] (ROmegaCut1) circle (0.1) ;
\draw[dark-green, fill = white] (ROmegaCut1Aux) circle (0.1) ;
\draw[red, fill = red] (GOmegaCut1Aux) circle (0.1) ;
\draw[dark-green, fill = white] (GOmegaCut2) circle (0.1) ;
\draw[red, fill = red] (ROmegaCut2) circle (0.1) ;
\draw[dark-green, fill = white] (ROmegaCut2Aux) circle (0.1) ;
\draw[red, fill = red] (GOmegaCut2Aux) circle (0.1) ;
\node[below right] at (GOmegaCut2) {$g_\omega^\ell$};
\node[above right] at (ROmegaCut2) {$r_\omega^\ell$};
\node[below left] at (GOmegaCut1) {$g_\omega^r$};
\node[above left] at (ROmegaCut1) {$r_\omega^r$};
\end{tikzpicture}
\caption{Cutting $S$ along $\omega$}
\label{FCutSurface}
\end{figure}
\end{definition}
\begin{remark}
It is easy to observe that $(S_\omega, M_\omega, P_\omega)$ is a well-defined marked surface because the new marked points are added in way that ensures that $\color{dark-green}{\circ}$-marked points and $\color{red} \bullet$-marked points still alternate on the component of $\partial S_\omega$ that they lie on.
\end{remark}
\begin{remark}
It is always possible to reconstruct $(S, M, P)$ from $(S_\omega, M_\omega, P_\omega)$ by identifying the segments of $\partial S_\omega$ from $g_\omega^\ell$ to $r_\omega^\ell$ and from $g_\omega^r$ to $r_\omega^r$.
\end{remark}
We observe that if $\omega_1$ and $\omega_2$ are two dividing arcs on $(S, M, P)$ that do not intersect or share a common endpoint, it is possible to cut $(S, M, P)$ along them simultaneously. This observation is used in the following definition:
\begin{definition}\label{DGoodCut}
Let $(S, M, P)$ be marked surface. A collection of dividing arcs up to homotopy, $\Omega$, is called a \textit{good cut} if:
\begin{enumerate}[(i)]
\item any two different arcs in $\Omega$ neither intersect nor meet at endpoints,
\item no connected component of the cut surface $(S_\Omega, M_\Omega, P_\Omega)$ contains an left added marked point and an right added marked point,
\item no connected component of the cut surface $(S_\Omega, M_\Omega, P_\Omega)$ is trivial, \textit{i.e.} homeomorphic to an open disk with only two marked points, one $\color{dark-green}{\circ}$-point and $\color{red} \bullet$-point, on its boundary.
\end{enumerate}
\end{definition}
\begin{proposition}\label{PDissectionGoodCut}
Dividing arcs of a bipartite admissible dissection $(\Gamma, p)$ form a good cut $\Omega_{(\Gamma, p)}$ of the associated marked surface. On the other hand, a good cut $\Omega$ gives rise to a bipartite admissible dissection $(\Gamma_\Omega, p_\Omega)$ of the associated marked surface such that the dividing arcs of $(\Gamma_\Omega, p_\Omega)$ are the same as the arcs in $\Omega$.
\end{proposition}
\begin{proof}
Suppose that $(\Gamma, p)$ is a bipartite admissible dissection and that $\Omega_{(\Gamma, p)}$ is the associated set of dividing arcs. We will prove that $\Omega_{(\Gamma, p)}$ is a good cut of the underlying marked surface. Firstly, we observe that dividing arcs of $(\Gamma, p)$ lie in different connected components of the complement of $\Gamma$ in $S \, \backslash \, P$. Therefore, two dividing arcs of $\Gamma$ may not intersect. For the same reason, they may also not meet at $\color{red} \bullet$-marked point endpoints. If two dividing arcs of $(\Gamma, p)$ were to meet at a $\color{dark-green}{\circ}$-marked point, it would mean that in the counter-clockwise order there would be arcs $\gamma_1, \gamma_2, \gamma_3, \gamma_4$ with $p(\gamma_i) = 1, i \in \{1, 3\}$ and $p(\gamma_i) = 2, i \in \{2, 4\}$, a contradiction.
Assume now that, some connected component $\mathcal{Q}$ of the complement of $\Omega_{(\Gamma, p)}$ in $S \, \backslash \, P$ has both left and right added marked points on its boundary. It is easy to observe that if there is $g_\omega^\ell \in \partial \mathcal{Q}$ and $g_\omega^r \in \partial \mathcal{Q}$ for $\omega \in \Omega_{(\Gamma, p)}$, then we have $r_\omega^\ell \in \partial \mathcal{Q}$ and $r_\omega^r \in \partial \mathcal{Q}$, respectively, and vice versa because dividing arcs of $(\Gamma, p)$ do not intersect.
So there are $g_{\omega_1}^\ell, r_{\omega_1}^\ell, g_{\omega_2}^r$, and $r_{\omega_2}^r$ on the boundary of $\mathcal{Q}$ for some $\omega_1, \omega_2 \in \Omega_{(\Gamma, p)}$. This means that there are $\color{dark-green}{\circ}$-arcs $\gamma_1, \gamma_2 \in \Gamma$ such that $p(\gamma_i) = i$ that lie in $\mathcal{Q}$. Because $\mathcal{Q}$ is connected, we can take a $\color{dark-green}{\circ}$-arc $\sigma$ that connects $\color{dark-green}{\circ}$-marked points which are endpoints of the two dividing arcs pertaining to $\gamma_1$ and $\gamma_2$. Without loss of generality, $\sigma$ starts in the connected component of the complement of $\Gamma$ in $S \, \backslash \, P$ where $\omega_1$ lies, and $\sigma$ ends in the connected component of the complement of $\Gamma$ in $S \, \backslash \, P$ where $\omega_2$ lies.
Now, we consider a sequence, $\gamma_1^{(1)}, \dots, \gamma_{m_1}^{(1)}, \gamma_1^{(2)}, \dots, \gamma_{m_2}^{(2)}, \dots, \gamma_1^{(n)}, \dots, \gamma_{m_n}^{(n)}$, of $\color{dark-green}{\circ}$-arcs in $\Gamma$ such that $\sigma$ is homotopic to the iterated concatenation thereof as constructed in the proof of Proposition \ref{PBipartiteDissectionDecomposition}. The $\color{dark-green}{\circ}$-arc $\sigma$ in such a way that $\gamma_1^{(1)} = \gamma_1$ and $\gamma_{m_n}^{(n)} = \gamma_1$. As $p(\gamma_1) = 1$ and $p(\gamma_2) = 2$, the concatenation of $\gamma_1^{(1)}, \dots, \gamma_{m_1}^{(1)}, \gamma_1^{(2)}, \dots, \gamma_{m_2}^{(2)}, \dots, \gamma_1^{(n)}, \dots, \gamma_{m_n}^{(n)}$, which is $\sigma$, has to cross a dividing arc in $\Omega_{(\Gamma, p)}$. We have thus obtained a contradiction with the assumption that $\sigma$ is connected in the complement of $\Omega_{(\Gamma, p)}$ in $S \, \backslash \, P$.
To establish that $\Omega_{(\Gamma, p)}$ is a good cut, it remains to be shown that no connected component of its complement in $S \backslash P$ is trivial. However, for each $\omega \in \Omega_{(\Gamma, p)}$ there are two $\color{dark-green}{\circ}$-arcs $\color{dark-green}{\circ}$-arcs $\gamma_1, \gamma_2 \in \Gamma$ such that they share a $\color{dark-green}{\circ}$-endpoint with $\omega$ and that $p(\gamma_i) = i$. No dividing may cross either $\gamma_1$ or $\gamma_2$, so $\gamma_1$ lies in the component of the complement of $\Omega_{(\Gamma, p)}$ in $S \backslash P$ that has the left added marked point $g\_omega^\ell$ on this boundary, similarly for $\gamma_2$. Because every component of the complement of $\Omega_{(\Gamma, p)}$ in $S \backslash P$ contains an added marked point, it contains a $\color{dark-green}{\circ}$-arc, and so it is non-trivial.
Denote the marked surface $(S, M, P)$. Consider the connected components of the complement of $\Omega$ in $S \, \backslash \, P$. Each connected components of the complement of $\Omega$ in $S \, \backslash \, P$ is a marked surface and has an admissible dissection.
We choose an admissible dissection for each connected component and observe that together, as a set $\Gamma_\Omega$, they from an admissible dissection of the marked surface $(S, M, P)$. Arcs in admissible dissections of components that come from components with only left added marked points on their boundary are assigned $2$ by $p_\Omega$; other arcs, coming from components with only right added marked points are assigned $1$ by $p_\Omega$.
The $\color{dark-green}{\circ}$-arcs in $\Gamma_\Omega$ may clearly only meet at ther common $\color{dark-green}{\circ}$-endpoints since this is the case for every constituent connected component of the complement of $\Omega$ in $S \, \backslash \, P$, which are glued on segments on their boundary. Each connected component of the complement of $\Gamma_\Omega$ in $S \, \backslash \, P$ is glued from at most two connected components of the complement of $\Gamma_\Omega$ in the respective constituent connected components of the complement of $\Omega$ in $S \, \backslash \, P$ in the following way:
\begin{figure}[H]
\centering
\begin{tikzpicture}[x = 1cm, y = 1cm]
\node[regular polygon, regular polygon sides = 5, minimum size = 3 cm, shape border rotate = 18] (Pen1) at (-2,0) {};
\foreach \x in {1,...,5}
{\node[coordinate, name = {Pen1Node\x}] at (Pen1.corner \x) {};}
\node[regular polygon, regular polygon sides = 5, minimum size = 3 cm, shape border rotate = -18] (Pen2) at (2,0) {};
\foreach \x in {1,...,5}
{\node[coordinate, name = {Pen2Node\x}] at (Pen2.corner \x) {};}
\draw (Pen1Node5) arc (36:-108:1.5) node[midway, coordinate, name = Pen1Red] {};
\draw[dark-green] (Pen1Node5) -- (Pen1Node1);
\draw[dark-green, dotted] (Pen1Node1) -- (Pen1Node2);
\draw[dark-green] (Pen1Node2) -- (Pen1Node3);
\draw (Pen2Node2) arc (144:288:1.5) node[midway, coordinate, name = Pen2Red] {};
\draw[dark-green] (Pen2Node4) -- (Pen2Node5);
\draw[dark-green, dotted] (Pen2Node5) -- (Pen2Node1);
\draw[dark-green] (Pen2Node1) -- (Pen2Node2);
\fill[pattern = north east lines] (Pen1Node3) arc (-108:36:1.5) to[out = 20, in = 160] (Pen2Node2) arc (144:288:1.5) to[out = 200, in = -20] (Pen1Node3);
\foreach \x in {1,2,3,5}
{\draw[dark-green, fill = white] ({Pen1Node\x}) circle (0.1);}
\foreach \x in {1,2,4,5}
{\draw[dark-green, fill = white] ({Pen2Node\x}) circle (0.1);}
\draw[red, fill = red] (Pen1Red) circle (0.1);
\draw[red, fill = red] (Pen2Red) circle (0.1);
\node[below left] at (Pen1Node5) {$g^r_\omega$};
\node[below right] at (Pen2Node2) {$g^\ell_\omega$};
\node[above left] at (Pen1Red) {$r^r_\omega$};
\node[above right] at (Pen2Red) {$r^\ell_\omega$};
\end{tikzpicture}
\begin{tikzpicture}[x = 1cm, y = 1cm]
\path (0,1) -- (0,0.5);
\draw[->, decorate, decoration = {snake, segment length = 2mm, amplitude = 0.25mm}] (0,0.5) -- (0,-0.5) node[] {};
\end{tikzpicture}
\begin{tikzpicture}[x = 1cm, y = 1cm]
\def5{5}
\def3{3}
\pgfmathsetmacro{\penfactor}{0.5*cos(180/5)}
\pgfmathsetmacro{\penonecenter}{-3*\penfactor}
\pgfmathsetmacro{\pentwocenter}{3*\penfactor}
\node[regular polygon, regular polygon sides = 5, minimum size = 3 cm, shape border rotate = 18] (Pen1) at (\penonecenter,0) {};
\foreach \x in {1,...,5}
{\node[coordinate, name = {Pen1Node\x}] at (Pen1.corner \x) {};}
\node[regular polygon, regular polygon sides = 5, minimum size = 3 cm, shape border rotate = -18] (Pen2) at (\pentwocenter,0) {};
\foreach \x in {1,...,5}
{\node[coordinate, name = {Pen2Node\x}] at (Pen2.corner \x) {};}
\draw (Pen1Node3) to[out = 45, in = 135] node[midway, coordinate, name = Red] {} (Pen2Node4);
\draw[dark-green] (Pen1Node5) -- (Pen1Node1);
\draw[dark-green, dotted] (Pen1Node1) -- (Pen1Node2);
\draw[dark-green] (Pen1Node2) -- (Pen1Node3);
\draw[dark-green] (Pen2Node4) -- (Pen2Node5);
\draw[dark-green, dotted] (Pen2Node5) -- (Pen2Node1);
\draw[dark-green] (Pen2Node1) -- (Pen2Node2);
\draw[dashed] (Pen1Node5) -- (Red);
\fill[pattern = north east lines] (Pen1Node3) to[out = 45, in = 135] node[midway, coordinate, name = Red] {} (Pen2Node4) -- cycle;
\foreach \x in {1,2,3,5}
{\draw[dark-green, fill = white] ({Pen1Node\x}) circle (0.1);}
\foreach \x in {1,4,5}
{\draw[dark-green, fill = white] ({Pen2Node\x}) circle (0.1);}
\draw[red, fill = red] (Red) circle (0.1);
\node[below left] at (Pen1Node5) {$g_\omega$};
\node[above left] at (Red) {$r_\omega$};
\path (Pen1Node5) -- (Red) node[midway, right] {$\omega$};
\end{tikzpicture}
\caption{Gluing connected components of of the complement of $\Omega$ in $S \, \backslash \, P$.}
\label{FGluingComponentsDissection}
\end{figure}
Hence every connected component of the complement of $\Gamma_\Omega$ in $S \, \backslash \, P$ has a single $\color{red} \bullet$-point, and $\Gamma_\Omega$ is an admissible dissection.
If two arcs in different parts of $(\Gamma_\Omega, p_\Omega)$ meet, they have to meet at a $\color{dark-green}{\circ}$-endpoint of one of the dividing arcs in $\Omega$. The $\color{dark-green}{\circ}$-arc in part $2$ lies follows the dividing arc in the counterclockwise order around the common endpoint because the component to which the $\color{dark-green}{\circ}$-arc belong lies to the left of the dividing arc; whereas, the $\color{dark-green}{\circ}$-arc in part $1$ has to precede in that order. This is enough to conclude that $(\Gamma_\Omega, p_\Omega)$ is a bipartite admissible dissection as well as that $\Omega$ is the set of dividing arcs of $(\Gamma_\Omega, p_\Omega)$.
\end{proof}
\begin{remark}
The proof that $\Gamma_\Omega$ may be also performed numerically using the following inductive argument as in the proof of Proposition 1.11 in \cite{amiot2019complete}. By $\color{dark-green}{\circ}$-arcs in admissible dissection equals $|M_{\color{dark-green}{\circ}}| + |P| + b + 2g - 2$ arcs, where $g$ is the genus of $S$ and $b$ is the number of connected components of $\partial S$. Consider $\omega$ a dividing arc of $S$ and the cut surface $S_\omega$. If $\omega$ connects two different connected components of the boundary $\partial S$ (Case 3 in the proof of Proposition 1.11 in \cite{amiot2019complete}), then the number of connected components of $S_\omega$, $b_\omega$, equals $b-1$. We have an extra $\color{dark-green}{\circ}$-marked point, so $(M_\omega)_{\color{dark-green}{\circ}} = M_{\color{dark-green}{\circ}} + 1$. It is easy to observe that the genus of $S_\omega$, $g_\omega$, and the number $\color{red} \bullet$-punctures remain unchanged, so $g = g_\omega$ and $|P_\omega| = |P|$. We get that any admissible dissection of $S_\omega$ has to have the same number of $\color{dark-green}{\circ}$-arcs as that of $S$. If $\omega$ starts and ends on the same boundary component, we need to distinguish whether $\omega$ is separating or non-separating. Provided that $\omega$ is non-separating (Case 1 in the proof of Proposition 1.11 in \cite{amiot2019complete}), the cut surface $S_\omega$ has a single connected component with $b+1$ boundary components and genus $g-1$. It has an extra $\color{dark-green}{\circ}$-marked point and the same number of punctures. Because $|(M_\omega)_{\color{dark-green}{\circ}}| + |P_\omega| + b_\omega + 2g_\omega - 2 = (|M_{\color{dark-green}{\circ}}| + 1) + |P| + (b + 1) + 2(g - 1) - 2 = |M_{\color{dark-green}{\circ}}| + |P| + b + 2g - 2$, the any admissible dissection of $S_\omega$ has to have the same number of $\color{dark-green}{\circ}$-arcs as that of $S$. Finally, if $\omega$ is separating (Case 6 in the proof of Proposition 1.11 in \cite{amiot2019complete}), then $S_\omega$ has two connected components $S_\omega^{(1)}$ and $S_\omega^{(2)}$ with combined genus, $g_\omega^{(1)} + g_\omega^{(2)}$, equal to $g$ and combined number of boundary components, $b_\omega^{(1)} + b_\omega^{(2)}$, equal to $b+1$. They also have an additional marked point together, in other words $|(M_\omega^{(1)})_{\color{dark-green}{\circ}}| + |(M_\omega^{(2)})_{\color{dark-green}{\circ}}| = |M_{\color{dark-green}{\circ}}|$. Therefore the number of $\color{dark-green}{\circ}$-arcs in a pair admissible dissections of $S_\omega^{(1)}$ and $S_\omega^{(2)}$ equals to $|(M_\omega^{(1)})_{\color{dark-green}{\circ}}| + |P_\omega^{(1)}| + b_\omega^{(1)} + 2g_\omega^{(1)} - 2 + |(M_\omega^{(2)})_{\color{dark-green}{\circ}}| + |P_\omega^{(2)}| + b_\omega^{(2)} + 2g_\omega^{(2)} - 2 = (|M_{\color{dark-green}{\circ}}| + 1) + |P| + (b + 1) + 2g - 4$, which is the same as for $S$.
\end{remark}
\begin{theorem}\label{TDecompositionsGoodCuts}
There is a one-to-one correspondence between semiorthogonal decompositions of $\Kb(\Lambda\textrm{-}\mathrm{proj})$ and good cuts of the marked surface associated to $\Lambda$.
\end{theorem}
\begin{proof}
For the purposes of this proof, we denote $\mathcal{S} \mathcal{O}$ the set of semiorthogonal decompositions of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ and $\mathcal{G} \mathcal{C}$ the set of cuts of $(S, M, P)$.
At first, we construct a map $F\colon \mathcal{S} \mathcal{O} \to \mathcal{G} \mathcal{C}$. Let us have $\langle \mathcal{L}, \mathcal{R} \rangle$, a semiorthogonal decomposition of $\Kb(\Lambda\textrm{-}\mathrm{proj})$. Consider the following approximation of $R_i \to P_i \to L_i \to R_i[1]$ such that and $R_i \in \mathcal{R}$, $L_i \in \mathcal{L}$, and $P_i$ is one of the indecomposable direct summands of $\Lambda$. Theorem \ref{TProjectivesDecomposition} paired with Lemma 3.10 in \cite{opper2018geometric} yields that: $$R_i \cong \bigoplus_{j} P_{(\varrho_j^{(i)}, g_j^{(i)})} \, \, \, \mbox{and} \, \, \, L_i \cong \bigoplus_{j'} P_{(\lambda_{j'}^{(i)}, \ell_{j'}^{(i)})}$$ for $\color{dark-green}{\circ}$-arcs $\varrho_j^{(i)}$ and $\lambda_{j'}^{(i')}$; these $\color{dark-green}{\circ}$-arcs meet only at endpoints, and that $\varrho_j^{(i)}$ may meet $\lambda_{j'}^{(i')}$ at a $\color{dark-green}{\circ}$-point only if $\varrho_j^{(i)}$ follows $\lambda_{j'}^{(i')}$ in the counter-clockwise order around a common endpoint.
These $\color{dark-green}{\circ}$-arcs gives rise to a bipartite collection $(\Gamma, p)$ with $p(\lambda_{j'}^{(i')}) = 1$ and $p(\varrho_j^{(i)}) = 2$ by possibly removing redundant $\color{dark-green}{\circ}$-arcs so that each connected component of its complement in contains at least one $\color{red} \bullet$-point.
It easy to observe that the removed $\color{dark-green}{\circ}$-arcs are simply concatenations of the remaining $\color{dark-green}{\circ}$-arcs, and the objects corresponding to the removed $\color{dark-green}{\circ}$-arcs can be generated using the objects corresponding to the remaining $\color{dark-green}{\circ}$-arcs for both parts respectively. This holds because if there is a component of of the complement of without a $\color{red} \bullet$-marked point, then all $\color{dark-green}{\circ}$-arcs on its boundary have to be either $\lambda_{j'}^{(i')}$ or $\varrho_j^{(i)}$ as discussed in the proof of Proposition \ref{PBipartiteDissectionDecomposition}.
Proposition \ref{PBipartiteDissectionCompletion} assures existence of a bipartite admissible dissection $(\Gamma', p')$ extending $(\Gamma, p)$. If $\langle \mathcal{L}_{(\Gamma', p')}, \mathcal{R}_{(\Gamma', p')} \rangle$ is the semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ associated to $(\Gamma', p')$ by Proposition \ref{PBipartiteDissectionDecomposition}, we obtain that $\mathcal{L} \subseteq \mathcal{L}_{(\Gamma', p')}$ and $\mathcal{R} \subseteq \mathcal{R}_{(\Gamma', p')}$ necessitating that the inclusions are equalities.
We note that $(\Gamma, p)$ must have already been a bipartite admissible dissection. Suppose that we have a $\color{red} \bullet$-arc $\psi$. Since the $\color{dark-green}{\circ}$-arcs $\sigma_i$ corresponding to indecomposable projectives $P_i$ form an admissible dissection, there exists $\sigma_i$ such that $\psi$ and $\sigma_i$ intersect. The uniqueness of approximation triangles in $\langle \mathcal{L}, \mathcal{R} \rangle = \langle \mathcal{L}_{(\Gamma', p')}, \mathcal{R}_{(\Gamma', p')} \rangle$ per Proposition \ref{PSemiorthogonalDecomposition} means that we can use Proposition \ref{PBipartiteDissectionTriangles} to show that $\sigma_i$ homotopic to an iterated concatenation of $\color{dark-green}{\circ}$-arcs $\lambda_{j'}^{(i)}$ and $\varrho_{j}^{(i)}$. It is an easy consequence that $\psi$ has to cross some of these $\color{dark-green}{\circ}$-arcs. We can therefore conclude that $(\Gamma, p)$ separates all $\color{red} \bullet$-points, and so it is a bipartite admissible dissection.
We set $F(\langle \mathcal{L}, \mathcal{R} \rangle)$ equal to the set of dividing arcs $(\Gamma', p')$. The bipartite admissible dissection $(\Gamma', p')$ may not be unique, but all possible choices of $(\Gamma', p')$ have the same dividing arcs. Therefore, the map $F\colon \mathcal{S} \mathcal{O} \to \mathcal{G} \mathcal{C}$ is well defined.\\\\
Conversely, we construct a map $G: \mathcal{G} \mathcal{C} \to \mathcal{S} \mathcal{D}$. Take a good cut $\Omega$ of the underlying marked surface. By Proposition \ref{PDissectionGoodCut}, there exists a bipartite admissible dissection $(\Gamma', p')$ such that its set of dividing arcs equals $\Omega$. By Propostion \ref{PBipartiteDissectionDecomposition}, we can associate a semiorthogonal decomposition $G(\Omega)$ of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ to $\Omega$. Since two bipartite admissible dissections with same dividing arcs give rise to the same semiorthogonal decomposition by Proposition \ref{PBipartiteDissectionsEquivalence}, the map $G: \mathcal{G} \mathcal{C} \to \mathcal{S} \mathcal{D}$ is well-defined.\\\\
Proposition \ref{PDissectionGoodCut} gives us that two bipartite admissible dissections give rise to the same semiorthogonal decomposition by Proposition \ref{PBipartiteDissectionsEquivalence} if and only if they have the same dividing arcs. This fact is then used to assure that $F$ and $G$ are mutually inverse bijections.
\end{proof}
\subsection{Semiorthogonal decompositions of \texorpdfstring{$\Db(\Lambda \textrm{-}\mathrm{mod})$}{Db(Lambda-mod)} and \texorpdfstring{$\Kb(\Lambda \textrm{-}\mathrm{proj})$}{Kb(Lambda-proj)}}
Finally, we prove that there is a bijective correspondence between semiorthogonal decompositions of $\Db(\Lambda \textrm{-}\mathrm{mod})$ and semiorthogonal decompositions of $\Kb(\Lambda \textrm{-}\mathrm{proj})$, which goes from left to right by restriction. This results represents the final step in establishing the bijective correspondence between semiorthogonal decompositions of $\Db(\Lambda \textrm{-}\mathrm{mod})$ and good cuts of the underlying marked surface $(S, M, P)$.
\begin{theorem}\label{TSemiorthogonalDecompositionRestriction}
A semiorthogonal decomposition of $\Db(\Lambda\textrm{-}\mathrm{mod})$ restricts to a semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$. On the other hand, a semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ can be uniquely extended to a semiorthogonal decomposition of $\Db(\Lambda\textrm{-}\mathrm{mod})$.
\end{theorem}
\begin{proof}
At first, we prove that a semiorthogonal decomposition of $\Db(\Lambda\textrm{-}\mathrm{mod})$ restricts to a semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$. The first part of the argument is similar to the argument made in the proof of Theorem \ref{TDecompositionsGoodCuts}, when constructing the map from semiorthogonal decompositions of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ to the good cuts of the underlying marked surface.
Suppose that $\langle \mathcal{L}, \mathcal{R} \rangle$ is a semiorthogonal decomposition of $\Db(\Lambda\textrm{-}\mathrm{mod})$. Consider the following approximation of $R_i \to P_i \to L_i \to R_i[1]$ such that and $R_i \in \mathcal{R}$, $L_i \in \mathcal{L}$, and $P_i$ is one of the indecomposable direct summands of $\Lambda$. Theorem \ref{TProjectivesDecomposition} paired with Lemma 3.10 in \cite{opper2018geometric} yields that: $$R_i \cong \bigoplus_{j} P_{(\varrho_j^{(i)}, g_j^{(i)})} \, \, \, \mbox{and} \, \, \, L_i \cong \bigoplus_{j'} P_{(\lambda_{j'}^{(i')}, \ell_{j'}^{(i')})}$$ for $\color{dark-green}{\circ}$-arcs $\varrho_j^{(i)}$ and $\lambda_{j'}^{(i')}$; these $\color{dark-green}{\circ}$-arcs meet only at endpoints, and that $\varrho_j^{(i)}$ may meet $\lambda_{j'}^{(i')}$ at a $\color{dark-green}{\circ}$-point only if $\varrho_j^{(i)}$ follows $\lambda_{j'}^{(i')}$ in the counter-clockwise order around a common endpoint.
Clearly, $\mathcal{L} \cap \Kb(\Lambda \textrm{-}\mathrm{proj})$ contains all objects corresponding to $\lambda_{j'}^{(i')}$, and $\mathcal{R} \cap \Kb(\Lambda \textrm{-}\mathrm{proj})$ contains all objects corresponding to $\varrho_j^{(i)}$. Also, there are no homomorphisms from $\mathcal{R} \cap \Kb(\Lambda \textrm{-}\mathrm{proj})$ to $\mathcal{R} \cap \Kb(\Lambda \textrm{-}\mathrm{proj})$. Because $\mathcal{L} \cap \Kb(\Lambda \textrm{-}\mathrm{proj})$ and $\mathcal{R} \cap \Kb(\Lambda \textrm{-}\mathrm{proj})$ generate $\Kb(\Lambda \textrm{-}\mathrm{proj})$ via its generator $\Lambda$, which is the direct sum of all indecomposable projectives $P_i$, they form a semiorthogonal decomposition. Therefore, $\langle \mathcal{L}, \mathcal{R} \rangle$ restricts to a semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$.
Now, it remains to be proved that a semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$ can be uniquely extended to a semiorthogonal decomposition of $\Db(\Lambda\textrm{-}\mathrm{mod})$.
Let $\langle \mathcal{L}, \mathcal{R} \rangle$ be a semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$. Our strategy is to find enough addditonal indecomposable objects in $\Db(\Lambda\textrm{-}\mathrm{mod})$ to add to $\mathcal{L}$ and $\mathcal{R}$, so that enriched $\mathcal{L}$ and $\mathcal{R}$ generated the bounded derived category, but there are still no morphisms from the enriched $\mathcal{R}$ to enriched $\mathcal{L}$ (cf. Proposition \ref{PSemiorthogonalDecomposition}). This enrichment will be constructed in a way that will assure its uniqueness.
The indecomposable objects of $\Db(\Lambda\textrm{-}\mathrm{mod})$ that are not in $\Kb(\Lambda \textrm{-}\mathrm{proj})$; these are exactly objects corresponding to infinite string complexes. In the geometric model, infinite string complexes are represented by arcs between a $\color{dark-green}{\circ}$-point and a $\color{red} \bullet$-puncture or two $\color{red} \bullet$-punctures (Theorem 2.12 and Remark 1.20 in \cite{opper2018geometric}). Such arcs infinitely wrap around the $\color{red} \bullet$-puncture which is its endpoint in the counter-clockwise direction.
Using Theorem \ref{TDecompositionsGoodCuts}, we can find a bipartite admissible dissection $(\Gamma, p)$ of the underlying marked surface $(S, M, P)$ that gives $\langle \mathcal{L}, \mathcal{R} \rangle$ by Proposition \ref{PBipartiteDissectionDecomposition}.
Consider the complement of $\Gamma$ in $S \, \backslash \, P$. Since $\Gamma$ is an admissible dissection of the marked surface, each $\color{red} \bullet$-puncture lies in a component of the complement of $\Gamma$ in $S \, \backslash \, P$ whose boundary is composed solely of arcs in $\Gamma$ and their endpoints (Proposition 1.12 in \cite{amiot2019complete}). It follows from the fact that $(\Gamma, p)$ is a bipartite admissible dissection that all arcs enclosing a component with a $\color{red} \bullet$-puncture need to belong to the same part.
For each $\color{red} \bullet$-puncture $t \in P_{\color{red} \bullet}$, we choose a $\color{dark-green}{\circ}$-endpoint of an $\color{dark-green}{\circ}$-arc in $\Gamma$ that lies on the boundary of the connected component $\mathcal{P}_t$ of the complement of $\Gamma$ in $S \backslash P$ pertaining to $t$, and we consider a $\color{dark-green}{\circ}$-infinite arc $\pi_t$ from this $\color{dark-green}{\circ}$-endpoint that infinitely wraps around the $\color{red} \bullet$-puncture $t$ in the counterclockwise direction and is contained within $\mathcal{P}_t$.
We now observe that every $\color{dark-green}{\circ}$-infinite arc on the marked surface is a concatenation of a $\color{dark-green}{\circ}$-arc and at most two arcs $\pi_s, \pi_t$ for $\color{red} \bullet$-punctures $t, s \in P$. This decomposition is induced simply by taking the complement of $\Gamma$ in $S \backslash P$.
Consider $\mathcal{L}'$ and $\mathcal{R}'$ where $\mathcal{L}'$ and $\mathcal{R}'$ are generated by $\mathcal{L}$ and $\mathcal{R}$ and $\pi_t$ such that $\mathcal{P}_t$ is enclosed by $\color{dark-green}{\circ}$-arcs in the first part and the second part, respectively. Our observation above can be combined with Theorem 2.1 in \cite{opper2018geometric} and Proposition \ref{PBipartiteDissectionDecomposition} to yield that the triangulated category generated by $\mathcal{L}'$ and $\mathcal{R}'$ contains all infinite string objects string objects of $\Db(\Lambda \textrm{-}\mathrm{mod})$. The triangulated category generated by $\mathcal{L}'$ and $\mathcal{R}'$ equals $\Db(\Lambda \textrm{-}\mathrm{mod})$ because it contains all the indecomposable objects thereof.
At this stage, we can extend a semiorthogonal decomposition of $\Kb(\Lambda \textrm{-}\mathrm{proj})$; finally, we will demonstrate this extension is unique.
Suppose that $\langle \tilde{\mathcal{L}}, \tilde{\mathcal{R}} \rangle$ is semiorthogonal decomposition of $\Db(\Lambda \textrm{-}\mathrm{mod})$ that extends $\langle \mathcal{L}, \mathcal{R} \rangle$. For each $\color{red} \bullet$-puncture $t \in P$, we consider the approximation $R_{(\pi_t, d_t)} \to P_{(\pi_t, d_t)} \to L_{(\pi_t, d_t)} \to R_{(\pi_t, d_t)}[1]$ of $P_{(\pi_t, d_t)}$ for some grading $d_t$ that exists by Proposition \ref{PSemiorthogonalDecomposition}. Our aim is to show that this approximation is trivial, and $P_{(\pi_t, d_t)} = L_{(\pi_t, d_t)}$ or $P_{(\pi_t, d_t)} = R_{(\pi_t, d_t)}$ based on whether the border of $\mathcal{P}_t$ consists of $\color{dark-green}{\circ}$-arcs in $\Gamma$ in the first part or the second part, respectively. This will assure that $\langle \tilde{\mathcal{L}}, \tilde{\mathcal{R}} \rangle = \langle \mathcal{L}', \mathcal{R}' \rangle$.
Let us have a $\color{red} \bullet$-puncture $t \in P$ and suppose that all the objects corresponding to arcs on the boundary of $\mathcal{P}_t$ belong to $\mathcal{L}$. To prove that $P_{(\pi_t, d_t)} = L_{(\pi_t, d_t)}$, it is enough to show that there is no non-trivial morphism $f\colon R \to P_{(\pi_t, d_t)}$ for any $R \in \mathcal{R}$. Without loss of generality, we may assume that $R$ is indecomposable and that $R$ is isomorphic to $P_{(\sigma, h)}$, for $\sigma$, either a $\color{dark-green}{\circ}$-arc, an $\color{dark-green}{\circ}$-infinite arc, or an arc representing an infinite string complex, with some grading $h$, or $P_{(\sigma, h, M)}$, for a closed arc $\sigma$ representing a one-dimensional band complex with grading $h$ and $M$ the associated isomorphism class of $k[x]$-modules (cf. Theorem \ref{TIndecomposablesStringsBands} and Theorem 2.12 in \cite{opper2018geometric}).
By Theorem 3.3 in \cite{opper2018geometric}, the map $f$ corresponds to an intersection of $\pi_t$ and $\sigma$ or a common endpoint provided that $P_{(\sigma, h)} \in \Kb(\Lambda \textrm{-}\mathrm{proj})$. If $\pi_t$ and $\sigma$ shared a $\color{dark-green}{\circ}$-endpoint inducing a map $P_{(\sigma, h)} \to P_{(\pi_t, d_t)}$, it would mean that $\sigma$ would precede $\pi_t$ in the counter-clockwise order around it. However, as $\pi_t$ is succeeded in the counter-clockwise order around its $\color{dark-green}{\circ}$-endpoint by a $\color{dark-green}{\circ}$-arc $\gamma \in \Gamma$, part of boundary of $\mathcal{P}_t$. This would yield an existence of a map $P_{\sigma, h} \to P_{(\gamma, c)}$ for some grading $c$ (Remark 3.8 in \cite{opper2018geometric}), a contradiction given that $P_{(\gamma, c)} \in \tilde{\mathcal{L}}$.
So any potential map must come from an intersection of $\sigma$ and $\pi_t$ in the interior of the underlying marked surface, and, up to homotopy, $\sigma$ needs to be contained in $\mathcal{P}_t$; otherwise, $\sigma$ would have an intersection with a $\color{dark-green}{\circ}$-arc $\gamma \in \Gamma$ on the boundary of $\mathcal{P}_t$, which would imply an existence of a map $R \to P_{(\gamma, c)}$ for some grading $c$, a contradiction given that $P_{(\gamma, c)} \in \tilde{\mathcal{L}}$.
We can assume that $\sigma$ lies in $\mathcal{P}_t$. If $\sigma$ is a $\color{dark-green}{\circ}$-arc, this guarantees that it has a $\color{dark-green}{\circ}$-endpoint on the boundary of $\mathcal{P}_t$. However, there needs to be an arc $\gamma$ on the boundary of $\mathcal{P}_t$ that follows $\sigma$ in the counter-clockwise order around the $\color{dark-green}{\circ}$-endpoint implying an existence of a map $R \to P_{(\gamma, c)}$ for some grading $c$, a contradiction as $P_{(\gamma, c)} \in \tilde{\mathcal{L}}$. If $\sigma$ is a closed arc, it is, up to homotopy, a loop that winds around $t$ a given number of times as $\mathcal{P}_t$ is homotopy equivalent to a circle whose fundamental group is $\mathbb{Z}$. Such a closed arc cannot be equipped with a grading to yield an object of $\Db(\Lambda \textrm{-}\mathrm{mod})$ as each time a loop winds around $t$ the grading decreases or increases by more than $1$ (Definition 2.10 and Remark 2.11 in \cite{opper2018geometric}):
\begin{figure}[H]
\centering
\begin{tikzpicture}[x = 1cm, y = 1cm]
\pgfmathsetmacro{5}{4}
\pgfmathsetmacro{\sideshalf}{5/2}
\node[regular polygon, regular polygon sides = 4, minimum size = 3cm] (Pol) at (0,0) {};
\foreach \x in {1,...,5}
{\node[coordinate, name = {PolNode\x}] at (Pol.corner \x) {};}
\foreach \x in {1,...,\sideshalf}
{\pgfmathtruncatemacro{\xstart}{2*\x - 1}
\pgfmathtruncatemacro{\xend}{2*\x}
\draw[dark-green] ({PolNode\xstart}) -- ({PolNode\xend}) node[coordinate, midway, name = {PolNodeAux\xstart}] {} ;}
\foreach \x in {1,...,\sideshalf}
{\pgfmathtruncatemacro{\xstart}{2*\x}
\pgfmathtruncatemacro{\xend}{mod(2*\x + 1, 5)}
\draw[dark-green, dotted] ({PolNode\xstart}) -- ({PolNode\xend}) node[coordinate, near start, name = {PolNodeAuxRight\xstart}] {} node[coordinate, near end, name = {PolNodeAuxLeft\xstart}] {};}
\foreach \x in {1,...,5}
{\draw[dark-green, fill = white] ({PolNode\x}) circle (0.1);}
\node[coordinate, name = Puncture] at (0,0) {};
\draw[red, fill = red] (Puncture) circle (0.1);
\pgfmathsetmacro{\redlinefactor}{1.5}
\foreach \x in {1,...,\sideshalf}
{\pgfmathtruncatemacro{\xstart}{2*\x - 1}
\draw[red] (Puncture) -- ($\redlinefactor*({PolNodeAux\xstart})$) node[coordinate, near start, name = {GradingAux\xstart}] {};}
\foreach \x in {1,...,\sideshalf}
{\pgfmathtruncatemacro{\xstart}{2*\x}
\draw[red] (Puncture) -- ($\redlinefactor*({PolNodeAuxLeft\xstart})$) node[coordinate, near start, name = {GradingAuxLeft\xstart}] {};}
\foreach \x in {1,...,\sideshalf}
{\pgfmathtruncatemacro{\xstart}{2*\x}
\draw[red] (Puncture) -- ($\redlinefactor*({PolNodeAuxRight\xstart})$) node[coordinate, near start, name = {GradingAuxRight\xstart}] {};}
\pgfmathsetmacro{\gradingfactor}{1.7}
\foreach \x in {1,...,\sideshalf}
{\pgfmathtruncatemacro{\xstartone}{2*\x}
\pgfmathtruncatemacro{\xstarttwo}{mod(2*\x + 1, 5)}
\path ({GradingAuxLeft\xstartone}) edge[->, bend right, shorten >= 0.05cm, shorten <= 0.05cm] node[midway, name = {LabelAuxLeft\xstartone}] {} ({GradingAux\xstarttwo});
\path (Puncture) -- ($\gradingfactor*({LabelAuxLeft\xstartone})$) node[at end] {$-1$};}
\foreach \x in {1,...,\sideshalf}
{\pgfmathtruncatemacro{\xstart}{2*\x}
\path ({GradingAuxRight\xstart}) edge[->, bend right, shorten >= 0.05cm, shorten <= 0.05cm] node[midway, name = {LabelAuxRight\xstart}] {} ({GradingAuxLeft\xstart});
\path (Puncture) -- ($\gradingfactor*({LabelAuxRight\xstart})$) node[at end] {$-?$};}
\foreach \x in {1,...,\sideshalf}
{\pgfmathtruncatemacro{\xstartone}{2*\x - 1}
\pgfmathtruncatemacro{\xstarttwo}{2*\x}
\path ({GradingAux\xstartone}) edge[->, bend right, shorten >= 0.05cm, shorten <= 0.05cm] node[midway, name = {LabelAux\xstartone}] {} ({GradingAuxRight\xstarttwo});
\path (Puncture) -- ($\gradingfactor*({LabelAux\xstartone})$) node[at end] {$-1$};}
\end{tikzpicture}
\label{FGradingAroundPucnture}
\caption{Changes in grading for graded curves in $\mathcal{Q}_t$}
\end{figure}
This can be observed by taking the admissible dissection $\Delta$ of the underlying marked surface that encodes the structure of $\Lambda$. The $\color{red} \bullet$-point puncture $t$ is enclosed in a connected component $\mathcal{Q}_t$ of the complement of $\Delta$ in $S \backslash P$ by $\color{dark-green}{\circ}$ $\delta_1, \dots, \delta_n \in \Delta$ that follow each other in the counter-clockwise order around $t$. In the dual admissible $\Delta^*$ (cf. Proposition \ref{PAdmissibleDissectionProperties}), there is a $\color{red} \bullet$-arc $\delta^*_i$ for each $\delta_i$ such that crossing two following dual $\color{red} \bullet$-point arcs in the counter-clockwise order decreases the grading by 1 and vice versa by Definition \ref{DGradedCurves}. A loop that winds around $t$ in $\mathcal{P}_t$ can be assumed to lie in $\mathcal{Q}_t$ as well. So every time, this loop winds around its grading would increase or decrease by $n \geq 1$; such a loop cannot represent a one-dimensional band object as it is not possible to grade it properly.
Lastly, $\sigma$ may be a $\color{dark-green}{\circ}$-infinite arc. By our discussion above, $\sigma$ needs to be contained in $\mathcal{P}_t$, otherwise crossing a boundary $\color{dark-green}{\circ}$-arc of $\mathcal{P}_t$ yielding a morphism from $\tilde{\mathcal{R}}$ to $\tilde{\mathcal{L}}$. In this instance, however, it has have a $\color{dark-green}{\circ}$-endpoint on the boundary of $\mathcal{P}_t$ and a boundary $\color{dark-green}{\circ}$-arc of $\mathcal{P}_t$ following it in the counter-clockwise order around the $\color{dark-green}{\circ}$-enpoint, which is also not permitted by our discussion above.
The other case, when the boundary of $\mathcal{P}_t$ is formed by $\color{dark-green}{\circ}$-arcs in $\Gamma$ that give rise to objects in $\mathcal{R}$, is discussed dually.
\end{proof}
\begin{theorem}\label{TDecompositionsGoodCuts2}
There is a one-to-one correspondence between semiorthogonal decompositions of $\Db(\Lambda\textrm{-}\mathrm{mod})$ and good cuts of the marked surface associated to $\Lambda$.
\end{theorem}
\begin{proof}
This a simple combination of Theorem \ref{TSemiorthogonalDecompositionRestriction} and Theorem \ref{TDecompositionsGoodCuts}.
\end{proof}
\subsection{Semiorthogonal decompositions of \texorpdfstring{$\Db(\Lambda \textrm{-}\mathrm{mod})$}{Db(Lambda-mod)} with multiple terms}
We conclude this section by a discussion on how the results on characterization of two-term semiorthogonal decompositons in Theorem \ref{TDecompositionsGoodCuts2} extends to semiorthogonal decomposition of $\Db(\Lambda \textrm{-}\mathrm{mod})$ with more than two terms.
\begin{definition}\label{DProperSequenceCuts}
Let $(S, M, P)$ be marked surface. A sequence of good cuts $(\Omega_1, \dots, \Omega_n)$, is called a \textit{proper} if the good cuts are pair-wise different up to homotopy and if for every $1 \leq i \leq n-1$:
\begin{enumerate}[(i)]
\item each $\omega \in \Omega_{i+1}$ that lies on a connected component of the complement of $\Omega_{i}$ in $S \backslash P$ with right added marked points is homotopic to a component of the boundary thereof;
\item each $\omega \in \Omega_{i}$ that lies on a connected component of the complement of $\Omega_{i_+1}$ in $S \backslash P$ with left added marked points is homotopic to a component of the boundary thereof.
\end{enumerate}
\end{definition}
\begin{proposition}\label{PProperSequenceDecomposition}
Let $(S, M, P)$ be marked surface, and let $(\Omega_1, \Omega_2)$ be a proper sequence of good cuts. Denote $\langle \mathcal{L}_i, \mathcal{R}_i \rangle$ the semiorthogonal decomposition of $\Db(\Lambda \textrm{-}\mathrm{mod})$ corresponding to $\Omega_i$ by Theorem \ref{TDecompositionsGoodCuts2}, for $i = 1, 2$. Then, $\mathcal{L}_1 \subsetneq \mathcal{L}_2$.
\end{proposition}
\begin{proof}
Consider a $\color{dark-green}{\circ}$-arc $\gamma$ on $(S, M, P)$ whose corresponding objects in $\Db(\Lambda \textrm{-}\mathrm{mod})$ lie in $\mathcal{L}_1$ of the semiorthogonal decomposition $\langle \mathcal{L}_1, \mathcal{R}_1 \rangle$ thereof. The $\color{dark-green}{\circ}$-arc $\gamma$ is contained within a connected component $\mathcal{P}$ of the complement of $\Omega_1$ in $S \backslash P$. By definition, the component $\mathcal{P}$ would be with right added marked points. Pick $\omega \in \Omega_1$ that lies on the boundary of $\mathcal{P}$ and a $\color{dark-green}{\circ}$-arc $\gamma'$ such that: it begins in $g_\omega$, the $\color{dark-green}{\circ}$-endpoint of $\omega$, and it is contained in $\mathcal{P}$. Therefore, $\gamma'$ has to precede $\omega$ in the counter-clockwise order around $g_\omega$.
Because $(\Omega_1, \Omega_2)$ is proper, the connected component $\mathcal{P}$ does not contain any $\omega \in \Omega_2$ not homotopic to a segment of its boundary. For this reason, neither $\gamma$ nor $\gamma'$ cross an arc in $\Omega_2$, and the corresponding objects of $\Db(\Lambda \textrm{-}\mathrm{mod})$ need to both lie either in $\mathcal{L}_{\Omega_2}$ or in $\mathcal{R}_{\Omega_2}$ by Lemma \ref{LCrossingDividingArcs} combined with Proposition \ref{PDissectionGoodCut}.
Furthermore, as $(\Omega_1, \Omega_2)$ is proper, we may assume that without loss of generality $\omega$ lies within a connected component $\mathcal{Q}$ of the complement of $\Omega_2$ in $S \backslash P$ with right added marked points or on a boundary thereof. If $\omega$ lies on $Q$, then so does $\gamma'$, which does not cross any arc in $\Omega_2$ and $\gamma'$ precedes $\omega$ in the counter-clockwise order around $g_\omega$. Hence, the objects of $\Db(\Lambda \textrm{-}\mathrm{mod})$ corresponding to $\gamma'$ lie in $\mathcal{L}_{\Omega_2}$; it is also the case for the objects of $\Db(\Lambda \textrm{-}\mathrm{mod})$ corresponding to $\gamma$, so $\mathcal{L}_1 \subseteq \mathcal{L}_2$.
The fact that $\mathcal{L}_1 \subsetneq \mathcal{L}_2$ follows from the assumption that $\Omega_1$ and $\Omega_2$ are not equal up to homotopy. Theorem \ref{TDecompositionsGoodCuts2} then assures that the corresponding semiorthogonal decompositions differ as well.
\end{proof}
\begin{theorem}\label{TDecompositionsProperSequences}
There is a one-to-one correspondence between semiorthogonal decompositions of $\Db(\Lambda\textrm{-}\mathrm{mod})$ with arbitrarily many terms and proper sequences of good cuts of the marked surface associated to $\Lambda$.
\end{theorem}
\begin{proof}
At first, we note that it is a straight-forward consequence of Proposition \ref{PSemiorthogonalDecomposition} that a semiorthogonal decomposition $\langle \mathcal{C}_1, \dots, \mathcal{C}_n \rangle$ of $\mathcal{T}$ uniquely corresponds to a sequence of two-term semirothogonal decompositions $\langle \mathcal{L}_i, \mathcal{R}_i \rangle$ of $\mathcal{T}$, for $1 \leq i \leq n-1$, such that $\mathcal{L}_i$ is the full triangulated subcategory of $\mathcal{T}$ generated by $\mathcal{C}_1, \dots, \mathcal{C}_i$ and $\mathcal{R}_i$ is the full triangulated subcategory of $\mathcal{T}$ generated by $\mathcal{C}_{i+1}, \dots, \mathcal{C}_n$, and vice versa.
Combining Theorem \ref{TDecompositionsGoodCuts2} with Proposition \ref{PProperSequenceDecomposition}, we obtain that there is a one-to-one correspondence with proper sequence of good cuts of the marked surface $(S, M, P)$ associated to $\Lambda$ and a sequence $\langle \mathcal{L}_i, \mathcal{R}_i \rangle$ of semiorthogonal decompositions of $\Db(\Lambda \textrm{-}\mathrm{mod})$ with $\mathcal{L}_i \subsetneq \mathcal{L}_{i+1}$ for each $i$ smaller than the length of the sequence. This correspondence is carried forward to a one-to-one correspondence between proper sequences of good cuts by our observation above.
\end{proof}
\bibliographystyle{plain}
| 3f9d56753a35b6fd2249d8da0da64c52d66673e7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{Introduction}
Recently, Federated Learning (FL) is known as a novel distributed learning methodology for enhancing communication efficiency and ensuring privacy in traditional centralized one \cite{2017-FL-FederatedLearning}. However, the most challenge of this method for client models is non-independent and identically distributed (non-IID) data, which leads to divergence into unknown directions. Inspired by this, various works on handling non-IID were proposed in
\cite{2020-FL-FedProx,2021-FL-FedDyne,2021-FL-FedU,2020-FL-Scaffold,2021-FL-FedNova,2021-FL-FedGen,2022-FL-Wasserstein}. However, these works mainly rely on arbitrary configurations without thoroughly understanding the models' behaviors, yielding low-efficiency results. Aiming to fulfil this gap, in this work, we propose a new hierarchical FL framework using information theory by taking a deeper observation of the model's behaviors, and this framework can be realized for various FL systems with heterogeneous data. In addition, our proposed framework can trigger the FL system to be more scalable, controllable, and accessible through hierarchical architecture. Historically, anytime a new segment (i.e., a new group of clients) is integrated into the FL network, the entire network must be retrained from the beginning. Nevertheless, with the assistance of LKD, the knowledge is progressively transferred during the training process without information loss owing to the empirical gradients towards the newly participated clients' dataset.
The main contributions of the paper are summarized as follows. \textbf{(1)} We show that conventional FLs performance is unstable in heterogeneous environments due to non-IID and unbalanced data by carefully analyzing the basics of Stochastic Gradient Descent (SGD). \textbf{(2)} We propose a new multi-teacher distillation model, Label-Driven Knowledge Distillation (LKD), where teachers can only share the most certain of their knowledge. In this way, the student model can absorb the most meaningful information from each teacher. \textbf{(3)} To trigger the scalability and robustness against non-IID data in FL, we propose a new hierarchical FL framework, subbed Full-stack Federated Learning (F2L). Moreover, to guarantee the computation cost at the global server, F2L architecture integrates both techniques: LKD and FedAvg aggregators at the global server. To this end, our framework can do robust training by LKD when the FL process is divergent (i.e., at the start of the training process). When the training starts to achieve stable convergence, FedAvg is utilized to reduce the server's computational cost while retaining the FL performance. \textbf{(4)} We theoretically investigate our LKD technique to make a brief comparison in terms of performance with the conventional Multi-teacher knowledge distillation (MTKD), and in-theory show that our new technique always achieves better performance than MTKD. \textbf{(5)} We validate the practicability of the proposed LKD and F2L via various experiments based on different datasets and network settings. To show the efficiency of F2L in dealing with non-IID and unbalanced data, we provide a performance comparison and the results show that the proposed F2L architecture outperforms the existing FL methods. Especially, our approach achieves comparable accuracy when compared with FedAvg (\cite{2017-FL-FederatedLearning}) and higher $7-20\%$ in non-IID settings.
\section{Related Work}
\label{sec:related-works}
\subsection{Federated Learning on non-IID data}
\label{sec:fl-non-iid}
To narrow the effects of divergence weights, some recent studies focused on gradient regularization aspects \cite{2020-FL-FedProx,2021-FL-FedDyne,2021-FL-FedU,2020-FL-Scaffold,2021-FL-FedNova,2021-FL-FedGen,2022-FL-Wasserstein}. By using the same conceptual regularization, the authors in \cite{2020-FL-FedProx,2021-FL-FedDyne}, and \cite{2021-FL-FedU} introduced the FedProx, FedDyne, and FedU, respectively, where FedProx and FedDyne focused on pulling clients' models back to the nearest aggregation model while FedU's attempted to pull distributed clients together. To direct the updated routing of the client model close to the ideal server route, the authors in \cite{2020-FL-Scaffold} proposed {SCAFFOLD} by adding a control variate to the model updates. Meanwhile, to prevent the aggregated model from following highly biased models, the authors in \cite{2021-FL-FedNova} rolled out FedNova by adding gradient scaling terms to the model update function. Similar to \cite{2021-FL-FedU}, the authors in \cite{2022-FL-Wasserstein} launched the WALF by applying Wasserstein metric to reduce the distances between local and global data distributions. However, all these methods are limited in providing technical characteristics. For example, \cite{2021-FL-FedNova} demonstrated that FedProx and FedDyne are ineffective in many cases when using pullback to the globally aggregated model. Meanwhile, FedU and WAFL have the same limitation on making a huge communication burden. Aside from that, FedU also faces a very complex and non-convex optimization problem.
Regarding the aspect of knowledge distillation for FL, only the work in \cite{2021-FL-FedGen} proposed a new generative model of local users as an alternative data augmentation technique for FL. However, the majority drawback of this model is that the training process at the server demands a huge data collection from all users, leading to ineffective communication.
Motivated by this, we propose a new FL architecture that is expected to be more elegant, easier to implement, and much more straightforward. Unlike \cite{2021-FL-FedU, 2021-FL-FedDyne,2020-FL-Scaffold}, we utilize the knowledge from clients' models to extract good knowledge for the aggregation model instead of using model parameters to reduce the physical distance between distributed models. Following that, our proposed framework can flexibly handle weight distance and probability distance in an efficient way, i.e., $\Vert p^k(y=c)-p(y=c) \Vert$ (please refer to Appendix~\ref{appendix:SGDissues}).
\subsection{Multi-Teacher Knowledge Distillation}
\label{sec:multi-teacher-knowledge-distillation}
MTKD is an improved version of KD (which is presented in Appendix~\ref{appendix:knowledge-distillation}), in which multiple teachers work cooperatively to build a student model. As shown in \cite{2018-TL-KD-OnFly-NativeEnsemble}, every MTKD technique solves the following problem formulation:
\begin{align}
\textbf{P1}:\min \mathcal{L}_m^\textit{KL}
= \sum^R_{r=1} \sum^C_{l=1}\hat{p}(l|\boldsymbol{X},\boldsymbol{\omega}^r,T)\log{\frac{\hat{p}(l|\boldsymbol{X},\boldsymbol{\omega^r},T)}{\hat{p}(l|\boldsymbol{X},\boldsymbol{\omega}^g,T)}},
\label{eq:general-mtkd}
\end{align}
here, $r\in \{R\}$ are the teachers' indices. By minimizing \textbf{P1}, the student $\hat{p}^g$ can attain knowledge from all teachers. However, when using MTKD, there are some problems in extracting the knowledge distillation from multiple teachers. In particular, the process of distilling knowledge in MTKD is typically empirical without understanding the teacher's knowledge (i.e., aggregating all KL divergences between each teacher and the student). Therefore, MTKD is unable to exploit teachers' detailed predictions for the KD (e.g., \cite{2021-TL-MultiTeacher-MultiLevel}, \cite{2019-DL-EnsembleKD}, \cite{2018-TL-KD-OnFly-NativeEnsemble}, \cite{2017-DL-EnsembleTeachers}, \cite{2020-TL-HydraKD}). Another version of MTKD, KTMDs can only apply for a better teachers to distill knowledge (e.g., \cite{2019-DL-Customize-Student-HeterogenousTeachers}, \cite{2021-DL-Student-Customized-KD}, \cite{2022-DL-ConfidenceAware-KD}, \cite{2021-DL-DenselyGuidedKD}). For example, as provided in \cite[eq.~6]{2019-DL-Customize-Student-HeterogenousTeachers}, the student only selects the best teacher to operate the knowledge distillation. Visually, this technique is the same as the way of selecting a teacher among a set of teachers to carry out a single teacher distillation. Therefore, the student's performance is always bounded by the best teacher's performance. Another popular research direction in MTKD is to leverage the advantage of the gap between teachers' hidden class features. However, owing to the lack of explanatory knowledge in teachers' hidden layers, the method in \cite{2021-DL-Student-Customized-KD} cannot obtain better student performance when compared to their teachers. Generally, current MTKD techniques cannot extract good knowledge from different customer models, leading to weight divergence in FL.
\section{Full-stack Federated Learning}
\label{sec:IV-ProposedAlgorithm}
\subsection{The F2L framework}
\label{sec:IV-C-MultiteacherKD}
The main objective of our work is to design a hierarchical FL framework, in which a global server manages a set of distinct regional servers. Utilizing hierarchical FL, our proposed algorithm can achieve computation and computation efficiency. The reason is that Hierarchical FL makes the clients to train sectionally before making the global aggregation \cite{2020-FL-Hierarchical, 2020-FL-HierarchicalClustering}. Consequently, FL inherits every advantage from mobile edge intelligence concept over traditional non-hierarchical networks (e.g., communication efficiency, scalability, controlability) \cite{2020-5G-Survey,2016-IoT-Survey,2020-FLMEC-Survey}. At the end of each knowledge-sharing episode, the regions (which are supervised by regional servers) cooperate and share their knowledge (each region functions as a distinguished FL system, with a set amount of communication rounds per episode).
In each episode, each region randomly selects a group of clients from the network to carry out the aggregation process (e.g., FedAvg, FedProx); therefore, each region functions as a small-scale FL network. As a result, there are always biases in label-driven performance by applying random sampling on users per episode (see Appendix~\ref{appendix:proof-on-sampling-and-data-distribution}).
Given the random sampling technique applied to the regional client set, the regions always have different regional data distributions.
Consequently, various label-driven performances of different teachers might be achieved.
At the global server, our goal is to extract good performance from regional teachers while maintaining the salient features (e.g., understanding of the regional data distributions) of all regions.
As a result, we can capture useful knowledge from diverse regions in each episode using our proposed innovative knowledge distillation technique (which is briefly demonstrated in Section~\ref{sec:IV-D-LabelDrivenKD}). We train the model on the standard dataset on the central server to extract knowledge from multiple teachers into the global student model. The preset data pool on the server $\mathcal{S}$ is used to verify the model-class reliability and generate pseudo labels.
The system model is illustrated in Fig.~\ref{fig:hierarchicalFL}, and thoroughly described in Appendix~\ref{sec:system-model}.
\begin{figure*}[t]
\centering
\includegraphics[width = 0.8\linewidth]{imageLib/LKD-FullStack-FL-2.pdf}
\caption{The architecture of our F2L framework.}
\label{fig:hierarchicalFL}
\end{figure*}
The pseudo algorithm for F2L is demonstrated in Algorithm~\ref{alg:F2L}. When the FL process suffers from client-drift \cite{2020-FL-Scaffold} (i.e., the distribution of label-driven accuracies of different regions have large variance), the F2L framework applies LKD to reduce the class-wise performance gaps between regions (i.e., the regions with better performance on a particular class share their knowledge to regions with low performance). As a result, the FL network achieves a significantly faster convergence when utilizing LKD (which is intensively explained in Section~\ref{sec:IV-D-LabelDrivenKD}.) for the global aggregation process. When the generalization gap between regions is considerably reduced (i.e., $\Vert\max_r{\beta^c_r} - \min_r{\beta^c_r}\Vert \leq \epsilon$), our F2L network becomes vanilla FL to reduce computation and communication costs. To this end, our method can achieve computation efficiency while showing robustness to the non-IID data in the network. Additionally, whenever a new set of clients are added into the network and makes positive contributions to the FL system (e.g., $\Vert\max_r{\beta^c_r} - \min_r{\beta^c_r}\Vert \geq \epsilon$ where $\Vert\max_r{\beta^c_r}\Vert$ a corresponding to the new region's performance) the LKD aggregator can be switched back to improve the FL system's performance over again.
\subsection{Label-driven Knowledge Distillation}
\label{sec:IV-D-LabelDrivenKD}
To extract knowledge from multiple teachers to the global student model, we train the model on the standard dataset on the central server. The preset data pool on the server $\mathcal{S}$ is used to verify the model-class reliability and generate pseudo labels. In our work, the MTKD process executes two main tasks: (1) extracting the teachers' knowledge and (2) maintaining the students' previous performance.
To comprehend the LKD method, we first revisit the conventional MTKD, where the probabilistic output is calculated by model $\boldsymbol{\omega}$ on $x_i$, the prediction label $c$ is $\hat{p}(l|x_i,\boldsymbol{\omega},T,c)$ and its relation is:
\begin{align}\label{LKD-surrogateOutput}
\hat{p}(l|x_i,\boldsymbol{\omega},T,c) =
\begin{cases}
\hat{p}(l|x_i,\boldsymbol{\omega},T), & \text{if } \text{argmax}\left[ \hat{p}(l|x_i,\boldsymbol{\omega},T)\right] = c, \\
0, & \text{otherwise}.
\end{cases}
\end{align}
On the one hand, we aim to transfer knowledge from different regional models to the global one. Inspired by \cite{2015-DL-KnowledgeDistillation}, we use the Kullback–Leibler (KL) divergence between each regional teacher and the global logits as a method to estimate the difference between two models' performance. The relationship is expressed as follows:
\begin{align}\label{LKD-TeaStu-Loss-Weighted}
\mathcal{L}_r^\textit{KL} = \sum^C_{c=1} &\beta^c_r \sum^{S^r_c}_{i=1} \sum^C_{l=1} \hat{p}^r(l|x_i,\boldsymbol{\omega}^r,T,c)
\times\log{\frac{\hat{p}^r(l|x_i,\boldsymbol{\omega}^r,T,c)}{\hat{p}^g(l|x_i,\boldsymbol{\omega}^g,T,c)}},
\end{align}
where $S$ is the number of samples of the fixed dataset $\mathcal{S}$ on the server. $(\boldsymbol{X}^r_\text{alg}, \boldsymbol{Y}^r_\text{alg})$ is the dataset which is pseudo labeled and aligned by regional model $r$ and $(\boldsymbol{X}^r_\text{alg}[c], \boldsymbol{Y}^r_\text{alg}[c])$ represents the set of data with size of $S_c^r$ labeled by the model $r$ as $c$. Although the same preset dataset is utilized on every teacher model, the different pseudo labeling judgments from different teachers lead to the different dataset tuples. The process of identifying $S_c^r$ is demonstrated in Algorithm~\ref{alg:L-SampleAlign}. Because the regional models label on the same dataset $S$, we have $\sum_{c=1}^C S^r_c = S$ for all regional models. $D_\textit{KL}^c(\hat{p}^r||\hat{p}^g)$ is the $c$ label-driven KL divergence between model $r$ and model $g$.
On the other hand, we aim to guarantee that the updated global model does not forget the crucial characteristics of the old global model. Hence, to measure the divergence between the old and the updated model, we introduce the following equation:
\begin{align}\label{LKD-OldNew-Loss-Weighted}
\mathcal{L}_{\boldsymbol{\omega}_\textit{upd}}^\textit{KL}
= \sum^C_{c=1} &\beta^c_{\boldsymbol{\omega}_\textit{old}} \sum^{S^r_c}_{i=1} \sum^C_{l=1} \hat{p}^g(l|x_i,\boldsymbol{\omega}^g_\textit{old},T,c)
\times\log{\frac{\hat{p}^g(l|x_i,\boldsymbol{\omega}^g_\textit{old},T,c)}{\hat{p}^g(l|x_i,\boldsymbol{\omega}^g_\textit{new},T,c)}},
\end{align}
where $\boldsymbol{\omega}_\textit{old}$ is the old parameters set of the global model which is distilled in the last episode of F2L. More details about the label-driven knowledge distillation are discussed in Appendix~\ref{appendix:labeldriven-knowledgedistillation}.
To compare the performance between LKD and MTKD, we consider the following assumption and lemmas:
\begin{lemma}
Given $\tau^c_r$ is the $c$-label driven predicting accuracy on model $r$. Let $\sigma^2_{r,c}, \mu_{r,c}$ be the model's variance and mean, respectively. The optimal value of variance and mean on student model (i) $\sigma^{*2}_{LKD,g,c}, \mu^*_{LKD,g,c}$ yields $\sigma^{*2}_{\textrm{LKD},g,c} = \frac{1}{\sum_{r=1}^{R} e^{\tau^c_r}} \sum^R_{r=1} e^{\tau_r^c}\sigma^2_{r,c}$, and $\mu^{*}_{\textrm{LKD},g,c} = \frac{1}{\sum_{r=1}^{R} e^{\tau^c_r}} \sum^R_{r=1} e^{\tau_r^c}\mu_{r,c}.$.
\label{lemma:optimal-LKD}
\end{lemma}
\textit{Proof:} The proof is provided in Appendix~\ref{appendix:lemma-1}.
\begin{assumption}
Without loss of generality, we consider $R$ distinct regional models whose accuracy satisfy the following prerequisites $\sigma^2_{1,c} \leq \sigma^2_{2,c} \leq \ldots \leq \sigma^2_{R,c}$, and $|\mu_{1,c} - \Bar{\mu}_c| \leq |\mu_{2,c} - \Bar{\mu}_c| \leq \ldots \leq |\mu_{R,c} - \Bar{\mu}_c|$ ($\Bar{\mu}_c$ is denoted as an empirical global mean of the dataset on class $c$).
\end{assumption}
\begin{lemma}
Given the set of models with variance satisfy $\sigma^2_{1,c} \leq \sigma^2_{2,c} \leq \ldots \leq \sigma^2_{R,c}$, the models' accuracy have the following relationship $\tau^c_1 \geq \tau^c_2 \geq \ldots \geq \tau^c_R$.
\label{lemma:relationship-accuracy-variance}
\end{lemma}
\textit{Proof.} The proof can be found in Appendix~\ref{appendix:relationship-accuracy-variance}.
\begin{theorem}
Let $\sigma^{*2}_{\textrm{LKD},g,c} $ be the class-wise variance of the student model, and $\sigma^{*2}_{\textrm{MTKD},g,c}$ be the class-wise variance of the model of teacher $r$, respectively. We always have the student's variance using LKD technique always lower than that using MTKD:
\begin{align}
\sigma^{*2}_{\textrm{LKD},g,c}
\leq
\sigma^{*2}_{\textrm{MTKD},g,c}.
\label{eq:LKD-vs-MTKD}
\end{align}
\label{theorem:labeldriven-knowledgedistillation-analysis}
\end{theorem}
\textit{Proof}: For the complete proof see Appendix \ref{appendix:labeldriven-knowledgedistillation-analysis}.
\begin{theorem}
Let $\mu^{*}_{\textrm{LKD},g,c} $ be the empirical $c$-class-wise mean of the student model, and $\mu^{*}_{\textrm{MTKD},g,c}$ be the empirical $c$-class-wise mean of the model of teacher $r$, respectively. We always have the student's empirical mean using LKD technique always closer to the empirical global dataset's class-wise mean ($\Bar{\mu}_c$) than that using MTKD:
\begin{align}
|\mu^{*}_{\textrm{LKD},g,c} - \Bar{\mu}_c|
\leq
|\mu^{*}_{\textrm{MTKD},g,c} - \Bar{\mu}_c|.
\label{eq:LKD-vs-MTKD-mean}
\end{align}
\label{theorem:labeldriven-knowledgedistillation-analysis-mean}
\end{theorem}
Given Theorems~\ref{theorem:labeldriven-knowledgedistillation-analysis} and \ref{theorem:labeldriven-knowledgedistillation-analysis-mean}, we can prove that our proposed LKD technique can consistently achieve better performance than that of the conventional MTKD technique. Moreover, by choosing the appropriate LKD allocation weights, we can further improve the LKD performance over MTKD.
Due to space limitation, we defer the proof to Appendix \ref{appendix:labeldriven-knowledgedistillation-analysis-mean}.
\subsection{Class Reliability Scoring}
The main idea of class reliability variables $\beta^c_r$, $\beta^c_{\boldsymbol{\omega}_\textit{old}}$ in LKD is to weigh the critical intensity of the specific model. Therefore, we leverage the attention design from \cite{2017-DL-Attention} to improve the performance analysis of teachers' label-driven.
For regional models with disequilibrium or non-IID data, the teachers only teach the predictions relying upon their specialization. The prediction's reliability can be estimated by leveraging the validation dataset on the server and using the function under the curve (AUC) as follows:
\begin{equation}\label{LKD-teacher_weight}
\begin{split}
\beta^c_r = \frac{\text{exp}(f_\textit{AUC}^{c,r} T_{\boldsymbol{\omega}})}{\sum^R_{r=1}\text{exp}(f_\textit{AUC}^{c,r} T_{\boldsymbol{\omega}} )},
\end{split}
\end{equation}
where $f^{c,r}_\text{AUC}$ denotes the AUC function on classifier $c$ of the regional model $r$. Since AUC provides the certainty that a specific classifier can work on a label over the rest, we use the surrogate softmax function to weigh the co-reliability among the same labeling classifiers on different teacher models. For simplicity, we denote $\beta^c_{\boldsymbol{\omega}_\textit{old}}$ as the AUC on each labeling classifier:
\begin{equation}\label{LKD-oldModel_weight}
\begin{split}
\beta^c_{\boldsymbol{\omega}_\textit{old}} = \frac{\text{exp}(f_\textit{AUC}^{c,\boldsymbol{\omega}_\text{old}} T_{\boldsymbol{\omega}})}{\text{exp}(f_\textit{AUC}^{c,\boldsymbol{\omega}_\text{new}} T_{\boldsymbol{\omega}})+\text{exp}(f_\textit{AUC}^{c,\boldsymbol{\omega}_\text{old}} T_{\boldsymbol{\omega}} )}.
\end{split}
\end{equation}
In the model update class reliability, instead of calculating the co-reliability between teachers, \eqref{LKD-oldModel_weight} compares the performance of the previous and current global models. Moreover, we introduce a temperatured value for the class reliability scoring function, denoted as $T_{\boldsymbol{\omega}}$. By applying a large temperatured value, the class reliability variable sets $\beta^c_r$, and $\beta^c_{\boldsymbol{\omega}_\textit{old}}$ make a higher priority on the better performance (i.e., the label-driven performance on class $c$ from teacher $r$, e.g., $f_\textit{AUC}^{c,r}$ in equation~\eqref{LKD-teacher_weight} or class $c$ from old model $\boldsymbol{\omega}_\text{old}$ in equation~\eqref{LKD-oldModel_weight}). By this way, we can preserve the useful knowledge which is likely ignored in the new distillation episode. The more detailed descriptions of class reliability scoring are demonstrated in Algorithm~\ref{alg:C-Reliability}.
\subsection{Joint Multi-teacher Distillation for F2L}
We obtain the overall loss function for online distillation training by the proposed F2L:
\begin{align}
\mathcal{L}_\text{F2L} = \lambda_1 \sum^R_{r=1}\mathcal{L}^\textit{KL}_r + \lambda_2\mathcal{L}^{KL}_{\boldsymbol{\omega}_\textit{upd}} + \lambda_3\mathcal{L}^g_\textit{CE},
\end{align}
where $\lambda_1, \lambda_2, \lambda_3$ are the scaling coefficients of the three terms in the joint loss function. The first and second terms imply the joint LKD from the regional teacher models and the updating correction step, respectively. Moreover, to ensure that knowledge the student receives from teachers is accurate and can be predicted accurately in practice, we need to validate the quality of the student model on the real data. Thus, we also compute the “standard” loss between the student and the ground-truth labels of the train dataset. This part is also known as the hard estimator, which is different from the aforementioned soft-distillator. The hard loss equation is as follows:
\begin{align}
\mathcal{L}^g_\textit{CE} &= H(y,\hat{p}(l|\boldsymbol{X},\boldsymbol{\omega}^g,T))
= \sum^C_{l=1} y_l \log{\hat{p}(l|\boldsymbol{X},\boldsymbol{\omega}^g,T)}.
\end{align}
\begin{algorithm}[h]
\caption{F2L framework}
\label{alg:F2L}
\begin{algorithmic}
\State {\bfseries Require:} Initialize clients' weights, global aggregation round, number of regions $R$, arbitrary $\epsilon$.
\While{not converge}
\For{all regions $r\in \{1,2,\dots,R\}$}
\For{all user in regions}
\State Apply FedAvg on regions $r$.
\EndFor
\State Send regional model $\boldsymbol{\omega}^r$ to the global server.
\EndFor
\If{reach global aggregation round}
\If{$\Vert\max_r{\beta^c_r} - \min_r{\beta^c_r}\Vert \geq \epsilon$ where $\beta = \{\beta^1_r,\dots,\beta^C_r\}\vert^R_{r=1}$ from Algorithm~\ref{alg:C-Reliability}}
\State Apply LKD as described in Algorithm~\ref{alg:LD-KD}
\Else
\State $\boldsymbol{\omega}^g = 1/R \sum^R_{r=1} \boldsymbol{\omega}^r$.
\EndIf
\EndIf
\EndWhile
\end{algorithmic}
\end{algorithm}
We use the temperature coefficient $T = 1$ to calculate the class probability for this hard loss. The overall training algorithm for LKD is illustrated in Algorithm~\ref{alg:LD-KD}. In terms of value selection for scaling coefficients, the old global model can be considered as an additional regional teacher's model in the same manner, in theory. Therefore, $\lambda_2$ should be chosen as:
\begin{align}\label{LKD-lambda2-choice}
\lambda_2 =
\begin{cases}
\frac{1}{R}\lambda_1, & \text{if update distillation in \eqref{LKD-OldNew-Loss-Weighted} is considered,} \\
0, & \text{otherwise},
\end{cases}
\end{align}
where $R$ is the number of regions decided by our hierarchical FL settings. With respect to $\lambda_3$, the value is always set as:
\begin{align}\label{LKD-lambda1-choice}
\lambda_3 =
\begin{cases}
1 - \frac{R+1}{R}\lambda_1, & \text{if update distillation in \eqref{LKD-OldNew-Loss-Weighted} is considered,} \\
1 - \lambda_1, & \text{otherwise}.
\end{cases}
\end{align}
\subsection{Discussions on the extent of protecting privacy}
In its simplest version, our proposed F2L framework, like the majority of existing FL approaches, necessitates the exchange of models between the server and each client, which may result in privacy leakage due to, for example, memorization present in the models.
Several existing protection methods can be added to our system in order to safeguard clients against enemies. These include adding differential privacy \cite{2017-FL-DifferentiallyPrivate} to client models and executing hierarchical and decentralized model fusion by synchronizing locally inferred logits, for example on random public data, as in work \cite{2019-FL-Cronus}. We reserve further research on this topic for the future.
\section{Experimental Evaluation}
\label{sec:V-ExperimentalEvaluation}
\subsection{Comparison with FL methods}
\label{sec:V-B-Comparison}
\begin{table*}[t]
\addtolength{\tabcolsep}{-3pt}
\centering
\caption{The top-1 test accuracy of different baselines on different data settings. The $\alpha$ indicates the non-IID degree of the dataset (the lower value of $\alpha$ means that the data is more heterogeneous).\\}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
Dataset & FedAvg & FedGen & FedProx & Fed- & F2L & FedAvg & FedGen & FedProx & Fed- & F2L \\
& & & & Distill & (Ours) & & & & Distill & (Ours) \\
\hline \hline
\cline{1-11}
& \multicolumn{5}{|c|}{Dirichlet ($\alpha=1$)} & \multicolumn{5}{|c|}{Dirichlet ($\alpha=0.1$)} \\
\cline{1-11}
EMNIST & $71.66$& $78.70$& $70.77$ & $75.56$& $\textbf{81.14}$ & $59.10$ & $68.24$& $58.88$ & $46.03$ & $\textbf{68.31}$ \\
CIFAR-10 & $60.48$& $59.21$& $63.72$ & $62.36$& $\mathbf{71.22}$ & $47.07$& $47.08$& $47.05$ & $45.67$ & $\mathbf{55.22}$ \\
CIFAR-100 & $36.17$& $40.26$ & $36.3$ & $34.88$ & $\mathbf{50.33}$ & $21.31$& $28.96$& $20.43$ & $16.15$ & $\mathbf{31.07}$ \\
CINIC-10 & $65.23$& $71.61$& $65.15$ & $67.77$ & $\mathbf{74.85}$ & $47.55$& $52.35$& $48.2$ & $47.1$ & $\mathbf{57.12}$\\
CelebA & $70.82$& $75.43$& $71.07$ & $68.59$ & $\mathbf{81.65}$ & $63.58$& $70.14$& $66.33$ & $62.91$ & $\mathbf{74.14}$\\
\hline
\end{tabular}
\label{tab:F2L-Comparisons}
\end{table*}
We run the baselines (see Section~\ref{sec:V-A-ExperimentSetup}) and compare with our F2L. Then, we evaluate the comparisons under different non-IID ratio. More precisely, we generate the IID data and non-IID data with two different Dirichlet balance ratio: $\alpha = \{1, 10\}$. The comparison results are presented in Table~\ref{tab:F2L-Comparisons}. As shown in Table~\ref{tab:F2L-Comparisons}, the F2L can outperform the four baselines with a significant increase in accuracy. The reason for this phenomenon is that the LKD technique selectively draws the good features from regional models to build a global model. Hence, the global model predicts the better result on each different class and the entire accuracy of the global model then increases tremendously. The significant impact when applying LKD to distill different teachers to one student is shown in Table~\ref{tab:LKD-Distillation-Efficiency}.
\subsection{Computation Efficiency of F2L}
\label{sec:V-C-F2L-CommCompEfficiency}
To evaluate the computation efficiency of our proposed F2L process, we compare our F2L process with 3 benchmarks: (i) F2L-noFedAvg (aggregator only consists of LKD), (ii) vanilla FL (FL with flatten architecture and FedAvg as an aggregator), and (iii) flatten LKD (FL with flatten architecture based with LKD as an alternate aggregator). Fig.~\ref{fig:F2L-performance} shows that the F2L system can achieve convergence as good as the F2L-noFedAvg. The reason is that: after several communication rounds, the distributional distance between regions is reduced thanks to the LKD technique. Hence, the efficiency of the LKD technique on the data is decreased. As a consequence, the LKD technique shows no significant robustness over FedAvg aggregator. In the non-hierarchical settings, the flatten LKD and FedAvg reveal under-performed compared to the proposed hierarchical settings. We assume that the underperformance above comes from the data shortage of clients' training models. To be more detailed, the clients' dataset are considerably smaller than that of the \say{regional dataset}. Thus, the regional models contain more information than the clients' models. We believe that: in the LKD technique, teachers' models require a certain amount of knowledge to properly train a good student (i.e., the global model).
\begin{figure}[!h]
\centering
\subfigure[\label{fig:F2L-performance}]{\includegraphics[width=0.3\linewidth]{imageLib/F2L-performance.pdf}}
\subfigure[\label{fig:F2L-compcost}]{\includegraphics[width=0.25\linewidth]{imageLib/F2Lcompcost.pdf}}
\subfigure[\label{fig:F2L-scalability}]{\includegraphics[width=0.3\linewidth]{imageLib/F2L-scalability.pdf}}
\caption{Performance benchmarks of F2L under different settings. Fig.~\ref{fig:F2L-performance} reveals the convergence. Fig.~\ref{fig:F2L-compcost} shows the computational cost, and Fig.~\ref{fig:F2L-scalability} demonstrates the F2L convergence when a new set of clients are added into the FL system (i.e., at communication round $100$).}
\label{fig:F2Ltest}
\end{figure}
Given the convergence rate from Fig.~\ref{fig:F2L-performance} and the computation cost at the server on Fig.~\ref{fig:F2L-compcost}, we can see that, by using the adaptive switch between LKD and FedAvg in F2L, we can achieve significant computational efficiency at the aggregation server. Note that F2L can substantially increase performance and computation efficiency compared with non-hierarchical architecture.
\subsection{Scalability}
\label{sec:V-C-F2L-Scalability}
This section evaluates the F2L scalability. To do so, we inject a group of clients with non-IID data into our FL system after $100$ rounds (when the convergence becomes stable). Note that the FL system has never trained these data. The detailed configurations of our experimental assessments can be found in Appendix~\ref{sec:V-A-ExperimentSetup}. As it can be seen from Fig.~\ref{fig:F2L-scalability}, when a new group of clients are added to the FL system, the vanilla FL shows a significant drop in terms of convergence. The reason is because of the distribution gap between the global model's knowledge and knowledge of the clients' data. Whenever new data with unlearned distribution is added to a stable model, the model will make considerable gradient steps towards the new data distribution. Thus, the FedAvg takes considerable learning steps to become stable again. In contrast, in F2L system, the learning from newly injected regions does not directly affect the learning of the whole FL system. Instead, the knowledge from the new domains is selectively chosen via the LKD approach. Thus, the LKD process does not suffer from information loss when new clients with non-IID data are added to the FL system.
\subsection{LKD Analysis}
\label{sec:V-C-LKD-Analysis}
In this section, we evaluate the LKD under various settings to justify the capability of LKD to educate the good student from the normal teachers. Our evaluations are as follows.
\textbf{Can student outperform teachers?}
To verify the efficiency of LKD with respect to enhancing student performance, we first evaluate F2L on MNIST, EMNIST, CIFAR-$100$, CINIC-$10$, CelebA dataset. The regions are randomly sampled from the massive FL network. In this way, we only evaluate the performance of LKD on random teachers. Table~\ref{tab:LKD-Distillation-Efficiency} shows top-1 accuracy on the regional teacher and student models. The results reveal that LKD can significantly increase the global model performance compared with that of the regional models. Moreover, the newly distilled model can work well under each regional non-IID data after applying the model update.
\begin{table}[t]
\renewcommand{\arraystretch}{1.25}
\caption{Top-1 accuracy of F2L on $5$ datasets MNIST, EMNIST, CIFAR-100, , CINIC-$10$ and CelebA. The data's heterogeneity is set at $\alpha = 0.1$ on CIFAR-100, MNIST, CINIC-$10$ and CelebA. We use EMNIST ``unbalanced'' to evaluate in this test. The ``before update'' and ``after update'' denote the teacher models' accuracies before and after the global distillation, respectively.}
\centering
\small\addtolength{\tabcolsep}{-3pt}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}} \\ \end{tabular}} &
\multicolumn{2}{c|}{MNIST}
& \multicolumn{2}{c|}{EMNIST}
& \multicolumn{2}{c|}{CIFAR-100}
& \multicolumn{2}{c|}{CINIC-10}
& \multicolumn{2}{c|}{Celeb-A} \\ \cline{2-11}
& before & after & before & after & before & after & before & after & before & after \\ \hline
& update & update & update & update & update & update & update & update & update & update \\ \hline
Teacher 1 & $61.02$ & $\mathbf{95.19}$ & $73.27$ & $\mathbf{84.09}$ & $20.11$ & $\mathbf{35.41}$ & $43.8$ & $\mathbf{46.59}$ & $62.37$ & $\mathbf{67.98}$ \\ \hline
Teacher 2 & $92.49$ & $\mathbf{98.22}$ & $78.80$ & $\mathbf{83.62}$ & $18.82$ & $\mathbf{31.2}$ & $42.15$ & $\mathbf{46.01}$ & $63.79$ & $\mathbf{72.33}$ \\ \hline
Teacher 3 & $81.60$ & $\mathbf{97.63}$ & $80.5$ & $\mathbf{84.10}$ & $22.40$ & $\mathbf{34.93}$ & $40.02$ & $\mathbf{42.15}$ & $64.05$ & $\mathbf{69.44}$ \\ \hline
G-student & $\mathbf{98.71}$ & & $\mathbf{84.11}$ & & $\mathbf{37.68}$ & & $\mathbf{47.65}$ & & $\mathbf{70.12}$ & \\ \hline
\end{tabular}
\label{tab:LKD-Distillation-Efficiency}
\end{table}
\newline
To make a better visualization for the LKD's performance, we reveal the result of LKD on EMNIST dataset in terms of confusion matrix as in Fig.~\ref{fig:ConfusionMatrix}. As it can be seen from the figure, the true predictions is represented by a diagonals of the matrices. A LKD performance is assumed to be well predicted when the value on diagonals is high (i.e., the diagonals' colors is darker), and the off-diagonals is low (i.e., the off-diagonals' colors are lighter). As we can see from the four figures, there are a significant reduce in the off-diagonals' darkness in the student performance (i.e., Fig.~\ref{fig:student-later-distill-evaluation}) comparing to the results in other teachers (i.e., Figures~\ref{fig:teacher1-prior-distill-evaluation}, \ref{fig:teacher2-prior-distill-evaluation}, and \ref{fig:teacher3-prior-distill-evaluation}). Therefore, we can conclude that our proposed MTKD techniques can surpass the teachers' performance as we mentioned in Section~\ref{sec:related-works}.
\begin{figure}[!h]
\centering
\subfigure[\label{fig:teacher1-prior-distill-evaluation}]{\includegraphics[width=0.235\linewidth]{imageLib/Teacher1-preUpdate.pdf}}
\subfigure[\label{fig:teacher2-prior-distill-evaluation}]{\includegraphics[width=0.235\linewidth]{imageLib/Teacher2-preUpdate.pdf}}
\subfigure[\label{fig:teacher3-prior-distill-evaluation}]{\includegraphics[width=0.235\linewidth]{imageLib/Teacher3-preUpdate.pdf}}
\subfigure[\label{fig:student-later-distill-evaluation}]{\includegraphics[width=0.235\linewidth]{imageLib/Student-postUpdate.pdf}}
\caption{The illustrative results of LKD on EMNIST dataset. Confusion matrices show the effectiveness of joint distillation on regional models. Figures (a), (b), and (c) are the confusion matrix before distillation of teacher's predictions in region $1$, $2$, and $3$, respectively (see Appendix~\ref{sec:V-A-ExperimentSetup}). Fig.~(d) is the confusion matrix of predictions after distillation of student. The matrix diagonal demonstrates the true-predicted label of the model.}
\label{fig:ConfusionMatrix}
\end{figure}
\textbf{Teachers can really educate student?}
We evaluate LKD under different soft-loss coefficients $\lambda_1$ while the hard-loss factor is set at $\lambda_3 = 1-\lambda_1$ (the scaling value $\lambda_2$ is set to $0$). Thus, we can justify whether the robust performance of LKD comes from the joint distillation from teachers or just the exploitation of data-on-server training. We evaluate LKD on six scaling values $\lambda_1 = \{0, 0.001, 0.01, 0.1, 0.5, 1\}$. We evaluate on three dataset, including EMNIST, CIFAR-10, and CIFAR-100, and summarize the results in Tables~\ref{tab:LKD-DistillationTest-EMNIST}, \ref{tab:LKD-DistillationTest-CIFAR10} and \ref{tab:LKD-DistillationTest-CIFAR100} in \textbf{Appendices}. We can see from the three tables that the LKD cap off with $\lambda_3=0.01$. Especially, when $\lambda_3=1$ (which means the LKD acts as a vanilla cross-entropy optimizer), the model accuracy reduces notably. This means that the LKD only uses hard-loss as a backing force to aid the distillation. Thus, our LKD is appropriate and technically implemented.
\textbf{Required training sample size for joint distillation.} To justify the ability of LKD under a shortage of training data, we evaluate LKD with six different data-on-server settings: $\sigma = \{1, 1/2, 1/4, 1/6, 1/8, 1/10\}$, where $\sigma$ is the sample ratio when compared with the original data-on-server as demonstrated in Table~\ref{tab:data-configuration}. As we can see from the implementation results in three Tables~\ref{tab:LKD-SampleTrainTest-EMNIST}, \ref{tab:LKD-SampleTrainTest-CIFAR10}, and \ref{tab:LKD-SampleTrainTest-CIFAR100} in \textbf{Appendices}, the F2L is demonstrated to perform well under a relatively small data-on-server. To be more specific, we only need the data-on-server to be $4$ times lower than the average data-on-client to achieve a robust performance compared with the vanilla FedAvg. However, we suggest using the data-on-server to be larger than the data from distributed clients to earn the highest performance for LKD. Moreover, due to the ability to work under unlabeled data, the data-on-server does not need to be all labeled. We only need a small amount of labeled data to aid the hard-loss optimizer. Thus, the distillation data on the server can be updated from distributed clients gradually.
\section*{Broader Impact and Limitation}
Due to the hierarchical framework of our proposed F2L, each sub-region acts like an independent FL process. Therefore, our F2L is integrable with other current methods, which means that we can apply varying FL techniques (e.g., FedProx, FedDyne, FedNova, HCFL \cite{2022-FL-HCFL}) into distinct sub-regions to enhance the overall F2L framework. Therefore, architecture search (e.g., which FL technique is suitable for distinct sub FL region) for the entire hierarchical network is essential for our proposed framework, which is the potential research for the future work. Moreover, the hierarchical framework remains unearthed. Therefore, a potentially huge amount of research directions is expected to be investigated (e.g., resource allocation \cite{2022-FL-ResourceAllocation-Compression, 2022-FL-StragglingPrivacy, 2021-FL-CodedFL,2021-FL-LargeScale}, and task offloading in hierarchical FL \cite{2021-FL-EnergyEfficiency}). However, our LKD technique still lacks of understanding about the teachers' models (e.g., how classification boundaries on each layer impact on the entire teachers' performance). By investigating in explainable AI, along with layer-wise performance, we can enhance the LKD, along with reducing the unlabeled data requirements for the distillation process in the future work.
\section{Conclusion}
In this research, we have proposed an FL technique that enables knowledge distillation to extract the-good-feature-only from clients to the global model. Our model is capable of tackling the FL's heterogeneity efficiently. Moreover, experimental evaluations have revealed that our F2L model outperforms all of the state-of-the-art FL baselines in recent years.
| 41dec9aeeeb7f51d864817b0201a192f1e0bc2d4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $q$ be a positive integer and $a\in(\mathbb{Z}/q\mathbb{Z})^*$. We define $p(q,a)$ as the least prime occurring in the arithmetic progression $\{qn+a, n\in\mathbb{N}\}$. Under the Riemann Hypothesis for the $L$-Dirichlet series of characters $\chi$ mod $q$, it can be proved that
\vspace{-2mm}
$$p(q,a)\ll(\phi(q)\log q)^2,$$
\vspace{2mm}
\noindent
where $\phi$ denotes the Euler totient function. Unconditionally, using the Siegel-Walfisz Theorem, one can show that
\vspace{-2mm}
$$p(q,a)\ll_{\varepsilon}\exp(p^{\varepsilon})$$
\vspace{2mm}
\noindent
for any $\varepsilon>0$. Moreover, if there are no Siegel zeros, this bound reduces to
\vspace{-1mm}
$$p(q,a)\ll q^{c\log q}$$
\vspace{2mm}
\noindent
for some $c>0$. So, in 1944, it came as a big surprise when Linnik \cite{lin1, lin2} proved that there exist positive universal constants $C$ and $L$ such that
\vspace{-2mm}
$$p(q,a)\leq Cq^L$$
\vspace{2mm}
\noindent
for any choice of $q\in\mathbb{N}$ and $a\in(\mathbb{Z}/q\mathbb{Z})^*$. One of the steps that lead Linnik to this breakthrough, now known as Linnik's theorem, was the following estimate.
Assume that $x\geq q^2$. If there exists an exceptional character $\chi$ corresponding to a Siegel zero, then
\vspace{-2mm}
\begin{equation}\label{linnest}
\sum_{\substack{n\leq x\\n\equiv a\mod{q}}}\Lambda(n)=\frac{x}{\phi(q)}+\frac{\chi(a)}{\phi(q)}\sum_{n\leq x}\Lambda(n)\chi(n)+O\bigg(\frac{x^{1-c_1/\log(2q)}}{\phi(q)}+\frac{xe^{-c_2\sqrt{\log x}}}{\phi(q)}\bigg),
\end{equation}
\noindent
where $c_1$ and $c_2$ are two positive absolute constants. If the exceptional character $\chi$ does not exist, then (\ref{linnest}) remains true, but the character sum of the right-hand side is omited.
Linnik's work on $p(q,a)$ was later simplified, as was done by Bombieri in \cite{bomb}, but the new proofs, including Linnik's original proof, relied in one form or another on the following three ingredients.
\vspace{2mm}
\begin{itemize}
\item The classical zero-free region;
\item A log-free zero-density estimate;
\item The exceptional zero repulsion, also known, as the Deuring-Heilbronn phenomenon, stating that it is possible to enlarge the classical zero-free region when it contains a Siegel zero.
\end{itemize}
\vspace{2mm}
\noindent
Proofs that make use of these three principles can be found in modern treatments, like the one which is presented in \cite[Chapter 18]{iwko}.
Nonetheless, in the last years, some proofs that avoid these tools have appeared. For example, in 2002, such a proof was developed by Elliot in \cite{el}. Later, in 2016, Granville, Harper and Soundararjan \cite{ghal} studied pretentiously the distribution of multiplicative functions on arithmetic progressions and as a consequence of their general results, they were able to show a weaker form of (\ref{linnest}). In turn, this served as the stepping stone for another new proof of Linnik's theorem which circumvented the combination of the three aforementioned principles. Another pretentious proof of Linnik's theorem is presented in \cite[Chapter 27]{dimb} and a basic element of the proof is a flexible variant of (\ref{linnest}) where every prime is weighted with $1/p$ instead of $\log p$. Even though the alternative approaches recover Linnik's theorem, they do not provide a pretentious proof of the same quantitative strength as (\ref{linnest}).
To this end, in this paper, we apply pretentious methods and prove Linnik's estimate (\ref{linnest}) with a refined error term where a Korobov-Vinogradov type term replaces $\exp\{-c_2\sqrt{\log x}\}$. In particular, we prove the following theorem.
\begin{thm}\label{main}
Let $q\geq 1$ be an integer and consider a real number $x\geq q^2$. For any character $\chi$ mod $q$, we set
$$L_q(1,\chi)=L(1,\chi)\prod_{p\leq q}(1-\chi(p)/p).$$
\vspace{1mm}
\noindent
We also define $\mathcal{R}_q$ as the set of real, non-principal characters mod $q$ and we take a character $\psi$ such that $L_q(1,\psi)=\min_{\chi\in\mathcal{R}_q}L_q(1,\chi)$. Then, for any $a\in(\mathbb{Z}/q\mathbb{Z})^*$, we have that
\begin{align*}
\sum_{\substack{n\leq x\\n\equiv a\mod{q}}}\Lambda(n)=\frac{x}{\phi(q)}+\frac{\psi(a)}{\phi(q)}\sum_{n\leq x}\psi(n)\Lambda(n)+O\bigg(\frac{x^{1-C_1/\log(2q)}}{\phi(q)}+\frac{xe^{-C_2(\log x)^{3/5}(\log \log x)^{-3/5}}}{\phi(q)}\bigg),
\end{align*}
\vspace{3mm}
\noindent
where $C_1$ and $C_2$ are two positive absolute constants.
\end{thm}
We might bound the sum $\sum_{n\leq x}\psi(n)\Lambda(n)$ and simplify Theorem \ref{main}. Indeed, by referring to Theorem 1.6(a) of \cite{oldk}, it follows that there exist positive constants $c'$ and $c''$ such that
$$\sum_{n\leq x}\psi(n)\Lambda(n)\ll x^{1-c'L_q(1,\psi)/\log(2q)}+xe^{-c''\sqrt{\log x}},$$
\noindent
for all $x\geq q^2$. The same theorem provides also information about the size of the quantity $L_q(1,\psi)$ that is involved in the bound above. It is known that there exists a constant $\delta\in(0,1)$ such that $L(\cdot,\psi)$ has at most one zero $\beta$ in $[1-\delta/\log q,1)$. If such a zero does not exist, we set $\beta=1-\delta/\log q$. In either case, Theorem 1.6(a) of \cite{oldk} claims that $L_q(1,\psi)\asymp (1-\beta)\log q$. Therefore, we arrive at the following consequence of Theorem \ref{main}.
\begin{cor}
Under the considerations and assumptions of Theorem \ref{main}, we have that
$$\sum_{\substack{n\leq x\\n\equiv a\mod{q}}}\Lambda(n)=\frac{x}{\phi(q)}+O\bigg(\frac{x^{1-\alpha_1/\log(2q)}}{\phi(q)}+\frac{x^{1-\alpha_2L_q(1,\psi)/\log(2q)}}{\phi(q)}+\frac{xe^{-\alpha_3\sqrt{\log x}}}{\phi(q)}\bigg),$$
\vspace{1mm}
\noindent
for some positive contants $\alpha_1, \alpha_2$ and $\alpha_3$. Moreover, there exists a constant $\delta\in(0,1)$ such that $L(\cdot,\psi)$ has at most one zero $\beta$ in $[1-\delta/\log q,1)$. If such a zero does not exist, we put $\beta=1-\delta/\log q$ and in any case, $L_q(1,\psi)\asymp (1-\beta)\log q$.
\end{cor}
\vspace{1mm}
We prove Theorem \ref{main} in Section 5. Its proof borrows ideas from the pretentious large sieve that Granville, Harper and Soundararjan developed in \cite{ghal}. It is also inspired by techniques of Koukoulopoulos from his work on bounded multiplicative functions with small partial sums \cite{oldk}. These techniques were deployed recently in a similar fashion by the author in \cite{sach}, where he generalized the work of Koukoulopoulos for classes of divisor-bounded multiplicative functions.
The rest of the text is divided into four sections and the first three serve as preparatory material for the last one where we prove Theorem \ref{main}.
In subsequent work, we intend to extend the recent results of \cite{kousou} and \cite{sach} by applying the core methods of this paper to determine the structure of divisor-bounded multiplicative functions whose partial sums are small on arithmetic progressions.
\subsection*{Notation} Throughout the text, for an integer $n>1$, we denote its smallest prime factor by $P^-(n)$. For $n=1$, we define $P^-(1)=+\infty$. For $m\in\mathbb{N}$, the symbol $\tau_m$ will denote the $d$-fold divisor function given as $\tau_m(n)=\sum_{d_1\cdots d_m=n}1$ for all $n\in\mathbb{N}$.
\subsection*{Acknowledgements} The author would like to thank his advisor, Dimitris Koukoulopoulos, for his constant guidance and for his suggestions who lead to a strengthening of Theorem \ref{main}. He would also like to express his gratitude towards the Stavros Niarchos Foundation for all its generous financial support during the making of this work.
\vspace{1mm}
\section{Preliminary Sieving Results}
In this section we prove two auxiliary sieving results. For the end of their proofs, the following theorem of Shiu \cite{shiu} will be necessary. This theorem may be seen as an analogue of the Brun-Titchmarsch inequality for multiplicative functions.
\begin{thm}\label{shiu}
Fix $m\in\mathbb{N}$ and $\varepsilon>0$. Given any choice of $q\in\mathbb{N}, a\in(\mathbb{Z}/q\mathbb{Z})^*,$ real numbers $x\geq y\geq 1$ with $y/q\geq x^{\varepsilon}$ and a multiplicative function $f$ such that $0\leq f\leq \tau_m$, we have that
$$\sum_{\substack{x-y<n\leq x\\n\equiv a\mod{q}}}f(n)\ll_{m,\varepsilon}\frac{y}{q}\exp\bigg\{\sum_{\substack{p\leq x\\p\nmid q}}\frac{f(p)-1}{p}\bigg\}.$$
\end{thm}
\vspace{1mm}
\begin{lem}\label{chlem}
Given a $q\in\mathbb{N}$, let $\chi$ be a Dirichlet character modulo $q$. For $j\in\mathbb{N}\cup\{0\}$ and any real numbers $x\geq y\geq (10q)^{100}$, we have
$$\sum_{\substack{n\leq x\\P^-(n)>y}}\chi(n)(\log n)^j=\mathds{1}_{\chi=\chi_0}\cdot\bigg(\int_y^x(\log t)^j\mathrm{d} t\bigg)\prod_{p\leq y}\bigg(1-\frac{1}{p}\bigg)+O\bigg(\frac{(\log x)^jx^{1-\kappa/\log y}}{\log y}\bigg),$$
\noindent
where $\chi_0$ is the principal character modulo $q$ and $\kappa>0$ is an absolute constant.
\end{lem}
\begin{proof}
For any $w\geq y$, it is known \cite[Lemma 22.2, p.224]{dimb} that
\begin{equation}\label{chi1}
\sum_{\substack{n\leq w\\P^-(n)>y}}\chi(n)=\mathds{1}_{\chi=\chi_0}\cdot w\prod_{p\leq y}\bigg(1-\frac{1}{p}\bigg)+R_y(w),
\end{equation}
\vspace{2mm}
\noindent
where $R_y(w)\ll(\log y)^{-1}w^{1-\kappa/\log y}$ for some sufficiently small constant $\kappa>0$. Using the Riemann-Stieltjes integral, we have
\begin{align}\label{st}
\sum_{\substack{n\leq x\\P^-(n)>y}}\chi(n)(\log n)^j&=\int_y^x(\log t)^j\mathrm{d}\Big(\sum_{\substack{n\leq t\\P^-(n)>y}}\chi(n)\Big)\\
&=\mathds{1}_{\chi=\chi_0}\cdot\bigg(\int_y^x(\log t)^j\mathrm{d} t\bigg)\prod_{p\leq y}\bigg(1-\frac{1}{p}\bigg)+\int_y^x(\log t)^j\mathrm{d} R_y(t).\nonumber
\end{align}
\vspace{1mm}
\noindent
Now, if $x\geq y^A$ for some large $A>0$, a simple integration by parts implies that
\begin{align*}
\int_y^x(\log t)^j\mathrm{d} R_y(t)\ll \frac{(\log x)^jx^{1-\kappa/\log y}}{\log y}+\frac{j}{\log y}\int_y^x(\log t)^{j-1}t^{-\kappa/\log y}\mathrm{d} t.
\end{align*}
\vspace{2mm}
\noindent
But, since
\begin{align*}
j\int_y^x(\log t)^{j-1}t^{-\kappa/\log y}\mathrm{d} t\leq jx^{1-\kappa/\log y}\int_1^x\frac{(\log t)^{j-1}}{t}\mathrm{d} t=(\log x)^jx^{1-\kappa/\log y},
\end{align*}
\vspace{2mm}
\noindent
we deduce that
\vspace{-2mm}
\begin{align}\label{est}
\int_y^x(\log t)^j\mathrm{d} R_y(t)\ll \frac{(\log x)^jx^{1-\kappa/\log y}}{\log y}.
\end{align}
\noindent
We insert (\ref{est}) in (\ref{st}) and complete the proof of the lemma when $x\geq y^A$. It only remains to establish the lemma in the case $y\leq x<y^A$. In this case, $\log x/\log y\asymp 1$, and so applying Theorem \ref{shiu}, we have that
$$\bigg|\sum_{\substack{n\leq x\\P^-(n)>y}}\chi(n)(\log n)^j\bigg|\leq (\log x)^j\sum_{\substack{n\leq x\\P^-(n)>y}}1\ll \frac{x(\log x)^j}{\log y}\ll \frac{(\log x)^jx^{1-\kappa/\log y}}{\log y}.$$
\noindent
This means that the lemma does hold in the range $y\leq x<y^A$ as well and this finishes the proof.
\end{proof}
\vspace{1mm}
\begin{lem}\label{logap}
Let $j$ be a non negative integer. For any $q\in\mathbb{N}, a\in(\mathbb{Z}/q\mathbb{Z})^*$ and real numbers $x\geq y\geq 2q$, there exists a constant $\lambda>0$ such that
\begin{equation}\label{lem2.3}
\sum_{\substack{n\leq x,P^-(n)>y\\n\equiv a \mod{q}}}(\log n)^j=\frac{1}{\phi(q)}\bigg(\int_y^x(\log t)^j\mathrm{d} t\bigg)\prod_{p\leq y}\bigg(1-\frac{1}{p}\bigg)+O\bigg(\frac{(\log x)^jx^{1-\lambda/\log y}}{\phi(q)\log y}\bigg).
\end{equation}
\end{lem}
\begin{proof}
Let $d$ be a positive integer such that $(d,q)=1$. Because of the Chinese Remainder Theorem, the system of linear congruences $n\equiv a\mod{q}, n\equiv 0\mod{d}$ is equivalent to $n\equiv a^*\mod{qd}$ for some $a^*\in(\mathbb{Z}/(qd)\mathbb{Z})$. Since $\sum_{n\leq x,n\equiv a^*\mod{qd}}1=x/(qd)+O(1)$, partial summation implies that
\begin{align}\label{arel}
\sum_{\substack{n\leq x,d\mid n\\n\equiv a\mod{q}}}(\log n)^j=\sum_{\substack{n\leq x\\n\equiv a^*\mod{qd}}}(\log n)^j&=\frac{x}{qd}\bigg\{(\log x)^j-j\int_1^x(\log t)^{j-1}\mathrm{d} t\bigg\}+O((\log x)^j)\nonumber\\
&=\frac{x}{qd}\int_1^x(\log t)^j\mathrm{d} t+O((\log x)^j).
\end{align}
If $x\geq y^A$ for an appropriately large exponent $A>0$, then using (\ref{arel}) for a standard application of the fundamental lemma of sieve theory \cite[Theorem 18.11, p. 190]{dimb} with the set of integers $\mathcal{A}=\{n\leq x, n\equiv a\mod{q}\}$ and the set of primes $\mathcal{P}=\{p\leq y, p\nmid q\}$, we deduce that
\begin{align}\label{int}
\sum_{\substack{n\leq x,P^-(n)>y\\n\equiv a \mod{q}}}(\log n)^j=\frac{1}{\phi(q)}\bigg(\int_1^x(\log t)^j\mathrm{d} t\bigg)\prod_{p\leq y}\bigg(1-\frac{1}{p}\bigg)+O\bigg(\frac{(\log x)^jx^{1-\lambda/\log y}}{\phi(q)\log y}\bigg).
\end{align}
\vspace{2mm}
\noindent
But, $\prod_{p\leq y}(1-1/p)\ll (\log y)^{-1}$ and
\begin{align*}
\int_1^y(\log t)^j\mathrm{d} t<y(\log x)^j\ll (\log x)^jx^{1-\lambda/\log y}
\end{align*}
\vspace{2mm}
\noindent
for $x\geq y^A$ with $A$ being sufficiently large. Hence, relation (\ref{int}) yields (\ref{lem2.3}) whenever $x\geq y^A$. When $y\leq x<y^A$, a similar argument as the one that we developed at the end of the proof of Lemma \ref{chlem} shows that
$$\sum_{\substack{n\leq x,P^-(n)>y\\n\equiv a \mod{q}}}(\log n)^j\ll\frac{(\log x)^jx^{1-\lambda/\log y}}{\phi(q)\log y}.$$
\vspace{2mm}
\noindent
The lemma is now proven.
\end{proof}
\vspace{1mm}
\section{Existing Bounds for Sifted $L$-Dirichlet Series}
Let us consider a real number $y\geq 1$ and a positive integer $q$. For $s\in\mathbb{C}$ with $\Re(s)>1$, we define the $y$-rough Dirichlet series of a Dirichlet character $\chi$ modulo $q$ as
\begin{equation}\label{sifL}
L_y(s,\chi):=\sum_{\substack{n\geq 1\\P^-(n)>y}}\frac{\chi(n)}{n^s}=L(s,\chi)\prod_{p\leq y}(1-\chi(p)/p^s).
\end{equation}
\noindent
In (\ref{sifL}), the series is absolutely convergent and its rightmost side implies the meromorphical continuation of $L_y(\cdot,\chi)$ on the whole complex plane (with one pole at $1$ only in the case of the principal character $\chi_0$).
The present section constitutes a collection of some existing upper and lower bounds for the values of a $y$-rough Dirichlet series. We will use all these bounds in the proof of Theorem \ref{main}.
The theorems listed below are stated without their proofs, as they are already part of the literature. We only provide a reference for each one of them. To simplify their statements, we introduce the notation
$$V_t:=\exp\{100(\log (3+|t|))^{2/3}(\log\log(3+|t|))^{1/3}\}\quad (t\in\mathbb{R}).$$
\vspace{3mm}
We start with the following theorem \cite[Lemma 4.1]{ppnt} concerning an upper bound for the derivatives of $L(\cdot,\chi)$.
\begin{thm}\label{ub}
Let $q$ be a positive integer and $\chi$ be a non-principal character modulo $q$. Let also $j\in\mathbb{N}$ and $s=\sigma+it$ with $\sigma>1$ and $t\in\mathbb{R}$. For $y\geq qV_t$, we have that
$$|L_y^{(j)}(s,\chi)|\ll j!(C\log y)^j,$$
\vspace{2mm}
\noindent
where $C>0$ is an absolute constant.
\end{thm}
\vspace{1mm}
The next result \cite[Lemma 4.2]{ppnt} equips us with lower bounds for $L_y(s,\chi)$.
\begin{thm}\label{lb}
Fix a positive integer $q$ and let $\chi$ be a character modulo $q$. Let $s=\sigma+it$ with $\sigma>1$ and $t\in\mathbb{R}$ and consider the real number $y\geq qV_t$.
\begin{enumerate}[(a)]
\item If $\chi$ is not real, then $\abs{L_y(s,\chi)}\gg 1$.
\item If $\chi$ is real and non-principal, then $\abs{L_y(s,\chi)}\gg L_y(1,\chi).$
\end{enumerate}
\end{thm}
\vspace{1mm}
We close this section with a theorem \cite[Lemma 27.5, p. 291]{dimb} dealing with the size of $L_q(\sigma,\chi)$ for $\sigma\geq 1$ when $\chi\neq\psi$, where $\psi$ is defined as in the statement of Theorem \ref{main}.
\begin{thm}\label{ldnS}
Let $q$ be a positive integer and let $\mathcal{R}_q$ be the set of real, non-principal characters modulo $q$. If we take a character $\psi$ such that $L_q(1,\psi)=\min_{\chi\in\mathcal{R}_q}L_q(1,\chi)$ and $\mathcal{C}_q:=\{\chi\mod{q}:\chi\neq \chi_0, \psi\}$, then $|L_y(\sigma,\chi)|\asymp 1$ for all $\chi\in\mathcal{C}_q, y\geq q$ and $\sigma\geq 1$.
\end{thm}
\vspace{1mm}
\section{A Mean Value Theorem}
The goal of this section is to establish a mean value thoerem for the derivatives of the Dirichlet series
$$\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)}{n^s}.$$
\vspace{2mm}
\noindent
Such a result will be essential at the end of the proof of Theorem \ref{main}. To prove it, we are making use of the following lemma which is due to Montgomery \cite[Theorem 3, p. 131]{mon}.
\begin{lem}\label{mon}
Let $A(s)=\sum_{n\geq 1}a_nn^{-s}$ and $B(s)=\sum_{n\geq 1}b_nn^{-s}$ be two Dirichlet series which converge for $\Re(s)>1.$ If $|a_n|\leq b_n$ for all $n\in \mathbb{N}$, then
\begin{equation*}
\int_{-T}^T|A(\sigma+it)|^2\mathrm{d} t\leq 3\int_{-T}^T|B(\sigma+it)|^2\mathrm{d} t,
\end{equation*}
\vspace{3mm}
\noindent
for any $\sigma>1$ and any $T\geq 0$.
\end{lem}
\begin{lem}\label{mvt}
Let us consider an integer $q\geq 1$ and a real number $\varepsilon>0$. We further consider a real number $y\geq q^{1+\varepsilon}$. Then for $j\in\mathbb{N}\cup\{0\}, T\geq 1, 1<\sigma<2$ and $a\in(\mathbb{Z}/q\mathbb{Z})^*$, we have
$$\int_{|t|>T}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\frac{\mathrm{d} t}{t^2}\ll_{\varepsilon}\frac{(\log 4)^j(2j)!}{\phi(q)^2(\sigma-1)^{2j+1}T}.$$
\end{lem}
\begin{proof}
First, for $k\in\mathbb{N}\cup\{0\}$, because of Lemma \ref{mon}, we observe that
\begin{align*}
\int_{k-1/2}^{k+1/2}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\mathrm{d} t&=\int_{-1/2}^{1/2}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^jn^{-ik}}{n^{\sigma+it}}\bigg|^2\mathrm{d} t\\
&\leq 3\int_{-1/2}^{1/2}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\mathrm{d} t.\nonumber
\end{align*}
\noindent
Consequently,
\begin{align}\label{lines}
\int_{|t|>T}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\frac{\mathrm{d} t}{t^2}&\leq \sum_{|k|>T-1/2}\int_{k-1/2}^{k+1/2}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\frac{\mathrm{d} t}{t^2}\nonumber\\
&\leq 4\sum_{|k|>T/2}\frac{1}{k^2}\int_{k-1/2}^{k+1/2}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\mathrm{d} t\nonumber\\
&\ll\bigg(\sum_{k>T/2}\frac{1}{k^2}\bigg)\cdot \int_{-1/2}^{1/2}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\mathrm{d} t\nonumber\\
&\ll\frac{1}{T}\int_{-1/2}^{1/2}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\mathrm{d} t.
\end{align}
Now, we focus on estimating the integral at the last line of (\ref{lines}). We will do this by adopting the rather standard technique which is used for proving similar mean value theorems. We consider the function $\Phi:\mathbb{R}\rightarrow \mathbb{R}$ given by the formula $\Phi(x)=(2\pi\sin(x/4))^2x^{-2}$ for all $x\in\mathbb{R}^*$ and $\Phi(0)=\pi^2/4$. Notice that $\Phi(x)\geq 1$ on $[-1/2,1/2]$. So, if $\hat{\Phi}$ is the Fourier transform of $\Phi,$ then
\begin{align}\label{afw}
\int_{-1/2}^{1/2}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\mathrm{d} t&\leq \int_{\mathbb{R}}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\Phi(t)\mathrm{d} t\nonumber\\
&=\sum_{\substack{m\equiv a\mod{q}\\n\equiv a\mod{q}\\P^-(m),P^-(n)>y}}\frac{\Lambda(n)\Lambda(n)(\log m)^j(\log n)^j}{m^{\sigma}n^{\sigma}}\hat{\Phi}(\log(m/n)),
\end{align}
\noindent
where we arrived at the last line by expanding the square and by interchanging the order of summation and integration. The Fourier transform $\hat{\Phi}$ is an even function, because $\Phi$ is also even. Therefore, we may bound the last line of (\ref{afw}) by twice the sum
\begin{equation}\label{sum}
\sum_{\substack{m\equiv a\mod{q}\\n\leq m, n\equiv a\mod{q}\\P^-(m),P^-(n)>y}}\frac{\Lambda(n)\Lambda(n)(\log m)^j(\log n)^j}{m^{\sigma}n^{\sigma}}\hat{\Phi}(\log(m/n)).
\end{equation}
\noindent
The Fourier transform $\hat{\Phi}$ is continuous and compactly supported on $[-1/2,1/2]$. Moreover, for $n\leq m$, we have that $|n-m|\leq m\log(m/n)$. Therefore, the sum (\ref{sum}) is smaller than or equal to
\begin{align*}
\sum_{\substack{m\equiv a\mod{q}\\P^-(m)>y}}\frac{\Lambda(m)(\log m)^j}{m^{\sigma}}&\sum_{\substack{n\leq m, |n-m|\leq m/2\\n\equiv a\mod{q}}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma}}\\
&\ll\sum_{\substack{m\equiv a\mod{q}\\P^-(m)>y}}\frac{\Lambda(m)(\log m)^{2j}}{m^{2\sigma}}\sum_{\substack{m/2\leq n\leq m\\n\equiv a\mod{q}}}\Lambda(n)\\
&\ll_{\varepsilon}\frac{1}{\phi(q)}\sum_{\substack{m\equiv a\mod{q}\\P^-(m)>y}}\frac{\Lambda(m)(\log m)^{2j}}{m^{2\sigma-1}}.
\end{align*}
\noindent
For the last estimate we made use of the Brun-Titchmarsch theorem. Its application was allowed, because $\Lambda(1)=0$, which means that the condition $P^-(m)>y$ implies that $m>y\geq q^{1+\varepsilon}$. According to all the above, relation (\ref{afw}) becomes
\begin{equation}\label{1stbe}
\int_{-1/2}^{1/2}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^j}{n^{\sigma+it}}\bigg|^2\mathrm{d} t\ll_{\varepsilon}\frac{1}{\phi(q)}\sum_{\substack{m\equiv a\mod{q}\\P^-(m)>y}}\frac{\Lambda(m)(\log m)^{2j}}{m^{2\sigma-1}}.
\end{equation}
We continue by bounding the sum of the right-hand side of (\ref{1stbe}). By decomposing this sum in dyadic intervals, we get that
\begin{align*}
\sum_{\substack{m\equiv a\mod{q}\\P^-(m)>y}}\frac{\Lambda(m)(\log m)^{2j}}{m^{2\sigma-1}}&\leq (\log 4)^j\sum_{r>\log y/\log 2}\frac{r^{2j}}{2^{r(2\sigma-1)}}\sum_{\substack{2^r\leq m<2^{r+1}\\mathfrak{m}\equiv a\mod{q}}}\Lambda(m)\\
&\ll_{\varepsilon}\frac{(\log 4)^j}{\phi(q)}\sum_{r\geq 0}\frac{r^{2j}}{4^{r(\sigma-1)}},
\end{align*}
\noindent
where we applied the Brun-Titchmarsch theorem for the last step, since $2^r>y\geq q^{1+\varepsilon}$. Using the $2j$-th derivative of the geometric series, it follows that
\begin{equation*}
\sum_{r\geq 0}\frac{r^{2j}}{4^{r(\sigma-1)}}\leq \sum_{r\geq 0}\frac{(r+2j)!}{4^{r(\sigma-1)}r!}=\frac{(2j)!}{(1-4^{1-\sigma})^{2j+1}}\ll\frac{(2j)!}{(\sigma-1)^{2j+1}},
\end{equation*}
\vspace{2mm}
\noindent
as $1-4^{1-\sigma}\gg \sigma-1$ for $\sigma\in(1,2)$ by the mean value theorem. Thus,
\begin{equation}\label{ests}
\sum_{\substack{m\equiv a\mod{q}\\P^-(m)>y}}\frac{\Lambda(m)(\log m)^{2j}}{m^{2\sigma-1}}\ll_{\varepsilon}\frac{(\log 4)^j(2j)!}{\phi(q)(\sigma-1)^{2j+1}}.
\end{equation}
We now combine (\ref{lines}), (\ref{1stbe}) and (\ref{ests}) and conclude the proof of the lemma.
\end{proof}
\vspace{1mm}
\section{Proof of Theorem \ref{main}}
\subsection{A useful lemma} Before embarking on the proof of Theorem \ref{main}, which is the main objective of this section, we first state a lemma that will be useful.
\begin{lem}\label{lasl}
Let $k\in\mathbb{N}$, $D$ be an open set of $\mathbb{C}$, $s\in D$ and $F:D\rightarrow\mathbb{C}$ be a function which is differentiable $k$ times at $s$. We further assume that $F(s)\neq 0$ and we set
$$M=\max_{1\leq j\leq k}\bigg\{\frac{1}{j!}\bigg|\frac{F^{(j)}}{F}(s)\bigg|\bigg\}^{1/j}\quad and\quad N=\max_{1\leq j\leq k}\bigg\{\frac{1}{j!}\bigg|\Big(\frac{F^{'}}{F}\Big)^{(j-1)}(s)\bigg|\bigg\}^{1/j}.$$
\vspace{2mm}
\noindent
Then $M/2\leq N\leq 2M$.
\end{lem}
\begin{proof}
\cite[Lemma 9.1]{oldk}
\end{proof}
\subsection{Proof of Theorem \ref{main}:} We may assume that $q\geq 2$ and that $x\geq q^B$ for some sufficiently large real number $B>0$. Indeed, for $q=1$, it is the prime number theorem with the error term provided by Korobov and Vinogradov and if $q^2\leq x<q^B$, then $\log x/\log q\asymp 1$ and the theorem follows from a trivial application of the Brun-Titchmarsch inequality. Now, let $k$ be a positive integer and set $y=(10q)^{100}V_T$ with $T=\exp\{2L(\log x)^{3/5}(\log\log x)^{2/5}\}$ for a large constant $L>0$.
Since $\log =\Lambda\ast 1$, we have that
\begin{align}\label{1'}
\sum_{\substack{n\leq x\\n\equiv a \mod{q}\\P^-(n)>y}}\log n=\sum_{\substack{m\leq \sqrt{x}\\P^-(m)>y}}\sum_{\substack{\ell\leq \frac{x}{m}\\\ell\equiv a\overline{m}\mod{q} \\P^-(\ell)>y}}\Lambda(\ell)+\sum_{\substack{\ell\leq \sqrt{x}\\P^-(n)>y}}\Lambda(\ell)\sum_{\substack{\sqrt{x}<m\leq \frac{x}{\ell}\\mathfrak{m}\equiv a\overline{\ell}\mod{q}\\P^-(m)>y}}1.
\end{align}
\vspace{2mm}
\noindent
For $\chi\in\{\chi_0,\psi\}$, it is true that $\log\cdot\chi=(\Lambda\cdot\chi)\ast\chi$. Therefore, we similarly have that
\begin{align}\label{2'}
\sum_{\substack{n\leq x\\P^-(n)>y}}\chi(n)\log n=\sum_{\substack{m\leq \sqrt{x}\\P^-(m)>y}}\chi(m)\sum_{\substack{\ell\leq \frac{x}{m}\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)+\sum_{\substack{\ell\leq \sqrt{x}\\P^-(n)>y}}\Lambda(\ell)\chi(\ell)\sum_{\substack{\sqrt{x}<m\leq \frac{x}{\ell}\\P^-(m)>y}}\chi(m).
\end{align}
\vspace{2mm}
\noindent
Because of Lemmas \ref{chlem} and \ref{logap}, we have that
\begin{align*}
\sum_{\substack{n\leq x\\n\equiv a \mod{q}\\P^-(n)>y}}\log n-\frac{1}{\phi(q)}\sum_{\substack{n\leq x\\P^-(n)>y}}\chi_0(n)\log n-\frac{\psi(a)}{\phi(q)}\sum_{\substack{n\leq x\\P^-(n)>y}}\psi(n)\log n\ll\frac{x^{1-c_1/\log y}}{\phi(q)},
\end{align*}
\vspace{2mm}
\noindent
where $c_1=\min\{\kappa,\lambda\}/2$. Consequenlty, if we put
\begin{align*}
\Delta(u,y;q,b):=\sum_{\substack{\ell\leq u\\\ell\equiv b\mod{q}\\P^-(\ell)>y}}\Lambda(\ell)-\frac{1}{\phi(q)}\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi_0(\ell)-\frac{\psi(b)}{\phi(q)}\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\psi(\ell)
\end{align*}
\noindent
and
\begin{align*}
\Delta^*(u,y;q,b):=\sum_{\substack{m\leq u\\mathfrak{m}\equiv b\mod{q}\\P^-(m)>y}}1-\frac{1}{\phi(q)}\sum_{\substack{m\leq u\\P^-(m)>y}}\chi_0(m)-\frac{\psi(b)}{\phi(q)}\sum_{\substack{m\leq u\\P^-(m)>y}}\psi(m)
\end{align*}
\vspace{2mm}
\noindent
for simplicity, then (\ref{1'}) and (\ref{2'}) imply that
\begin{align*}
\sum_{\substack{m\leq\sqrt{x}\\P^-(m)>y}}\Delta\Big(\frac{x}{m},y;q,a\overline{m}\Big)&=-\sum_{\substack{\ell\leq\sqrt{x}\\P^-(\ell)>y}}\Lambda(\ell)\Delta^*\Big(\frac{x}{\ell},y;q,a\overline{\ell}\Big)\\
&+\sum_{\substack{\ell\leq\sqrt{x}\\P^-(\ell)>y}}\Lambda(\ell)\Delta^*(\sqrt{x},y;q,a\overline{\ell})+O\bigg(\frac{x^{1-c_1/\log y}}{\phi(q)}\bigg).
\end{align*}
\vspace{1mm}
\noindent
Since $x/\ell\geq\sqrt{x}\geq y$ for $\ell\leq \sqrt{x}$, we can apply Lemmas \ref{chlem} and \ref{logap} to bound $\Delta^*(x/\ell,y;q,a\overline{\ell})$ and $\Delta^*(\sqrt{x},y;q,a\overline{\ell})$. Doing so yields that
\begin{align*}
\sum_{\substack{\ell\leq\sqrt{x}\\P^-(\ell)>y}}\Lambda(\ell)\Big\{\Delta^*\Big(\frac{x}{\ell},y;q,a\overline{\ell}\Big)-\Delta^*(\sqrt{x},y;q,a\overline{\ell})\Big\}&\ll\frac{x^{1-c_1/\log y}}{\phi(q)\log y}\sum_{\substack{\ell\leq\sqrt{x}\\P^-(\ell)>y}}\frac{\Lambda(\ell)}{{\ell}^{1-c_1/\log y}}\\
&\ll\frac{x^{1-c_1/(2\log y)}}{\phi(q)\log y}\sum_{\ell\leq\sqrt{x}}\frac{\Lambda(\ell)}{\ell}\\
&\ll \frac{x^{1-c_1/(2\log y)}\log x}{\phi(q)\log y}\ll \frac{x^{1-c_1/(3\log y)}}{\phi(q)}.
\end{align*}
\noindent
Hence,
\begin{align}\label{rec}
\Delta(x,y;q,a)=-\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\Delta\Big(\frac{x}{m},y;q,a\overline{m}\Big)+O\bigg(\frac{x^{1-c_2/\log y}}{\phi(q)}\bigg),
\end{align}
\noindent
where $c_2=c_1/3$.
Now, we set $D=x^{1-\delta/\log y}$ for some sufficietly small $\delta\in(0,c_2)$. When $m\leq \sqrt{x}$ and $x\geq q^B$ with a suitably large exponent $B$, we have that $D/m\geq (x/m)^{\eta}q$ with $\eta=2\delta/(\log 2)$. Thus, for $t\in[x-D,x]$, an application of the Brun-Titchmarsch theorem gives
$$\sum_{\substack{\frac{t}{m}<\ell\leq\frac{x}{m}\\\ell\equiv a\overline{m}\mod{q}\\ P^-(\ell)>y}}\Lambda(\ell)\leq\sum_{\substack{\frac{x-D}{m}<\ell\leq\frac{x}{m}\\\ell\equiv a\overline{m}\mod{q}}}\Lambda(\ell)\ll\frac{D}{m\phi(q)}.$$
\noindent
Similarly,
$$\sum_{\substack{\frac{t}{m}<\ell\leq\frac{x}{m}\\P^-(\ell)>y}}\Lambda(\ell)\chi_0(\ell)\ll\frac{D}{m}\quad\,\,\,\,\text{and}\quad\sum_{\substack{\frac{t}{m}<\ell\leq\frac{x}{m}\\P^-(\ell)>y}}\Lambda(\ell)\psi(\ell)\ll\frac{D}{m},$$
\vspace{2mm}
\noindent
by bounding the characters trivially before making use of the Brun-Titchmarsh theorem. With these estimates, we deduce that
\begin{align}\label{trick}
\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\Delta\Big(\frac{x}{m},y;q,a\overline{m}\Big)&=\frac{1}{D}\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\int_{x-D}^x\Delta\Big(\frac{x}{t},y;q,a\overline{m}\Big)\mathrm{d} t\\
&+O\bigg(\frac{D}{\phi(q)}\sum_{\substack{m\leq\sqrt{x}\\P^-(m)>y}}\frac{1}{m}\bigg).\nonumber
\end{align}
\noindent
But,
$$\sum_{\substack{m\leq\sqrt{x}\\P^-(m)>y}}\frac{1}{m}\leq \prod_{y<p\leq\sqrt{x}}\bigg(1-\frac{1}{p}\bigg)^{-1}\ll \frac{\log x}{\log y},$$
\noindent
and so combination of (\ref{rec}) and (\ref{trick}) leads to
\vspace{-1mm}
\begin{align}\label{recres}
\Delta(x,y;q,a)=-\frac{1}{D}\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\int_{x-D}^x\Delta\Big(\frac{x}{t},y;q,a\overline{m}\Big)\mathrm{d} t+O\bigg(\frac{x\log x}{\phi(q)\log y}\bigg(\frac{2k\log y}{\delta\log x}\bigg)^{\!\!k+1\,}\bigg),
\end{align}
\vspace{1mm}
\noindent
since $x^{\delta/\log y}=\exp\{\delta\log x/\log y\}\geq (\delta\log x/\log y)^{k+1}/(k+1)!\geq (\delta\log x/(2k\log y))^{k+1}$.
By referring to the orthogonality of the characters modulo $q$, it follows that
\vspace{-1mm}
$$\Delta\Big(\frac{x}{t},y;q,a\overline{m}\Big)=\frac{1}{\phi(q)}\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0,\psi}}\overline{\chi}(a)\chi(m)\sum_{\substack{\ell\leq\frac{x}{t}\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell).$$
\noindent
Therefore,
\vspace{-1mm}
\begin{align*}
&\frac{1}{D}\bigg|\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\int_{x-D}^x\Delta\Big(\frac{x}{t},y;q,a\overline{m}\Big)\mathrm{d} t\bigg|\\
&\leq \frac{1}{D\phi(q)}\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0,\psi}}\bigg|\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\chi(m)\int_{x-D}^x\Big(\sum_{\substack{\ell\leq \frac{x}{t}\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big)\mathrm{d} t\bigg|\\
&=\frac{1}{D\phi(q)}\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0,\psi}}\bigg|\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\chi(m)m\int_{\frac{x-D}{m}}^{\frac{x}{m}}\Big(\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big)\mathrm{d} t\bigg|\\
&=\frac{1}{D\phi(q)}\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0, \psi}}\bigg|\int_{\frac{x-D}{\sqrt{x}}}^{\frac{x}{y}}\Big(\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big)\Big(\sum_{\substack{\frac{x-D}{t}<m\leq\frac{x}{t}\\P^-(m)>y}}\chi(m)m\Big)\mathrm{d} t\bigg|.
\end{align*}
\vspace{1mm}
\noindent
We move the absolute value inside the integral and then use the Cauchy-Schwarz inequality twice to obtain
\vspace{-1mm}
\begin{align}\label{dcs}
&\frac{1}{D}\bigg|\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\int_{x-D}^x\Delta\Big(\frac{x}{t},y;q,a\overline{m}\Big)\mathrm{d} t\bigg|\\
&\leq\frac{1}{D\phi(q)}\int_{\frac{x-D}{\sqrt{x}}}^{\frac{x}{y}}\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big|^2\bigg)^{1/2}\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\Big|\sum_{\substack{\frac{x-D}{t}<m\leq\frac{x}{t}\\P^-(m)>y}}\chi(m)m\Big|^2\bigg)^{1/2}\mathrm{d} t\nonumber\\
&\leq\frac{1}{D\phi(q)}\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{x-D}{\sqrt{x}}}^{\frac{x}{y}}\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big|^2\frac{\mathrm{d} t}{t^3}\bigg)^{1/2}\nonumber\\
&\times \bigg(\int_{\frac{x-D}{\sqrt{x}}}^{\frac{x}{y}}\sum_{\chi\mod{q}}\Big|\sum_{\substack{\frac{x-D}{t}<m\leq\frac{x}{t}\\P^-(m)>y}}\chi(m)m\Big|^2t^3\mathrm{d} t\bigg)^{1/2}.\nonumber
\end{align}
\noindent
However,
\vspace{-1mm}
\begin{align*}
\sum_{\chi\mod{q}}\Big|\sum_{\substack{\frac{x-D}{t}<m\leq\frac{x}{t}\\P^-(m)>y}}\chi(m)m\Big|^2&=\phi(q)\sum_{b\in(\mathbb{Z}/q\mathbb{Z})^*}\bigg(\sum_{\substack{\frac{x-D}{\sqrt{x}}<m\leq \frac{x}{t}\\mathfrak{m}\equiv b\mod{q}\\P^-(m)>y}}m\bigg)^2\\
&\leq \frac{\phi(q)x^2}{t^2}\sum_{b\in(\mathbb{Z}/q\mathbb{Z})^*}\bigg(\sum_{\substack{\frac{x-D}{\sqrt{x}}<m\leq\frac{x}{t}\\mathfrak{m}\equiv b\mod{q}\\P^-(m)>y}}1\bigg)^2\ll\frac{x^2D^2}{t^4(\log y)^2},
\end{align*}
\vspace{1mm}
\noindent
where the final step is a consequence of Theorem \ref{shiu} which was applied separately to each sum of the last line. So,
\vspace{-1mm}
$$\int_{\frac{x-D}{\sqrt{x}}}^{\frac{x}{y}}\sum_{\chi\mod{q}}\Big|\sum_{\substack{\frac{x-D}{t}<m\leq\frac{x}{t}\\P^-(m)>y}}\chi(m)m\Big|^2t^3\mathrm{d} t\ll\frac{x^2D^2}{(\log y)^2}\int_1^x\frac{\mathrm{d} t}{t}=\frac{x^2D^2\log x}{(\log y)^2}.$$
\vspace{1mm}
\noindent
This means that (\ref{dcs}) leads to the estimate
\vspace{-1mm}
\begin{align}\label{bP}
\frac{1}{D}\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\int_{x-D}^x\Delta\Big(\frac{x}{t},y;q,a\overline{m}\Big)\mathrm{d} t&\ll \frac{x\sqrt{\log x}}{\phi(q)\log y}\\
&\times\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{\sqrt{x}}{2}}^x\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big|^2\frac{\mathrm{d} t}{t^3}\bigg)^{1/2}.\nonumber
\end{align}
\vspace{1mm}
Now, we continue by establishing a bound for the sum of integrals at the second line of (\ref{bP}). For $t\in[\sqrt{x}/2,x]$, by partial summation and Chebyshev's estimates, we have that
\vspace{-2mm}
\begin{align*}
\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)&=O(\sqrt{t})+\int_{\sqrt{t}}^t(\log u)^{-k}\mathrm{d} \Big(\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big)\\
&=O(\sqrt{t})+(\log t)^{-k}\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\\
&+\int_{\sqrt{t}}^t\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\frac{\mathrm{d} t}{u(\log u)^{k+1}}\\
&\ll\sqrt{t}+M_0^k(\log x)^{-k}\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|\\
&+M_0^k(\log x)^{-(k+1)}\int_{\sqrt{t}}^t\Big|\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|\frac{\mathrm{d} u}{u},
\end{align*}
\noindent
where $M_0$ is some sufficiently large positive constant. Because of the basic inequality $3(\alpha^2+\beta^2+\gamma^2)\geq (\alpha+\beta+\gamma)^2$ for all $\alpha, \beta, \gamma\in\mathbb{R}$, we deduce that
\vspace{-2mm}
\begin{align*}
\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big|^2&\ll t+M_0^{2k}(\log x)^{-2k}\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|^2\\
&+M_0^{2k}(\log x)^{-2(k+1)}\bigg(\int_{\sqrt{t}}^t\Big|\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|\frac{\mathrm{d} u}{u}\bigg)^2.
\end{align*}
\noindent
But, upon noticing that
\vspace{-1mm}
$$\bigg(\int_{\sqrt{t}}^t\Big|\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|\frac{\mathrm{d} u}{u}\bigg)^2\leq\frac{\log t}{2}\int_{\sqrt{t}}^t\Big|\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|^2\frac{\mathrm{d} u}{u},$$
\vspace{1mm}
\noindent
by the Cauchy-Schwarz inequality, we conclude that
\vspace{-2mm}
\begin{align*}
&\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{\sqrt{x}}{2}}^x\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big|^2\frac{\mathrm{d} t}{t^3}\bigg)^{1/2}\\
&\ll\sqrt{\frac{\phi(q)}{x}}+M_0^k(\log x)^{-k}\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{\sqrt{x}}{2}}^x\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|^2\frac{\mathrm{d} t}{t^3}\bigg)^{1/2}\nonumber\\
&+M_0^k(\log x)^{-k}\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{\sqrt{x}}{2}}^x\bigg(\int_{\sqrt{t}}^t\Big|\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|^2\frac{\mathrm{d} u}{u}\bigg)\frac{\mathrm{d} t}{t^3}\bigg)^{1/2}.\nonumber
\end{align*}
\vspace{1mm}
\noindent
We use Fubini's theorem to interchange the order of integration, and so the double integral above equals
\vspace{-2mm}
\begin{align*}
\int_{\frac{\sqrt[4]{x}}{\sqrt{2}}}^x\Big|\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|^2\bigg(\int_u^{u^2}\frac{\mathrm{d} t}{t^3}\bigg)\frac{\mathrm{d} u}{u}\leq\int_{\frac{\sqrt[4]{x}}{2}}^x\Big|\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|^2\frac{\mathrm{d} u}{u^3}.
\end{align*}
\noindent
Hence,
\vspace{-2mm}
\begin{align*}
&\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{\sqrt{x}}{2}}^x\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big|^2\frac{\mathrm{d} t}{t^3}\bigg)^{1/2}\\
&\ll\sqrt{\frac{\phi(q)}{x}}+M_0^k(\log x)^{-k}\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{\sqrt[4]{x}}{2}}^x\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|^2\frac{\mathrm{d} t}{t^3}\bigg)^{1/2}\nonumber\\
&\ll\sqrt{\frac{\phi(q)}{x}}+M_0^k(\log x)^{-k}\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{\sqrt[4]{x}}{2}}^x\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|^2\frac{\mathrm{d} t}{t^{3+2/\log x}}\bigg)^{1/2}.\nonumber
\end{align*}
\noindent
Since Parseval's theorem for Dirichlet series guarantees that
$$\int_1^{\infty}\Big|\sum_{\substack{\ell\leq u\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)(\log \ell)^k\Big|^2\frac{\mathrm{d} u}{u^{3+2/\log x}}=\frac{1}{2\pi}\int_{\mathbb{R}}\Big|\bigg(\frac{L_y^{'}}{L_y}\bigg)^{(k)}(c+it,\chi)\Big|^2\frac{\mathrm{d} t}{c^2+t^2},$$
\noindent
with $c:=1+1/\log x$, we arrive at the following bound.
\begin{align}\label{afP}
&\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{\sqrt{x}}{2}}^x\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big|^2\frac{\mathrm{d} t}{t^3}\bigg)^{1/2}\\
&\ll\sqrt{\frac{\phi(q)}{x}}+M_0^k(\log x)^{-k}\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0, \psi}}\int_{\mathbb{R}}\Big|\bigg(\frac{L_y^{'}}{L_y}\bigg)^{(k)}(c+it,\chi)\Big|^2\frac{\mathrm{d} t}{c^2+t^2}\bigg)^{1/2}\nonumber
\end{align}
We now restrict our attention to the integrals involving the derivatives of $L_y^{'}/L_y$. We are going to split these integrals into two parts. In the first parts, we will be integrating over $|t|\leq T$. We will bound these parts by mainly using the results of Section 3. For the remaining parts, where we integrate over the range $|t|>T$, we will use Lemma \ref{mvt}. We start with the integrals over $|t|>T$ first.
It is true that
\begin{align*}
\sum_{\chi\mod{q}}\bigg|\bigg(\frac{L_y^{'}}{L_y}\bigg)^{(k)}(c+it,\chi)\bigg|^2=\phi(q)\sum_{b\in(\mathbb{Z}/q\mathbb{Z})^*}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^k}{n^{c+it}}\bigg|^2.
\end{align*}
\noindent
This may be proven by opening the square of the left hand-side and by using the orthogonality relations of the characters modulo $q$. So, now one can use Lemma \ref{mvt} to infer that
\begin{align}\label{smt}
&\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0, \psi}}\int_{|t|>T}\Big|\bigg(\frac{L_y^{'}}{L_y}\bigg)^{(k)}(c+it,\chi)\Big|^2\frac{\mathrm{d} t}{c^2+t^2}\\
&\leq\phi(q)\sum_{b\in(\mathbb{Z}/q\mathbb{Z})^*}\int_{|t|>T}\bigg|\sum_{\substack{n\equiv a\mod{q}\\P^-(n)>y}}\frac{\Lambda(n)(\log n)^k}{n^{c+it}}\bigg|^2\frac{\mathrm{d} t}{t^2}\ll \frac{(c_3k)^{2k}(\log x)^{2k+1}}{T},\nonumber\end{align}
\noindent
where $c_3=2\sqrt{\log 4}$.
Our treatment for the integrals corresponding to the large values of $t$ is complete and we turn our focus to the integrals whose range of integration is $|t|\leq T$. For a character $\chi\notin\{\chi_0,\psi\}$, using Lemma \ref{lasl}, it follows that
\begin{align}\label{afld}
&\int_{|t|\leq T}\Big|\bigg(\frac{L_y^{'}}{L_y}\bigg)^{(k)}(c+it,\chi)\Big|^2\frac{\mathrm{d} t}{c^2+t^2}\\
&\leq4^k((k+1)!)^2\sum_{j=1}^{k+1}(j!)^{-\frac{2(k+1)}{j}}\int_{|t|\leq T}\Big|\frac{L_y^{(j)}}{L_y}(c+it,\chi)\Big|^{\frac{2(k+1)}{j}}\frac{\mathrm{d} t}{c^2+t^2}.\nonumber
\end{align}
\vspace{3mm}
\noindent
Since $\chi\notin\{\chi_0,\psi\}$ and $y=(10q)^{100}V_T>qV_t$ when $|t|\leq T$, a proper combination of Theorems \ref{lb} and \ref{ldnS} implies that $\abs{L_y^{-1}(c+it,\chi)}\ll 1$. Moreover, $|L_y^{(j)}(c+it,\chi)|\ll j!(C\log y)^j$ for all $j\in\{1,\ldots,k+1\}$, as can be seen from Theorem \ref{ub}. Hence, (\ref{afld}) gives
\begin{align}\label{1rel}
&\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0,\psi}}\int_{|t|\leq T}\Big|\bigg(\frac{L_y^{'}}{L_y}\bigg)^{(k)}(c+it,\chi)\Big|^2\frac{\mathrm{d} t}{c^2+t^2}\\
&\ll (2C)^{2k}((k+1)!)^2\sum_{j=1}^{k+1}\frac{(\log y)^{2(k+1-j)}}{(j!)^2}\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0,\psi}}\int_{\mathbb{R}}\abs{L_y^{(j)}(c+it,\chi)}^2\frac{\mathrm{d} t}{c^2+t^2}.\nonumber
\end{align}
\vspace{2mm}
\noindent
From an application of Parseval's theorem, it follows that
\begin{align}\label{lP}
\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0, \psi}}\int_{\mathbb{R}}\abs{L_y^{(j)}(c+it,\chi)}^2\frac{\mathrm{d} t}{c^2+t^2}\leq\int_y^{\infty}\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0}}\Big|\sum_{\substack{n\leq u\\P^-(n)>y}}\chi(n)(\log n)^j\Big|^2\frac{\mathrm{d} u}{u^3}.
\end{align}
\noindent
For $b\in(\mathbb{Z}/q\mathbb{Z})^*$ and $u\geq y$, Lemmas \ref{chlem} and \ref{logap} yield that
\begin{align}\label{1''}
\sum_{\substack{n\leq u\\n\equiv b \mod{q}\\P^-(n)>y}}(\log n)^j-\frac{1}{\phi(q)}\sum_{\substack{n\leq u\\P^-(n)>y}}\chi_0(n)(\log n)^j\ll\frac{(\log u)^ju^{1-c_1/\log y}}{\phi(q)\log y}.
\end{align}
\noindent
Since $\int_y^v(\log t)^j\mathrm{d} t\leq v(\log v)^j$ for $v\geq y$, these lemmas also imply that
\begin{align}\label{2''}
\sum_{\substack{n\leq u\\n\equiv b \mod{q}\\P^-(n)>y}}(\log n)^j+\frac{1}{\phi(q)}\sum_{\substack{n\leq u\\P^-(n)>y}}\chi_0(n)(\log n)^j\ll\frac{u(\log u)^j}{\phi(q)\log y}.
\end{align}
\noindent
We combine (\ref{1''}) and (\ref{2''}) with the elementary identity $w^2-z^2=(w-z)(w+z)$ and infer that
\begin{align*}
&\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0}}\Big|\sum_{\substack{n\leq u\\P^-(n)>y}}\chi(n)(\log n)^j\Big|^2\\
&=\phi(q)\sum_{b\in(\mathbb{Z}/q\mathbb{Z})^*}\bigg\{\Big(\sum_{\substack{n\leq u\\n\equiv b \mod{q}\\P^-(n)>y}}(\log n)^j\Big)^2-\Big(\frac{1}{\phi(q)}\sum_{\substack{n\leq u\\P^-(n)>y}}\chi_0(n)(\log n)^j\Big)^2\bigg\}\\
&\ll\frac{(\log u)^{2j}u^{2-c_1/\log y}}{(\log y)^2},
\end{align*}
\vspace{3mm}
\noindent
for all $j\in\{1,\ldots,k+1\}$. We plug this estimate into (\ref{lP}) and obtain
\vspace{1mm}
\begin{align*}
\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0, \psi}}\int_{\mathbb{R}}\abs{L_y^{(j)}(c+it,\chi)}^2\frac{\mathrm{d} t}{c^2+t^2}&\ll\frac{1}{(\log y)^2}\int_y^{\infty}(\log u)^{2j}u^{-1-c_1/\log y}\mathrm{d} u\\
&=c_1^{-2j-1}(\log y)^{2j-1}\Gamma(2j+1)\leq c_1^{-3}(2j)!(\log y)^{2j-1},
\end{align*}
\vspace{2mm}
\noindent
for $j\in\{1,\ldots,k+1\}$. Now, with this last bound, (\ref{1rel}) turns into
\begin{align*}
\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0,\psi}}\int_{|t|\leq T}\Big|\bigg(\frac{L_y^{'}}{L_y}\bigg)^{(k)}(c+it,\chi)\Big|^2\frac{\mathrm{d} t}{c^2+t^2}\ll (2C)^{2k}((k+1)!)^2(\log y)^{2k+1}\sum_{j=1}^{k+1}\binom{2j}{j}.
\end{align*}
\noindent
But, the series $\sum_{n\geq 1}\binom{2n}{n}5^{-n}$ converges because of the ratio test, and so
$$\sum_{j=1}^{k+1}\binom{2j}{j}\leq 5^{k+1}\sum_{j=1}^{k+1}\binom{2j}{j}5^{-j}\leq25^k\sum_{j\geq 1}\binom{2j}{j}5^{-j}\ll 25^k.$$
\vspace{2mm}
\noindent
Therefore, we complete the estimation of the integrals over $|t|\leq T$ by arriving at the bound
\begin{align}\label{lart}
\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0,\psi}}\int_{|t|\leq T}\Big|\bigg(\frac{L_y^{'}}{L_y}\bigg)^{(k)}(c+it,\chi)\Big|^2\frac{\mathrm{d} t}{c^2+t^2}\ll (10C)^{2k}((k+1)!)^2(\log y)^{2k+1}.
\end{align}
Putting (\ref{smt}) and (\ref{lart}) together, we deduce that
\begin{align*}
\sum_{\substack{\chi\mod{q}\\\chi\neq\chi_0,\psi}}\int_{\mathbb{R}}\Big|\bigg(\frac{L_y^{'}}{L_y}\bigg)^{(k)}(c+it,\chi)\Big|^2\frac{\mathrm{d} t}{c^2+t^2}\ll (10C)^{2k}((k+1)!)^2(\log y)^{2k+1}+\frac{(c_3k)^{2k}(\log x)^{2k+1}}{T}.
\end{align*}
\vspace{2mm}
\noindent
We insert this into (\ref{afP}) and conclude that
\begin{align}\label{afc}
&\bigg(\sum_{\substack{\chi\mod{q}\\\chi\neq \chi_0, \psi}}\int_{\frac{\sqrt{x}}{2}}^x\Big|\sum_{\substack{\ell\leq t\\P^-(\ell)>y}}\Lambda(\ell)\chi(\ell)\Big|^2\frac{\mathrm{d} t}{t^3}\bigg)^{1/2}\\
&\ll\sqrt{\frac{\phi(q)}{x}}+(10M_0C)^k(k+1)!\frac{(\log y)^{k+\frac{1}{2}}}{(\log x)^k}+\frac{(c_3M_0k)^k(\log x)^{k+\frac{1}{2}}}{\sqrt{T}}.\nonumber
\end{align}
\vspace{3mm}
\noindent
In turn, (\ref{afc}), when used in (\ref{bP}), implies that
\begin{align}\label{alend}
\frac{1}{D}\sum_{\substack{1<m\leq\sqrt{x}\\P^-(m)>y}}\int_{x-D}^x\Delta\Big(\frac{x}{t},y;q,a\overline{m}\Big)\mathrm{d} t\ll\frac{x}{\phi(q)}\bigg(\frac{\log x}{\log y}\bigg)^{3/2}\bigg\{\bigg(\frac{M_1\ell\log y}{\log x}\bigg)^{\!\!\ell}+\frac{(c_3M_0\ell)^{\ell}}{\sqrt{T}}\bigg\},
\end{align}
\vspace{2mm}
\noindent
where $\ell=k+1$ and $M_1=10M_0C$. Note that we omitted the term stemming from $\sqrt{\phi(q)}/\sqrt{x}$, since we are working with a $x\geq q^B$ for a sufficiently large $B$. Now, we combine (\ref{recres}) with (\ref{alend}) and obtain that
\begin{align}\label{forend}
\Delta(x,y;q,a)\ll \frac{x}{\phi(q)}\bigg(\frac{\log x}{\log y}\bigg)^{3/2}\bigg\{\bigg(\frac{M_2\ell\log y}{\log x}\bigg)^{\!\!\ell}+\frac{(c_3M_0\ell)^{\ell}}{\sqrt{T}}\bigg\},
\end{align}
\vspace{2mm}
\noindent
where $M_2=\max\{M_1,2/\delta\}$.
If $C'=eM_2$, then by choosing
$$\ell=\flbgg{\frac{\log x}{C'\log y}},$$
\noindent
it follows that
$$\bigg(\frac{M_2\ell\log y}{\log x}\bigg)^{\!\!\ell}\leq \frac{\log x}{M_2\log y}x^{-1/(C'\log y)}.$$
\vspace{3mm}
\noindent
Since $T=\exp\{2L(\log x)^{3/5}(\log\log x)^{2/5}\}$, we have $\log V_T\asymp (\log x)^{2/5}(\log\log x)^{3/5}$, and so with the above integer $\ell$, we deduce that
\begin{align}\label{opt1}
\bigg(\frac{\log x}{\log y}\bigg)^{3/2}\bigg(\frac{M_2\ell\log y}{\log x}\bigg)^{\!\!\ell}&\ll x^{-\frac{1}{2C'\log y}}\ll x^{-\frac{1}{4C'\log (2q)}}+e^{-\frac{(\log x)^{3/5}}{4C'(\log\log x)^{3/5}}}.
\end{align}
\vspace{2mm}
\noindent
In order to justify the last step, note that the first term $x^{-1/(4C'\log (2q))}$ of the rightmost side dominates $x^{-1/(2C'\log y)}$ when $V_T\geq q$, while the second term majorizes the expression $x^{-1/(2C'\log y)}$ when $q\geq V_T$. We also observe that $\ell\leq \log x/\log V_T\ll (\log x)^{3/5}(\log\log x)^{-3/5}$ and that $\log\ell\leq\log\log x$. Hence, there exist some positive constants $c_4$ and $c_5$ such that
\begin{align}\label{opt2}
\bigg(\frac{\log x}{\log y}\bigg)^{3/2}\frac{(c_3M_0\ell)^{\ell}}{\sqrt{T}}&\leq \exp\bigg\{\ell\log \ell+\log(c_3M_0)\ell+\frac{3}{2}\log\log x-L(\log x)^{3/5}(\log\log x)^{2/5}\bigg\}\nonumber\\
&\leq\exp\{c_4(\log x)^{3/5}(\log\log x)^{1-9/15}-L(\log x)^{3/5}(\log\log x)^{2/5}\}\nonumber\\
&\leq\exp\{-c_5(\log x)^{3/5}(\log\log x)^{2/5}\},
\end{align}
\vspace{3mm}
\noindent
because the constant $L$ in the definition of $T$ is suffiently large.
With the selection of $\ell$ that we made, we have the estimates (\ref{opt1}) and (\ref{opt2}) and then (\ref{forend}) becomes
\begin{align*}
\Delta(x,y;q,a)\ll\frac{x^{1-c_6/\log(2q)}}{\phi(q)}+\frac{xe^{-c_6(\log x)^{3/5}(\log \log x)^{-3/5}}}{\phi(q)},
\end{align*}
\vspace{2mm}
\noindent
where $c_6=1/(4C')$. The proof of the theorem is almost complete. It only remains to observe that $\Delta(x,y;q,a)$ does not differ much from
$$\sum_{\substack{n\leq x\\n\equiv a\mod{q}}}\Lambda(n)-\frac{x}{\phi(q)}-\frac{\psi(a)}{\phi(q)}\sum_{n\leq x}\Lambda(n)\psi(n).$$
\vspace{2mm}
First,
$$\sum_{\substack{n\leq x\\n\equiv a\mod{q} \\P^-(n)>y}}\Lambda(n)=\sum_{\substack{n\leq x\\n\equiv a\mod{q}}}\Lambda(n)+O\bigg(\frac{y\log x}{\log y}\bigg),$$
\noindent
because
\begin{align*}
0\leq \sum_{\substack{n\leq x\\n\equiv a\mod{q}}}\Lambda(n)-\sum_{\substack{n\leq x\\n\equiv a\mod{q} \\P^-(n)>y}}\Lambda(n)=\sum_{\substack{p^{\nu}\leq x\\p\leq y}}\Lambda(p^{\nu})\leq \pi(y)\log x\ll\frac{y\log x}{\log y}.
\end{align*}
\noindent
Similarly,
$$\sum_{\substack{n\leq x\\P^-(n)>y}}\Lambda(n)\chi(n)=\sum_{n\leq x}\Lambda(n)\chi(n)+O\bigg(\frac{y\log x}{\log y}\bigg)$$
\vspace{1mm}
\noindent
for $\chi\in\{\chi_0,\psi\}$. In addition, there exists a constant $c_7>0$ such that
$$\sum_{n\leq x}\Lambda(n)\chi_0(n)=x+O(x\exp\{-c_7(\log x)^{3/5}(\log\log x)^{-1/5}\}),$$
\vspace{1mm}
\noindent
and so, when $x\geq q^B$ for a large $B>0$, with the $y$ that we had chosen, we conclude that
\begin{align*}
&\Delta(x,y;q,a)=\sum_{n\leq x}\Lambda(n)-\frac{x}{\phi(q)}-\frac{\psi(a)}{\phi(q)}\sum_{n\leq x}\Lambda(n)\psi(n)\\
&\ll\frac{x^{1-c_6/\log(2q)}}{\phi(q)}+\frac{xe^{-c_8(\log x)^{3/5}(\log \log x)^{-3/5}}}{\phi(q)}
\end{align*}
\vspace{3mm}
\noindent
with $c_8=\min\{c_6,c_7\}$. The theorem has now been proven.
\bibliographystyle{alpha}
| f61fe6409e7cb1e44ff30329fc83987690b28454 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection*{Du~Bois property}{\ }
\noindent
Let $M$ be a compact K\"ahler manifold. One of the useful consequences of the Hodge
decomposition is the surjectivity of the natural map
\[
H^i(M, \c)\onto H^i(M, \o_M).
\]
Roughly speaking, projective varieties with Du~Bois singularities form the largest
class that is stable under natural operations (small deformations, products, general
hyperplane sections) where the above surjectivity still holds.
For our curent purposes the Du~Bois property can be handled as a black box. We list
in Paragraph~\ref{db.sings.props.say} all the properties that we use. We give
references to the original papers; \cite[Chap.6]{kk-singbook} is a suitable general
introduction.
The original and most useful definition is rather complicated; see the papers
\cite{DuBois81, MR2339829, Kovacs11d} or \cite[Sec.6.1]{kk-singbook}. The following
characterization emphasizes that Du~Bois is a generalization of seminormality. We can
take (\ref{schwede.thm}.\ref{item:2}) as our definition. (We use `$X$ is Du~Bois' and
`$X$ has Du~Bois singularities' as synonyms.) We state the version given in
\cite[6.4]{Kovacs-Schwede11}.
\begin{defn-thm}\cite{MR2339829}\label{schwede.thm}
Let $X$ be reduced and $Y\supset X$ a smooth space containing it. Let
$\pi:Y'\to Y$ be an embedded log resolution of $X$, that is, $Y'$ is smooth and
$E:=\red \pi^{-1}(X)$ is a simple normal crossing divisor. Then $X$ is
\begin{enumerate}
\item\label{item:1} seminormal iff $\pi_*\o_E=\o_X$, and
\item\label{item:2} Du~Bois iff $\pi_*\o_E=\o_X$ and $R^i\pi_*\o_E=0$ for
$i>0$.\qed
\end{enumerate}
\end{defn-thm}
In particular, if $X$ is Du~Bois, then it is reduced and seminormal.
If $X$ is smooth then the blow-up $Y':=B_XY\to Y$ shows that $X$ is Du~Bois.
\begin{say}[Properties of Du~Bois singularities that we
use]\label{db.sings.props.say}
We work either with algebraic spaces of finite type over a field of characteristic
0, or with complex analytic spaces. Note that we allow them to have irreducible
components of different dimensions.
\medskip\noindent {\it Property~\thethm.0.} Smooth implies
Du~Bois. Du~Bois implies reduced and seminormal.
\medskip\noindent {\it Property~\thethm.1.} \cite{k-db, k-db2} Let $(X, \Delta)$ be
an log canonical pair and $V\subset X$ a union of some of its log canonical
centers. Then $V$ is Du~Bois. More generally, this holds for log canonical centers
of crepant log structures, as in Definition~\ref{crep.l.s.def}.
\medskip
\noindent {\it Property~\thethm.2.} \cite[2.11-12]{MR2784747} Let
$X_1, X_2\subset X$ be closed subspaces. If 3 of
$\{X_1\cap X_2, X_1, X_2, X_1\cup X_2\}$ are Du~Bois, then so is the 4th.
\medskip
\noindent {\it Property~\thethm.3.} \cite[1.6]{k-db}, \cite[3.3]{Kovacs11c} and
\cite[6.27]{kk-singbook}. Let $p: Y\to X$ be a proper surjective morphism,
$V\subset X$ a closed, reduced subscheme, and $D:= \supp p^{-1}(V)$. Assume that
$\o_X(-V)\to R p_*\o_Y(-D)$ has a left inverse and $Y, D$ are Du~Bois. Then $X$ is
Du~Bois $\Leftrightarrow$ $V$ is Du~Bois.
\end{say}
In applications they key is to find examples where
Property~\ref{db.sings.props.say}.3 applies. The following gives most known cases.
\begin{thm}\label{1/6.sn.lem}
Let $f:Y\to Z$ be a projective morphism with connected fibers between normal
spaces. Assume that $(Y, \Delta)$ is $\q$-factorial, dlt and
$K_Y+\Delta\sim_{f,\r}0$. Let $D$ be an effective $\z$-divisor such that
$\rdown{\Delta}\subset \supp D\subset \supp \Delta$ and $-D$ is $f$-semiample. Set
$V=f(\supp D)$. Then $\o_Z(-V)\to R f_*\o_Y(-D)$ has a left inverse
\end{thm}
Note that, since $-D$ is $f$-semiample, $D$ does not dominate $Z$. Thus
$V\subsetneq Z$ and $\o_Z(-V)$ makes sense.
\begin{proof} If $f$ is birational then
the proof is much simpler, and worth doing separately.
Choose $\epsilon>0$ such that $\Theta:=\Delta-\epsilon D$ is effective. Note that
\[
-D\sim_{f,\r} K_Y+\Theta+(1-\epsilon)(-D),
\eqno{(\ref{1/6.sn.lem}.1)}
\]
$(Y, \Theta)$ is klt and $(1-\epsilon)(-D)$ is $f$-semiample.
In the birational case, the general form of Grauert-Riemenschneider vanishing gives
that $R^if_*\o_Y(-D)=0$ for $i>0$; see \cite[2.68]{km-book} for the algebraic case
and \cite[3.7]{MR946250} for the analytic one. Thus
$R f_*\o_Y(-D)\,{\simeq}_{qis}\, f_*\o_Y(-D)=\o_Z(-V)$.
If $D$ does not dominate $Z$, then the assumption $\rdown{\Delta}\subset \supp D$
implies that the generic fiber is klt. Also, $D=\supp f^{-1}(V)$, since $-D$ is
$f$-nef and the fibers are connected. We can now use \cite[3.1]{k-hdi2}, more
precisely the form given in \cite[10.41]{kk-singbook}, to get the required left
inverse.
\end{proof}
\subsection*{The klt case of Theorem~\ref{DB.epsilon.thm}}{\ }
\begin{say
\label{DB.epsilon.thm.klt.pf}
The proof is short and follows \cite{k-1/6}.
First we show the special case when $\supp V$ is a divisor; see
Lemma~\ref{DB.epsilon.div.lem}.
In general, we find a dlt modification $g:(Y, \Delta_Y)\to (X, \Delta)$ such that
$D:=g^{-1}(V)$ is a divisor and $\mld(D, Y, \Delta_Y)=\mld(V, X, \Delta)$; see
Proposition~\ref{thm:one-div-models}. Choose $\epsilon>0$ such that
$\Theta:=\Delta_Y-\epsilon D$ is effective. Note that
\[
-D\sim_{g,\r} K_Y+\Theta+(1-\epsilon)(-D), \eqno{(\ref{DB.epsilon.thm.klt.pf}.1)}
\]
$(Y, \Theta)$ is klt.
If $-D$ is $g$-nef, then
Grauert-Riemenschneider vanishing applies to $R^i g_*\o_Y(-D)$.
We can achive these
after running a suitable MMP; see
Lemma~\ref{mmp.on.crep.D.lem}.
Thus we may assume that
$\o_X(-V)\cong R g_*\o_Y(-D)$.
$V$ is now
Du~Bois by (\ref{db.sings.props.say}.3). \qed
\end{say}
\begin{lem} \label{DB.epsilon.div.lem}
Theorem~\ref{DB.epsilon.thm} holds if $X$ is $\q$-factorial and
$V$ has pure codimension 1.
\end{lem}
\begin{proof} Write $V=\cup_{i\in I} D_i$ where the $D_i\subset X$ are irreducible divisors. Note that $\mld(D, X, \Delta)=1-\coeff_{D}\Delta$ for any irreducible divisor $D$.
Thus we can write $\Delta=\sum_{i\in I} d_iD_i+\Delta'$ where $d_i>1-\lcg(n)$ and $D_i\not\subset\supp \Delta'$.
Since $X$ is $\q$-factorial,
$(X, \sum_{i\in I}d_i D_i)$ is also lc, hence so is $(X, \sum_{i\in I} D_i)$ by Definition~\ref{epsilon.defn}.
Note that each $D_i$ is a log canonical center of $(X, \sum_{i\in I} D_i)$, so
$\cup_iD_i$ is Du~Bois by (\ref{db.sings.props.say}.1).
\end{proof}
\begin{prop} \cite[1.38]{kk-singbook}\label{thm:one-div-models}
Let $(X,\Delta)$ be log canonical,
and $\{E_i:i\in I\}$ finitely many exceptional divisors over $X$ such that
$-1\leq a(E_i,X,\Delta)< 0$.
Then there is a $\q$-factorial, dlt
modification $g:\bigl(Y,\Delta_Y\bigr)\to (X,\Delta)$
such that
\begin{enumerate}
\item the $\{E_i:i\in I\}$ are among the exceptional divisors of $g$, and
\item every other exceptional divisor $F$ of $g$ has discrepancy $-1$. \qed
\end{enumerate}
\end{prop}
\begin{lem}\label{mmp.on.crep.D.lem}
Let $f:Y\to Z$ be a projective morphism between normal spaces. Assume that
$(Y, \Delta)$ is $\q$-factorial, dlt and $K_Y+\Delta\sim_{f,\q}0$. Let $D$ be an
effective $\z$-divisor such that $ \supp D\subset \supp \Delta$. Then the
$(-D)$-MMP runs and terminates in a good minimal model if $D$ does not dominate
$Z$.
\end{lem}
\begin{proof} The $(-D)$-MMP is the same as the $(-\epsilon D)$-MMP, which in turn
agrees with the $(K_Y+\Delta -\epsilon D)$-MMP since $K_Y+\Delta\sim_{f,\q}0$.
If $f$ is birational, the $(K_Y+\Delta -\epsilon D)$-MMP runs and terminates by
\cite{birkar11, MR3032329}.
If $f$ is not birational, then the generic fiber of $(Y,\Delta -\epsilon D)\to X$
is the same as the generic fiber of $(Y,\Delta)\to X$, and the latter is a good
minimal model by assumption. Thus the MMP for $(Y,\Delta -\epsilon D)\to X$ runs
and terminates by \cite{MR3032329}.
The above references work for varieties; see \cite{dvp, k-nqmmp} for algebraic
spaces of finite type, \cite{fujino2022minimal} for analytic spaces and
\cite{lyu-mur} for the most general settings.
\end{proof}
\subsection*{Crepant log structures}{\ }
\noindent
For the general case of Theorem~\ref{DB.epsilon.thm}, we first study what the above
proof gives. Keeping in mind the inductive arguments of \cite{k-db}, we do this for
crepant log structures. The end result is Lemma~\ref{DB.epsilon.crep.lem}. Then
induction and repeated use of (\ref{db.sings.props.say}.2) completes the proof in
Proposition~\ref{W.db.ind.lem}.
\begin{defn}\label{crep.l.s.def}
A {\it crepant log structure} is a dominant, projective morphism with connected fibers
$g:(Y, \Delta)\to Z$, where $(Y, \Delta)$ is lc, $Z$ is normal and
$K_Y+ \Delta\sim_{g, \r}0$.
If $(X, \Delta)$ is lc, then the identity
$(X, \Delta)\to X$ is a crepant log structure.
As a generalization of (\ref{db.sings.props.say}.1), $Z$ is Du~Bois
\cite[6.31]{kk-singbook}.
For an irreducible $V\subset Z$ we define $\mld(V,Y,\Delta)$ as the infimum of the
numbers $1+a(E,Y,\Delta)$ where $E$ runs through all divisors over $Y$ that
dominate $V$.
As in Definition~\ref{crep.log.str.mld.defn}, if $V\subset Z$ is a closed subset
with irreducible components $V_i$, then we set
$\mld(V,Y,\Delta):=\max_i \bigl\{\mld(V_i,Y,\Delta)\bigr\}$.
We will use the following property proved in \cite{k-1/6}, see also
\cite[7.5]{kk-singbook}.
\begin{equation}
\label{eq:2}
\mld(V_1\cap V_2,Y,\Delta)\leq \mld(V_1,Y,\Delta)+\mld(V_2,Y,\Delta).
\end{equation}
\end{defn}
The following generalization of Theorem~\ref{DB.epsilon.thm} is better suited for
induction.
\begin{thm}\label{DB.epsilon.crep.thm}
Let $g:(Y, \Delta)\to Z$ be a crepant log structure. Set $n=\dim Y$ and let
$V\subset Z$ be a closed subset such that $\mld(V, Y,\Delta)<\lcg(n)$. Then $V$
has Du~Bois singularities.
\end{thm}
\noindent
Next we see what the method of (\ref{DB.epsilon.thm.klt.pf}) gives for crepant log
structures.
\begin{notation}\label{nota.not.1}
Let $g:(Y, \Delta)\to Z$ be a crepant log structure. For a closed subset
$Z_1\subset Z$, let $Z_1^{\diamond}\subset Z_1$ denote the union of those log
canonical centers of $(Y,\Delta)$ that are contained in $Z_1$, but are nowhere
dense in it.
Note that $Z^{\diamond}$ is the non-klt locus of $g:(Y, \Delta)\to Z$.
If $Z_1$ itself is the union of log canonical centers of $(Y,\Delta)$, then $Z_1$
is seminormal and $Z_1\setminus Z_1^{\diamond}$ is normal by \cite{ambro, fuj-book}
and \cite[4.32]{kk-singbook}.
\end{notation}
\begin{lem} \label{DB.epsilon.crep.lem}
Let $g:(Y, \Delta)\to Z$ be a crepant log structure with klt generic fiber. Set $n=\dim Y$ and let $V\subset Z$ be a closed subset such that $\mld(V, Y,\Delta)<\lcg(n)$.
Then $V\cup Z^{\diamond}$ has Du~Bois singularities.
\end{lem}
\begin{proof} By Proposition~\ref{thm:one-div-models}, we may assume that
$(Y, \Delta)$ is $\q$-factorial, dlt, and there is a divisor $ D=\sum D_i\subset Y$
such that $\rdown{\Delta}\subset D$, $\mld(D, Y, \Delta)<\lcg(n)$, and
$g(D)=V\cup Z^{\diamond}$. After running the MMP for $K_Y+ \Delta-\epsilon D$ for
some $\epsilon>0$ as in Lemma~\ref{mmp.on.crep.D.lem}, we may also assume that $-D$
is $g$-semiaple. As in (\ref{1/6.sn.lem}.1),
\[
-D\sim_{g,\r} K_Y+(\Delta-\epsilon D)+(1-\epsilon)(-D),
\eqno{(\ref{DB.epsilon.crep.lem}.1)}
\]
where $(Y, \Delta-\epsilon D)$ is klt and $(1-\epsilon)(-D)$ is $g$-semiample.
Also note that $D=\supp g^{-1}(V)$, since $-D$ is $g$-nef and the fibers are
connected. We can now use Theorem~\ref{1/6.sn.lem} to get that
\[
\o_Z\bigl(-(V\cup Z^{\diamond})\bigr)\to R g_*\o_Y(-D)
\eqno{(\ref{DB.epsilon.crep.lem}.2)}
\]
has a left inverse. As we noted in Definition~\ref{crep.l.s.def}, $Z$ is Du~Bois.
By (\ref{db.sings.props.say}.3) these imply that $V\cup Z^{\diamond}$ is Du~Bois.
\end{proof}
\begin{cor}\label{W+lc.db.lem}
Let $g:(Y,\Delta)\to X$ be a crepant log structure of dimension $n$, and
$Z\subset X$ a union of some of its log canonical centers. Let $V\subset Z$ be a
closed subset such that $\mld(V, Y,\Delta)<\lcg(n)$. Then $V\cup Z^{\diamond}$ is
Du~Bois.
\end{cor}
\begin{proof} We may assume that $(Y,\Delta)$ is $\q$-factorial and dlt.
For each irreducible component $Z_i\subset Z$, let $Y_i\subset Y$ be a minimal
dimensional log canonical center of $(Y,\Delta)$ that dominates $Z_i$. Set
$\Theta_i:=\diff^*_{Y_i}\Delta$ as in \cite[4.18.4]{kk-singbook}.
Let $\pi_i:\bar Z_i\to Z_i$ denote the normalization. Stein factorization of
$Y_i\to Z_i$ gives $g_i: Y_i\to \widetilde Z_i$ and
$\tau_i:\widetilde Z_i\to \bar Z_i$. The
$g_i: \bigl(Y_i, \Theta_i\bigr)\to \widetilde Z_i$ are crepant log structures with
klt general fibers.
Precise inversion of adjunction \cite[7.10]{kk-singbook} shows that
$(\pi_i\circ \tau_i)^{-1}(Z^{\diamond})=(\widetilde Z_i)^{\diamond}$ and
$\mld\bigl(\widetilde V_i, Y_i, \Theta_i\bigr)\leq \mld(V, Y,\Delta)$, where
$\widetilde V_i:=(\pi_i\circ \tau_i)^{-1}(V)$.
The $\widetilde Z_i$ are Du~Bois by (\ref{db.sings.props.say}.1), and the
$\widetilde V_i\cup\widetilde Z_i^{\diamond}$ are Du~Bois by
Lemma~\ref{DB.epsilon.crep.lem}.
Set $\bar V_i:=\pi_i^{-1}(V)$, $\bar Z_i^{\diamond}:=\pi_i^{-1}(Z^{\diamond})$ and
$\bar Z^{\diamond}:=\cup_i \bar Z_i^{\diamond}$.
The normalized trace map splits $\o_{\bar Z_i}\DOTSB\lhook\joinrel\to (\tau_i)_*\o_{\widetilde Z_i}$,
and hence also splits
\[
\o_{\bar Z_i}\bigl(-(\bar V_i\cup\bar Z_i^{\diamond})\bigr)\DOTSB\lhook\joinrel\to
(\tau_i)_*\o_{\widetilde Z_i}\bigl(-(\widetilde V_i\cup\widetilde
Z_i^{\diamond})\bigr).
\]
Using the first splitting and applying (\ref{db.sings.props.say}.3) to
$\bigl(\widetilde Z_i, \emptyset\bigr)\to \bigl(\bar Z_i, \emptyset\bigr)$ shows
that $\bar Z_i$ is Du~Bois. Applying (\ref{db.sings.props.say}.3) and the second
splitting now gives that $\bar V_i\cup \bar Z_i^{\diamond}$ is Du~Bois.
As we noted in (\ref{nota.not.1}), $Z$ is seminormal and normal outside
$Z^{\diamond}$, thus $\o_Z(-Z^{\diamond})=\pi_*\o_{\bar Z}(-\bar Z^{\diamond})$,
hence
\[
\o_Z\bigl(-(V\cup Z^{\diamond})\bigr)=\pi_*\o_{\bar Z}\bigl(-(\bar V\cup\bar
Z^{\diamond})\bigr).
\]
Since $Z^{\diamond}$ is Du~Bois by induction on the dimension, the first splitting
shows that $Z$ is Du~Bois. Using (\ref{db.sings.props.say}.3) and the second
splitting gives that $V\cup Z^{\diamond}$ is Du~Bois.
\end{proof}
\subsection*{Proof of Theorems~\ref{DB.epsilon.thm} and \ref{DB.epsilon.crep.thm}}{\
}
\noindent
Since Theorem~\ref{DB.epsilon.crep.thm} implies Theorem~\ref{DB.epsilon.thm},
all that remains is to formulate a variant of Theorem~\ref{DB.epsilon.crep.thm}
that allows for induction on the dimension.
The strongest version would use the language of quasi-log structures as in \cite{fuj-book}. They appear implicitly in
the proof of Proposition~\ref{W.db.ind.lem}, but our approach works well enough.
Note that Theorem~\ref{DB.epsilon.crep.thm} is the $Z=X$ special case of
Proposition~\ref{W.db.ind.lem}. Thus the proof of Proposition~\ref{W.db.ind.lem}
yields Theorem~\ref{DB.epsilon.crep.thm}.
\begin{prop}\label{W.db.ind.lem} Let $g:(Y,\Delta)\to X$ be a crepant log structure of dimension $n$, and $Z\subset X$ a union of some of its log canonical centers, allowing $Z=X$. Let
$V\subset Z$ be a closed subset such that $\mld(V, Y,\Delta)<\lcg(n)$. Then
$V$ is Du~Bois.
\end{prop}
\begin{proof} The proof is by induction on $\dim Z$. If $\dim Z=0$ then $V$ is a
union of smooth points, hence Du~Bois.
Write $V=V_1\cup V_2$, where $V_2\subset Z^{\diamond}$ and none of the irreducible
components of $V_1$ is contained in $ Z^{\diamond}$.
Note that $V_1\cup Z^{\diamond}$ is Du~Bois by (\ref{W+lc.db.lem}), and so is
$Z^{\diamond}$. Furthermore, $\mld(V_1\cap Z^{\diamond}, X,\Delta)<\lcg(n)$ by
(\ref{eq:2}), hence $V_1\cap Z^{\diamond} $ is Du~Bois by induction since
$\dim Z^{\diamond}<\dim Z$. Thus $V_1$ is Du~Bois by (\ref{db.sings.props.say}.2).
Next $\mld(V\cap Z^{\diamond}, X,\Delta)<\lcg(n)$ by (\ref{eq:2}), hence
$V\cap Z^{\diamond} $ is Du~Bois by induction. We already checked that $V_1$ and
$V_1\cap Z^{\diamond} = V_1\cap (V\cap Z^{\diamond}) $ are Du~Bois. Thus
$V=V_1\cup (V\cap Z^{\diamond}) $ is Du~Bois by (\ref{db.sings.props.say}.2).
\end{proof}
\subsection*{Conjectures and comments}{\ }
\noindent
In the proof of Theorem~\ref{DB.epsilon.thm}, instead of $\mld(V, Y,\Delta)<\lcg(n)$,
we only use the assumption that $\mld(V_i, Y,\Delta)<\lcg(n-1)$ if $V_i$ is contained
in a log canonical center, and $\mld(V_i, Y,\Delta)<\lcg(n)$ otherwise. This suggests
that the following should be true.
\begin{ques}\label{DB.epsilon.crep.ques}
Let $g:(Y, \Delta)\to Z$ be a crepant log structure. Let $V\subset Z$ be a closed
subset with irreducible components $V_i$. Let $Z_i\supset V_i$ be the minimal log
canonical center that contains $V_i$ (we allow $Z_i=X$). Assume that
$\mld(V_i, Y,\Delta)<\lcg\bigl(\dim(Z_i)\bigr)$ for every $i$. Is $V$ necessarily
Du~Bois?
\end{ques}
\noindent
A related question is the following.
\begin{conj}\label{near.lc.=.lc.q}
Let $(X, \Delta)$ be a quasi-projective, log canonical pair of dimension $n$. Let
$V\subset X$ be a closed subset such that $\mld(V, X,\Delta)<\lcg(n)$ and $V$
contains all log canonical centers of $(X, \Delta)$.
Then there is a log canonical pair $(X, \Theta)$ such that $V$ is the union of all
log canonical centers of $(X, \Theta)$.
\end{conj}
Note that usually one can not choose $\Theta\geq \Delta$, as shown by the
2-dimensional example $\bigl(\a^2, (1-\eta)(x=0)+(1-\eta)(y=0)+\eta(x=y)\bigr)$.
Also, if $Z$ is a log canonical center of $(X, 0)$, then it is also a log canonical
center of any $(X, \Theta)$, so the assumption that $V$ contain all log canonical
centers of $(X, \Delta)$ is necessary in many cases.
\medskip
If $(X, \Delta)$ is klt, then a proof of Conjecture~\ref{near.lc.=.lc.q} is given in \cite{k-dano}. Together with \cite{k-db, k-db2}, this gives another proof of
the klt case of Theorem~\ref{DB.epsilon.thm}.
However, even the full conjecture does not seem to imply
Theorem~\ref{DB.epsilon.thm}, since we get no information about
those $V_i$ that are contained in a log canonical center of $(X, \Delta)$.
A positive answer to the following stronger version would imply
Theorem~\ref{DB.epsilon.thm}.
\begin{ques}\label{near.lc.=.lc.q.2}
Let $(X, \Delta)$ be a quasi-projective, log canonical pair of dimension $n$.
Is there a log canonical pair
$(X, \Theta)$ such that every
irreducible subvariety satisfying $\mld(V, X,\Delta)<\lcg(n)$
is a log canonical center of $(X, \Theta)$?
\end{ques}
Note that usually we can not achieve that the log canonical centers are exactly the
$\{V\colon \mld(V, X,\Delta)<\lcg(n)\}$. Indeed, any intersection of log canonical centers is a union of log canonical centers, but this does not hold for the
$\mld(V, X,\Delta)<\lcg(n)$ condition.
\begin{ack}
Partial financial support to JK was provided by the NSF under grant number
DMS-1901855. SK was supported in part by NSF Grants DMS-1951376 and
DMS-2100389, and a Simons Fellowship (Award Number 916188).
\end{ack}
\def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def\dbar{\leavevmode\hbox to 0pt{\hskip.2ex
\accent"16\hss}d} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$''${$''$}
\def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'$} \def\cprime{$'$} \def\cprime{$'${$'$}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
| f7d47a39dd056b82d0a57eaa9acb59a3a39adad4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Legal information retrieval is a specialized task in natural language processing (NLP) that involve retrieving relevant legal documents \footnote{In this paper, "text" and "document" are used interchangeably} given a query. Compared with traditional text retrieval, legal text retrieval is more difficult since legal documents are often long and complicated. On the other hand, questions are often highly complex and need a person with good expertise to give the correct answer.
Legal information retrieval for English documents has received a lot of attentions. COLIEE \footnote{\url{https://sites.ualberta.ca/~rabelo/COLIEE2022/}} is an annual competition about legal information extraction/entailment or JURIX \footnote{\url{http://jurix.nl/conferences/}} is an annual conference on legal AI.
However, there are very few studies on Vietnamese legal information retrieval. Recent studies of Vietnamese legal text retrieval have been mainly on CNN architecture combined with attention mechanism \cite{kien-etal-2020-answering,nguyen2022toward}, using fine-tuned Multilingual transformer model like XLM-RoBERTa \cite{nguyen2022toward}. In recent years, monolingual transformer models like PhoBERT \cite{nguyen-tuan-nguyen-2020-phobert}, ViBERT \cite{tran2020improving} have beat the state-of-the-art (SOTA) in many NLP tasks in Vietnamese like Part-of-Speech(POS tagging), Named-Entity Recognition(NER), and Dependency Parsing(DP). Sentence-BERT(SBERT) \cite{reimers2019sentence} and models evolved from it \cite{thakur-etal-2021-augmented} have recently reached significant results on several tasks like semantic searching or information retrieval. In this paper, we focus on exploring fine-tuned sentence-transformer monolingual models for Vietnamese legal text retrieval.
The contribution of this paper is two-fold. First, we propose a novel pipeline of multi-stage information retrieval based on sentence-transformer for Vietnamese legal texts. Experiments were conducted and the results show that the proposed pipeline significantly outperforms existing models. Second, empirical analyses of the effects of multiple factors including language models(LM), word segmentation methods, and ranking scores on the performance of information retrieval are performed. The results show that SPhoBERT-large language model, word-based segmentation, and system with combine ranking scores yield the best result in information retrieval for Vietnamese legal documents.
\section{Background and Related Work}
In this paper, we focus on answering legal queries at the article level. Given a legal query, our goal is to retrieve all the relevant articles, that can be used as the fundamental to answer the query.
Many approaches, especially for ad-hoc text retrieval have been proposed, from past decades to recent years.
\subsubsection{Non-neural approaches} These methods, decide relevance based on the frequency and occurrence of the words in the query and the documents. Non-neural approaches do not perform well in the semantic searching problem due to lexical mismatching in the answers. However, they are still useful for current SOTA models. BM25 \cite{10.1145/1031171.1031181} and tf-idf \cite{SALTON1988513} are well-known among all while BM25+\cite{robertson1994some} is currently the most effective in the approach. To overcome the lexical mismatching challenge, dense embedding is used to represent queries and documents. The main idea of this method was proposed with the LSI approach \cite{Deerwester1990IndexingBL}.
\subsubsection{Attention mechanism approaches} Attention mechanism makes the model able to focus more on the main keywords or informative sentences in the original text. Kien and Nguyen et al. signed a simple attentive convolution neural network for Vietnamese legal text retrieval {\cite{kien-etal-2020-answering}}. Nguyen also used this method in his dissertation{\cite{nguyen2022toward}}. Their retrieval system can capture both local and global contexts to construct to build representation vectors.
\subsubsection{Transformer cross-encoder approaches} BERT \cite{devlin2018bert} brings breakthroughs in NLP. The cross-encoder approach is the first widely approach of BERT in information retrieval. Both query and document will be passed simultaneously to the BERT network. Birch \cite{akkalyoncu-yilmaz-etal-2019-applying} - the system combines lexical matching and the cross-encoder approach for better performance. Several other studies \cite{lee-etal-2019-latent,karpukhin-etal-2020-dense,laskar-etal-2020-contextualized} applied the cross-encoder approach and reached significant results. BERT-PLI \cite{shao2020bert} is a retrieval system based on these approaches. However, these approaches require lots of time for training and costly computation resource.
\subsubsection{Transformer bi-encoder approaches} Cross-encoder approaches are costly and time-consuming for training. Motivated by this, Reimers and Gurevych presented SBERT which uses SiameseBERT-network to represent semantically meaningful sentence embeddings. SBERT produce vector embedding for each sentence independently. When we want to compare two sentences, we just need to calculate the cosine similarity of the two existing vectors. The authors of SBERT train the bi-encoder from BERT and compare two input sentences using cosine similarity. For question answering(QA) or IR tasks, many studies developed from or applied SBERT \cite{SBERT-WK,condor2021automatic} and get high performance.According to recent research by Gao and Callan \cite{gao-callan-2021-Condenser} standard LMs' internal attention structure is not ready to use for dense encoders. The authors proposed a new transformer architecture, Condenser as an improvement for the bi-encoder approach. For applying bi-encoder approach for Vietnamese, to the best of our knowledge, there is only one paper \cite{ha2021utilizing} on bi-encoder approach and the solution \footnote{\url{https://github.com/CuongNN218/zalo_ltr_2021}} for an annual competition of artificial intelligence.
\section{Sentence-transformers based Multi-stage information retrieval for Vietnamese legal text}
The general idea of our approach is to use both lexical matching and semantic searching to improve the performance of our system. For lexical matching, we used BM25+ in package rank\_bm25 library \footnote{\url{https://pypi.org/project/rank-bm25/}}. For semantic searching, we trained sentence-transformer models by contrastive learning. For each query in training dataset, the label of the relevant article to query is 1 while the negative sample is 0. To get the negative samples, we took the top-k highest ranking score in the previous training round. Then the samples for each query was k+(number of positive articles)as positive(pos) and negative(neg) pairs. We used BM25+ and then trained the sentence-transformer model for three rounds. Our pipeline for training are shown in Figure \ref{pipeline for training}
\begin{figure}
\raggedright
\centering
\includegraphics[scale=0.6]{figures/training_process.png
\caption{Our proposed pipeline for training}
\label{pipeline for training}
\end{figure}
\section{Experiments}
\subsection{Datasets}
The dataset used in our paper is from the original dataset in the paper published by Kien et al. \cite{kien-etal-2020-answering}.
We clean the data to reduce the noise in the legal corpus. The removed laws and articles do not appear in QA dataset while QA dataset remains the original so that data cleaning does not affect the evaluation results.
Finally, we obtained the legal corpus containing 8,436 documents with 114,177 articles.
In legal corpus, we concatenated title and text in each article as a "very long" sentence.
\begin{table}
\parbox{.45\linewidth}{
\centering
\caption{Distribution of articles length by syllables}
\label{len_syl}
\begin{tabular}{ccc}
\hline
Length & Amount & Proportion\\
\hline
< 100 & 32950 & 28.86\% \\
\hline
101 - 256 & 43923 & 38.47\% \\
\hline
257 - 512 & 23799& 20.84\%\\
\hline
513+ & 9483 & 11.83\% \\
\hline
\end{tabular}
}
\hfill
\parbox{.45\linewidth}{
\centering
\caption{Distribution of articles length by words}
\label{len_word}
\begin{tabular}{ccc}
\hline
Length & Amount & Proportion\\
\hline
< 100 & 43763 & 38.33\% \\
\hline
101 - 256 & 42800 & 37.48\% \\
\hline
257 - 512 & 18636& 16.33\%\\
\hline
513+ & 6111 & 7.86\% \\
\hline
\end{tabular}
}
\end{table}
Not like English, Vietnamese syllables and word tokens are different. For example, 4-syllable written text "Bài báo khoa học" (scientific paper) form 2 words "Bài\_báo \raisebox{-0.4ex}{\footnotesize{paper}} khoa\_học \raisebox{-0.4ex}{\footnotesize{scientific}}". We used VNCoreNLP \cite{vu-etal-2018-vncorenlp} for word segmentation. We divided the corpus into 4 types of lengths: < 100, 101-256,257-512, and 513+. 256 is the maximum number of tokens supported by PhoBERT while 512 is the maximum number of tokens supported by ViBERT. Table \ref{len_syl} and table \ref{len_word} show percentage of articles based on length.
For QA dataset, we found that only 1,709 articles (about 1.5\% of the whole articles) in the legal corpus appear and about 95\% of questions in both training and testing datasets have one to three relevant articles.
\subsection{Experimental procedure}
We use a single NVIDIA Tesla P100 GPU via Google Colaboratory to train all the models. Figure \ref{experiment procedure} presents an overview of the experimental procedure in this paper.
\begin{figure}[!h]
\raggedright
\centering
\includegraphics[scale=0.6]{figures/procedure.png
\caption{Experiment procedure}
\label{experiment procedure}
\end{figure}
\subsubsection{Monolingual pre-trained models}
\begin{itemize}
\item PhoBERT pre-training approach is based on RoBERTa \cite{Liu2019RoBERTaAR} which is using Dynamic Masking to create masked tokens. Because two PhoBERT versions are trained on large-scale Vietnamese dataset, they perform great results on several NLP tasks in Vietnamese like NER, POS tagging, and DP.
\item ViBERT leverages checkpoint from mBERT \cite{devlin2018bert} and continues training on 10 GB of Vietnamese news data. ViBERT pre-training approach based on BERT which has model architecture is a multilayer bidirectional Transformer encoder. ViBERT now is supporting sequence lengths up to 512 tokens, this is an advantage compared to only 256 tokens supported by PhoBERT.
\end{itemize}
\subsubsection{Word-level and syllable-level approaches} In a field that contains many semantic challenges such as law, we want to evaluate how words and syllables of Vietnamese affect the results of models. We trained transformer models on both word-level and syllable-level independently in every below step.
\subsubsection{Fine-tuning monolingual language models} We used our legal corpus to fine-tune masked language modeling of monolingual language models. We set gradient\_accumulation\_steps = 4, train\_batch\_size = 8, eval\_batch\_size = 8 and epochs = 20.
\subsubsection{Fine-tuning Condenser} After training mask language modeling, we used both versions of PhoBERT to train Condenser. We used source code \footnote{\url{ https://github.com/luyug/Condenser}} for creating data in Condenser's form and fine-tuning Condenser from PhoBERT.
\subsubsection{Lexical matching} For lexical matching, we used BM25+. VNCoreNLP was used for word segmentation. Stopwords were also removed to make the model focus on important words. Our goal after this step is to get the list of the top highest lexical similarity articles for each query.
\subsubsection{Training sentence-transformer } We trained 3 rounds for sentence-transformer. The choosing method we mentioned above, for short, we will call the number of negative sample each round is k. In the first round, we picked 35 negative samples from BM25+. These negative samples almost are high lexical similarity to queries but low semantic similarity. In the second round, we picked 20 negative samples from the first round. These negative samples have better semantic similarity to queries. In the third round, we pick 15 negative samples from the second round. These negative samples are really close semantic to queries. We set batch\_size=8, epochs = 4, learning\_rate = 1e-5. For loss function, we used contrastive loss \cite{hadsell2006dimensionality}
\subsubsection{Evaluation metrics}
The first measure we use to evaluate the performance of the system is (macro)recall@20, where 20 is the number of the top selected articles.
The second evaluation metric we use is F2. The F2-measure is shown below:
\[Precision\raisebox{-0.7ex}{\footnotesize{i-th}} = \frac{\text{the number of correctly retrieved articles of query i-th}}{\text{the number of retrieved articles of query i-th}} \]
\[Recall\raisebox{-0.7ex}{\footnotesize{i-th}} = \frac{\text{the number of correctly retrieved articles of query i-th}}{\text{the number of relevant articles of query i-th}} \]
\[F2\raisebox{-0.7ex}{\footnotesize{i-th}} = \frac{\text{5*Precision\raisebox{-0.7ex}{\footnotesize{i-th}}*Recall\raisebox{-0.7ex}{\footnotesize{i-th}}}}{\text{4*Precision\raisebox{-0.7ex}{\footnotesize{i-th}}+Recall\raisebox{-0.7ex}{\footnotesize{i-th}}}} \]
\[F2= \text{average of (F2\raisebox{-0.7ex}{\footnotesize{i-th}})} \]
\subsection{Results of BM25+ and sentence-transformer}
The models we selected can only compute scores between query-rule pairs. The performance of the model on F2 will decrease if we choose too many articles. In order to get good results on the F2 measure, we have to set a suitable threshold. Let the highest score between each query and all articles as highest\_score. The articles would be chosen if their scores in range [highest\_score-threshold, highest\_score]. For short, we rewrite sentence-transformer Condenser trained from PhoBERT-base and PhoBERT-large respectively SConPBB and SConPBL in the below table.
\begin{table}
\caption{Results of BM25+ and sentence-transformer}
\label{single model}
\centering
\begin{tabular}{|l|l|l|l|c|l|c|}
\hline
\textbf{Model} & \textbf{Recall@20} & \textbf{F2} \\
\hline
BM25+ & 0.557 & 0.221 \\
\hline
SPhoBERT-base(syl) & 0.944 & 0.700\\
\hline
SConPBB(syl) & 0.953 & 0.697 \\
\hline
SPhoBERT-large(syl) & 0.940 & 0.715\\
\hline
SConPBL(syl) & 0.953 & 0.719 \\
\hline
SViBERT(syl) & 0.942 & 0.679\\
\hline
SPhoBERT-base(word) & 0.954 & 0.721\\
\hline
SConPBB(word) & 0.957 & 0.704 \\
\hline
SPhoBERT-large(word) & 0.954 & 0.727\\
\hline
SConPBL(word) & 0.954 & 0.724 \\
\hline
SViBERT(word) & 0.921 & 0.669\\
\hline
\end{tabular}
\end{table}
The result in Table \ref{single model} shows that: sentence-transformer models outperform BM25+. However, we want to build a retrieval system with both lexical matching and semantic searching approaches. We consider the combining score of these two approaches. More than 95\% of questions have one to three relevant articles. We visualized the average score of the top 3 highest scores of BM25+.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.7]{figures/bm25_top3.png}
\caption{Average score of top 3 highest scores of BM25+}
\label{fig:top3_BM25_avg}
\end{figure}
\subsection{Retrieval system using BM25+ and sentence-transformer}
Sentence-transformer uses the cosine similarity(cos\_sim) of the two vectors query and article to rank.Cos\_sim is always in range [-1,1]. According to Figure \ref{fig:top3_BM25_avg} average score of BM25+ is much larger than cos\_sim while semantic approach makes a big gap. Combining score for ranking in previous work like (1- $\alpha $)*lexical\_score + $\alpha$ * semantic\_score will not work or take much time to fine the $\alpha$. We proposed two combining score: sqrt(BM25\_score)*cos\_sim and BM25\_score*cos\_sim to rank our retrieval systems. Similar to single model, we used suitable threshold for retrieval system to get best results on F2. For short, we named the retrieval system by a name of sentence-transformer model within it.
\begin{table}[!h]
\caption{Results of retrieval systems }
\label{retrieval_sys}
\centering
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
&\multicolumn{2}{| c |}{sqrt(BM25\_score)*cos\_sim} &\multicolumn{2}{| c |}{BM25\_score*cos\_sim}\\
\hline
\textbf{System} & \textbf{Recall@20} & \textbf{F2} & \textbf{Recall@20} & \textbf{F2} \\
\hline
SPhoBERT-base(syl) & 0.958 & 0.715 & 0.960 & 0.703\\
\hline
SConPBB(syl) & 0.967 & 0.712 & 0.965 & 0.692 \\
\hline
SPhoBERT-large(syl) & 0.951 & 0.726 & 0.955 & 0.706\\
\hline
SConPBL(syl) & 0.953 & 0.729 & 0.956 & 0.710 \\
\hline
SViBERT(syl) & 0.951 & 0.695 & 0.963 & 0.684\\
\hline
\hline
SPhoBERT-base(word) & 0.963 & 0.725 & 0.964 & 0.706\\
\hline
SConPBB(word) & 0.962 & 0.719 & 0.964 & 0.700 \\
\hline
SPhoBERT-large(word) & 0.967 & \textbf{0.741} & \textbf{0.970} & 0.721\\
\hline
SConPBL(word) & 0.960 & 0.738 & 0.967 & 0.718 \\
\hline
SViBERT(word) & 0.948 & 0.683 & 0.950 & 0.661\\
\hline
\end{tabular}
\end{table}
Based on the results from Tables \ref{single model}, \ref{retrieval_sys}, we had three observations:
\begin{itemize}
\item If LMs were pre-trained on word-level data like PhoBERT and CondenserPhoBERT, trained sentence-transformer models on word-level data bring better results.
\item If LM were pre-trained on syllable-level data like ViBERT, trained sentence-transformer model on syllable-level data bring better results.
\item The system with both lexical matching and semantic searching performances better than sentence-transformer model within it.
\item Sqrt(BM25\_score)*cos\_sim is the best ranking score for the system on F2 while BM25\_score*cos\_sim is the best on recall@20.
\end{itemize}
\subsection{Comparison with previous Vietnamese legal text retrieval systems}
Our best performing system on F2 is SPhoBERT-large(word) system with ranking score: sqrt(BM25\_score)*cos\_sim. On Recall@20, our best performing system is SPhoBERT-large(word) system with ranking score: BM25\_score*cos\_sim. For short, we call our best retrieval system on F2 is our best system 1 and our best retrieval system on Recall@20 is our best system 2.
\begin{table}
\parbox{.45\linewidth}{
\centering
\caption{Comparison with previous Vietnamese legal text retrieval systems on F2}
\label{CompareF2}
\begin{tabular}{|l|c|}
\hline
\textbf{System} & \textbf{F2} \\
\hline
Our best system 1 & \textbf{0.741}\\
\hline
Attentive CNN \cite{nguyen2022toward} & 0.4774 \\
\hline
XLM-RoBERTa \cite{nguyen2022toward} & 0.2006\\
\hline
\end{tabular}
}
\hfill
\parbox{.45\linewidth}{
\centering
\caption{Comparison with previous Vietnamese legal text retrieval systems on Recall@20}
\label{CompareRecall@20}
\begin{tabular}{|p{3.5cm}|c|}
\hline
\textbf{System} & \textbf{Recall@20} \\
\hline
Our best system 2 & \textbf{0.970}\\
\hline
ElasticSearch + Attentive CNN \cite{kien-etal-2020-answering} & 0.825 \\
\hline
Birch(256 first words) \cite{kien-etal-2020-answering} & 0.763\\
\hline
Birch(title) \cite{kien-etal-2020-answering} & 0.783\\
\hline
\end{tabular}
}
\end{table}
Table \ref{CompareF2} and Table \ref{CompareRecall@20} illustrate the results of our best systems compared with the previous systems \cite{kien-etal-2020-answering,nguyen2022toward}. It is proved that our system yields the best performance on both metrics.
\subsection{Analysis by the number of relevant articles related to queries}
We used best retrieval system from each sentence-transformer model to analysis effect of the number of relevant articles to queries in both metrics.
\begin{figure}[!h]
\raggedright
\centering
\includegraphics[scale=0.45]{figures/F2_num_art.png
\caption{Effect of the number of relevant articles to queries on F2}
\label{F2_num}
\end{figure}
\begin{figure}[!h]
\raggedright
\centering
\includegraphics[scale=0.45]{figures/recall20_num_art.png
\caption{Effect of the number of relevant articles to queries on recall@20}
\label{recall_num}
\end{figure}
Our observations based on the results from Figures \ref{F2_num}, \ref{recall_num}:
\begin{itemize}
\item The more relevant articles relate to the query, the lower F2 of the systems there. Part of the reason for this is because we set the threshold pretty low. So the range [highest\_score-threshold, highest\_score] would not be large enough to cover many the relevant articles.
\item Queries with (3,5,7) relevant articles make the systems with recall@20 significantly lower than queries with (2,4,6) relevant articles. Most of the recall@20 results of systems are pulled down because of queries with (2,4,6) relevant articles.
\item SViBERT system achieves much better results than other systems on the recall@20 measure in queries with (5,6) relevant articles.
\item SPhoBERT-large achieved the best results because it outperformed other systems in queries with (3,5,7) relevant articles.
\end{itemize}
\subsection{Data-driven error analysis}
There are two main errors we found during the experiment.
First, wrong crawling relevant articles for query in the dataset. After a thorough discussion, we all agreed that the answer to this question was wrong. The example in Table \ref{tab:wrong_crawling} shows that relevant article was wrongly collected. The question is about person who is obliged to execute the judgment returns the non-implementation papers while relevant article is about principles of auction of land use rights. We did not find any correlation between the question and the answer.
\begin{table}
\caption{Example of wrong crawling relevant articles}
\centering
\begin{tabularx}{\textwidth}{|>{\hsize=.15\hsize\centering\arraybackslash}X|X|>{\hsize=.1\hsize\centering\arraybackslash}X|}
\cline{1-2}
\textbf Q/A/AC & \multicolumn{1}{c|}{\textbf{Content}}\\
\cline{1-2}
Query(Q) & Người có nghĩa vụ phải thi hành án trả giấy tờ không thực hiện thì cơ quan thi hành án dân sự cưỡng chế như thế nào ? \textit({If the person who is obliged to execute the judgment returns the non-implementation papers, how does the civil judgment enforcement agency coerce ?)}\\
\cline{1-2}
Answer(A) & Article 117 in law id:45/2013/QH13 \\
\cline{1-2}
Article content (AC) & Nguyên tắc đấu giá quyền sử dụng đất \textit({Principles of auction of land use rights)} \newline 1 . Đấu giá quyền sử dụng đất được thực hiện công khai , liên tục , khách quan , trung thực , bình đẳng , bảo vệ quyền và lợi ích hợp pháp của các bên tham gia. \textit({Auction of land use rights is conducted publicly, continuously, objectively, honestly, equally, protecting the legitimate rights and interests of the participating parties)} \newline 2. Việc đấu giá quyền sử dụng đất phải đúng trình tự , thủ tục theo quy định của pháp luật về đất đai và pháp luật về đấu giá tài sản\textit({The auction of land use rights must comply with the order and procedures prescribed by the law on land and the law on asset auction.)} \\
\cline{1-2}
\end{tabularx}
\label{tab:wrong_crawling}
\end{table}
Second, very hard semantic queries and relevant articles. These queries often miss-matching lexical with their relevant articles. Both of them are also long and contain difficult specialized words. To be able to understand the semantic of these queries and articles will still be a big challenge for retrieval systems.
\begin{table}[!h]
\centering
\caption{Example of very hard semantic query and relevant articles}
\label{tab:hard_semantic}
\begin{tabularx}{\textwidth}{|>{\hsize=.15\hsize\centering\arraybackslash}X|X|>{\hsize=.1\hsize\centering\arraybackslash}X|}
\cline{1-2}
\textbf Q/A/AC & \multicolumn{1}{c|}{\textbf{Content}}\\
\cline{1-2}
Query(Q) & Có thể kê biên tài sản riêng của giám đốc để đảm bảo trả nợ cho công ty không ? \textit({Can the director's personal assets be distraint to ensure the repayment of the company's debt ?)}\\
\cline{1-2}
Answer(A) & Article 74 in law id:91/2015/QH13 \\
\cline{1-2}
Article content (AC) & Pháp nhân \newline 1. Một tổ chức được công nhận là pháp nhân khi có đủ các điều kiện sau đây \textit{(An organization is recognized as a "pháp nhân" when the following conditions are satisfied)} a ) Được thành lập theo quy định của Bộ Luật này , luật khác có liên quan; \textit({a) Established under the provisions of this law and other relevant laws;)} b ) Có cơ cấu tổ chức theo quy định tại Điều 83 của bộ luật này;\textit{(b) Having an organizational structure as prescribed in Article 83 of this law;)} c) Có tài sản độc lập với cá nhân , pháp nhân khác và tự chịu trách nhiệm bằng tài sản của mình ; \textit{(c) Having assets independent of other individuals or legal entities and taking responsibility for their own property;)} \textbf{c) Có tài sản độc lập với cá nhân , pháp nhân khác và tự chịu trách nhiệm bằng tài sản của mình ; \textit{(c) Having assets independent of other individuals or legal entities and taking responsibility for their own property;)}} d ) Nhân danh mình tham gia quan hệ pháp luật một cách độc lập. \textit{(d) Independently participate in legal relations in their own name.)} \newline 2 . Mọi cá nhân , pháp nhân đều có quyền thành lập pháp nhân , trừ trường hợp luật có quy định khác . \textit{(2 . Every individual and juridical person has the right to establish a juridical person, unless otherwise provided for by law.)}\\
\cline{1-2}
\end{tabularx}
\end{table}
In Table \ref{tab:hard_semantic}, "pháp nhân" is a specialized word in Vietnamese legal. To the best of our knowledge, the word has the closest meaning to it in English is corporation. We found the answer for the query in clause c) of this article. This article and query are lexical mismatching, so retrieval system will not work as well as sentence-transformer model. We used SPhoBERT-large which is the best semantic searching model to evalute cos\_sim between question and clause c) only. The result is 0.351. This shows that our model is still limited in the face of questions and answers with high semantic difficulty.
\section{Conclusion and future work}
In this paper, we proposed multi-stage information retrieval based on sentence-transformers for Vietnamese legal texts. We also compare the best performance model to the Attentive CNN model (previous SOTA in this QA dataset). The results indicate that our models outperform previous SOTA in both evaluation metrics.
In future work, we will try other techniques for long articles like summary, and keyword extraction. We will also improve the models to learn the to learn the relationship between queries and rules of very high semantic complexity.
\bibliographystyle{splncs04}
| 4c850bdaafe04adb46eb23cb849aab802ae9da3f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Quantum interference is one of the most vivid illustrations of the wave nature of quantum mechanical systems. It shows up whenever multiple paths of different phases or multiple energy levels are present, from double slit experiments and weak localization to Landau-Zener interference. Quantum interference also underlies many important applications, whether it is atomic clocks \cite{Degen2017} or Shor's factorization algorithm \cite{Shor1994, Ekert1996, Shor1997}, and would conceivably be a key feature to many future quantum coherent devices.
One example of the consequences of quantum interference is coherent population trapping (CPT), as a result of destructive interference between different transition paths, and first observed in a three-level atom in an optical experiment \cite{Gray1978}.
In such a three-level system, two states are coupled to a third, intermediate, state. When the driving fields are properly detuned for the two allowed transitions, a superposition of the first two states emerges and is decoupled from the intermediate states.
Such a superposition is called a ``dark state'' since a system trapped in this state would not respond to the probe field, leading to interesting phenomena such as electromagnetically induced transparency \cite{Harris1990}.
By adiabatically tuning the controls of the dark state (such as the amplitude and phase of the driving fields), one can perform rapid state initialization and Stimulated Raman Adiabatic Passage (STIRAP) with suppressed excitation \cite{Xu2007, Rogers2014, Gaubatz1990, Vitanov2017}.
Compared to state manipulation based on resonant driving, STIRAP has significant advantages in robustness, while maintaining high degree of efficiency and selectivity, and is thus of great importance in quantum information processing \cite{Gaubatz1990, Vitanov2017}.
CPT has been extensively studied in various physical systems since its first observation, and widely employed in both metrology and quantum engineering, from atomic cooling \cite{Aspect1988} and atomic clock \cite{Vanier2005}, to high-sensitivity magnetometry \cite{Scully1992, Nagel1998} and quantum state manipulation \cite{Gaubatz1990, Rogers2014, Vitanov2017}. In condensed matter systems, CPT has been demonstrated in doped crystals \cite{Goto2007, Klein2008}, ultracold atoms \cite{Takekoshi2014, Molony2014}, color centers in diamond \cite{Golter2014}, superconducting circuits \cite{Kelly2010, Xu2016, Kumar2016}, microelectromechanical systems \cite{Knappe2008} and self-assembled quantum dots \cite{Xu2008, Brunner2009}.
In contrast, in electrically-controlled gate-defined quantum dots (QDs), which are viewed as one of the promising platforms for quantum computing \cite{Hendrickx2021, Adam2022, Noiri2022, Xue2022, Zwerver2022, Camenzind2022}, CPT has received less attention. This is due to differences in energy structure and the way electromagnetic field couples to the system, even though QDs, particularly double quantum dots (DQD), provide intriguing possibilities for CPT.
For example, in an atom, the optical field participates in CPT in two aspects, to couple the energy levels, and to serve as a probe.
On the other hand, DQD energy levels are tunable, and its state can be readout via electronic methods, so that CPT without driving fields and CPT measurement without a probe field are both possible.
In this letter, we demonstrate CPT in a gate-defined DQD without drive or probe field, and also with one longitudinal drive.
We focus on a singlet-triplet (ST) system \cite{Petta2005, Pioro-Ladriere2008, Foletti2009,Huang2019, Hendrickx2020} in a DQD device for holes, and measure the leakage current through the DQD in the Pauli Spin Blockade (PSB) regime \cite{Kouwenhoven1997, Ono2002, Petta2005}. In the absence of driving, we observe a sharp dip in the leakage current at zero bias, which can be attributed to the formation of dark states and the occurrence of CPT.
When the ST system is driven longitudinally through detuning $\varepsilon$, we again observe CPT under proper conditions, and the physical picture can be explained using an effective Hamiltonian \cite{Bukov2015, Oka2019, Weitenberg2021, Zhou2022} for this longitudinally driven system (with distinctive features compared to the transversely driven atomic systems, such as the modulation of the effective couplings and the observation of the odd-even effect \cite{Stehlik2014, Danon2014, Zhou2022}). Importantly, with controls provided by the driving, we show that STIRAP is feasible.
\begin{figure}[tbp]
\centering
\includegraphics[width=8.6cm]{fig1.pdf}
\caption{{\bf Scanning electron micrograph of the device and transport current with respect to gate voltages.}
(a) Scanning electron micrograph of a device of the same structure with the one used in experiment. The nanowire is outlined by the dashed box, of which the cross-section is sketched in the inset. QDs are formed when voltages are applied appropriately.
The device is exposed to a uniform magnetic field, and microwave drive is applied to gate $U_2$.
(b) Transport current as a function of gate voltages. The dashed box indicates the current suppression caused by PSB, and detuning $\varepsilon$ is defined accordingly, as shown by the arrow.
(c) Energy structure of ST system as a function of $\varepsilon$, where the anticrossings suggest the couplings between states.
(d) Illustration for establishing the transport current. The tunneling events are indicated by the gray dashed arrows, which represent the tunnel of holes in and out of QDs.
Black solid arrows labeled by $t$ or $t_{so}$ are for the couplings between relative states.
We assume $\Gamma\gg t,~t_{so}$, and treat the process of jumping in and out by jumping operators $|s_z s_z\rangle\langle S_{20}|$ with $|s_z\rangle = |\uparrow\rangle, |\downarrow\rangle$.
}
\label{fig-1}
\end{figure}
Our sample is fabricated in a Ge hut nanowire \cite{Vukusic2017, Li2017, Li2018, Vukusic2018, Gao2020, Xu2020a, Xu2020, Zhang2021, Liu2022}.
The DQD is defined in the nanowire by depositing Aluminum electrodes above, which control the numbers of holes in each dot and the coupling strength in between. Figure \ref{fig-1} (a) shows a scanning electron micrograph of the device, where the black dashed box indicates the location of the nanowire.
Sketch of the cross-section along the nanowire is presented in the inset, where the DQD potential, from interdot detuning to tunnel barriers, is tuned by gate voltages $U_i$. In addition, microwave drive is applied to $U_2$ gate when we study driven dynamics.
A uniform magnetic field is applied perpendicular to the nanowire, as shown in Fig.~\ref{fig-1} (a). The device is cooled in a dilution refrigerator with a base temperature of $200$ mK.
We perform transport measurement through the DQD. Figure \ref{fig-1} (b) shows the current as a function of gate voltages $U_2$ and $U_4$, which determine the energy levels and thus the occupation of the two dots. When gate voltages are set in the region outlined by the black dashed trapezoid in Fig.~\ref{fig-1} (b), we observe PSB, when current is blocked due to spin configurations of the holes instead of Coulomb interaction \cite{Kouwenhoven1997, Ono2002, OnoPRL2004, Petta2005}. In Fig.~\ref{fig-1} (b) we define the interdot detuning $\varepsilon$ along the black solid arrow. When $\varepsilon$ is tuned deeper into negative beyond the PSB regime, excited orbital states are involved in the transition and lift the PSB, leading to a jump in current. Based on the width of the PSB region in Fig.~\ref{fig-1} (b), we estimate the magnitude of the excitation energy of the DQD to be $\sim 1 \text{meV}$.
While our DQD is still in the multi-hole regime, in the PSB regime the system is well described by an effective two-hole model near the $(N+2,M+0)$ to $(N+1,M+1)$ charge transition, where $N$ and $M$ refer to the nominal number of core holes in the left and right QDs that do not participate in transport. More specifically, the low-energy DQD Hamiltonian can be expanded in the basis of five effective two-hole states
$\{|\uparrow\uparrow\rangle,~
|\uparrow\downarrow\rangle,~
|\downarrow\uparrow\rangle,~
|\downarrow\downarrow\rangle,~
|S_{20}\rangle\}$.
Here $|S_{20}\rangle$ denotes a singlet state $|S\rangle = (|\uparrow\downarrow\rangle-|\downarrow\uparrow\rangle)/\sqrt2$ with a charge configuration of $(N+2,M+0)$, while the charge configuration of the other four states has the two valence holes evenly distributed between the dots as $(N+1,M+1)$.
The Hamiltonian in this basis is given by
\begin{equation}
H =
\begin{pmatrix}
\bar E_z &0 &0 &0 & t_{so}\\
0& \delta E_z & 0& 0& t\\
0& 0& -\delta E_z & 0& -t\\
0& 0& 0& -\bar E_z & t_{so}\\
t_{so} & t & -t & t_{so}& \varepsilon
\end{pmatrix},
\label{eq-H}
\end{equation}
where $\varepsilon$ is the detuning defined in Fig.~\ref{fig-1} (b), $\bar E_z=(g_1+g_2)\mu_B B/2$, $\delta E_z=(g_1-g_2)\mu_B B/2$, with $g_{1,2}$ the g-factors of the hole spins in the two QDs, $\mu_B$ the Bohr magneton and $B$ the external magnetic field strength.
Spin-flip tunneling ($t_{so}$) couples $|S_{20}\rangle$ to $|T_\pm\rangle$, while $t$ gives the spin-preserved inter-dot tunnel coupling strength.
Figure \ref{fig-1} (c) presents the low-energy spectrum as a function of $\varepsilon$ when $B\ne 0$, and Fig.~\ref{fig-1} (d) gives a schematic illustration of possible tunneling events between source/drain and QDs. Notice that in general $t_{so} \ll t$, even for a hole system like ours, as can be qualitatively seen in the contrast of the current amplitude between the PSB regime and the non-PSB-non-Coulomb-blockade regime shown in Fig.~\ref{fig-1} (b). Nevertheless, the presence of $t_{so}$ means that in our DQD, PSB is not complete, and a finite leakage current is always present through the DQD \cite{Ono2002, OnoPRL2004}, as shown by the elevated current in the PSB regime compared to the surrounding Coulomb-blocked areas in Fig.~\ref{fig-1} (b). In the rest of this study, our focus will be on how the leakage current in the PSB regime varies with system parameters and possible driving.
The form of the ST Hamiltonian $H$ is reminiscent of the one for the original observation of CPT \cite{Gray1978}. In particular, when $B=0$, the system should have three dark states that are orthogonal to $|S_{20}\rangle$, the final state for the transitions here:
$(|\uparrow\uparrow\rangle-|\downarrow\downarrow\rangle)/\sqrt2$,
$(|\uparrow\downarrow\rangle+|\downarrow\uparrow\rangle)/\sqrt2$
and $(\sin\theta|\uparrow\uparrow\rangle-\cos\theta|\uparrow\downarrow\rangle+\cos\theta|\downarrow\uparrow\rangle + \sin\theta|\downarrow\downarrow\rangle)/2$, where $\theta=\arctan(t/t_{so})$. With these states all in the $(N+1,M+1)$ charge configuration decoupled from the $|S_{20}\rangle$ state in the $(N+2, M)$ configuration, PSB should be strengthened and the leakage current should be suppressed.
In other words, CPT in our DQD should manifest itself as a suppression in the leakage current.
\begin{figure}[tbp]
\centering
\includegraphics[width=8.6cm]{fig2.pdf}
\caption{{\bf CPT in ST system without microwave drive.}
(a) Transport current as a function of strength of magnetic field $B$ and detuning (controlled by $U_4$).
An abnormal current suppression is observed in the vicinity of $B=0$, which can be attributed to CPT.
In addition, the zoom-in of the area in the green solid box is shown in the inset, where an apparent dip can be observed.
(b) Cross-section in the position of $\varepsilon=0$, as indicated by black dashed line in (a).
Measured current from experiment is represented via scatters, while the red solid curve is for theory fitting.
The experiment result is well captured by the fitting, from which we estimate the parameters of the system as presented in the main text.
}
\label{fig-2}
\end{figure}
To experimentally demonstrate the possible existence of these dark states and therefore CPT, we measured transport current through the DQD versus the applied field $B$ and interdot detuning $\varepsilon$ (controlled by $U_4$), as is shown in Fig.~\ref{fig-2} (a).
This method has been used to characterize mechanisms leading to leakage current through PSB, such as hyperfine interaction, spin-orbit coupling \cite{Nadj-Perge2012, Zhang2021}, \textit{etc.}.
In a significant deviation from previous results, we observe an abrupt and narrow current suppression at zero field, as is shown in Fig.~\ref{fig-2} (a).
The cross-section at $\varepsilon=0$ is measured repeatedly, and the result is shown in Fig.~\ref{fig-2} (b).
The signal is clearly a narrow sharp dip on a broad peak (instead of a superposition of two peaks), which is a distinct feature of CPT.
To understand our observation, we have carried out a detailed theoretical analysis. Specifically, we model sequential tunneling through the DQD by operator
$T = i\sum_{s_z,s_z'}|s_zs_z'\rangle\langle S_{20}|+h.c.$, where the summation runs over all possible states in the $|(1, 1)\rangle$ configuration, and $h.c.$ stands for Hermitian conjugate \cite{Stoof1996}.
The transport current is then evaluated as $I \propto \text{tr}(T\rho)$, where $\rho$ is the steady-state density matrix of the DQD, obtained from the Lindblad master equation (LME)
\begin{equation}
\frac{d\rho}{dt} = -\frac{i}{\hbar}[H, \rho] + \sum_k \left(L_k\rho L_k^\dagger - {1\over2}\{L^\dagger_kL_k, \rho\}\right),
\label{eq-LME}
\end{equation}
with $L_k$ the Lindblad operators representing relaxation and dephasing in the DQD. In our device, the ST system undergoes a transport cycle of $|S_{20}\rangle\to|(1, 0)\rangle\to|(1, 1)\rangle\to|S_{20}\rangle$ as indicated by the arrows in Fig.~\ref{fig-1} (d). In practice, the dot-reservoir tunneling rate labeled by $\Gamma$ in Fig.~\ref{fig-1} (d) is in the order of $10$ GHz, which is significantly greater than the interdot tunnel coupling strength $t$ and $t_{so}$. The steady state population of $|S_{20}\rangle$ is thus very small, since it would rapidly relax to $|(1,0)\rangle$ and the DQD would then quickly reload into one of the $|(1,1)\rangle$ states. Under this condition, a hole jumping out of the left dot and another jumping into the right dot happen almost simultaneously, with the whole reloading process captured by the Lindblad operator $L_{s_z s_z'}^r=\sqrt\Gamma|s_zs_z'\rangle\langle S_{20}|$, with $\Gamma$ the dot-reservoir tunneling rate and $s_z^{(\prime)} = \uparrow,~\downarrow$. In addition to this relaxation process, we also consider dephasing of all the energy levels, depicted by the Lindblad operator $L_{i}^d = \sqrt\gamma|i\rangle\langle i|$ with $i$ covering all the basis states of $H$.
The steady-state of the system can be calculated straightforwardly by vectorizing the LME (see Appendix Sec. 1).
To estimate the parameters of our device, we calculate the mean squared error $\text{MSE} = \frac{1}{N}\sum_{i=1}^N|I_i^e-I_i^s|^2$, where the summation runs over all experimental measurements, with $I^e$ and $I^s$ referring to experimental data and numerical simulations, respectively. Experimental parameters are then estimated by searching for the minimal MSE in a reasonable regime. Simulation results are shown in Fig.~\ref{fig-2} (b) by the solid line, which agrees very well with the experiment data. From our simulation results, we estimate the following parameters for our device: $\Gamma\approx 27.5$ GHz, $\gamma \approx 0.655$ GHz, $t_{so}/\hbar\approx 5.32$ GHz and $t/\hbar \approx 2.09$ GHz \cite{Xu2020, Zhang2021, Liu2022}.
\begin{figure}[]
\centering
\includegraphics[width=8.4cm]{fig3.pdf}
\caption{{\bf Odd-even effect and CPT in ST system with longitudinal driving.}
Transport current as a function of driving frequency and strength of magnetic field is presented, in which enhancement or suppression in transport current can be observed for different harmonics $E_i-E_j=nhf$. Both first- and second-order harmonics are observed in our experiment, the spectrum is relevant to process labeled in the left inset.
Zoom-in of the area in green solid box is presented in the right inset, in which we see the peaks caused by first-order single-spin rotation and dip caused by second-order two-spin rotation.
In addition, the dip in the vicinity of $B\approx 0$ can be attributed to CPT, the width of which is modulated by driving filed.
}
\label{fig-3}
\end{figure}
Our experimental measurements, supported by numerical simulations, demonstrate that coherent population trapping can indeed occur in a DQD even without microwave driving. However, this observed CPT has limited utility due to the degeneracy among the dark states in the five-level system.
On the other hand, a recent theoretical study \cite{Zhou2022} has shown that a longitudinally driven (at angular frequency $\omega$) ST system can be described by an effective Hamiltonian of a similar structure as conventional CPT systems, near the harmonic resonance condition $E_i-E_j\approx n\hbar\omega$, where $E_{i,j}$ is the energies of the $|(1,1)\rangle$ states. In such a driven ST system, the resonant driving field allows us to selectively create individual dark states and avoid degeneracy.
This driven CPT could then enable useful applications such as STIRAP. Moreover, the effective coupling is modulated by driving field, further increasing the tunability of the ST system. According to our theoretical investigation, such modulation is reflected in two aspects. Firstly, the modulation cased by a sinusoidal drive leads to an odd-even effect \cite{Stehlik2014, Danon2014, Zhou2022}, i.e., for odd (even) orders of harmonics, the resonance signals are measured as a peak (dip) (see Appendix Sec. 2). Secondly, the modulation on the effective coupling also inevitably leads to the CPT being modulated at the same time.
We have performed our transport measurement with a microwave applied to the $U_2$ gate, longitudinally driving the interdot detuning of the DQD. In Fig.~\ref{fig-3}, the transport current at $\varepsilon=0$ as a function of driving frequency and external magnetic field strength is presented. Here we clearly observe current enhancement and suppression caused by the various resonances, indicated by the arrows in the inset of Fig.~\ref{fig-3}. These resonances include single-spin rotations labeled by the relevant single-dot g-factor $g_{1,2}$, and a two-spin rotation labeled by $(g_1+g_2) / 2$.
From the spectrum, we extract g-factors of the spins in the left and right dot as $g_1 \approx 4.5$ and $g_2 \approx 3.0$.
Our observation here clearly indicates the presence of the odd-even effect for single-spin rotations: for $n = 1$, a resonance yields an enhancement in current, while for $n = 2$, the current is suppressed.
\begin{figure}[]
\centering
\includegraphics[width=8.6cm]{fig4.pdf}
\caption{{\bf Modulation of CPT by driving field.}
(a) Experiment data in the vicinity of $B=0$ mT (left panel) and numerical simulation (right panel). The dip in the middle indicates the modulation on CPT introduced by the longitudinal drive.
(b) The half width at half maximum $w$ is extracted by fitting with a Lorentzian function.
$w$ as a function of driving frequency is then presented, with an error bar given by $ 95\%$ confidence interval of the fitting.
The result fitted from LME ($H_\text{eff}$) is presented by the red (black) solid line, which agrees with the experiment fitting.
}
\label{fig-4}
\end{figure}
In addition to the odd-even effect, we also observe a clear dip in current caused by CPT in the vicinity of $B=0$ in Fig.~\ref{fig-3}.
For a driven ST system, CPT can be observed whenever $E_i-E_j = n\hbar\omega$ while $E_i - E_{S} = m\hbar\omega$, where the subscript $i,~j$ represent $(1,1)$ states and $S$ refers to $|S_{20}\rangle$.
This condition can be fulfilled when $\varepsilon =0$ and $B=0$, where a suppression in current is observed as presented in the left panel of Fig.~\ref{fig-4} (a).
We also perform a numerical simulation based on the time-dependent LME with a driving amplitude $A/\hbar = 190$ GHz, and the result is presented in the right panel of Fig.~\ref{fig-4} (a). A consequence of the longitudinal driving we employed is a modulation of the effective couplings when varying the driving frequency. To reveal this modulation, we extract the half-width-at-half-maximum $w$ as a function of driving frequency $f$ by fitting the current dip at $B = 0$ with a Lorentzian function.
The fitted result is displayed in Fig.~\ref{fig-4} (b) by scatters, where the error bar is given by the $95\%$ confidence interval of the fitting.
Results from the LME and $H_\text{eff}$ are also presented in Fig.~\ref{fig-4} (b) by red and black solid curves.
The $H_\text{eff}$ used here is accurate up to $\mathcal{O}(t_{(so)}J_\nu/\omega)$, in which the coupling strengths are replaced by the modulated ones, i.e., $t_{(so)}\to t_{(so)}J_0(A/\omega)$ (See Appendix Sec. 3).
Figure \ref{fig-4} (b) shows that results from both LME and $H_\text{eff}$ are qualitatively consistent with the experiment data, though $H_\text{eff}$ can be non-convergent when $J_0(A/\omega)=0$ due to the approximations used in its derivation.
The agreement here shows that the clear modulation on CPT originates from the longitudinal driving. Such a modulation could enable possible applications such as STIRAP in similar systems \cite{Zhou2022}.
It is worth noting that, both odd-even effect and CPT yields a suppression in current, though they are fundamentally different.
Odd-even effect is a direct result of the modulated effective coupling in the form of Bessel function, and is a coherent destruction of tunneling between $|(1,1)\rangle$ and $|S_{20}\rangle$.
In contrast, CPT does not require modulation of the effective coupling. Instead, it requires that all the relevant energy levels are on resonance, i.e. all diagonal terms in $H_\text{eff}$ are equal, leading to a coherent interference involving all levels.
In summary, we report experimental observation of CPT without driving field based on a five-level ST system for holes in a nanowire device.
The CPT manifests itself as a sharp suppression in the leakage current in the Pauli Spin Blockade regime, and is due to the presence of multiple dark states under certain conditions.
By applying a longitudinal drive, we have also demonstrated the possibility of selectively creating a dark state and the associated CPT, with the added benefit of an increase in tunability for the effective couplings. The longitudinal driving also leads to other interesting features to our experimental observations,
from an odd-even effect in the tunnel current to a CPT modulated by the varying driving field frequency.
Our results clearly demonstrate the potential tunability of a longitudinally driven system \cite{Zhou2022}, open up the possibility for STIRAP based quantum gates, and broaden the capacity of the ST system in quantum simulation and quantum computation applications.
\section{acknowledgment}
This work is supported the National Natural Science Foundation of China (Grants Nos. 12074368, 92165207, 12034018 and 61922074), the Anhui Province Natural Science Foundation (Grant No.2108085J03), and the USTC Tang Scholarship. X.H. acknowledges financial support by U.S. ARO through Grant No. W911NF1710257, and this work was partially carried out at the USTC Center for Micro and Nanoscale Research and Fabrication.
| 5ce3b067074287da06bce1ac065b9fe8290a9afd | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The sharing of medical datasets is essential in enabling the cross-hospital flow of medical information and improving the quality of medical services~\cite{kumar2021integration}.
However, sharing healthcare datasets between different hospitals faces several thorny issues.
Firstly, privacy protection has been a severe issue hindering the process when sharing medical image datasets from different hospitals~\cite{kaissis2020secure}.
Second, sharing large-scale high-resolution medical image datasets increases transmission and storage costs~\cite{zhou2021models}.
Therefore, the solution to these problems will significantly promote the development of medical dataset sharing.
\par
Dataset distillation can synthesize a small dataset such that models trained on it achieve comparable performance with the original large dataset~\cite{wang2018datasetdistillation}.
Although dataset distillation has been proposed for distilling some simple datasets, such as MNIST and CIFAR10, its effectiveness in high-resolution complex medical datasets has not yet been proved~\cite{zhao2021datasetcondensation}.
Medical dataset distillation may have potential advantages for solving the existing medical dataset sharing problems~\cite{li2020soft}.
For example, the size of distilled medical image datasets can be significantly compressed, and distilled images generated from noise are automatically anonymized~\cite{dong2022privacy}.
Therefore, it is desirable to explore the potential of dataset distillation for medical dataset sharing and contribute to real-world applications.
\par
In this paper, we propose a novel dataset distillation-based method for medical dataset sharing.
COVID-19 and its variants have rapidly spread worldwide, influencing the health and life of billions of people~\cite{mofijur2021impact}.
X-ray is widely used in clinical because of its high speed and low cost.
Detecting COVID-19 from chest X-ray images is perhaps one of the fastest and easiest ways~\cite{minaee2020deep}.
However, sharing COVID-19 datasets between different hospitals also has the above-mentioned problems.
We perform experiments on a COVID-19 chest X-ray image dataset to prove the effectiveness of the proposed method.
Experimental results show that we can achieve high COVID-19 detection performance even using scarce anonymized chest X-ray images, hopeful of solving existing medical dataset sharing problems.
The concept of this study is shown in Figure~\ref{fig1}.
\begin{figure}[t]
\centering
\includegraphics[width=13.5cm]{Image/Concept1.png}
\caption{The concept of this study.}
\label{fig1}
\end{figure}
\section{Methodology}
\begin{table}[t]
\centering
\caption{COVID-19 detection accuracy when using different numbers of distilled images. IPC denotes images per class.}
\label{tab1}
\begin{tabular}{lcccccc}
\hline
IPC & 1 & 2 & 3 & 5 & 10 & 20 \\\hline
Accuracy
& 52.5\% & 76.4\% & 77.0\% & 79.3\% & 82.2\% & \bfseries{82.7\%} \\\hline
\end{tabular}
\end{table}
The objective of our method is to have the parameters of the student network trained on the distilled dataset match the parameters of the teacher networks trained on the original dataset.
Before the distillation process, we first train $T$ teacher networks on the original COVID-19 dataset $\mathcal{D}$ and obtain their parameters~\cite{cazenavette2022dataset}.
These time sequences of parameters $\{\theta_{i}\}^{I}_{0}$ are defined as teacher parameters.
Also, network parameters trained on the distilled dataset $\mathcal{D}_{c}$ at each training step $i$ are defined as student parameters $\tilde{\theta}_{i}$.
Our method aims to distill chest X-ray images that induce network parameters similar to those learned from the original COVID-19 dataset (given the same initial values).
In the distillation process, student parameters are initialized as $\tilde{\theta}_{i}=\theta_{i}$ by sampled from
one of the teacher parameters at a random step $i$.
Then we perform gradient descent updates on the student parameters $\tilde{\theta}$ with respect to the cross-entropy loss $\ell$ of the distilled dataset $\mathcal{D}_{c}$ as follows:
\begin{equation}
\tilde{\theta}_{i+j+1} = \tilde{\theta}_{i+j} - \alpha\nabla\ell(\mathcal{A}(\mathcal{D}_{c});\tilde{\theta}_{i+j}),
\end{equation}
where $j$ and $\alpha$ represent the number of gradient descent updates and the trainable learning rate, respectively.
$\mathcal{A}$ represents a differentiable data augmentation module that can improve the distillation performance, which was proposed in~\cite{zhao2021differentiatble}.
Since the data augmentation used during distillation is differentiable, it can be propagated back through the augmentation layers to the distilled dataset.
Then we get the teacher parameters $\theta_{i+K}$ from $K$ gradient descent updates after the parameters used to initialize the student network.
The final loss $\mathcal{L}$ calculate the normalized $L_{2}$ loss between updated student parameters $\tilde{\theta}_{i+J}$ and teacher parameters $\theta_{i+K}$ as follows:
\begin{equation}
\mathcal{L} = \frac{|| \tilde{\theta}_{i+J}-\theta_{i+K} ||^{2}_{2}} {|| \theta_{i}-\theta_{i+K} ||^{2}_{2}},
\end{equation}
Finally, we minimize the loss $\mathcal{L}$ and backpropagate the gradient through all $J$ updates to the student network for obtaining the optimized distilled dataset $\mathcal{D}^{\ast}_{c}$.
Since the distilled chest X-ray images are generated from noise and have different distribution or visual similarities from the original images, they are automatically anonymized.
After obtaining the distilled dataset $\mathcal{D}^{\ast}_{c}$, we can share it with different hospitals and train neural networks for high-accuracy COVID-19 detection.
\section{Experiments}
\begin{table}[t]
\centering
\caption{COVID-19 detection accuracy of different methods.}
\label{tab2}
\begin{tabular}{lccccccc}
\hline
Method & \bfseries{Ours} & SKD & BYOL & SimSiam & MAE & Transfer & From Scratch \\\hline
Accuracy
& \bfseries{82.7\%} & 74.2\% & 68.3\% & 66.8\% & 62.3\% & 53.9\% & 28.4\% \\\hline
\end{tabular}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=12cm]{Image/Distilled_Images.png}
\caption{Examples of real and distilled images.}
\label{fig2}
\end{figure}
The dataset used in our study has four classes, i.e., COVID-19 (C), Lung Opacity (L), Normal (N), and Viral Pneumonia (V)~\cite{rahman2021exploring}.
The number of images in each class is 3616, 6012, 10192, and 1345, respectively.
The resolution of chest X-ray images is 224 $\times$ 224, and we resized it to 112 $\times$ 112 for distillation or training networks.
The number of pre-trained teacher networks $T$ was set to 100.
And we set the number of distilled images as 1, 2, 3, 5, 10, and 20 images per class.
The network structure used in this study is a simple ConvNet with depth-5 and width-128, which is often used in the dataset distillation task.
For comparative methods, we used several SOTA self-supervised learning methods, including SKD~\cite{li2022self}, BYOL~\cite{grill2020bootstrap}, SimSiam~\cite{chen2021exploring} and MAE~\cite{he2022masked}.
We also used transfer learning from ImageNet~\cite{deng2009imagenet} and training from scratch as baseline methods.
We randomly selected 42 images per class (1\% of the training set) for these comparative methods.
Except for the MAE method used ViT-Large~\cite{dosovitskiy2020vit}, all other methods used ResNet50~\cite{he2022masked} as the backbone network.
\par
The test accuracy of COVID-19 detection are shown in Tables~\ref{tab1} and~\ref{tab2}.
From Table~\ref{tab1}, we can see that the accuracy of our method increased accordingly as the number of distilled images grew.
Table~\ref{tab2} shows that our method achieved high COVID-19 detection accuracy even when using scarce distilled chest X-ray images.
Furthermore, our method drastically outperformed other SOTA methods with a simpler network and fewer training images.
Figure~\ref{fig2} shows some examples of real and distilled images.
We can see that the distilled images are entirely visually different from the original images, which shows the anonymization effectiveness of the proposed method.
\section{Conclusion}
We have proposed a novel dataset distillation-based method for medical dataset sharing.
Since the size of the distilled medical image dataset has been significantly compressed and the images are also anonymized, the sharing of medical datasets between different hospitals will be more efficient and secure.
Experimental results on a COVID-19 chest X-ray image dataset show the advantage of our method compared to other SOTA methods.
\section*{Potential Negative Societal Impact}
The findings of this paper show the effectiveness of dataset distillation for medical dataset sharing.
Although the experimental results are promising, the proposed method should be verified on other medical datasets of different diseases for any potential bias.
Additionally, since the computational overhead of training and storing teacher parameters is relatively high, which may not necessarily be available in low-resource settings.
\section*{Acknowledgement}
This study was supported in part by AMED Grant Number JP21zf0127004, the Hokkaido University-Hitachi Collaborative Education and Research Support Program, and the MEXT Doctoral Program for Data-Related InnoVation Expert Hokkaido University (D-DRIVE-HU) Program.
This study was conducted on the Data Science Computing System of Education and Research Center for Mathematical and Data Science, Hokkaido University.
\bibliographystyle{plainnat}
| 15553e1fafd62c06f3966f5879c542e0b55947a6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Introduction}
The Earth's temperate climate has been maintained through the geochemical carbon cycle in which weathering plays a key role \cite{Walker+1981}.
On the present-day Earth, weathering takes place efficiently on land.
Lands exist on the Earth because the planet has a moderate amount of seawater accounting for 0.023\% of the planet's total mass.
On planets having tens of times more seawater than the Earth, weathering could not work efficiently enough to maintain temperate climates (e.g., refs.\cite{Abbot+2012,Alibert2014,Nakayama+2019}).
Regarding the origin of seawater, a widespread idea is that the Earth's seawater was brought by water-laden or icy planetesimals (e.g., ref.\cite{Genda2016}).
Based on this idea, one would naturally predict a bimodal distribution of planetary water contents, since initial planetesimals are distinctly different in water content between the regions interior and exterior to the snowline in a protoplanetary disc, which was demonstrated by refs.\cite{Tian+Ida2015,Miguel+2019}, as described above.
Alternatively, water can be secondarily produced in a primordial atmosphere of nebular origin through reaction of atmospheric hydrogen with oxidising minerals from the magma ocean, which is formed because of the atmospheric blanketing effect\cite{Ikoma+Genda2006}, thereby enriching the primordial atmosphere with water.
By assuming effective water production, we recently showed that nearly-Earth-mass planets can acquire sufficient amounts of water for their atmospheric vapour to survive in harsh UV environments around pre-main-sequence M stars~\cite{Kimura+Ikoma2020}.
The results suggest that including this water production process significantly affects the predicted water amount distribution of exoplanets in the habitable zone around M dwarfs.
Our planetary population synthesis model, which follows the evolution of planets' masses, radii, and orbits by combining empirical laws for several components of the planet formation process based on the planetesimal accretion hypothesis,
is similar to those of refs.~\cite{Ida+2013,Ida+2018,Miguel+2019} but includes the movement of snowline location due to the thermal evolution of the protoplanetary disc, the accumulation of atmospheres of nebular origin (i.e., the primordial atmospheres) also before the core mass becomes critical, and the effects of water production in the primordial atmospheres\cite{Kimura+Ikoma2020}.
Furthermore we have updated the treatment of some of the processes involved in planet formation/evolution according to improved understandings (See Method).
In this study, we have performed Monte Carlo calculations for $1 \times 10^9$~yr with 10000 different initial conditions for a given set of the input parameters (See Method), and investigated the frequency distribution of water content among the synthesised planets.
\section*{Results}
\subsection*{Planetary mass and semi-major axis distribution}
\begin{figure*}
\centering
\includegraphics{Ma_03M_2.png}
\caption{Effects of water enrichment in the primordial atmospheres on atmospheric accumulation and planetary growth. The two panels show the
planet mass ($M_\mathrm{p}$) vs. semi-major axis ($a_\mathrm{p}$) distributions of the synthesised planets orbiting $0.3M_\odot$ stars at 1~Gyr after the beginning of calculations for the water mass fraction in the primordial atmosphere of (a) $X_{\text{H}_2\text{O}}=0.0$ and (b) $X_{\text{H}_2\text{O}}=0.8$, respectively.
The symbols are colour-coded according to the
atmospheric mass ($M_\text{atm}$) relative to the planet's total mass ($M_\text{p}$).
Note that planets without any atmosphere are shown in black.
The dashed box in each panel shows the region of nearly Earth-mass planets in the present-day habitable zone (HZ-NEMPs; see text for the definition).
}
\label{fig:Ma_M03}
\end{figure*}
Firstly our planetary population synthesis calculations demonstrate that water enrichment in the primordial atmosphere has
a great effect on atmospheric accumulation
of low-mass planets such as sub-Earths and super-Earth (see Fig.~\ref{fig:Ma_M03}).
The overall distributions of planetary masses, $M_\text{p}$, and semi-major axes, $a$, for unenriched (panel (a)) and enriched (panel (b)) atmospheres are similar to each other. However, it turns out that sub/super-Earths with enriched atmospheres consequently have more massive atmospheres in relatively cool regions (see green symbols in panel (b)). This is because H$_2$O enrichment leads to increasing the atmospheric mean molecular weight and the effective heat capacity through condensation and chemical reactions, making the atmosphere denser \cite{Kimura+Ikoma2020}.
\subsection*{Water content distribution}
\begin{figure}
\centering
\includegraphics{histMw03_PDF.pdf}
\caption{Probability density distribution of the water mass fraction in the synthesized planets with mass of $0.3$--$3M_\oplus$ ("nearly Earth-mass planets") orbiting at between 0.1 and 0.2~au ("habitable zone") around M dwarfs of $0.3M_\odot$ at the age of 1~Gyr.
The panels show the results for the cases with (a) enriched atmospheres ($X_{\rm H_2O}=0.8$), (b) unenriched atmospheres ($X_{\rm H_2O}=0.0$), and (c) the same assumptions and settings as ref.~\cite{Tian+Ida2015}.
Planets without any water are shown in the leftmost bar in each panel for drawing purpose.
The vertical dashed line indicates the ocean mass fraction for the present-day Earth ($\num{2.3e-4}$).}
\label{fig:hist_M03}
\end{figure}
\begin{figure}
\centering
\includegraphics{histMw_PDF_Mstar.pdf}
\caption{Same as Fig.~\ref{fig:hist_M03}~(a) and (b), but for different stellar masses, (a) 0.1$M_\odot$, (b) 0.3$M_\odot$, and (c) 0.5$M_\odot$.
The orange and blue bars show the results for enriched and unenriched atmospheres (water mass fraction in the primordial atmosphere being $X_{\text{H}_2\text{O}}=0.8$ and 0.0), respectively.}
\label{fig:hist_Mstar}
\end{figure}
Such an increase of low-mass planets with relatively massive H$_2$O-enriched atmospheres greatly affects the occurrence of aqua planets in the habitable zones around M dwarfs.
Figure~\ref{fig:hist_M03} shows the probability density distributions of the water mass fraction in planets with mass of 0.3--3~$M_\oplus$ in the \textit{present-day} habitable zone (HZ) around 0.3~$M_\odot$ M dwarfs for the cases with enriched ($X_{\text{H}_2\text{O}}=0.8$; panel (a)) and unenriched ($X_{\text{H}_2\text{O}}=0.0$; panel (b)) atmospheres (See \textit{Method} for the definition of the density probability).
Such planets are hereafter called the habitable-zone nearly-Earth-mass planets (HZ-NEMPs).
Note that the present-day habitable zone is defined as the habitable zone around the star with age of 5~Gyr.
For comparison, the result obtained under the same assumptions and settings as ref.~\cite{Tian+Ida2015} is also shown in panel (c).
Firstly, compared to the previous model~\cite{Tian+Ida2015} (panel (c)),
our planetary population synthesis models produce HZ-NEMPs with a much wider range of water contents even for $X_{\text{H}_2\text{O}}=0.0$ (in particular, those with water mass fraction of $\lesssim$10\%).
This is due to the shift in location of the snowline with the protoplanetary disc's thermal evolution, which is not included in the previous model.
Initially, when viscous heating dominates, the snowline is located at $\sim 1$~au.
As viscous heating diminishes and, then, stellar irradiation dominates, the snowline migrates inward, reaching $\sim 0.2$~au, corresponding to the outer edge of the prersent-day HZ, on a timescale of Myrs (see Supplementary Figure 1 and Section 1.1 in the Supplementary Information for details).
Here, our model assumes that
once the snowline passes through inward, water vapour immediately condenses onto rocky planetesimals, making planetesimals with an ice-rock ratio of unity.
Thus, planets originally formed in this region can acquire small amounts of ice, depending on the timescale of planetary mass growth, orbital migration, and the snowline migration.
Effects of this assumption are discussed in Section 1.2 in the Supplementary Information, with Supplement Fig.2.
The water enrichment in the primordial atmosphere, which is of special interest in this study, is found to bring about further increase in water contents of HZ-NEMPs (see panel~(a) of Fig.~\ref{fig:hist_M03}).
Because of the enrichment, even without capturing icy planetesimals, the rocky planets obtain water from their primordial atmosphere.
As demonstrated in Fig.~\ref{fig:Ma_M03} (b), the water enrichment enhances the accumulation of the primordial atmospheres. Consequently, the captured H$_2$/He and secondarily produced H$_2$O are large in amount enough to survive the subsequent atmospheric photo-evaporation process. This mechanism forms rocky planets with water mass fraction of $< 1~\%$ inside the snowline.
In contrast to the previous models\cite{Tian+Ida2015,Miguel+2019} which predict the absence of HZ-NEMPs with the Earth-like water content ($2.3 \times 10^{-4}$), our model with enriched primordial atmospheres predicts that a significant number of such HZ-NEMPs are formed.
It is noted that the abundance of HZ-NEMPs with Earth-like water contents is significantly affected by the protoplanetary disc conditions, especially the disc lifetime.
Most of the HZ-NEMPs with small amounts ($\lesssim$ 1~wt.\%) of water are found to form in discs with lifetime of $\lesssim$ 3~Myr.
In longer-lived discs, the outer icy planets significantly migrate inward, pushing the inner rocky planets closer to the disc inner edge.
As a result, only ice-dominant planets exist in HZ.
\subsection*{Dependence on central stellar mass}
Similarly to Fig.~\ref{fig:hist_M03}, Fig.~\ref{fig:hist_Mstar} shows the probability density distribution for HZ-NEMPs orbiting stars of three different masses (0.1, 0.3 and 0.5$M_\odot$), around which the habitable zones are located at 0.03--0.07~au, 0.1--0.2~au and 0.2--0.4~au, respectively~\cite{Kopparapu+2014}, in their main-sequence phase.
As found in this figure, the larger the stellar mass, the higher the abundance of HZ-NEMPs with small water contents ($\lesssim$ 1~wt.\%).
This is due to two effects: Firstly, the region that corresponds to the HZ when the host star is on its main sequence is located at a larger orbital distance for larger stellar mass.
In that region of a protoplanetary disc, therefore, protoplanets can grow larger around more massive stars;
consequently, the protoplanets can obtain more atmospheric gas, thereby producing more water.
Secondly, the less massive the host star, the higher the stellar XUV irradiation in the habitable zone during the pre-main sequence phase. Therefore, the planets in the HZ undergo severer loss of the atmosphere and are less likely to keep water.
In the cases with unenriched atmospheres, the distribution also differs depending on stellar mass (dashed bars).
The distributions for 0.1 and 0.3$M_\odot$ stars are similar to each other, whereas the planets around 0.5$M_\odot$ stars have a much wider range of the water contents, and the distribution is rather similar to that for enriched atmospheres.
This indicates that many of the HZ-NEMPs around 0.5$M_\odot$ stars have obtained relatively large amounts of icy planetesimals during their formation.
This is because the snowline comes closer to the habitable zone around more massive stars; for 0.5$M_\odot$ stars, the former comes inside the latter.
\section*{Discussion}
\begin{figure*}
\centering
\includegraphics{RpFwater_03M.pdf}
\caption{Planetary radius, $R_\text{p}$ vs. the water content of the ``nearly Earth-mass planets'' in the habitable zone (HZ-NEMPs) around $0.3M_\odot$ M dwarfs in the case of enriched atmospheres with the primordial-atmospheric water mass fraction $X_{\rm H_2O}=0.8$.
The symbols are colour-coded according to the atmospheric mass relative to the planetary core mass.
Atmosphere-free planets are shown by black points.
Planets without any water are piled up at $M_\text{water}/M_\text{core}=10^{-5}$ for drawing purposes.
}
\label{fig:Rp_Fwater}
\end{figure*}
Our planetary population synthesis calculations have shown that the HZ-NEMPs orbiting M dwarfs can have diverse water contents by the effects of the disc's thermal evolution and the water enrichment in the primordial atmosphere.
Especially, the latter effect enables about 1\% of the HZ-NEMPs to have water amounts comparable to the Earth oceans, provided the atmospheric water mass fraction of 80~\% is achieved.
When water is produced through the chemical reaction between the atmospheric hydrogen and oxides in the magma ocean~\cite{Ikoma+Genda2006}, the resultant water mass fraction $X_{\text{H}_2\text{O}}$ depends on the kind of oxide; the equilibrium partial pressure ratio $P_{\text{H}_2\text{O}}/P_{\text{H}_2}$ is on the order of unity if the normal iron oxides such as w\"{u}stite are available from the magma.
The ratio $P_{\text{H}_2\text{O}}/P_{\text{H}_2}=1$ corresponds to about 80 wt.\% of H$_2$O in the H$_2$-He-H$_2$O atmosphere with the solar He/H ratio (=0.385~\cite{Lodders+2009}).
Thus, the assumed value of $X_{\text{H}_2\text{O}}=0.8$ is feasible if the entire atmospheric hydrogen is equilibrated with the oxides.
We should note that our results and conclusion hardly change for $X_{\text{H}_2\text{O}} \gtrsim 0.5$, because the atmospheric mass significantly increases in that range (see ref.~\cite{Kimura+Ikoma2020} for details).
We further discuss the implication of this water production process in the primordial atmosphere for the Earth in Section 2.3 in the Supplement Information.
Our planetary population synthesis model is based on the planetesimal accretion scenario, and does not consider pebbles.
The pebble accretion may affect our results and conclusion mainly in the following three ways: changes in the disc gas composition, in the oxidation state of the planetary rocky components, and in the planetary mass distribution.
Firstly, drifting pebbles can enhance the water content of disc gas inside the snowline through the sublimation of ice~\cite{Booth+2017}.
However, this change in the disc gas composition would hardly affect the total mass of the water produced in the primordial atmosphere when the atmosphere is equilibrated with the magma ocean.
Secondly, pebbles drifting from outside the snowline would be more oxidised than solids originally existing inside the snowline~\cite{Lodders2003}.
Since the equilibrium partial pressure of the water vapour in the atmosphere is higher for more oxidised magma, the accretion of the oxidised rocky components would work in favour of the water production and the formation of planets with small water content.
Finally, a pebble accretion scenario may significantly change the resultant $M_\text{p}$-$a$ distribution of the synthesised planets~\cite{Brugger+2020}.
Since the water amount in the atmosphere largely depends on the planetary mass, the mass distribution of the planets in the habitable zone affects the water amount distribution.
Further studies including pebbles are needed to quantify these effects.
Our models place constraints on the radii of temperate
aqua planets, which will be useful for ongoing and future exploration of habitable planets around M dwarfs.
Figure~\ref{fig:Rp_Fwater} shows the relationships between the radius and water content of HZ-NEMPs with enriched atmospheres orbiting $0.3M_\odot$ M dwarfs.
Most of the HZ-NEMPs of $> 1.3 R_\oplus$ are shown in bluish colours, indicating that those planets have thick hydrogen-rich atmospheres ($\gtrsim 10^{-3} M_\mathrm{p}$).
Given those thick atmospheres bring about such a strong blanketing effect that an H$_2$O layer below the atmosphere would be, if any, in a super-critical state,
those planets are unlikely to be habitable.
The other HZ-NEMPs of $> 1.3 R_\oplus$ (black symbols) retain large amounts of water ($\gtrsim 0.1 M_\mathrm{p}$) and
are far from Earth-like (but worth atmospheric characterisation with JWST~\cite{Gardner+2006} and Ariel \cite{Tinetti+2018} for verifying our theoretical prediction).
The HZ-NEMPs of 0.7--1.3~$R_\oplus$ (shown in black) have lost their hydrogen atmospheres completely, ending up with rocky planets covered with oceans.
It turns out that those planets are diverse in water content and do include planets with Earth-like water content.
Several climate studies argue the amounts of seawater appropriate for temperate climates, considering the effects of seafloor weathering, high-pressure ice, water cycling and heterogeneous surface water distribution \cite{Abe+2011,Abbot+2012,Kaltenegger+2013,Alibert2014,Nakayama+2019,Moore+Cowan2020}. According to those studies, the appropriate seawater amount ranges from $\sim$0.1 to 100 times that of the Earth.
From an observational point of view, it would be important to exclude planets unlikely to be habitable in advance. Among those HZ-NEMPs, there are relatively low-mass rocky planets that have deep oceans with high-pressure ice; such planets are unlikely to have temperate climates~\cite{Kaltenegger+2013,Alibert2014,Nakayama+2019}.
Meanwhile, one can identify rocky planets with ocean mass fractions $\gtrsim$~100 times the present-day Earth's one, which are also likely uninhabitable,
if planetary masses and radii are measured within
$\lesssim$ 20\% and 5\% accuracies, respectively.
In Fig.~\ref{fig:Rp_Fwater}, about 25~\% of the HZ-NEMPs of 0.7--1.3~$R_\oplus$ are such ocean-dominated planets.
The remaining 75\% would be identified as ``water-poor rocky planets'' (i.e., rocky planets without hydrogen-rich atmospheres nor thick oceans),
which are capable of having Earth-like temperate climates as long as they have little amounts of seawater~\citep[$\gtrsim$ 0.001\%;][]{Abe+2011,Kodama+2019,Moore+Cowan2020}.
Thus, excluding the completely dry planets, which account for about 95\% of those remaining HZ-NEMPs, the HZ-NEMPs with appropriate amounts of seawater for habitability are estimated to account for $\sim$5~\% of the ``water-poor rocky planets'' orbiting $0.3M_\odot$ M dwarfs.
This frequency becomes higher for larger stellar mass, and around $0.5M_\odot$ stars, for example, more than 10~\% of the water-poor rocky planets are expected to have the appropriate amounts of seawater.
Note that the appropriate water amount for temperate climate can be even wider, considering the possibility of the water sequestration in the mantle~\cite{Cowan+Abbot2014} and of further water loss due to the absence of silicate weathering~\cite{Abbot+2012}. On the other hand, for tidally locked planets, water can be trapped as ice on the night side, leading to atmospheric collapse, if the planet has only a small amount of the atmosphere~\cite{Joshi1997,Wordsworth+2011}.
In this case, a temperate climate would be difficult to achieve.
Whether the planets with seawater contents focused on in this study can actually sustain a temperate climate needs to be addressed in the future.
Finally, the significance of the above occurrence rate of aqua planets with temperate climates may be realised by comparison with the number of HZ-NEMPs possibly detected by ongoing and future survey missions such as TESS and PLATO. TESS is estimated to detect $\sim$10 HZ-NEMPs around early M dwarfs of 0.3-0.5~M$_\odot$~\cite{Kunimoto+2022}, and PLATO will detect $\sim$100 HZ-NEMPs~\cite{Rauer+2014}.
Multiplying the number of possible detections by the predicted occurrence rate, our model predicts that 5-10 aqua planets with temperate climates will be discovered in the 2020s, whereas the previous model predicting the occurrence rate of < 0.01\% \cite{Tian+Ida2015} suggests that the discovery is hopeless.
\section*{Method}
Here we describe the details about our planetary population synthesis model, which is based on the planetesimal-driven core accretion scenario.
The model includes the evolution of the central star,
the evolution and radial structure of the protoplanetary disc
the growth of solid cores,
the accumulation and loss of atmospheres,
orbital migration,
and dynamical interactions between protoplanets,
including resonance trapping
and orbital instabilities for multi-protoplanet systems and their outcomes
In addition, we describe the calculation method for the radius of the isolated planet during its thermal evolution
and the initial conditions and parameters for generating the planetary populations
We discuss the effects of the updated components of the model from the previous study~\cite{Tian+Ida2015} and the consistency with the observed exoplanet populations in the Section 2.1 and 2.2 in the Supplementary Information, together with Supplementary Figs. 3 and 4.
\subsection*{Stellar Evolution} \label{sec:star}
From the table provided by ref.\cite{Baraffe+1998}, we take the stellar radius $R_*$, luminosity $L_*$ and effective temperature $T_*$
as a function of stellar mass $M_*$ and age for the solar metallicity with He abundance of $Y=0.275$.
Those data are used for calculating the protoplanetary disc temperature, planetary equilibrium temperature, and atmospheric photoevaporation rate in our planetary population synthesis calculations.
\subsection*{Protoplanetary Disc} \label{sec:disc}
Our planetary population synthesis calculations need
the surface density distributions of gas and solids (or planetesimals) in the protoplanetary disc and their evolution, and also need the disc midplane temperature.
Those are calculated in a similar way to ref. \cite{Emsenhuber+2021a}, which is summarised below.
\subsubsection*{Gas disc profile and evolution}
\begin{center}
\textit{(i) Initial profiles}
\end{center}
The initial profile of disc gas surface density is
assumed as~\citep{Veras+Armitage2004,Andrews+2010}
\begin{align}
\begin{split}
\Sigma_\text{g}^{(t=0)} &= \Sigma_\text{g0}\qty(\frac{r}{r_0})^{-q_\text{g}}
\exp[ -\qty(\frac{r}{r_\text{disc}})^{2-q_\text{g}} ] \\
& \qquad
\times \qty( 1 - \sqrt{\frac{r_\text{in}}{r}}),
\label{eq:Sigmag0}
\end{split}
\end{align}
where $\Sigma_\text{g0}$ is the gas surface density at orbital radius $r$ = $r_0$,
$r_\text{disc}$ is the gas disc
characteristic radius beyond which the surface density decays exponentially,
and $r_\text{in}$ is the inner edge radius.
This is the so-called self-similar solution with the turbulent viscosity $\nu_\text{acc} \propto r^{q_\text{g}}$~\citep{Lynden-Bell+Pringle1974,Hartmann+1998}, including the smooth cutoff at $r=r_\text{in}$.
We set $q_\text{g}=0.9$, which is inferred from observations of protoplanetary discs~\citep{Andrews+2010}, and $r_0=1$~au.
The surface density at $r$ = $r_0$, $\Sigma_\text{g0}$, is calculated from the total disc gas mass $M_\text{disc}$ as
\begin{equation}
\Sigma_\text{g0} = \frac{2-q_\text{g}}{2\pi} M_\text{disc}
r_0^{-q_\text{g}} r_\text{disc}^{q_\text{g}-2}.
\end{equation}
This is obtained by integrating Eq.~\eqref{eq:Sigmag0} multiplied by $2\pi r$ (i.e., $2\pi r \Sigma_\text{g}^{(t=0)}$) from $r=0$ to $r_\text{disc}$, ignoring the exponential and inner edge cutoff terms.
\begin{center}
\textit{(ii) Viscous diffusion}
\end{center}
The evolution of the disc gas surface density $\Sigma_\text{g}$ is assumed to occur via radial viscous diffusion, photo-evaporation, and absorption by embedded planets and is expressed as~\citep[e.g.,][]{Lynden-Bell+Pringle1974}
\begin{equation}
\pdv{\Sigma_\text{g}}{t} - \frac{3}{r}
\pdv{r}( r^{1/2} \pdv{r}(\nu_\text{acc} \Sigma_\text{g} r^{1/2}))
= -\dot{\Sigma}_\text{pe} - \dot{\Sigma}_\text{planet}
\label{eq:dSigma_dt_basic},
\end{equation}
where $-\dot{\Sigma}_\text{pe}$ and $-\dot{\Sigma}_\text{planet}$ are the
sink terms due to photo-evaporation and absorption by planets, respectively.
We adopt the $\alpha$-prescription for the turbulent viscosity:
$\nu_\text{acc}=\alpha_\text{acc} c_\text{s}H_\text{disc}$ \citep{Shakura+Sunyaev1973},
where $c_\text{s} = \sqrt{k_\text{B}T_\text{disc}/(\mu m_\text{H})}$ is the isothermal sound speed
and $H_\text{disc}=c_\text{s}/\Omega_\text{K}$ is the disc scale height.
Here $\mu$ is the mean molecular weight, which is assumed to be 2.34 for gas with solar abundances, $T_\text{disc}$ is the midplane temperature calculated below, $k_\text{B}$ and $m_\text{H}$ are the Boltzmann constant ($=\SI{1.38e-16}{erg.K^{-1}}$) and the hydrogen atomic mass ($=\SI{1.66e-24}{g}$), respectively, and $\Omega_\text{K}$ is the Keplerian frequency. We treat $\alpha_\mathrm{acc}$ as an input parameter.
Equation~\eqref{eq:dSigma_dt_basic} is solved with a log-uniform grid with
500 points ($N_\mathrm{grid,disc}$) extending from $r_\text{in}$ to $r_\text{max}=1000$~au.
The boundary conditions are $\Sigma_\text{g}(r_\text{in})=0$ and $\Sigma_\text{g}(r_\text{max})=0$.
\begin{center}
\textit{(iii) Photo-evaporation}
\end{center}
We consider photo-evaporation processes caused by UV irradiation from the central star (internal photo-evaporation) and from nearby massive stars (external photo-evaporation).
The total photo-evaporation rate $\dot{\Sigma}_\text{pe}$ is determined by
the sum of the two photo-evaporation rates.
The external photo-evaporation rate is calculated by the prescription from ref. \cite{Matsuyama+2003}.
The model assumes that the FUV photons from nearby massive stars evaporate the disc gas uniformly only in the regions exterior to the effective gravitational radius.
The gravitational radius for the dissociated gas is defined as
\begin{equation}
r_\text{g,I} = \frac{GM_*}{c_\text{s,I}^2},
\end{equation}
where $G$ is the gravitational constant ($=\SI{6.67e-8}{cm^3.g^{-1}.s^2} $); $c_\text{s,I}$ is the sound speed with temperature of $1 \times 10^3$~K and mean molecular weight of
1.35~\citep{Matsuyama+2003}, given that the hydrogen is fully-dissociated and the He/H ratio is equal to solar.
Assuming that the disc gas is uniformly removed only from the regions exterior to the effective gravitational radius $\beta_\text{I}r_\text{g,I}$, the external photo-evaporation rate is given by
\begin{equation}
\dot{\Sigma}_\text{pe,ext} =
\begin{cases}
0 &\qfor r < \beta_\text{I}r_\text{g,I} \\
\displaystyle{
\frac{\dot{M}_\text{wind}}{\pi(r_\text{max}^2-\beta_\text{I}^2r_\text{g,I}^2)}} & \quad \text{otherwise},
\end{cases}
\label{eq:dSigma_pe_ext}
\end{equation}
where we assume $r_\text{max}=1000$~au and regard $\dot{M}_\text{wind}$ as an input parameter providing the total mass loss rate.
For the effective gravitational radius, analytical estimates~\citep[e.g.,][]{Liffman2003} and numerical results~\cite[e.g.,][]{Begelman+1983a,Adams+2004,Font+2004} show that $\beta_\text{I}=$0.1--0.2 would be appropriate.
Here we set $\beta_\text{I}=0.14$ following ref.~\cite{Emsenhuber+2021a}.
Next, the internal photo-evaporation rate is calculated
from the following equation \citep[e.g.,][]{Hollenbach+1994,Clarke+2001}
\begin{equation}
\dot{\Sigma}_\text{pe,int}(r) =
\begin{cases}
0 &\qfor r < \beta_\text{II}r_\text{g,II} \\
2 c_\text{s,II}n_0(r) m_\text{H} & \quad \text{otherwise}.
\end{cases}
\end{equation}
Here, $r_\text{g,II}$ and $c_\text{s,II}$ are the gravitational radius and sound speed for ionised gas of temperature $1 \times 10^{4}$~K and mean molecular weight 0.68~\citep[assuming the fully ionised hydrogen and the He/H ratio of solar;][]{Hollenbach+1994,Liffman2003,Matsuyama+2003}, respectively.
We also set $\beta_\text{II}=0.14$, as well as $\beta_\text{I}$.
The number density of the ionised hydrogen $n_0$ is
given by
\begin{equation}
n_0(r) = \num{3.1e5} \qty(\frac{r_\text{g,II}}{1~\si{au}})
\qty(\frac{\Phi}{10^{40}~\si{s^{-1}}})^{1/2}
\qty(\frac{r}{1~\si{au}})^{-5/2},
\label{eq:n0}
\end{equation}
following the fitting formula derived in ref. \cite{Hollenbach+1994}, with $\Phi$ the ionising EUV photon luminosity.
Finally, the total photo-evaporation rate is given by
\begin{equation}
\dot{\Sigma}_\text{pe} = \dot{\Sigma}_\text{pe,ext} + \dot{\Sigma}_\text{pe,int}.
\end{equation}
\begin{center}
\textit{(iv) Absorption by planets}
\end{center}
The sink term for absorption of disc gas by planets $\dot{\Sigma}_\text{planet}$ is simply given by
\begin{equation}
\dot{\Sigma}_\text{planet} = \frac{\dot{M}_\text{env}}{4\pi a_\text{p} R_\text{H}},
\label{eq:dSigma_planet}
\end{equation}
where $\dot{M}_\text{env}$ is the runaway gas accretion rate of the planet given by Eq.~\eqref{eq:gas_acc_rate}, and $a_\text{p}$ and $R_\text{H}$ are the semi-major axis and Hill radius of the planet, respectively.
Equation~\eqref{eq:dSigma_planet} is calculated for grids inside the Hill radius of each planet.
\subsubsection{Disc temperature}\label{sec:disk_temp}
The disc midplane temperature $T_\text{disc}$ is determined by the combination of viscous heating and direct and indirect stellar irradiation and can be expressed as~\citep[e.g.,][]{Emsenhuber+2021a}
\begin{equation}
\sigma T_\text{disc}^4 = \sigma T_\text{vis}^4 + \sigma T_\text{irr}^4
+ \sigma T_\text{eq}^4 \exp(-\tau_r),
\end{equation}
where $T_\mathrm{vis}$ and $T_\mathrm{irr}$ are the temperatures that would be achieved if only viscous heating or indirect stellar irradiation, respectively, was available, and $T_\text{eq}$ is the equilibrium temperature due to the direct stellar irradiation through the disc midplane, and $\tau_r$ is the radial optical depth at the disc midplane.
The temperature $T_\text{vis}$ is given by \citep[e.g.,][]{Nakamoto+Nakagawa1994,Hueso+Guillot2005}
\begin{equation}
\sigma T_\text{vis}^4 = \frac{1}{2}\qty(\frac{3}{8}\tau_\text{R} + \frac{1}{2\tau_\text{P}})\dot{E},
\label{eq:Tdisc_NN}
\end{equation}
where $\tau_\text{R} (= \kappa_\text{R}\Sigma_\text{g})$ and $\tau_\text{P}$ are the Rosseland and Planck mean optical depths, respectively, and $\dot{E}$ is the viscous energy generation rate per unit surface area.
The Rosseland mean opacity $\kappa_\text{R}$ is given by the analytical fitting of the opacity of dust grains provided by ref. \cite{Bell+Lin1994}.
We ignore the gas opacity and simply set $\kappa_\text{R} = 0~\si{cm^2/g}$ after the evaporation of dust grains.
Also, we simply set $\tau_\text{P} = \max(2.4\tau_\text{R},0.5)$
\citep{Nakamoto+Nakagawa1994,Hueso+Guillot2005,Suzuki+2016}.
The minimum value of 0.5 is adopted so that the coefficient of $\dot{E}$ in Eq.~\eqref{eq:Tdisc_NN} approaches to unity in the optically thin limit.
The viscous heating rate $\dot{E}$ is
\begin{equation}
\dot{E} = \frac{9}{4}\nu_\text{acc} \Sigma_\text{g} \Omega_\text{K}^2.
\end{equation}
The temperature due to indirect irradiation $T_\text{irr}$ is given by \citep{Kusaka+1970,Adams+1988,Ruden+Pollack1991}
\begin{align}
\begin{split}
T_\text{irr}^4 &= T_*^4
\left[ \frac{2}{3\pi}\qty(\frac{R_*}{r})^3 \right. \\
& \qquad \left. + \frac{1}{2}\qty(\frac{R_*}{r})^2
\frac{H_\text{disc}}{r}\qty(\dv{\ln H_\text{disc}}{\ln r}-1)\right].
\label{eq:Tirr}
\end{split}
\end{align}
For computational simplicity, we adopt a constant value of $\dv*{\ln H_\text{disc}}{\ln r}=9/7$, which is the approximate equilibrium solution for a flaring disc \citep{Chiang+Goldreich1997}.
Finally, $T_\text{eq}$ and $\tau_r$ are given by
\begin{equation}
T_\text{eq}^4 = \frac{L_*}{16\pi \sigma r^2},
\label{eq:Teq}
\end{equation}
and
\begin{equation}
\tau_r = \int_0^r \kappa_\text{R} \rho_\text{disc} \dd{r},
\end{equation}
where $\rho_\text{disc} = \Sigma_\text{gas}/\sqrt{2\pi}H_\text{disc}$ is the gas density at the disc midplane.
Since all of the
temperatures above change with time, the H$_2$O snowline, defined as the radial location for
$T_\text{disc}=170$~K,
moves in the protoplanetary disc as follows.
In the early stage of the disc evolution, the disc midplane temperature $T_\text{disc}$ is almost
equal to $T_\text{vis}$ and generally much higher than $T_\text{eq}$ because of the large viscous heating rate.
Thus, the snowline locates farther than the present position.
As the disc evolves, $T_\text{disc}$ becomes comparable to $T_\text{irr}$, which is much smaller than $T_\text{vis}$, and the snowline moves inward on a $\sim$Myr timescale.
Then, at the timing of the disc gas dissipation, $T_\text{disc}$ jumps up to $T_\text{eq}$ in a quite short time (typically $\sim 10^{4}$~yrs).
Since $T_\text{eq} > T_\text{irr}$ holds in many cases, the snowline moves outward in this stage.
Finally, the snowline slightly moves according to the evolution of the stellar luminosity (generally inward around
M dwarfs).
\subsubsection
Distribution and evolution of planetesimals}
We set the initial distribution of planetesimals (or solids) in terms of
surface density as
\begin{equation}
\Sigma_\text{s}^{(t=0)} = \eta_\text{ice} \Sigma_\text{rock}^{(t=0)},
\end{equation}
where $\eta_\text{ice}$ is the enhancement factor associated with H$_2$O condensation (see below) and $\Sigma_\text{rock}^{(t=0)}$ is the initial surface density of rocks, which is given by
\begin{equation}
\Sigma_\text{rock}^{(t=0)} = \Sigma_\text{s0}\qty(\frac{r}{r_0})^{-q_\text{s}}
\exp\left[ -\qty(\frac{r}{r_\text{solid}})^2 \right]
\qty( 1 - \sqrt{\frac{r_\text{in}}{r}});
\label{eq:Sigma_solid}
\end{equation}
$\Sigma_\text{s0}$ is the
initial surface density of rocks at $r=r_0$,
and $r_\text{solid}$ is the
characteristic radius of the existence region of planetesimals (or called the planetesimal disc).
We set $q_\text{s}=1.5$ \citep{Birnstiel+2012,Birnstiel+Andrews2014} and $r_\text{solid}=0.5r_\text{disc}$ \citep{Ansdell+2018}.
Similarly to $\Sigma_\text{g0}$, $\Sigma_\text{s0}$ is calculated by
\begin{equation}
\Sigma_\text{s0} = \frac{2-q_\text{s}}{2\pi \eta_\text{ice}} M_\text{solid}
r_0^{-q_\text{s}} r_\text{solid}^{q_\text{s}-2}.
\end{equation}
Here $M_\text{solid}$ is the total mass of planetesimals initially existing in the entire protoplanetary disc, which
is given by
\begin{equation}
M_\text{solid} = 10^{\text{[Fe/H]}}Z_{\odot}M_\text{disc},
\end{equation}
where [Fe/H] is the disc metallicity relative to the solar one, which is assumed equal to the central star's one, and $Z_\odot$ is the solar metallicity, which is set to 0.015~\citep{Lodders2003}.
The enhancement factor $\eta_\text{ice}$ is given by
\begin{equation}
\eta_\text{ice} =
\begin{cases}
1 & \qfor r < r_\text{ice}, \\
2 & \qfor r > r_\text{ice},
\end{cases}
\label{eq:etaice}
\end{equation}
where $r_\text{ice}$ is the snowline position
This is similar to the prescription in \cite{Ida+2013}, but we assume $\eta_\text{ice}=2$ beyond the snowline \citep{Lodders2003}, instead of 4.2 in their original prescription. We ignore snowlines for condensates other than H$_2$O.
The planetesimal distribution changes due to the accretion and scattering by protoplanets, and also due to the sublimation of ice as the snowline moves.
Assuming
that $\Sigma_\text{rock}$ is only affected by the first two processes, we calculate its evolution by
\begin{equation}
\dv{\Sigma_\text{rock}}{t} = -\frac{\dot{M}_\text{core}+\dot{M}_\text{scat}}{2\pi \eta_\text{ice} a_\text{p} \Delta a_\text{FZ}},
\label{eq:dSigma_rock}
\end{equation}
where $\dot{M}_\text{core}$ and $\dot{M}_\text{scat}$ are the planetesimal accretion and scattering rates by the protoplanet, respectively, and $\Delta a_\text{FZ}$ is the full-width of the feeding zone of the planet, which is assumed to be 10~$R_\mathrm{H}$ \citep{Kokubo+Ida2002}.
The prescriptions for $\dot{M}_\text{core}$ and $\dot{M}_\text{scat}$ are described in section~\ref{sec:solid_core_growth}.
To calculate Eq.~\eqref{eq:dSigma_rock}, we set a log-uniform grid, which is different from that for disc gas, with $N_\text{grid,solid}=1000$ extending from $r_\text{in}$ to $5r_\text{solid}$.
To evaluate $\Sigma_\text{s}(t)$, we first integrate Eq.~\eqref{eq:dSigma_rock} to derive $\Sigma_\text{rock}(t)$ for all grids inside the feeding zone of each planet.
If a grid is in the feeding zone of multiple planets, the right hand side of Eq.~\eqref{eq:dSigma_rock} is calculated for all the involved planets and the results are summed up.
At the same time, $\eta_\text{ice}(t)$ for each grid is evaluated using $T_\text{disc}(t)$.
Then, the planetesimal surface density is derived by $\Sigma_\text{s}(t) = \eta_\text{ice}(t)\Sigma_\text{rock}(t)$
\subsection*{Solid Core Growth} \label{sec:solid_core_growth}
We assume that protoplanetary solid cores grow by the accretion of planetesimals.
The mass growth rate is given by
\begin{equation}
\dv{M_\text{core}}{t} = \Omega_\text{K} \bar{\Sigma}_\text{s} R_\text{H}^2 p_\text{col}
\end{equation}
with $\bar{\Sigma}_\text{s}$ the mean surface density of planetesimals in the planet's feeding zone and $p_\text{col}$ the collision probability of planetesimals;
$p_\text{col}$ is given by \citep{Inaba+2001}
\begin{equation}
p_\text{col} = \min \qty(p_\text{col,med}, \qty(p_\text{col,high}^{-2}+p_\text{col,low}^{-2})^{-1/2} ),
\end{equation}
with
\begin{align}
p_\text{col,high} &= \frac{1}{2\pi} \qty(\frac{R_\text{cap}}{R_\text{H}})^2
\qty(
\mathcal{F}(i_\text{plt}/e_\text{plt}) + \frac{6R_\text{H}\mathcal{G}(i_\text{plt}/e_\text{plt})}{R_\text{cap}\tilde{e}_\text{plt}^2}
), \\
p_\text{col,med} &= \frac{1}{4\pi \tilde{i}_\text{plt}} \qty(\frac{R_\text{cap}}{R_\text{H}})^2
\qty(
17.3 + \frac{232R_\text{H}}{R_\text{cap}}
),\\
p_\text{col,low} &= 11.3\sqrt{R_\text{cap}/R_\text{H}}.
\end{align}
Here $i_\text{plt}$ and $e_\text{plt}$ are respectively the inclination and eccentricity of planetesimals,
and $\tilde{i}_\text{plt}=i_\text{plt}/h$ and $\tilde{e}_\text{plt} = e_\text{plt}/h$ are the reduced inclination and eccentricity, respectively, with $h = R_\text{H}/a_\text{p}$.
We assume $i_\text{plt} = e_\text{plt}/2$, in which case $\mathcal{F}(i_\text{plt}/e_\text{plt}) = 17.3$ and $\mathcal{G}(i_\text{plt}/e_\text{plt}) = 38.2$
~\citep[see][for the explicit form of $\mathcal F(x)$ and $\mathcal G(x)$]
{Greenzweig+Lissauer1990,Greenzweig+Lissauer1992,Inaba+2001}.
The value of $\tilde{e}_\text{plt}$ around a protoplanet is assumed equal to the equilibrium eccentricity determined by the balance of viscous stirring by the planet and the damping due to the disc gas drag.
Following the analytic estimates by ref.\cite{Kokubo+Ida2002}, it is expressed as
\begin{align}
\begin{split}
\tilde{e}_\text{plt}
= 4.1
&\qty(\frac{\rho_\text{disc}}{\SI{1e-9}{g/cc}})^{-1/5}
\qty(\frac{\rho_\text{plt}}{\SI{3}{g/cm^3}})^{2/15} \\
& \times\qty(\frac{a_\text{p}}{1~\si{au}})^{-1/5}
\qty(\frac{m_\text{plt}}{10^{20}\si{g}})^{1/15},
\end{split}
\label{eq:tilde_e_plt}
\end{align}
where $\rho_\text{plt}=3.0~\si{g.cm^{-3}}$ and $m_\text{plt}=10^{20}~\si{g}$ are the material density and mass of planetesimals.
Finally,
we calculate the effective capture radius enhanced by the atmospheric gas drag, $R_\text{cap}$,
following ref.~\cite{Inaba+Ikoma2003};
\begin{equation}
R_\text{plt} = \frac{3}{2}
\frac{v_\text{ran}^2 + 2GM_\text{p}/R_\text{cap}}
{v_\text{ran}^2 + 2GM_\text{p}/R_\text{H}}
\frac{\rho(R_\text{cap})}{\rho_\text{plt}} R_\text{cap},
\label{eq:Rcap}
\end{equation}
with $R_\text{plt}$ the planetesimal's radius derived from $m_\text{plt}$ and $\rho_\text{plt}$, $v_\text{ran} = e_\text{plt}a_\text{p}\Omega_\text{K}$ the random velocity of the planetesimals and $\rho(R_\text{cap})$ the envelope gas density at the radius of $R_\text{cap}$.
After the disc gas dissipation, some of the planetesimals that encountered a planet are ejected from the system, instead of colliding with the planet.
This ejection rate can be estimated by comparing the collisional cross section and the scattering cross section, and is given by~\citep{Ida+Lin2004}
\begin{equation}
\dot{M}_\text{scat} = \qty(\frac{v_\text{surf}}{v_\text{esc}})^4 \dot{M}_\text{core},
\end{equation}
where $v_\text{esc}=\sqrt{2GM_*/a_\text{p}}$ is the escape velocity from the host star and $v_\text{surf}=\sqrt{GM_\text{p}/R_\text{cap}}$ is the surface velocity of the planet.
This is used for the calculation of the evolution of $\Sigma_\text{rock}$ (Eq.~\eqref{eq:dSigma_rock}).
\subsection*{Atmospheric Accumulation and Loss} \label{sec:atmosphere}
\subsubsection*{Purely hydrostatic equilibrium phase}
During vigorous planetesimal accretion, the atmospheric mass is rather small because of significant energy deposition. Then, the atmosphere is in the hydrostatic equilibrium and thermally steady state \citep[e.g.,][]{Ikoma+2000}. We solve the standard set of equations for stellar structure, namely,
\begin{align}
\pdv{P}{M_R} &= -\frac{GM_R}{4\pi R^4}, \label{eq:dPdm} \\
\pdv{T}{M_R} &= -\frac{GM_R}{4\pi R^4}\frac{T}{P} \nabla , \label{eq:dTdm} \\
\pdv{R}{M_R} &= \frac{1}{4\pi R^2 \rho} \label{eq:drdm},
\end{align}
where $P$, $T$, and $\rho$ are the pressure, temperature, and density of the atmosphere (or envelope) gas, respectively, $R$ is the radial distance to the planet's centre, and $M_R$ is the total mass inside the sphere of radius $R$.
In addition to the above equations, we use the ideal equation of state (or the $P$-$T$ relationship) for chemical equilibrium mixtures composed of H- and O-bearing molecules and He, taking H$_2$O sublimation/condensation into account \citep[see][for the details]{Kimura+Ikoma2020}. Also, $\nabla$ is the temperature gradient, $\dv*{\log T}{\log P}$, for radiative diffusion or convection (dry or moist adiabat).
In the early stages of planetary accretion, the temperature is high enough at the bottom of the atmosphere to keep
a global magma ocean~\citep[][]{Ikoma+Genda2006,Kimura+Ikoma2020}. The atmosphere-magma interaction produces volatiles, modifying the atmospheric composition significantly.
Therefore, we assume that the water-producing reaction effectively occurs and that the produced water vapour is uniformly mixed in the atmosphere.
Hereafter, we call such an atmosphere at this phase the \textit{vapour-mixed atmosphere},
and denote its mass by $M_\text{mix}$.
We calculate the atmospheric mass $M_\mathrm{mix}$ in the same way as ref.~\cite{Kimura+Ikoma2020}, treating
the water mass fraction in the vapour-mixed atmosphere $X_{\text{H}_2\text{O}}$ as an input parameter.
We set $T = T_\text{disc}$ and $P = P_\text{disc}$ at $R = \min(R_\text{B}, R_\text{H})$, where $R_\text{B}$ and $R_\text{H}$ are the Bondi and Hill radii, respectively.
The luminosity $L$, which is assumed constant in the entire atmosphere, is equal to that from the solid core; namely,
\begin{equation}
L_\text{core} = L_\text{acc} + L_\text{cool} + L_\text{radio},
\label{eq:Lcore}
\end{equation}
where $L_\text{acc}$, $L_\text{cool}$ and $L_\text{radio}$ are the luminosity due to the planetesimal accretion, the solid core cooling and radioactive decay, respectively. The accretion luminosity
\begin{equation}
L_\text{acc} = \frac{G M_\text{core} \dot{M}_\text{core}}{R_\text{core}},
\label{eq:Lacc}
\end{equation}
where
$\dot{M}_\text{core}$ is the planetesimal accretion rate, and $R_\text{core}$ is the solid planet radius.
By assuming that the core surface temperature $T_\text{surf}$ and the disc gas temperature $T_\text{disc}$ are related in the form of $\dv*{T_\text{surf}}{\rho_\text{disc}}=T_\text{disc}/4$, as indicated by the analytical solution of the fully-radiative atmosphere,
$L_\text{cool}$ is given by~\citep{Ikoma+Hori2012}
\begin{equation}
L_\text{cool} = \frac{M_\text{core} C_\text{rock} T_\text{disc}}{4\tau_\text{disc}},
\label{eq:Lcool}
\end{equation}
where $C_\text{rock}$ is the specific heat of rock $(=\SI{1.2e7}{erg \, g^{-1}K^{-1}})$, and $\tau_\text{disc}$ is the disc dissipation timescale.
We set $\tau_\text{disc}$ = $1 \times 10^4$~yr in our simulations, since disc dissipation occurs on a timescale of $\sim 10^4$--$10^5$~yr due to photoevaporation in the final stage of disc evolution.
Increasing this value by an order of magnitude has little effect on our results.
While the above equation was derived assuming a thin radiative atmosphere, the vapour-rich primordial atmosphere considered in this study is thick even in the final stage of the disc evolution, and the lower layer is often convective.
Even in that case, the changing rate of temperature at the bottom of the atmosphere is about the same as that at the radiative-convective boundary.
Therefore, Eq.~\eqref{eq:Lcool} is still considered to be a good approximation for describing the cooling luminosity.
The radiogenic luminosity $L_\text{radio}$ is simply set to $\num{2e20}(M_\text{core}/M_\oplus)~\si{erg/s}$~\citep{Guillot+1995}.
\subsubsection*{Quasi-static contraction phase}
Once planetesimal accretion stops,
the vapour-mixed atmosphere contracts gravitationally and further accumulation of disc gas occurs.
Thus, we have to integrate the equation of entropy change, in addition to Eqs.~(\ref{eq:dPdm})-(\ref{eq:drdm});
\begin{equation}
\pdv{L_R}{M_R} = \dot{\varepsilon} -T \dv{S}{t},
\label{eq:dLdm}
\end{equation}
where $L_R$ is the total energy flux passing through a spherical surface of radius $R$ (or luminosity), $\dot{\varepsilon}$ is the energy generation rate per unit mass, and $S$ is the specific entropy.
We assume that the accumulating disc gas composed predominantly of H and He never mixes with the lower vapour-mixed layer and, thus, the mass of the vapour-mixed atmosphere ($M_\text{mix}$) is conserved in this phase.
Hereafter, the accumulated mass of disc gas is denoted by $M_\text{HHe}$.
Therefore we calculate the quasi-static contraction of the two-layer atmosphere in this phase.
The numerical integration of Eq.~(\ref{eq:dLdm}) is done based on the total energy conservation approximation adopted in several previous studies on the formation of gas giants~\citep{Papaloizou+Nelson2005,Mordasini+2012b,Fortier+2013,Piso+Youdin2014,Venturini+2016}.
The total atmospheric energy conservation between the time $t-\Delta t$ and $t$ being considered, the following relation holds (see ref.~\cite{Piso+Youdin2014} for the derivation):
\begin{align}
\begin{split}
& \frac{E_\text{env}(t)-E_\text{env}(t-\Delta t)}{\Delta t} \\
&\qquad = L_\text{core}
+ e_\text{gas} \frac{M_\text{env}(t) - M_\text{env}(t-\Delta t)}{\Delta t} - L
\end{split}
\label{eq:energy_conserv}
\end{align}
or
\begin{equation}
\Delta t =
\frac{E_\text{env}(t)-E_\text{env}(t-\Delta t)-e_\text{gas} \left[M_\text{env}(t) - M_\text{env}(t-\Delta t)\right]}
{L_\mathrm{core}-L}
\label{eq:energy_conserv2}
\end{equation}
where $e_\text{gas}$ is
the total energy per unit mass (i.e., the sum of the specific internal energy and the gravitational energy)
of the disc gas at the outer boundary, $M_\text{env} = M_\text{mix}+M_\text{HHe}$
is the total envelope mass, and $E_\text{env}$ is the envelope's total energy defined by
\begin{equation}
E_\text{env} = \int_{M_\text{core}}^{M_\text{p}} \qty( u - \frac{GM_R}{R}) \dd{M_R};
\end{equation}
$u$ is the specific internal energy.
In Eq.~\eqref{eq:energy_conserv}, we have neglected the work done by the atmospheric surface (i.e. the boundary between the atmosphere and disc gas) for simplicity.
This term generally accounts for only a few \% relative to the other terms \citep[e.g.,][]{Lee+2014} and hardly affects the results of this study.
Once planetesimal accretion is over, the luminosity decreases with the mass growth of the envelope until reaching the critical luminosity below which no hydrostatic solution is found \citep[][]{Ikoma+2000,Hubickyj+2005}. Thus, for $L (t)$ = $\max [0.95 L (t-\Delta t), L_\mathrm{core}]$, we integrate Eqs.~(\ref{eq:dPdm})-(\ref{eq:drdm}) to calculate
$E_\text{env}(t)$ and $M_\text{env}(t)$, assuming $L_R$ = $L (t)$, and thereby calculate $\Delta t$ from Eq.~(\ref{eq:energy_conserv2}).
The assumption of $L_R = L$ is valid until the critical luminosity while invalid after that \citep[][]{Ikoma+2000}.
The H$_2$O fraction $X_{\text{H}_2\text{O}}$ is set to zero for $M_R >M_\text{core}+ M_\text{mix}$, and is set to the input value otherwise.
Once the critical luminosity is reached, we shift to the runaway gas accretion regime (see the following section).
\subsubsection*{Runaway gas accretion phase} \label{sec:runaway_gas_acc}
The runaway gas accretion phase is divided into three sub-phases \citep[e.g.,][]{Tanigawa+Ikoma2007}. In the earlier phase the gas accretion is regulated by the Kelvin-Helmholtz contraction of the atmosphere.
Numerical simulations show that the contraction timescale is strongly dependent on planet mass \citep[][]{Tajima+1997,Ikoma+2000}.
Here we assume that the growth timescale, $\tau_\mathrm{KH}$, increases with the cube of planetary mass, following the previous planetary population synthesis models \citep[e.g.,][]{Ida+2013}.
\begin{equation}
\tau_\text{KH} = 1 \times 10^9 \qty(\frac{M_\text{p}}{M_\oplus})^{-3}~\si{yr}.
\label{eq:tKH}
\end{equation}
While \citet{Tajima+1997} showed a stronger dependence, our results are insensitive to the choice of the power index, provided the index is smaller than $-3$, partly because the duration of this phase is short.
When the planet grows large enough to open a gap in the surrounding disc, the growth is limited by the gas supply from the flow in the gap.
The supply rate is given by~\citep{Tanigawa+Ikoma2007,Tanigawa+Tanaka2016,Tanaka+2020}
\begin{equation}
\dv{M_\text{p}}{t} = D \Sigma_\text{g,gap},
\label{eq:dMp_dt_gap}
\end{equation}
where the coefficient $D$ is empirically given by
\begin{equation}
D = 0.29\qty(\frac{H_\text{disc}}{a_\text{p}})^{-2}
\qty(\frac{M_\text{p}}{M_*})^{4/3} a^2 \Omega_\text{K},
\end{equation}
and $\Sigma_\text{g,gap}$ is the gas surface density at the bottom of the gap, which is also empirically given by \citep{Kanagawa+2015}
\begin{equation}
\Sigma_\text{g,gap} = \frac{\Sigma_\text{g}}{1+0.04K},
\end{equation}
with $K$ being
\begin{equation}
K = \qty(\frac{M_\text{p}}{M_*})^2 \qty(\frac{h}{a})^{-5}
\alpha_\text{vis}^{-1}.
\label{eq:Kanagawa}
\end{equation}
Here, $\alpha_\text{vis}$ is a parameter for disc turbulent viscosity, which is not necessarily the same as $\alpha_\text{acc}$.
When the wind driven accretion is dominant in the global angular momentum transfer, $\alpha_\text{acc}$ is larger than $\alpha_\text{vis}$ by about one order of magnitude~\citep{Simon+2013,Armitage+2013,Hasegawa+2017}.
Thus, we set $\alpha_\text{vis}=0.1\alpha_\text{acc}$ in this study.
Equation~\eqref{eq:dMp_dt_gap} is valid during there is sufficient disc gas outside the gap.
As the disc depletes, the gas accretion rate is limited by the global disc accretion rate $\dot{M}_\text{disc}$ given by
\begin{equation}
\dot{M}_\text{disc} = 6\pi r^{1/2}
\pdv{r}(\Sigma_\text{g} \nu_\text{acc} r^{1/2}).
\end{equation}
Considering these processes, the runaway gas accretion rate is calculated by
\begin{equation}
\dv{M_\text{p}}{t} = \min\qty(\frac{M_\text{p}}{\tau_\text{KH}}, D\Sigma_\text{g,gap}, \dot{M}_\text{disc}).
\label{eq:gas_acc_rate}
\end{equation}
Note that the solid accretion never occurs in this phase.
\subsection*{Atmospheric Thermal Evolution and Loss} \label{sec:radius}
\subsubsection*{Atmospheric Thermal Evolution}
The planetary radius $R_\text{p}$ is expressed by
\begin{equation}
R_\text{p} = R_\text{core} + \Delta R_\text{env},
\end{equation}
where $R_\text{core}$ is the solid core radius and $\Delta R_\text{env}$ is the thickness of the envelope.
The solid core is assumed to consist of iron, silicate, and, if present, ice.
Following ref.\cite{Zeng+2019}, we calculate the core radius as
\begin{equation}
R_\text{core}
= R_\text{rock}(1+0.55 f_\text{ice} - 0.14f_\text{ice}^2),
\label{eq:Rcore}
\end{equation}
where $f_\text{ice}$ is the ice mass fraction and $R_\text{rock}$ is the pure rocky (iron+silicate) core radius, which is calculated from ref.~\cite{Fortney+2007} as
\begin{align}
\begin{split}
R_\text{rock}
&= (0.0592f_\text{rock}+0.0975) (\log M_\text{core})^2 \\
& \qquad + (0.2337f_\text{rock}+0.4938) \log M_\text{core} \\
& \qquad + (0.3102f_\text{rock}+0.7932),
\end{split}
\label{eq:Rrock}
\end{align}
with $f_\text{rock}$ being the Si/(Si+Fe) mass ratio set to the Earth-like value of 0.66 in this study.
To evaluate the planetary radius including the envelope, we calculate the quasi-static thermal evolution of the planet after the disc gas dissipation.
For numerical convenience, we treat the upper part of the envelope (simply called the atmosphere below) separately from its deeper part (called the envelope), following previous evolution models of ice giant planets \citep[e.g.,][]{Kurosaki+Ikoma2017}.
The envelope structure is calculated by directly integrating Eqs.~\eqref{eq:dPdm}--\eqref{eq:drdm} and Eq.~\eqref{eq:dLdm} for a given water mass fraction $X_{\text{H}_2\text{O}}$.
For the equation of state, we use SCvH~\citep{Saumon+1995} for H-He and SESAME~\citep{Lyon+Johnson1992} for H$_2$O and mix H-He and H$_2$O according to the additive-volume law and the ideal mixing for entropy.
The inner boundary condition for the envelope structure is
\begin{equation}
L_R = L_\text{core}, \quad R = R_\text{core} \qquad \text{at}\, M_R = M_\text{core}.
\end{equation}
Here $R_\text{core}$ is calculated with Eq.~\eqref{eq:Rcore}.
The core luminosity $L_\text{core}$ is the same as Eq.~\eqref{eq:Lcore} with $L_\text{acc}=0$.
We also calculate $L_\text{cool}$ self-consistently during the quasi-static evolution with
\begin{align}
\begin{split}
L_\text{cool} &= -M_\text{core} C_\text{rock}\dv{T_\text{surf}}{t} \\
&= -M_\text{core} C_\text{rock}\frac{T_\text{surf}(t) - T_\text{surf}(t-\Delta t)}{\Delta t}
\end{split}
\end{align}
where $T_\text{surf}$ is the temperature at the solid core surface (i.e. the bottom of the envelope).
Since $T_\text{surf}(t)$ is determined by the envelope structure, we should find a self-consistent value of $L_\text{cool}$ iteratively.
The outer boundary condition for the envelope structure is given in terms of pressure and temperature, respectively, by
\begin{equation}
P = P_\text{out}, \quad T = T_\text{out} \qquad \text{at}\, M_R = M_\text{p}.
\end{equation}
To evaluate $P_\text{out}$ and $T_\text{out}$, we calculate the radiative-convective structure of the atmosphere in the same way as~\cite{Kurosaki+Ikoma2017}.
The atmosphere is assumed to be plane parallel and its mass and thickness are negligible compared to the total planetary mass and radius, respectively.
Thus, the gravitational acceleration $g=GM_\text{p}/R_\text{out}^2$ is constant through the atmosphere, with $R_\text{out}$ denoting the radius at the boundary between the atmosphere and the envelope (note that it is not equal to the planetary radius $R_\text{p}$; see below).
This boundary is assumed to locate at the radius where the optical depth for the stellar visible radiation, $\tau_\text{vis}$, is 10,
so that the temperature gradient in the envelope is hardly affected by the irradiation from the central star.
The temperature in the radiative region is calculated by the analytical formula given by ref.~\cite{Matsui+Abe1986b};
\begin{equation}
\begin{split}
\sigma T^4 &= F_\text{p} \frac{D\tau_\text{IR}+1}{2}\\
& \quad+\frac{\sigma T_\text{eq}^4}{2}
\qty[
1 + \frac{D}{\gamma} + \qty( \frac{\gamma}{D} - \frac{D}{\gamma})
\exp(-\tau_\text{vis})
],
\end{split}
\end{equation}
where $\tau_\text{IR}$ is the optical depth for infrared radiation, $F_\text{p}=L_\text{p}/4\pi R_\text{out}^2$ is the net energy flux from the planet, with $L_\text{p}$ denoting the luminosity at the top of the envelope, $T_\text{eq}$ is the equilibrium temperature calculated with Eq.~\eqref{eq:Teq}, $D=3/2$ is the diffusivity factor arising from the angular dependence of the radiation flux, and $\gamma = \kappa_\text{vis}/\kappa_\text{IR}$ with $\kappa_\text{vis}$ and $\kappa_\text{IR}$ being the opacities for visible and infrared radiation, respectively.
We set $\gamma=0.1$ following \cite{Kurosaki+Ikoma2017}.
The radiative-convective boundary (i.e. tropopause) is determined so that the temperature and the radiation flux connects continuously, using the same numerical procedure as \cite{Kurosaki+Ikoma2017}.
As a result, $P_\text{out}$ and $T_\text{out}$
are determined when $L_\text{p}$ and $R_\text{out}$ are given from the envelope structure calculation.
Thus, we find a self-consistent set of $(P_\text{out}, T_\text{out}, L_\text{cool}, L_\text{p}, R_\text{out})$ at the time $t$ for a given $M_\text{p}$ with iterations.
Then we define the planetary radius $R_\text{p}$ as the pressure level of 10~mbar, and the envelope thickness is derived simply by $\Delta R_\text{env} = R_\text{p}-R_\text{core}$.
If the planet has experienced the runaway accretion phase, however, the thick envelope with nebular composition exists instead of the water-enriched envelope.
The radius in this case is calculated with the fitting formula given by \cite{Lopez+Fortney2014} as;
\begin{align}
\begin{split}
\Delta R_\text{env} &= 2.06R_\oplus
\qty(\frac{M_\text{p}}{M_\oplus})^{-0.21}
\qty(\frac{f_\text{env}}{0.05})^{0.59} \\
&\qquad \times \qty(\frac{F_\text{p}}{F_\oplus})^{0.044}
\qty(\frac{t}{\si{yr}})^{-0.11} + \Delta R_\text{corr},
\end{split}
\end{align}
where $f_\text{env}=M_\text{env}/M_\text{p}$ and $F_\text{p}$ is the stellar insolation.
$\Delta R_\text{corr}$ is the correction factor defined as
\begin{equation}
\Delta R_\text{corr} = \frac{9k_\text{B}T_\text{eq}}{g \mu_\text{neb}m_\text{H}},
\end{equation}
with $g$ the gravitational acceleration and $\mu_\text{neb}=2.34$ the mean molecular weight of the nebular gas.
\subsubsection*{Atmospheric Photoevaporation}
After the disc dispersal, the atmospheric escape (or photoevaporation) occurs because of stellar XUV irradiation.
We assume that the atmospheric photoevaporation starts once the radial optical depth at the planet's location measured from the central star, $\tau_r$, becomes less than unity.
We use the fitting formula by \cite{Kubyshkina+2018b,Kubyshkina+2018a} for the escape rate, which yields higher values of the escape rate than those of the energy-limited escape rate when the escape parameter is small ($\lesssim 10$).
Although the atmospheres in our calculations are enriched with water vapour, we simply adopt the formula derived for hydrogen-dominated atmospheres.
Since the enrichment with water leads to raising the value of the escape parameter and, thus, reducing the escape rate, our calculations are expected to overestimate the escape rate.
The fitting formula is expressed in the form of
\begin{equation}
\dot{M}_\text{esc} = e^\beta
\qty(\frac{F_\text{XUV}}{\si{erg.cm^{-2}.s^{-1}}})^{\alpha_1}
\qty(\frac{a}{\si{au}})^{\alpha_2}
\qty(\frac{R_\text{p}}{R_\oplus})^{\alpha_3}
\Lambda^k,
\end{equation}
where $\beta, \alpha_1, \alpha_2, \alpha_3, k$ are the fitting coefficients~\citep[see][for their values]{Kubyshkina+2018a}, $F_\text{XUV}$ is the incoming XUV flux, and $\Lambda$ is the escape parameter.
The stellar XUV luminosity, $L_\text{XUV}$, is taken from the table given by \cite{Johnstone+2021}.
Since the table just gives the best-fit relation between the stellar $L_\text{XUV}$ and time, we set $L_\text{XUV} = 10^{\delta \log L_\text{XUV}} L_\text{XUV,bf}$ to account for the variation between observed stars, according with \cite{Johnstone+2021}.
Here, $L_\text{XUV,bf}$ is the best-fit value of XUV luminosity given in the table, and $\delta \log L_\text{XUV}$ is the deviation factor which follows the normal distribution with $\mu=0$ and $\sigma = 0.359$.
\subsection*{Orbital Migration} \label{sec:migration}
The planetary orbital migration rate can be calculated by
\begin{equation}
\dv{a_\text{p}}{t} = \frac{2\Gamma}{M_\text{p} a_\text{p} \Omega_\text{K}},
\label{eq:da_dt}
\end{equation}
where $\Gamma$ is the total torque acting on the planet.
In the type-I regime, the torque is given as the sum of the Lindblad ($\Gamma_\mathrm{L}$), corotation ($\Gamma_\mathrm{C})$, and thermal ($\Gamma_\mathrm{T}$) components:
\begin{equation}
\Gamma = \Gamma_\text{L} + \Gamma_\text{C} + \Gamma_\text{T}.
\label{eq:Gamma_typeI}
\end{equation}
We adopt the formula for the Lindblad torque derived by \cite{Jimenez+Masset2017} through 3D hydrodynamic simulations as
\begin{equation}
\Gamma_\text{L} = (-2.34+1.50\beta_{T}-0.10\beta_{\Sigma})f(\chi_\text{P})\Gamma_0,
\end{equation}
where $\beta_{T}=\dv*{\log T_\text{disc}}{\log r}$, $\beta_{\Sigma}=\dv*{\log \Sigma_\text{g}}{\log r}$, $\chi_\text{P}$ is the thermal diffusion coefficient at the disc midplane, and $\Gamma_0$ is the characteristic torque expressed as
\begin{equation}
\Gamma_0 = \qty(\frac{M_\text{p}}{M_*})^2 \qty(\frac{H_\text{disc}}{a_\text{p}})^{-2}\Sigma_\text{g} a_\text{p}^4 \Omega_\text{K}^2.
\end{equation}
The function $f(\chi_\text{P})$ transitions from $1/\gamma$ (`adiabatic regime') to unity (`locally isothermal regime') as $\chi_\text{P}$ increases (i.e. the thermal diffusion of the disc gas becomes effective).
Here $\gamma$ is the specific heat ratio of the disc gas.
See \cite{Jimenez+Masset2017} for the explicit form of $\chi_\text{P}$ and $f(\chi_\text{P})$.
We also use the formula for the corotation torque derived by \citet{Jimenez+Masset2017}, which is expressed as the sum of four components:
\begin{equation}
\Gamma_\text{C} = \Gamma_\text{C,vor}+\Gamma_\text{C,ent}+\Gamma_\text{C,temp}+\Gamma_\text{C,vv},
\end{equation}
where $\Gamma_\text{C,vor}$, $\Gamma_\text{C,ent}$, and $\Gamma_\text{C,temp}$ are the torques arising from the radial gradient of vortensity, entropy, and temperature.
The last term $\Gamma_\text{C,vv}$ is associated with the viscously created vortensity arising during the horse-shoe U-turns~\citep[see][for more detailed explanations]{Masset+Casoli2009a,Masset+Casoli2009b,Masset+Casoli2010}.
Each of the first three torques is calculated by the combination of two terms; for example,
\begin{equation}
\Gamma_\text{C,vor} = (1-\epsilon_b) \Gamma_\text{vor,lin} + \epsilon_b \Gamma_\text{vor,hs},
\end{equation}
where $\Gamma_\text{vor,lin}$ and $\Gamma_\text{vor,hs}$ are the vortensity-related components of the linear corotation torque and the horseshoe torque, respectively.
The horseshoe torque $\Gamma_\text{vor,hs}$ is the product of the unsaturated horseshoe drag $\Gamma_\text{vor,uhs}$ and the saturation function $\mathcal{F}_\text{v}$ (i.e., $\Gamma_\text{vor,uhs} \mathcal{F}_\text{v}$); the latter expresses the saturation of horseshoe drag and depends on horseshoe width.
The blending coefficient $\epsilon_b$ is derived in \cite{Masset+Casoli2010} by fitting their numerical results.
Note that the saturation function and the blending coefficient are different among $\Gamma_\text{C,vor}$, $\Gamma_\text{C,ent}$ and $\Gamma_\text{C,temp}$.
The final term $\Gamma_\text{C,vv}$
has only the component of the horseshoe torque (i.e., no linear corotation torque)
because this torque is only arising from the horse-shoe region.
Again, see \cite{Jimenez+Masset2017} for the explicit forms of these torques.
The thermal torque $\Gamma_\text{T}$ arises when the horseshoe gas cools during the U-turn by thermal diffusion (cold torque) and when the horseshoe gas is heated by the luminous planet (heating torque).
We calculate the thermal torque using the analytical formulation by \cite{Masset2017};
\begin{equation}
\Gamma_\text{T} = 1.61\frac{\gamma-1}{\gamma}\eta \qty(\frac{H_\text{disc}}{\lambda_c})
\qty(\frac{L_\text{p}}{L_\text{c}}-1)\Gamma_0,
\end{equation}
where $\eta = -\beta_{\Sigma}/3 - \beta_{T}/6 + 1/2$, $\lambda_c = \sqrt{\chi_\text{P}/(q\Omega_\text{K}\gamma)}$ with $q = M_\text{p}/M_*$, $L_\text{p}$ is the planetary luminosity,
and $L_\text{c} = 4\pi GM_\text{p}\chi_\text{P}\rho_\text{disc}/\gamma$ is the critical planetary luminosity above which the positive heating torque exceeds the negative cold torque.
Note that, as discussed in \cite{Masset2017}, the thermal torque acts effectively only when
$M_\text{p} < \chi_\text{P}c_\text{s}/G$ is satisfied.
This condition means that the thermal diffusion timescale across the Bondi radius is shorter than the acoustic timescale.
If $M_\text{p}>\chi_\text{P}c_\text{s}/G$, we set $\Gamma_\text{thermal} = 0$.
When the planet grows massive enough to open a gap in the disc, its orbital migration transitions to the type-II regime.
Recently, \cite{Kanagawa+2018} found that the migration rate in this regime can be expressed in the similar way as that in type I regime by replacing gas surface density $\Sigma_\text{g}$ with $\Sigma_\text{gap}$ and removing the corotation torque.
To connect the both regimes smoothly, we express the total torque in the type II regime as
\begin{equation}
\Gamma = \frac{\Gamma_\text{L} + (\Gamma_\text{C} + \Gamma_\text{T})\exp(-K/K_t)}
{1+0.04K},
\label{eq:Gamma_typeII}
\end{equation}
where $K_t=20$ is the typical value of $K$ where the disc gas gap becomes deep enough \citep{Kanagawa+2018}.
This is similar to the formula given in \cite{Kanagawa+2018},
but the formulae of $\Gamma_\text{L}$ and $\Gamma_\text{C}$ are different, and $\Gamma_\text{T}$ is also added.
Since Eq.~\eqref{eq:Gamma_typeII} becomes equal to Eq.~\eqref{eq:Gamma_typeI} when $K$ is small (i.e., the planet mass is small), we always use Eq.~\eqref{eq:Gamma_typeII} in the calculations of this study.
\subsection*{Resonance trapping} \label{sec:resonance}
Planets or planetary embryos with converging orbits can be captured in mean-motion resonances with each other; this phenomenon is called resonance trapping \citep{Murray+Dermott1999}.
In this study, the resonance trapping process is included in a similar way to refs. \cite{Ida+Lin2010} and \cite{Ida+2013}.
The treatment differs depending on whether the pair includes a giant planet or not---the definition of a `giant planet' is given in Section~\ref{sec:dyanamic_interact},
while an `embryo' refers to a celestial body other than the giant planet.
\subsubsection*{Case with two embryos}
When the orbits of two adjacent embryos (denoted by $i$ and $j$) are converging and get close enough to each other, their separation ($b = |a_i-a_j|$) expands impulsively at every conjunction.
The change in separation via the orbital repulsion is estimated by a linear analysis as~\citep{Goldreich+Tremaine1982,Hasegawa+Nakazawa1990}
\begin{equation}
\Delta b = 30\qty(\frac{b}{r_\text{H}})^{-5} r_\text{H},
\end{equation}
where $r_\text{H} = ((M_i+M_j)/3M_*)^{1/3} a_{ij}$ with $a_{ij} = \sqrt{a_i a_j}$.
The time interval between two conjunctions (i.e., the synodic period) is approximately given by
\begin{equation}
T_\text{syn} \simeq \frac{2\pi}{(\dv*{\Omega_\text{K}}{r})b}
\simeq \frac{4\pi a_{ij}}{3b\Omega_\text{K}}.
\end{equation}
Therefore, the expansion rate of the separation is found to be
\begin{equation}
\dv{b}{t} \sim \frac{\Delta b}{T_\text{syn}}
= 7 \qty( \frac{b}{r_\text{H}} )^{-4}
\qty( \frac{r_\text{H}}{a_{ij}} )^2 v_\text{K}.
\end{equation}
When the converging speed of the two orbits $\Delta v_\text{mig}$
($\equiv |(\dot{a}_j)_\text{mig}-(\dot{a}_i)_\text{mig}|$) becomes equal to $\dv*{b}{t}$, the two embryos are assumed to stop approaching, ending up captured in a mean-motion resonance with each other.
The resultant separation is, thus,
\begin{equation}
b_\text{trap} = 0.16 \qty( \frac{M_i+M_j}{M_\oplus} )^{1/6}
\qty( \frac{\Delta v_\text{mig}}{v_\text{K}} )^{-1/4} r_\text{H}.
\end{equation}
Note that we do not specify in which mean motion resonance those two embryos are really captured, but assume the mean motion resonance with the width nearly equal to $b_\text{trap}$.
In the planetary population synthesis calculations, we calculate $b_\text{trap}$ for all the converging pairs. If $b < b_\text{trap}$ is satisfied, the pair is treated as captured in the resonance.
If $b_\text{trap} < 2\sqrt{3}r_\text{H}$, the two embryos or planets collide and merge with each other.
After trapped in the resonance, the pair migrates keeping the ratio of their orbital periods (i.e., the ratio of their semi-major axes) unchanged.
In that case, their migration rates are different from Eq.~\eqref{eq:da_dt}.
We calculate the loss of the angular momenta of the resonantly trapped planets via the interaction between the respective planets and the disc,
and, then, redistribute the loss to both planets so that the planets migrate with the fixed semi-major axis ratio.
The migration rate of the resonantly trapped planet $i$
denoted by $v_\text{mig,trap}$ and that of the planet $j$
by $C_{ij}v_\text{mig,trap}$ ($C_{ij} \equiv a_j/a_i$),
that angular momentum loss rate, $\dot{\mathcal{L}}_\text{trap}$, can be expressed as
\begin{align}
\dot{\mathcal{L}}_\text{trap}
&= \dv{t}(M_i \sqrt{GM_*a_i}+M_j\sqrt{GM_*a_j}) \\
&= \frac{M_i \sqrt{GM_*a_j}+C_{ij}M_j\sqrt{GM_*a_i}}{2a_{ij}} v_\text{mig,trap}.
\end{align}
Here we assume that the eccentricities of both planets are negligible.
Thus, $v_\text{mig,trap}$ becomes
\begin{equation}
v_\text{mig,trap} =
\frac{2a_{ij} (\Gamma_{i}+\Gamma_{j})}{M_i \sqrt{GM_*a_j}+C_{ij}M_j\sqrt{GM_*a_i}},
\end{equation}
where $\Gamma_i$ and $\Gamma_j$ are the torques exerted on the planets $i$ and $j$, respectively, by disc gas, calculated with Eq.~\eqref{eq:Gamma_typeII}.
The orbital migration terminates at the disc inner edge.
The subsequent planets moving inward are trapped in the resonance, and multiple planets line up near the disc inner edge.
However, if a heavy enough planet joins such a resonance chain,
the planet pushes planets ahead of itself and, in particular, the innermost one into the disc cavity.
Following \cite{Ida+2013}, we adopt the condition for this `leakage' of the planet based on \cite{Ogihara+2010}; namely, once the following condition is satisfied, the planet at the disc edge halts the migration of the subsequent planet:
\begin{equation}
\frac{e_1}{2\pi}\frac{M_1 \mathcal{L}_1}{\tau_e(M_1)}
-\sum_{i=1}^{N_\text{edge}} \frac{M_i \mathcal{L}_i}{2\tau_\text{mig}(M_i)} > 0,
\label{eq:ecc_trap_cond}
\end{equation}
where $M_i$ is the mass of the $i$-th planet ($i=1$ is the planet at the inner edge), $\mathcal{L}_i \sim \sqrt{GM_*a_i}$ is the orbital angular momentum, $\tau_\text{mig}$ is the migration timescale, $\tau_e$ is the eccentricity damping timescale due to migration, which is given by~\cite{Tanaka+Ward2004} as
\begin{equation}
\tau_e = \frac{1}{0.78}\qty(\frac{M_\text{p}}{M_*})^{-1}
\qty(\frac{\Sigma_\text{g} a^2}{M_*})^{-1}
\qty(\frac{h}{a})^4 \Omega_\text{K}^{-1},
\end{equation}
$e_1$ is the eccentricity of the inner-most planet, which is fitted by ref.~\cite{Ogihara+2010} as
\begin{equation}
e_1 = 0.02 \qty(\frac{\tau_e/\tau_\text{mig}}{10^{-3}})^{1/2},
\end{equation}
and, finally, $N_\text{edge}$ is the number of planets trapped near the disc inner edge.
If Eq.~\eqref{eq:ecc_trap_cond} is not satisfied, the inner-most planet is pushed into the disc cavity.
We assume that the semi-major axis ratio for this pushed-out planet and the next planet is kept constant, until orbital crossing or a giant impact occurs.
\subsubsection*{Case including giant planets}
The resonance trapping conditions for a pair of an embryo and a giant planet and that of two giant planets are the same as in the two-embryo case.
The difference is in the outcome that happens when $b_\text{trap} < 2 \sqrt{3} r_\text{H}$. Note that these cases are rare in planetary synthesis models for M dwarfs because giant planets are rarely formed.
A pair of two giant planets, if $b_\text{trap} < 2\sqrt{3}r_\text{H}$, undergo close encounters.
The post-process is calculated following the procedure for dynamical interactions between protoplanets (see Section~\ref{sec:dyanamic_interact} (ii)).
For a pair of an embryo and a giant planet, the trapping condition is calculated based on \cite{Shiraishi+Ida2008}, which gives a criterion for an embryo to enter the feeding zone of a giant planet.
This condition is expressed by comparing the decreasing rate in the separation of two bodies relative to the Hill radius of the giant planet $v_\text{H}$ and the eccentricity damping rate of the embryo $v_\text{damp}$.
Given that an embryo of mass $M_\text{E}$ and semi-major axis $a_\text{E}$ interacts with a giant planet of $M_\text{G}$ and $a_\text{G}$,
$v_\text{H}$ is given by
\begin{align}
v_\text{H} = \dv{\tilde{b}^2}{t}
= \frac{2\tilde{b}}{h_\text{G}}\frac{a_\text{E}}{a_\text{G}}
\qty(
\frac{\dot{a}_\text{E}}{a_\text{E}} -
\frac{\dot{a}_\text{G}}{a_\text{G}}
)
-\frac{2\tilde{b}^2}{3}\frac{\dot{M}_\text{G}}{M_\text{G}},
\end{align}
where $\tilde{b} = (a_\text{E}-a_\text{G})/h_\text{G}a_\text{G}$ with $h_\text{G}=(M_\text{G}/3M_*)^{1/3}$.
On the other hand, $v_\text{damp}$ is expressed as
\begin{equation}
v_\text{damp} = -\frac{(e_\text{E}/h_\text{G})}{\tau_\text{damp}}
\end{equation}
with the damping timescale given by~\citep{Artymowicz1993,Iwasaki+2002}
\begin{align}
\begin{split}
\tau_\text{damp} &= 400~\si{yr}
\qty( \frac{H_\text{disc}/a}{0.03} )^{4}
\qty( \frac{\Sigma_\text{g}}{1000~\si{g/cm^2}} )^{-1} \\
& \quad \times \qty( \frac{M_\text{E}}{M_\oplus} )^{-1}
\qty( \frac{a_\text{E}}{1~\si{au}} )^{-1/2}
\qty( \frac{M_*}{M_\odot} )^{-1/2}.
\end{split}
\end{align}
If $|v_\text{H}|>|v_\text{damp}|$ is satisfied, the embryo enters the feeding zone of the giant planet and undergoes close encounters.
The post-process is also calculated following the procedure given in Section~\ref{sec:dyanamic_interact} (ii).
\subsection*{Dynamic interaction of multi-body systems and its outcome}\label{sec:dyanamic_interact}
As the disc gas depletes, the typical timescale for eccentricity damping via gas drag becomes larger than the timescale for the orbit destabilisation and crossing of the closest pair to happen (called the orbital crossing timescale).
In this case, orbital crossing occurs between the pair.
We use the semi-analytical model from \cite{Ida+Lin2010} and \cite{Ida+2013} regarding the orbital repulsion, merging events, and gravitational scatterings to calculate the resultant semi-major axis and planetary mass \citep[see][for the details]{Ida+Lin2010,Ida+2013}.
Here we briefly summarise the numerical procedure.
The treatment of dynamical interactions differs depending on the number of giant planets in the system $N_\text{giant}$: (i) $N_\text{giant}=0,1$, (ii) $N_\text{giant}=2$, and (iii) $N_\text{giant} \ge 3$.
Here we define `giant planets' as the planets that satisfy both conditions (1) $M_\text{p}>30M_\oplus$ and (2) $e_\text{esc} > 1$, where
\begin{align}
e_\text{esc} &= \frac{v_\text{esc}}{v_\text{K}}
= \frac{\sqrt{2GM_\text{p}/R_\text{p}}}{\sqrt{GM_*/a_\text{p}}} \notag \\
\begin{split}
&= 1.6 \qty( \frac{M_\text{p}}{M_\text{J}} )^{1/3}
\qty( \frac{\bar{\rho}}{1~\si{g/cm^3}} )^{1/6}\\
& \qquad \qquad
\times \qty( \frac{a_\text{p}}{1~\si{au}} )^{1/2}
\qty( \frac{M_*}{M_\odot})^{-1/2},
\end{split}
\end{align}
with $M_\text{J}$ the Jovian mass and $\bar{\rho}$ the mean density of the planet.
\begin{enumerate}[label=(\roman*)]
\item If the system has no or only one giant planet, we first calculate the orbital crossing timescale for every adjacent pair of embryos $(i,j)$.
The timescale $\tau_\text{cross}$ follows the fitting formula given by \cite{Zhou+2007} as
\begin{equation}
\log(\frac{\tau_\text{cross}}{T_\text{K}}) = A + B \log(\frac{b}{2.3r_\text{H}}),
\label{eq:tau_cross}
\end{equation}
where $T_\text{K}$ is the Keplerian period at the semi-major axis of $a=\sqrt{a_j a_i}$, $b=|a_i-a_j|$, $r_\text{H} = ((M_i + M_j)/3M_*)^{1/3} \min(a_i,a_j)$, and
\begin{align}
\begin{split}
A &= -2.0 + e_0 -0.27 \log \mu \\
&\quad + 0.51i_0+0.19i_0\log \mu + 0.03i_0 (\log \mu)^2 ,
\end{split}
\\
\begin{split}
B &= 18.7 + 1.1\log\mu - (16.8 + 1.2\log \mu) e_0 \\
& \quad - 0.28i_0 + 0.19i_0 \log\mu,
\end{split}
\\
e_0 &= \frac{1}{2}\frac{e_i + e_j}{b}a, \quad
\mu = \frac{1}{2} \frac{M_i + M_j}{M_*}.
\end{align}
Here $i_0$ is the mean inclination of the two planets in a unit of degree. The inclination of a planet in a unit of radian is set to the half of the eccentricity. The planetary eccentricity before any orbital crossing events is assumed to follow the Rayleigh distribution with root mean square (rms) of $\sigma = (M_i/3M_*)^{1/3}$.
Then, after the time interval equal to $\tau_\text{cross}$, the pair $(i,j)$ undergoes orbital crossing or sometimes ends up merging.
Their resultant masses, semi-major axes, and eccentricities are calculated following the procedure presented in \cite{Ida+Lin2010}.
They are evaluated so that each of the total mass, orbital energy, and Laplace-Runge-Lenz vector is conserved.
Finally, if the orbit of the embryo is within $3.5R_\text{H}$ of the giant planet, the embryo is scattered by the giant planet, and its semi-major axis and eccentricity are modified again, following ref.\cite{Ida+2013}.
\item If the system has two giant planets (1 and 2) and their separation is $b = |a_1-a_2| < 2\sqrt{3}r_\text{H}$, with $r_\text{H}=((M_1+M_2)/3M_*)^{1/3}\sqrt{a_1 a_2}$, then,
the orbital instability occurs.
The orbital elements after the instability are calculated following `Two Giants Case' in \cite{Ida+2013}.
If the orbital instability occurs, all other embryos are assumed to be ejected from the system.
Otherwise, the interactions between embryos are calculated in the same way as in Case (i).
\item If the system has more than two giant planets, the orbital crossing timescale is calculated for every pair of giant planets using Eq.~\eqref{eq:tau_cross}.
If any of the derived $\tau_\text{cross}$ is larger than the total integration time, the giant planets do not interact with each other, and only the interaction between embryos are calculated following Case (i).
Otherwise, the orbital instability occurs after $\tau_\text{cross}$, and the resultant masses and orbital elements are derived from the `Three Giants Case' model in \cite{Ida+2013}.
All the other embryos are ejected from the system in this case.
\end{enumerate}
\subsection*{Initial Conditions and Parameters} \label{sec:setting}
To start the planetary population synthesis simulations, we perform random samplings of the initial mass $M_\text{disc}$, radius $r_\text{disc}$, metallicity [Fe/H], and inner edge radius $r_\text{in}$ of the protoplanetary gas disc,
the external photo-evaporation rate $\dot{M}_\text{wind}$,
and the initial masses and semi-major axes of embryos in the following way.
We use the default subroutine \verb|random_number| in FORTRAN90 to generate random numbers of uniform distribution in $[0,1]$.
To generate random numbers following the normal and Rayleigh distributions, we use Monty Python method and inverse transform method, respectively.
\subsubsection*{Initial conditions for protoplanetary disc} \label{sec:init_cond_disc}
We determine the initial properties of the protoplanetary disc by scaling recent observation results for stars of $\sim$~1~$M_\odot$. We adopt the fitting formula for the disc gas mass for $1M_\odot$ stars,
the log-normal distribution with the mean $\log(\mu/M_*)=-1.49$ and the standard deviation $\sigma = 0.35$, which \citet{Emsenhuber+2021b} derived from observational results of \cite{Tychoniec+2018}.
The minimum and maximum disc masses are set to $\num{4e-3}M_\odot$ and $0.16M_\odot$, respectively, which roughly correspond to the lightest and heaviest samples in \cite{Tychoniec+2018}.
We also assume that the mean, minimum, and maximum disc masses are proportional to the stellar masses~\citep{Andrews+2013}.
The disc gas radius is calculated with
\begin{equation}
r_\text{disc} = 10 \qty(\frac{M_\text{disc}}{\num{2e-3}M_\odot})^{0.625}~\si{au},
\end{equation}
which is taken from the observational trend derived in \cite{Andrews+2010}.
The disc metallicity [Fe/H] follows the normal distribution with $\mu=-0.02$ and $\sigma = 0.22$ derived from \cite{Santos+2005}.
The range of the value is limited to $-0.6<$[Fe/H]$<0.5$.
We use the same distribution regardless of the stellar mass.
We assume that the disc inner edge
locates at the corotation radius where the Kepler rotation period is equal to the rotation period of the central star.
Here the stellar rotation period is assumed to follow the log-normal distribution with $\log (\mu \si{[days]})=0.676$ and $\sigma = 0.306$ based on the observational results of young stellar objects by \cite{Venuti+2017}.
The minimum of $r_\text{in}$ is set to the initial stellar radius $R_*$.
We also use the same distribution for all stellar types.
The external photo-evaporation rate $\dot{M}_\text{wind}$ generally depends on the population of nearby massive stars.
Following \cite{Burn+2021}, we set the distribution of $\dot{M}_\text{wind}$ so that the mean value of the resultant disc lifetime locates at $\sim$3Myr~\citep{Mamajek2009,Ansdell+2018} and that it has a deviation of about half an order, regardless of the stellar mass.
Since the disc gas radius $r_\text{disc}$ is determined only by the disc mass $M_\text{disc}$, the disc lifetime depends on $M_\text{disc}$ and $\dot{M}_\text{wind}$ for given $\alpha_\text{acc}$.
Then we find that, for a star of $0.3M_\odot$ and $\alpha_\text{acc}=\num{2e-3}$, the log-normal distribution with $\log(\mu ~\si{[M_\odot/yr]})=-6.0$ and $\sigma = 0.5$ accounts for the above distribution.
Also, the stellar mass dependence of $\dot{M}_\text{wind} \propto M_*^{1.4}$ is found to be suitable for the star in the range of $0.1M_\odot \le M_* \le 0.5M_\odot$.
\subsubsection*{Initial conditions for planetary embryos} \label{sec:init_cond_planet}
Initially, 50 planetary embryos with mass of 0.01~$M_\oplus$ are placed log-uniformly from $r_\text{in}$ to $r_\text{solid}$.
Here the initial separations of all the adjacent embryos are larger than the feeding zone width ($10r_\text{H}$) for the local isolation mass $M_\text{iso}$ given by~\citep{Kokubo+Ida2002}
\begin{equation}
M_\text{iso} = 0.16 \qty(\frac{\Sigma_\text{s}}{10~\si{g.cm^{-2}}})^{3/2}
\qty(\frac{a}{1~\si{au}})^{3/4} \qty(\frac{M_*}{M_\odot})^{1/2} ~M_\oplus.
\end{equation}
Therefore, in quite massive solid discs, the initial number of planetary embryos can be smaller than 50.
\subsubsection*{Input parameters}
The parameters and their fiducial values are summarised in Table~\ref{tab:parameter}.
These values are used in our calculations unless otherwise mentioned.
\subsection*{Probability density}
To show the distribution of water mass fraction in the synthesised planets (Figs.~\ref{fig:hist_M03} and \ref{fig:hist_Mstar}), we use the probability densities (PDs) calculated by
\begin{equation}
{\rm PD}(i)
= \frac{N_i}{N_{\rm tot}\Delta \log (M_{\rm water}/M_{\rm core})},
\end{equation}
where $N_i$ is the number of planets in the $i$-th bin, $N_{\rm tot}=\sum{N_i}$ is the total number of planets, and $\Delta \log(M_{\rm water}/M_{\rm core})$ is the bin width.
\section*{Data availability}
All data from the simulation are available at \url{https://github.com/TadahiroKimura/Kimura-Ikoma2022}.
Soure Data for each figure is provided with this paper.
\section*{Code availability}
The numerical code used in the current study is available from the corresponding author upon request only for the purpose of reproducing our results.
\section*{Acknowledgements}
This work is supported by JSPS KAKENHI Nos.~JP18H05439, JP21H01141 and JP22J11725. TK is a JSPS Research Fellow, and also
supported by International Graduate Program for Excellence in Earth-Space Science (IGPEES).
\section*{Author contributions}
Both authors contributed equally to this work. M.I. conceived the original idea and supervised this project. T.K. developed the entire model of planetary population synthesis partly using a few modules that M.I. had developed. T.K. carried out the numerical simulations and analyzed the simulation results. Both authors discussed the results and implications and wrote the paper.
\section*{Corresponding author}
Correspondence and requests for materials should be addressed to T. Kimura.
\section*{Competing interests}
The authors declare no competing interests.
\begin{table*}
\centering
\caption{Parameters used in calculations}
\label{tab:parameter}
\begin{tabular}{c|>{\centering}p{5cm}|c|c}\hline
Symbol & Meaning & Value & Eq. mainly used \\
\hline
$\alpha_\text{acc}$ &
parameter for effective turbulent viscosity & $\num{2.0e-3}$ &
Eq.~\eqref{eq:dSigma_dt_basic}\\% [0.8cm]
%
$\alpha_\text{vis}$ &
parameter for turbulent viscosity &
$\num{2.0e-4}$ & Eq.~\eqref{eq:Kanagawa}\\
%
$N_\text{grid,disc}$ &
number of grids for gas disc &
500 & \\
%
$N_\text{grid,solid}$ &
number of grids for solid disc &
1000 & \\
%
$r_\text{max}$ &
outer boundary radius for gas disc &
1000~au & Eq.~\eqref{eq:dSigma_pe_ext} \\
%
$r_\text{solid}$ &
solid disc radius &
$0.5r_\text{disc}$ & Eq.~\eqref{eq:Sigma_solid} \\
%
$\Phi$ &
ionising EUV photon luminosity &
$\num{1.0e40}~\si{s^{-1}}$ & Eq.~\eqref{eq:n0} \\
%
$m_\text{plt}$ &
planetesimal mass &
$\num{1.0e20}$~g & Eq.~\eqref{eq:tilde_e_plt} \\
%
$\rho_\text{plt}$ &
planetesimal material density &
$3.0~\si{g.cm^{-3}}$ & Eq.~\eqref{eq:tilde_e_plt} \\
%
$C_\text{rock}$ &
specific heat of solid core for constant volume&
$\SI{1.2e7}{erg/(g.K)}$ & Eq.~\eqref{eq:Lcool} \\
$K_t$ &
typical value of $K$ for which the corotation torque becomes ineffective &
20.0 & Eq.~\eqref{eq:Gamma_typeII} \\
%
\hline
\end{tabular}
\end{table*}
\clearpage
\section*{References}
| 8ede9ab47451d0a29a831a6d93ac367230857c46 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $k\geq 2$ be a~fixed integer. We recall that $d(n) := \sum_{d\mid n}1$ is the number of divisors of $n$, and $\omega (
n) :=\sum_{p\mid n}1$ is the number of distinct prime divisors
of $n$. We define the function $D_{k,\omega}(n) $ by
\begin{equation}
D_{k,\omega}(n) :=\frac{d(n)}{k^{\omega
(n)}}.
\end{equation}
Notice that for every fixed integer $k\geq 2$, the function $D_{k,\omega}(n) $ is multiplicative and for every prime number $p$ and
every integer $m$ the relation
\begin{equation}
D_{k,\omega}(p^{m}) = \frac{m+1}{k}, \label{11}
\end{equation}
holds. By using \eqref{11}, we get
$$
D_{k,\omega}(n) = \prod\limits_{p^{m}\Vert n}\frac{m+1}{k}
$$
where $p^{m}\Vert n$ means $p^{m}\mid $ $n$ and $p^{m+1}\nmid n$. In the
particular case $k = 2$, the function $D_{2,\omega}(n) $ is
exactly $D(n) = \dfrac{d(n)}{d^{\ast}(
n)},(\text{see~\cite{D.K}})$. For $k\geq 3$, we can
easily check that
\begin{equation}
\sum_{n\leq x}D_{k},_{\omega}(n) \ll _{k}x(\log x)
^{2/k-1}.
\end{equation}
Indeed, for any integer $n$, we have $D_{k},_{\omega}(n) \leq
d(n) \ll _{\varepsilon}n^{\varepsilon}$. Furthermore, the
hypotheses of Shiu's theorem are satisfied; see Theorem $1$ in~\cite{Shui}
and~\cite[p.1]{Nar-Ten}. One gets
$$
\sum_{n\leq x}D_{k},_{\omega}(n) \ll _{k}\frac{x}{\log x}\exp
\Bigl(\sum_{p\leq x}\frac{2}{kp}\Bigr).
$$
Now, by using Lemma $4.63$ in \cite{O.BORD}, it follows that
$$
\sum_{n\leq x}D_{k},_{\omega}(n) \ll _{k}\frac{x}{\log x}\exp
\Bigl(\frac{2}{k}\log (2e^{\gamma}\log x)\Bigr) \ll
_{k}x(\log x) ^{2/k-1}.
$$
\section{Main result}
In this section, we establish two results concerning the mean value of the
function $D_{k,\omega}(n)$. We begin by giving a~weaker result.
\begin{theorem}
Let $k\geq 2$ be a~fixed integer. For all $x\geq 1$ large enough, we have
\[
\sum_{n\leq x}D_{k,\omega}(n) = \frac{x(\log x)^{2/k-1}}{\Gamma
(2/k)}\prod\limits_{p}\Bigl(1-\frac{1}{p}\Bigr) ^{2/k}
\Bigl(1+\frac{2p-1}{kp(p-1) ^{2}}\Bigr) +O(x(\log
x)^{-1}) (\log \log x)^{4/k}.
\]
\end{theorem}
The proof of this result is based on Tulyaganov's theorem; this theorem is
summarized as follows:
\begin{theorem}\label{thm2}
Let $f$ be a~complex valued multiplicative function. Suppose there exists $ z\in\mathbb{C}$, independent of $p$, with $\vert z\vert \leq c_{1}$ and
\begin{enumerate}
\item[a)] $$\sum_{p\leq x}f(p)\log p = zx + O(xe^{-c_{2}\sqrt{\log x}}) $$
\item[b)] $$\sum_{p\leq x}\vert f(p)\vert \log p\ll x
$$
\item[c)]
$$
\sum_{p\leq x}\sum_{\alpha = 2}^{\infty}\frac{\vert
f(p^{\alpha})\vert \log p^{\alpha}}{p^{\alpha}}\ll (\log \log
x) ^{2}$$
\item[d)] $$\sum_{p}\frac{\vert f(p)\vert ^{2}\log p}{p^{2}}<c_{3}
$$
\end{enumerate}
for some real numbers $c_{1}$, $c_{2}$ and $c_{3}$. Then, for all $x\geq 1$
sufficiently large, we have
\begin{align*}
\sum_{n\leq x}f(n) =
{ }&{ }\frac{x(\log x)^{z -1}} {\Gamma (z )}\prod\limits_{p}\Bigl(1-\frac{1}{p}\Bigr)^{z }\Bigl(1+\sum_{\alpha = 1}^{\infty}\frac{f(p^{\alpha})}{p^{\alpha}}\Bigr) \Bigl\{1+O(\frac{(\log \log x) ^{2}}{\log x}) \Bigr\} \\
&{ }+ O(x(\log x)^{\max (0,\mathrm{Re}\,z-1) -1}) (\log \log x)^{2(A-\max (0,\mathrm{Re}\,z -1))},
\end{align*}
where $A>0$ satisfies
$$
\sum_{u<p\leq v}\vert f(p)\vert p^{-1}\leq A\log (\log
v/\log u) +O(1).
$$
\end{theorem}
\begin{proof}
This theorem is a~consequence of Theorem $4$ in~\cite{Tulyag}, where we take
$g = f$.
\end{proof}
\noindent To complete the demonstration of the main result we have the
following lemmas.
\begin{lemma}\label{lem1}
For any fixed integer $k\geq 2$, we have the estimate
$$
\sum_{p\leq x}\vert D_{k,\omega}(p) \vert \log p\ll
x.
$$
\end{lemma}
\begin{proof}
By Chebyshev's estimates~\cite{Disar}, we have
\[
\sum_{p\leq x}\vert D_{k,\omega}(p) \vert \log p = \frac{2}{k}\sum_{p\leq x}\log p<\frac{2}{k}(1.000081x) \ll x.
\qedhere
\]
\end{proof}
\begin{lemma}\label{lem2}
For any fixed integer $k\geq 2$, there is a~constant $c>0$, such that
$$
\sum_{p\leq x}D_{k,\omega}(p) \log p = \frac{{\footnotesize 2}}{k}x+O(xe^{-c\sqrt{\log x}}).
$$
\end{lemma}
\begin{proof}
We have
$$
\sum_{p\leq x}D_{k,\omega}(p) \log p = \frac{{\footnotesize 2}}{k}\sum_{p\leq x}\log p = \frac{{\footnotesize 2}}{k}\theta (x),
$$
and by Theorem $6.9$ in~\cite{H.L.M and R.C.V}, there is a~constant $c>0$
such that
$$
\theta (x) = x+O(xe^{-c\sqrt{\log x}}),
$$
which implies the desired result.
\end{proof}
\begin{lemma}\label{lem3}
For any fixed integer $k\geq 2$, we have
$$
\sum_{p}\frac{\vert D_{k,\omega}(p) \vert ^{2}}{p^{2}%
}\log p<\infty .
$$
\end{lemma}
\begin{proof}
We first check the inequality $\sum\limits_{m = 2}^{\infty}\dfrac{\log m}{m(m-1)}\leq \log 4$, and using the following
$$
\sum_{p}\frac{\log p}{p^{2}}<\sum_{m = 2}^{\infty}\frac{\log m}{m^{2}}\leq
\sum_{m = 2}^{\infty}\frac{\log m}{m(m-1)},
$$
then we have
\[
\sum_{p}\frac{\vert D_{k,\omega}(p) \vert ^{2}}{p^{2}%
}\log p = \frac{4}{k^{2}}\sum_{p}\frac{\log p}{p^{2}}
<\frac{4\log 4}{k^{2}}.
\qedhere
\]
\end{proof}
\begin{lemma}\label{lem4}
For any fixed integer $k\geq 2$, we have
$$
\sum_{p\leq x}\sum_{\alpha = 2}^{\infty}\frac{\vert D_{k,\omega}(
p^{\alpha}) \vert \log (p^{\alpha})}{p^{\alpha}}\leq \frac{28}{k}.
$$
\end{lemma}
\begin{proof}
For every integer $k\geq 3$, we write
\begin{align*}
\sum_{p\leq x}\sum_{\alpha = 2}^{\infty}\frac{\vert D_{k,\omega}(
p^{\alpha}) \vert \log (p^{\alpha})}{p^{\alpha}}
&= \frac{1}{k}\sum_{p\leq x}\log p\sum_{\alpha = 2}^{\infty}\frac{\alpha
(\alpha +1)}{p^{\alpha}} \\
&= \frac{1}{k}\sum_{p\leq x}\frac{\log p}{p}\sum_{\alpha = 2}^{\infty}\frac{%
\alpha (\alpha +1)}{p^{\alpha -1}},
\end{align*}
and the infinite series $\sum\limits_{\alpha = 2}^{\infty}\dfrac{\alpha
(\alpha +1)}{p^{\alpha -1}}$ converges to $\dfrac{2}{(
1-1/p) ^{3}}-2$, since
\begin{align*}
\sum_{p\leq x}\sum_{\alpha = 2}^{\infty}\frac{\vert D_{k,\omega}(
p^{\alpha}) \vert \log (p^{\alpha})}{p^{\alpha}}
&= \frac{2}{k}\sum_{p\leq x}\frac{3p^{2}-3p+1}{p(p-1)^{3}}\log p\\
&\leq \frac{28}{k}\sum_{p\leq x}\frac{\log p}{p^{2}}.
\end{align*}
By Lemma $70.1$ in~\cite{HALL -TENENBAUM}, we have $\sum\limits_{p}\dfrac{\log p}{p^{\alpha}}<\dfrac{1}{\alpha -1}$ for all $\alpha >1$, consequently
$$
\sum_{p\leq x}\sum_{\alpha = 2}^{\infty}\frac{\vert D_{k,\omega}(
p^{\alpha}) \vert \log (p^{\alpha})}{p^{\alpha}}<\frac{28}{k}.
$$
Finally, by Lemma \ref{lem1}, \ref{lem2}, \ref{lem3} and \ref{lem4} we have shown that the function $D_{k,\omega}(n) $ satisfies the conditions of Theorem \ref{thm2}. As
we have
$$
\sum_{u<p\leq v}\frac{\vert D_{k,\omega}(p) \vert}{p%
} = \frac{2}{k}\sum_{u<p\leq v}\frac{1}{p}\leq \frac{2}{k}\log \frac{\log v}{%
\log u}+O(1),
$$
then the constant $A$ in Theorem $2$ is $\frac{2}{k}$.
\end{proof}
The next result is improved over the previous one.
\begin{theorem}
Let $k\geq 2$ be a~fixed integer. For all $x\geq 1$ large enough, we have
$$
\sum_{n\leq x}D_{k,\omega}(n) = \frac{x(\log x)^{2/k-1}}{\Gamma
(2/k)}\prod\limits_{p}\Bigl(1-\frac{1}{p}\Bigr) ^{2/k}
\Bigl(1+\frac{2p-1}{k(p-1) ^{2}}\Bigr) +O_{k}(x(\log
x)^{2/k-2}).
$$
\end{theorem}
The demonstration is based on the following lemmas:
\begin{lemma}\label{lem5}
Let $k\geq 2$ be a~fixed integer. For every $s:= \sigma +it\in
\mathbb{C}
$ such that $\sigma >1$ and $L(s,D_{k,\omega}(n))
:= \sum\limits_{n = 1}^{\infty}\dfrac{D_{k,\omega}(n)}{n^{s}}$,
we have
$$
L(s,D_{k,\omega}(n)) = \zeta (s)
^{2/k}L(s,g_{k}),
$$
or $L(s,g_{k}) $ is a~series of Dirichlet absolutely convergent
in the half-plane $\sigma >\frac{1}{2}$.
\end{lemma}
\begin{proof}
If $\sigma >1$, then
\begin{align*}
L(s,D_{k,\omega}(n)) &= \prod\limits_{p}\Bigl(
1+\sum\limits_{\alpha = 1}^{\infty}\dfrac{D_{k,\omega}(p^{\alpha
})}{p^{\alpha s}}\Bigr) \\
&= \prod\limits_{p}\Bigl(1+\sum\limits_{\alpha = 1}^{\infty}\dfrac{\alpha
+1}{kp^{\alpha s}}\Bigr) \\
&= \prod\limits_{p}\Bigl(1+\dfrac{2p^{s}-1}{k(p^{s}-1) ^{2}}\Bigr),
\end{align*}
on the other hand we have
$$
\Bigl(1+\frac{2p^{s}-1}{k(p^{s}-1) ^{2}}\Bigr) = ((
1-p^{-s}) ^{-2/k}) \Bigl(1+\frac{h(s)}{k(
p^{s}-1) ^{2}}\Bigr),
$$
such that
$$
h(s) = (1-p^{-s}) ^{2/k}(kp^{2s}-2(
k-1) p^{s}+k-1) -k(p^{s}-1) ^{2}.
$$
Since
$$
(1-p^{-s}) ^{2/k} = 1-\frac{2}{kp^{s}}-\frac{k-2}{k^{2}p^{2s}}-O\Bigl(\frac{k}{p^{3\sigma}}\Bigr),
$$
he comes
\begin{align*}
h(s) &= \Bigl(1-\frac{2}{kp^{s}}-\frac{k-2}{k^{2}p^{2s}}-O\bigl(\frac{k}{p^{3\sigma}}\bigr)\Bigr) (kp^{2s}-2(
k-1) p^{s}+k-1) -k(p^{s}-1) ^{2} \\
&= 2\bigl(1-\frac{1}{k}\bigr) +O(p^{-\sigma}),
\end{align*}
which implies the announced result.
\end{proof}
\begin{lemma}[\cite{Selb}]\label{lem6}
Let $A>0$. Uniformly for $x\geq 2$ and $z\in
\mathbb{C}
$ such that $\vert z\vert \leq A$, we have
$$
\sum_{n\leq x}\tau _{z}(n) = \frac{x(\log x) ^{z-1}}{%
\Gamma (z)}+O_{A}(x(\log x) ^{\mathrm{Re}\,z-2}).
$$
$\tau _{z}(n) $ is the multiplicative function defined by $\tau
_{z}(p^{\alpha}) =
\binom{z+\alpha -1}{\alpha}$.
\end{lemma}
\begin{proof}[Proof of Theorem 3]
According to the Lemma \ref{lem5}, we have $D_{k,\omega} = \tau _{2/k}\ast g_{k}$.
Then, by Lemma \ref{lem6}
\begin{align*}
\sum_{n\leq x}D_{k,\omega}(n) &= \sum_{d\leq x}g_{k}(
d) \sum_{m\leq \frac{x}{d}}\tau _{2/k}(m) \\
&= \sum_{d\leq x}g_{k}(d) \Bigl(\frac{x(\log \frac{x}{d}%
) ^{2/k-1}}{d\Gamma (2/k)}+O_{k}\bigl(\frac{x}{d}(
\log \frac{x}{d}) ^{2/k-2}\bigr) \Bigr) \\
&= \sum_{d\leq x}g_{k}(d) \Bigl(\frac{x(\log x)
^{2/k-1}}{d\Gamma (2/k)}+O_{k}((\log x)
^{2/k-2}\log d) \\
&\qquad+O_{k}\bigl(\frac{x}{d}(\log \frac{x}{d})
^{2/k-2}\bigr) \Bigr) \\
&= \frac{x(\log x) ^{2/k-1}}{\Gamma (2/k)}\sum_{d\leq x}\frac{g_{k}(d)}{d}+O_{k}\Bigl(x(\log
x) ^{2/k-2}\sum_{d\leq x}\frac{\vert g_{k}(d)
\vert (1+\log d)}{d}\Bigr).
\end{align*}
The series $L(s,g_{k}) $ is absolutely convergent on the
half-plane $\sigma >\frac{1}{2}$, then for all $\varepsilon >0$
$$
\sum_{d\leq x}\vert g_{k}(d) \vert \ll
_{k,\varepsilon}x^{1/2+\varepsilon},
$$
hence by partial summation
$$
\sum_{d\leq x}\frac{\vert g_{k}(d) \vert (
1+\log d)}{d}\ll _{k,\varepsilon}x^{-1/2+\varepsilon}
$$
and therefore
$$
\sum_{n\leq x}D_{k,\omega}(n) = L(1,g_{k}) \frac{%
x(\log x)^{2/k-1}}{\Gamma (2/k)}+O_{k}(x(\log
x)^{2/k-2}) +O_{k,\omega}(x^{1/2+\varepsilon}).
$$
Which completes the demonstration.
\end{proof}
\section*{Acknowledgments}
The author would like to sincerely thank Professor Olivier Bordellès for his help and interest in this work and Professor Karl Dilcher for his generosity in reviewing this paper.
| 8775d123d75ebb6dc186ba51e79825de54b0e638 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Supermassive black holes (SMBHs) universally exist in galaxy centers.
Observationally there are fundamental relations between the mass of SMBHs and their host galaxy bulge luminosity ($M -$bulge) or stellar dispersion ($M-\sigma$)\citep{Dressler1988ApJ,Kormendy1993nag,Magorrian_1998,Ferrarese_2000,Gebhardt_2000,Kormendy:2013dxa}, which indicates the co-evolution between SMBHs and their host galaxies.
It has been argued that the growth of SMBHs is related to the hierarchical mergers with other SMBHs, gas accretion and feedback, or the combination of both mechanisms in a way that they evolve towards the observed fundamental relations \citep{Volonteri_2003,DiMatteo:2005ttp,Hopkins_2006, Volonteri2009_Msigma}. The exact picture remains an open question, for which observations will play an important role in testing theoretical models. In particular, the mergers of SMBH binaries should happen as a consequence of galaxy mergers.
In addition, there are substantial amount of evidences of galaxy mergers observed with electromagnetic observations \citep[see, e.g., ][for reviews]{Schweizer1996book, Lotz_2011,Conselice:2014joa}. The relevant key questions are how SMBHs form binaries and how the binaries evolve to merge, following the galaxy mergers.
The dynamical evolution of two SMBHs after their host galaxy mergers depends on the mass ratio $\mu$ between the stellar mass of the progenitor galaxies, and the nuclear environment they reside in the descendant galaxy \citep{Begelman1980Natur, Dotti2012, Colpi2014, Merritt2005review, Vasiliev2016}. For major mergers ($\mu \in \[1/4,1\]$ with $\mu$ defined as the smaller mass divided by the larger mass), the dynamical friction is efficient to migrate the SMBHs to the center of the common nucleus of the descendant galaxies \citep{Callegari:2008py}. This process brings SMBH pair at a wide separation to a distance of 1-10 parsec within a timescale of about $\sim 10^8$yr \citep{Begelman1980Natur, Yu2002, Mayer:2007vk}, and then a hard binary forms where the mutual gravitational force between SMBHs becomes larger than the environmental gravitational force from stars or gas disc.
For mergers with a small mass ratio, e.g., less than 1:10, the dynamical friction timescale could be larger than Hubble time and cause the secondary black hole to wander in the remnant galaxy \citep[see, e.g., ][]{Callegari_2011, Callegari:2008py}.
After the hard binary is formed, the further decay of the SMBH binary orbit may be driven by the ejection of stars in the ``loss cone" region which has small angular momentum. In the nucleus of remnant galaxies which host spherical stellar distribution, the SMBH binaries tend to eject stars in the loss cone region rapidly and then form a depleted zone \citep{Milosavljevic_2002, Dullo_2014}.
Since the re-population time scale of spherically distributed star clusters is usually longer than the Hubble time \citep{Yu2002}, the re-population efficiency is low so that further reduction of the binary separation is paused.
At this stage, the GW radiation is however not strong enough to drive the binary to merge within Hubble time. It appears to be a problem to efficiently drive the SMBH binary across the parsec scale to the regime where GW radiation dominates. This is usually referred to as the ``final parsec problem" \citep{Milosavljevi__2001,Yu2002, Milosavljevic2003b}.
On the other hand, there is circumstantial evidence that favors efficient SMBH binary coalescence. For example, the X-shaped radio sources (which are probably coalesced SMBH binaries with flipped jet directions \citep[][]{Dennett2002}) was observed to has a comparable rate to the expected rate of mergers of bright ellipticals, suggesting a quick coalescence following mergers \citep[][]{Merritt:2002hc}. In addition, if SMBH binaries fail to merge efficiently, binaries with stalled orbit evolution should be present in many bright ellipticals, and subsequent galaxy mergers will bring in additional SMBHs to form multiple SMBH systems. The multiple SMBH system is observed to be rare, although the theoretical predictions suggest multiple SMBH systems may be common \citep{Hoffman_Loeb2007,Haehnelt2002}.
The multi-body interaction may also slingshot eject SMBHs from the galactic centers, leading to much more scatter in the $M-\sigma$ and $M-$ bulge luminosity relations \citep[][]{Haehnelt2002}.
To reconcile with observations, several mechanisms have been proposed to drive the orbital separation of SMBH binaries across the parsec scale. For example, a non-spherical shape of the galaxy, especially the triaxial-shaped gravitational potential gives rise to additional torques that change the angular momenta of stars, thus the loss cone remains full and the interaction with stars will continue to sink SMBH binary till the sub-parsec regime \citep[][]{Yu2002, Merritt_2004, Holley-Bockelmann:2006gbs, Gualandris2016}.
Alternatively, in a gas-rich environment where SMBH binary interacts with the massive gaseous disk is another scenario that could efficiently drive SMBH binaries to GW-dominated regime \citep[see, e. g., ][]{Armitage2002, Dotti_2006_nucleardisc, Haiman_2009}. We will discuss these two scenarios in Sec. \ref{sec_delay_models}, in terms of their observational implication with GW detection. Other proposed models include three body interactions of SMBHs as a result of multiple mergers \citep{Hoffman_Loeb_2007} and refilling of the loss cone via star-star encounters \citep[][]{Yu2002, Milosavljevic2003}.
Different scenarios of the dynamical evolution of SMBH binary predict different delay times between galaxy mergers and the coalescence of SMBH binaries. The delay time is generally determined by the resident time the binary spends at their final-parsec stage, i.e., from several parsecs to sub-parsec separation before gravitational radiation takes over. To test the various dynamical models, observation of SMBH binaries at various evolution phases, especially at or below sub-parsec separation, is of key importance.
Current searches for SMBH binaries are mainly performed with electromagnetic (EM) observations. The observation at X-ray and radio band have already revealed a few SMBH pairs at wide separations, from several parsec to kilo-parsec, by directly resolving two individual emission cores \citep[e.g., ][]{Komossa2003_NGC6240,Rodriguez:2006th, fu2011kiloparsec}. At smaller angular separation bellowing the limits of the resolution of EM telescopes, SMBH binary candidates are indirectly found by diagnosing semi-periodic variations in the light curves \citep[e.g., ][]{Liu_2014,graham2015possible,Kharb_2017, Dey_2018, Britzen_2018,Bhatta_2019,Jiang_Yang_2022}, in the optical, X-ray, and radio waves.
The detectability of SMBH binaries or SMBH pairs at their evolution phases with multi-band EM waves is schematically shown in Fig.~\ref{SMBHB_evolution_phase}.
The direct resolution of double cores at X-ray and radio band confirm the existence of SMBH pairs/binaries at wide separations. The optical spectroscopy observation helps to prob individual nuclei by searching for double-peaked emission lines, while the optical light curves along are usually inconclusive to confirm the nature of emission sources, due to possible degeneracy with other origins \citep[see, e.g., ][]{fu2011kiloparsec, Wang_2010_Xray}.
It is expected that the next generation of Event Horizon Telescope with better angular resolution may find sub-parsec SMBH binaries with both direct and indirect detections \citep{D_Orazio_2018,Fang:2021xab}.
\begin{figure}
\centering
\includegraphics[height=5.9cm]{SMBHB_merge}
\caption{The observations of SMBH pairs and binaries at their different evolution phases after galaxy mergers. The continuously filled regions represent SMBH binaries/pairs that can be directly resolved by the EM telescopes (gray) \citep{Weisskopf_2000_Chandra,Perley_2011,Fu_2011,EventHorizonTelescope_2019I,Thatte_etal_2021,Breiding_2021} and can be detected through GWs \citep{2017eLISA, Yunes:2010sm, NANOGrav:2020bcs}. The regions filled with lines represent the indirect detection of SMBH binaries with EM observations.}
\label{SMBHB_evolution_phase}
\end{figure}
The gravitational waves (GWs) emitted by mergers of SMBH binaries are the most energetic astrophysical signals in the universe.
The space-based gravitational wave detectors, such as LISA (Laser Interferometer Space Antenna, \cite[][]{2017eLISA, eLISA:2013}) and Tianqin/Taiji \citep{luo2016tianqin,YueliangWu_taiji}, are designed to detect such events for binaries within the mass range $[10^5\Msun-10^8\Msun]$. The signals are so loud that LISA will be able to detect these events across almost all the relevant redshift space\citep[][]{2017eLISA}. It is predicted that LISA will be able to detect a few to several hundreds GW events from SMBH binary mergers \citep[see, e.g., ][]{Klein:2015hvg}.
In addition, LISA will also observe a few to a few thousands of Extreme Mass Ratio Inspirals (EMRIs) per year \citep{Babak:2017tow, Pan:2021ksp, Pan:2021oob}, which comprise a stellar-mass compact object orbiting around a SMBH. It is suggested that \citep{Yunes:2010sm} if the host black hole of an EMRI system is a component of a SMBH binary, the massive companion at the separation range of $\[0.001\text{pc},10\text{pc}\]$
might be measured through the EMRI waveform which is modulated by the binary-induced center-of-mass motion. According to \citet{Mazzolari:2022cho,Naoz:2022rru}, a relevant fraction of EMRI systems might form in the environment of a SMBH binary system due to an accelerated formation channel. At lower frequencies, Pulsar Timing Array (PTA, see, e.g., \cite{Hobbs_etal_2010}) is expected to detect GWs from loud, individual SMBH binaries or the GW background (GWB) contributed by unresolved SMBH binaries within the mass range $\[10^8\Msun,10^9\Msun\]$ which are inspiraling at a separation about $0.01 \text{pc}$. Fig. \ref{SMBHB_evolution_phase} also plots these GW detections over SMBH binaries.
Compared to EM observations, GW observations are often subjected to less observational systematics and the selection bias is usually better understood. These features are vital for constructing an accurate description of SMBH binary populations. As a result, although in principle all observations mentioned in Fig.~\ref{SMBHB_evolution_phase} should be incorporated into a coherent framework to infer the evolution path and population of SMBH binaries, in practice it is technically challenging for such a task. We shall only consider GW observations in this work.
The detection of GW singles will provide population information of SMBH binaries. So far, the non-detection of GWB from SMBH binaries in PTA band has already constrained both the SMBH binary population \citep{Grahammnras1726, NANOGrav:2019tvo} and the individual candidates \citep{Jenet_2004}, although a sign of common process was found in the PTA data \citep{NANOGrav_12d5, Middleton:2020asl}. In the future, LISA could be able to test the seeding and evolution models of SMBH evolution histories \citep[e.g., ][]{Klein:2015hvg, Chen:2022sae, Toubiana:2021iuw}.
We are interested in the delay time of SMBH binary coalescence after their host galaxy merge. By assuming the hierarchical merger model of SMBH formation and evolution following galaxy mergers, the merger rate of SMBH binary is determined by both the galaxy merger rate and the later time evolution histories. The galaxy merger rate could be observed from large-scale surveys \citep[e.g., ][]{Lotz_2011,Xu_2012,Mundy_2017, Casteels_etal_2014_GAMA, Driver_etal_2022GAMA, SDSS_merger_rate_2010, Davies_etal_DEVILS2018} or theoretically predicted from cosmological simulations \citep[e.g.,][]{Filip_2022_galaxy, OLeary_2021, Rodriguez-Gomez:2015aua}.
In addition, the mass of SMBHs is observed to be statistically related to the properties of their host galaxies, as mentioned before. Based on these observations and predictions, one could infer the formation rate of SMBH pairs/binaries synchronously after galaxy merge. Finally, the delay time distribution is able to be inferred by associating the formation rate distribution of SMBH pair/binary with the merger rate distribution of SMBH binary, through the hierarchical Bayesian approach.
LISA provides a promising opportunity for this task since it will provide the population information of SMBH binary merger events, such as the joint distribution of the redshift and mass etc., to a relatively high precision, thanks to its high single to noise ratio. On the other hand,
the delay time information is however unavailable from the stochastic gravitational wave background in PTA observation, due to the lack of detailed population information.
If PTA also detect a set of SMBH binary merger events, they can also be used to infer the delay time. It is also worth to note that PTA generally observes SMBH binaries with larger masses then those observed by LISA.
In this work, we construct a mock data set of LISA GW events and show how to infer the delay time of SMBH binary coalescence after galaxy mergers based on the mock data. The delay time distribution is inferred via the method of hierarchical Bayesian inference, by combining the data of LISA GW events and galaxy merger rates (observed from large-scale surveys or predicted by cosmological simulations), together with the intrinsic relationship between the binary mass and its host galaxy properties. We then compare different delay time models associated to various dynamical evolutionary scenarios, via the Bayesian model selection method.
This paper is organized as follows. In Sec. \ref{sec_2}, we discuss the merger rate of SMBH binaries as a consequence of their host galaxy mergers. We construct the merger rate of SMBH binaries in Sec. \ref{sec_2.4}, based on the galaxy merger rate (Sec. \ref{sec_2.1}), the relation between SMBH binary mass and the host galaxy property (e.g., the stellar mass, in Sec. \ref{sec_2.3}), and the delay time distribution. Different delay time models associated to different dynamical scenarios are discussed in Sec. \ref{delay_stellar}--\ref{gas_rich_scenario}, and a phenomenological description of delay time distribution is given in Sec. \ref{delay_time_models}.
In Sec. \ref{estimate_delay_distri}, we infer the delay time distribution from a mock LISA GW event set using the hierarchical Bayesian inference approach. Comparison between different delay time models associated to the corresponding dynamical scenarios are discussed in Sec. \ref{compare_delay_model}. The main results are discussed in Sec. \ref{conclusion}. In this article, we adopt the $\Lambda$ cold dark matter cosmology, with $\Omega_{M}=0.3$, $\Omega_{\Lambda}=0.7$, and $H_0=70\ {\text{km}}\ {\text{s}^{-1}\ \text{Mpc}}^{-1}$.
\section{Supermassive black hole binary mergers as a consequence of galaxy-galaxy mergers}
\label{sec_2}
According to the hierarchical merger model, SMBH binary merger events may happen as a consequence of galaxy-galaxy mergers. The merger rate of SMBH binaries is determined by the galaxy merger rate and the later-time binary evolution. The intrinsic relationship between SMBH mass and host galaxy properties also plays an important role in mapping the galaxy distributions to SMBH binary distributions. In this section, we construct the merger rate of SMBH binaries from the galaxy merger rate and the relationship between SMBH mass and host galaxy mass, together with a delay time distribution.
\subsection{Galaxy merger rate}
\label{sec_2.1}
The galaxy-galaxy merger rate has been theoretically predicted based on semi-analytical models \citep[e.g.,][]{Filip_2022_galaxy}, semi-empirical models \citep[e.g.,][]{OLeary_2021}, and hydrodynamical simulations \citep[e.g.,][]{Rodriguez-Gomez:2015aua}. Observationally it can be obtained using the fraction of close galaxy pairs and the averaged merging timescale \citep[see, e.g., ][]{Lotz_2011, Xu_2012, Mundy_2017}.
The merger timescale is still subject to large uncertainties and different approaches currently lead to non-converging predictions
\citep[see, e.g., ][and the references therein]{Rodriguez-Gomez:2015aua,OLeary_2021,Filip_2022_galaxy}.
However, it is reasonable to expect that by the time the GW data from LISA is available, improved theoretical modeling together with large scale surveys, e.g., the Galaxy And Mass Assembly Survey \citep[][]{Casteels_etal_2014_GAMA, Driver_etal_2022GAMA}, the Sloan
Digital Sky Survey \citep[][]{SDSS_merger_rate_2010}, and the Deep Extragalactic VIsible Legacy Survey \citep[][]{Davies_etal_DEVILS2018}, should provide a more consistent galaxy merger rate.
In the following context, in order to illustrate how the delay time can be measured, we adopt the galaxy merger rate presented from the Illustris simulation by \cite{Rodriguez-Gomez:2015aua}, in which they have summarized a fitting expression of galaxy merger rate per galaxy, $\frac{{\rm d}N}{{\rm d}t} (z, M_*, \mu)$, as a function of redshift $z$, the stellar mass of the descendant galaxy $M_{*}$, and the stellar mass ratio $\mu$ of the progenitor galaxies (see Table \ref{table_galaxy_merger_rate} in the Appendix \ref{sec.app} for details). The galaxy merger rate density is given by the galaxy merger rate per galaxy times with the galaxy number density, i.e., the galaxy stellar mass function or Schechter function $\phi(z, M_{*})$, which can be written down as follows
\begin{eqnarray}
{{\rm d}N_{\rm mergers} \over {\rm d}V_c \ {\rm d} \text{log}_{10}M_* \ {\rm d}t } = \frac{{\rm d}N}{{\rm d}t} (z,M_*,\mu) \times \phi(z,M_*) \,,
\end{eqnarray}
where $\phi(z,M_*):={{\rm d}N_{\rm number} \over {\rm d}V_c \ {\rm d} \text{log}_{10}M_* }$, $N_{\rm mergers}$ is the number of galaxy mergers within the parameter bins, $V_c$ is the cosmological co-moving volume.
The galaxy stellar mass functions in different redshift bins are taken from the recent observations given in \citep[][]{McLeod2021, Grazian2015AA, Stefanon_2021}. The details are shown in Table \ref{tab:Schechter_parameters}.
In the future, it is straightforward to update this part with more accurate galaxy merger rate descriptions.
\subsection{Delay time models of supermassive black hole binary mergers}
\label{sec_delay_models}
The delay time of SMBH binary coalescence is closely related to the ``final parsec problem", as it is mainly determined by the dynamical evolution time the SMBH binary spends to evolve from several parsec to sub-parsec separation. There are however different predictions of merger times considered in different dynamical scenarios, as suggested in various theoretical and numerical works \citep[see, e. g.,][]{Yu2002, Milosavljevic2003, Mayer:2007vk, Haiman_2009}, or \citep[see][for reviews]{Merritt2005review, Colpi2014}.
The observations of delay time through gravitational wave measurements will provide key insights into testing these models. In this work, we mainly focus on two scenarios for the matter of comparison, i.e., the stellar interaction and gas interaction scenarios, which are classified according to the dynamical evolution environment of the post-merger galaxies that the SMBH binaries reside in \citep[see, e. g., ][]{Begelman1980Natur, Yu2002, Milosavljevic2003, Mayer:2007vk, Haiman_2009, Antonini_2015}.
\subsubsection{Delay time of SMBH binary in stellar environment}
\label{delay_stellar}
In a dry (negligible gas influence) major merger, the dynamical friction from the background stellar bulge of the post-merger galaxy is capable of bringing the two SMBHs to a bound orbit within a timescale $t_{\rm df}\lesssim 10 \text{Myrs} \ \text{to} \ 100\text{Myrs}$.
The remnant galaxy after a major merger is expected to be substantially nonspherical that may host triaxial potential. The centrophilic orbits in triaxial galaxies can fill the loss cone of the phase space efficiently and harden the binary orbit towards the gravitational radiation stage. The hardening time scale is about a few $\text{Gyrs}$ that dominate the entire delay timescale in this scenario \citep[see, e. g., ][]{Yu2002,Khan_2011}. The detailed coalescence timescale after a bound binary SMBH is formed was estimated in \cite{Vasiliev_2015} using Monte Carlo
simulations, the results are as follows
\begin{eqnarray} \label{delaytime_stellar}
T_{\text{coal}} &\simeq& 1.7 \times 10^8 \text{yr} \times \left( r_{\text{infl}}\over 30 \text{pc} \right) ^{10+4\nu \over 5+\nu} \left( M\over 10^8 \Msun \right)^{- {5+3\nu \over 5+\nu}} \\ \nonumber
&& \times {\xi}^{- {4\over 5+\nu}} \left( 4 q \over (1+q)^2 \right)^{3\nu -1\over 5+\nu} 20^{\nu} \\ \nonumber
&& \times (1-e^2) \[ k+(1-k)(1-e^2)^4 \] \,,
\end{eqnarray}
where $M$ and $q$ are the mass and mass ratio of SMBH binary, $k=0.4 + 0.1\ \text{log}_{10} (M/10^8 \Msun)$, and the parameters $\nu$ and $\xi$ parameterize the hardening rate with values estimated from Monte Carlo realizations (See Table 1 of \cite{Vasiliev_2015} for values of $\nu$ and $\xi$ in different types of triaxial galaxy). The influence radius of the binary SMBH $r_{\text{infl}}$ is approximated to be $r_{\text{infl}}=1.5 \ M/\sigma^2$ ( see \cite{Merritt_2006} for a more refined definition of $r_{\text{infl}}$). One may simplify Eq.~(\ref{delaytime_stellar}) by dropping the mass ratio dependent term $\left( 4q \over (1+q)^2 \right)^{3\nu -1\over 5+\nu}$, since it is almost unity for major mergers.
By replacing $\sigma$ by $M$ according to the Faber–Jackson relation, the coalescence time can be parametrized to be
\begin{eqnarray} \label{coal_time_stallar_para}
T_{\text{coal}} &\simeq& \beta \left( M\over 10^6 \Msun\right)^{\alpha} \,,
\end{eqnarray}
where $\beta$ and $\alpha$ depends on the undetermined values of $\nu$, $\xi$ and $e$.
As the coalescence time $T_{\rm coal}$ generally dominates the lifetime of SMBH binary, we approximate the delay time in stellar environment by the coalescence time described in Eq.~(\ref{delaytime_stellar}) and (\ref{coal_time_stallar_para}):
\begin{eqnarray}
t_{\text{delay}, \text{stellar}} \sim T_{\text{coal}} \,.
\end{eqnarray}
Note that the power index of $M$, i.e. $\alpha$, only depends on $\nu$, and for all the listed values of $\nu$ in Table \ref{table_galaxy_merger_rate} of \cite{Vasiliev_2015}, $\alpha$ varies but always stays negative. For example, if we set $\mu=0.4$, $\nu=1/3$, and $e=0$, the coalescence time equals to
\begin{eqnarray} \label{stellar_delay}
T_{\text{coal}} &\simeq& 1 \text{Gyr} \left( M\over 10^6 \Msun\right)^{-1/16} \,.
\end{eqnarray}
We can characterize the behaviors of delay time, with $\beta$ parametrizing the delay time scale and $\alpha$ parameterizing the dependence on the mass of SMBH binary. These values may vary depending on the triaxial potential of the galaxy.
\subsubsection{Delay time of SMBH binary in gas-rich environment}
\label{gas_rich_scenario}
In gas-rich environments, the dynamical evolution is even more difficult to fully understand. We use a simplified delay time expression given in \cite{Antonini_2015} which assumes the delay is simply controlled by the viscous timescale of the nuclear gas. The predicted delay time in this gas-rich nuclear environment is \citep{Antonini_2015, Granato_2004}
\begin{eqnarray} \label{delay_time_gas}
t_{\text{delay}, \text{gas}} \sim \mathcal{R}_c t_{\rm dyn} \,,
\end{eqnarray}
where $\mathcal{R}_c$ is given by the critical Reynolds number in the range $\sim 10^2- 10^3$, and $t_{\rm dyn}$ is the dynamical time at the influence radius given by $t_{\rm dyn}= M/ \sigma^3$. Here we assume a benchmark value $\mathcal{R}_c = 10^3$ similar to \cite{Antonini_2015}. The delay time described by Eq.~(\ref{delay_time_gas}) is in the range $10\text{Myrs}$ to $ 100\text{Myrs}$, which is consistent with the studies of the coalescence time scale of SMBH binary in gas-rich environments \citep{Escala_2005,Dotti_2006_nucleardisc, Mayer:2007vk, Haiman_2009, Colpi2014}.
By replacing $\sigma$ with $M$ using the Faber–Jackson relation, we rewrite Eq.~(\ref{delay_time_gas}) to
\begin{eqnarray} \label{gas_delay}
t_{\text{delay}, \text{gas}} \sim 40 \ \text{Myr} \times \left({M\over 10^6 \Msun} \right )^{1/4} \,.
\end{eqnarray}
It is also worth noting that other models of gas-rich scenarios predict different delay times. For example, \cite{Goicovic:2016dul} considered the dynamical evolution of SMBH binaries interacting with infalling gas clumps by exchanging angular momentum through gas capture and accretion. \cite{Goicovic:2016dul} predicted a dynamical timescale of $0.1\text{Gyr}$ to $1\text{Gyr}$ for SMBH binaries to evolve into the GW emission regime.
Similar to the stellar scenario, we also phenomenologically parametrize the delay time in gas-rich scenario as
\begin{eqnarray}
t_{\text{delay}, \text{gas}} \sim \beta \left({M \over 10^6} \right)^{\alpha} \,.
\end{eqnarray}
In this way, the delay time $t_{\text{delay}, \text{gas}}$ and $t_{\text{delay}, \text{stellar}}$ predicted by different dynamical scenarios are both parameterized with $\alpha$ and $\beta$.
\subsubsection{Phenomenological delay time distribution models}
\label{delay_time_models}
The delay time in the stellar dynamical scenario and gas-rich scenario described in Eq.~(\ref{stellar_delay}) and Eq.~(\ref{gas_delay}) share the same form while differed by the underlying values of $\alpha$ and $\beta$.
In reality, the astrophysical environments of SMBH binaries are complicated so that the stellar scattering, the gas friction and/or other mechanisms may play different roles in different galaxies. To describe the statistical distribution of delay times in various galaxies, we use two phenomenological models for the delay time distribution: the Gaussian delay and the Power-law delay.
{\it Gaussian delay distribution.} In the ``Gaussian" delay distribution model, we parameterize the delay time distribution in the following way:
\begin{eqnarray} \label{gaussian_delay}
P_{\rm delay} (\tau, M | {\bf \Lambda}) = N( u, \sigma ) {2 \over 1+\text{erf}({ u \over \sqrt{2} \sigma }) }\,,
\end{eqnarray}
where $N( u, \sigma)$ denotes a Gaussian normal distribution with a mean value $ u=\beta ({M\over 10^6 \Msun})^{\alpha}$ and a standard deviation $\sigma$, and ${\bf \Lambda}$ here denotes $(\alpha,\beta,\sigma)$ which parameterize the delay distribution model in this case. The delay time $\tau$ takes the range $\[0, \infty\]$, so that a normalization constant ${2 \over 1+\text{erf}({u \over \sqrt{2} \sigma }) }$ in Eq.~(\ref{gaussian_delay}) is required to keep the total probability unity.
{\it Power-law delay distribution.} In the Power-law delay model, we parameterize the delay time distribution to be
\begin{eqnarray} \label{powerlaw_delay}
P_{\rm delay} (\tau, M | {\bf \Lambda}) \propto {1\over (\tau/\text{Gyr})^{\gamma \ \text{log}_{10} M + \kappa} }\,,
\end{eqnarray}
with $\tau$ takes value between $\[\tau_{\text{min}}, \tau_{\text{max}}\]$, where $\tau_{\text{min}}$ and $\tau_{\text{max}}$ are the lower and upper limit of delay time respectively, and here ${\bf \Lambda}=(\gamma, \kappa)$.
We set $\tau_{\text{min}}=0.001\text{Gyr}$ which is consistent with the dynamical friction timescale. The maximal delay time $\tau_{\text{max}}$ could in principle be larger than Hubble time, corresponding to the case that SMBH binary never merger.
In that case, we can set $\tau_{\text{max}}$ to an arbitrarily large number since the contribution from the tail in the power law distribution is small and negligible. Therefore we assume $\tau_{\text{max}}=30\text{Gyr}$.
\subsection{Relationship between Supermassive black hole mass and host galaxy property}
\label{sec_2.3}
The measurement of delay time distribution requires the comparison between the formation rate distribution of SMBH pairs/binaries (synchronously after galaxy mergers) and the merger rate distribution of SMBH binaries (through observation with GWs). The SMBH pair/binary formation rate relies on the galaxy merger rate and the fundamental relationship between the SMBH mass and the host galaxy properties, such as the $M$--bulge mass or the $M$--stellar velocity dispersion relations. In this work, we shall assume the relationship between the binary mass $M$ and the stellar mass $M_{*}$ of the descendant galaxy, i.e., the $M$--$M_*$ relationship. Observationally, the $M$--$M_*$ relationship is under certain divergence at different mass bins and subjected to uncertainties at high redshift bins \citep[e.g., ][]{Kormendy:2013dxa, Reines_2015, Ding_etal_2020}. For a simple illustration purpose, we shall take the $M$--$M_*$ relation from the observations of bulge dominated galaxies (where $M_{*} \sim M_{\text{bulge}}$), which gives \citep{Kormendy:2013dxa}:
\begin{eqnarray} \label{M_Mstar_relation}
\text{log}_{10}M= a + b\ \text{log}_{10}\left ({M_{*} \over 10^{11}\Msun} \right ) \,,
\end{eqnarray}
where $a=8.69 \pm 0.05$, $b=1.17 \pm 0.08$, and the intrinsic scatter is $\epsilon=0.28 \ \text{dex}$.
\subsection{Supermassive black hole binary merger rate}
\label{sec_2.4}
With the galaxy merger rate, the delay time model, and the relationship between the SMBH binary mass and stellar mass of descendant galaxy presented in Sec. \ref{sec_2.1}-- \ref{sec_2.3}, we now construct the merger rate distribution of SMBH binaries as follows
\begin{eqnarray} \label{BH_merger_rate1}
\mathcal{R}(M, t_{\text {L}} | {\bf \Lambda})= \int R_g [t_{\text {L}} + \tau, X(M)] {dX\over dM} P_{\text{delay}} (\tau, M | {\bf \Lambda}) d\tau,
\end{eqnarray}
where $t_{\text{L}}$ is the lookback time of SMBH binary merger, $R_g$ is the galaxy merger rate, and $X$ either denotes the stellar mass $M_{*}$, the bulge mass $M_{\text{bulge}}$, or the bulge luminosity $L_{\text{bulge}}$ of the host galaxy, which is a function of $M$ according the observed fundamental relationships. If the uncertainty of such relationship (mainly from the intrinsic scattering) is taken into account, Eq.~(\ref{BH_merger_rate1}) can be rewritten as
\begin{align} \label{BH_merger_rate2}
\mathcal{R}(M, t_{\text {L}} | {\bf \Lambda})\!\! =\!\!\! \int \!\!\!\, R_g (t_{\text {L}} \!\!+\!\! \tau, X) P_{\text{delay}} (\tau, M | {\bf \Lambda}) P (M| X) d X d\tau,
\end{align}
where the conditional probability $P(M|X)$ can be obtained through the joint distribution between $M$ and $X$ through $P(M,X)/P(X)$.
We take the galaxy merger rate $R_g (t_{\text{L}}, M_{*})$ in Eq.~(\ref{BH_merger_rate1}) (and (\ref{BH_merger_rate2})) as a marginalized distribution over mass ratio within the range $\[1/4,1\]$:
\\
\begin{eqnarray} \label{galaxy_merger_rate_t_Mg}
R_g (t_{\text{L}}, M_{*})&:=&{{\rm d}N_{\rm mergers} \over {\rm d}z \ {\rm d} \text{log}_{10}M_* \ {\rm d}t} \times { {\rm d}z\over {\rm d} t_{\text{L}}} \Big|_{z=z(t_{\text{L}})} \, \\ \nonumber
&=& \frac{{\rm d}N}{{\rm d}t} (z, M_*) \times \phi(z,M_{*}) \times {{\rm d}V_c \over {\rm d} t_{\text{L}} } \Big|_{z=z(t_{\text{L}})} \,,
\end{eqnarray}
where
\begin{eqnarray}
\frac{{\rm d}N}{{\rm d}t} (z, M_*)=\int^{1}_{0.25} \frac{{\rm d}N}{{\rm d}\mu \, {\rm d}t} (z, M_*, \mu)\ {\rm d} \mu \,,
\end{eqnarray}
as here we are considering major mergers. One can infer a similar range of mass ratio for SMBH binaries from the $M-M_{*}$ relation. We neglected the explicit dependence on the mass ratio of SMBH binary in the distribution based on several reasons. Firstly, for the delay time distribution determined by the interaction of SMBH binary with stars, such as the case described by Eq.~(\ref{delaytime_stellar}), the delay time only weakly depends on the mass ratio of the binary. Secondly, for the delay time predicted in gas-rich environment via the nuclear-disc-driven migration \citep{Dotti_2006_nucleardisc} or/and the binary-disc-driven migration \citep{Haiman_2009}, as has been discussed in Sec. \ref{gas_rich_scenario} with a simplified description, it also shows weak dependence on mass ratio in the case of major merger. Thirdly, physically major mergers are more successful to form compact/hard SMBH binaries \citep[e.g.,][]{Callegari_2011, Callegari:2008py, Colpi2014}. Finally, including the data of mass ratio is more computationally expensive. Since we use a interpolation function to replace the merger rate distribution (given in the form of convolution) in Eq.~(\ref{SMBHB_merger_rate_integ})
to estimate the delay time distribution via the method of hierarchical Bayesian inference, it requires computationally more expensive process to obtain the high dimensional interpolation function if we include the mass ratio as an extra dimension.
In this way, we constructed the SMBH binary merger rate $\mathcal{R}( M, z | {\bf \Lambda})$ by linking to the galaxy merger rate through the $M$-$M_{*}$ relationship and a delay time distribution, as
\begin{eqnarray} \label{SMBHB_merger_rate_integ}
\mathcal{R}( M, z | {\bf \Lambda}) &=& \mathcal{R}( M, t_{\text{L}} | {\bf \Lambda}) { {\rm d} t_{\text{L}} \over {\rm d}z }|_{t_{\text{L}}=t_{\text{L}}(z)} \\ \nonumber
%
&=& \int_{\tau_{\text{min}}}^{\tau_{\text{max}}} \! R_g (t_{\text{L}} \! + \! \tau, M_*) P_{\text{delay}} (\tau, M | {\bf \Lambda}) d\tau \\ \nonumber
&&\times \ { {\rm d} \text{log}_{10}M_* \over {\rm d} \text{log}_{10}M } \ {{\rm d} t_{\text{L}} \over {\rm d}z }\Big|_{t_{\text{L}}=t_{\text{L}}(z)} \\ \nonumber
%
%
\end{eqnarray}
where we have taken $X$ as $M_*$.
Before we move onto detailed calculations, we notice that the stellar mass functions given in Table \ref{tab:Schechter_parameters} are discontinuous piecewise functions fitted from observations at different redshift bins, which produce discontinuities in SMBH binary merger rates constructed with Eq.~(\ref{SMBHB_merger_rate_integ}), and also lead to auxiliary oscillations when the integration in Eq.~(\ref{SMBHB_merger_rate_integ}) is performed. To improve on these artificial issues, we smooth out the galaxy stellar mass function given in Table \ref{tab:Schechter_parameters} by averaging the stellar mass function in the transition regions between two neighbor bins.
\begin{figure}
\centering
\includegraphics[height=6.5cm]{Rbh_z.pdf}
\caption{The marginalized SMBH binary merger rate over mass as a function of redshift given by different delay time models. The SMBH binary merger rate is derived from Eq.~(\ref{SMBHB_merger_rate_integ}), considering the best fit value of the galaxy merger rate (Table~\ref{table_galaxy_merger_rate} and Table~\ref{tab:Schechter_parameters}), and the best fit of the relationship between the binary mass and host galaxy stellar mass (Eq.~(\ref{M_Mstar_relation})). The delay time models are phenomenologically parameterized with a Gaussian function or a power-law function given in Eq.~(\ref{gaussian_delay}) and Eq.~(\ref{powerlaw_delay}). The legend for different line styles denotes the delay parameters for different delay models.}
\label{RBH_z_models}
\end{figure}
The SMBH merger rate for several sample delay models is shown in Fig. \ref{RBH_z_models}.
In particular, the SMBH merger rate is marginalized over mass to be a function of redshift, in which we take the best fit of galaxy merger rate given in Table \ref{table_galaxy_merger_rate} and \ref{tab:Schechter_parameters}, the $M$--$M_{*}$ relation given in Eq.~(\ref{M_Mstar_relation}), and assume four different delay models. The first three delay models are ``Gaussian" delay distribution (described in Eq.~(\ref{gaussian_delay})) parameterized by three groups of parameters $(\alpha,\beta,\sigma)$.
In the case of the solid red line, we take $\alpha=-1/16$ and $\beta=\sigma=0.8 \text{Gyr}$ to represent the stellar environmental delay scenario discussed in Sec. \ref{delay_stellar}; In the dotted red line case, we assume $\alpha=1/4$ and $\beta=\sigma=0.04 \text{Gyr}$ to represent the gas-rich delay scenario discussed in Eq.~(\ref{gas_delay}); In the dashed red line, we take $\alpha=1/4$ and $\beta=\sigma=0.8\text{Gyr}$ to represent either stellar scenario with different mass dependence or gas-rich scenario with longer delay timescale as predicted in \cite{Goicovic:2016dul}.
As a comparison, we also consider a delay model with power-law delay distribution (described in Eq.~(\ref{powerlaw_delay})), as shown in the green line.
From Fig. \ref{RBH_z_models} we see that the ``Gaussian" delay model with delay timescale of $\mathcal{O}(\text{Gyrs})$ (solid and dashed red lines) induces an obvious change to SMBH merger rate compared to the case without delay (black line), by shifting the position of redshift where the peak locates to a smaller value, and simultaneously raising the amplitude of the peak. The case with delay time of order $\mathcal{O}(10 \text{Myrs})$ (dotted red line) lead to an almost indistinguishable merger rate distribution as the one without delay. The power-law delay model, such as the one shown in the green line, tends to broaden the distribution of merger rate towards smaller redshift while keeping the peak position less affected. The number of merger events predicted by these models is about 18 ${{\text{year}}^{-1}}$ -- 20 ${{\text{year}}^{-1}}$, which is consistent with the previous results \citep[e.g.,][]{Klein:2015hvg, Bhagwat:2021kwv}.
The variation of predicted binary merger rates in different models is small because the delay time scales from these delay models are approximately between $\mathcal{O}(1)$Gyr to $10$Myrs, so that most of the systems merge within Hubble time. The merger number per year is therefore close to the result without delay, as predicted under the same underlying galaxy merger rate and $M$--$M_*$ relation (which gives 19.5 $\text{year}^{-1}$).
\section{Estimating delay time distribution using Hierarchical Bayesian inference}
\label{estimate_delay_distri}
The discussion in Sec.~\ref{sec_2} shows how the SMBH binary merger distribution is influenced by the underlying delay time distribution. In particular, under the same assumption of the galaxy merger rate and the $M$--$M_*$ relationship, the SMBH binary merger distribution displays rather distinctive features as shown in Fig. \ref{RBH_z_models}. Thus, the delay time information should be encoded in the merger distribution of SMBH binaries, which can be inferred from the difference in the distribution between the SMBH binary merger rate and the formation rate of SMBH pairs/binaries. In this section, we discuss how to use the set of GW events to infer the underlying delay time distribution. For this purpose, we shall apply the framework of Hierarchical Bayesian Inference.
\subsection{Population analysis framework}
\label{sec_frame}
The hierarchical Bayesian approach is commonly used to infer the underlying distribution with a set of events statistically following such distribution.
The underlying distribution is usually parameterized with a mathematical expression with a few hyperparameters, which are constrained by the observed data using Bayesian Inference. This method has been previously applied to infer the population properties of binary black holes with LIGO events from GWTC \citep[][]{Abbott_2021_GWTC2,LIGO_2021_GWTC3}.
In this section we infer the distribution of delay time of SMBH mergers with mock data sets of LISA GW events via hierarchical Bayesian approach, following the similar way that the binary black hole population property is inferred with LIGO events. For the SMBH binaries considered here, the population distribution of LISA events is determined by the SMBH pair/binary formation rate combined with a delay time model, where the SMBH pair/binary formation rate is inferred from the galaxy merger rate (from observations or theoretical predictions) together with the relationship between the SMBH binary mass and host galaxy properties (from observation). The total number of events is modeled as an inhomogeneous Poisson process, and the joint likelihood function is given by
\begin{eqnarray} \label{def_hyper_likelihood}
\mathcal{L}(\{{\bf d}\} | {\bm \Lambda} ) &\propto& {N({\bf \Lambda})}^{N_{\text{det}}} e^{- N_{\text{exp}} ({\bf \Lambda}) } \times \\ \nonumber
&& \prod^{N_{\text{det}}}_{i=1} \int \mathcal{L}(d_i |{\bm \theta}) P ({\bm \theta}| {\bm \Lambda}) d {\bm \theta}
\end{eqnarray}
where $N_{\text{exp}}$ is the expected number of detections within the observation duration for the population model ${\bm \Lambda}$ ( $\bm \Lambda$ represents the hyperparameters of the given model), ${\bm \theta}=(m_1, m_2, z, ...)$ represents source parameters, $N({\bf \Lambda})$ is the predicted number of merger events under the model ${\bf \Lambda}$, $N_{\text{exp}}=\xi({\bf \Lambda}) N $ represents the expected detection number assuming a detection fraction $\xi({\bf \Lambda})$, $\mathcal{L}(d_i |{\bm \theta})$ is the likelihood function for each individual event, and $P ({\bm \theta}| {\bm \Lambda})$ denotes the population distribution corresponding to model $\bm \Lambda$.
The integral in Eq.~(\ref{def_hyper_likelihood}) can be evaluated by generating Monte-Carlo (MC) posterior samples ${\bm \theta}_i$ following the likelihood function $\mathcal{L}(d_i |{\bm \theta})$ in the parameter estimation process for each event. The expression can be rewritten as
\begin{eqnarray} \label{redef_likelihood}
\mathcal{L}(\{{\bf d}\}|{\bm \Lambda}) \propto
{N({\bf \Lambda})}^{N_{\text{det}}} e^{- N_{\text{exp}} ({\bf \Lambda}) }
\prod^{N_{\text{det}}}_{i=1} \langle {P ({\bm \theta}| {\bm \Lambda}) \over P_{\varnothing} ({\bm \theta})} \rangle
\end{eqnarray}
where $ \langle ... \rangle$ is the average over the Monte-Carlo samples ${\bm \theta}_i$ of the individual events, and here $P_{\varnothing} ({\bm \theta})$ is the default prior taken in the parameter estimation. The detection fraction $\xi({\bm \Lambda})$ here is taken to be one as we simply assume all the SMBH binary merger events within the $10^5-10^8 M_\odot$ mass range (the range of mass considered in the mock data set) are detectable by LISA. This assumption is based on the fact that LISA can detect SMBH binary mergers (with comparable mass) over a wide mass bin and across a large redshift \citep{2017eLISA}.
The posterior of the hyperparameters, $P ({\bm \Lambda} | \{ {\bf d} \})$, given the the population model ${\bm \Lambda}$ and data $\{{\bf d}\}$ is
\begin{eqnarray}
P ({\bm \Lambda} | \{ {\bf d} \}) \propto \mathcal{L}(\{{\bf d}\}|{\bm \Lambda}) P ({\bm \Lambda})\,.
\end{eqnarray}
We consider the population properties of binary mass $M$ and redshift $z$, i.e., ${\bm \theta}=(M,z)$.
The population distribution $P({\bm \theta}| {\bm \Lambda})$ is the normalized merger distribution of $ \mathcal{R}( M, z| {\bf \Lambda})$ (Eq.~(\ref{SMBHB_merger_rate_integ})), defined as
\begin{eqnarray} \label{population_prob}
P({\bm \theta}| {\bm \Lambda}) = {1\over \mathcal{N}({\bf \Lambda})}\mathcal{R}( M, z | {\bf \Lambda}) \,,
\end{eqnarray}
where
\begin{eqnarray}
\mathcal{N}({\bf \Lambda})=\int {\rm d} \text{log}_{10}M \ \text{d}z \ \mathcal{R}( M, z | {\bf \Lambda})
\end{eqnarray}
is the merger events per year of SMBH binaries. The predicted merger number $N({\bf \Lambda})$ is $\mathcal{N}({\bf \Lambda})$ times the duration of observation $T_{\text{det}}$, $N({\bf \Lambda})= \mathcal{N}({\bf \Lambda}) \ T_{\text{det}}$.
The hierarchical Bayesian inference analysis requires a definite distribution model of SMBH binary merger rate, which in turn requires a definite result of galaxy merger rate and the $M$--$M_*$ relation. In reality, the latter two components are still subject to large uncertainties, either from observations or from theoretical predictions.
To account for the influence of the uncertainties, we further extend the Likelihood in Eq.~(\ref{def_hyper_likelihood}) to a statistically averaged expression in which the uncertain fitting parameters satisfy their corresponding distributions. We denote ${\bm \lambda_1}$ as the fitting parameters in the galaxy merger rate, such as the parameters listed in Table \ref{table_galaxy_merger_rate} and \ref{tab:Schechter_parameters} for our assumption of galaxy merger model. Similarly, we use ${\bm \lambda_2}$ to denote the fitting parameters in the $M$--$M_*$ relationship, such as the one taken in Eq.~(\ref{M_Mstar_relation}). Then the averaged Likelihood takes the form
\begin{eqnarray} \label{redef_likelihood_errors}
\!\!\! \! \bar{\mathcal{L}}(\{{\bf d}\}|{\bm \Lambda}) &=& \! \int \! \mathcal{L}(\{{\bf d}\}|{\bm \Lambda}, {\bm \lambda_1}, {\bm \lambda_2}) P({\bm \lambda_1}) P( {\bm \lambda_2}) d {\bm \lambda_1} d {\bm \lambda_2} \nonumber \\
&=& \sum_{{\bm \lambda_1},{\bm \lambda_2}} \mathcal{L}(\{{\bf d}\}|{\bm \Lambda}, {\bm \lambda_1}, {\bm \lambda_2}) \,,
\end{eqnarray}
where $\mathcal{L}(\{{\bf d}\}|{\bm \Lambda}, {\bm \lambda_1}, {\bm \lambda_2})$ is the likelihood inherited from Eq.~(\ref{redef_likelihood}) associated with the galaxy merger rate $R_{\text{g}}(t_{\text{L}},M_{*} , {\bm \lambda_1})$ and the relationship $M_{*}(M ,{\bm \lambda_2})$, throuth Eq.~(\ref{population_prob}) and (\ref{SMBHB_merger_rate_integ}). The summation in the second equal is the average over parameters ${\bm \lambda_1}$ and ${\bm \lambda_2}$ which satisfy the corresponding distributions $P({\bm \lambda_1})$ and $P({\bm \lambda_2})$.
\begin{figure}
\centering
\includegraphics[height=6cm]{samples_Gaussian.pdf}
\caption{The samples of mock GW data assuming four years observation period of LISA, obtained from our ``Gaussian" delay model with $\alpha=-1/16$ and $\beta=\sigma=0.8\text{Gyr}$. The black points represent the total 78 GW events denoted by source parameters $M$ and $z$. The error bars of $M$ and $z$ in each GW event are estimated from the Fisher information matrix according to the LISA sensitivity curve. }
\label{data_samples_Gaussian_delay}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=6cm]{samples_PL.pdf}
\caption{Similar to \ref{data_samples_Gaussian_delay}, but for a power-law delay model with $\gamma=0.1$ and $\kappa=0.5$, and 73 GW events in total. }
\label{data_samples_powerlaw_delay}
\end{figure}
\subsection{Estimating the delay time distribution}
To illustrate the procedure discussed in Sec. \ref{sec_frame}, we generate a mock data set of LISA GW events for SMBH binary mass in the range $\[10^5 \Msun, 10^8 \Msun\]$, following the rate function $\mathcal{R}( M, z | {\bf \Lambda})$, where the best fit values are taken for ${\bm \lambda_1}$ and $ {\bm \lambda_2}$. In particular, we consider a ``Gaussian" delay model with parameters $\alpha=-1/16$ and $\beta=\sigma=0.8\text{Gyr}$, and a power-law delay model with parameters $\gamma=0.1$ and $\kappa=0.5$. The resulting SMBH merger rates predicted from these two delay models are shown in Fig. \ref{RBH_z_models}.
Assuming an observation duration of four years, the total number of events in each mock data set is 78 and 73 respectively. The binary mass M and redshift z of these mock GW events are shown in Fig. \ref{data_samples_Gaussian_delay} and \ref{data_samples_powerlaw_delay}, as sampled from $\mathcal{R}( M, z | {\bf \Lambda})$. The error bar or the posterior distribution for each parameter in an individual event is estimated according to the sensitive curve of LISA \citep{Barack:2003fp, Robson:2018ifk} and Fisher information matrix analysis using the phenomenological waveform \citep{Ajith_2007, Ajith_etal_2011}.
We now estimate the delay time distribution of SMBH binary mergers with these mock GW data (Fig.~\ref{data_samples_Gaussian_delay} and \ref{data_samples_powerlaw_delay}) using hierarchical Bayesian inference.
Note that the Likelihood function is generated from an integration function defined in Eq.~(\ref{SMBHB_merger_rate_integ}), and we are dealing with $N_{\text{det}}\times N_{\text{samples}}$ samples of GW data (where $N_{\text{samples}}$ is the number of samples of $M$ and $z$ in each GW event).
Thus for each MC sample of the hyperparameters ${\bf \Lambda}$, one needs to perform the integration for $N_{\text{det}}\times N_{\text{samples}}$ times for the Likelihood defined in Eq.~(\ref{redef_likelihood}).
If we need to generate $N_{\text{h-p}}$ samples for hyperparameters, it becomes $N_{\text{det}}\times N_{\text{samples}} \times N_{\text{h-p}}$, which is usually time-consuming. In addition, for the averaged Likelihood considered in Eq.~(\ref{redef_likelihood_errors}), it is even more computationally expensive, as it requires $N_{\text{det}}\times N_{\text{samples}} \times N_{\bf \lambda}$ times of integration for each MC sample of hyperparameters (where $N_{\bf \lambda}$ is number of samples of ${\bm \lambda_1}$ and ${\bm \lambda_2}$).
Therefore, to reduce the computational cost, first, we fix the galaxy merger rate and the $M$--$M_*$ relationship in the Likelihood function by taking the best fit value of their fitting parameters ${\bm \lambda_1}$ (Table \ref{table_galaxy_merger_rate} and \ref{tab:Schechter_parameters}) and ${\bm \lambda_2}$ (Eq.~(\ref{M_Mstar_relation})). Second, we replace the merger rate function $\mathcal{R}( M, z | {\bf \Lambda})$ in Eq.~(\ref{SMBHB_merger_rate_integ}) with an interpolation function to avoid the time-consuming integration.
To be more specific, we use the following Likelihood function to infer the delay parameters from our mock GW data:
\begin{eqnarray} \label{likelihood_parameter_estimate}
\mathcal{L}(\{{\bf d}\}|{\bm \Lambda}, {\bm {\lambda_1}}^{\text{bf}}, {\bm {\lambda_2}}^{\text{bf}}) \propto
\prod^{N_{\text{det}}}_{i=1} \langle {P ({\bm \theta}| {\bm \Lambda}, {\bm {\lambda_1}}^{\text{bf}}, {\bm {\lambda_2}}^{\text{bf}}) \over P_{\varnothing} ({\bm \theta})} \rangle \,.
\end{eqnarray}
The Likelihood in Eq.~(\ref{likelihood_parameter_estimate}) makes use of the distribution information of the merger rate without considering the information of the ``observed" number of GW events $N$ (or the amplitude of the merger rate). We have not found a significant difference between including and not including the N data in the Likelihood above. Rather, we find the uncertainties from some of the parameters in ${\bm \lambda_1}$ and $ {\bm \lambda_2}$ may result in order unity changes of $N({\bf \Lambda})$. Therefore $N({\bf \Lambda})$ is more vulnerable to the model uncertainty of galaxy mergers and galaxy-SMBH relations, so that we remove $N({\bf \Lambda})$-related terms in the Likelihood which may bring large errors for estimating delay parameters.
The uncertainties of the fitting parameters and the information of $N$ have been considered in the next section when we compare different delay models.
\begin{figure}
\centering
\includegraphics[height=8cm]{delay_posterior_gaussian}
\caption{The estimated posteriors of delay parameters in a ``Gaussian" delay distribution model from the mock GW data showed in Fig. \ref{data_samples_Gaussian_delay}.
The truth values of the hyperparameters
are $\alpha=-1/16$, $\beta=0.8 \text{Gyr}$ and $\sigma=0.8 \text{Gyr}$, denoted by crossed blue lines.
The black dashed lines boundary the $68\%$ confidence region.
}
\label{delay_posterior_gaussian}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=6.5cm]{confidence_delay_gaussian}
\caption{The confidence plot of the ``Gaussian" delay distribution for the estimated delay parameters is shown in Fig.~\ref{delay_posterior_gaussian}, taking the mass of the SMBH binary to be $M=2 \times 10^6 \Msun$. The filled blue area is the one sigma confidence region, and the orange line is the given truth delay distribution (with $\alpha=-1/16$, $\beta=0.8 \text{Gyr}$ and $\sigma= 0.8 \text{Gyr} $). }
\label{confidence_delay_gaussian}
\end{figure}
As illustrative examples, the posteriors of delay parameters for the ``Gaussian" delay and power-law delay models are shown in Fig. \ref{delay_posterior_gaussian} and \ref{delay_posterior_powerlaw} respectively. Based on the posteriors of the hyperparameters, the confidence plot of the delay functions can be displayed for various SMBH masses. For example, the recovered delay time distribution for SMBH binary with mass $M=2 \times 10^{6}\Msun$ are shown in Fig. \ref{confidence_delay_gaussian} and \ref{confidence_delay_powerlaw}. It is clear from these figures that the delay parameters and the delay time distributions are properly recovered within the uncertainties.
For the ``Gaussian" delay model, there is one more hyperparameter than the power-law delay model, and the resulting uncertainties for individual parameters are relatively larger. There is also a degeneracy between $\beta$ and $\sigma$ in the ``Gaussian" delay model as shown in Fig. \ref{delay_posterior_gaussian}, which is mainly due to the fact that the delay time scale is determined both by $\beta$ and $\sigma$ (Eq.~(\ref{gaussian_delay})). Fig. \ref{confidence_delay_gaussian} shows that the delay time distribution is better constrained at a larger delay time region than the smaller delay time region. This is due to the fact that the larger delay time induces a larger shift in the merger distribution over redshift, while the smaller delay time induces a smaller shift in redshift, thus leading to merger distributions being less distinguishable from each other.
In the Power-law delay model, the delay parameters $\gamma$ and $\kappa$ are properly recovered with the truth value of both parameters settled in the center of the posterior distribution, which is close to a Gaussian shape, as can be seen from Fig. \ref{delay_posterior_powerlaw}. As a result, the truth delay time distribution lies in the center of the confidence interval, as shown in Fig. \ref{confidence_delay_powerlaw}.
\begin{figure}
\centering
\includegraphics[height=6.5cm]{powerlaw_delay_posteriors}
\caption{The estimated posteriors of delay parameters in a power-law delay model from the mock GW data showed in Fig. \ref{data_samples_powerlaw_delay}. The truth values of the hyperparameters are $\gamma=0.1$ and $\kappa=0.5 \text{Gyr}$, denoted by crossed blue lines. The black dashed lines boundary the $68\%$ confidence region. }
\label{delay_posterior_powerlaw}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=6.5cm]{confidence_delay_powerlaw}
\caption{The confidence plot of the power-law delay time distribution for the estimated delay parameters shown in Fig. ~\ref{delay_posterior_powerlaw}, taken the mass of the SMBH binary to be $M=2 \times 10^6 \Msun$. The filled blue area is the one sigma confidence region, and the orange line is the given truth delay distribution (with $\gamma=0.1$ and $\kappa=0.5$). }
\label{confidence_delay_powerlaw}
\end{figure}
So far, we have shown that the delay time distribution of SMBH binary coalescence can in principle be recovered using hierarchical Bayesian inference from LISA GW data. This highlights the main point of this work, that the delay time distribution as an observable may be measured with GW events. We shall discuss the application of this measurement in the next section. Here we would like to emphasize one important caveat in the above analysis, that we have assumed perfect knowledge of the galaxy merger rate distribution and galaxy-SMBH mass relationship (by assuming best fit values for ${\bm \lambda_a}$ and ${\bm \lambda_2}$). The current understanding of these relations is still subject to large uncertainties from both observational and theoretical perspectives.
Therefore in a more thorough analysis, these uncertainties must be taken into account, which requires one to consider the averaged Likelihood function given in Eq.~(\ref{redef_likelihood_errors}). Because of the computational cost issue mentioned in the previous section, this part of procedure is not presented in this work. On the other hand, it is reasonable to expect future observations of galaxy mergers together with large-scale cosmological simulations will improve the accuracy of these relations, which is crucial to minimize the systematic error in the measurement of the delay time distribution.
\section{Comparing different delay time distribution models}
\label{compare_delay_model}
The recovered delay time distribution encodes important information about the evolution mechanism of SMBH binaries. As discussed in Sec. \ref{sec_delay_models}, different evolutionary scenarios predict different delay times, so that the measured delay time distribution may be used to test the predictions of different models. In reality each evolution model may be dominant in only part of events observed by gravitational waves. In that case one may alternatively choose to constrain the fraction of events in each channels.
In this section, we perform the comparison between models based on the ``observed" time delay distribution using the Bayesian Model Selection method. In particular, we compare the coalescence models of SMBH binaries which has been discussed in Sec.\ref{sec_delay_models}. From Sec.\ref{sec_delay_models}, we know that the delay time predicted from the dynamical scenarios in different astrophysical environment are characterised with different time scales or mass dependence. The realistic astrophysical environment the SMBH binary reside in after galaxy mergers is complex so that it is reasonable to expect that the delay time distribution is dispersed by astrophysical conditions. Using the ``Gaussian" delay distribution described in Eq.~(\ref{gaussian_delay}), we can phenomenologically assign different parameters to individual delay models as predicted from various coalescence scenarios. For example, the hyperparameter $\beta$ in the ``Gaussian" delay model characterizes the delay time scale, $\alpha$ parameterizes the dependence on the mass of SMBH binary by assuming a power-law relation, and $\sigma$ describes the dispersion of the delay time scale according to the statistical population of environment the SMBH binary resides in. We shall only discuss the ``Gaussian" model here as it seems to provide a more natural fit with these physical descriptions.
For illustration purposes, we have chosen three delay models predicted from three different dynamical scenarios. As listed in Table \ref{tab:gaussian_delay_models}, the first one is the stellar scenario with $\alpha=-1/16$ and $\beta=\sigma=0.8 \text{Gyr}$, characterizing the delay time resulted from the interaction between SMBH binary and stars, as discussed in Sec. \ref{delay_stellar}; the second model is the gas-rich scenario with $\alpha=1/4$ and $\beta=\sigma=0.04 \text{Gyr}$, characterizing the delay time due to the interaction of SMBH binary with environmental gas disk, as described in Eq.~(\ref{gas_delay}); the third model is given by $\alpha=1/4$ and $\beta=\sigma=0.8\text{Gyr}$ representing the delay time either coming from the stellar scenario with different mass dependence or from the gas-rich scenario with a longer delay timescale. The corresponding delay time distribution from these three models are plotted in Fig. \ref{delay_gaussian_examples}, assuming the binary mass to be $M=2\times 10^6 \Msun$. The blue line (model 1) suggests a delay time of order $10\text{Myrs}$, and the red (Model 1) and purple (Model 3) lines correspond to a delay time of $\mathcal{O}(\text{Gyr})$.
Fig. \ref{delay_gaussian_various_mass} illustrates the mass dependence of delay time distribution of Model 1 and Model 3, assuming the binary mass to be $M=2\times 10^5 \Msun$, $2\times 10^6 \Msun$, and $2\times 10^7 \Msun$ respectively. Although for $M=2\times 10^6 \Msun$, Model 1 and Model 3 share a similar delay time scale, for larger mass such as the $M=2\times 10^7 \Msun$ case they differ significantly, as shown in Fig. \ref{delay_gaussian_various_mass}. The delay time distribution in Model 3 is more sensitive to mass than Model 1, i.e., Model 3 predicts longer delay times for larger masses, while Model 1 has a weaker and opposite trend.
The resulting marginalized SMBH binary merger rate predicted from these three delay models has been shown previously in Fig. \ref{RBH_z_models} with the best fit values for parameters ${\bm \lambda_1}$ and ${\bm \lambda_2}$. The mock GW data we use in this analysis is assumed to be the same as the one given in Fig. \ref{data_samples_Gaussian_delay}, which is generated from the first delay-time model.
\begin{figure}
\centering
\includegraphics[height=7cm]{delay_distribution.pdf}
\caption{The delay time distribution for the three delay models considered in Table \ref{tab:gaussian_delay_models} parameterized as a ``Gaussian" distribution given in Eq.~(\ref{gaussian_delay}), taken the mass of the SMBH binary to be $2 \times10^6\Msun$.
}
\label{delay_gaussian_examples}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=7cm]{P_delay_compare_mass.pdf}
\caption{The delay time distribution referred to model 1 and model 3 in Table \ref{tab:gaussian_delay_models} for different masses of SMBH binary.
}
\label{delay_gaussian_various_mass}
\end{figure}
We compare the statistical significance of different delay models by calculating the Bayesian Odds ratio, which is given by
\begin{eqnarray} \label{Odds_ratio}
\mathcal{O}={ P({\bf \Lambda}_A|\{ {\bf d} \}) \over P({\bf \Lambda}_B|\{ {\bf d} \})} = { P(\{ {\bf d \} }| {\bf \Lambda}_A ) \over P(\{ {\bf d} \} | {\bf \Lambda}_B )} { P({\bf \Lambda}_A ) \over P( {\bf \Lambda}_B )}
= \mathcal{B} { P({\bf \Lambda}_A ) \over P( {\bf \Lambda}_B )} \,,
\end{eqnarray}
where ${\bf \Lambda}_i=(\alpha_i, \beta_i, \sigma_i)$ with $i=A, B$ is the corresponding value of hyperparameter for the delay model A or B, and $\mathcal{B}$ is the Bayes factor between the two models. Here, we assume $ P({\bf \Lambda}_A ) = P( {\bf \Lambda}_B )$, thus the Odds ratio $\mathcal
{O}$ equals $\mathcal{B}$. The Odds Ratio given in Eq.~(\ref{Odds_ratio}) is a statistical quantity that measures the relative degree being favored by data for two competing models. The standard range of Odds ratio in determining the strength of evidence is listed in Table \ref{tab:Bayes_factor}.
\begin{table}
{
\centering
\caption{The interpretation of the Odds ratio (or Bayes factor) in determining the strength of evidence the model is favored \citep[see][]{Kass_Raftery_1995}.}
\label{tab:Bayes_factor}
\begin{tabular}{ c c }
\hline
$\mathcal{O}$ & Strength of evidence \\
\hline
$<1$ & Negative \\[0.15cm]
1-3.2 & Not worth more than a bare mention \\[0.15cm]
3.2-10 & Weak \\[0.15cm]
10-100 & Strong \\[0.15cm]
$>100$ & Very strong/Decisive \\[0.15cm]
\hline
\end{tabular}
\par
}
\end{table}
\begin{table}
{
\centering
\caption{The delay parameters for different ``Gaussian" delay models, and the result of Odds ratio between models.}
\label{tab:gaussian_delay_models}
\begin{tabular}{ c | c c c | c }
\hline
Model & $\alpha$ & $\beta$ ($\text{Gyr}$) &
$\sigma$ ($\text{Gyr}$) & Odds ratio \\
\hline
1 & -1/16 & 0.8 & 0.8 & \\[0.15cm]
2 & 1/4 & 0.04 & 0.04 & $\sim 10^7$ \\[0.15cm]
3 & 1/4 & 0.8 & 0.8 & 327 or 524 \\[0.15cm]
\hline
\end{tabular}
\par
}
\begin{tablenotes}
\item {\bf Notes}:
Model 1 represents the delay time model predicted in a stellar dynamical scenario described in Sec. \ref{delay_stellar}, model 2 represents the case of a gas-rich dynamical scenario, as discussed in Sec. \ref{gas_rich_scenario}. Model 3 represents delay time model either from a stellar scenario with different mass dependence or the gas-rich scenario but with larger delay timescale. The last column lists the Odds ratio of Model 1 to Model 2 and Model 1 to Model 3 calculated from our averaged Likelihood function.
\end{tablenotes}
\end{table}
We use the averaged Likelihood defined in Eq.~(\ref{redef_likelihood_errors}) to calculate the Bayes factor and Odds ratio in Eq.~(\ref{Odds_ratio}), in which the modeling uncertainties from the fitting parameters ${\bm \lambda_1}$ and ${\bm \lambda_2}$ in galaxy merger rate and $M$--$M_*$ relationship are taken into account. Specifically, we generated $8000$ samples for each of the fitting parameters in ${\bm \lambda_1}$ and ${\bm \lambda_2}$ from their probability distribution functions (PDFs) $P({\bm \lambda_1})$ and $P({\bm \lambda_2})$. The best fit and error bars for these fitting parameters are given in Table \ref{table_galaxy_merger_rate}, \ref{tab:Schechter_parameters}, and Eq.~(\ref{M_Mstar_relation}). The PDF is then constructed with a Gaussian or a Gamma distribution, or a composition of two different Gaussian distribution such that it satisfies the corresponding best fit and error bar values for each fitting parameter. Each sample of $({\bm \lambda_1}, {\bm \lambda_2})$ generates one Likelihood $\mathcal{L}(\{{\bf d}\}|{\bm \Lambda}, {\bm \lambda_1}, {\bm \lambda_2})$, and the summation over these Likelihood functions leads to the averaged Likelihood $\bar{\mathcal{L}}(\{{\bf d}\}|{\bm \Lambda})$ defined in Eq.~(\ref{redef_likelihood_errors}). We have used two different definitions for the Likelihood for comparison, one is given by
\begin{eqnarray} \label{redef_likelihood1}
\mathcal{L}(\{{\bf d}\}|{\bm \Lambda}, {\bm \lambda}_1, {\bm \lambda}_2 ) &\propto&
{N({\bf \Lambda}, {\bm \lambda}_1, {\bm \lambda}_2)}^{N_{\text{det}}} e^{- N ({\bf \Lambda}, {\bm \lambda}_1, {\bm \lambda}_2) } \\ \nonumber
&&
\prod^{N_{\text{det}}}_{i=1} \langle {P ({\bm \theta}| {\bm \Lambda}, {\bm \lambda}_1, {\bm \lambda}_2) \over P_{\varnothing} ({\bm \theta})} \rangle
\end{eqnarray}
where the information of the number of GW events $N$ is included,
and another one is given by
\begin{eqnarray} \label{redef_likelihood2}
\mathcal{L}(\{{\bf d}\}|{\bm \Lambda}, {\bm \lambda}_1, {\bm \lambda}_2 ) \propto
\prod^{N_{\text{det}}}_{i=1} \langle {P ({\bm \theta}| {\bm \Lambda}, {\bm \lambda}_1, {\bm \lambda}_2) \over P_{\varnothing} ({\bm \theta})} \rangle
\end{eqnarray}
where the information of number of events (or the merger rate amplitude) is neglected.
The Odds ratio for the comparison between Model 1 and Model 2 is $\mathcal{O}(10^7)$ (with and without considering the factor of total number of events in the Likelihood function), indicating a decisive evidence favoring Model 1 over Model 2. This result is consistent with our expectation since the the mock GW data is sampled from Model 1, which predicts rather different scale of delay time as compared to Model 2 as shown in Fig. \ref{delay_gaussian_examples}. In fact, these two models also lead to rather distinct SMBH binary merger rates as functions of redshift shown in Fig. \ref{RBH_z_models}. On the other hand, the Odds ratio for Model 1 compared to Model 3 is approximately 327 (without including the total event number in the Likelihood function) and 524 (total event number included ) respectively, still indicating a decisive evidence supporting Model 1 over Model 2. In fact, the difference between Model 1 and 3 can be more clearly found in Fig. \ref{delay_gaussian_various_mass} where binaries with different masses are presented. As a result, Model 1 and 3 actually predict rather different joint distribution of $M$ and $z$ in the merger rate. Including the information about the total number of merger events in the Likelihood function further increases the evidence as expected. The delay time difference between Model 1 and Model 3 is no more than a few times (Fig. \ref{delay_gaussian_various_mass}), while it is a few orders of magnitude difference between Model 1 and Model 2 (Fig. \ref{delay_gaussian_examples}). That is the reason why the Odds ratio between them also differs by orders. Nevertheless, for the specific mock data set we have used, it is possible to statistically distinguish different models based on data. This provides a promising aspect for the application of delay time measurements in understanding SMBH binary evolution mechanisms.
\section{Conclusion and discussion}
\label{conclusion}
In this work we have discussed how to apply a set of SMBH bianry merger events detected by space-borne GW detectors to study the evolution mechanisms of SMBH binaries, through the measurement of delay times. In fact, the delay time between the galaxy merger and various stages of SMBH binary evolution may all be potential targets of opportunities, but it is generally difficult to address the selection effects in electromagnetic observations in order to obtain faithful information about the SMBH binary population and distribution. Thanks to the fact that SMBH binary mergers are the energetically loudest events in the universe, mergers within the right mass range are likely to be all identified by the GW detectors such as LISA. Therefore measuring the delay time till merger through GWs seems to be the most promising way to test different binary evolution models.
Such tests are illustrated in Sec.\ref{estimate_delay_distri} and Sec.\ref{compare_delay_model} with Mock GW data ($\sim 70$--$80$ events) generated assuming an observation period of four years. The mass and redshift measurement uncertainties are estimated with Fisher Information Matrix, using the sensitive curve of LISA \citep{Barack:2003fp, Robson:2018ifk} and the phenomenological waveform \citep{Ajith_2007, Ajith_etal_2011}. The population model of SMBH binary mergers is constructed with galaxy merger rate together with the intrinsic relationship between SMBH binary mass $M$ and descendant galaxy stellar mass $M_*$.
The delay time distribution is inferred via the framework of hierarchical Bayesian inference. By simply assuming the best fit value of galaxy merger rate (from an Illustris simulation \citep{Vicente2015} and observational results of Schechter function, see Tables \ref{table_galaxy_merger_rate} --\ref{tab:Schechter_parameters}) and the $M$--$M_*$ relationship gives in \citet{Kormendy:2013dxa}, our results (Fig. \ref{delay_posterior_gaussian}--\ref{confidence_delay_powerlaw}) show that the delay time distribution can be properly recovered within the uncertainties.
To compare different delay models, the method of Bayesian model selection is used to quantify the relative faithfulness of different models based on data. The result (Table \ref{tab:gaussian_delay_models}) shows that at least for the models discussed in Sec.\ref{compare_delay_model}, the statistical evidence to distinguish different delay models is significant. Of course, one possible caveat of this analysis - the systematic error of the galaxy merger rate and $M$--$M_*$ relationship, is only partially addressed by marginalizing over uncertainties of modeling parameters. This part ought to be updated with future observations of galaxy merger rate via large-scale surveys, and more precise and consistent cosmological simulations.
We have only considered a few delay-time models motivated by simple physical driving mechanisms.
In reality, the formation and evolution of SMBHs or SMBH binaries are influenced by more complex astrophysical conditions, such as the impact of seeding models of SMBH formations \citep{Klein:2015hvg, Toubiana:2021iuw}, the role of mass accretion for the individuals in SMBH binaries \citep{Callegari_2011}, SMBH binary mergers in multiple systems \citep{Hoffman_Loeb_2007}, etc. All these effects should in principle contribute to the final distribution of SMBH binary mergers, and should be considered in a systematic framework of population analysis. In addition, as we have only discussed events detected by LISA, it will be interesting and straightforward to extend the framework discussed here to also include possible PTA events.
\section*{Acknowledgements}
We acknowledge the use of the HPC Cluster in Perimeter Institute and the HPC Cluster of the National Supercomputing Center in Beijing. This work makes use of the open-sourced python package emcee \citep{Foreman2013emcee}. Y. F. is partly supported by the National Science Foundation
of China (NSFC) Grant No. 11721303, and the fellowship of China Postdoctoral Science Foundation No.
2021M690228. HY is supported by the Natural Sciences and Engineering Research
Council of Canada and in part by Perimeter Institute for Theoretical Physics.
Research at Perimeter Institute is supported in
part by the Government of Canada through the Department of
Innovation, Science and Economic Development Canada and
by the Province of Ontario through the Ministry of Colleges
and Universities.
| 42345858dcd25a75d3d63974ca83660f6c0488b8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
Applications of big data analytics have brought many new opportunities to economic research. With the developments of modern quantitative tools, economists are now able to analyze datasets that contain tens of millions of observations. One motivation of using datasets with ultra-large sample sizes is to deal with econometric models with a large number of parameters within the framework of simple parametric models. Suppose that there are $d$ parameters to estimate with a sample of size $n$. For instance, a parametric regression model can be as large as $d \sim 1,000$ because a large number of controls might be necessary. This could be viewed as a high-dimensional estimation problem if $n$ is relatively small, say several thousands.
In the recent decade, new machine learning techniques have been applied in economics to deal with high-dimensional estimation problems \citep{BCH:2014,AI:2019}; however, they arguably require more demanding conditions such as sparsity assumptions and inference is less straightforward.
This paper deals with the setting of ``ultra-large" sample size with $n\sim 10^7$, which is not an unusual sample size in applied microeconomic studies. With this size of the sample,
straightforward parametric estimators can be used with $d \sim 1,000$ under standard textbook assumptions.
With datasets containing millions of observations, one major challenge of parametric estimation and inference based on extreme estimators \citep[e.g.,][]{newey1994} is that they would require huge computing powers and memories that are often not accessible using ordinary personal computers.
While inference is desirable in empirical studies to quantify statistical uncertainty, it often requires more computationally demanding tasks than point estimation.
Typically, asymptotic normal inference with extreme estimators involves solving optimization problems and computing asymptotic covariance matrices. Computationally efficient implementation of both tasks is crucial with ultra-large datasets.
In this paper, we focus on such a challenging instance, namely, to run large-scale quantile regression, which
has been increasingly popular since the groundbreaking work of \citet{Koenker1978}.
One of the important economic applications of quantile regression is to study the wage structure
\citep[see, e.g., ][among others]{chamberlain1994quantile,Buchinsky:1994,Buchinsky:1998,Gosling:2000,angrist2006quantile,CFM:2013}.
The common data source for the U.S. wage structure is IPUMS USA. For instance, the 2000 5\% sample in IPUMS USA contains more than 14 million observations and more than 4 million even after restricting the sample to a subpopulation of working adults. We use the data from IPUMS USA to study the trends in the gender gap in terms of the college wage premium. In our empirical application, we aim to control for work experience by flexibly interacting workers' age with various state-level dummies, which creates over one thousand regressors, thereby motivating the need of ultra-large datasets.
We find that existing inference methods are not applicable due to either time or memory constraints, while our inference method provides new insight into the trends in the gender gap.
To the best of knowledge, our empirical illustration is the first to obtain confidence intervals based on quantile regression with more than 4 million observations and more than 1000 regressors.
We propose a very fast statistical inference framework to analyze cross-sectional data with millions of observations as typically the case for the IPUMS data by embracing the stochastic subgradient descent techniques, one of the most active research areas in machine learning. While we focus on the quantile regression framework, whose inference has been well known to be a hard problem, our proposed method can be easily generalized to other econometric frameworks.
\subsection{Standard inference for quantile regression}
Consider the setting of a linear quantile regression model for which
data $\{ Y_i \equiv (y_i,x_i) \in \mathbb{R}^{1+d} : i=1,\ldots,n\}$ are generated from
\begin{align}\label{model}
y_i = x_i'\beta^* + \varepsilon_i,
\end{align}
where
$\beta^*$ is a vector of unknown parameters
and the unobserved error $\varepsilon_i$ satisfies
$P(\varepsilon_i \leq 0|x_i)= \tau$ for a fixed quantile $\tau \in(0,1)$.
Note that $\beta^{*}$ is characterized by
\[
\beta^{*}:=\arg\min_{\beta\in\mathbb{R}^{d}}Q\left(\beta\right),
\]
where
$
Q(\beta):= \mathbb E [q(\beta, Y_i)]$ with the check function
$
q(\beta, Y_i):= (y_i-x_i'\beta)(\tau-I\{y_i-x_i'\beta\leq 0\}).
$
Here, $I(\cdot)$ is the usual indicator function.
The standard approach for estimating $\beta^{*}$ is to use the following M-estimator
originally proposed by \citet{Koenker1978}:
\begin{align}\label{def:m-estimator}
\widehat{\beta}_n := \arg\min_{\beta\in\mathbb{R}^{d}} \; \frac{1}{n} \sum_{i=1}^n q(\beta, Y_i).
\end{align}
See \citet{Koenker2005} and \citet{Koenker17} for a monograph and a review of recent
developments, respectively.
There have been two potential challenges for the inference of standard quantile regression. First is the computational issues. The optimization problem is typically reformulated to a linear programming problem, and solved using interior-point algorithms \citep[e.g.,][]{Portnoy:Koenker:97}.
The second challenge is associated with the statistical inference. As was shown by
\citet{Koenker1978}, the asymptotic distribution of the M-estimator is:
$$
\sqrt{n}(\widehat\beta_n-\beta^*) \overset{d}{\to} N(0, \tau(1-\tau)G^{-1}\mathbb Ex_ix_i' G^{-1}),\quad G= \mathbb E f_{\varepsilon}(0|x_i)x_ix_i'
$$
where $f_{\varepsilon}(\cdot|x_i)$ is the conditional distribution of $\varepsilon_i$ given $x_i$, assuming data are independent and identically distributed (i.i.d.).
Hence in the heteroskedastic setting, standard inference based on the ``plug-in'' method would require nonparametrically estimating the conditional density function. An alternative inference would be based on bootstrap.
One of the attempts to solving the computational/inference difficulties is to rely on the \emph{smoothing} idea, via either smoothing the estimating equation \citep{Horowitz:1998} or the convolution-type smoothing \citep{FGH,HE2021,TanWangZhou}. Both require a choice of the smoothing bandwidth. In particular, the convolution-type smoothing (\texttt{conquer} as coined by \citet{HE2021}) has received recent attention in the literature because the optimization problem is convex, so it is more scalable. Figure 1 of \citet{HE2021} shows that \texttt{conquer} works well for point estimations when the sample size ranges from $10^3\sim 10^6$ with the number of regressors $d \approx n^{1/2}$.
Meanwhile, in terms of inference, both types of smoothed estimators are first-order asymptotically equivalent to
the unsmoothed estimator $\widehat{\beta}_n$. Meanwhile,
the scale of the numerical studies of these solutions is less ambitious for the purpose of statistical inference. For instance, the largest model considered for inference in \citet{HE2021} is merely $(n, d) = (4000, 100)$ (see Figure 7 in their paper).
Therefore, there is a scalability gap between point estimation and inference in the literature.
In this paper, we aim to bridge this gap by proposing a method for fast inference.
\subsection{The proposed fast S-subGD framework}
We focus on the inference problem at the scale up to $(n, d) = (10^7, 10^3)$, which we refer to as an ``ultra-large" quantile regression problem.
This is a possible scale with IPUMS USA, but very difficult to deal with using benchmark inference procedures for quantile regression.
To tackle this large-scale problem, we estimate $\beta^*$ via stochastic (sub)gradient descent.
Our assumption on the asymptotic regime is that $d$ is fixed but $n$ grows to infinity.
Suppose that we have i.i.d. data from a large cross-sectional survey:
$$Y_1,..., Y_n,\quad Y_i=(x_i, y_i),$$
where the ordering of these observations is randomized.
Our method produces a sequence of ``estimators'', denoted by $\beta_i$, which is a solution path being updated as
$$
\beta_{i}=\beta_{i-1}-\gamma_{i}\nabla q\left(\beta_{i-1},Y_{i}\right).
$$
Here $Y_i$ is a ``new'' observation in the randomized data sequence. Also, $\nabla q\left(\beta_{i-1},Y_i\right)$ is a subgradient of the check function with respect to the current update, and $\gamma_i$ is a pre-determined learning rate. Then we take the average of the sequence
$$
\bar\beta_n:=\frac{1}{n}\sum_{i=1}^n\beta_i,
$$
which is called as the \citet{Polyak1990}-\citet{ruppert1988efficient} average in the machine learning literature.
It will be shown in this paper that
$\sqrt{n}(\bar{\beta}_{n} - \beta^*)$ is asymptotically normal
with the asymptotic variance same as that of the standard \citet{Koenker1978} estimator $\widehat{\beta}_n$. So it achieves the same first-order asymptotic efficiency.
The main novelty of this paper comes from how to conduct inference based on
$\bar{\beta}_{n}$.
We use a recursive transformation of $\beta_i$'s that is suitable for on-line updating to construct asymptotically pivotal statistics.
That is, we studentize $\sqrt{n}\left(\bar{\beta}_{n}-\beta^{*}\right)$ via a random scaling matrix $\widehat V_n$ whose exact form will be given later.
The resulting statistic is not asymptotically normal but \emph{asymptotically pivotal} in the sense that its asymptotic distribution is free of any unknown nuisance parameters; thus, its critical values are easily available.
Furthermore, the random scaling quantity $\widehat{V}_{n}$ does not require any additional inputs other than stochastic subgradient paths $\beta_i$, and can be computed very fast.
The main contribution of this paper is computational. We combine the idea of stochastic subgradient descent with random scaling and make large-scale inference for quantile regression much more practical. We demonstrate the usefulness of our method via Monte Carlo experiments. Empirically, we apply it to studying the gender gap in college wage premiums using the IPUMS dataset. The proposed inference method of quantile regression to the ultra-big dataset reveals some interesting new features. First, it shows heterogeneous effects over different quantile levels. Second, we find that the female college wage premium is significantly higher that that of male workers at the median, which is different from what have been concluded in the literature. In fact, although the S-subGD inference tends to be conservative than the standard normal approximation, it reveals statistically significant results while controlling over $10^3$ covariates to mitigate confounding effects. With more availability of such a large dataset, the S-subGD method will make it possible to obtain convincing empirical evidence for other analyses.
\subsection{The related literature}
Our estimator is motivated from the literature on \textit{online learning}, where data were collected in a streaming fashion, and as a new observation $Y_i$ arrives, the estimator is updated to $\beta_i$. Under the M-estimation framework with strongly convex, differentiable loss functions, \citet{polyak1992acceleration} showed that the \citet{Polyak1990}-\citet{ruppert1988efficient} average is asymptotically normal. More recently, \cite{fang2018online}, \cite{chen2020statistical}, and \cite{zhu2021online} studied the statistical inference problem and proposed methods that require consistent estimation of the asymptotic variance matrix, implementation of bootstrap, or the use of ``batch-means'' approach. Besides the smoothness assumption that excludes quantile regression, these inference solutions are not computationally attractive for datasets as large as ones this paper is concerned with.
The idea of random scaling is borrowed from
the time-series literature on fixed bandwidth heteroskedasticity and autocorrelation robust (HAR) inference \citep[e.g.,][]{kiefer2000simple,sun2008optimal,sun2014fixed,lazarus2018har} and has been recently adopted in the context
of stochastic gradient descent (SGD) in machine-learning problems: online inference for linear mean regression \citep{lee2021fast},
federated learning \citep{li2021statistical},
and
Kiefer-Wolfowitz methods \citep{ChenLai2021}
among others.
The current paper shares the basic inference idea with our previous work \citep{lee2021fast}; however, quantile regression is sufficiently different from mean regression and it requires further theoretical development that is not covered in the existing work.
In particular, our previous work is based on \citet{polyak1992acceleration}, which limits its analysis to
differentiable and strongly convex objective functions; however,
the check function $q(\beta, Y_i)$ for quantile regression is non-differentiable and is not strongly convex.
We overcome these difficulties by using the results given in \citet{gadat2022optimal}.
One could have adopted convolution-type smoothing for S-subGD as the corresponding objective function is differentiable and locally strongly convex. We do not go down this alternative route in this paper as the existence of a sequence of smoothing bandwidths converging to zero complicates theoretical analysis and it would be difficult to optimally choose the sequence of smoothing bandwidths along with the sequence of learning rates.
In the econometric literature, \cite{forneron2021estimation} and \citet{forneron2022estimation} developed SGD-based resampling schemes that deliver both point estimates and standard errors within the same optimization framework.
The major difference is that their stochastic path is a Newton-Raphson type which requires computing the Hessian matrix or its approximation. Hence, their framework is not applicable to quantile regression because the check function is not twice differentiable. In other words, for quantile regression inferences with over millions of observations, estimating the inverse Hessian matrix is a task that we particularly would like to avoid.
As an alternative to full sample estimation, one may consider sketching \citep[e.g., see][for a review from an econometric perspective]{Lee:Ng:2020}. For example, \citet{Yang:QR:ICML} proposed a fast randomized algorithm for large-scale quantile regression and solved the problem of size $n \sim 10^{10}$ and $d = 12$ by randomly creating a subsample of about $n = 10^5$. However, the theoretical analysis carried out in \citet{Yang:QR:ICML} is limited to approximations of the optimal value of the check function and does not cover the issue of inference. In mean regression models, the precision of a sketched estimator depends on the subsample size and is typically worse than that of the full sample estimator \citep{Lee:Ng:2020,Lee:Ng:2022}.
\subsection{Notation}
Let $a'$ and $A'$, respectively, denote the transpose of vector $a$ and matrix $A$.
Let $| a |$ denote the Euclidean norm of vector $a$
and $\| A \|$ the Frobenius norm of matrix $A$.
Also, let $\ell^{\infty}\left[0,1\right]$ denote the set of bounded continuous functions on $[0,1].$
Let $I(A)$ be the usual indicator function, that is $I(A) = 1$ if $A$ is true and 0 otherwise.
For a symmetric, positive definite matrix $S$, let $\lambda_{\min}(S)$ denote its smallest eigenvalue.
\section{A Fast Algorithm for Quantile Inference}\label{sec:algorithm}
\subsection{Stochastic subgradient descent (S-subGD)}
In this section, we describe our proposed inference algorithm.
The general idea behind the algorithm is proposed in our previous work \citep{lee2021fast} but it is now tailored to run quantile regression with cross-sectional data $\{Y_1,..., Y_n \}$, where $Y_i=(x_i, y_i)$.
First, we randomize the ordering of the observed data:
$$
Y_1=(x_1, y_1),..., Y_n=(x_n, y_n).
$$
We start with an initialized estimator, denoted by $\beta_0$, using a method which we shall discuss later. Then produce a sequence of S-subGD solution path which updates according to the rule:
\begin{align}\label{eq:SGD1}
\beta_{i}=\beta_{i-1}-\gamma_{i}\nabla q\left(\beta_{i-1},Y_{i}\right),
\end{align}
where the updating subgradient depends on a \textit{single ``next'' observation} $Y_i$:
\begin{align}\label{eq:SS}
\nabla q(\beta, Y_i) := x_i [I\{y_i \leq x_i'\beta\} - \tau].
\end{align}
This is the subgradient
of $q\left(\beta,Y_i\right)$ with respect to $\beta$. Here $\beta_{i-1}$ is the ``current'' estimate, which is updated to $\beta_{i}$ when the next observation $Y_i$ comes into play.
Also $\gamma_i$ is a pre-determined step size, which is also called a learning rate and is assumed to have the form
$\gamma_{i}:=\gamma_0 i^{-a} $ for some constants $\gamma_0 > 0$ and $a \in (1/2,1)$. Then we take the average of the sequence
$
\bar\beta_n:=\frac{1}{n}\sum_{i=1}^n\beta_i
$ as the final estimator.
Computing this estimator does not require storing the historical data; thus, it is very fast and memory-efficient even when $n\sim 10^7$. In fact, as we shall see below, the gain of memory-efficiency is much more substantial for inference.
\subsection{The pivotal statistic}
The usual inference based on a consistent estimator of the asymptotic variance for $\sqrt{n}(\bar\beta_n-\beta^*)$ is very computationally demanding. For instance, one needed to estimate the conditional density function. Instead of pursuing a consistent estimator for the asymptotic variance, we apply the random scaling using the \emph{fixed-b} standardization. Motivated by the fact that the solution path $\beta_i$ is recursively updated, we apply the random scaling approach by defining:
\begin{align}\label{def:random-scaling}
\widehat{V}_{n} := \frac{1}{n}\sum_{s=1}^{n}
\left\{ \frac{1}{\sqrt{n}} \sum_{i=1}^{s} \left( \beta_{i}-\bar{\beta}_{n} \right) \right \}
\left\{ \frac{1}{\sqrt{n}} \sum_{i=1}^{s} \left( \beta_{i}-\bar{\beta}_{n} \right) \right \}'.
\end{align}
Inference will be conducted based on the standardization using $\widehat V_n.$ Computing $\widehat V_n$ can be very efficient even if $n\sim 10^7$ or larger, since it can be recursively computed as detailed later.
Once $\bar{\beta}_{n}$ and $\widehat{V}_{n}$ are obtained, it is straightforward to carry out inference. For example, for the $j$ th component of $\bar \beta_n$, let $\widehat V_{n,jj}$ denote the $(j,j)$ th diagonal entry of $\widehat V_n$. The t-statistic is then
\begin{align}\label{t-stat}
\frac{\sqrt{n}\left(\bar{\beta}_{n,j}-\beta_{j}^{*}\right)}{\sqrt{\widehat{V}_{n,jj}}}
\end{align}
whose asymptotic distribution is mixed normal and symmetric around zero, and we shall formally derive it in the next section.
The mixed normal asymptotic distribution for the t-statistic in \eqref{t-stat} is the same as the distribution of the statistics observed in the estimation of the cointegration vector by \citet{johansen1991estimation} and \citet{Abadir:Paruolo:97} and in the heteroskedasticity and autocorrelation robust inference in \citet{kiefer2000simple}. They are different statistics but have the identical distribution as functions of the standard Wiener process as shown by \citet{Abadir:Paruolo:02}.
We can use the t-statistic to construct
the $(1-\alpha)$ asymptotic confidence interval for the $j$-th element $\beta_{j}^{*}$ of $\beta^{*}$ by
\[
\left[
\bar{\beta}_{n,j} - \textrm{cv} (1- \alpha/2) \sqrt{\frac{\widehat{V}_{n,jj}}{n}}, \;
\bar{\beta}_{n,j} + \textrm{cv} (1- \alpha/2) \sqrt{\frac{\widehat{V}_{n,jj}}{n}} \;
\right],
\]
where the critical value $\textrm{cv} (1- \alpha/2)$ is tabulated in \citet[Table I]{Abadir:Paruolo:97}. For easy reference, we reproduce the critical values in Table~\ref{tab:cv}.
When $\alpha = 0.05$, the critical value is 6.747.
Critical values for testing linear restrictions $H_{0}: R\beta^{*} = c$
are given in \citet[Table II]{kiefer2000simple}.
\begin{table}[htbp]
\caption{Asymptotic critical values of the t-statistic}
\begin{center}
\begin{tabular}{lcccc}
\hline
Probability & 90\% & 95\% & 97.5\% & 99\% \\
Critical Value & 3.875 & 5.323 & 6.747 & 8.613 \\
\hline
\end{tabular}
\end{center}
\label{tab:cv}
\begin{minipage}{1\textwidth}
{Note.
The table gives one-sided asymptotic critical values that satisfy
$\mathrm{Pr}( \hat{t} \leq c ) = p$ asymptotically, where
$p \in \{0.9, 0.95, 0.975, 0.99\}$.
Source: \citet[Table I]{Abadir:Paruolo:97}.
\par}
\end{minipage}
\end{table}%
\subsection{Practical details for memory efficiency}
While computing $(\widehat V_n,\bar\beta_n)$ is straightforward for datasets of medium sample sizes, it is not practical when $n\sim 10^7$ or even larger size, as storing the entire solution path $\{\beta_i\}$ can be infeasible for memories of usual personal computing devices.
As the main advantage of our random-scaling based S-subGD approach is its computational efficiency, we demonstrate that we can also construct a solution path for $(\widehat V_n,\bar\beta_n)$ along the way of updating $\beta_i$, so all relevant quantities of the proposed inference can be obtained recursively.
Specifically, Let $(\bar\beta_{i-1},\beta_{i-1})$ be the current update when we use up to $i-1$ observations. We then update using the new observation $Y_i$ by:
\begin{eqnarray}
\beta_i&=& \beta_{i-1} - \gamma_i \nabla q(\beta_{i-1}, Y_i)\label{eq.7beta}\\
\bar\beta_i&=& \bar\beta_{i-1}\frac{i-1}{i} + \beta_i\frac{1}{i}\label{eqa7}\\
\widehat V_i&=& i^{-2}\left(A_i-\bar\beta_i b_i' -b_i\bar\beta_i'+\bar\beta_i\bar\beta_i'\sum_{s=1}^is^2\right),\label{eqa8}
\end{eqnarray}
where the intermediate quantities $A_i$ and $b_i$ are also updated sequences:
$$
A_i=A_{i-1} +i^2\bar\beta_i\bar\beta_i',\quad b_i= b_{i-1} + i^2\bar\beta_i.
$$
Therefore computing $(\bar\beta_n, \widehat V_n)$ can be cast recursively until $i=n$, and does not require saving the entire solution path of $\beta_i$.
A good initial value could help achieve the computational stability in practice, though it does not matter in the theoretical results.
We recommend applying \texttt{conquer} in \citet{HE2021} to the subsample (e.g. 5\% or 10\%) to estimate the initial value, which we adopted in our numerical experiments.
Specifically, we start with a smooth quantile regression proposed in \citet{FGH} and \citet{HE2021}:
$$
\beta_0=\arg\min\sum_{i\in \mathcal S}q_{n}(\beta, Y_i)
$$
where $q_n$ is a convolution-type-smoothed checked function, and $\mathcal S$ is a randomized subsample.
The \texttt{conquer} algorithm
is very fast to compute a point estimate of $\beta_0$ but it is much less scalable with respect to inference (see the Monte Carlo results later in the paper).
For the intermediate quantities $A_i$ and $b_i$, we set $A_0 = 0$ and $b_0 = 0$.
Another approach is the burn-in method that applies the S-subGD to some initial observations, and throw them away for the main inference procedure.
In our experiments, initial values estimated by \texttt{conquer} outperform those by the burn-in method because \texttt{conquer} tend to produce high-quality initial solutions.
\subsection{Sub-vector inference} In most empirical studies, while many covariates are being controlled in the model, of interest are often just one or two key independent variables that are policy-related. In such cases, we are particularly interested in making inference for subvectors, say $\beta^*_1$, of the original vector $\beta^*$. Sub-vector inference is not often direct target of interest in the quantile regression literature, partially because inference regarding the full vector $\beta^*$ and then converting to the subvector is often scalable up to the medium sample size. But this is no longer the case for the ultra-large sample size that is being considered here.
It is straightforward to tailor our updating algorithm to focusing on sub-vectors. Continue denoting $\beta_i$ as the $i$-th update of the full vector. While the full vector $\beta_i$ still needs to be computed in the S-subGD, both the Polyak-Ruppert average and the random scale can be updated only up to the scale of the sub-vector. Specifically, let $\bar\beta_{i, sub}$ denote the sub-vector of $\bar\beta_i$, corresponding to the sub-vector of interest. Also let $\widehat V_{i,sub}$ denote the sub-matrix of $\widehat V_i$; both are updated according to the following rule:
\begin{eqnarray}
\beta_i&=& \text{updated the same as (\ref{eq.7beta})}\\
\bar\beta_{i,sub}&=& \bar\beta_{i-1, sub}\frac{i-1}{i} + \beta_{i,sub}\frac{1}{i}\label{eqa7-sub}\\
\widehat V_{i,sub}&=& i^{-2}\left(A_{i,sub}-\bar\beta_{i,sub} b_{i,sub}' -b_{i,sub}\bar\beta_{i,sub}'+\bar\beta_{i,sub}\bar\beta_{i,sub}'\sum_{s=1}^is^2\right) \\
A_{i,sub}&=&A_{i-1,sub} +i^2\bar\beta_{i,sub}\bar\beta_{i,sub}',\quad b_{i,sub}= b_{i-1,sub} + i^2\bar\beta_{i,sub}.\label{end123}
\end{eqnarray}
In most cases, steps (\ref{eqa7-sub})-(\ref{end123}) only involve small dimensional objects.
\begin{comment}
\section{The cluster-dependent setting}
In this section we consider a cluster-dependent setting:
$$
y_{i,t} = x_{i,t}'\beta^*+\varepsilon_{i,t},\quad i=1...n, t=1...T_i
$$
where $P(\varepsilon_{i,t}\leq 0|x_{i,t})=\tau$. Here there are $n\to\infty$ clusters, and within each cluster we have finite many ($T_i$) observations. Extending to the cluster-dependence is desirable for two reasons: first, instead of updating whenever a new observation $(Y_{it}):=(x_{i,t}, y_{i,t})$ becomes available, we can treat all individuals within a cluster as a ``batch'', and update $\beta^*$ batch-by-batch. This can further speed up the algorithm. Second, our cross-sectional assumptions require i.i.d. observations, which should be revised to allow dependence within-clusters, since it is desirable in economic studies that units within clusters should influence one another.
Our theory assumes observations are independent across clusters, but allows arbitrary dependence within clusters.
Inference for quantile models that allows arbitrary within-cluster dependence has been studied by \cite{hagemann2017cluster}, who proposed a bootstrap procedure that resamples purturbations of the first order condition (the wild bootstrap). Here we consider the same cluster structure, but practically with the number of cluster as large as $n\sim 10^7$, we need a much more scalable method to conduct inference, which is the goal being studied in this section.
Specifically, we update the estimator at the $i$ level:
\begin{align}\label{eq:SGDpanel}
\beta_{i}=\beta_{i-1}-\gamma_{i}\frac{1}{T_i}\sum_{t=1}^{T_i}\nabla q\left(\beta_{i-1},Y_{i,t}\right),
\end{align}
where $q(\beta, Y_{i,t}) := x_{i,t} [I\{y_{i,t} \leq x_{i,t}'\beta\} - \tau] $. The final estimator is still $\bar\beta_n=\frac{1}{n}\sum_{i=1}^n\beta_i$, and with the solution path $\beta_i$, we use the same random scaling $\widehat V_n$ as in the cross-sectional setting. Both $(\bar\beta_n,\widehat V_n)$ can be updated using (\ref{eqa7})-(\ref{eqa8}), and obtain the pivotal statistic that is ready for inference. For instance, the $(1-\alpha)$ asymptotic confidence interval for the $j$-th element $\beta_{j}^{*}$ of $\beta^{*}$ by
\[
\left[
\bar{\beta}_{n,j} - \textrm{cv} (1- \alpha/2) \sqrt{\frac{\widehat{V}_{nT,jj}}{n}}, \;
\bar{\beta}_{n,j} + \textrm{cv} (1- \alpha/2) \sqrt{\frac{\widehat{V}_{nT,jj}}{n}} \;
\right]
\]
with the same critical value as in the cross-sectional setting.
\end{comment}
\section{Asymptotic Theory}\label{sec:theory}
In this section, we present asymptotic theory that underpins our inference method.
We begin by stating the regularity conditions for quantile regression.
\begin{asm}[Conditions for Quantile Regression]\label{asm:qr}
Suppose that $\{ Y_i \equiv (y_i, x_i) \}_{i=1}^n$ are independent and identically distributed (i..i.d.), generated from
$$
y_{i}= x_i'\beta^*_\tau + \varepsilon_i,
$$
where $P(\varepsilon_i\leq 0|x_i)= \tau$ for a fixed quantile $\tau \in(0,1)$,
$f_{\varepsilon}(\cdot|x_i)$ denotes the conditional density of $\varepsilon_i$ given $x_i$,
and its partial derivative
$\frac{d}{de}f_{\varepsilon}(\cdot|x_i)$ exists.
Furthermore, assume that
\begin{enumerate}[(i)]
\item
there exist positive constants $\epsilon$ and $c_0$ such that
$$
\inf_{ | \beta-\beta^* |<\epsilon }\lambda_{\min} \left( \mathbb E [x_ix_i'f_{\varepsilon}(x_i'(\beta-\beta^*)|x_i)] \right) > c_0,
$$
\item $\sup_b\mathbb E [ \|x_i\|^3A(b, x_i) ] <C$ for some constant $C < \infty$, where $$A(b, x_i) := \left|\frac{d}{de}f_{\varepsilon}(x_i'b|x_i)\right| + f_{\varepsilon}(x_i'b|x_i),$$
\item $\mathbb E[ (\|x_i\|^6+1)\exp(\|x_i\|^2) ]<C$ for some constant $C < \infty$,
\end{enumerate}
\end{asm}
Condition (i) can be viewed as a sort of global identifiability condition, which is reasonable as $\beta \mapsto q(\beta, Y_i)$ is convex.
Conditions (ii) and (iii) impose some moment conditions and are satisfied, for example, if $x_i$ are bounded
and
both $f_{\varepsilon}(\cdot|x_i)$ and $\frac{d}{de}f_{\varepsilon}(\cdot|x_i)$ are uniformly bounded.
Under a different setting (with smoothed and globally convex loss functions), \citet{lee2021fast} established a functional central limit theorem (FCLT) for the aggregated solution path of SGD. We
extend our previous result to the quantile regression model. Under Assumption~\ref{asm:qr}, we establish the following \emph{functional} central limit theorem (FCLT):
\begin{equation}\label{eq4}
\frac{1}{\sqrt{n}}\sum_{i=1}^{\left[nr\right]}\left(\beta_{i}-\beta^{*}\right)\Rightarrow \Upsilon^{1/2}W\left(r\right),\quad r\in\left[0,1\right],
\end{equation}
where $\Rightarrow$ stands for the weak convergence in $\ell^{\infty}\left[0,1\right]$,
$W\left(r\right)$ stands for a vector of the independent standard
Wiener processes on $\left[0,1\right]$,
and
$ \Upsilon := H^{-1}SH^{-1}$, with
$
S := \mathbb E [x_i x_i'] \tau(1-\tau)
$ and $H := \mathbb E [x_i x_i' f_{\varepsilon}(0|x_i)]$.
The FCLT in \eqref{eq4} states that the partial sum of the recursively updated estimates $\beta_i$ converges weakly to a rescaled Wiener process, with the scaling matrix equal to a square root of the asymptotic variance of the usual quantile regression estimator $\widehat{\beta}_n$.
Building on the FCLT, we propose
a large-scale inference procedure.
More generally, for any $\ell \leq d$ linear restrictions
\[
H_{0}: R\beta^{*} = c,
\]
where $R$ is an $(\ell \times d)$-dimensional known matrix of rank $\ell$
and
$c$ is an $\ell$-dimensional known vector,
the conventional Wald test based
on $\widehat{V}_{n}$ becomes asymptotically pivotal.
We formally state the main theoretical results in the following theorem.
\begin{thm}[Main Theorem]
\label{thm:Wald:qr} Suppose that
$H_{0}: R\beta^{*} = c$ holds with
$\mathrm{rank}(R)=\ell$. Under Assumption~\ref{asm:qr},
the FCLT in (\ref{eq4}) holds and
\begin{eqnarray*}
&&n\left(R\bar{\beta}_{n}-c\right)'\left(R\widehat{V}_{n}R'\right)^{-1}\left(R\bar{\beta}_{n}-c\right)\overset{d}{\to}W\left(1\right)'\left(\int_{0}^{1}\bar{W}(r)\bar{W}(r)'dr\right)^{-1}W\left(1\right),
\end{eqnarray*}
where $W$ is an $\ell$-dimensional vector of the standard Wiener
processes and $\bar{W}\left(r\right):=W\left(r\right)-rW\left(1\right)$.
\end{thm}
As an important special case of Theorem \ref{thm:Wald:qr}, the t-statistic defined in \eqref{t-stat} converges in distribution to the following pivotal limiting distribution:
for each $j = 1,\ldots,d$,
\begin{align}\label{t-stat-limit}
\frac{\sqrt{n}\left(\bar{\beta}_{n,j}-\beta_{j}^{*}\right)}{\sqrt{\widehat{V}_{n,jj}}}
\overset{d}{\to}
W_1\left(1\right) \left[ \int_{0}^{1} \left\{ W_1\left(r\right)-rW_1\left(1\right) \right\}^2 dr\right]^{-1/2},
\end{align}
where
$W_1$ is a one-dimensional standard Wiener process.
\begin{comment}
\subsection{The cluster-dependent setting}
We consider the ``large $N$, small $T_i$'' asymptotic scenario, so there are many clusters each with finite cluster size. As for the dependence assumption, we assume observations are independent across clusters but allow arbitrary dependence within clusters.
\begin{asm}\label{asm:qrpanel}
Suppose that $\{ Y_{i,t} \equiv (y_{i,t}, x_{i,t}) \}$ are independent and identically distributed across $i$, and generated from
$$
y_{i,t}= x_{i,t}'\beta^*_\tau + \varepsilon_{i,t},
$$
where $P(\varepsilon_{i,t}\leq 0|x_{i,t})= \tau$ for a fixed quantile $\tau \in(0,1)$. Let $Q(\beta):= \mathbb E [q(\beta, Y_{i,t})]$ for all $i$ and $t$ and $f_{\varepsilon}(\cdot|x_{i,t})$ denote the conditional density of $\varepsilon_{i,t} $ given $x_{i,t}$,
and its partial derivative
$\frac{d}{de}f_{\varepsilon}(\cdot|x_{i,t})$ exist.
Furthermore, assume that
\begin{enumerate}
\item
there exist positive constants $\epsilon$ and $c_0$ such that
$$
\inf_{ | \beta-\beta^* |<\epsilon }\lambda_{\min} \left( \mathbb E [x_{i,t}x_{i,t}'f_{\varepsilon}(x_{i,t}'(\beta-\beta^*)|x_{i,t})] \right) > c_0,
$$
\item $\sup_b\mathbb E [ \|x_{i,t}\|^3A(b, x_{i,t}) ] <C$ for some constant $C < \infty$, where $ A(b, x_{i,t}) := \left|\frac{d}{de}f_{\varepsilon}(x_{i,t}'b|x_{i,t})\right| + f_{\varepsilon}(x_{i,t}'b|x_{i,t}),$
\item $\mathbb E[ (\| T^{-1} \sum_{t=1}^{T} x_{i,t}\|^6+1)\exp(\|T^{-1} \sum_{t=1}^{T}x_{i,t}\|^2) ]<C$ for some constant $C < \infty$,
\end{enumerate}
\end{asm}
\begin{thm}[Main Theorem]
\label{thm:Waldpanel} Suppose that
$H_{0}: R\beta^{*} = c$ holds with
$\mathrm{rank}(R)=\ell$. Under Assumption~\ref{asm:qrpanel},
\begin{eqnarray*}
&&n\left(R\bar{\beta}_{n}-c\right)'\left(R\widehat{V}_{n}R'\right)^{-1}\left(R\bar{\beta}_{n}-c\right)\overset{d}{\to}W\left(1\right)'\left(\int_{0}^{1}\bar{W}(r)\bar{W}(r)'dr\right)^{-1}W\left(1\right),
\end{eqnarray*}
where $W$ and $\bar{W}\left(r\right) $ as as the same as in Theorem \ref{thm:Wald:qr}.
\end{thm}
\end{comment}
\section{Monte Carlo Experiments}\label{sec:MC}
In this section we investigate the performance of the S-subGD random scaling method via Monte Carlo experiments.
The main question of these numerical experiments is whether the proposed method is feasible for a large scale model.
The simulation is based on the following data generating process:
\begin{align*}
y_{i} = x_i'\beta + \varepsilon_i~~\mbox{for}~~i=1,\ldots,n,
\end{align*}
where $x_i$ is a $(d+1)$-dimensional covariate vector whose first element is 1 and the remaining $d$ elements are generated from $\mathcal{N} (0,I_d)$. The error term $\varepsilon_i$ is generated from $\mathcal N(0,1)$, and the parameter value $\beta$ is set to be $(1,\ldots,1)$.
Without loss of generality, we focus on the second element of $\beta$.
To accommodate a large scale model, the sample size varies in $n \in \{10^5, 10^6, 10^7\}$, and the dimension of $x$ varies in $d\in \{10, 20, 40, 80, 160, 320, 1000\}$.
The initial value $\beta_0$ is estimated by the convolution-type smoothed quantile regression (\texttt{conquer}) in \citet{HE2021} and we do not burn in any observations.
The learning rate is set to be $\gamma=\gamma_0 t^{-a}$ with $\gamma_0=1$ and $a=0.501$.
The simulation results are summarized from $1000$ replications of each design.
We compare the performance of the proposed method with four additional alternatives:
\begin{description}
\item[(i)] S-subGD: the proposed S-subGD random scaling method.
\item[(ii)] QR: the classical quantile regression method.
Given the scale of models, we apply the Frisch-Newton algorithm after preprocessing and the Huber sandwich estimate under the local linearity assumption of the conditional density function by selecting `pfn' and `nid' options in R package \texttt{quantreg} (CRAN version 5.88).
\item[(iii)] CONQUER-plugin: estimates the parameter using the conquer method, and estimates the asymptotic variance by plugging in the parameter estimates; implemented using the R package \texttt{conquer} (CRAN version 1.3.0).
\item[(iv)] CONQUER-bootstrap: estimates the parameter using the conquer method, and applies the multiplier bootstrap method (see \citet{HE2021} for details); implemented using the R package \texttt{conquer}. We set the number of bootstrap samples as 1000.
\item[(v)] SGD-bootstrap: estimates using the S-subGD method and conduct the inference by an online bootstrap method (see \citet{fang2018online} for details). We set the number of bootstrap samples as 1000.
\end{description}
\begin{figure}[!htbp]
\caption{Computation time} \label{fig:time}
\centering
\vskip10pt
\includegraphics[width=\textwidth]{figure/fig_time.png}
\includegraphics[width=\textwidth]{figure/fig_rel_time.png}
\end{figure}
\begin{figure}[!htbp]
\caption{Coverage and CI Length} \label{fig:ci}
\centering
\vskip10pt
\includegraphics[width=\textwidth]{figure/fig_coverage.png}
\includegraphics[width=\textwidth]{figure/fig_ci_length.png}
\end{figure}
We use the following performance measures: the computation time, the coverage rate, and the length of the 95\% confidence interval.
Note that the nominal coverage probability is 0.95.
The computation time is measured by 10 replications on a separate desktop computer, equipped with an AMD Ryzen Threadripper 3970X CPU (3.7GHz) and a 192GB RAM.
Other measures are computed by 1000 replications on the cluster system composed of several types of CPUs.
We set the time budget as 10 hours for a single estimation and the RAM budget as 192GB.
Figures \ref{fig:time}--\ref{fig:ci} summarize the simulation result. To save space, we report the results of $d=20, 80,$ and $320$. We provide tables for all the simulation designs in the appendix.
First, we observe that S-subGD is easily scalable to a large simulation design. On the contrary, some methods cannot complete the computation within the time/memory budget. For instance, CONQUER-plugin does not work for any $d$ when $n=10^6$ or $n=10^7$ because of the memory budget. QR shows a similar issue when $n=10^7$ and $d=320$. CONQUER-bootstrap cannot meet the 10-hour time budget for a single replication when $n=10^7$ and $d=320$. They are all denoted as `NA' in the figures.
Second, S-subGD outperforms the alternative in terms of computational efficiency. The bottom panel of Figure \ref{fig:time} shows the relative computation time, and the efficiency scales are 10's or 100's levels. In case of CONQUER-plugin, it is slower than S-subGD in the scale of 1000's. If we convert them into the actual computation time, S-subGD takes 16 seconds for the design of $n=10^6$ and $d=320$. However, QR, CONQUER-bootstrap, and SGD-bootstrap take 1374 seconds (22 minutes 54 seconds), 4113 seconds (1 hour 8 minutes 33 seconds), and 1643 seconds (27 minutes 23 seconds) on average of 10 replications.
Third, all methods shows satisfactory results in terms of the coverage rate and the length of the confidence interval. The top panel of Figure \ref{fig:ci} reports that coverage rates are all around 0.95 although S-subGD and SGD-bootstrap are slightly under-reject and over-reject when $d=320$. The average lengths of the confidence interval are reported on the bottom panel. The performance of S-subGD is comparable to QR, CONQUER-plugin, and CONQUER-bootstrap. SGD-bootstrap shows much larger CI lengths, especially when $d=320$.
Finally, we stretch the computational burden to $d=1000$ and $n=10^7$ and check the performance of S-subGD. Table \ref{tb_SGD_rs} reports the simulation result along with different sizes of $d$. S-subGD still completes the computation within a reasonable time range. The average computation time is under 762 seconds (12 minutes 42 seconds). The coverage rate and CI length becomes slightly worse than those of smaller $d$'s but they still fall in a satisfactory range.
In sum, S-subGD performs well when we conduct a large scale inference problem in quantile regression. It is scalable to the range where the existing alternatives cannot complete the computation within a time/memory budget. In addition to computational efficiency, the coverage rate and CI lengths are also satisfactory. Thus, S-subGD provides a solution to those who hesitate to apply quantile regression with a large scale data for the computational burden.
\begin{table}[htb]
\centering
\caption{Performance of S-subGD: $n=10^7$}
\label{tb_SGD_rs}
\begin{tabular}{c S[table-format=3.2]S[table-format=1.3]S[table-format=1.4]}
\hline
{$d$} & {Time (sec.)} & {Coverage Rate} & {CI Length} \\
\hline
10 & 5.87 & 0.965 & 0.0020 \\
20 & 11.05 & 0.955 & 0.0020 \\
40 & 21.86 & 0.954 & 0.0020 \\
80 & 43.12 & 0.952 & 0.0020 \\
160 & 81.35 & 0.953 & 0.0021 \\
320 & 166.40 & 0.963 & 0.0011 \\
1000 & 762.16 & 0.925 & 0.0461 \\
\hline
\end{tabular}
\end{table}
\section{Inference for the College Wage Premium}
The study of the gender gap in the college wage premium has been a long-standing important question in labor economics. The literature has pointed out a stylized fact that the higher college wage premium for women as the major cause for attracting more women to attend and graduate from colleges than men (e.g., \cite{goldin2006homecoming, chiappori2009investment}). Meanwhile, \citet{hubbard2011phantom} pointed out a bias issue associated with the ``topcoded'' wage data, where the use of the sample from Current Population Survey (CPS) often censors the wage data at a maximum value. As a remedy for it, \citet{hubbard2011phantom} proposed to use quantile regression that are robust to censoring. Analyzing the CPS data during 1970-2008, he found no gender difference in the college wage premium later years once the topcoded bias has been accounted for.
We revisit this problem by estimating and comparing the college wage premium between women and men. While the data also contains the topcoded issue, the quantile regression is less sensitive to the values of upper-tail wages. By applying the new S-subGD inference to the ultra-big dataset we are using, we aim at achieving the following goals in the empirical study: (1) identify (if any) the heterogeneous effects across quantiles; (2) understand the trends in the college wage premium respectively for female and male; and (3) understand the gender difference in the college wage premium.
We use the data from IPUMS USA that contains several millions of workers. The main motivation of using ultra-large datasets for wage regressions is to deal with high-dimensional regressors within a simple parametric quantile regression model. As has been crucially pointed out in the literature, work experience is
an important factor in wage regressions, but is difficult to be measured precisely. One typical means of controlling for the experience is to flexibly interact workers' age with various state-level dummies, which would create over one thousand regressors, so that a very large sample size would be desirable to obtain precise estimates.
But the ultra-large dataset makes most existing inference procedures for quantile regression fail to work.
\subsection{The data}
We use the samples over six different years (1980, 1990, 2000-2015) from IPUMS USA at \url{https://usa.ipums.org/usa/}.
In the years from 1980 to 2000, we use the 5\% State sample which is a 1-in-20 national random sample of the population.
In the remaining years, we use the American Community Survey (ACS) each year. The sampling ratio varies from 1-in-261 to 1-in-232 in 2001-2004, but it is set to a 1-in-100 national random sample of the population after 2005.
To balance the sample size, we bunch the sample every 5 year after 2001.
We also provide the estimation results using separate but smaller samples in 2005, 2010, and 2015 in the appendix, which is similar to those reported in this section.
We restrict our sample to $White$, $18 \le Age \le 65$, and $Wage \ge \$62$, which is a half of minimum wage earnings in 1980 ($\$3.10 \times 40 \mbox{hours} \times 1/2$).
$Wage$ denotes the implied weekly wage that is computed by dividing yearly earnings by weeks worked last year.
Since 2006, weeks worked last year are available only as an interval value and we use the midpoint of an interval.
We only consider full-time workers who worked more than 30 hours per week.
Then, we compute the \textit{real} wage using the personal consumption expenditures price index (PCEPI) normalized in 1980.
The data cleaning leaves us with 3.6-4.7 million observations besides 2001-2005, where we have around 2.5 million observations.
Table \ref{tb:summary_stat} reports some summary statistics on the key variables, where $Educ$ denotes an education dummy for some college or above.
The table confirms that female college education has increased substantially over the years and female workers received more college education than male workers since 1990.
\begin{table}[ht]
\caption{Summary Statistics} \label{tb:summary_stat}
\centering
\begin{tabular}{crcccc}
\hline
Year & Sample Size & $\mathbb E(Female$) & $\mathbb E(Educ$) & $\mathbb E(Educ$|$Male$) & $\mathbb E(Educ$|$Female$) \\
\hline
1980 & 3,659,684 & 0.390 & 0.433 & 0.444 & 0.416 \\
1990 & 4,192,119 & 0.425 & 0.543 & 0.537 & 0.550 \\
2000 & 4,479,724 & 0.439 & 0.600 & 0.578 & 0.629 \\
2001-2005 & 2,493,787 & 0.447 & 0.642 & 0.619 & 0.670 \\
2006-2010 & 4,708,119 & 0.447 & 0.663 & 0.631 & 0.701 \\
2011-2015 & 4,542,874 & 0.447 & 0.686 & 0.646 & 0.735 \\
\hline
\end{tabular}
\end{table}
\subsection{The quantile regression model}
We use the following baseline model:
\begin{footnotesize}
\begin{align*}
\log(Wage_i) = \beta_0 + \beta_1 Female_i + \beta_2 Educ_i + \beta_3 Female_i\cdot Educ_i + \theta_1'X_i + \theta_2' (X_i\cdot Female_i) + \varepsilon_i,
\end{align*}
\end{footnotesize}where $Wage_i$ is a real weekly wage in 1980 terms, $Female_i$ is a female dummy, $Educ_i$ is an education dummy for some college or above, and $X_i$ is a vector of additional control variables.
For control variable $X_i$, we use 12 age group dummies with a four-year interval, 51 states dummies (including D.C.), and their interactions. Note that $(X_i\cdot Female_i)$ implies that there exist up to 3-way interactions.
The model contains 1226 covariates in total.
We also add 4 additional year dummies for the 5-year combined samples after 2001.
We estimate the model using the proposed S-subGD method over 9 different quantiles: $\tau=0.1,0.2,\ldots,0.9$.
We obtain the starting value of S-subGD Inference by applying CONQUER to the 10\% subsample.\footnote{Recall that $X_i$ contains a large number of dummy variables in this empirical application. If the subsample size is too small, CONQUER may face a singularity issue and cannot compute the initial value. Alternatively, one may use a randomzied algorithm of \citet{Yang:QR:ICML}.
}
We do not burn in any sample and set the learning rate as $\gamma_0 t^{-a}$ with $\gamma_0=1$ and $a=0.501$.
We use the Compute Canada cluster system equipped with the AMD Milan CPUs whose clock speed is 2.65 GHz and with 650 GB of RAMs.
For comparison, we also try to estimate the model using the standard QR method with the `quantreg' R package, but it fails to compute the result in the given cluster environments.
\subsection{The results}
Figures \ref{fig:premium_design_5yr_full}--\ref{fig:diff_5yr_full} and Table \ref{tb:median_5yr_full} summarize the estimation results.
We also provide the full estimation results in the appendix.
\begin{figure}[!htbp]
\centering
\caption{College Wage Premium:
Combining 5-Year Data
}
\label{fig:premium_design_5yr_full}
\hskip15pt
\begin{tabular}{c c c}
\includegraphics[scale=0.23]{figure/chart_0.1_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.2_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.3_03_FullT_5yr.png} \\
\includegraphics[scale=0.23]{figure/chart_0.4_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.5_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.6_03_FullT_5yr.png} \\
\includegraphics[scale=0.23]{figure/chart_0.7_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.8_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.9_03_FullT_5yr.png}
\end{tabular}
\end{figure}
\begin{figure}
\centering
\caption{Difference of College Premium: Combining 5-Year Data}
\label{fig:diff_5yr_full}
\hskip15pt
\begin{tabular}{c c c}
\includegraphics[scale=0.23]{figure/diff_chart_0.1_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.2_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.3_03_FullT_5yr.png} \\
\includegraphics[scale=0.23]{figure/diff_chart_0.4_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.5_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.6_03_FullT_5yr.png} \\
\includegraphics[scale=0.23]{figure/diff_chart_0.7_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.8_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.9_03_FullT_5yr.png}
\end{tabular}
\end{figure}
Our results share several interesting features about the estimated college wage premium. First, some degrees of heterogeneity are found over different quantiles. At the upper tail quantile, $\tau=0.9$, male college premium is slightly higher than the female one since 2000, while female college premiums are higher over all years in other quantiles. This shows an interesting feature about high-income individuals that cannot be viewed by the mean regression, though sensible economic interpretations might be subject to debates. Also, note that the 95\% confidence intervals are wider for tail quantiles ($\tau=0.1$ and 0.9, respectively).
Second, college premiums increase over time for both male and female. This result coincides with the overall findings in the literature. It is interesting that college premiums get flatter since 2010 for lower quantiles ($\tau \le 0.5$).
Third, when we focus on the median ($\tau=0.5$) as reported in Table \ref{tb:median_5yr_full}, we observe that the female-male college premium difference shows an inverse ``U-shape" pattern.
In addition, the difference is always significantly positive, which is different from the result in \citet{hubbard2011phantom}.
He found the gender difference to be insignificant starting from around the year 2000 from the CPS data.
We also note that the computation time is within a reasonable range meanwhile it is not feasible to estimate the model with the alternative approaches.
Finally, the use of ultra-big dataset and the S-subGD inference in quantile regression provide several features for the college premium analysis. Although the S-subGD inference tends to be conservative than the standard normal approximation at the 95\% level, it reveals statistically significant results while controlling over $10^3$ covariates to mitigate confounding effects. With more availability of such a large dataset, the S-subGD method will make it possible to obtain convincing empirical evidence for other analyses.
\begin{table}[htb]
\centering
\caption{College Wage Premium: $\tau=0.5$}\label{tb:median_5yr_full}
\begin{tabular}{ccccc}
\hline
Year & Female & Male & Difference & Time (min.) \\
\hline
\underline{$\tau = 0.5$}\\
1980 & 0.2365 & 0.1988 & 0.0377 & 29.4 \\
& [0.2294,0.2435] & [0.1945,0.2030] & [0.0291,0.0463] & \\
1990 & 0.3667 & 0.2962 & 0.0705 & 34.2 \\
& [0.3603,0.3732] & [0.2942,0.2982] & [0.0634,0.0777] & \\
2000 & 0.4101 & 0.3439 & 0.0662 & 36.7 \\
& [0.4056,0.4146] & [0.3372,0.3506] & [0.0552,0.0772] & \\
2001-2005 & 0.4468 & 0.3854 & 0.0613 & 20.2 \\
& [0.4369,0.4567] & [0.3765,0.3944] & [0.0554,0.0673] & \\
2006-2010 & 0.4791 & 0.4271 & 0.0520 & 47.7 \\
& [0.4748,0.4834] & [0.4174,0.4368] & [0.0454,0.0585] & \\
2011-2015 & 0.4957 & 0.4498 & 0.0458 & 46.0 \\
& [0.4887,0.5027] & [0.4455,0.4542] & [0.0348,0.0568] & \\
\hline
\end{tabular}
\flushleft{\footnotesize Notes. The male college premium is from $\hat{\beta}_2$ and the female college premium is from $\hat{\beta}_2+\hat{\beta}_3$. Thus, the college premium difference between male and female workers is from $\hat{\beta}_3$. }
\end{table}
\section{Conclusions}
We have proposed an inference method for large-scale quantile regression to analyze datasets whose sizes are of order $(n, p) \sim (10^7, 10^3)$. Our method runs very fast, based on the stochastic sub-gradient descent updates, and constructs asymptotically pivotal statistics via random scaling.
There are a couple of extensions worth pursuing in future research.
First, we may build on \citet{chernozhukov2022fast} to develop fast inference for the quantile regression process.
Second, while we focus on the regular quantile regression models where the quantile is assumed to be bounded away from both zero and one, our framework is also potentially applicable to extreme quantile regression models \citep[see, e.g.,][for a review]{chernozhukov2016extremal}. The ability of handling ultra-large datasets is also appealing for extreme quantile regression models because the resulting sample size is considerably larger at the extreme quantiles. A formal treatment for this case is out of the scope of this paper, and we leave it for future studies.
\section{Introduction}\label{sec:intro}
Applications of big data analytics have brought many new opportunities to economic research. With the developments of modern quantitative tools, economists are now able to analyze datasets that contain tens of millions of observations. One motivation of using datasets with ultra-large sample sizes is to deal with econometric models with a large number of parameters within the framework of simple parametric models. Suppose that there are $d$ parameters to estimate with a sample of size $n$. For instance, a parametric regression model can be as large as $d \sim 1,000$ because a large number of controls might be necessary. This could be viewed as a high-dimensional estimation problem if $n$ is relatively small, say several thousands.
In the recent decade, new machine learning techniques have been applied in economics to deal with high-dimensional estimation problems \citep{BCH:2014,AI:2019}; however, they arguably require more demanding conditions such as sparsity assumptions and inference is less straightforward.
This paper deals with the setting of ``ultra-large" sample size with $n\sim 10^7$, which is not an unusual sample size in applied microeconomic studies. With this size of the sample,
straightforward parametric estimators can be used with $d \sim 1,000$ under standard textbook assumptions.
With datasets containing millions of observations, one major challenge of parametric estimation and inference based on extreme estimators \citep[e.g.,][]{newey1994} is that they would require huge computing powers and memories that are often not accessible using ordinary personal computers.
While inference is desirable in empirical studies to quantify statistical uncertainty, it often requires more computationally demanding tasks than point estimation.
Typically, asymptotic normal inference with extreme estimators involves solving optimization problems and computing asymptotic covariance matrices. Computationally efficient implementation of both tasks is crucial with ultra-large datasets.
In this paper, we focus on such a challenging instance, namely, to run large-scale quantile regression, which
has been increasingly popular since the groundbreaking work of \citet{Koenker1978}.
One of the important economic applications of quantile regression is to study the wage structure
\citep[see, e.g., ][among others]{chamberlain1994quantile,Buchinsky:1994,Buchinsky:1998,Gosling:2000,angrist2006quantile,CFM:2013}.
The common data source for the U.S. wage structure is IPUMS USA. For instance, the 2000 5\% sample in IPUMS USA contains more than 14 million observations and more than 4 million even after restricting the sample to a subpopulation of working adults. We use the data from IPUMS USA to study the trends in the gender gap in terms of the college wage premium. In our empirical application, we aim to control for work experience by flexibly interacting workers' age with various state-level dummies, which creates over one thousand regressors, thereby motivating the need of ultra-large datasets.
We find that existing inference methods are not applicable due to either time or memory constraints, while our inference method provides new insight into the trends in the gender gap.
To the best of knowledge, our empirical illustration is the first to obtain confidence intervals based on quantile regression with more than 4 million observations and more than 1000 regressors.
We propose a very fast statistical inference framework to analyze cross-sectional data with millions of observations as typically the case for the IPUMS data by embracing the stochastic subgradient descent techniques, one of the most active research areas in machine learning. While we focus on the quantile regression framework, whose inference has been well known to be a hard problem, our proposed method can be easily generalized to other econometric frameworks.
\subsection{Standard inference for quantile regression}
Consider the setting of a linear quantile regression model for which
data $\{ Y_i \equiv (y_i,x_i) \in \mathbb{R}^{1+d} : i=1,\ldots,n\}$ are generated from
\begin{align}\label{model}
y_i = x_i'\beta^* + \varepsilon_i,
\end{align}
where
$\beta^*$ is a vector of unknown parameters
and the unobserved error $\varepsilon_i$ satisfies
$P(\varepsilon_i \leq 0|x_i)= \tau$ for a fixed quantile $\tau \in(0,1)$.
Note that $\beta^{*}$ is characterized by
\[
\beta^{*}:=\arg\min_{\beta\in\mathbb{R}^{d}}Q\left(\beta\right),
\]
where
$
Q(\beta):= \mathbb E [q(\beta, Y_i)]$ with the check function
$
q(\beta, Y_i):= (y_i-x_i'\beta)(\tau-I\{y_i-x_i'\beta\leq 0\}).
$
Here, $I(\cdot)$ is the usual indicator function.
The standard approach for estimating $\beta^{*}$ is to use the following M-estimator
originally proposed by \citet{Koenker1978}:
\begin{align}\label{def:m-estimator}
\widehat{\beta}_n := \arg\min_{\beta\in\mathbb{R}^{d}} \; \frac{1}{n} \sum_{i=1}^n q(\beta, Y_i).
\end{align}
See \citet{Koenker2005} and \citet{Koenker17} for a monograph and a review of recent
developments, respectively.
There have been two potential challenges for the inference of standard quantile regression. First is the computational issues. The optimization problem is typically reformulated to a linear programming problem, and solved using interior-point algorithms \citep[e.g.,][]{Portnoy:Koenker:97}.
The second challenge is associated with the statistical inference. As was shown by
\citet{Koenker1978}, the asymptotic distribution of the M-estimator is:
$$
\sqrt{n}(\widehat\beta_n-\beta^*) \overset{d}{\to} N(0, \tau(1-\tau)G^{-1}\mathbb Ex_ix_i' G^{-1}),\quad G= \mathbb E f_{\varepsilon}(0|x_i)x_ix_i'
$$
where $f_{\varepsilon}(\cdot|x_i)$ is the conditional distribution of $\varepsilon_i$ given $x_i$, assuming data are independent and identically distributed (i.i.d.).
Hence in the heteroskedastic setting, standard inference based on the ``plug-in'' method would require nonparametrically estimating the conditional density function. An alternative inference would be based on bootstrap.
One of the attempts to solving the computational/inference difficulties is to rely on the \emph{smoothing} idea, via either smoothing the estimating equation \citep{Horowitz:1998} or the convolution-type smoothing \citep{FGH,HE2021,TanWangZhou}. Both require a choice of the smoothing bandwidth. In particular, the convolution-type smoothing (\texttt{conquer} as coined by \citet{HE2021}) has received recent attention in the literature because the optimization problem is convex, so it is more scalable. Figure 1 of \citet{HE2021} shows that \texttt{conquer} works well for point estimations when the sample size ranges from $10^3\sim 10^6$ with the number of regressors $p \approx n^{1/2}$.
Meanwhile, in terms of inference, both types of smoothed estimators are first-order asymptotically equivalent to
the unsmoothed estimator $\widehat{\beta}_n$. Meanwhile,
the scale of the numerical studies of these solutions is less ambitious for the purpose of statistical inference. For instance, the largest model considered for inference in \citet{HE2021} is merely $(n, p) = (4000, 100)$ (see Figure 7 in their paper).
Therefore, there is a scalability gap between point estimation and inference in the literature.
In this paper, we aim to bridge this gap by proposing a method for fast inference.
\subsection{The proposed fast S-subGD framework}
We focus on the inference problem at the scale up to $(n, d) = (10^7, 10^3)$, which we refer to as an ``ultra-large" quantile regression problem.
This is a possible scale with IPUMS USA, but very difficult to deal with using benchmark inference procedures for quantile regression.
To tackle this large-scale problem, we estimate $\beta^*$ via stochastic (sub)gradient descent.
Our assumption on the asymptotic regime is that $d$ is fixed but $n$ grows to infinity.
Suppose that we have i.i.d. data from a large cross-sectional survey:
$$Y_1,..., Y_n,\quad Y_i=(x_i, y_i),$$
where the ordering of these observations is randomized.
Our method produces a sequence of ``estimators'', denoted by $\beta_i$, which is a solution path being updated as
$$
\beta_{i}=\beta_{i-1}-\gamma_{i}\nabla q\left(\beta_{i-1},Y_{i}\right).
$$
Here $Y_i$ is a ``new'' observation in the randomized data sequence. Also, $\nabla q\left(\beta_{i-1},Y_i\right)$ is a subgradient of the check function with respect to the current update, and $\gamma_i$ is a pre-determined learning rate. Then we take the average of the sequence
$$
\bar\beta_n:=\frac{1}{n}\sum_{i=1}^n\beta_i,
$$
which is called as the \citet{Polyak1990}-\citet{ruppert1988efficient} average in the machine learning literature.
It will be shown in this paper that
$\sqrt{n}(\bar{\beta}_{n} - \beta^*)$ is asymptotically normal
with the asymptotic variance same as that of the standard \citet{Koenker1978} estimator $\widehat{\beta}_n$. So it achieves the same first-order asymptotic efficiency.
The main novelty of this paper comes from how to conduct inference based on
$\bar{\beta}_{n}$.
We use a recursive transformation of $\beta_i$'s that is suitable for on-line updating to construct asymptotically pivotal statistics.
That is, we studentize $\sqrt{n}\left(\bar{\beta}_{n}-\beta^{*}\right)$ via a random scaling matrix $\widehat V_n$ whose exact form will be given later.
The resulting statistic is not asymptotically normal but \emph{asymptotically pivotal} in the sense that its asymptotic distribution is free of any unknown nuisance parameters; thus, its critical values are easily available.
Furthermore, the random scaling quantity $\widehat{V}_{n}$ does not require any additional inputs other than stochastic subgradient paths $\beta_i$, and can be computed very fast.
The main contribution of this paper is computational. We combine the idea of stochastic subgradient descent with random scaling and make large-scale inference for quantile regression much more practical. We demonstrate the usefulness of our method via Monte Carlo experiments. Empirically, we apply it to studying the gender gap in college wage premiums using the IPUMS dataset. The proposed inference method of quantile regression to the ultra-big dataset reveals some interesting new features. First, it shows heterogeneous effects over different quantile levels. Second, we find that the female college wage premium is significantly higher that that of male workers at the median, which is different from what have been concluded in the literature. In fact, although the S-subGD inference tends to be conservative than the standard normal approximation, it reveals statistically significant results while controlling over $10^3$ covariates to mitigate confounding effects. With more availability of such a large dataset, the S-subGD method will make it possible to obtain convincing empirical evidence for other analyses.
\subsection{The related literature}
Our estimator is motivated from the literature on \textit{online learning}, where data were collected in a streaming fashion, and as a new observation $Y_i$ arrives, the estimator is updated to $\beta_i$. Under the M-estimation framework with strongly convex, differentiable loss functions, \citet{polyak1992acceleration} showed that the \citet{Polyak1990}-\citet{ruppert1988efficient} average is asymptotically normal. More recently, \cite{fang2018online}, \cite{chen2020statistical}, and \cite{zhu2021online} studied the statistical inference problem and proposed methods that require consistent estimation of the asymptotic variance matrix, implementation of bootstrap, or the use of ``batch-means'' approach. Besides the smoothness assumption that excludes quantile regression, these inference solutions are not computationally attractive for datasets as large as ones this paper is concerned with.
The idea of random scaling is borrowed from
the time-series literature on fixed bandwidth heteroskedasticity and autocorrelation robust (HAR) inference \citep[e.g.,][]{kiefer2000simple,sun2008optimal,sun2014fixed,lazarus2018har} and has been recently adopted in the context
of stochastic gradient descent (SGD) in machine-learning problems: online inference for linear mean regression \citep{lee2021fast},
federated learning \citep{li2021statistical},
and
Kiefer-Wolfowitz methods \citep{ChenLai2021}
among others.
The current paper shares the basic inference idea with our previous work \citep{lee2021fast}; however, quantile regression is sufficiently different from mean regression and it requires further theoretical development that is not covered in the existing work.
In particular, our previous work is based on \citet{polyak1992acceleration}, which limits its analysis to
differentiable and strongly convex objective functions; however,
the check function $q(\beta, Y_i)$ for quantile regression is non-differentiable and is not strongly convex.
We overcome these difficulties by using the results given in \citet{gadat2022optimal}.
One could have adopted convolution-type smoothing for S-subGD as the corresponding objective function is differentiable and locally strongly convex. We do not go down this alternative route in this paper as the existence of a sequence of smoothing bandwidths converging to zero complicates theoretical analysis and it would be difficult to optimally choose the sequence of smoothing bandwidths along with the sequence of learning rates.
In the econometric literature, \cite{forneron2021estimation} and \citet{forneron2022estimation} developed SGD-based resampling schemes that deliver both point estimates and standard errors within the same optimization framework.
The major difference is that their stochastic path is a Newton-Raphson type which requires computing the Hessian matrix or its approximation. Hence, their framework is not applicable to quantile regression because the check function is not twice differentiable. In other words, for quantile regression inferences with over millions of observations, estimating the inverse Hessian matrix is a task that we particularly would like to avoid.
As an alternative to full sample estimation, one may consider sketching \citep[e.g., see][for a review from an econometric perspective]{Lee:Ng:2020}. For example, \citet{Yang:QR:ICML} proposed a fast randomized algorithm for large-scale quantile regression and solved the problem of size $n \sim 10^{10}$ and $d = 12$ by randomly creating a subsample of about $n = 10^5$. However, the theoretical analysis carried out in \citet{Yang:QR:ICML} is limited to approximations of the optimal value of the check function and does not cover the issue of inference. In mean regression models, the precision of a sketched estimator depends on the subsample size and is typically worse than that of the full sample estimator \citep{Lee:Ng:2020,Lee:Ng:2022}.
\subsection{Notation}
Let $a'$ and $A'$, respectively, denote the transpose of vector $a$ and matrix $A$.
Let $| a |$ denote the Euclidean norm of vector $a$
and $\| A \|$ the Frobenius norm of matrix $A$.
Also, let $\ell^{\infty}\left[0,1\right]$ denote the set of bounded continuous functions on $[0,1].$
Let $I(A)$ be the usual indicator function, that is $I(A) = 1$ if $A$ is true and 0 otherwise.
For a symmetric, positive definite matrix $S$, let $\lambda_{\min}(S)$ denote its smallest eigenvalue.
\section{A Fast Algorithm for Quantile Inference}\label{sec:algorithm}
\subsection{Stochastic subgradient descent (S-subGD)}
In this section, we describe our proposed inference algorithm.
The general idea behind the algorithm is proposed in our previous work \citep{lee2021fast} but it is now tailored to run quantile regression with cross-sectional data $\{Y_1,..., Y_n \}$, where $Y_i=(x_i, y_i)$.
First, we randomize the ordering of the observed data:
$$
Y_1=(x_1, y_1),..., Y_n=(x_n, y_n).
$$
We start with an initialized estimator, denoted by $\beta_0$, using a method which we shall discuss later. Then produce a sequence of S-subGD solution path which updates according to the rule:
\begin{align}\label{eq:SGD1}
\beta_{i}=\beta_{i-1}-\gamma_{i}\nabla q\left(\beta_{i-1},Y_{i}\right),
\end{align}
where the updating subgradient depends on a \textit{single ``next'' observation} $Y_i$:
\begin{align}\label{eq:SS}
\nabla q(\beta, Y_i) := x_i [I\{y_i \leq x_i'\beta\} - \tau].
\end{align}
This is the subgradient
of $q\left(\beta,Y_i\right)$ with respect to $\beta$. Here $\beta_{i-1}$ is the ``current'' estimate, which is updated to $\beta_{i}$ when the next observation $Y_i$ comes into play.
Also $\gamma_i$ is a pre-determined step size, which is also called a learning rate and is assumed to have the form
$\gamma_{i}:=\gamma_0 i^{-a} $ for some constants $\gamma_0 > 0$ and $a \in (1/2,1)$. Then we take the average of the sequence
$
\bar\beta_n:=\frac{1}{n}\sum_{i=1}^n\beta_i
$ as the final estimator.
Computing this estimator does not require storing the historical data; thus, it is very fast and memory-efficient even when $n\sim 10^7$. In fact, as we shall see below, the gain of memory-efficiency is much more substantial for inference.
\subsection{The pivotal statistic}
The usual inference based on a consistent estimator of the asymptotic variance for $\sqrt{n}(\bar\beta_n-\beta^*)$ is very computationally demanding. For instance, one needed to estimate the conditional density function. Instead of pursuing a consistent estimator for the asymptotic variance, we apply the random scaling using the \emph{fixed-b} standardization. Motivated by the fact that the solution path $\beta_i$ is recursively updated, we apply the random scaling approach by defining:
\begin{align}\label{def:random-scaling}
\widehat{V}_{n} := \frac{1}{n}\sum_{s=1}^{n}
\left\{ \frac{1}{\sqrt{n}} \sum_{i=1}^{s} \left( \beta_{i}-\bar{\beta}_{n} \right) \right \}
\left\{ \frac{1}{\sqrt{n}} \sum_{i=1}^{s} \left( \beta_{i}-\bar{\beta}_{n} \right) \right \}'.
\end{align}
Inference will be conducted based on the standardization using $\widehat V_n.$ Computing $\widehat V_n$ can be very efficient even if $n\sim 10^7$ or larger, since it can be recursively computed as detailed later.
Once $\bar{\beta}_{n}$ and $\widehat{V}_{n}$ are obtained, it is straightforward to carry out inference. For example, for the $j$ th component of $\bar \beta_n$, let $\widehat V_{n,jj}$ denote the $(j,j)$ th diagonal entry of $\widehat V_n$. The t-statistic is then
\begin{align}\label{t-stat}
\frac{\sqrt{n}\left(\bar{\beta}_{n,j}-\beta_{j}^{*}\right)}{\sqrt{\widehat{V}_{n,jj}}}
\end{align}
whose asymptotic distribution is mixed normal and symmetric around zero, and we shall formally derive it in the next section.
The mixed normal asymptotic distribution for the t-statistic in \eqref{t-stat} is the same as the distribution of the statistics observed in the estimation of the cointegration vector by \citet{johansen1991estimation} and \citet{Abadir:Paruolo:97} and in the heteroskedasticity and autocorrelation robust inference in \citet{kiefer2000simple}. They are different statistics but have the identical distribution as functions of the standard Wiener process as shown by \citet{Abadir:Paruolo:02}.
We can use the t-statistic to construct
the $(1-\alpha)$ asymptotic confidence interval for the $j$-th element $\beta_{j}^{*}$ of $\beta^{*}$ by
\[
\left[
\bar{\beta}_{n,j} - \textrm{cv} (1- \alpha/2) \sqrt{\frac{\widehat{V}_{n,jj}}{n}}, \;
\bar{\beta}_{n,j} + \textrm{cv} (1- \alpha/2) \sqrt{\frac{\widehat{V}_{n,jj}}{n}} \;
\right],
\]
where the critical value $\textrm{cv} (1- \alpha/2)$ is tabulated in \citet[Table I]{Abadir:Paruolo:97}. For easy reference, we reproduce the critical values in Table~\ref{tab:cv}.
When $\alpha = 0.05$, the critical value is 6.747.
Critical values for testing linear restrictions $H_{0}: R\beta^{*} = c$
are given in \citet[Table II]{kiefer2000simple}.
\begin{table}[htbp]
\caption{Asymptotic critical values of the t-statistic}
\begin{center}
\begin{tabular}{lcccc}
\hline
Probability & 90\% & 95\% & 97.5\% & 99\% \\
Critical Value & 3.875 & 5.323 & 6.747 & 8.613 \\
\hline
\end{tabular}
\end{center}
\label{tab:cv}
\begin{minipage}{1\textwidth}
{Note.
The table gives one-sided asymptotic critical values that satisfy
$\mathrm{Pr}( \hat{t} \leq c ) = p$ asymptotically, where
$p \in \{0.9, 0.95, 0.975, 0.99\}$.
Source: \citet[Table I]{Abadir:Paruolo:97}.
\par}
\end{minipage}
\end{table}%
\subsection{Practical details for memory efficiency}
While computing $(\widehat V_n,\bar\beta_n)$ is straightforward for datasets of medium sample sizes, it is not practical when $n\sim 10^7$ or even larger size, as storing the entire solution path $\{\beta_i\}$ can be infeasible for memories of usual personal computing devices.
As the main advantage of our random-scaling based S-subGD approach is its computational efficiency, we demonstrate that we can also construct a solution path for $(\widehat V_n,\bar\beta_n)$ along the way of updating $\beta_i$, so all relevant quantities of the proposed inference can be obtained recursively.
Specifically, Let $(\bar\beta_{i-1},\beta_{i-1})$ be the current update when we use up to $i-1$ observations. We then update using the new observation $Y_i$ by:
\begin{eqnarray}
\beta_i&=& \beta_{i-1} - \gamma_i \nabla q(\beta_{i-1}, Y_i)\label{eq.7beta}\\
\bar\beta_i&=& \bar\beta_{i-1}\frac{i-1}{i} + \beta_i\frac{1}{i}\label{eqa7}\\
\widehat V_i&=& i^{-2}\left(A_i-\bar\beta_i b_i' -b_i\bar\beta_i'+\bar\beta_i\bar\beta_i'\sum_{s=1}^is^2\right),\label{eqa8}
\end{eqnarray}
where the intermediate quantities $A_i$ and $b_i$ are also updated sequences:
$$
A_i=A_{i-1} +i^2\bar\beta_i\bar\beta_i',\quad b_i= b_{i-1} + i^2\bar\beta_i.
$$
Therefore computing $(\bar\beta_n, \widehat V_n)$ can be cast recursively until $i=n$, and does not require saving the entire solution path of $\beta_i$.
A good initial value could help achieve the computational stability in practice, though it does not matter in the theoretical results.
We recommend applying \texttt{conquer} in \citet{HE2021} to the subsample (e.g. 5\% or 10\%) to estimate the initial value, which we adopted in our numerical experiments.
Specifically, we start with a smooth quantile regression proposed in \citet{FGH} and \citet{HE2021}:
$$
\beta_0=\arg\min\sum_{i\in \mathcal S}q_{n}(\beta, Y_i)
$$
where $q_n$ is a convolution-type-smoothed checked function, and $\mathcal S$ is a randomized subsample.
The \texttt{conquer} algorithm
is very fast to compute a point estimate of $\beta_0$ but it is much less scalable with respect to inference (see the Monte Carlo results later in the paper).
For the intermediate quantities $A_i$ and $b_i$, we set $A_0 = 0$ and $b_0 = 0$.
Another approach is the burn-in method that applies the S-subGD to some initial observations, and throw them away for the main inference procedure.
In our experiments, initial values estimated by \texttt{conquer} outperform those by the burn-in method because \texttt{conquer} tend to produce high-quality initial solutions.
\subsection{Sub-vector inference} In most empirical studies, while many covariates are being controlled in the model, of interest are often just one or two key independent variables that are policy-related. In such cases, we are particularly interested in making inference for subvectors, say $\beta^*_1$, of the original vector $\beta^*$. Sub-vector inference is not often direct target of interest in the quantile regression literature, partially because inference regarding the full vector $\beta^*$ and then converting to the subvector is often scalable up to the medium sample size. But this is no longer the case for the ultra-large sample size that is being considered here.
It is straightforward to tailor our updating algorithm to focusing on sub-vectors. Continue denoting $\beta_i$ as the $i$-th update of the full vector. While the full vector $\beta_i$ still needs to be computed in the S-subGD, both the Polyak-Ruppert average and the random scale can be updated only up to the scale of the sub-vector. Specifically, let $\bar\beta_{i, sub}$ denote the sub-vector of $\bar\beta_i$, corresponding to the sub-vector of interest. Also let $\widehat V_{i,sub}$ denote the sub-matrix of $\widehat V_i$; both are updated according to the following rule:
\begin{eqnarray}
\beta_i&=& \text{updated the same as (\ref{eq.7beta})}\\
\bar\beta_{i,sub}&=& \bar\beta_{i-1, sub}\frac{i-1}{i} + \beta_{i,sub}\frac{1}{i}\label{eqa7-sub}\\
\widehat V_{i,sub}&=& i^{-2}\left(A_{i,sub}-\bar\beta_{i,sub} b_{i,sub}' -b_{i,sub}\bar\beta_{i,sub}'+\bar\beta_{i,sub}\bar\beta_{i,sub}'\sum_{s=1}^is^2\right) \\
A_{i,sub}&=&A_{i-1,sub} +i^2\bar\beta_{i,sub}\bar\beta_{i,sub}',\quad b_{i,sub}= b_{i-1,sub} + i^2\bar\beta_{i,sub}.\label{end123}
\end{eqnarray}
In most cases, steps (\ref{eqa7-sub})-(\ref{end123}) only involve small dimensional objects.
\begin{comment}
\section{The cluster-dependent setting}
In this section we consider a cluster-dependent setting:
$$
y_{i,t} = x_{i,t}'\beta^*+\varepsilon_{i,t},\quad i=1...n, t=1...T_i
$$
where $P(\varepsilon_{i,t}\leq 0|x_{i,t})=\tau$. Here there are $n\to\infty$ clusters, and within each cluster we have finite many ($T_i$) observations. Extending to the cluster-dependence is desirable for two reasons: first, instead of updating whenever a new observation $(Y_{it}):=(x_{i,t}, y_{i,t})$ becomes available, we can treat all individuals within a cluster as a ``batch'', and update $\beta^*$ batch-by-batch. This can further speed up the algorithm. Second, our cross-sectional assumptions require i.i.d. observations, which should be revised to allow dependence within-clusters, since it is desirable in economic studies that units within clusters should influence one another.
Our theory assumes observations are independent across clusters, but allows arbitrary dependence within clusters.
Inference for quantile models that allows arbitrary within-cluster dependence has been studied by \cite{hagemann2017cluster}, who proposed a bootstrap procedure that resamples purturbations of the first order condition (the wild bootstrap). Here we consider the same cluster structure, but practically with the number of cluster as large as $n\sim 10^7$, we need a much more scalable method to conduct inference, which is the goal being studied in this section.
Specifically, we update the estimator at the $i$ level:
\begin{align}\label{eq:SGDpanel}
\beta_{i}=\beta_{i-1}-\gamma_{i}\frac{1}{T_i}\sum_{t=1}^{T_i}\nabla q\left(\beta_{i-1},Y_{i,t}\right),
\end{align}
where $q(\beta, Y_{i,t}) := x_{i,t} [I\{y_{i,t} \leq x_{i,t}'\beta\} - \tau] $. The final estimator is still $\bar\beta_n=\frac{1}{n}\sum_{i=1}^n\beta_i$, and with the solution path $\beta_i$, we use the same random scaling $\widehat V_n$ as in the cross-sectional setting. Both $(\bar\beta_n,\widehat V_n)$ can be updated using (\ref{eqa7})-(\ref{eqa8}), and obtain the pivotal statistic that is ready for inference. For instance, the $(1-\alpha)$ asymptotic confidence interval for the $j$-th element $\beta_{j}^{*}$ of $\beta^{*}$ by
\[
\left[
\bar{\beta}_{n,j} - \textrm{cv} (1- \alpha/2) \sqrt{\frac{\widehat{V}_{nT,jj}}{n}}, \;
\bar{\beta}_{n,j} + \textrm{cv} (1- \alpha/2) \sqrt{\frac{\widehat{V}_{nT,jj}}{n}} \;
\right]
\]
with the same critical value as in the cross-sectional setting.
\end{comment}
\section{Asymptotic Theory}\label{sec:theory}
In this section, we present asymptotic theory that underpins our inference method.
We begin by stating the regularity conditions for quantile regression.
\begin{asm}[Conditions for Quantile Regression]\label{asm:qr}
Suppose that $\{ Y_i \equiv (y_i, x_i) \}_{i=1}^n$ are independent and identically distributed (i..i.d.), generated from
$$
y_{i}= x_i'\beta^*_\tau + \varepsilon_i,
$$
where $P(\varepsilon_i\leq 0|x_i)= \tau$ for a fixed quantile $\tau \in(0,1)$,
$f_{\varepsilon}(\cdot|x_i)$ denotes the conditional density of $\varepsilon_i$ given $x_i$,
and its partial derivative
$\frac{d}{de}f_{\varepsilon}(\cdot|x_i)$ exists.
Furthermore, assume that
\begin{enumerate}[(i)]
\item
there exist positive constants $\epsilon$ and $c_0$ such that
$$
\inf_{ | \beta-\beta^* |<\epsilon }\lambda_{\min} \left( \mathbb E [x_ix_i'f_{\varepsilon}(x_i'(\beta-\beta^*)|x_i)] \right) > c_0,
$$
\item $\sup_b\mathbb E [ \|x_i\|^3A(b, x_i) ] <C$ for some constant $C < \infty$, where $$A(b, x_i) := \left|\frac{d}{de}f_{\varepsilon}(x_i'b|x_i)\right| + f_{\varepsilon}(x_i'b|x_i),$$
\item $\mathbb E[ (\|x_i\|^6+1)\exp(\|x_i\|^2) ]<C$ for some constant $C < \infty$,
\end{enumerate}
\end{asm}
Condition (i) can be viewed as a sort of global identifiability condition, which is reasonable as $\beta \mapsto q(\beta, Y_i)$ is convex.
Conditions (ii) and (iii) impose some moment conditions and are satisfied, for example, if $x_i$ are bounded
and
both $f_{\varepsilon}(\cdot|x_i)$ and $\frac{d}{de}f_{\varepsilon}(\cdot|x_i)$ are uniformly bounded.
Under a different setting (with smoothed and globally convex loss functions), \citet{lee2021fast} established a functional central limit theorem (FCLT) for the aggregated solution path of SGD. We
extend our previous result to the quantile regression model. Under Assumption~\ref{asm:qr}, we establish the following \emph{functional} central limit theorem (FCLT):
\begin{equation}\label{eq4}
\frac{1}{\sqrt{n}}\sum_{i=1}^{\left[nr\right]}\left(\beta_{i}-\beta^{*}\right)\Rightarrow \Upsilon^{1/2}W\left(r\right),\quad r\in\left[0,1\right],
\end{equation}
where $\Rightarrow$ stands for the weak convergence in $\ell^{\infty}\left[0,1\right]$,
$W\left(r\right)$ stands for a vector of the independent standard
Wiener processes on $\left[0,1\right]$,
and
$ \Upsilon := H^{-1}SH^{-1}$, with
$
S := \mathbb E [x_i x_i'] \tau(1-\tau)
$ and $H := \mathbb E [x_i x_i' f_{\varepsilon}(0|x_i)]$.
The FCLT in \eqref{eq4} states that the partial sum of the recursively updated estimates $\beta_i$ converges weakly to a rescaled Wiener process, with the scaling matrix equal to a square root of the asymptotic variance of the usual quantile regression estimator $\widehat{\beta}_n$.
Building on the FCLT, we propose
a large-scale inference procedure.
More generally, for any $\ell \leq d$ linear restrictions
\[
H_{0}: R\beta^{*} = c,
\]
where $R$ is an $(\ell \times d)$-dimensional known matrix of rank $\ell$
and
$c$ is an $\ell$-dimensional known vector,
the conventional Wald test based
on $\widehat{V}_{n}$ becomes asymptotically pivotal.
We formally state the main theoretical results in the following theorem.
\begin{thm}[Main Theorem]
\label{thm:Wald:qr} Suppose that
$H_{0}: R\beta^{*} = c$ holds with
$\mathrm{rank}(R)=\ell$. Under Assumption~\ref{asm:qr},
the FCLT in (\ref{eq4}) holds and
\begin{eqnarray*}
&&n\left(R\bar{\beta}_{n}-c\right)'\left(R\widehat{V}_{n}R'\right)^{-1}\left(R\bar{\beta}_{n}-c\right)\overset{d}{\to}W\left(1\right)'\left(\int_{0}^{1}\bar{W}(r)\bar{W}(r)'dr\right)^{-1}W\left(1\right),
\end{eqnarray*}
where $W$ is an $\ell$-dimensional vector of the standard Wiener
processes and $\bar{W}\left(r\right):=W\left(r\right)-rW\left(1\right)$.
\end{thm}
As an important special case of Theorem \ref{thm:Wald:qr}, the t-statistic defined in \eqref{t-stat} converges in distribution to the following pivotal limiting distribution:
for each $j = 1,\ldots,d$,
\begin{align}\label{t-stat-limit}
\frac{\sqrt{n}\left(\bar{\beta}_{n,j}-\beta_{j}^{*}\right)}{\sqrt{\widehat{V}_{n,jj}}}
\overset{d}{\to}
W_1\left(1\right) \left[ \int_{0}^{1} \left\{ W_1\left(r\right)-rW_1\left(1\right) \right\}^2 dr\right]^{-1/2},
\end{align}
where
$W_1$ is a one-dimensional standard Wiener process.
\begin{comment}
\subsection{The cluster-dependent setting}
We consider the ``large $N$, small $T_i$'' asymptotic scenario, so there are many clusters each with finite cluster size. As for the dependence assumption, we assume observations are independent across clusters but allow arbitrary dependence within clusters.
\begin{asm}\label{asm:qrpanel}
Suppose that $\{ Y_{i,t} \equiv (y_{i,t}, x_{i,t}) \}$ are independent and identically distributed across $i$, and generated from
$$
y_{i,t}= x_{i,t}'\beta^*_\tau + \varepsilon_{i,t},
$$
where $P(\varepsilon_{i,t}\leq 0|x_{i,t})= \tau$ for a fixed quantile $\tau \in(0,1)$. Let $Q(\beta):= \mathbb E [q(\beta, Y_{i,t})]$ for all $i$ and $t$ and $f_{\varepsilon}(\cdot|x_{i,t})$ denote the conditional density of $\varepsilon_{i,t} $ given $x_{i,t}$,
and its partial derivative
$\frac{d}{de}f_{\varepsilon}(\cdot|x_{i,t})$ exist.
Furthermore, assume that
\begin{enumerate}
\item
there exist positive constants $\epsilon$ and $c_0$ such that
$$
\inf_{ | \beta-\beta^* |<\epsilon }\lambda_{\min} \left( \mathbb E [x_{i,t}x_{i,t}'f_{\varepsilon}(x_{i,t}'(\beta-\beta^*)|x_{i,t})] \right) > c_0,
$$
\item $\sup_b\mathbb E [ \|x_{i,t}\|^3A(b, x_{i,t}) ] <C$ for some constant $C < \infty$, where $ A(b, x_{i,t}) := \left|\frac{d}{de}f_{\varepsilon}(x_{i,t}'b|x_{i,t})\right| + f_{\varepsilon}(x_{i,t}'b|x_{i,t}),$
\item $\mathbb E[ (\| T^{-1} \sum_{t=1}^{T} x_{i,t}\|^6+1)\exp(\|T^{-1} \sum_{t=1}^{T}x_{i,t}\|^2) ]<C$ for some constant $C < \infty$,
\end{enumerate}
\end{asm}
\begin{thm}[Main Theorem]
\label{thm:Waldpanel} Suppose that
$H_{0}: R\beta^{*} = c$ holds with
$\mathrm{rank}(R)=\ell$. Under Assumption~\ref{asm:qrpanel},
\begin{eqnarray*}
&&n\left(R\bar{\beta}_{n}-c\right)'\left(R\widehat{V}_{n}R'\right)^{-1}\left(R\bar{\beta}_{n}-c\right)\overset{d}{\to}W\left(1\right)'\left(\int_{0}^{1}\bar{W}(r)\bar{W}(r)'dr\right)^{-1}W\left(1\right),
\end{eqnarray*}
where $W$ and $\bar{W}\left(r\right) $ as as the same as in Theorem \ref{thm:Wald:qr}.
\end{thm}
\end{comment}
\section{Monte Carlo Experiments}\label{sec:MC}
In this section we investigate the performance of the S-subGD random scaling method via Monte Carlo experiments.
The main question of these numerical experiments is whether the proposed method is feasible for a large scale model.
The simulation is based on the following data generating process:
\begin{align*}
y_{i} = x_i'\beta + \varepsilon_i~~\mbox{for}~~i=1,\ldots,n,
\end{align*}
where $x_i$ is a $(d+1)$-dimensional covariate vector whose first element is 1 and the remaining $d$ elements are generated from $\mathcal{N} (0,I_d)$. The error term $\varepsilon_i$ is generated from $\mathcal N(0,1)$, and the parameter value $\beta$ is set to be $(1,\ldots,1)$.
Without loss of generality, we focus on the second element of $\beta$.
To accommodate a large scale model, the sample size varies in $n \in \{10^5, 10^6, 10^7\}$, and the dimension of $x$ varies in $d\in \{10, 20, 40, 80, 160, 320, 1000\}$.
The initial value $\beta_0$ is estimated by the convolution-type smoothed quantile regression (\texttt{conquer}) in \citet{HE2021} and we do not burn in any observations.
The learning rate is set to be $\gamma=\gamma_0 t^{-a}$ with $\gamma_0=1$ and $a=0.501$.
The simulation results are summarized from $1000$ replications of each design.
We compare the performance of the proposed method with four additional alternatives:
\begin{description}
\item[(i)] S-subGD: the proposed S-subGD random scaling method.
\item[(ii)] QR: the classical quantile regression method.
Given the scale of models, we apply the Frisch-Newton algorithm after preprocessing and the Huber sandwich estimate under the local linearity assumption of the conditional density function by selecting `pfn' and `nid' options in R package \texttt{quantreg} (CRAN version 5.88).
\item[(iii)] CONQUER-plugin: estimates the parameter using the conquer method, and estimates the asymptotic variance by plugging in the parameter estimates; implemented using the R package \texttt{conquer} (CRAN version 1.3.0).
\item[(iv)] CONQUER-bootstrap: estimates the parameter using the conquer method, and applies the multiplier bootstrap method (see \citet{HE2021} for details); implemented using the R package \texttt{conquer}. We set the number of bootstrap samples as 1000.
\item[(v)] SGD-bootstrap: estimates using the S-subGD method and conduct the inference by an online bootstrap method (see \citet{fang2018online} for details). We set the number of bootstrap samples as 1000.
\end{description}
\begin{figure}[!htbp]
\caption{Computation time} \label{fig:time}
\centering
\vskip10pt
\includegraphics[width=\textwidth]{figure/fig_time.png}
\includegraphics[width=\textwidth]{figure/fig_rel_time.png}
\end{figure}
\begin{figure}[!htbp]
\caption{Coverage and CI Length} \label{fig:ci}
\centering
\vskip10pt
\includegraphics[width=\textwidth]{figure/fig_coverage.png}
\includegraphics[width=\textwidth]{figure/fig_ci_length.png}
\end{figure}
We use the following performance measures: the computation time, the coverage rate, and the length of the 95\% confidence interval.
Note that the nominal coverage probability is 0.95.
The computation time is measured by 10 replications on a separate desktop computer, equipped with an AMD Ryzen Threadripper 3970X CPU (3.7GHz) and a 192GB RAM.
Other measures are computed by 1000 replications on the cluster system composed of several types of CPUs.
We set the time budget as 10 hours for a single estimation and the RAM budget as 192GB.
Figures \ref{fig:time}--\ref{fig:ci} summarize the simulation result. To save space, we report the results of $d=20, 80,$ and $320$. We provide tables for all the simulation designs in the appendix.
First, we observe that S-subGD is easily scalable to a large simulation design. On the contrary, some methods cannot complete the computation within the time/memory budget. For instance, CONQUER-plugin does not work for any $d$ when $n=10^6$ or $n=10^7$ because of the memory budget. QR shows a similar issue when $n=10^7$ and $d=320$. CONQUER-bootstrap cannot meet the 10-hour time budget for a single replication when $n=10^7$ and $d=320$. They are all denoted as `NA' in the figures.
Second, S-subGD outperforms the alternative in terms of computational efficiency. The bottom panel of Figure \ref{fig:time} shows the relative computation time, and the efficiency scales are 10's or 100's levels. In case of CONQUER-plugin, it is slower than S-subGD in the scale of 1000's. If we convert them into the actual computation time, S-subGD takes 16 seconds for the design of $n=10^6$ and $d=320$. However, QR, CONQUER-bootstrap, and SGD-bootstrap take 1374 seconds (22 minutes 54 seconds), 4113 seconds (1 hour 8 minutes 33 seconds), and 1643 seconds (27 minutes 23 seconds) on average of 10 replications.
Third, all methods shows satisfactory results in terms of the coverage rate and the length of the confidence interval. The top panel of Figure \ref{fig:ci} reports that coverage rates are all around 0.95 although S-subGD and SGD-bootstrap are slightly under-reject and over-reject when $d=320$. The average lengths of the confidence interval are reported on the bottom panel. The performance of S-subGD is comparable to QR, CONQUER-plugin, and CONQUER-bootstrap. SGD-bootstrap shows much larger CI lengths, especially when $d=320$.
Finally, we stretch the computational burden to $d=1000$ and $n=10^7$ and check the performance of S-subGD. Table \ref{tb_SGD_rs} reports the simulation result along with different sizes of $d$. S-subGD still completes the computation within a reasonable time range. The average computation time is under 762 seconds (12 minutes 42 seconds). The coverage rate and CI length becomes slightly worse than those of smaller $d$'s but they still fall in a satisfactory range.
In sum, S-subGD performs well when we conduct a large scale inference problem in quantile regression. It is scalable to the range where the existing alternatives cannot complete the computation within a time/memory budget. In addition to computational efficiency, the coverage rate and CI lengths are also satisfactory. Thus, S-subGD provides a solution to those who hesitate to apply quantile regression with a large scale data for the computational burden.
\begin{table}[htb]
\centering
\caption{Performance of S-subGD: $n=10^7$}
\label{tb_SGD_rs}
\begin{tabular}{c S[table-format=3.2]S[table-format=1.3]S[table-format=1.4]}
\hline
{$d$} & {Time (sec.)} & {Coverage Rate} & {CI Length} \\
\hline
10 & 5.87 & 0.965 & 0.0020 \\
20 & 11.05 & 0.955 & 0.0020 \\
40 & 21.86 & 0.954 & 0.0020 \\
80 & 43.12 & 0.952 & 0.0020 \\
160 & 81.35 & 0.953 & 0.0021 \\
320 & 166.40 & 0.963 & 0.0011 \\
1000 & 762.16 & 0.925 & 0.0461 \\
\hline
\end{tabular}
\end{table}
\section{Inference for the College Wage Premium}
The study of the gender gap in the college wage premium has been a long-standing important question in labor economics. The literature has pointed out a stylized fact that the higher college wage premium for women as the major cause for attracting more women to attend and graduate from colleges than men (e.g., \cite{goldin2006homecoming, chiappori2009investment}). Meanwhile, \citet{hubbard2011phantom} pointed out a bias issue associated with the ``topcoded'' wage data, where the use of the sample from Current Population Survey (CPS) often censors the wage data at a maximum value. As a remedy for it, \citet{hubbard2011phantom} proposed to use quantile regression that are robust to censoring. Analyzing the CPS data during 1970-2008, he found no gender difference in the college wage premium later years once the topcoded bias has been accounted for.
We revisit this problem by estimating and comparing the college wage premium between women and men. While the data also contains the topcoded issue, the quantile regression is less sensitive to the values of upper-tail wages. By applying the new S-subGD inference to the ultra-big dataset we are using, we aim at achieving the following goals in the empirical study: (1) identify (if any) the heterogeneous effects across quantiles; (2) understand the trends in the college wage premium respectively for female and male; and (3) understand the gender difference in the college wage premium.
We use the data from IPUMS USA that contains several millions of workers. The main motivation of using ultra-large datasets for wage regressions is to deal with high-dimensional regressors within a simple parametric quantile regression model. As has been crucially pointed out in the literature, work experience is
an important factor in wage regressions, but is difficult to be measured precisely. One typical means of controlling for the experience is to flexibly interact workers' age with various state-level dummies, which would create over one thousand regressors, so that a very large sample size would be desirable to obtain precise estimates.
But the ultra-large dataset makes most existing inference procedures for quantile regression fail to work.
\subsection{The data}
We use the samples over six different years (1980, 1990, 2000-2015) from IPUMS USA at \url{https://usa.ipums.org/usa/}.
In the years from 1980 to 2000, we use the 5\% State sample which is a 1-in-20 national random sample of the population.
In the remaining years, we use the American Community Survey (ACS) each year. The sampling ratio varies from 1-in-261 to 1-in-232 in 2001-2004, but it is set to a 1-in-100 national random sample of the population after 2005.
To balance the sample size, we bunch the sample every 5 year after 2001.
We also provide the estimation results using separate but smaller samples in 2005, 2010, and 2015 in the appendix, which is similar to those reported in this section.
We restrict our sample to $White$, $18 \le Age \le 65$, and $Wage \ge \$62$, which is a half of minimum wage earnings in 1980 ($\$3.10 \times 40 \mbox{hours} \times 1/2$).
$Wage$ denotes the implied weekly wage that is computed by dividing yearly earnings by weeks worked last year.
Since 2006, weeks worked last year are available only as an interval value and we use the midpoint of an interval.
We only consider full-time workers who worked more than 30 hours per week.
Then, we compute the \textit{real} wage using the personal consumption expenditures price index (PCEPI) normalized in 1980.
The data cleaning leaves us with 3.6-4.7 million observations besides 2001-2005, where we have around 2.5 million observations.
Table \ref{tb:summary_stat} reports some summary statistics on the key variables, where $Educ$ denotes an education dummy for some college or above.
The table confirms that female college education has increased substantially over the years and female workers received more college education than male workers since 1990.
\begin{table}[ht]
\caption{Summary Statistics} \label{tb:summary_stat}
\centering
\begin{tabular}{crcccc}
\hline
Year & Sample Size & $\mathbb E(Female$) & $\mathbb E(Educ$) & $\mathbb E(Educ$|$Male$) & $\mathbb E(Educ$|$Female$) \\
\hline
1980 & 3,659,684 & 0.390 & 0.433 & 0.444 & 0.416 \\
1990 & 4,192,119 & 0.425 & 0.543 & 0.537 & 0.550 \\
2000 & 4,479,724 & 0.439 & 0.600 & 0.578 & 0.629 \\
2001-2005 & 2,493,787 & 0.447 & 0.642 & 0.619 & 0.670 \\
2006-2010 & 4,708,119 & 0.447 & 0.663 & 0.631 & 0.701 \\
2011-2015 & 4,542,874 & 0.447 & 0.686 & 0.646 & 0.735 \\
\hline
\end{tabular}
\end{table}
\subsection{The quantile regression model}
We use the following baseline model:
\begin{footnotesize}
\begin{align*}
\log(Wage_i) = \beta_0 + \beta_1 Female_i + \beta_2 Educ_i + \beta_3 Female_i\cdot Educ_i + \theta_1'X_i + \theta_2' (X_i\cdot Female_i) + \varepsilon_i,
\end{align*}
\end{footnotesize}where $Wage_i$ is a real weekly wage in 1980 terms, $Female_i$ is a female dummy, $Educ_i$ is an education dummy for some college or above, and $X_i$ is a vector of additional control variables.
For control variable $X_i$, we use 12 age group dummies with a four-year interval, 51 states dummies (including D.C.), and their interactions. Note that $(X_i\cdot Female_i)$ implies that there exist up to 3-way interactions.
The model contains 1226 covariates in total.
We also add 4 additional year dummies for the 5-year combined samples after 2001.
We estimate the model using the proposed S-subGD method over 9 different quantiles: $\tau=0.1,0.2,\ldots,0.9$.
We obtain the starting value of S-subGD Inference by applying CONQUER to the 10\% subsample.\footnote{Recall that $X_i$ contains a large number of dummy variables in this empirical application. If the subsample size is too small, CONQUER may face a singularity issue and cannot compute the initial value. Alternatively, one may use a randomzied algorithm of \citet{Yang:QR:ICML}.
}
We do not burn in any sample and set the learning rate as $\gamma_0 t^{-a}$ with $\gamma_0=1$ and $a=0.501$.
We use the Compute Canada cluster system equipped with the AMD Milan CPUs whose clock speed is 2.65 GHz and with 650 GB of RAMs.
For comparison, we also try to estimate the model using the standard QR method with the `quantreg' R package, but it fails to compute the result in the given cluster environments.
\subsection{The results}
Figures \ref{fig:premium_design_5yr_full}--\ref{fig:diff_5yr_full} and Table \ref{tb:median_5yr_full} summarize the estimation results.
We also provide the full estimation results in the appendix.
\begin{figure}[!htbp]
\centering
\caption{College Wage Premium:
Combining 5-Year Data
}
\label{fig:premium_design_5yr_full}
\hskip15pt
\begin{tabular}{c c c}
\includegraphics[scale=0.23]{figure/chart_0.1_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.2_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.3_03_FullT_5yr.png} \\
\includegraphics[scale=0.23]{figure/chart_0.4_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.5_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.6_03_FullT_5yr.png} \\
\includegraphics[scale=0.23]{figure/chart_0.7_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.8_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/chart_0.9_03_FullT_5yr.png}
\end{tabular}
\end{figure}
\begin{figure}
\centering
\caption{Difference of College Premium: Combining 5-Year Data}
\label{fig:diff_5yr_full}
\hskip15pt
\begin{tabular}{c c c}
\includegraphics[scale=0.23]{figure/diff_chart_0.1_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.2_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.3_03_FullT_5yr.png} \\
\includegraphics[scale=0.23]{figure/diff_chart_0.4_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.5_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.6_03_FullT_5yr.png} \\
\includegraphics[scale=0.23]{figure/diff_chart_0.7_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.8_03_FullT_5yr.png} & \includegraphics[scale=0.23]{figure/diff_chart_0.9_03_FullT_5yr.png}
\end{tabular}
\end{figure}
Our results share several interesting features about the estimated college wage premium. First, some degrees of heterogeneity are found over different quantiles. At the upper tail quantile, $\tau=0.9$, male college premium is slightly higher than the female one since 2000, while female college premiums are higher over all years in other quantiles. This shows an interesting feature about high-income individuals that cannot be viewed by the mean regression, though sensible economic interpretations might be subject to debates. Also, note that the 95\% confidence intervals are wider for tail quantiles ($\tau=0.1$ and 0.9, respectively).
Second, college premiums increase over time for both male and female. This result coincides with the overall findings in the literature. It is interesting that college premiums get flatter since 2010 for lower quantiles ($\tau \le 0.5$).
Third, when we focus on the median ($\tau=0.5$) as reported in Table \ref{tb:median_5yr_full}, we observe that the female-male college premium difference shows an inverse ``U-shape" pattern.
In addition, the difference is always significantly positive, which is different from the result in \citet{hubbard2011phantom}.
He found the gender difference to be insignificant starting from around the year 2000 from the CPS data.
We also note that the computation time is within a reasonable range meanwhile it is not feasible to estimate the model with the alternative approaches.
Finally, the use of ultra-big dataset and the S-subGD inference in quantile regression provide several features for the college premium analysis. Although the S-subGD inference tends to be conservative than the standard normal approximation at the 95\% level, it reveals statistically significant results while controlling over $10^3$ covariates to mitigate confounding effects. With more availability of such a large dataset, the S-subGD method will make it possible to obtain convincing empirical evidence for other analyses.
\begin{table}[htb]
\centering
\caption{College Wage Premium: $\tau=0.5$}\label{tb:median_5yr_full}
\begin{tabular}{ccccc}
\hline
Year & Female & Male & Difference & Time (min.) \\
\hline
\underline{$\tau = 0.5$}\\
1980 & 0.2365 & 0.1988 & 0.0377 & 29.4 \\
& [0.2294,0.2435] & [0.1945,0.2030] & [0.0291,0.0463] & \\
1990 & 0.3667 & 0.2962 & 0.0705 & 34.2 \\
& [0.3603,0.3732] & [0.2942,0.2982] & [0.0634,0.0777] & \\
2000 & 0.4101 & 0.3439 & 0.0662 & 36.7 \\
& [0.4056,0.4146] & [0.3372,0.3506] & [0.0552,0.0772] & \\
2001-2005 & 0.4468 & 0.3854 & 0.0613 & 20.2 \\
& [0.4369,0.4567] & [0.3765,0.3944] & [0.0554,0.0673] & \\
2006-2010 & 0.4791 & 0.4271 & 0.0520 & 47.7 \\
& [0.4748,0.4834] & [0.4174,0.4368] & [0.0454,0.0585] & \\
2011-2015 & 0.4957 & 0.4498 & 0.0458 & 46.0 \\
& [0.4887,0.5027] & [0.4455,0.4542] & [0.0348,0.0568] & \\
\hline
\end{tabular}
\flushleft{\footnotesize Notes. The male college premium is from $\hat{\beta}_2$ and the female college premium is from $\hat{\beta}_2+\hat{\beta}_3$. Thus, the college premium difference between male and female workers is from $\hat{\beta}_3$. }
\end{table}
\section{Conclusions}
We have proposed an inference method for large-scale quantile regression to analyze datasets whose sizes are of order $(n, p) \sim (10^7, 10^3)$. Our method runs very fast, based on the stochastic sub-gradient descent updates, and constructs asymptotically pivotal statistics via random scaling.
There are a couple of extensions worth pursuing in future research.
First, we may build on \citet{chernozhukov2022fast} to develop fast inference for the quantile regression process.
Second, while we focus on the regular quantile regression models where the quantile is assumed to be bounded away from both zero and one, our framework is also potentially applicable to extreme quantile regression models \citep[see, e.g.,][for a review]{chernozhukov2016extremal}. The ability of handling ultra-large datasets is also appealing for extreme quantile regression models because the resulting sample size is considerably larger at the extreme quantiles. A formal treatment for this case is out of the scope of this paper, and we leave it for future studies.
| 3bb5a7da821e561113f344e28afcd5045cf42b13 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec:1}
Generally, compact stars are the final stage of the stellar evolution, that could be formed because of the radial pressure from core nuclear fusion that is bigger that gravitational forces. Gravitational compact objects include many various objects, such as white dwarfs, neutron stars, black holes and naked singularities. Moreover, compact star could also include more exotic objects, such as strange stars (made of strange quarks), gravitational condensate stars (non-singular three-layer alternative to the black hole) etc. Such compact objects usually were studied in the General Theory of Relativity (further - GR), but there are also present a couple of works in the modified gravity formalism (for example, in the $f(\mathcal{R})$ gravity \cite{Pretel:2022plg,Jimenez:2021wik,Numajiri:2021nsc,Sharif:2021emv,Abbas:2021bid}, in the $f(\mathcal{R},\mathcal{T})$ gravity \cite{Bhar:2021uqr,Ahmed:2021fav,Kumar:2021vqa,Bhar:2021iog,Pretel:2021kgl} and finally in the teleparallel modified gravity \cite{deAraujo:2021zma,Solanki:2021fzo} and references therein). Gravity modification could solve many problems of GR, namely dark energy problem, inflation and late time accelerated expansion. Apart from presenting new geometrodynamical terms in the Einstein-Hilbert (further - EH) action integral, we could also introduce additional matter fields to solve that problems. In the current paper we modified GR Lagrangian by using Bose-Einstein Condensate, Kalb-Ramond and $U(1)$ gauge fields.
\subsection{Bose-Einstein Condensate as Dark Matter / Dark Energy model}
As a first additional matter field of our consideration we have chosen the case with Bose-Einstein condensate. Bose-Einstein condensate is created from the set of Bose gas particles in the same ground state at the very low temperatures near the absolute zero. It is known that Bose-Einstein condensates could be formed from the Cold Dark Matter (CDM) axions by thermalisation process \cite{PhysRevLett.103.111301}. Such BEC dark matter could viably recreate galaxy rotational curves, which was showed in the work of \cite{Craciun2020} on the example of SPARC galactic rotation curves data. Finally, BE condensates could not only act as DM, but also as dark energy, which was shown in the following study \cite{Das_2015}. As well, it was reported by \cite{Fukuyama:2009vzr} that Bose-Einstein condensation phase of the boson field could be present in the early universe, and later that could lead to the formation of dark matter - dark energy unification. It also worth to notice that in the aforementioned model matter density fluctuations are located in the allowed by observational data bounds.
Finally, there was some work done in the field of Bose-Einstein condensate stars, generally adopting hydrodynamical representation for BEC wave function and using polytropic Equation of State. For example, BEC stars were investigated in the works of \cite{Mukherjee:2014kqa,Danila:2015qla,Madarassy:2014jfa}.
In the current paper we will use an assumption of zero temperature BEC, namely pure BEC and an assumption of small repulsion strength. Only with these assumptions Gross–Pitaevskii equation could properly describe wave function of the Bose-Einstein condensate (for more details, see phase transition diagram at the Figure (\ref{fig:222})). On the phase transition diagram, there is a couple of phases of Boson fluids present, related to the different values of temperature $T$ and repulsion strength. It could be noticed that fluid with $T\gg0$ and relatively small repulsion strength has superfluid behavior (i.e. vanishing viscosity). However, for smaller values of temperature (assuming that the fluid is in thermodynamical equilibrium), Bose-Einstein Condensate could be formed (pure BEC appears at $T=0$). On the other hand, if we assume big values of repulsion strength, fluid becomes insulator.
\begin{figure}[!htbp]
\centering
\begin{tikzpicture}[scale=0.4]
\message{Phase diagrams 2^^J}
\def\xtick#1#2{\draw[thick] (#1)++(0,.2) --++ (0,-.4) node[below=-.5pt,scale=0.7] {#2};}
\def\ytick#1#2{\draw[thick] (#1)++(.2,0) --++ (-.4,0) node[left=-.5pt,scale=0.7] {#2};}
\coordinate (Z) at (0,0);
\coordinate (N1) at (0,7);
\coordinate (N2) at (11,10);
\coordinate (N3) at (0,4.5);
\coordinate (NE) at (12,10);
\coordinate (NW) at (0,10);
\coordinate (SE) at (12,0);
\coordinate (W) at (0,5);
\coordinate (S) at (6,0);
\coordinate (C) at (12,7);
\coordinate (T) at (6,0);
\def(T) to[out=100,in=20] (N1){(T) to[out=100,in=20] (N1)}
\def(T) to[out=100,in=20] (N3){(T) to[out=100,in=20] (N3)}
\def(T) to[out=100,in=160] (C){(T) to[out=100,in=160] (C)}
\def(0,5.5) -- (12,5.5){(0,5.5) -- (12,5.5)}
\path[name path=SLL] (T) to[out=100,in=20] (N3);
\path[name path=SL] (T) to[out=100,in=20] (N1);
\path[name path=LG] (T) to[out=100,in=160] (C);
\path[name path=atm] (0,5.5) -- (12,5.5);
\fill[mylightblue] (T) to[out=100,in=20] (N1) -- (N1) -- (Z) -- cycle;
\fill[blue!5] (T) to[out=100,in=20] (N3) -- (N3) -- (Z) -- cycle;
\fill[mylightred] (T) to[out=100,in=160] (C) -- (C) -- (SE) -- cycle;
\node at (1.5,3) {BEC};
\node at (1.5,5.5) {SF};
\node at (9,3) {Insulator};
\fill (T) circle (4pt);
\draw[thick] (T) to[out=100,in=160] (C);
\draw[thick] (T) to[out=100,in=20] (N1);
\draw[thick,dashed] (T) to[out=100,in=20] (N3);
\draw[thick] (Z) rectangle (NE);
\node[left=3pt,above,rotate=90] at (W) {$T$ $[K]$};
\node[below=3pt] at (S) {Repulsion strengh};
\end{tikzpicture}
\caption{Phase transition diagram for Boson fluids at small temperature}
\label{fig:222}
\end{figure}
\subsection{Kalb-Ramond field and the rise of bouncing cosmology}
Second additional matter field is fully antisymmetric rank-2 tensor field, namely Kalb-Ramond (or B) field. Such fields as Kalb-Ramond one is essential in reproduction of the low energy effective string actions \cite{PhysRevLett.89.121101}. Moreover, massless KR field could occur in the critical and massless string spectrum due to the compactification to four dimensions \cite{Schwarz:2000ew}. Additionally, this field could be the source of spontaneous Lorentz Symmetry Breaking (LSB), if self-interaction potential and non-vanishing vacuum expectation value (VEV) are assumed \cite{PhysRevD.81.065028}. In the pioneering work \cite{Hell_2022} it was shown that KR field could be dual to the massive spin-1 field, namely Proca field. This property could be used in the study of exotic Proca, Boson-Proca stars.
It is also worth to notice that KR field could induce cosmological bounce, which was shown by \cite{nair2021kalbramond} in the modified teleparallel cosmology and that black holes with KR VEV background behave like Reissner-N\" ordstrom black holes, despite the charge is absent \cite{Lessa:2019bgi}.
\subsection{Gauge field background}
Gauge field imposing $U(1)$ local gauge symmetry (namely photon gauge field) is the third and the last exotic matter field that we consider in our study. Stars surrounded by such gauge fields are called gauged boson stars and now these kind of stars and their shells are widely studied in the literature \cite{Kunz2021mbm,Herdeiro:2021jgc,Liu:2020uaz,SalazarLandea:2016bys}. In this articles numerical evidence for the existence of gauged boson stars in asymptotically flat/asymptotically AdS spacetimes was given and it was reported that such solutions similarly to the Kerr-Newmann black holes preserve nonzero electric charge and magnetic dipole momentum \cite{Kichakova2013}. As well, asymptotically AdS boson stars play important role in the AdS/CFT (Conformal Field Theory) correspondence.
\subsection{Article Organisation}
Our article is organised as follows: in the Section (\ref{sec:1}) we provide a general introduction into the topic of compact (relativistic) objects and their kind, into the gravity modification and problems of GR. As well, we discuss each exotic matter field of our consideration in the separated subsections. On the other hand, in the Section (\ref{sec:22}) we derive field equations for perfect fluid stress-energy tensor in terms of Einsteinian gravity and present the formalism that is being used in our study, introduce viable Finch-Skea metric potentials. Further, in the next Section (\ref{sec:2}) we probe the Finch-Skea compact stars in the presence of pure Bose-Einstein condensate. As well, in the Section (\ref{sec:3}) we investigate our Kalb-Ramond spherically symmetric stellar solutions, in the Section (\ref{sec:4}) case with minimally coupled to gravity gauge field imposing $U(1)$ local gauge symmetry. As a final note, in the Section (\ref{sec:5}) we provide concluding remarks on the main topics of our study.
\section{Finch-Skea stars and Einstein gravity} \label{sec:22}
For the regular GR gravity, Einstein-Hilbert action integral reads:
\begin{equation}
\mathcal{S}[g,\Gamma,\Psi_i]=\int_\mathcal{M}d^4x\sqrt{-g}\frac{1}{2\kappa^2}(\mathcal{R}+\mathcal{L}(\Psi_i))
\label{eq:1}
\end{equation}
where $\mathcal{R}$ is common Ricci scalar curvature, $g=\det g_{\mu\nu}=\prod^{3}_{\mu,\nu=0}g_{\mu\nu}$ is the determinant of metric tensor and $\kappa$ is the well-known Einstein gravitational constant. Finally, we define $\mathcal{L}(\Psi_i)$ as a Lagrangian density for additional matter fields $\Psi_i$, $\Gamma$ is torsionless metric-affine connection. Varying the aforementioned EH action with respect to the metric tensor inversion $g^{\mu\nu}$, we could get the set of Einstein Field Equations (further - EFE's):
\begin{equation}
G_{\mu\nu}=\kappa T_{\mu\nu}
\end{equation}
for the case with flat background spacetime. In the equation above $T_{\mu\nu}$ is defined as stress-energy tensor
\begin{equation}
T_{\mu\nu} = -\frac{2}{\sqrt{-g}}\frac{\delta(\sqrt{-g}\mathcal{L}(\Psi_i))}{\delta g^{\mu\nu}}
\label{eq:3}
\end{equation}
In our paper stress-energy-momentum tensor is assumed to be anisotropic and perfect fluid one, and therefore
\begin{equation}
T_{\mu}^{\nu} = (\rho+p_t)U_\mu U^\nu - p_t\delta^\nu_\mu + (p_r-p_t)V_\mu V^\nu
\end{equation}
Here, as usual $\rho$ is energy density and $p_r$, $p_t$ are radial, tangential pressures respectively, $U_\mu$ is normalised by $U^\mu U_\mu=-1$ timelike four-velocity and $V_\mu$ is spacelike (radial) four-vector.
Let us consider spherically symmetric line element with metric signature $(+;-;-;-)$:
\begin{equation}
ds^2 =e^{\nu(r)}dt^2-e^{\lambda(r)}dr^2-r^2d\Omega^2_{D-2}
\end{equation}
where $e^{\nu(r)}$ and $e^{\lambda(r)}$ are metric potentials and $d\Omega^2_{D-2}$ is the $D-2$ dimensional unit sphere line element:
\begin{equation}
\begin{gathered}
d\Omega^2_{D-2}=d\theta_1^2+\sin^2\theta_1d\theta_2^2+\sin^2\theta_1\sin^2\theta_2d\theta_3^2+...+\bigg(\prod^{D-3}_{j=1}\sin^2\theta_j\bigg)d\theta^2_{D-2}
\end{gathered}
\end{equation}
In the further investigation we will restrict our analysis to the $D=(3+1)$ dimensions. For that case, stress-energy-momentum tensor components are \cite{Bhar2017}
\begin{equation}
\kappa \rho=\frac{1-e^{-\lambda(r)}}{r^2}+\frac{e^{-\lambda(r)}\lambda'(r)}{r}
\end{equation}
\begin{equation}
\kappa p_r=\frac{e^{-\lambda(r)}-1}{r^2}+\frac{e^{-\lambda(r)}\nu'(r)}{r}
\end{equation}
\begin{equation}
\kappa p_t = e^{-\lambda}\left(\frac{\nu''(r)}{2}+\frac{\nu'(r)^2}{4}-\frac{\nu'(r)\lambda'(r)}{4}+\frac{\nu'(r)-\lambda'(r)}{2r}\right)
\end{equation}
\subsection{Finch-Skea symmetry}
Throughout the paper we will consider that our compact star solution impose Finch-Skea symmetry and have physically viable, non-singular metric potentials of form \cite{Finch_1989}:
\begin{equation}
e^{\nu(r)}=\left(A+\frac{1}{2}Br\sqrt{r^2C}\right)^2
\end{equation}
\begin{equation}
e^{\lambda(r)}=\left(1+Cr^2\right)
\end{equation}
where $A$, $B$ and $C$ have constant values and are called Finch-Skea coefficients. Using Finch-Skea ansatz above, we could rewrite field equations:
\begin{equation}
\kappa \rho=\frac{C \left(C r^2+3\right)}{\left(C r^2+1\right)^2}
\end{equation}
\begin{equation}
\kappa p_r=-\frac{C \left(2 A \sqrt{C r^2}+B r \left(C r^2-4\right)\right)}{\left(C r^2+1\right)
\left(2 A \sqrt{C r^2}+B C r^3\right)}
\end{equation}
\begin{equation}
\kappa p_t = \frac{C \left(B r \left(C r^2+4\right)-2 A \sqrt{C r^2}\right)}{\left(C r^2+1\right)^2
\left(2 A \sqrt{C r^2}+B C r^3\right)}
\end{equation}
In the following subsection we are going to derive exact solutions for Finch-Skea coefficients using junction conditions at the boundary.
\subsection{Junction conditions}
Extra matching conditions for spherically symmetric relativistic objects were provided by Goswami et al. \cite{PhysRevD.90.084011}, and it was shown that constraints on the thermodynamic properties and stellar structure are purely mathematical \cite{Pandya2021}. For spacetime with axymptotically flat background, exterior region could be considered as a Schwarzschild’s spacetime:
\begin{equation}
ds^2 = \bigg(1-\frac{2M}{r}\bigg)dt^2 - \bigg(1-\frac{2M}{r}\bigg)^{-1}dr^2 - r^2 d\theta^2 - r^2 \sin^2 \theta d\phi^2
\end{equation}
where $M$ indicates total gravastar mass. The above line element imply conditions of metric potential continuity at the boundary:
\begin{equation}
\mathrm{Continuity\;of\;}g_{tt}:\quad 1-\frac{2M}{R}=e^{\nu(r)}
\end{equation}
\begin{equation}
\mathrm{Continuity\;of\;}\frac{\partial g_{tt}}{\partial r}:\quad \frac{2M}{R^2}=-B \left(2 A \sqrt{C R^2}+B C R^3\right)
\end{equation}
\begin{equation}
\mathrm{Continuity\;of\;}g_{rr}:\quad \bigg(1-\frac{2M}{R}\bigg)^{-1}=e^{\lambda(r)}
\end{equation}
\begin{equation}
\mathrm{Vacuum\;condition}:\quad p\rvert_{r=R}=0
\end{equation}
The solutions for above continuity conditions are:
\begin{equation}
A=\frac{3 M-2 R}{2 \sqrt{R} \sqrt{R-2 M}}
\end{equation}
\begin{equation}
B=\frac{\sqrt{\frac{M}{R-2 M}} \sqrt{R-2 M}}{\sqrt{2} R^{3/2}}
\end{equation}
\begin{equation}
C=\frac{2 M}{R^2 (R-2 M)}
\end{equation}
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{Potentials.pdf}
\caption{Finch-Skea interior and exterior metric potentials with adopted junction conditions for compact star PSRJ1416-2230}
\label{fig:6555}
\end{figure}
We as well plotted metric potentials using junction conditions defined above on the Figure (\ref{fig:6555}) for relativistic star PSRJ1416-2230, which has mass $M+1.97$ and Radius $R=9.69$. Now we could proceed further and define Bose-Einstein condensate (BEC) field minimally coupled to gravity.
\section{Minimally coupled Bose-Einstein condensate} \label{sec:2}
Generally, the ground state of Bose-Einstein condensate could be described by the complex scalar field \cite{Fagnocchi:2010sn}. If we assume Minimally-coupled to Einsteinian gravity Bose-Einstein field, classical GR Lagrangian (\ref{eq:1}) transforms as
\begin{equation}
\mathcal{S}[g,\Gamma,\Psi_i,\hat{\phi}]=\int_\mathcal{M}d^4x\sqrt{-g}\frac{1}{2\kappa^2}(\mathcal{R}+\mathcal{L}(\Psi_i)+\mathcal{L}(\hat{\phi}))
\label{eq:23}
\end{equation}
where we define \cite{Bettoni_2014}
\begin{equation}
\mathcal{L}(\hat{\phi})=-g_{\mu\nu}\nabla^\mu\hat{\phi}^\dagger \nabla^\nu \hat{\phi}-m^2\hat{\rho}-U(\hat{\rho})
\end{equation}
Here, obviously $\hat{\phi}$ is Bose-Einstein field, $\hat{\phi}^\dagger$ is its complex conjugate, $m^2$ is scalar Bose field mass and finally $U(\hat{\rho})$ is the so-called self-interaction potential, which could be expanded in a series with many-body interaction terms (external potential $V(\hat{\rho})$ for Bose field is absent) \cite{PhysRevD.9.3320,PhysRevD.9.3357}:
\begin{equation}
U(\hat{\rho})=\frac{\lambda_2}{2}\hat{\rho}^2+\frac{\lambda_3}{6}\hat{\rho}^3+...-\frac{\lambda_2}{8}T^2\hat{\rho}+\frac{\pi^2}{90}T^4\hat{\rho}^2
\end{equation}
where first term corresponds to the regular two particle interaction, $\rho=\hat{\phi}^\dagger\hat{\phi}$ is the Bose field probability density. For the sake of simplicity, in the definition of self-interaction potential we will use only the terms up to quadratic, and therefore we redefine set of self-interaction coupling coefficients to $\eta$. As well, we assume that our Bose-Einstein condensate is pure, and therefore $T=0$. Using (\ref{eq:3}) we could easily obtain stress-energy tensor for our Bose field:
\begin{equation}
\begin{gathered}
T_{\mu\nu}^{\hat{\phi}}=\nabla_\mu \hat{\phi}^\dagger \nabla_\nu \hat{\phi}+\nabla_\mu \hat{\phi}\nabla_\nu \hat{\phi}^\dagger -g_{\mu\nu}\bigg(g^{\alpha\beta}\nabla_\alpha \hat{\phi}^\dagger \nabla_\beta \hat{\phi} +m^2\hat{\rho}+U(\hat{\rho})\bigg)
\end{gathered}
\end{equation}
Since we are working with scalar field, for the first order covariant derivatives further we imply proper transformation $\nabla_\mu \hat{\phi}\to \partial_\mu \hat{\phi}$. To numerically derive Bose field it is also useful to introduce modified massive Klein-Gordon equation:
\begin{equation}
g_{\mu\nu}\nabla^\mu \nabla^\nu \hat{\phi}-\bigg(m^2+U'(\hat{\rho})\bigg)\hat{\phi}=0
\end{equation}
Using the assumption of harmonic time dependence we could impose the transformation
\begin{equation}
\hat{\phi}=\exp(-i\omega t)\phi(r)
\end{equation}
where $\phi(r)$ is the real function that depends only on radial coordinate $r$, and therefore $\hat{\rho}=\hat{\phi}\hat{\phi}^\dagger=\phi^2$. Using the aforementioned assumption. Klein-Gordon equation reduces to Gross–Pitaevskii–like equation \cite{Matos:2016ryp}:
\begin{equation}
i\hslash \nabla_t \hat{\phi}=\bigg(-\frac{\hslash^2}{2m}\nabla^i\nabla_i+\underbrace{mV_{\mathrm{ext}}}_\text{0}+\eta\hat{\rho}^2\bigg)\hat{\phi}
\end{equation}
where
\begin{equation}
\nabla_i\hat{\phi} = \frac{1}{\sqrt{-g}}\partial_i(\sqrt{-g}g^{ij}\partial_i\hat{\phi})
\end{equation}
and $V_{\mathrm{ext}}$ is external integral. We will solve the equation above numerically for the set of initial conditions $\phi(0)= 10$ and $\phi'(0)=C$. Solutions for above equation have viable behavior for $\omega\gg|\eta|\land m$ and for relatively small values of this parameters. We plot the oscillating solution for $\phi(r)$ on the Figure (\ref{fig:1}).
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{BEC.pdf}
\caption{(\textit{left plot}) Solutions for scalar field $\phi(r)$ governed by Gross–Pitaevskii equation, (\textit{right plot}) probability density for Bose-Einstein condensate. To obtain solutions, we assumed that $m=\eta=10^{-8}$, $\omega=10^{-2}$. As well, mass and radius were similar to compact star PSRJ1416-2230 ($M=1.97$ and $R=9.69$)}
\label{fig:1}
\end{figure}
Finally, we could derive field equations from modified Einstein-Hilbert action (\ref{eq:23}):
\begin{equation}
\kappa (\rho+\rho^{\hat{\phi}})=\frac{1-e^{-\lambda(r)}}{r^2}+\frac{e^{-\lambda(r)}\lambda'(r)}{r}
\end{equation}
\begin{equation}
\kappa (p_r+p_r^{\hat{\phi}})=\frac{e^{-\lambda(r)}-1}{r^2}+\frac{e^{-\lambda(r)}\nu'(r)}{r}
\end{equation}
\begin{equation}
\kappa (p_t+p_t^{\hat{\phi}}) = e^{-\lambda}\left(\frac{\nu''(r)}{2}+\frac{\nu'(r)^2}{4}-\frac{\nu'(r)\lambda'(r)}{4}+\frac{\nu'(r)-\lambda'(r)}{2r}\right)
\end{equation}
where \cite{doi:10.1142/S0218271821300068}
\begin{equation}
\rho^{\hat{\phi}} = -e^{-\nu}\omega^2\phi^2 - e^{-\lambda}(\phi')^2 + V(\phi)
\end{equation}
\begin{equation}
p_r^{\hat{\phi}} = -e^{-\nu}\omega^2\phi^2 - e^{-\lambda}(\phi')^2 - V(\phi)
\end{equation}
\begin{equation}
p_t^{\hat{\phi}} = -e^{-\nu}\omega^2\phi^2 + e^{-\lambda}(\phi')^2 - V(\phi)
\end{equation}
Here, we define $V(\Phi)=m^2\hat{\rho}^2/2+U(\hat{\rho})$.
In the next subsection we are going to probe the behavior of Finch-Skea star minimally coupled to Bose-Einstein condensate.
\subsection{Energy Conditions}
Energy conditions are probes of relativistic model viability. Generally, there exists four energy conditions, namely Weak Energy Condition (WEC), Null Energy Condition
(NEC), Strong Energy Condition (SEC) and Dominant Energy Condition (DEC). For model to be physically plausible, all of the aforementioned energy conditions must be satisfied at the every point of manifold. EC's, derived from temporal Raychaudhuri equations are defined as follows:
\begin{itemize}
\item Null Energy Condition (NEC): $\rho +p_{r}\geq 0$ and $\rho +p_{t}\geq 0
$
\item Weak Energy Condition (WEC) $\rho >0$ and $\rho +p_{r}\geq 0$ and $%
\rho +p_{t}\geq 0$
\item Dominant Energy Condition (DEC): $\rho -|p_{r}|\geq 0$ and $\rho
-|p_{t}|\geq 0$
\item Strong Energy Condition (SEC): $\rho +p_{r}+2p_{t}\geq 0$
\end{itemize}
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{BECConds.pdf}
\caption{Null, Dominant and Strong energy conditions for Finch-Skea PSRJ1416-2230 star minimally coupled to Bose-Einstein condensate (we assumed that $m=\eta=10^{-8}$, $\omega=10^{-2}$)}
\label{fig:3}
\end{figure}
We plot energy conditions for Finch-Skea PSRJ1416-2230 star minimally coupled to Bose-Einstein condensate on the Figure (\ref{fig:3}). As one may obviously notice, NEC is validated for tangential and radial pressures, DEC is validated for both pressure kinds, SEC is obeyed everywhere. It is worth to state that Null, Dominant and Strong Energy Conditions will be still satisfied for every relatively small and positive values of Bose field mass squared, frequency and self-coupling constant $\eta$.
\subsection{Gradients and Equation of State}
To probe the nature of matter content in the Finch-Skea star interior region we will use the Eqution of State (further - EoS) parameter, which is defined as
\begin{equation}
\omega_r=p_r/\rho,\quad \omega_t=p_t/\rho
\end{equation}
As well, we could depict the energy density, anisotropic pressure gradients by simply calculating $\rho'(r)$ and $p_r'(r)$, $p_t'(r)$. We illustrated EoS and stress-energy tensor component gradients on the Figure (\ref{fig:4}). As we see, at the stellar origin radial EoS is asymptotically ($\phi'(0)\to-\infty$) described by stiff Zeldovich fluid and tangential one by dark-energy fluid, but for all negative values of $\phi'(0)$ initial condition radial fluid is regular one in the bounds $\omega\in(0,1)$ and tangential is quintessence in the bounds $\omega\in(-1,0)$. On the other hand, gradient of energy density, anisotropic pressures could have both positive and negative values in the stellar interior, but generally because of the asymptotical flatness, gradients vanish at the region $r\to \infty$.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{EoS.pdf}
\caption{(\textit{left plot}) Equation of State for radial and tangential pressures (parameter values are the same as for EC's), (\textit{right plot}) Gradient of the energy density and anisotropic pressures}
\label{fig:4}
\end{figure}
\subsection{Tolman-Oppenheimer-Volkoff stability}
Stability of the matter content in the stellar interior could be investigated using the well known Tolman-Oppenheimer-Volkoff equilibrium condition, which is given below in its modified form \cite{Oppenheimer1939,Poncede1993,Rahaman2014,Tolman1939}:
\begin{equation}
-\frac{dp_{r}}{dr}-\frac{\nu^{'}(r)}{2}(\rho+p_{r})+\frac{2}{r}(p_{t}-p_{r})+F_{\mathrm{ex}}=0
\label{eq:34}
\end{equation}
As one may obviously notice, in the modified form of TOV equilibrium condition present one extra force, namely $F_{\mathrm{ex}}$, which is present because of the stress-energy-momentum tensor discontinuity ($\nabla^\mu T_{\mu\nu}\neq0$) to hold relativistic object stable. As well, in classical and modified TOV's present three additional forces: hydrodynamical $F_\mathrm{H}$, gravitational $F_{\mathrm{G}}$ and anisotropic $F_{\mathrm{A}}$:
\begin{equation}
F_H=-\frac{dp_{r}}{dr},\;\;\;\;\;\;\;\;F_A=\frac{2}{r}(p_{t}-p_{r}), \;\;\;\;\;\;\;\;F_G=-\frac{\nu^{'}}{2}(\rho+p_{r}),
\end{equation}
and therefore, we could easily rewrite MTOV (\ref{eq:34}):
\begin{equation}\label{t}
F_A+F_G+F_H+F_{ex}=0
\end{equation}
We plot the solutions for each force present in MTOV on the Figure (\ref{fig:5}). It is noticable that both hydrodynamical force $F_H$ and extra force $F_{ex}$ have positive values at the whole interior radial domain, gravitational and anisotropic forces are always negative and asymptotically vanish. Hence, for Finch-Skea star in the presence of Bose-Einstein condensate modified TOV equilibrium is satisfied (relativistic star is stable) if extra force is present.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{TOVBEC.pdf}
\caption{Modified TOV forces for Finch-Skea star minimally coupled to Bose-Einstein condensate}
\label{fig:5}
\end{figure}
\subsection{Adiabatic index}
Adiabatic index is the tool that could be used to probe relativistic object via adiabatic perturbations. Firstly, it was introduced in the work of Chandrasekhar \cite{Chandrasekhar1964}. Chandrasekhar predicted that for the relativistic system to be stable the adiabatic index should exceed $4/3$. This adiabatic index is defined as \cite{Maurya2017}:
\begin{equation}
\Gamma = \frac{p_r+\rho}{p_r}\frac{dp_r}{d\rho}
\label{eq:31}
\end{equation}
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{AdiabaticBEC.pdf}
\caption{(\textit{first row}) Adiabatic index for Finch-Skea star minimally coupled to Bose-Einstein condensate, (\textit{second row}) speed of sound for Finch-Skea star minimally coupled to Bose-Einstein condensate}
\label{fig:76}
\end{figure}
Adiabatic index solutions are plotted on the Figure (\ref{fig:76}). Generally, for the same values of parameters, as it was used in the energy conditions subsection, $\Gamma$ constraint satisfied at the area near origin and envelope for small $\phi'(0)<0$ and satisfied everywhere for $\phi'(0)\ll0$. Moreover, stability holds for every relatively small and positive number of $\omega$, $\eta$ and $m$.
\subsection{Surface redshift}
Finally, we could also define surface redshift:
\begin{equation}
\mathcal{Z}_s = |g_{tt}|^{-1/2}-1
\label{eq:34}
\end{equation}
Which for anisotropic matter distribution must not exceed the value of 2. Routinely, we plotted numerical solution for surface redshift for different compact stars on the Figure (\ref{fig:6}). Judging by the plotted data we could conclude that our solution satisfy surface gravitational redshift constraint.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.6\textwidth]{SurfaceBEC.pdf}
\caption{Surface (gravitational) redshift for Finch-Skea stellar interior (parameter values are the same as for EC's)}
\label{fig:6}
\end{figure}
\subsection{Sound of speed}
Another important criterion of matter content viability is the so-called speed of sound. This quantity is defined in the following way (\textbf{derived from the ultra-relativistic hydrodynamics}):
\begin{equation}
v^2 = \frac{dp_r}{d\rho}\leq c^2 =1
\end{equation}
\textbf{Inequality above needs to be satisfied in order to respect the causality constraints.} Solution for speed of sound is plotted on the third plot of Figure (\ref{fig:76}). As we see, inequality $v^2\leq c^2$ always holds, which is necessary condition. Also, it is important to state that asymptotically $\phi'(0)\to-\infty$ our fluid becomes massless, since $v^2=c^2$.
Since we already probed all of the necessary parameters for minimally coupled Bose-Einstein condensate, we could proceed to the minimally coupled Kalb-Ramond field case.
\section{Minimally coupled Massless Kalb-Ramond field} \label{sec:3}
In this section, we are going to investigate the compact stars admitting Finch-Skea symmetry in the presence of background Kalb-Ramond (further - KB) field. For that particular case, total Einstein-Hilbert action integral is given below:
\begin{equation}
\mathcal{S}[g,\Gamma,\Psi_i,\hat{\phi},B_{\mu\nu}]=\int_\mathcal{M}d^4x\sqrt{-g}\frac{1}{2\kappa^2}(\mathcal{R}+\mathcal{L}(\Psi_i)))+\int_\mathcal{M}d^4x\sqrt{-g} H_{\mu\nu\alpha}H^{\mu\nu\alpha}
\label{eq:23}
\end{equation}
In the action above, $H_{\mu\nu\alpha}$ is defined as a Kalb-Ramond field strength:
\begin{equation}
H_{\mu\nu\alpha}=\partial _\mu B_{\nu\alpha}+\partial _\nu B_{\alpha\mu}+\partial _\alpha B_{\mu\nu}
\end{equation}
where obviously $B_{\mu\nu}$ is the antisymmetric rank 2 tensor, namely Kalb-Ramond field. For KB field we have a set of two EoMs, which are given as follows \cite{PhysRevD.77.044030,PhysRevD.52.5877,2018EPJC...78..531D}:
\begin{equation}
\frac{1}{\sqrt{-g}}\partial_\mu (\sqrt{-g}H^{\mu\nu\alpha})=0
\label{eq:41}
\end{equation}
\begin{equation}
\epsilon^{\mu\nu\alpha\beta}\partial_\mu (\sqrt{-g}H_{\nu\alpha\beta})
\end{equation}
Here, $\epsilon^{\mu\nu\alpha\beta}$ is the well known Levi-Cevita symbol, which is for Minkowskian spacetime equal to unity if it's indices are an even permutation of $(1234)$, equal to $-1$ if indices are odd permutation of $(1234)$ and vanish if one of the indices repeat. For KB field, stress-energy-momentum tensor reads \cite{maluf2022bianchi}:
\begin{equation}
T^{\mu\nu}_B=\frac{1}{2}H^{\alpha\beta\mu}H^{\nu}_{\;\alpha\beta}-\frac{1}{2}g^{\mu\nu}H^{\alpha\beta\gamma}H_{\alpha\beta\gamma}
\end{equation}
where we consider that there is no external potential present. Solving field equation (\ref{eq:41}), we could obtain the solution for Kalb-Ramond field strength \cite{nair2021kalbramond}:
\begin{equation}
H_{\mu\nu\alpha}=\epsilon^{\mu\nu\alpha\beta}\partial_\beta \phi
\end{equation}
Here, $\phi$ is the real scalar field, whose evolution is governed by the following equation:
\begin{equation}
\nabla_\mu\nabla^\mu\phi =0
\end{equation}
Which is exactly massless wave equation. As usual, we will assume only radial coordinate dependence for real scalar field and as well use set of initial conditions to solve wave equation above numerically: $\phi(0)=0.1$ and $\phi'(0)=C$ (if we assume that radial derivative of scalar field vanish at the origin, scalar field will have constant solution). To numerically plot the results in the presence of KB field we will vary the values of scalar field first order radial derivative at the origin $\phi'(0)$. Finally, using previously defined stress-energy-momentum tensor for KB field we could write down new set of (modified) Einstein Field Equations:
\begin{equation}
\begin{gathered}
\rho=e^{\lambda (r)-4 \nu (r)} \phi '(r)^2-\frac{e^{-\lambda (r)} \left(r \lambda
'(r)+e^{\lambda (r)}-1\right)}{r^2}
\end{gathered}
\end{equation}
\begin{equation}
\begin{gathered}
p_r =-\frac{e^{-\lambda (r)} \left(r \nu '(r)+1\right)}{r^2}+\frac{1}{r^2}+3 e^{-3 \lambda (r)}
\phi '(r)^2
\end{gathered}
\end{equation}
\begin{equation}
\begin{gathered}
p_t =\frac{e^{-\lambda (r)} \left(\left(r \nu '(r)+2\right) \left(\lambda '(r)-\nu '(r)\right)-2 r
\nu ''(r)\right)}{4 r}-\frac{e^{\lambda (r)} \phi '(r)^2}{r^8}
\end{gathered}
\end{equation}
Since our spacetime admits spherical symmetry, without the loss of generality we could work only in the equatorial plane adopting $\theta=\pi/2$.
\subsection{Energy conditions}
Energy conditions for Finch-Skea star minimally coupled to Kalb-Ramond field are illustrated on the Figure (\ref{fig:8}). Obviously, from that data we could conclude that all of the energy conditions are violated. This conditions could be satisfied if we will take another values of free parameters $\omega$, initial conditions $\phi(0)$ and $\phi'(0)$, but for that case speed of sound will exceed $c^2$, which is forbidden.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{CondsKB.pdf}
\caption{Null, Dominant and Strong energy conditions for Finch-Skea PSRJ1416-2230 star minimally coupled to Kalb-Ramond field}
\label{fig:8}
\end{figure}
\subsection{Gradients and Equation of State}
In addition to the energy conditions, we as well probed the equation of state for our stellar matter content and energy density, anisotropic pressure radial derivatives behavior. Results are consequently plotted on the Figure (\ref{fig:9}). EoS has regular behavior ($0<\omega<1$) for radial pressure and is negative (quintessence) for tangential pressure (near the stellar core), regular near envelope. Finally, gradients generally are positive for energy density and radial pressure, negative for tangential one.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{EoSKB.pdf}
\caption{Equation of state and radial derivative of stress-energy tensor components for Finch-Skea star coupled to KB field}
\label{fig:9}
\end{figure}
\subsection{TOV stability}
It is also convenient to investigate the TOV stability of our Finch-Skea stellar solution coupled to KB field. Graphical results of such investigation are located on the four plots of Figure (\ref{fig:10}). It is worth to notice that only hydrodynamical force have negative values, gravitational, anisotropic and extra forces therefore have positive values.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{KBForces.pdf}
\caption{Forces present in the modified TOV equilibrium for Finch-Skea star coupled to KB field}
\label{fig:10}
\end{figure}
\subsection{Adiabatic index}
As a final note, in this subsection we are going to probe the behavior of adiabatic index for our stellar solutions. Results of numerical investigation are properly plotted on the Figure (\ref{fig:11}). One may notice that our stellar solution is stable everywhere except the regions where adiabatic index numerical solution diverge.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{AdiabaticKB.pdf}
\caption{($first\;row$) Adiabatic index for anisotropic Finch-Skea star coupled to KB field, ($second\;row$) sound of speed squared for anisotropic Finch-Skea star coupled to KB field}
\label{fig:11}
\end{figure}
\subsection{Speed of sound}
As the final note on minimal KB field in this subsection we are going to investigate the speed of sound for perfect fluid matter content inside Finch-Skea compact star. Results of such investigation are properly plotted on the last, third plot of Figure (\ref{fig:11}). As we see, judging by the data from the plot our fluid is relativistic, but speed of sound does not exceed the speed of light, which is required for fluid to be viable. As well, it is worth to notice that the values of $v^2$ are generally smaller in the presence of KB field (in relation to the classical GR).
\section{$U(1)$ Gauge background field cosmology}\label{sec:4}
In this section we present the final model of our consideration, namely gauged boson Finch-Skea star. This theory has a a gauge field $A_\mu$ admitting $U(1)$ unitary symmetry and massless complex scalar field $\Phi$ with conical potential $V(|\Phi|)=\lambda |\Phi|$. For that kind of model, Lagrangian looks exactly like:
\begin{equation}
\mathcal{S}[g,\Gamma,\Psi_i,\hat{\phi},A_\mu,\Phi]=\int_\mathcal{M}d^4x\sqrt{-g}\frac{1}{2\kappa^2}(\mathcal{R}+\mathcal{L}(\Psi_i)))-\frac{1}{4}\int_\mathcal{M}d^4x\sqrt{-g}F^{\mu\nu}F_{\mu\nu} -\int_\mathcal{M}d^4x\sqrt{-g}(\mathcal{D}_\mu \Phi)^*(\mathcal{D}^\mu \Phi)-V(|\Phi|)
\label{eq:49}
\end{equation}
where,
\begin{equation}
\mathcal{D}_\mu \Phi = \partial_\mu \Phi+ieA_\mu \Phi
\end{equation}
\begin{equation}
F_{\mu\nu}=\partial_\mu A_\nu - \partial_\nu A_\mu
\end{equation}
Asterisk about expression with gauge derivative denotes usual complex conjugation. For electromagnetic field $F_{\mu\nu}$ and massless scalar field $\Phi$ equations of motion could be easily derived from the minimal action:
\begin{equation}
\mathcal{D}_\mu(\sqrt{-g}F^{\mu\nu})=-ie\sqrt{-g}[\Phi^*(\mathcal{D}^\nu\Phi)-\Phi(\mathcal{D}^\nu\Phi)^*]
\end{equation}
\begin{equation}
\mathcal{D}_\mu(\sqrt{-g}\mathcal{D}^\mu\Phi)=\frac{\lambda}{2}\sqrt{-g}\frac{\Phi}{|\Phi|}
\end{equation}
\begin{equation}
[ \mathcal{D}_\mu(\sqrt{-g}\mathcal{D}^\mu\Phi)]^*=\frac{\lambda}{2}\sqrt{-g}\frac{\Phi^*}{|\Phi|}
\end{equation}
Finally, it is also useful to define stress-energy-momentum tensor for gauge and complex scalar fields respectively
\cite{PhysRevD.93.044014}:
\begin{equation}
T_{\mu\nu}^F=E_{\mu\nu}=-g_{\mu\nu}\mathcal{L}_F+\frac{2\partial \mathcal{L}_F}{\partial g^{\mu\nu}}=F_{\mu\alpha}F_\nu^{\;\alpha}-\frac{1}{4}g_{\mu\nu}F_{\alpha\beta}F^ {\alpha\beta}
\label{eq:55}
\end{equation}
\begin{equation}
T_{\mu\nu}^\Phi=(\mathcal{D}_\mu\Phi)^*(\mathcal{D}_\nu\Phi)+(\mathcal{D}_\mu\Phi)(\mathcal{D}_\nu\Phi)^*-g_{\mu\nu}[(\mathcal{D}_\alpha\Phi)^*(\mathcal{D}_\beta\Phi)g^{\alpha\beta}-g_{\mu\nu}\lambda |\Phi|]
\label{eq:56}
\end{equation}
Throughout the paper we assume vanishing electromagnetic strength tensor $F_{\mu\nu}=0$. For that purpose we could use the anzatz below:
\begin{equation}
\Phi(t,r)=\phi(r)e^{i\omega t},\quad A_{\mu}(x^\mu)dx^\mu = A(r)dt
\end{equation}
Therefore, field equations for gauged boson Finch-Skea fluid sphere are expressed in the following form \cite{PhysRevD.88.024053}:
\begin{equation}
\kappa\rho= \left[e^{-\nu}
\left(\omega+qA\right)^2\right] \phi^2+\frac{ e^{-\lambda-\nu}
(A')^2}{2}+ \phi'^2e^{-\lambda}
\end{equation}
\begin{equation}
\kappa p_r =
\left[-e^{-\nu} \left(\omega+qA\right)^2\right]
\phi^2+\frac{e^{-\lambda-\nu}
(A')^2}{2}- \phi'^2e^{-\lambda}
\end{equation}
\begin{equation}
\kappa p_t =\left[-e^{-\nu} \left(\omega+qA\right)^2\right]
\phi^2-\frac{e^{-\lambda-\nu} (A')^2}{2}+ \phi'^2e^{-\lambda}
\end{equation}
Also, we could properly derive field equations for gauge (Maxwell) and scalar (Klein-Gordon) fields:
\begin{equation}
A''+\left( \frac{2}{r}-\frac{\nu'+\lambda'}{2}\right)A'-2 q
e^{\lambda} \phi^2 \left(\omega+qA\right)=0
\end{equation}
\begin{equation}
\phi''+\left( \frac{2}{r}+\frac{\nu'-\lambda'}{2}\right)\phi'+
e^{\lambda} \left[
\left(\omega+qA\right)^{2}e^{-\nu}\right]\phi=0
\end{equation}
As an initial conditions we have chosen the case with $\phi(0)=A(0)=const$ and $\phi'(0)=A'(0)=0$. In the next subsection we are going to match interior stellar spacetime and exterior Reissner-N\"ordrstrom one.
\subsection{Junction Conditions}
Since, we added minimally coupled $U(1)$ gauge field to our total Lagrangian, we need to redefine the junction conditions in the presence of interior charge. This time we will match interior and exterior spacetimes using the Reissner-N\" ordstrom metric tensor (which describes charged spherically symmetric vacuum), for which line element reads:
\begin{equation}
ds^2 = \bigg(1-\frac{2M}{R}+\frac{Q^2}{R^2}\bigg)dt^2 - \bigg(1-\frac{2M}{R}+\frac{Q^2}{R^2}\bigg)^{-1}dr^2 - r^2 d\theta^2 - r^2 \sin^2 \theta d\phi^2
\end{equation}
Here, $Q$ is the total charge of the system, which is defined as
\begin{equation}
Q=\int^R_0j^t\sqrt{-g}dr
\end{equation}
where $j^t$ is the only one non-vanishing component of four-current (Noether current) in the static spacetime. This current could be written as follows
\begin{equation}
j^\mu = -ie[\Phi(\mathcal{D}^\mu\Phi)^*-\Phi^*(\mathcal{D}^\mu\Phi)]
\end{equation}
With the use of spherically symmetric line element of stellar spacetime we could rewrite the expression for total charge:
\begin{equation}
Q=8 \pi q\int^{R}_{0}dr r^2
\left(\omega+qA\right)\phi^2e^{\frac{\lambda-\nu}{2}}
\end{equation}
To numerically solve field equations and coupled differential equations for scalar and gauge fields we need to make and assumption of value for charge $Q$. Since we are working in the RS spacetime, $Q<M$.
Using RS line element we could impose continuity conditions at the boundary:
\begin{equation}
\mathrm{Continuity\;of\;}g_{tt}:\quad 1-\frac{2M}{R}+\frac{Q^2}{R^2}=e^{\nu(r)}
\end{equation}
\begin{equation}
\mathrm{Continuity\;of\;}\frac{\partial g_{tt}}{\partial r}:\quad \frac{2M}{R^2}-\frac{2Q^2}{R^3}=-B \left(2 A \sqrt{C R^2}+B C R^3\right)
\end{equation}
\begin{equation}
\mathrm{Continuity\;of\;}g_{rr}:\quad \bigg(1-\frac{2M}{R}+\frac{Q^2}{R^2}\bigg)^{-1}=e^{\lambda(r)}
\end{equation}
\begin{equation}
\mathrm{Vacuum\;condition}:\quad p\rvert_{r=R}=0
\end{equation}
Solutions for above conditions are introduced below:
\begin{equation}
\begin{gathered}
A=\frac{\sqrt{R (R-2 M)+Q^2}}{R}-\frac{1}{2} B R \sqrt{C R^2}
\end{gathered}
\end{equation}
\begin{equation}
\begin{gathered}
B=\frac{C \left(M R-Q^2\right)}{\left(C R^2\right)^{3/2} \sqrt{R (R-2 M)+Q^2}}
\end{gathered}
\end{equation}
\begin{equation}
\begin{gathered}
C=\frac{1}{-2 M R+Q^2+R^2}-\frac{1}{R^2}
\end{gathered}
\end{equation}
We will use these junction conditions to numerically investigate our model in the next subsections.
\subsection{Energy Conditions}
Routinely, Null, Dominant and Strong energy conditions are plotted on the Figure (\ref{fig:13}). It is obvious that in our case only one energy condition, namely null energy condition for radial pressure is satisfied. Some of other violated EC's could be satisfied for bigger $\omega$ and different initial conditions for $A(r)$ and $\phi(r)$, but in that case $v^2>c^2$, which is unacceptable. Also, it is worth to notice that for Noether charge obeying inequality $|Q|<M$ if we will vary it's values, results will not change seriously and EC's will be still validated/violated.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{U1Conds.pdf}
\caption{Null, Dominant and Strong energy conditions for Finch-Skea PSRJ1416-2230 star minimally coupled to $U(1)$ gauge field and massless complex scalar field. To numerically derive $T_{\mu\nu}$ components we use assumptions $Q=0.5$ and $\omega=10^{-6}$}
\label{fig:13}
\end{figure}
\subsection{Gradients and Equation of State}
As usual, in this subsection we are going to study the radial derivatives of energy density, pressures and equation of state for Finch-Skea star minimally coupled to $U(1)$ gauge field, massless complex scalar field. Results of such investigation are illustrated on the Figure (\ref{fig:14}). From the plots we could conclude that radial fluid behave like dark energy and tangential like stiff fluid. On the other hand, gradients are positive for radial pressure and negative for energy density, tangential pressure.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{EoSU1.pdf}
\caption{Equation of state and radial derivative of stress-energy tensor components for Finch-Skea star coupled to $U(1)$ gauge field and massless complex scalar field. To numerically derive that solutions we use assumptions $Q=0.5$ and $\omega=10^{-6}$}
\label{fig:14}
\end{figure}
\subsection{TOV stability}
As well, we want to probe the stability of our stellar matter content through the well known modified TOV equation. We properly locate graphical results of TOV forces on the Figure (\ref{fig:15}). For particular case in the presence of gauge field, all of the forces except extra one are negative.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{ForcesU1.pdf}
\caption{Forces present in the modified TOV equation for Finch-Skea star coupled to $U(1)$ gauge field and massless complex scalar field. To numerically derive that solutions we use assumptions $Q=0.5$ and $\omega=10^{-6}$}
\label{fig:15}
\end{figure}
\subsection{Adiabatic index}
In this subsection we are going to study the stability of our stellar matter content from continuous adiabatic perturbations. Numerical solution for adiabatic index is located on the first plot of the Figure (\ref{fig:15}). As we see here, adiabatic index satisfy the $\Gamma>4/3$ constraint, so our stellar solutions in the presence of gauge field could be considered as stable from adiabatic perturbations.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{AdiabU1.pdf}
\caption{Anisotropic adiabatic index and speed of sound for Finch-Skea star coupled to $U(1)$ gauge field and massless complex scalar field. To numerically derive that solutions we use assumptions $Q=0.5$ and $\omega=10^{-6}$}
\label{fig:16}
\end{figure}
\subsection{Speed of Sound}
Finally, we as well probed the speed of sound on the last plot of Figure (\ref{fig:16}). As we see, in our case fluid is ultrarelativistic, but does not exceed unity, which is required.
\subsection{Surface redshift}
Since the junction conditions changed, surface redshift also will vary. Graphical representation of surface redshift is therefore placed on the Figure (\ref{fig:17}) for different values of total charge $Q$. As we noticed during the numerical analysis, surface redshift for charged stars is smaller that for uncharged ones, and consequently as $|Q|\to\infty$, $\mathcal{Z}_s\to 0$.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.6\textwidth]{SurfaceU1.pdf}
\caption{Surface (gravitational) redshift for gauged Finch-Skea stellar interior (parameter values are the same as for EC's). On the plot solid line represent the solution with $Q=0.9$ and dashed for $Q=1$}
\label{fig:17}
\end{figure}
\section{Conclusions}\label{sec:5}
In this article, we investigated the spherically symmetric and static stellar solutions imposing Finch-Skea symmetry in the presence of such exotic matter fields as Bose-Einstein Condensate, Kalb-Ramond antisymmetric tensor field and gauge field with local $U(1)$ gauge symmetry. To numerically solve numerous equations and graphically plot the results of such investigation we have used compact star PSRJ1416-2230, which has the mass $M=1.69M_\odot$ and radius $R=9.69R_\odot$. In this section we are going to mention main results of our study:
\begin{itemize}
\item \textbf{Energy Conditions}: for Bose-Einstein Condensate case Null and Strong energy conditions for both radial and tangential pressures were satisfied, so there is no exotic fluid in the stellar interior present. On the other hand, for the sake of causality every of the energy conditions mentioned in the paper were violated for Kalb-Ramond field and only radial NEC was validated for gauge field. We have already plotted energy conditions for each exotic matter field of our consideration on the Figures (\ref{fig:3}), (\ref{fig:8}) and (\ref{fig:13})
\item \textbf{Equation of State}: for BE Condensate radial EoS parameter is regular and asymptotically $\phi'(0)\to-\infty$ describes Zeldovich fluid, tangential asymptotically describes $\Lambda$CDM dark energy like fluid. For Kalb-Ramond fluid situation differs, radial EoS is regular and vanish at the envelope, tangential one is negative at the stellar core and regular at the intermediate, envelope regions. Finally, for gauge field radial EoS describes DE-like fluid for any $|q|\in\mathbb{R}$ and tangential describes fluid with EoS which slightly deviates from the Zeldovich one and as $q\to0$, $\omega\to1$. For more detailed information and graphical representation of aforementioned results, refer to the Figures (\ref{fig:4}), (\ref{fig:9}) and (\ref{fig:14}) first row.
\item \textbf{Gradients of perfect fluid stress-energy tensor components}: radial gradients were negative for energy density and radial pressure and positive for tangential one in the presence of BEC. For KB field everything is opposite, namely $\rho'(r)\land p_r'(r)\geq0$ and $p_t(r)\leq0$. Finally, for $U(1)$ gauge electromagnetic field energy density and tangential gradients were negative, radial was positive. Numerical evidence for provided statements could easily be found at the Figures (\ref{fig:4}), (\ref{fig:9}) and (\ref{fig:14})
\item \textbf{TOV equilibrium}: we have as well probed the dynamical stability of our stellar object with minimally coupled exotic fields. For BEC as we noticed only hydrodynamical and extra forces were positive, the rest were negative and as well the vast contribution to the total TOV provided fluid anisotropy. On the other hand, for KB field there were only one negative force, namely hydrodynamical one, which was also the biggest one. Finally, for gauge field only extra force were positive. As usual, results are properly illustrated on the Figures (\ref{fig:5}), (\ref{fig:10}) and (\ref{fig:15})
\item \textbf{Adiabatic index}: in the current study we also studied the stability of anisotropic matter inside compact star from adiabatic perturbations. As it was shown in this paper, $\Gamma>4/3$ for relatively big and negative initial condition for scalar field $\phi'(0)$ in the BEC case. Moreover, for KB case star is stable everywhere except the diverging regions. Ar final, for gauge field star is also stable everywhere. $\Gamma$ is plotted over the whole interior radial domain on the Figures (\ref{fig:76}), (\ref{fig:11}) and (\ref{fig:16})
\item \textbf{Sound of speed}: the last quantity that we are going to discuss is the speed of sound. For object to obey causality condition, $v^2\leq c^2 =1$. For the BEC, KB and gauge fields this necessary condition was validated everywhere, as expected. Squared speed of sound is therefore plotted on the second plot of Figures (\ref{fig:76}), (\ref{fig:11}) and (\ref{fig:16})
\end{itemize}
Now, it will be handful to compare our results with the existing ones obtained for other compact star geometries. There were written numerous papers on compact stars within the General Theory of Relativity, however, here we are going to concentrate on the solutions, that are close to our own. For example, one could introduce Einstein-Klein-Gordon stars (compact stellar solutions endangered by scalar field), Einstein-Maxwell stars (analogically, stars endangered by Maxwell field that usually respects some gauge symmetries) or Boson/Proca stars as the closest "siblings" to our solutions.
For Einstein-Maxwell Buchdahl stars, good work in the GR theory is \cite{2021Ap&SS.366...26P}. In comparison to our results for stars with $U(1)$ background field, their solution respected all energy conditions. Moreover, because of the isotropy of their stellar solution, anisotropic forces in TOV vanished, hydrodynamical one were positive. Speed of sound squared had similar behaviour to ours, but the rate of change with $r$ was significantly bigger, EoS parameter was positive within the bounds of $\omega\in(0,1)$. Finally, the adiabatic index also had a lot bigger values, but inequality $\Gamma>4/3$ holds for both ours and their solutions.
On the other hand, bosonic stellar solutions unfortunately does not have such comprehensive analysis in the existing literature within General Theory of Relativity. However, there are some papers written within the Einstein-Scalar-Gauss-Bonnet (ESGB) theory, which is the viable theory of gravitation. For example, in the work of \cite{https://doi.org/10.48550/arxiv.2112.13391}, isotropic ESGB stellar solutions were probed. By comparing aforementioned model with our BEC stellar configuration, we could conclude that EC's for both models had practically the same behavior and asymptotically vanish. But in the EGSB case, $v^2$ and $\Gamma$ are monotonically decreasing functions while in our case they grow nearby stellar envelope. Moreover, for the case with Gauss-Bonnet corrections present, surface redshift were negative while in our case it was positive and decrease up to the envelope.
At the moment, there are no comprehensive studies (with the investigation of EC's, EoS, adiabatic and TOV stability) within the GR for such stellar configurations as Boson, Proca stars as well. However, from paper \cite{sym12122032} we could conclude that similar to our case, both Boson and Proca stars has positive energy density with the peak at the stellar origin and that $\rho\to0$ asymptotically with radial coordinate, which were observed for our case as well. Besides, there were carried out the investigation of magnetised BEC stars in \cite{Angulo:2022gpj}. In comparison to our BEC star solution, magnetised one had smaller surface redshift and both positive pressure and density with $\rho\gg p$, so that NEC/WEC/SEC and DEC are satisfied.
From the aforementioned results and graphical representations it is clear that our stellar solution is of non-singular nature and exhibits interesting properties in the physically interesting and viable models, such as GR expanded using exotic matter fields (BEC dark matter, KB field from string theory and gauge fields from electromagnetism). However, in the future studies it will be interesting to probe such model with Dirac spinor fields, QCD Axions and in the presence of non-homogeneous oscillating axion field with large scale correlations.
\section*{Data Availability Statement}
There are no new data associated with this article.
\section*{Acknowledgments}
PKS acknowledges National Board for Higher Mathematics (NBHM) under Department of Atomic Energy (DAE), Govt. of India for financial support to carry out the Research project No.: 02011/3/2022 NBHM(R.P.)/R\&D II/2152 Dt.14.02.2022. Sokoliuk O. performed the work in frame of the "Mathematical modeling in interdisciplinary research of processes and systems based on intelligent supercomputer, grid and cloud technologies" program of the NAS of Ukraine. We are very much grateful to the honorable referee and to the editor for the illuminating suggestions that have significantly improved our work in terms of research quality, and presentation.
| a610aa776b019ba4ff67600dd92089ed965e3c7f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:Introduction}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{Figures/Fig1.pdf}
\caption{ Motion transfer results. (a) is generated by traditional motion transfer model trained on source domain videos only and (b) is generated by our proposed MAA model}
\label{fig:figure-1}
\end{figure}
Given a source image and a driving video of the same object, motion transfer~(\textit{a.k.a.} image animation) aims to generate a synthesized video that mimics the motion of the driving video while preserving the appearance of the source image. It recently received increasing attention, due to its potential applications in real-world scenarios, such as face swapping\cite{wiles2018x2face,siarohin2019first,xu2021move}, dance transferring\cite{chan2019everybody}, \emph{etc}\onedot} \def\vs{\emph{vs}\onedot.
Many works in this field focus on the single-domain motion transfer \cite{siarohin2019animating,siarohin2019first,siarohin2021motion}, where the driving video and source image come from the same domain. However, in real applications, there are often requirements to transfer motion among different domains. For example, as shown in Fig~\ref{fig:figure-1}, the e-commerce companies might be interested in animating a fashion model to attract consumers by learning the robot dance from a Mixamo character. However, due to the differences in shape and cloth between the Mixamo character and the fashion model, traditional single-domain motion transfer approaches often produce notable artifacts in the synthesized image (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, failing to preserve the human shape of the fashion model (\emph{cf}\onedot} \def\Cf{\emph{C.f}\onedot~Fig.~\ref{fig:figure-1} (a))).
To this end, in the present work, we study the cross-domain motion transfer problem and propose a novel Motion and Appearance Adaptation (MAA) approach to address this issue. Specifically, traditional motion transfer methods usually take two arbitrary frames of the same video as source image and driving frame for learning motion with a reconstruction loss, because the two frames share the same appearance and shape. However, such training mode cannot be directly applied to the cross-domain motion transfer because no ground-truth is available. In our proposed MAA approach, we build a cyclic reconstruction pipeline inspired by CycleGAN\cite{zhu2017unpaired} and cross-identity\cite{jeon2020cross}. In particular, given a source image and a driving frame obtained from different domains, we first obtain a synthesized image using a basic motion transfer (MT) model, \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, the model in\cite{siarohin2019first} or\cite{siarohin2021motion}. We next arbitrarily take another frame from the driving video as the source image and the synthesized image as a driving frame, and input them into the basic MT model to produce the second synthesized image. Because the second synthesized image should mimic both the motion and appearance of original driving frame, a cyclic reconstruction loss can be applied for training. In this way, we obtain a motion transfer model for cross-domain motion transfer.
Moreover, since the source image and driving frame are drawn from different domains, while the topology of the object structure (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, the skeleton) is similar, the configurations of the object structure (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, the human body shape) often deviate. When doing motion transfer, we should be aware of such difference and keep the object shape of synthesized image be similar to the source image while unaffected by the driving frame. For this purpose, we design a shape-invariant motion adaptation module and a structure-guided appearance consistency module to regularize the basic motion transfer model.
Specifically, in the shape-invariant motion adaptation module, we design an angle consistency loss to enforce the angles of the corresponding object parts in the synthesized image to be similar to those of the driving frame, such that the motion of this frame can be mimicked well without changing the object shape. In the structure-guided appearance consistency module, we extract image patches from the synthesized image and the source images based on the object structure and enforce the corresponding patches to be similar; this ensures that the appearance of the synthesized image and the source image are consistent, even though the motions of the two images are different.
The entire process can be trained in an end-to-end manner, and finally our MAA model can effectively perform motion transfer across domains while also properly preserving the shape and appearance of the object (\emph{cf}\onedot} \def\Cf{\emph{C.f}\onedot Fig.~\ref{fig:figure-1} (b)).
We validate our proposed approach on two pairs of datasets: the human body datasets Mixamo-Video to Fashion-Video\cite{zablotskaia2019dwnet} and the human face datasets Vox-Celeb\cite{nagrani2017voxceleb} to Cufs\cite{wang2008face}. Extensive experimental results demonstrate the effectiveness of our proposed approach. Our source code will be released soon.
\section{Related Work}
\label{sec:Related Work}
\textbf{Motion Transfer}: Current motion transfer approaches can be categorized into two types: model-based and model-free approaches. The model-based approaches mainly focus on human body pose transfer\cite{ma2017pose,ma2018disentangled,balakrishnan2018synthesizing}, which utilize a pre-trained pose estimator or key point detector to extract the pose of driving image as a guidance information. And a number of researchers followed such setting\cite{liu2019liquid,ren2020deep,zhu2019progressive,li2019dense,huang2021few,kappel2021high}. Moreover, a series of works apply this model-based pattern on human facial expression transfer\cite{burkov2020neural,chen2020puppeteergan,gu2020flnet}. Like body pose transfer, they also employ a pre-trained facial landmark detector to model the facial expression.
The model-free approaches\cite{siarohin2019animating,siarohin2019first,jeon2020cross,siarohin2021motion,tao2022structure} does not rely on pre-trained third-party models, and extend the model-based method to arbitrary objects. Aliaksandr \emph{et al}\onedot\cite{siarohin2019animating} proposed a model-free motion transfer model Monky-Net that can apply motion transfer on arbitrary objects with an unsupervised key point detector trained by reconstruction loss \cite{jakab2018unsupervised}. Aliaksandr \emph{et al}\onedot\cite{siarohin2019first} further improved Monkey-Net to FOMM to solve the large motion problem. The unsupervised key point detector is also utilized in FOMM, with local affine transformations being added for motion modeling. A generator module is utilized to generate final result with the warped source image feature. Subin \emph{et al}\onedot\cite{jeon2020cross} proposed pose attention mechanism with an unsupervised key point detector to model motion. Recently, Aliaksandr \emph{et al}\onedot\cite{siarohin2021motion} improved FOMM with an advanced motion model and background motion model to MRAA. Although promising results are achieved for the single domain motion transfer, these methods might suffer from performance degradation when the source image and driving video come from different domains, where a considerable appearance difference often exists. Recently, Wang \emph{et al}\onedot\cite{wang2020self} used encoder based motion transfer approach which can be applied to the cross-domain scenario, and better results are achieved compared with the single domain motion transfer Monkey-Net model. In contrast, our proposed MAA approach is a general framework, and can integrate traditional motion transfer model like FOMM and MRAA to produce excellent results for large motion.
\textbf{Domain Adaptation:} Many works have been proposed to handle the scenario where the training and test data comes from different domains for different computer vision tasks , \emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, classification~\cite{DMCD}, semantic segmentation~\cite{Liu_2022_CVPR,Liu_2021_ICCV,9616392_Dong,What_Transferred_Dong_CVPR2020}, object detection~\cite{chen2018domain,deng2021unbiased}, pose estimation\cite{li2021synthetic,zhang2021keypoint}, \emph{etc}\onedot} \def\vs{\emph{vs}\onedot. A majority of works were developed to learn domain-invariant features using the domain adversarial learning~\cite{tzeng2017adversarial,ganin2015unsupervised}. Cross-domain motion transfer is more complicated, since we need to capture motion from the the driving video while preserving the appearance from the source domain. Nevertheless, the strategies proposed in traditional domain adaptation works might be useful to help motion transfer. For example, we apply the cyclic training pipeline inspired by CycleGAN~\cite{zhu2017unpaired}, and build our patch-based appearance consistency module based on Patch-GAN~\cite{isola2017image}.
\section{Methodology}
\label{sec:Methodology}
In this section, we present our Motion and Appearance Adaptation approach for cross-domain motion transfer. Formally, let us denote a driving video as $V_d = \{I_d^{i}|_{i=1}^T\}$, where each $I_d^i$ is a driving frame, while a source image is denoted as $I_s$; thus, the task of motion transfer is to synthesize a new video $\hat{V}_d = \{\hat{I}_d^{i}|_{i=1}^T\}$, where each $\hat{I}_d$ adequately captures the object motion in the corresponding driving frame $I_d^i$ while also preserving the object appearance of the source image $I_s$.
The appearance of an object roughly consists of two aspects, \emph{shape} and \emph{texture}. The shape largely refers to its geometric property (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, length, slimness, etc.), while the texture usually means how the object looks like regardless of its shape (\emph{e.g}\onedot} \def\Eg{\emph{E.g}\onedot, dresses with different colors). Traditional motion transfer methods generally assume that the driving frame and the source image are derived from the same domain, where they implicitly suppose the object shapes are similar. Consequently, when the source image is derived from a new domain with different object shapes, these methods often fail to preserve the shape of the object in the source image.
In this work, we study the cross-domain motion transfer problem, in which the source image and driving frame are from different domains. In other words, there might be considerable differences in appearance between them in terms of both shape and texture. An example is given in Fig.~\ref{fig:figure-1}, where both the clothes and body shapes of the fashion model and the Mixamo character exhibit notable differences.
In what follows, we first present an overview of the pipeline of our proposed MAA approach in \cref{sec:overview}, after which we present the shape-invariant motion adaptation (SIMA) module and structure-guided appearance consistency~(SGAC) module in \cref{sec:SIMA} and \cref{sec:SGAC} respectively; these effectively learning the motion and appearance from the driving frame and source image, respectively.
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{Figures/Fig2.pdf}
\caption{The pipeline of our proposed method. The left-hand side is the architecture of the traditional single domain motion transfer model FOMM\cite{siarohin2019first}, which is used as a basic motion transfer model in our approach. Moreover, the right-hand side is the framework of our proposed MAA method where we design a cyclic reconstruction loss (CYC), a shape-invariant motion adaptation (SIMA) and a structure-guided appearance consistency (SGAC) module}
\label{fig:fig-pipeline}
\end{figure*}
\subsection{Overview}
\label{sec:overview}
We design a cyclic training pipeline for cross-domain motion transfer, as shown in the right-hand part of \cref{fig:fig-pipeline}. The pipeline consists of a basic motion transfer model, our proposed shape-invariant motion adaptation module and structure-guided appearance consistency module, and a cyclic loss.
\textbf{Basic Motion Transfer Model:} The basic motion transfer (MT) model follows the traditional motion transfer model\cite{siarohin2019first,siarohin2021motion}. We here illustrate the basic MT model by taking FOMM\cite{siarohin2019first} as an example, and other models like\cite{siarohin2021motion} can be similarly integrated into our pipeline.
As shown in the left-hand part of \cref{fig:fig-pipeline}, traditional motion transfer methods typically employ a reconstruction training mode for learning and synthesizing motion. During the training phase, they select two arbitrary frames from the driving video as the source image and driving frame, which are used as input of the MT model. For each image, the motion keypoints and their local affine transformation are extracted using a motion estimator, where the motion keypoints can be conceptualized as the centroids of moving object parts. The dense motion flow from the source image to the driving frame can therefore be estimated using their motion keypoints and affine transformations. In the next step, the dense motion flow is used to warp the feature map of the source image, and produce the synthesized image $\hat{I}_d^i$ using the image generator. A perceptual loss is used as the reconstruction loss after the image generator to ensure that the synthesized image $\hat{I}_d^i$ fully reconstructs the driving frame $I_d^i$ as in\cite{siarohin2019first}:
\begin{equation}
\mathcal{L}_{r} = \sum_{l=k}^{K} | F_{l}(\hat{I}_d^i) - F_{l}(I_d^i)|
\label{eq:reconstruction loss}
\end{equation}
where $ F_{l}(\cdot) $ is feature map output by the $l$-th layer of a pre-trained VGG-19 network\cite{simonyan2014very}.
Researchers have proposed different method\cite{siarohin2021motion} to improve the motion estimator in order to more precisely extract motion information, yet the motion representation (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, motion keypoints and affine transformations) remains similar. In the interests of simplicity, we depict only the motion keypoints in \cref{fig:fig-pipeline}, which are related to our MAA approach. Readers can refer to\cite{siarohin2019first} for further details.
\textbf{Cyclic Training Pipeline:} In cross-domain motion transfer, the source image and driving video are obtained from different domains. So it is undesirable to pick a frame in the driving video as a source image and apply the reconstruction loss after the image generator, as the model will inevitably be overfitted to the driving video, which will lead to artifacts in the synthesized image.
To address this issue, we build a cyclic reconstruction framework inspired by the CycleGAN\cite{zhu2017unpaired} and cross-identity\cite{jeon2020cross}. As shown in the right-hand side of Fig~\ref{fig:fig-pipeline}, we employ two basic basic MT models that share the same parameters. Given a source image $I_s$ and a driving frame $I_d^{i}$, we first obtain a synthesized image $I_p$ using the basic MT model. Since there is no ground-truth for the synthesized image, the reconstruction loss cannot be used for $I_p$.
We then take the synthesized image $I_p$ as a driving frame, along with an arbitrary frame $I_d^{j}$ as the source image, and these are input again into the basic MT model to produce another synthesized image $I_c$. Intuitively, $I_c$ should mimic the motion of $I_p$, as well as $I_d^{i}$, since we expect $I_p$ to mimic the motion of $I_d^{i}$. At the same time, $I_c$ should also preserve the appearance of $I_d^{j}$, as well as $I_d^{i}$, which is derived from the same driving video as $I_d^{j}$. This allows us to employ $I_d^{i}$ and the cyclically generated $I_c$ to create a reconstruction loss for training. More specifically, we employ the perceptual loss
similarly as in \cref{eq:reconstruction loss}:
\begin{equation}
\mathcal{L}_{c} = \sum_{l=k}^{K} | F_{l}(I_c) - F_{l}(I_d^i)|
\label{eq:cyclic reconstruction loss}
\end{equation}
While the cyclic reconstruction loss enables us to train the motion transfer model in the cross-domain setting, this is only a weak supervision that cannot fully guarantee a satisfactory result. We therefore further introduce the shape-invariant motion adaptation module and patch-based appearance model to regularize the motion transfer process, which will be explained in more detail below.
\subsection{Shape-invariant Motion Adaptation}
\label{sec:SIMA}
\begin{figure}
\centering
\includegraphics[width=0.5\linewidth]{Figures/Fig3.pdf}
\caption{Illustration of our shape-invariant motion adaptation module. The top row show the structure topology, and the bottom two rows represents the motion adaptation stage using structured and unstructured keypoints}
\label{fig:skeleton}
\end{figure}
Due to the significant appearance difference between the source image $I_s$ and the driving frame $I_d$, the generated synthesized image $I_p$ often fails to adequately capture the object motion in the driving frame $I_d$. We therefore propose to directly regularize the object pose in $I_p$ with that in $I_d$ based on the extracted motion keypoints.
However, due to the diversity of the object shapes in $I_p$ and $I_d$, it is undesirable to directly regularize the consistency of their keypoint positions. We therefore propose to discover the intrinsic topology of the object, then regularize the included angles between adjacent object parts of two objects.
\textbf{Structure Topology Discovery:} To discover the intrinsic object topology, for each driving video, we employ a pre-trained basic MT model to extract the motion keypoints of all frames in the video. Because the motion keypoints roughly describe the objects' moving body parts, two keypoints can be considered to be adjacent if their distance does not change substantially between different frames.
Formally, given a driving frame $I_d$, we denote its motion keypoints as $ \mathcal{K}_d = \{{\k_{d}^{i}}|_{i=1}^K\}$, where $K$ is the number of motion keypoints. For each pair of keypoints $\k_{d}^{i}$ and $\k_{d}^j$, we calculate their $\ell_2$ distance $d_{i,j} = \ell_2(\k_{d}^i, \k_{d}^j)$, where $i \neq j$. The average distance across all frames of all driving videos can then be computed as $\bar{d}_{i,j} = \frac{1}{T}\sum_{t=1}^T d^{(t)}_{i,j}$, where $d^{(t)}_{i,j}$ is the distance in the $t$-th frame, while $T$ is the total number of video frames. Finally, we calculate the total distance diversity of $\k_{d}^{i}$ and $\k_{d}^j$ as follows:
\begin{equation}
\begin{split}
&v_{i,j} = \sum_{t=1}^T | d^{(t)}_{i,j} - \bar{d}_{i,j} |
\end{split}
\end{equation}
Intuitively, the distance diversity describes the stability of the connection between two keypoints. The smaller the distance diversity $v_{i,j}$, the higher the likelihood that the two keypoints will be adjacent. We then use the distance diversities to construct a structure topology graph $G$, where the nodes are keypoints, and the edges are defined according to the distance diversities. Specifically, we define the edge value as follows:
\begin{equation}
e_{i,j}=
\begin{cases}
\frac{(v_{i,j} - \eta)^{2}}{\eta^{2}}, & v_{i,j}<\eta,\\
0, & \mbox{otherwise}
\end{cases}
\end{equation}
where $\eta$ is a threshold, and we filter out the edges with high distance diversities, as these imply that the two keypoints are unlikely to be adjacent. Note that the edge value $e_{i,j}$ is within the range of $[0, 1]$. It can be seen as a measurement of the strength of the connection between two keypoints. We will demonstrate that it can also be used as a weight when we regularize the keypoints between driving frame and synthesized image.
Moreover, it is possible that not all keypoints are connected in a single graph; we select the largest graph as our structure topology graph $G$. We refer to the keypoints in $G$ as the structured keypoints and the others as unstructured keypoints. For improved convenience of presentation, we denote the set of structured keypoints as $\cS$ and their edges as $\cE$, the structure topology graph can be presented as $G = \{\cS, \cE\}$. For unstructured keypoints, we retain only the keypoints and discard their edges, since their connectivities are weak, and denote the set of unstructured keypoints as $\cU$. We present an illustration of the structure topology discovery in the top row of \cref{fig:skeleton}.
\textbf{Regularizing Structured Keypoints:}
Given a driving frame $I_d$ and the synthesized $I_p$, we extract their keypoints $\cK_d = \{\k_d\}$ and $\cK_p=\{\k_p\}$ using the basic MT model. To regularize the keypoints in the driving frame $I_d$ and the synthesized $I_p$, we instantiate the structure topology $G$ using the extracted keypoints $\cK_d$ and $\cK_p$, respectively. Taking the driving frame as an example, the instantiated graph is presented as $G_d = \{\cS_d, \cE_d\}$; here, $\cS_d$ is the set of structured keypoints in $I_d$, while $\cE_d$ is the set of corresponding edges which are calculated based on the Euclidean distances between keypoints. The instantiated graph of the synthesized image $G_p = \{\cS_d, \cE_d\}$ can be similarly defined. We illustrate the instantiated graphs in the top of Fig~\ref{fig:skeleton}.
When examining the structured keypoints, we can observe that considerable differences exist in terms of object shape; this validates our analysis that it is not preferable to directly regularize the keypoint positions. However, the pose can be portrayed as the included angle of each triplet of the connected keypoints in the structure graph.
Specifically, taking the driving frame as an example, let us define a triplet of connected keypoints as $\t_d = \{\k_d^{i}, \k_d^{j}, \k_d^{k}\}$, where both $\k_d^j$ and $\k_d^k$ are connected to $\k_d^{i}$. We further denote the set of all keypoint triplets in $G_d$ as $\cT_d = \{\t_d^{n}|_{n=1}^N\}$, where $N$ is the total number of triplets. Similarly, we define the set of triplets for the synthesized image as $\cT_p = \{\t_p^{n}|_{n=1}^N\}$.
For each triplet $\t_d^{n}$ (\emph{resp.}, $\t_p^{n}$ ), we calculate its included angle and denote it by $\alpha(\t_d^n)$ (\emph{resp.}, $\alpha(\t_p^n)$). We then regularize the consistency of the corresponding included angles for structured keypoints in the driving frame and the synthesized image as follows:
\begin{equation}
\begin{split}
\cL_{rs} = \frac{1}{N}\sum_{n=1}^N \gamma_n| \alpha(\t_{d}^n) - \alpha(\t_{p}^n) |
\end{split}
\end{equation}
where $\gamma_n$ is the weight for the $n$-th triplet. We calculate $\gamma_n$ using the edge values in the topology graph $G$. Specifically, given any triplet $\t = \{\k^{i}, \k^{j}, \k^{k}\}$ in the topology graph, the weight is computed as $\gamma = e_{i,j}e_{i,k}$. As the edge represents the strength of the connections between two keypoints, it is reasonable to employ the multiplication of the two edges that formed the included angle as the weight for regularization.
\textbf{Regularizing Unstructured Keypoints:} Similarly, given a driving frame $I_d$ and the synthesized $I_p$, we identify their unstructured keypoints $\cU_d$ and $\cU_p$, respectively. Since these unstructured keypoints are disjoint, we constrain them by encoding their included angles with the object centroid. Taking the driving frame as an example, for each pair of keypoints $(\k_d^i, \k_d^j)$ in $\cU_d$, we construct a triplet $\hat{\t}_d = (\k_d^i, \k_d^c, \k_d^j)$ in which $\k_d^c$ is the object centroid, and further denote the included angle as $\beta(\hat{\t}_d)$. We similarly define the corresponding included angle for the synthesized image as $\beta(\hat{\t}_p)$. We then regularize the consistency of the corresponding included angles for the structured keypoints in the driving frame and the synthesized image as follows:
\begin{equation}
\begin{split}
\cL_{ru} = \frac{1}{\hat{N}}\sum_{n=1}^{\hat{N}}|\beta(\hat{\t}_{d}^n) - \beta(\hat{\t}_p^n) |
\end{split}
\end{equation}
where $\hat{N}$ is the number of constructed triplets using structured keypoints in each image.
Combining the loss of structured and unstructured keypoints, the total loss of our shape-invariant motion adaptation loss can be written as follows:
\begin{equation}
L_{ma} = L_{rs} + L_{ru}
\end{equation}
\begin{figure}
\centering
\includegraphics[width=0.6\linewidth]{Figures/Fig4.pdf}
\caption{Illustration of our structure-guided appearance consistency module}
\label{fig:patch}
\end{figure}
\subsection{Structure-Guided Appearance Consistency Module}
\label{sec:SGAC}
We now consider how the appearance of the synthesized image $I_p$ might be enforced to be similar to that of the source image $I_s$. Note that the object poses in $I_p$ and $I_s$ are different, as we have enforced $I_p$ to mimic the pose of the driving frame. We therefore propose structure-guided appearance consistency module to regularize the appearance consistency of object parts in $I_p$ and $I_c$ to avoid impacting the learned object pose of $I_p$
In particular, we use the predicted motion keypoints to extract image patches of fixed size from both images. After collecting the patches from $ {I}_{p}$ (\emph{resp.}, $I_s$), a discriminator $\cD$ is then introduced to enforce the appearance consistency between the corresponding patches by means of an adversarial training strategy, as shown in \cref{fig:patch}. The aim of the discriminator is to determine whether the input patches are from ${I}_{p}$ or $I_s$ by minimizing a cross-entropy loss, while the generation model $\cB$ (\emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, the basic MT model) aims at generating pseudo-images $\cB(I_{s}) $, which are difficult to distinguish from the source image $ I_{s}$ by maximizing the cross-entropy loss. Formally, we express the loss of our patch-based appearance consistency module as follows:
\begin{equation}
\begin{split}
L_{ac} = & \log \cD(V(I_s)) +\log(1-\cD(V(\cB(I_s))))
\end{split}
\end{equation}
where $V(\cdot) $ represents the patch extraction operation.
\subsection{Summary}
We combine all losses together to train our proposed MAA model in an end-to-end manner. The overall objective function can be written as follows,
\begin{equation}
\cL = \cL_{r} + \cL_{c} + \lambda_{ma}\cL_{ma} - \lambda_{ac}L_{ac}
\end{equation}
where $\lambda_{ma}$ and $\lambda_{ac}$ are tradeoff parameters used to balance the losses. Due to the existence of the discriminator, we optimize the overall loss in an adversarial training manner, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot, $\min_{\cB}\max_{\cD}\cL$. Detailed training loop is presented in Supplementary materials.
\section{Experiment}
\label{sec:Experiment}
\subsection{Datasets}
We conduct experiments for two types of object including human body and human face. For the human body animation, we transfer motion from Mixamo-Video to Fashion-Video, and for human face animation we transfer motion from Vox-Celeb to CUHK Face Sketch~(Cufs).
\textbf{Mixamo-Video Dataset} is a synthetic human dancing video dataset newly constructed by ourselves. We collect $15$ characters of 3D human body models and $46$ dancing sequences from the mixamo~\cite{mixamo.com} website, then render dancing videos for these characters and dancing sequences, leading to $ 15\times 46 = 690 $ videos in total with resolution of $ 256\times 256 $. We split ten of the characters as training set and the rest as test set, \emph{i.e}\onedot} \def\Ie{\emph{I.e}\onedot $460$ and $230$ videos, respectively. Details of the dataset are presented in supplementary materials, and we will release the dataset soon.
\textbf{Fashion-Video Dataset} is a video dataset for showing clothes. It contains $500$ training videos and $100$ testing videos with size of $ 256\times 256 $.Although it is a video dataset, We use it as an image dataset by selecting one frame per video randomly in training stage.
\textbf{Vox-Celeb} is a video dataset of human talking. It consists of $12,331$ training videos and $444$ testing videos resized to $ 256\times 256 $.
\textbf{CUHK Face Sketch~(Cufs)} is an image dataset of human face sketches. The dataset contains $305$ images where training set and test set have $250$ and $45$ images \emph{resp.}. Each image is a sketch drawn by an artist based on a photo taken in a frontal pose with a natural expression. We also resize those images into the size of $ 256\times 256 $.
\begin{table}[]
\centering
\caption{ Quantity results comparison of our method with source only FOMM model and MRAA model. The lower FID and AED values are the better}
\begin{tabular}{c|cc|cc}
\hline
& \multicolumn{2}{l|}{Mixamo $\longrightarrow$ Fashion} & \multicolumn{2}{l}{Vox $\longrightarrow$ Cufs} \\
& FID $\downarrow$ & AED $\downarrow$ & FID $\downarrow$ & AED $\downarrow$ \\ \hline
MRAA & 177.3 & 0.376 & 127.1 & 0.764 \\ \hline
FOMM & 175.9 & 0.359 & 112.5 & 0.693 \\ \hline
Ours~(MRAA) & 72.1 & 0.289 & 86.5 & 0.627 \\ \hline
Ours~(FOMM) & \textbf{61.7} & \textbf{0.274} & \textbf{50.1} & \textbf{0.573} \\ \hline
\end{tabular}
\label{tab:FID-AED}
\end{table}
\subsection{Quantitative Results}
\textbf{Metrics:} As the ground-truth video are not available in cross-domain motion transfer, to quantitatively assess the synthesized videos, we employ two metrics for evaluate generative models as follows
\begin{itemize}
\item \textbf{Fr\'{e}chet Inception distance (FID)\cite{heusel2017gans}} This score indicates the overall quality of generated frames, it compares the feature statistics of generated frames and real images, then calculates the distance between them.
\item \textbf{Average Euclidean Distance (AED)}\cite{siarohin2019first} Considering the generated images share the same identity with source images, we utilize AED to evaluate the identity similarity between them. It also computes the feature distance between two input images. Specifically, a pre-trained person re-identification network~\cite{hermans2017defense} and a pre-trained facial identification network~\cite{amos2016openface} are used to extract identity feature representations for human body and human face dataset, respectively.
\end{itemize}
\textbf{Results:}
As aforementioned, unsupervised motion transfer models like FOMM\cite{siarohin2019first} or MRAA\cite{siarohin2021motion} can be integrated into our MAA framework as the basic motion transfer model. We conduct experiments by respectively using FOMM and MRAA as our basic motion transfer model, and take the original FOMM and MRAA as the corresponding baseline for comparison. For both methods, the baseline models are trained on the driving video dataset without considering the cross-domain issue. Note that the newly proposed modules in our MAA model are only used in training stage, and the model in the test stage share the same architecture with the baseline FOMM or MRAA model.
We report the results for Mixamo-Video $ \rightarrow $ Fashion-Video and Vox $ \rightarrow $ Cufs in~\cref{tab:FID-AED}. Comparing with the FOMM and MRAA model, our proposed MAA approach achieves a much better performance. In particular, compared with FOMM, we achieve a FID of $61.7$ and an AED of $0.274$ for Mixamo-Video $ \rightarrow $ Fashion-Video, and $50.1$ \vs $112.5$ and $0.573$ \vs $0.693$ for Vox $ \rightarrow $ Cufs, respectively. Compared with MRAA, we achieve a FID of $72.1$ and an AED of $0.289$ for Mixamo-Video $ \rightarrow $ Fashion-Video, and $86.5$ \vs $127.1$ and $0.627$ \vs $0.764$ for Vox $ \rightarrow $ Cufs, respectively. Note that, for both FID and AED metrics, smaller value is better. The large improvement indicates that the cross-domain motion transfer is challenging for the traditional FOMM and MRAA method, while our MAA model works well on the cross-domain scenario. We observe that MRAA performs worse than FOMM in the cross-domain motion transfer task, although the previous work shows MRAA usually performs better than FOMM in the traditional single-domain motion transfer\cite{siarohin2021motion}. This possibly dues to that the PCA based motion estimation in the MRAA method is non-parametric and less flexible for cross-domain motion transfer.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Figures/Fig5.pdf}
\caption{Visualization results of FOMM,MRAA and ours method on human body datasets}
\label{fig:body-result}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{Figures/Fig6.pdf}
\caption{Visualization results of FOMM,MRAA and ours method on human face datasets}
\label{fig:face-result}
\end{figure}
\subsection{Qualitative Results}
We visualize the generated results to gain an intuitive assessment of FOMM, MRAA and our MAA models for cross-domain human body and human face animation in \cref{fig:body-result} and \cref{fig:face-result}, respectively. In each figure, two pairs of results are visualized in the left and right parts, respectively. Driving frames extracted from the test video are displayed on the top row, while the source images are showed at the most left column of each part.
For human body animation, as shown in \cref{fig:body-result}, the results generated by the FOMM and MRAA model obviously suffer from domain shift problem. Although the motion of driving video is roughly captured, the human body shape of source image is rarely preserved, and notable artifacts can be observed in almost all frames of the synthesized video. In contrast, our MAA model is able to capture the motion of the driving frames while properly preserving the appearance of the source image.
For human face animation, as shown in \cref{fig:face-result}, the FOMM and MRAA model could generate results with a rough motion of driving frames and a similar facial appearance with source image. However, the quality of synthesized image are not satisfactory where artifacts are obvious to observe. For example, artifacts on glasses and heads can be observed for FOMM results as highlighted in the red bounding boxes. These differences in qualitative results clearly demonstrate the effectiveness of our proposed MAA model for cross-domain motion transfer.
\subsection{Ablation Study}
To study the impact of our proposed modules, we further conduct ablation study on both human body and human face datasets. The FOMM is used as the basic motion transfer model. The quantitative results are shown in \cref{tab:ablation}, where 'w/o CYC', 'w/o SIMA' and 'w/o SGAC' means removing the cyclic training pipeline, shape-invariant motion adaptation and structure-guided appearance consistency of FOMM model, respectively.
\begin{table}[]
\centering
\caption{Ablation results comparison of FOMM and our ablated models.}
\begin{tabular}{c|cc|cc}
\hline
& \multicolumn{2}{l|}{Mixamo $\longrightarrow$ Fashion} & \multicolumn{2}{l}{Vox $\longrightarrow$ Cufs} \\
& FID $\downarrow$ & AED $\downarrow$ & FID $\downarrow$ & AED $\downarrow$ \\ \hline
FOMM & 175.9 & 0.359 & 112.5 & 0.693 \\ \hline
w/o CYC & 136.9 & 0.354 & 74.1 & 0.633 \\ \hline
w/o SIMA & 80.2 & 0.303 & 60.7 & 0.622 \\ \hline
w/o SGAC & 67.7 & 0.284 & 55.2 & 0.603 \\ \hline
Ours~(FOMM) & \textbf{61.7} & \textbf{0.274} & \textbf{50.1} & \textbf{0.573} \\ \hline
\end{tabular}
\label{tab:ablation}
\end{table}
For both human body and human face animation, as shown in \cref{tab:ablation}, we observe considerable performance drops on both AED and FID for w/o CYC, which again confirms the importance to explicitly consider the cross-domain issue when performing motion transfer across domains. Similar observations can be obtained on human face dataset, which is also confirmed in the qualitative results in \cref{fig:ablation}. Other ablation settings w/o SIMA and w/o SGAC also degrade the performance considerably, which validates the necessity of using the two modules for generating satisfactory synthesized video in cross-domain motion transfer.
To show the effect of each module intuitively, we further visualize the synthesized results in \cref{fig:ablation}. We observe that the result of w/o CYC has richer details than that of FOMM model. For example, the face and the clothes are clearer. However, compared with our final MAA result, it still drops important motion and appearance information. Moreover, we observe the result of w/o SIMA are able to preserve relative rich appearance information, however, the pose of driving frames are not transferred properly without the help of motion consistency module. For example, artifacts can be observed for the poses of the arms and heads as highlighted in the blue rectangles. And, on the third row, w/o~SGAC performs well in pose transferring but fails to preserve source image appearance without SGAC, especially for the details of human face as highlighted in red rectangles. These observations confirm the effectiveness of the modules proposed in our MAA approach.
\begin{figure}
\centering
\includegraphics[width=0.45\linewidth]{Figures/Fig7.pdf}
\caption{Visualized ablation study results on the human body datasets}
\label{fig:ablation}
\end{figure}
\subsection{User Study}
To further evaluate our model, we additionally conduct a user study. In particular, we randomly select 50 pairs of source domain driving videos and target domain source images for both human body animation and human face animation, and generate result videos in an ablation setting. For each dataset, we compare results of our final MAA model with those of FOMM and three ablation methods, respectively. The comparison are evaluated by 25 users according to three aspects, motion, appearance~(denoted as app. in \cref{tab:User-Study}) and overall, respectively.
The user preferences are shown in \cref{tab:User-Study}. We observe that all scores are above $0.5$, which means our results are preferred by the majority of users for all aspects in all settings. For motion aspect, fewer users prefer w/o SIMA than other settings when compared with MAA model on both datasets ($0.748$ vs. $0.717$ and $0.679$ for human body, and $0.571$ vs. $0.704$ for human face), which indicates SIMA improves the motion of generated results. For appearance aspect, fewer user prefer w/o SGAC than other ablation settings when compared with MAA model in human body dataset ($0.715$ vs. $0.711$ and $0.699$), which indicates SGAC contributes to appearance of generated results.
\begin{table}[]
\centering
\caption{User study results. We compare the Ours (FOMM) model to every ablation model, and the values represent the user preferences to Ours (FOMM) model}
\begin{tabular}{c|ccc|ccc}
\hline
& \multicolumn{3}{c|}{Mixamo $\longrightarrow$ Fashion} & \multicolumn{3}{c}{Vox $\longrightarrow$ Cufs} \\
& motion & appearance & overall & motion & appearance & overall \\ \hline
FOMM & 0.888 & 0.983 & 0.978 & 0.845 & 0.792 & 0.875 \\ \hline
w/o CYC & 0.717 & 0.699 & 0.732 & 0.571 & 0.615 & 0.626 \\ \hline
w/o SIMA & 0.748 & 0.711 & 0.702 & 0.704 & 0.655 & 0.675 \\ \hline
w/o SGAC & 0.679 & 0.715 & 0.725 & 0.593 & 0.617 & 0.575 \\ \hline
\end{tabular}
\label{tab:User-Study}
\end{table}
\section{Conclusion}
\label{sec:Conclusion}
In this paper, we propose a Motion and Appearance Adaptation~(MAA) approach for cross-domain motion transfer. In MAA, we design a shape-invariant motion adaptation module to enforce the consistency of the angles of object parts in two images to capture the motion information. Meanwhile, we introduce a structure-guided appearance consistency module to regularize the similarity between the patches of the synthesized image and the source image. The experimental results demonstrates the effectiveness of our proposed method.
\begin{FlushLeft}\textbf{Acknowledgement}. This work is supported by the Major Project for New Generation of AI under Grant No. 2018AAA0100400, the National Natural Science Foundation of China (Grant No. 62176047), Sichuan Science and Technology Program (No. 2021YFS0374, 2022YFS0600), Beijing Natural Science Foundation (Z190023), and Alibaba Group through Alibaba Innovation Research Program. This work is also partially supported by the Science and Technology on Electronic Information Control Laboratory.\end{FlushLeft}
\bibliographystyle{splncs04}
| da56fe47e094c85d6aded2965d40795d608a9ffe | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Sampling from a probability distribution given its score function, i.e., the gradient of the log-density, is an active area of research in machine learning. Its applications range far and wide, from Bayesian learning \citep{welling2011} to learning energy-based models \citep{song2021ebm}, synthesizing new high-quality data \citep{dhariwal2021}, and so on. Typical examples of traditional score-based samplers are Markov chain Monte Carlo (MCMC) methods such as Langevin dynamics \citep{langevin1908} and Hamiltonian Monte Carlo \citep{neal2011}.
Recent developments in score matching with deep neural networks (DNNs) have made it possible to estimate scores of high-dimensional distributions such as those of natural images \citep{song2019score}. However, natural data distributions are often sharp and multi-modal, rendering na\"{i}ve application of traditional MCMC methods impractical. Specifically, MCMC methods tend to skip over or get stuck at local high-density modes, producing biased samples \citep{levy2018}.
Diffusion models \citep{dickstein2015,ho2020,song2021ddim} depart from MCMC and use the concept of diffusion, the process of gradually corrupting data into noise, to generate samples. \citet{song2021} observed that for each diffusion process, there is a reverse stochastic differential equation (SDE) and an ordinary differential equation (ODE). Hence, given a noise sample, integrating the reverse-S/ODE produces a data sample. Only a time-dependent score function of the data during the diffusion process is required to simulate the reverse process.
This discovery generated great interest in finding better ways to integrate reverse-S/ODEs. For instance, \citet{song2021} uses black-box ODE solvers with adaptive stepsizes to accelerate sampling. Furthermore, multitude of recent works on score-based generative modeling focus on improving reverse-S/ODE integrators \citep{martineau2021,lu2022,karras2022,zhang2022}.
In this work, we develop an orthogonal approach to accelerating score-based sampling. Specifically, we propose Denoising MCMC (DMCMC) which combines MCMC with reverse-S/ODE integrators. MCMC is used to generate samples $\{(\bm{x}_n,t_n)\}$ in the product space of data $\bm{x}$ and variance exploding (VE) diffusion time $t$ / noise level $\sigma$ (see Fig.~\ref{fig:dmcmc_example} top panel). Since all modes are connected in the product space, MCMC mixes well. Then, a reverse-S/ODE integrator solves the reverse-S/ODE starting at $\bm{x}_n$ from time $t = t_n$ to $t = 0$. Since MCMC explores high-density regions, the MCMC chain stays close to the data manifold, so $t_n$ tends to be close to $0$, i.e., noise level tends to be small (see Fig.~\ref{fig:dmcmc_example} top and bottom panels). Thus, integrating the reverse-S/ODE from $t = t_n$ to $t = 0$ is much faster than integrating the reverse-S/ODE from maximum time $t = T$ to $t = 0$ starting from noise. This leads to a significant acceleration of the sampling process.
\begin{figure}[t]
\centering
\begin{subfigure}{0.9\linewidth}
\includegraphics[width=1.0\linewidth]{figures/diag.png}
\end{subfigure}
\begin{subfigure}{0.49\linewidth}
\includegraphics[width=1.0\linewidth]{figures/1a-min.png}
\end{subfigure}
\hfill
\begin{subfigure}{0.49\linewidth}
\includegraphics[width=1.0\linewidth]{figures/1b-min.png}
\end{subfigure}
\caption{\textbf{Top:} a conceptual illustration of a VE diffusion model sampling process and DMCMC sampling process. VE diffusion models integrate the reverse-S/ODE starting from maximum diffusion time / maximum noise level. So, samples are often noisy with small computation budget due to large truncation error. DMCMC produces an MCMC chain which travels close to the image manifold (compare the noise level $\sigma$). So, the MCMC samples can be denoised to produce high-quality data with relatively little computation. \textbf{Bottom:}
Visualization of sampling processes without (left) and with (right) DMCMC on CelebA-HQ-256 under a fixed computation budget.}
\label{fig:dmcmc_example}
\end{figure}
Our contributions can be summarized as follows.
\begin{itemize}
\item We introduce the product space of data and diffusion time, and develop a novel score-based sampling framework called Denoising MCMC on the product space. Our framework is general, as any MCMC, any VE process noise-conditional score function, and any reverse-S/ODE integrator can be used in a plug-and-play manner.
\item We develop Denoising Langevin Gibbs (DLG), which is an instance of Denoising MCMC that is simple to implement and is scalable. The MCMC part of DLG alternates between a data update step with Langevin dynamics and a noise level prediction step, so all that DLG requires is a pre-trained noise-conditional score network and a noise level classifier.
\item We verify the effectiveness of DLG by accelerating six reverse-S/ODE integrators. Notably, combined with the integrators of \citet{karras2022}, DLG achieves state-of-the-art results. On CIFAR10 in the limited number of score function evaluation (NFE) setting, we obtain $3.86$ FID with $\approx 10$ NFE and $2.63$ FID with $\approx 20$ NFE. On CelebA-HQ-256, we have $6.99$ FID with $\approx 160$ NFE, which is currently the best result with score-based models. The computation cost of evaluating a noise level classifier is negligible, so we obtain acceleration essentially for free.
\end{itemize}
\section{Background}
\subsection{Denoising Score Matching}
Given a distribution $p(\bm{x})$, a noise level $\sigma$, and a perturbation kernel $p_\sigma(\bm{x}\mid\tilde{\bm{x}}) = \mathcal{N}(\bm{x} \mid \tilde{\bm{x}}, \sigma^2 \bm{I})$, solving the denoising score matching objective \citep{vincent2011}
\begin{align}
\min_\theta \mathbb{E}_{p(\tilde{\bm{x}})} \mathbb{E}_{p_\sigma(\bm{x} \mid \tilde{\bm{x}})} \left[ \| {\boldsymbol s}_\theta(\bm{x}) - \nabla_{\bm{x}} \log p_\sigma(\bm{x} \mid \tilde{\bm{x}}) \|_2^2 \right] \label{eq:dsm_orig}
\end{align}
yields a score model ${\boldsymbol s}_\theta(\bm{x})$ which approximates the score of $\int p_\sigma(\bm{x} \mid \tilde{\bm{x}}) p(\tilde{\bm{x}}) \, d\tilde{\bm{x}}$.
Denoising score matching was then extended to train Noise Conditional Score Networks (NCSNs) ${\boldsymbol s}_\theta(\bm{x}, \sigma)$ which approximate the score of data smoothed at a general set of noise levels by solving
\begin{align}
\min_\theta \mathbb{E}_{\lambda(\sigma)} \mathbb{E}_{p(\tilde{\bm{x}})} \mathbb{E}_{p_{\sigma}(\bm{x} \mid \tilde{\bm{x}})} \left[ \| {\boldsymbol s}_\theta(\bm{x},\sigma) - \nabla_{\bm{x}} \log p_{\sigma}(\bm{x} \mid \tilde{\bm{x}}) \|_2^2 \right] \label{eq:dsm_cont}
\end{align}
where $\lambda(\sigma)$ can be a discrete or a continuous distribution over $(\sigma_{\min},\sigma_{\max})$ \citep{song2019score,song2021}. We note $\int p_\sigma(\bm{x} \mid \tilde{\bm{x}}) p(\tilde{\bm{x}}) \, d\tilde{\bm{x}}$ approaches $p(\bm{x})$ as $\sigma \rightarrow 0$, since the perturbation kernel $p_\sigma(\bm{x} \mid \tilde{\bm{x}})$ converges to the Dirac delta function centered at $\bm{x}$.
\subsection{Markov Chain Monte Carlo (MCMC)}
Given an unnormalized version of $p(\bm{x})$ or the score function $\nabla_{\bm{x}} \log p(\bm{x})$, MCMC constructs a Markov chain in the data space whose stationary distribution is $p(\bm{x})$. An MCMC which uses the unnormalized density is the Metropolis-Hastings MCMC \citep{metropolis1953,hastings1970} that builds a Markov chain by sequentially accepting or rejecting proposal distribution samples according to a density ratio. A popular score-based MCMC is Langevin dynamics \citep{langevin1908}. Langevin dynamics generates a Markov Chain $\{\bm{x}_n\}_{n = 1}^\infty$ using the iteration
\begin{align}
\bm{x}_{n+1} = \bm{x}_n + (\eta / 2) \cdot \nabla_{\bm{x}} \log p(\bm{x}_n) + \sqrt{\eta} \cdot \bm{\epsilon} \label{eq:langevin}
\end{align}
where ${\boldsymbol \epsilon} \sim \mathcal{N}(\bm{0},\bm{I})$. $\{\bm{x}_n\}_{n = 1}^\infty$ converges to $p(\bm{x})$ in distribution for an appropriate choice of $\eta$.
To sample from a joint distribution $p(\bm{x},\bm{y})$, we may resort to Gibbs sampling \citep{geman1984}. Given a current Markov chain state $(\bm{x}_n,\bm{y}_n)$, Gibbs sampling produces $\bm{x}_{n+1}$ by sampling from $p(\bm{x} \mid \bm{y}_n)$ and $\bm{y}_{n+1}$ by sampling from $p(\bm{y} \mid \bm{x}_{n+1})$. The sampling steps may be replaced with MCMC. Hence, Gibbs sampling is useful when conditional distributions are amenable to MCMC.
\textbf{Annealed MCMC.} Despite their diversity, MCMC methods often have difficulty crossing low-density regions in high-dimensional multimodal distributions. For Langevin dynamics, at a low-density region, the score function vanishes in \eqref{eq:langevin}, resulting in a meaningless diffusion. Moreover, natural data often lies on a low-dimensional manifold. Thus, once Langevin dynamics leaves the data manifold, it becomes impossible for Langevin dynamics to find its way back.
One way to remedy this problem is to use annealing, i.e., constructing a sequence of increasingly smooth and wide distributions and running MCMC at different levels of smoothness. As smoothness is increased, disjoint modes merge, so MCMC can cross over to other modes. Annealing has been used to empower various types of MCMC \citep{geyer1995,neal2001}. In this work, we shall refer to the collection of MCMC that use annealing as annealed MCMC.
An instance of annealed MCMC is annealed Langevin dynamics (ALD) \citep{song2019score}. For a sequence of increasing noise levels $\{\sigma_i\}_{i = 1}^N$, Langevin dynamics is sequentially executed with $\int p_{\sigma_i}(\bm{x} \mid \tilde{\bm{x}}) p(\tilde{\bm{x}}) \, d\tilde{\bm{x}}$ in place of $p(\bm{x})$ in \eqref{eq:langevin} for $i = N, N - 1, \ldots, 1$. Since $p(\bm{x})$ smoothed at a large noise level has wide support and connected modes, ALD overcomes the pitfalls of vanilla Langevin dynamics. However, ALD has the drawback that thousands of iterations are required to produce a single batch of samples.
\subsection{Diffusion Models}
\textbf{Diffusion models and differential equations.} Diffusion models opened up a new avenue towards fast sampling with score functions via SDEs and ODEs \citep{song2021}. Suppose data is distributed in $\mathbb{R}^d$. Given a diffusion process of data sample $\bm{x}_0 \sim p(\bm{x})$ into a sample from a simple prior noise distribution, the trajectory of data during diffusion can be described with an It\^{o} SDE
\begin{align}
d\bm{x} = \bm{f}(\bm{x},t) \, dt + g(t) \, d\bm{w} \label{eq:SDE}
\end{align}
for some drift coefficient $\bm{f} : \mathbb{R}^d \times [0,T] \rightarrow \mathbb{R}^d$, diffusion coefficient $g : [0,T] \rightarrow \mathbb{R}$, and Brownian motion $\bm{w}$. Here, $T$ is the diffusion termination time. With initial condition $\bm{x}(0) = \bm{x}_0$, integrating \eqref{eq:SDE} from time $t = 0$ to $t = T$ produces a sample from the prior distribution.
For each diffusion SDE, there exists a corresponding reverse-SDE:
\begin{align}
d\bm{x} = [\bm{f}(\bm{x},t) - g(t)^2 \nabla_{\bm{x}} \log p_t(\bm{x})] \, dt + g(t) \, d\bar{\bm{w}} \label{eq:RSDE}
\end{align}
where $p_t(\bm{x})$ is the density of $\bm{x}(t)$ evolving according to \eqref{eq:SDE} and $\bar{\bm{w}}$ is a Brownian motion if time flows from $t = T$ to $t = 0$. Given a sample $\bm{x}_T$ from the prior distribution, integrating \eqref{eq:RSDE} with initial condition $\bm{x}(T) = \bm{x}_T$ from $t = T$ to $t = 0$ results in a sample from $p(\bm{x})$. Moreover, to each reverse-SDE, there exists a corresponding deterministic reverse-ODE
\begin{align}
d\bm{x} = \left[\bm{f}(\bm{x},t) - (1/2) \cdot g(t)^2 \nabla_{\bm{x}} \log p_t(\bm{x}) \right] \, dt
\end{align}
which also can be integrated from $t = T$ to $t = 0$ to produce samples from $p(\bm{x})$.
Diffusion models generate data by simulating the reverse of the diffusion process, i.e., by solving the reverse-S/ODE of the diffusion process. Initial works on diffusion models \citep{dickstein2015,ho2020} used computationally expensive ancestral sampling to solve the reverse differential equations. Later works discovered that using adaptive numerical integrators to solve the reverse-S/ODE could accelerate the sampling process. This led to great attention on developing better reverse-S/ODE integrators \citep{martineau2021,song2021,lu2022,karras2022,zhang2022}. Our work is orthogonal to such works as focus on finding good initialization points for integration via MCMC. Hence, a better integration technique directly translates to even better generative performance when plugged into Denoising MCMC.
\textbf{Variance exploding (VE) diffusion model.} A VE diffusion model considers the diffusion process
\begin{align}
d\bm{x} = \sqrt{\frac{d[\sigma^2(t)]}{dt}} \, d\bm{w}. \label{eq:VESDE}
\end{align}
where $\sigma(t)$ increases monotonically with $t$ from $\sigma_{\min}$ to $\sigma_{\max}$. The data distribution evolves as
\begin{align}
\textstyle p_t(\bm{x}) = \int p_{\sigma(t)}(\bm{x} \mid \tilde{\bm{x}}) p(\tilde{\bm{x}}) \, d\tilde{\bm{x}}
\end{align}
so if $\sigma_{\min}$ is sufficiently small, $p_0(\bm{x}) \approx p(\bm{x})$, and if $\sigma_{\max}$ is sufficiently large, so variance explodes, $p_T(\bm{x}) \approx \mathcal{N}(\bm{x} \mid \bm{0}, \sigma_{\max}^2 \bm{I})$. If we have a score model ${\boldsymbol s}_\theta(\bm{x},\sigma)$ trained with \eqref{eq:dsm_cont}, $\nabla_{\bm{x}} \log p_t(\bm{x}) \approx {\boldsymbol s}_\theta(\bm{x}, \sigma(t))$. It follows that with $\bm{x}_T \sim \mathcal{N}(\bm{x} \mid \bm{0}, \sigma_{\max}^2 \bm{I})$, we may integrate the reverse-S/ODE corresponding to \eqref{eq:VESDE} with $\bm{x}(T) = \bm{x}_T$ from $t = T$ to $t = 0$ using a score model to generate data.
In the next section, we bridge MCMC and reverse-S/ODE integrators with VE diffusion to form a novel sampling framework that improves both MCMC and diffusion models.
\section{Denoising Markov Chain Monte Carlo (DMCMC)} \label{sec:dmcmc}
From here on, we denote the data distribution as $p(\bm{x})$ and its domain as $\mathcal{X} \subseteq \mathbb{R}^d$. Diffusion runs from time $t = 0$ to $t = T$, and noise scale $\sigma(t)$ increases monotonically from $\sigma_{\min}$ to $\sigma_{\max}$, such that $\sigma(0) = \sigma_{\min}$ and $\sigma(T) = \sigma_{\max}$. We denote the range of $\sigma(t)$ as $\SS = [\sigma_{\min},\sigma_{\max}]$. The Gaussian perturbation kernel is denoted as $p_\sigma(\bm{x} \mid \tilde{\bm{x}}) = \mathcal{N}(\bm{x} \mid \tilde{\bm{x}}, \sigma^2 \bm{I})$. The distribution of $\bm{x}(t)$ following the VE diffusion \eqref{eq:VESDE} is denoted as $p_t(\bm{x})$, and recall that $p_t(\bm{x}) = \int p_{\sigma(t)}(\bm{x} \mid \tilde{\bm{x}}) p(\tilde{\bm{x}}) \, d\tilde{\bm{x}}$.
We now develop a general framework called Denoising MCMC (DMCMC) which combines MCMC with reverse-S/ODE integrators. The construction of DMCMC is comprised of two steps. In the first step, we build MCMC on the product space $\mathcal{X} \times \SS$, i.e., $\mathcal{X}$ augmented by the smoothness parameter $\sigma$. Since $\sigma(t)$ is a monotone increasing function, this is equivalent to augmenting the data space with diffusion time $t$. In the second step, we incorporate denoising steps, where we denoise MCMC samples via reverse-S/ODE integrators.
\subsection{Construction Step 1: MCMC on the Product Space $\mathcal{X} \times \mathcal{S}$}
Suppose $p(\bm{x})$ is a high-dimensional multimodal distribution, supported on a low-dimensional manifold. If the modes are separated by wide low-density regions, MCMC can have difficulty moving between the modes. Indeed, convergence time for such distributions can grow exponential in dimension $d$ \citep{roberts2001}. Intuitively, for MCMC to move between disjoint modes, the Markov Chain would have to step off the data manifold. However, once MCMC leaves the data manifold, the density or the score vanishes. Then, most random directions produced by the proposal distribution do not point to the manifold. Thus, MCMC gets lost in the ambient space, whose volume grows exponentially in $d$.
Annealing via Gaussian smoothing, used in both ALD and VE diffusion, circumvents this problem. As $p(\bm{x})$ smoothed with perturbation kernel $p_\sigma(\bm{x} \mid \tilde{\bm{x}})$ of increasing $\sigma$, the modes of $p(\bm{x})$ grow wider and start to connect. Thus, MCMC can easily transition between modes. However, running MCMC in the manner of ALD is inefficient since we do not know how many iterations within each noise level is sufficient. To address this problem, we propose to augment $\mathcal{X}$ with the smoothness scale $\sigma$ and run MCMC in the product space $\mathcal{X} \times \SS$ such that MCMC automatically controls the value of $\sigma$. Below, we formally describe MCMC on $\mathcal{X} \times \SS$.
Let us define the $\sigma$-conditional distribution
\begin{align}
\textstyle \hat{p}(\bm{x} \mid \sigma) \coloneqq \int p_\sigma(\bm{x} \mid \tilde{\bm{x}}) p(\tilde{\bm{x}}) \, d\bm{\tilde{x}}. \label{eq:cond}
\end{align}
We also define a prior $\hat{p}(\sigma)$ on $\SS$. Then by the Bayes' Rule,
\begin{align}
\hat{p}(\bm{x},\sigma) = \hat{p}(\bm{x} \mid \sigma) \cdot \hat{p}(\sigma).
\end{align}
Here, $\hat{p}(\sigma)$ reflects our preference for how much time we want the MCMC chain to stay at a particular level of $\sigma$. MCMC with $\hat{p}(\bm{x},\sigma)$ will produce samples $\{(\bm{x}_n,\sigma_n)\}$ in $\mathcal{X} \times \SS$ such that
\begin{align}
\sigma_n \sim \hat{p}(\sigma), \qquad \bm{x}_n \sim \hat{p}(\bm{x} \mid \sigma_n). \label{eq:sample_dist}
\end{align}
Hence, if $\sigma_n \gg \sigma_{\min}$, $\bm{x}_n$ will be a noisy sample, i.e., a sample corrupted with Gaussian noise of variance $\sigma_n^2$, and if $\sigma_n \approx \sigma_{\min}$, $\bm{x}_n$ will resemble a sample from $p(\bm{x})$.
Since our goal is to generate samples from $p(\bm{x})$, na\"{i}vely, we can keep samples $(\bm{x}_n,\sigma_n)$ with $\sigma_n \approx \sigma_{\min}$ and discard other samples. However, this could lead to a large waste of computation resources. In the next section, we incorporate reverse-S/ODE integrators to avert this problem.
\subsection{Construction Step 2: Incorporating Denoising Steps}
Let us recall that integrating the reverse-S/ODE for the VE diffusion SDE Eq. (\ref{eq:VESDE}) from time $t = T$ to $t = 0$ sends samples from $p_T(\bm{x})$ to samples from $p_0(\bm{x}) \approx p(\bm{x})$. In general, integrating the reverse-SDE or ODE from time $t = t_2$ to $t = t_1$ for $t_1 < t_2$ sends samples from $p_{t_2}(\bm{x})$ to samples from $p_{t_1}(\bm{x})$ \citep{song2021}. We use this fact to denoise MCMC samples from $\hat{p}(\bm{x},\sigma)$.
Suppose we are given a sample $(\bm{x}_n,\sigma_n) \sim \hat{p}(\bm{x},\sigma)$. With $t_n \coloneqq \sigma^{-1}(\sigma_n)$, \eqref{eq:sample_dist} tells us
\begin{align}
\bm{x}_n \sim p_{t_n}(\bm{x})
\end{align}
so integrating the reverse-S/ODE with initial condition $\bm{x}(t_n) = \bm{x}_n$ from $t = t_n$ to $t = 0$ produces a sample from $p_0(\bm{x}) \approx p(\bm{x})$. Here, we note that any reverse-S/ODE solver may be used to carry out the integration.
Given an MCMC chain $\{(\bm{x}_n,\sigma_n)\}$ in $\mathcal{X} \times \SS$, MCMC is biased towards high-density regions of $\mathcal{X}$, so the sequence $\{\bm{x}_n\}$ will generally stay close to the data manifold, except when traversing between disjoint modes. This means $\sigma_n \ll \sigma_{\max}$ for most $n$, or in other words, $t_n \ll T$ for most $n$. So, the average length of integration intervals will tend to be much shorter than $T$. Thus, a numerical integrator can integrate the reverse-S/ODE with less truncation error given the same computation budget \citep{numanalysis}. Equivalently, less computation budget is required to reach the same truncation error level. This idea is illustrated in Figure \ref{fig:dmcmc_example}.
\section{Denoising Langevin Gibbs (DLG)}
In Section \ref{sec:dmcmc}, we described an abstract framework, DMCMC, for accelerating score-based sampling by combining MCMC and reverse-S/ODE integrators. We now develop a concrete instance of DMCMC. As the second construction step of DMCMC is simple, we only describe the first step.
Na\"{i}vely, we could extend denoising score matching \eqref{eq:dsm_orig} to estimate the score $\hat{s}_\theta(\bm{x},\sigma) : \mathbb{R}^d\times\mathbb{R} \rightarrow \mathbb{R}^d\times\mathbb{R}$ of $\hat{p}(\bm{x},\sigma)$ and apply Langevin dynamics in the first step of DMCMC. But, this would prevent us from using pre-trained score models, as we would have to solve (for some small $\nu > 0$)
\begin{align}
\min_\theta \mathbb{E}_{\bm{\epsilon} \sim \mathcal{N}(\bm{0}_{d+1},\nu^2 \bm{I}_{d+1})} \mathbb{E}_{\hat{p}(\bm{x}, \sigma)} [\|\hat{s}_\theta(\bm{x}-\epsilon_{1:d},\sigma-\epsilon_{d+1}) - \bm{\epsilon}/\nu^2\|_2^2]
\end{align}
which differs from \eqref{eq:dsm_cont}. Gibbs sampling provides a simple path around this problem. Let us recall that given a previous MCMC iterate $(\bm{x}_n,\sigma_n)$, Gibbs sampling proceeds by alternating between an $\bm{x}$ update step $\bm{x}_{n+1} \sim \hat{p}(\bm{x} \mid \sigma_n)$ and a $\sigma$ update step $\sigma_{n+1} \sim \hat{p}(\sigma \mid \bm{x}_{n+1})$. Below, we describe our score-based sampling algorithm, Denoising Langevin Gibbs (DLG).
\textbf{Updating $\bm{x}$.} Suppose we are given an MCMC iterate $(\bm{x}_n,\sigma_n)$ and a score model $s_\theta(\bm{x},\sigma)$ from \eqref{eq:dsm_cont}. We generate $\bm{x}_{n+1}$ by a Langevin dynamics step on $\hat{p}(\bm{x} \mid \sigma_n)$. Specifically, by \eqref{eq:cond},
\begin{align}
\nabla_{\bm{x}} \log \hat{p}(\bm{x} \mid \sigma_n) \approx s_\theta(\bm{x}, \sigma_n) \label{eq:dlg_lang}
\end{align}
and so an Langevin dynamics update on $\bm{x}$, according to \eqref{eq:langevin} is
\begin{align}
\bm{x}_{n+1} = \bm{x}_n + (\eta/2) \cdot s_\theta(\bm{x}_n,\sigma_n) + \sqrt{\eta} \cdot \bm{\epsilon}
\end{align}
for $\bm{\epsilon} \sim \mathcal{N}(\bm{0},\bm{I})$. Here, we call $\eta$ the step size.
\textbf{Updating $\sigma$.} We now have $\bm{x}_{n+1}$ and need to sample $\sigma_{n+1} \sim \hat{p}(\sigma \mid \bm{x}_{n+1})$. To this end, we first train a DNN noise level classifier $q_{\phi}(\sigma \mid \bm{x})$ to approximate $\hat{p}(\sigma \mid \bm{x})$ by solving
\begin{align}
\max_{\phi} \mathbb{E}_{\hat{p}(\bm{x},\sigma)} [\log q_{\phi}(\sigma \mid \bm{x})].
\end{align}
Specifically, we discretize $[\sigma_{\min},\sigma_{\max}]$ into $M$ levels $\tau_1 = \sigma_{\min} < \tau_2 < \cdots < \tau_M = \sigma_{\max}$. Given $\tau_m$ where $1 \leq m \leq M$, $m$ serves as the label and clean training data corrupted by Gaussian noise of variance $\tau_m^2$ serves as the classifier input. The classifier is trained to predict $m$ by minimizing the cross entropy loss. Having trained a noise level classifier, we sample $\sigma_{n+1}$ by drawing an index $m$ according to the classifier output probability for $\bm{x}_{n+1}$ and setting $\sigma_{n+1} = \tau_m$. In practice, using the index of largest probability worked fine. We denote this process as $\sigma_{n+1} \sim q_{\phi}(\sigma \mid \bm{x}_{n+1})$.
\subsection{Practical Considerations}
\textbf{Computation cost of $\sigma$ prediction.} We found that using shallow neural networks for the noise classifier $q_{\phi}$ was sufficient to accelerate sampling. Concretely, using a neural net with four convolution layers and one fully connected layer as the classifier, one evaluation of $q_{\phi}$ was around $100 \sim 1000$ times faster than one evaluation of the score model ${\boldsymbol s}_\theta$. So, when comparing sampling methods, we only count the number of score function evaluations (NFE). We also note that the training time $q_{\phi}$ was negligible compared to the training time of $s_\theta$. For instance, on CelebA-HQ-256, training $q_{\phi}$ with the aforementioned architecture for 100 epochs took around 15 minutes on an RTX 2080 Ti.
\textbf{Initialization points for DLG.} Theoretically, MCMC chain $\{(\bm{x}_n,\sigma_n)\}_{n = 1}^\infty$ will converge to $\hat{p}(\bm{x},\sigma)$ regardless of the initialization point $(\bm{x}_0,\sigma_0)$. However, we found it was beneficial to set $\bm{x}_0$ close to the image manifold and set $\sigma_0 \sim q_{\phi}(\sigma \mid \bm{x}_0)$. Indeed, theory also shows that setting initialization points close to the stationary distribution could significantly accelerate convergence of the Markov chain \citep{dalalyan2017,dwivedi2019}. In practice, we set $\bm{x}_0$ by generating a sample starting from a noise distribution, adding Gaussian noise of variance $0.25$, and running Gibbs sampling for a few iterations. The NFE involved in generating $\bm{x}_0$ is included in the final per-sample average NFE computation for DLG when comparing methods in Section \ref{sec:exp}. But, we note that this cost vanishes in the limit of infinite sample size.
\textbf{Reducing autocorrelation.} Autocorrelation in MCMC chains, i.e., correlation between consecutive samples in the MCMC chain, could reduce the sample diversity of MCMC. A typical technique to reduce autocorrelation is to use every $n_{skip}$-th samples of the MCMC chain for some $n_{skip} > 1$. For DMCMC, this means we denoise every $n_{skip}$-th sample. So, if we use $n_{den}$ NFE to denoise MCMC samples, the average NFE for generating a single sample is around $n_{skip} + n_{den}$.
\textbf{Choosing iterates to apply denoising.} The MCMC chain can be partitioned into blocks which consist of $n_{skip}$ consecutive samples. Using every $n_{skip}$-th sample of the MCMC chain corresponds to denoising the last iterate of each block. Instead, to further shorten the length of integration, within each block, we apply denoising to the sample of minimum noise scale $\sigma$.
\textbf{Choice of prior $\hat{p}(\sigma)$.} We use $\hat{p}(\sigma) \propto 1/\sigma$ to drive the MCMC chain towards small values of $\sigma$.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{figures/mix.png} \\
\caption{Mixing analysis of DLG with a mixture of Gaussians.}
\label{fig:mixing}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{figures/4.png}
\caption{Sampling acceleration of DLG on CIFAR10. FID of notable points are written in the corresponding color. \textbf{Top row}: deterministic integrators. \textbf{Bottom row}: stochastic integrators.}
\label{fig:cifar10_acc}
\end{figure}
\section{Experiments} \label{sec:exp}
\subsection{Mixing of DMCMC Chains}
For DMCMC to successfully generate diverse high-quality data, DMCMC must mix, i.e., traverse between disjoint modes of $p(\bm{x})$. We provide experimental evidence that DMCMC indeed mixes as a consequence of running MCMC in the product space $\mathcal{X} \times \SS$. Specifically, we run fifty Langevin dynamics chains and fifty DLG chains on a mixture of Gaussians (MoG) with $1k$ modes at CIFAR10 images. All chains are initialized at a single mode. For each method, we compute the mode coverage of the samples, the class distribution of the samples, and the autocorrelation of sample image class sequence. Since the noise conditional score function can be calculated analytically for MoGs, this setting decouples sampler performance from score model approximation error.
Figure \ref{fig:mixing} shows the results. In the left panel, we observe that Langevin dynamics is unable to escape the initial mode. Increasing the step size $\eta$ of Langevin dynamics caused the chain to diverge. On the other hand, DLG successfully captures all modes of the distribution. DLG samples cover all $1k$ modes at chain length $432$. Middle panel provides evidence that DLG samples correctly reflect the statistics of the data distribution. Finally, the right panel indicates that the DLG chain moves freely between classes, i.e., distant modes. These observations validate our claim that DLG mixes well.
\subsection{Accelerating Image Generation with Score Networks}
We compare six integrators with and without DLG on CIFAR10 and CelebA-HQ-256 image generation. The deterministic integrators are: the deterministic integrator of \citet{karras2022} (KAR1), the probability flow integrator of \citet{song2021}, and the RK45 solver. The stochastic integrators are: the stochastic integrator of \citet{karras2022} (KAR2), the reverse diffusion integrator of \citet{song2021}, and the Euler-Maruyama method. We use the Fr\'{e}chet Inception Distance (FID) \citep{heusel2017} to measure sample quality. For CIFAR10, we generate $50k$ samples, and for CelebA-HQ-256, we generate $10k$ samples. We use pre-trained score models of \citet{song2021}.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{figures/5.png}
\caption{Sampling acceleration of DLG on CelebA-HQ-256. A dot indicates an integrator without DLG, and a cross of the same color indicates corresponding integrator combined with DLG. Dotted lines indicate performance improvement due to DLG.}
\label{fig:celeba_acc}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.497\linewidth]{figures/sample_van.png}
\hspace{0.5mm}
\includegraphics[width=0.485\linewidth]{figures/sample_dlg.png}
\caption{Non-cherry-picked samples on CelebA-HQ-256 using the settings for Fig. \ref{fig:celeba_acc}. Each row shows samples for an integrator without (left col.) and with (right col.) DLG.}
\label{fig:samples}
\end{figure}
\textbf{CIFAR10.} Figure \ref{fig:cifar10_acc} shows the results on CIFAR10. We make two important observations. First, DLG successfully accelerates all six integrators by a non-trivial margin. In particular, if an integrator without DLG already performs well, the integrator combined with DLG outperforms other integrators combined with DLG. For instance, compare the results for KAR1 with those of other deterministic integrators. Second, DLG improves the performance lower bound for some deterministic integrators. Specifically, while the performance of KAR1 and RK45 saturates at around $4$ FID, KAR1 and RK45 combined with DLG achieve better results around $2.4$ FID.
\textbf{CelebA-HQ-256.} Figure \ref{fig:celeba_acc} shows the results on CelebA-HQ-256. We observe that DLG improves computational efficiency and sample quality simultaneously. Indeed, in Figure \ref{fig:samples}, we observe remarkable improvements in sample quality despite using fewer NFE. This demonstrates the scalability of DLG to generating high-resolution images. We also note that we did not perform an exhaustive search of DLG hyper-parameters for CelebA-HQ-256, so fine-tuning could yield better results.
\textbf{Achieving SOTA.} DLG combined with KAR1 sets a new SOTA record for CIFAR10 in the limited number of NFE setting: $3.86$ FID with $10.11$ NFE and $2.63$ FID with $20.11$ NFE which beats the results of \citet{zhang2022}, $4.17$ FID with $10$ NFE and $2.86$ FID with $20$ NFE. DLG combined with KAR2 sets a new record on CelebA-HQ-256 among score-based models: $6.99$ FID with $158.96$ NFE which beats the current best result of \citet{kim2022}, $7.16$ FID with $4000$ NFE.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{figures/3.png}
\caption{Ablation study of DLG with the deterministic integrator of \citet{karras2022}. Dots indicate the points of lowest FID.}
\label{fig:ablation}
\end{figure}
\subsection{Ablation Study}
In this section, we perform an ablation study of the components of DLG. Based on the observations, we construct a generic strategy for choosing the hyper-parameters. Given a reverse-S/ODE integrator, DLG is determined by three hyper-parameters: total NFE per sample $n$, NFE spent on denoising samples $n_{den}$, and Langevin dynamics step size $\eta$. We fix the integrator to be the deterministic sampler of \citet{karras2022} and observe the effect of each hyper-parameter on CIFAR10 image generation. We observed the same general trend with other samplers.
\textbf{$\bm{\eta}$ vs. $\bm{n}_{\bm{den}}/ \bm{n}$.} In the left panel of Figure \ref{fig:ablation}, we fix NFE and vary $\eta$ and $n_{den}/n$. $n_{den}$ governs individual sample quality, and $n_{skip} = n - n_{den}$ governs sample diversity. Thus, we observe optimal FID is achieved when $n_{den}/n$ has intermediate values, not extreme values near $0$ or $1$. Also, lower $n_{den}/n$ is needed to attain optimality for lower $\eta$. This is because lower $\eta$ means the MCMC chain travels closer to the image manifold at the cost of slower mixing.
\textbf{NFE vs. $\bm{n}_{\bm{den}}/ \bm{n}$.} In the middle panel of Figure \ref{fig:ablation}, we fix $\eta$ and vary NFE and $n_{den}/n$. We observe two trends. First, in the small NFE regime, where $10 \leq \text{NFE} \leq 50$, it is beneficial to decrease $n_{den}/n$ as NFE increases. Second, in the large NFE regime, where $\text{NFE} > 50$, it is beneficial to increase $n_{den}/n$ as NFE increases. This is because if $n_{skip}$ is sufficiently large, MCMC chain starts producing essentially independent samples, so increasing $n_{skip}$ further provides no gain.
Also, as we increase NFE, the set of $n_{den}/n$ which provides near-optimal performance becomes larger. Moreover, we see that most of the time, optimal FID is achieved when $n_{den} / n > 0.5$, i.e., when $n_{den} > n_{skip}$. So, a reasonable strategy for choosing $n_{den}$ given $\eta$ and NFE budget $n$ is to find smallest $n_{skip}$ which produces visually distinct samples, and then allocate $n_{den} = n - n_{skip}$.
\textbf{$\bm{\eta}$ vs. NFE.} In the right panel of Figure \ref{fig:ablation}, we choose optimal (in terms of FID) $n_{den}/n$ for each combination of $\eta$ and NFE. We see choosing overly small or large $\eta$ leads to performance degradation. If $\eta$ is within a certain range, we obtain similarly good performance. In the case of CIFAR10, we found it reasonable to set $\eta \in [0.05,1.0]$.
To choose $\eta$ for data of general dimension $d$, we define a value $\kappa \coloneqq \eta / \sqrt{d}$ called displacement per dimension. If we see \eqref{eq:dlg_lang}, Gaussian noise of zero mean and variance $\eta$ is added to the sample at each update step. Gaussian annulus theorem tells us that a high-dimensional Gaussian noise with zero mean and variance $\eta$ has Euclidean norm approximately $\eta \sqrt{d}$. So, the average displacement of the sample per dimension by the random noise is around $\eta / \sqrt{d}$. Since $\kappa$ is a dimension-independent value, given $\kappa$ and $d$, we can set $\eta = \sqrt{d} \kappa$. On CIFAR10, we have $\eta \in [0.05,1.0]$, which translates to $\kappa \in [0.0009, 0.018]$. This means, on CelebA-HQ-256, we can choose $\eta \in [0.4,8.0]$. If the sampler was inefficient, we chose a smaller $\kappa$ to trade-off diversity for sample quality.
\section{Conclusion}
In this work, we proposed DMCMC which combines MCMC with reverse-S/ODE integrators. This has led to significant improvements for both MCMC and diffusion models. For MCMC, DMCMC allows Markov chains to traverse between disjoint modes. For diffusion models , DMCMC accelerates sampling by reducing the average integration interval length of reverse-S/ODE. We developed a practical instance of DMCMC, called DLG, which uses Langevin dynamics along with Gibbs sampling as the choice of MCMC. We demonstrated the practicality and scalability of DLG through various experiments. In particular, DLG achieved state-of-the-art results on CIFAR10 and CelebA-HQ-256. Overall, our work opens up an orthogonal approach to accelerating score-based sampling. We leave exploration of other kinds of MCMC or diffusion process in DMCMC as future work.
\newpage
| 24ac643f66bd8479ba9b5fd79c6bda104a20d89e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{METHOD}
In this work, we investigate magnetotransport properties of 4LG encapsulated by hexagonal boron nitride (hBN). The electronic band structure of 4LG can be characterized by a set of hopping parameters $\gamma_{0}$-$\gamma_{5}$ [Fig.~$1$(a)] and energy imbalance between dimmer and non-dimmer sites $\delta$~\cite{koshino2011landau}. The low energy band structure of 4LG, shown in Fig.~$1$(b), comprises two BLG-like subbands with different effective masses~\cite{koshino2010interlayer,koshino2011landau,yagi2018low}. We denote the light-mass subband by b and the heavy-mass subband by B. The presence of skewed lattice sites hopping $\gamma_{3}$ induces trigonal warping. The next-nearest interlayer hopping parameters $\gamma_{2}$ and $\gamma_{5}$ cause the b and B subbands to overlap and hybridize at low energy as shown in Fig.~$1$(b)~\cite{koshino2011landau, shi2018tunable}. The remaining hopping parameters $\gamma_{4}$ and $\delta$ generate electron-hole asymmetry in the band structure.
\begin{figure*}
\includegraphics{Klanurak_figure2.pdf}
\caption{(a) conductance as a function of $n$ and $B$. (b) ${\rm d} g/ {\rm d} n$ numerically calculated from data in (a). The numbers at the right border indicate filling factors. (c) LL spectra in 4LG calculated from TB model. The filling factors associated with some energy gaps are labeled. The labels at the right border denote quantum states for each LL. At high field, LLs from the b and B subbands are in blue and red, respectively. At low field, LLs from both subbands are in black due to LL hybridization. (d) Calculated density of states of 4LG as a function of $n$ and $B$. \label{fig:wholeLL}}
\end{figure*}
In our devices, we use a dry transfer technique to assemble the heterostructures and etch them in CHF$_3$/O$_2$ plasma. Edge contacts are defined by electron beam lithography and formed by sputtering 80-nm Mo~\cite{wang2013one}. The standard lock-in technique is used to investigate magnetotransport properties. All measurements are performed at 2.4~K unless stated otherwise. Figure~$1$(c) shows an optical image and a schematic diagram of a device.
We first examine two-terminal conductance $g$ as a function of back-gate voltage $V_{\rm BG}$ at magnetic field $B=0$ [Fig.~$1$(d), black line]. We find that the conductance curve exhibits multiple local minimums, which are associated with band edges and Lifshitz transitions near zero energy in the band structure of 4LG as shown in Fig.~$1$(b)~ \cite{shi2018tunable,Hirahara20184LGpeaks}. At finite $B$, we observe conductance oscillations in $V_{\rm BG}$ due to the QHE [Fig.~$1$(d)]. Our conductance does not develop into well-defined plateaus even at $7$~T, likely due to relatively high contact resistance ($\sim$1~k$\Omega$) and geometry effect on two-terminal conductance which leads to the distortion of the QH plateaus \cite{abanin2008conformal,williams2009quantum}. At high field, the conductance exhibits a single minimum at which we associate to the charge neutrality point (CNP) of the sample. Near-zero gate voltage for the CNP indicates pristine quality of our samples.
To further investigate the magnetotransport properties of the system, we measure $g$ as a function of $V_{\rm BG}$ and $B$ [Fig.~$2$(a)]. Charge carrier density $n$ is determined from the period of conductance oscillations at high $B$. The Landau fan diagram displays rich features associated with LLs from the b and B subbands. To see the oscillations more clearly, we calculate ${\rm d} g/ {\rm d} n$ [Fig.~$2$(b)]. The dark lines correspond to energy gaps with associated filling factors $\nu$ shown in the figure. Multiple LL crossings are evident. For instance, the dark line for $\nu=12$, seen clearly at $7$~T, disappears at about $5$~T and reemerges again at $3$~T. The absence of the dark line for $\nu = 12$ between $3$ and $5$ T is the result of the LL crossing.
To gain a better understanding of the Landau fan diagram, we calculate energies of LLs as a function of $B$ using the tight-binding (TB) model [Fig.~$2$(c)]. In this plot, we assume that potential difference between layers is zero and use $\gamma_{0}=3.1$, $\gamma_{1}=0.39$, $\gamma_{2}=-0.022$, $\gamma_{3}=0.315$, $\gamma_{4}=0.12$, $\gamma_{5}=0.018$ and $\delta=0.020$ eV. These TB parameters are determined by matching the LL crossing positions from the experiment at low density with those from the calculation~\cite{taychatanapat2011quantum}. We note that our device has a single gate. As we induce higher carrier density via back gate, 4LG is inevitably subject to a stronger displacement field which induces larger potential difference between layers. Therefore, our simulation gives a good agreement with data from low density at which potential difference is still small.
From the spectra, LLs of 4LG at high $B$ can be viewed as a combination of two sets of BLG-like LLs from the b and B subbands~\cite{koshino2011landau,yin2017landau}. At high energy, a LL energy is approximately linear in $B$, as expected from bilayer nature of the subbands. The energy spacing of the LLs from the light-mass band b is larger than that of the heavy-mass band B because cyclotron frequency is inversely proportional to effective mass. The mixing between LLs due to $\gamma_2$ and $\gamma_5$ parameters and trigonal warping effect from $\gamma_3$ parameter leads to hybridization gaps, more visible at low energy~\cite{koshino2011landau}. We label each LL in Fig.~$2$(c) with two indices, indicating the subband (B or b) and LL index $n$ ($n+$ for electron-like and $n-$ for hole-like LLs). For the zeroth index LL, we label them as (b/B, 0) and (b/B, $-1$). These two LLs are degenerate in bilayer graphene but the degeneracy is slightly lifted in 4LG with a small energy gap of $\sim$2~meV at $7$~T. The numbers inside LL energy gaps in Fig.~$2$(c) indicate values of $\nu$ associated with the gaps. Each LL has four-fold spin and valley degeneracy~\cite{koshino2011landau}.
To compare our data with the calculation, we simulate density of states as a function of $B$ and $n$ from the LL spectra in Fig.~$2$(c) to obtain the plot in Fig.~$2$(d). Here, we assume a Lorentzian line-shape for each LL with broadening of 1.5 meV, estimated from a LL gap from the b subband at the onset of the oscillations of $0.5$~T [Fig.~$2$(b)]. The simulation captures main features of the experimental data in Fig.~$2$(b) such as the positions of LL crossing at low density and a position of the horizontal line in the electron side which originates from the zeroth LL of the light-mass band.
However, some discrepancies exist between our data and simulation. In our data, some LLs are two-fold degenerate. For example, in Fig.~$2$(b) at $6$~T, we observe the filling factor sequences of \{$8$, $10$, $12$\} and \{$24$, $26$, $32$\} which imply LLs with degeneracy of $2$ and $6$. Comparing the positions of these LLs with the LL spectra in Fig.~$2$(c), we conclude that each of the 4-fold degenerate (b,~$0$) and (b,~$1+$) split into $2$ LLs with $2$-fold degeneracy while the LLs from the B subband remain 4-fold degenerate. The $6$-fold degeneracy observed when filling factors change from $26$ to $32$ is the result of LL crossing between the $2$-fold degenerate LL from the splitting of (b,~$1+$) and the $4$-fold degenerate (B,~$5+$).
\begin{figure}
\centering
\includegraphics{Klanurak_figure3.pdf}
\caption{(a) LLs energy of the bulk 4LG as a function of $B$ with potential difference between adjacent layers of $16$~meV. The black and red lines are LLs for $K$ and $K'$, respectively. (b) Calculated density of states of the LLs in (a) as a function of $n$ and $B$.}
\label{fig:finiteU}
\end{figure}
In 4LG, a 2-fold degenerate LL can occur by applying a displacement field to generate a potential difference between layers which breaks inversion symmetry and lifts valley degeneracy. To capture the effect of the potential difference on LLs, we simulate LL spectra using a constant value of the potential difference between adjacent layers of $16$~meV and the same set of the hopping parameters used in Fig.~$2$(c). We note that, in our measurement, a value of potential difference will vary as we change density (see more details in Supplemental Material~\cite{RefSM}). The black and red lines in Fig.~$3$(a) represent LLs from $K$ and $K'$ valleys, respectively. The valley splittings of the LLs from the B~subband are much smaller than LL broadening and therefore they continue to appear $4$-fold degenerate in our measurement. However, the splittings are much more pronounced for LLs from the b subband, consistent with the data in which the LL splittings are observed in the b subband only. Although we can explain most observed features within a single-electron picture, we cannot completely rule out interaction-induced LL splitting. For example, the (b,~$0$) and (b,~$-1$) are so close in energy that they should experience similar value of the potential difference. However, we observe the splitting of (b,~$0$) but not (b,~$ -1$) which may suggest that other symmetry-breaking mechanisms are involved.
\begin{figure*}
\centering
\includegraphics{Klanurak_figure4.pdf}
\caption{(a) Color map of ${\rm d} g/{\rm d} n$ at low $n$ and $B$. The integer numbers indicate filling factors of the dark diagonal lines. (b) Conductance oscillations at $n = 0$ as a function of $B$ at various temperatures. The temperature at which the oscillations disappear ($\sim$20~K) is consistent with the size of the $\nu = 0$ gap. (c) The bulk LL spectra of 4LG (black lines). The blue and red lines are LL spectra of b and B subbands, respectively, calculated without band hybridization. (d) Diagrams of LL edge states at $B_{c1}^<$ and $B_{c1}^>$, indicated by pink and orange lines in (c). We omit the zeroth LLs (B,~$0$) and (B,~$-1$) (two red horizontal lines). }
\label{fig:zeroFF}
\end{figure*}
A more surprising discrepancy between our data and the simulation occurs at zero filling factor. Figure~\ref{fig:zeroFF}(a) shows a plot of ${\rm d}g/{\rm d}n$ at low $n$ and $B$ while Figure~\ref{fig:zeroFF}(b) displays $g$ at $n = 0$ as a function of $B$ at various temperatures. We observe oscillations in magnetoconductance clearly along the zero density in all three devices we have measured (see data in Supplemental Material~\cite{RefSM}). Typically, the conductance oscillations in QHE occur when Fermi energy passes through a LL. However, from the calculation of the LL spectra from the TB model [Fig.~\ref{fig:zeroFF}(c), black lines], there is no LL crossing inside $\nu = 0$. As a result, the conductance should exhibit no oscillation in $B$ at $\nu = 0$, contradicting our results.
To resolve this discrepancy, we examine the complex nature of low-energy LL spectra. Due to the band inversion in 4LG at zero magnetic field [see Fig.~\ref{fig:bandstruc}(b)], hole states from the b subband reside at higher energy than electron states from the B subband for wave vectors around zero. As $B$ increases, energies of hole-like LLs decrease while those of electron-like LLs increase. The opposite magnetic dependence of LL energies leads to a series of crossings and anti-crossings which manifests as three adjacent LLs braided together at low field [see Fig.~4(c), black lines]. These anti-crossings, whose energy gaps depend on $\gamma_2$ and $\gamma_5$ parameters, are the result of hybridization between hole-like LLs and electron-like LLs from the b and B subbands, respectively. A braiding of three LLs is the consequence of the trigonal warping effect from $\gamma_3$ parameter. The effect causes the anti-crossings to occur when the LL indices of the unperturbed LLs are the same or differ by multiples of $3$~\cite{Serbyn2013hybrid,campos2016TLG, Shimazaki2016LandauTrilayer,Zibrov2018Gully}.
To identify the underlying LLs that hybridize into the braided LLs, we calculate LL spectra by setting the mixing terms between the subbands to zero~\cite{koshino2011landau}. The result is displayed in Fig.~\ref{fig:zeroFF}(c). The blue and red lines represent LLs of the b and B subbands, respectively. Comparing the LLs with and without the mixing terms, we find that the energy gap at $\nu = 0$ originates from two different mechanisms with a crossover at the critical field $B_{c1}$ of $\sim$2.1~T. For $B > B_{c1}$, the $\nu=0$ gap is a trivial LL gap which is always present regardless of LL hybridization. For $B < B_{c1}$, the $\nu=0$ gap emerges from a series of hybridizations between hole-like (b,~$i-$) and electron-like (B,~$i+$) for $i = 1,2,\ldots$.
Let us focus on a $\nu = 0$ gap at $B_{c1}$ arising from the hybridization between (b,~$1-$) and (B,~$1+$). At $B_{c1}^< < B_{c1}$, the hybridized LLs at higher energy (solid blue star) and lower energy (solid red triangle) in Fig.~\ref{fig:zeroFF}(c) are mostly dominated by hole-like (b,~$1-$) and electron-like (B,~$1+$), respectively. As we increase $B$ beyond $B_{c1}$, the admixture of each hybridized LL gradually changes and the situation becomes reversed. Now, the higher-energy LL (hollow red triangle) evolves into the electron-like (B,~$1+$) while the lower-energy LL (hollow blue star) turns into the hole-like (b,~$1-$). Therefore, as we increase $B$, the characteristic of the higher-energy LL switches from electron-like to hole-like LL while that of the lower-energy LL changes from hole-like to electron-like LL. As a result, the edge states change their behavior significantly across a hybridization gap.
Figure~\ref{fig:zeroFF}(d) illustrates edge state diagrams at $B_{c1}^<$ and $B_{c1}^>$, indicated by pink and orange lines in Fig.~\ref{fig:zeroFF}(c). Due to a confining potential, the energy of an electron-like LL will bend up near the edge while that of a hole-like LL will bend down. At $B_{c1}^>$ [Fig.~$4$(d), right], an energy ordering of the LLs is a conventional one in which the electron-like (B,~$1+$) has higher energy than the hole-like (b,~$1-$). In this case, the energies of both LLs will bend away from each other near the edge. Therefore, these two LLs do not contribute any edge state to the system at $\nu = 0$. However, at $B_{c1}^<$, we have an inverted energy ordering of the LLs in which the energy of hole-like (b,~$1-$) is higher than that of the electron-like (B,~$1+$) in the bulk. Near the edge, their energies will bend toward each other (down for (b,~$1-$) and up for (B,~$1+$)) [Fig.~\ref{fig:zeroFF}(d), left]. We therefore obtain two counter-propagating edge states even though the filling factor is zero in the bulk. We note that these counter-propagating edge states are likely not helical since they can interact via mixing terms. As a result, we expect an energy gap to open at the crossing.
As we lower $B$ further, we encounter another hybridization gap between (b,~$2-$) and (B,~$2+$) at $B_{c2} \sim 1.2$~T [Fig.~$4$(a-c)]. With the same argument as the $B_{c1}$ case, the number of edge states will increase from $2$ to $4$ when $B$ drops below $B_{c2}$ because two hole-like (b,~$1-$) and (b,~$2-$) now sit at higher energy than two electron-like (B,~$1+$) and (B,~$2+$). In general, as we move through the hybridization gap between (b,~$i-$) and (B,~$i+$), the number of edge states changes from $2(i-1)$ to $2i$. We find that the positions of the conductance peaks at $\nu = 0$ in Fig.~\ref{fig:zeroFF}(b) are in excellent agreement with the theoretical positions of $B_{ci}$ in Fig.~4(c) which are magnetic fields at which a slope is zero for a hybridized LL separating $\nu = 0$ and $4$. For $B > B_{c1}$, the energy gap at $\nu = 0$ turns into a trivial LL gap and there is no further inversion of electron-like and hole-like LLs for LL indices $|n| \geq 1$. Therefore, the number of edge states stays constant and we no longer observe any oscillation [Fig.~\ref{fig:zeroFF}(b)]. Similarly, following the $\nu = 12$ line in Fig.~\ref{fig:zeroFF}(a), we observe magnetoconductance oscillations when $B \lesssim 1.2$~T even though the $\nu = 12$ gap remains finite in the bulk. We find that this $\nu = 12$ gap below $1.2$~T arises from a series of hybridization gaps between (b,~$i-$) and (B,~$(i+3)+$) while the gap above $1.2$~T is a trivial LL gap.
We therefore conclude that the oscillations arise from the change in the number of edge states in a hybridization gap between electron-like and hole-like LLs. We emphasize that a hybridization gap between LLs of the same type will not lead to an oscillation because the number of edge states remains unchanged~\cite{Shimazaki2016LandauTrilayer,Zhang2006Anticrossing}. Our results show that it is not sufficient to predict Shubnikov-de Haas oscillations from Landau level spectra only. One needs to consider if a gap is a trivial Landau level gap or a hybridization gap between electron-like and hole-like LLs to obtain a complete picture of the oscillations. We expect our result to be useful for understanding Landau level spectra of other few layer graphene systems because of the band inversion in their band structures.
For helical edge modes, two-terminal conductance in the hybridization gap regime along $\nu = 0$ should appear as steps with quantized values of $4N e^2/h$ where $N$ is the number of edge states and a factor of $4$ comes from spin and valley degeneracy~\cite{young2014tunable,sanchez2017helical,che2020helical}. For instance, when $\nu = 0$ and $B_{c2} < B < B_{c1}$, we expect conductance of $8e^2/h$ from two counter-propagating edge states from (b,~$1-$) and (B,~$1+$) but the measured value is less than $2 e^2/h$. A few mechanisms could contribute to the observed low value of conductance. Since the edge states in our system are not helical and they counterpropagate on the same edge, these two states could mix and tunnel to each other. As a result, they form 1D localized states and conductance is no longer quantized at $4N e^2/h$ because the edge states do not have a perfect transmission~\cite{Lee1985}. An interaction between the edge states could induce a small gap at the Fermi energy, reducing conductance further. In addition, our conductance appears oscillatory which is likely due to the geometry effect observed in a long sample for two-terminal measurement~\cite{abanin2008conformal, williams2009quantum}.
In summary, we study magnetotransport properties of 4LG. We observe LL crossings between the b and B subbands. At finite displacement field, we find that the LLs in the b~subband become valley polarized while those in the B subband remain valley degenerate, in agreement with the TB calculation with finite potential difference. At low $n$ and $B$, the band inversion gives rise to a series of bulk hybridization gaps between electron- and hole-like LLs. As a result, alternating characteristic of the hybridized LLs between electron and hole states leads to the change in the number of edge states and manifests as magnetoconductance oscillations in our measurement. Finally, our proposed mechanism for magnetoconductance oscillations should be applicable to other Bernal-stacked multilayer graphenes since they also host similar band inversion~\cite{Hirahara20184LGpeaks,Horii20194LG}.
\begin{acknowledgments}
We thank K. Jaruwongrungsee for experimental help and S. Hodak for useful discussion. This research has been primarily supported by the Research Fund for DPST graduate with First Placement (Grant no.~002/2015), the NSRF via the Program Management Unit for Human Resources \& Institutional Development, Research and Innovation (Grant no.~B05F640152), and National Research Council of Thailand (NRCT) and Chulalongkorn University (Grant no.~N42A650266). K.W. and T.T. acknowledge support from JSPS KAKENHI (Grant Numbers 19H05790, 20H00354 and 21H05233).
\end{acknowledgments}
| 402b92802b7edb1255dc844fdac73e1164d1e63d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Brownian motion of particles was first observed and described in 1827 by Brown. This random, uncontrolled motion of particles in a fluid, driven by the bombardment of ambient fluid molecules, gives rise to the macroscopic phenomenon of diffusion \citep{Einstein1905On} and convincing evidence of atomistic theory of matter. \citep{Perrin1908Les} These seminal works of Einstein and Perrin entail a fundamental understanding of Brownian motions of rigid spherical particles in a Newtonian fluid. In particular, the intimate relation is now well understood between the seemingly uncontrolled microscopic random motions of microparticles and the macroscopic diffusion of a solute in a solvent, i.e., the upscaling from the microscopic processes to macroscopic properties.
Nowadays, many advanced technologies necessitate non-standard Brownian processes of microstructured particles in a complex medium. For instance, drug release from swellable polymer involves Brownian motions in a porous elastic medium, giving rise to non-Fickian diffusion behavior.~\citep{ende1995} Also, the diffusion of drug molecules in the tumor interstitium~\citep{Jain1987tumor} and the diffusion of univalent cations and anions out of liquid crystals of lecithin~\citep{Bangham1965ion} require consideration of viscoelasticity and anisotropy of the medium and multiphysical interactions between the particle and ambient medium. These non-standard Brownian processes demand a more comprehensive analysis of motions of microstructured particles in a complex medium and its ramification in macroscopic anomalous diffusions.
With the advancement of experimental techniques including fluorescence recovery and confocal microscopy, a great number of experiments have been made to Brownian motions of various microstructured particles in a complex medium. Perrin~\citep{Perrin1934Mouvement,Perrin1936Mouvement} has observed the Brownian motion of ellipsoidal particle as early as in 1930's.
Han {\em et al.}\citep{Han2006Brownian} analyzed and measured the two dimensional Brownian motions of ellipsoidal particles in water, which was later extended to quasi-two-dimension to account for the effects of shape anisotropy and confinement. \citep{Han2009Quasi} In the field of nanotechnology, the Brownian motions of a variety of nano-particles are visualized and measured, including
single-walled carbon nanotubes, \citep{Duggal2006carbonnanotube} copper oxide nanorods, \citep{Cheong2010Rotational} carbon nanofibers, \citep{Bhaduri2008carbonnanofiber} etc. The diffusion behavior of microstructured particles or microorganisms have been characterized, ranging from colloidal trimer, \citep{Kraft2013colloidal} graphene flakes, \citep{Marago2010Brownian} boomerang particles, \citep{Chakrabarty2014Brownian} actin filaments, \citep{Koster2005Brownian} to Leptospira interrogans.~\citep{Koens2014Leptospira}
Conventional analysis of Brownian motions starts from a spherical rigid homogeneous particle in a Newtonian fluid.
The central result of Einstein\citep{Einstein1905On} and Perrin~\citep{Perrin1908Les} is termed as the Stokes-Einstein's relation: $D=\mu k_B T$, where $D$, $\mu$, $k_B$, and $T$ are the macroscopic diffusivity, microscopic mobility, Boltzmann's constant, and absolute temperature, respectively. Combined with Stokes's formula for the mobility, the Stokes-Einstein's relation provides early measurement of the Boltzmann's constant (or Avogadro constant) and decisive evidence of atomistic theory of matter. A microstructured particle, however, could have heterogeneous density and complex shape. From a microscopic viewpoint, the hydrodynamic interaction between the microstructured particle and ambient fluid can no longer be sufficiently captured by a single scalar mobility $\mu$. Nevertheless, at the absence of external field we expect particles would diffuse isotropically with $D=\langle x(t)^2\rangle/2t$ if the time scale $t$ is long enough to smear out all microscopic short-time rotations of the particles. From this viewpoint, we expect two distinct regions of diffusion behaviors delimited by the {\em crossover time} scale $t_{\rm cross}$.~\citep{Han2006Brownian} On a short time-scale $t<t_{\rm cross}$, the initial orientation of the anisotropic particle has a significant influence on movements of the particle, resulting in an anisotropic and time-dependent diffusivity tensor ${\mathbf D}\sim \langle {\mathbf x}(t)\otimes {\mathbf x}(t)\rangle/2t$. On the other hand, if the time-scale $t \gg t_{\rm cross}$, the classic Stokes-Einstein's relation should hold for some ``effective'' or ``average'' mobility $\mu^{\rm eff}$, provided that the ambient fluid is isotropic and there is no external field breaking the rotational symmetry.
In an effort to quantify the crossover time scale and generalize the Stokes-Einstein's relation, Brenner~\cite{Brenner1964extension, Brenner1967Coupling} conducted a systematic study on the Brownian motions of particles of arbitrary shapes. His analysis revealed two important effects of shape anisotropy: (i) the Stokes' formula for mobility of a spherical particle shall be replaced by an anisotropic mobility tensor for a particle of general shape, and (ii) the {\em hydrodynamic center}, i.e., the point of action of the resultant hydrodynamic forces on the non-rotating particle, in general may not coincide with the center of mass of the particle. For particles with hydrodynamic center coinciding with the center of mass,~\citep{Brenner1964extension, Bernal1980, Harvey1980} the translational and rotational motions are uncoupled and solutions to the associated stochastic differential equations are manageable.~\citep{Han2006Brownian, Cheong2010Rotational, Chakrabarty2014Brownian, Yuan2021} For particles of general shapes and heterogeneities, the translational and rotational degrees of freedom are intrinsically coupled, besides the technical difficulties arising from the stochastic forces of thermal agitations and nonlocality of hydrodynamic forces on the particle. A major theme of this work is to analyze this coupled stochastic equations of motions, quantify the crossover time scale $t_{\rm cross}$, and achieve generalized Stokes-Einstein's relations for the long-time diffusivity in terms of geometrical and physical properties of the particle and ambient fluid.
A second theme of our current work is to explore the effects of external fields on the diffusivity of microstructured particles. External electric or magnetic field can be remotely applied to suspensions and easily controlled. From a non-stochastic viewpoint, external fields have two major effects on microstructured particles: (i) suspension particles in general move along field lines to either concentrate to (or dilate from) regions of larger field strength if the field is nonuniform, and (ii) suspension particles tend to align with field lines. The movements driven by external fields, though inevitably randomized by thermal agitations of ambient fluid molecules, are expected to have an overall effect on the cross-over time scale and the long-time diffusivity. The ability of manipulating the motion or diffusion of particles in suspension has inspired many applications. For instance, an electric field can assist the scalable fabrication of composites incorporating vertically-aligned carbon nanotubes \citep{Castellano2015Electrokintics} or well-organized boron-nitride nanotubes.~ \citep{Cetindag2017Surface} A magnetic field can be used to tune the acoustic attenuation of the suspensions of subwavelength-sized nickel particles.~ \citep{Yuan2017Tunable} Further, the field-controlled motion or diffusion of particles in suspension paves new anvenues for applications in induced-charge electrophoresis,~\citep{Squires2006Breaking} contactless characterization of the semiconductor,~ \citep{Yuan2019Contactless} separation of the particle,~ \citep{Doiparticlesorting} etc.
To understand the field-controlled diffusion of particles, Grima et al. \citep{Grima2007Brownian} studied the quasi-two-dimensional Brownian motion of an ellipsoidal rigid particle in the electrophoresis and microconfinement and found that the long-time diffusivity of an asymmetric particle is anisotropic in the presence of external forces and depends on the shape of the particle. Gu\"{e}ll {\em et al.}\citep{Guell2010Anisotropic} studied the diffusion of an ellipsoidal particle in an external rotating magnetic field and showed the correlation of the crossover time scale with the frequency and amplitude of the field. Aurell {\em et al.}\citep{Aurell2016Diffusion} used a systematic multi-scale technique to study the diffusion of an ellipsoid under the application of a constant external force and found that the diffusivity parallel to the direction of the force field is $\frac{4}{3}$ of that perpendicular to the direction the force field. Obasanjo \citep{Obasanjo2016Response} experimentally studied the electro-rotation and electro-orientation of the particle and measured the electric-field-tuned diffusion coefficients of ellipsoidal particles in the AC field. Segovia-Guti\'{e}rrez {\em et al.} \citep{Segovia2019Diffusion} experimentally studied the rotational and translational dynamics of trimers in a quasi-two-dimensional system under a random light field and found the coupling between the rotational motion and the translational motion of trimers relies on the length of the scale of the particle. In our prior work,~\citep{Yuan2021} we studied the diffusion of carbon nanotube under an aligning AC electric field, gave an explicit formula to the anisotropic diffusivity tensor, and experimentally validated the formula. In particular, we found that the diffusion coefficient parallel to the field increases and the diffusion coefficient perpendicular to the field decreases as the field strength increases. The trace of diffusivity tensor tensor remains as constant. In all these works, the hydrodynamic center of the particle coincides with the center of mass because of axis-symmetry, and hence the translational and rotational motions are essentially decoupled.
In the this work we focus on the long-time diffusivity of microstructured particles of general shapes and heterogeneities in an external field. Starting from the microscopic stochastic equations of motions with coupled translational and rotational degrees of freedom, we find the coarse-grained Fokker-Planck equation that governs the evolution of the Probability Distribution Function (PDF) in the configurational space consisting of position, orientation, linear velocity and angular velocity of the particle. From the classical Boltzmann's distribution for equilibrium states, we determine the fluctuation coefficients associated with Brownian forces. Next, we neglect the effects of inertia and consider a Lagevin-type equation for the position and orientation of the particle. Again by studying the associated Fokker-Planck equation, we extract the macroscopic long-time diffusivity in terms of properties of microstructured particles and ambient fluid. This approach is versatile and self-contained and could be applied to study more general diffusion phenomena, e.g., deformable particles in a visco-elastic ambient medium.
On the technical side, the generalized Stokes-Einstein's relations we obtain for the long-time diffusivity of general microstructured particles involve evaluations of the mobility tensors and integrals over the continuous group of all spatial orientations (or rigid rotations) ${\rm SO}(3)$. The former concerns the classical Stokes' flow around a particle of arbitrary shape with certain linear velocity and angular velocity and has been addressed in a number of earlier works.~\cite{Brenner1964extension, Brenner1967Coupling} In particular, explicit solutions to the mobility tensor are available for ellipsoids.~\cite{Brenner1964extension, Brenner1967Coupling} Integrations over the continuous group ${\rm SO}(3)$ are achieved by a special parametrization of ${\rm SO}(3)$ that is closely related to the quaternion algebras.\citep{Quaternions} Upon employing approximations based on the mobility tensor of ellipsoids and quaternion algebras, we achieve explicit formulas for long-time diffusivitives of general heterogeneous ellipsoids and two spheroids bonded by some ligaments. These results could be further applied to improve the design and functionality of microstructured particles in colloidal systems and create fundamental understanding of the anomalous diffusion of macromolecules in complex media.
The paper is organized as follows. In Sec.~\ref{sec:EOM} we present our approach from microscopic equations of motions for Brownian motions to Fokker-Planck equations for macroscopic evolution of PDF in configurational space. In Sec.~\ref{sec:Laginvin} we analyze the Fokker-Planck equations, quantify the crossover time scale, and obtain a formula for the long-time diffusivity of microstructured particles under an external field. In Sec.~\ref{Sec:app} we present explicit results for the long-time diffusivity of heterogeneous spheroids, which generalizes the classical Stokes-Einstein's formula to account for the effects of shape anisotropy and heterogeneities. Based on these explicit results, we then study the diffusion of a pair of spheroids undergoing a transition from an initial non-equilibrium state to the final equilibrium state. The non-standard scaling behavior of MSD versus time of these processes may be a good model for understanding the anomalous diffusions, e.g., migration of macromolecules and cells in the complex medium.
\noindent {\em Notation.}\; Direct notation is employed for brevity and transparency of physical interpretation whenever possible. { Frequently, recognizing that many readers may be more familiar with the index notation, we also presented translations in index form to illustrate details of the calculations.} Vectors and tensors are denoted by bold symbols such as ${\mathbf x}, {\mathbf v}, {\mathbf Q}$, etc., while scalars are denoted by $\psi, \eta$, etc. For a vector field ${\mathbf v}$, in index form the gradient operator ${\rm grad} {\mathbf v}$ is equivalent to $({\mathbf v})_{i, j}$ with $i$ $(j)$ being the first (second) index.
When index notation is in use, the convention of summation over repeated index are followed.
\section{Equation of motion} \label{sec:EOM}
We consider the Brownian motion of a particle in a Newtonian fluid under the application of some external field. Denote by ${\mathbf x}$ the position of the center of mass with respect to a fixed global frame $\{{\mathbf e}_i: i=1,2,3\}$, ${\mathbf v}=\dot{{\mathbf x}}(t) $ is the velocity of the center of mass of the particle, (``$\cdot$'' denotes $\frac{\mathrm{d}}{\mathrm{d} t}$) ${{\boldsymbol{\omega}}}$ the angular velocity of the particle, and
$m$ (resp. ${\mathbf I}\in {\mathbb{R}}^{3\times 3}_{\rm sym}$) the mass (resp. the moment of inertia with respect to the center of mass and the global frame). The presence of an external field in general exerts certain external force (${\mathbf g}^{e} \in\mathbb{R}^3$) and torque
($ {{{\bf \tau}}}^e\in\mathbb{R}^3$) on the microparticle. In addition, the bombardments of ambient molecules on the particle give rise to a random force ${\mathbf g}^{B}$ and random torque ${{{\bf \tau}}}^B $.
Suppose the particle moves in the fluid with linear velocity ${\mathbf v}$ and angular velocity ${{\boldsymbol{\omega}}}$. The particle generates a flow in the ambient fluid and consequently, suffers from a drag force ${\mathbf g}^d$ and torque ${{\bf \tau}}^d$ because of the viscosity of the fluid. Since the Reynold number is small, it suffices to consider Stokes' flow and conclude that the force ${\mathbf g}^d$ and torque ${{\bf \tau}}^d$ on the particle are related with the linear velocity ${\mathbf v}$ and angular velocity ${{\boldsymbol{\omega}}}$ by a linear transformation:
\begin{eqnarray}\label{frictiontotal}
\begin{bmatrix}
{\mathbf g}^d\\
{{\bf \tau}}^d
\end{bmatrix}=-{\mathbf R}
\begin{bmatrix}
{\mathbf v}\\
{{\boldsymbol{\omega}}}
\end{bmatrix},
\end{eqnarray}
where the drag matrix ${\mathbf R}\in {\mathbb{R}}^{6\times 6}_{\rm sym}$ is symmetric and positive-definite. For future convenience, we write the drag matrix in a block form as
\begin{eqnarray} \label{eq:bfRblock}
\begin{split}
{\mathbf R}=
\begin{bmatrix}
{\mathbf R}^\tt &{\mathbf R}^{\rm tr} \\
( {\mathbf R}^{\rm tr})^T &{\mathbf R}^{\rm rr} \\
\end{bmatrix} ,
\end{split}
\end{eqnarray}
where ${\mathbf R}^{(\tt, {\rm rr}, {\rm tr})}\in{\mathbb{R}} ^{3\times3}$ pertains to the translational motion, the rotational motions, and the coupling between translational and rotational motions, respectively.
From the classic rigid-body mechanics, the equation of motion for the particle can be expressed as
\begin{eqnarray}\label{eq:kinetic anisotropic000}
\begin{split}
&\frac{\mathrm{d}}{\mathrm{d} t}( m{\mathbf v}) =-{\bf R}^\tt {\mathbf v} -{\bf R}^{\rm tr} {{\boldsymbol{\omega}}}+{\mathbf g}^e +{\mathbf g}^B,\\
&\frac{\mathrm{d}}{\mathrm{d} t} ({\mathbf I} \dot{{{\boldsymbol{\omega}}}})= - {{\mathbf R}}^{\rm rr} {{\boldsymbol{\omega}}}-{{\mathbf R}}^{{\rm tr}^T}{\mathbf v}+{{{\bf \tau}}}^e+{{{\bf \tau}}}^B.
\end{split}
\end{eqnarray}
For an orthonormal body frame $\{{\mathbf f}_i: i=1,2,3\}$ fixed on the particle, let ${\mathbf Q}(t)={\mathbf e}_i\otimes{\mathbf f}_i(t)\in{\rm SO}(3)$ be the associated orthogonal matrix
and
\begin{eqnarray*}
\begin{split}
& ({\hat{\bf R}}^{(\tt, {\rm tr}, {\rm rr})}, {\hat{\bf I}}, {\hat{\bfsigma}}_r)={\mathbf Q}({\mathbf R}^{(\tt, {\rm tr}, {\rm rr})}, {\mathbf I}, {{\boldsymbol{\sigma}}}_r){\mathbf Q}^T ,\\
& (\hat{{{\boldsymbol{\omega}}}}, \hat{{{\bf \tau}}}^e,\hat{{{\bf \tau}}}^B)={\mathbf Q}({{{\boldsymbol{\omega}}}}, {{{\bf \tau}}}^e,{{{\bf \tau}}}^B)
\end{split}
\end{eqnarray*}
the tensors and vectors with respect to the body frame $\{{\mathbf f}_i: i=1,2,3\}$. Then the equation of motion~\eqref{eq:kinetic anisotropic000} can be rewritten as~\citep{Yuan2021}
\begin{eqnarray}\label{eq:kinetic anisotropic}
\begin{split}
&m\dot{{\mathbf v}}=-{\mathbf Q}^T\hat{\bf R}^\tt {\mathbf Q} {\mathbf v} -{\mathbf Q}^T\hat{\bf R}^{\rm tr} {\hat{\boldsymbol \omega}}+{\mathbf g}^e +{\mathbf g}^B,\\
&{\hat{\bf I}} \dot{{\hat{\boldsymbol \omega}}}= - \hat{{\mathbf R}}^{\rm rr} {\hat{\omega}}- \hat{{\mathbf R}}^{{\rm tr}^T} {\mathbf Q} {\mathbf v}+\hat{{{\bf \tau}}}^e+\hat{{{\bf \tau}}}^B,
\end{split}
\end{eqnarray}
where the term $ {\hat{\boldsymbol \omega}}\times {\hat{\bf I}}{\hat{\boldsymbol \omega}}$ has been neglected as compared with ${\hat{\bf I}} \dot{{\hat{\boldsymbol \omega}}}$.~\citep{Yuan2021}
We remark that the drag matrices ${\hat{\bf R}}^{(\tt, {\rm tr}, {\rm rr})}$ and moment of inertia ${\hat{\bf I}}$ with respect to the body frame are independent of the time or the motion of the particle. They are geometrical and physical properties of the particle and the ambient fluid. In this work we will focus on the case that the translational and rotational motions of the particle are intrinsically coupled in the sense that ${\hat{\bf R}}^{\rm tr}\neq{\bf 0}$.
We now determine the random force ${\mathbf g}^B$ and torque $\hat{{{\bf \tau}}}^B$ from thermal agitations. Presumably, these random force and torque on the particle is independent of the external fields. Therefore, for the purpose of fixing ${\mathbf g}^B$ and $\hat{{{\bf \tau}}}^B$ it is convenient to analyze \eqref{eq:kinetic anisotropic} at the absence of the external field, i.e.,
\begin{eqnarray}\label{eq: for velocity}
\begin{split}
\dot{{\mathbf v}}&=-\frac{1}{m}{\mathbf Q}^T\hat{\bf R}^\tt {\mathbf Q} {\mathbf v} -\frac{1}{m}{\mathbf Q}^T\hat{\bf R}^{\rm tr} {\hat{\boldsymbol \omega}} +\frac{1}{m}{\mathbf g}^B,\\
\dot{{\hat{\boldsymbol \omega}}}&= - {\hat{\bf I}}^{-1}\hat{{\mathbf R}}^{\rm rr} {\hat{\omega}}- {\hat{\bf I}}^{-1}\hat{{\mathbf R}}^{{\rm tr}^T}{\mathbf Q} {\mathbf v}+{\hat{\bf I}}^{-1}\hat{{{\bf \tau}}}^B.\\
\end{split}
\end{eqnarray}
To proceed, we assume that
the generalized random force can be expressed as
\begin{eqnarray}\label{eq: sigmatildeexpression}
\begin{bmatrix}\frac{1}{m}{\mathbf g}^B\\{\hat{\bf I}}^{-1}\hat{{{\bf \tau}}}^B\end{bmatrix}=\tilde{\bm{\sigma}}\tilde{\bm{\xi}},
\end{eqnarray}
where $\tilde{\bm{\sigma}}\in {\mathbb{R}}^{6\times 6}$ is called the fluctuation coefficients, and $\tilde{\bm{\xi}}(t) $ represents normalized six-dimensional uncorrelated white noises. In other words, the process $\tilde{\bm{\xi}}(t) $ satisfies
\begin{eqnarray} \label{eq:xidef}
\langle \tilde{\xi}(t)\rangle =0,\qquad
\langle \tilde{\xi}_i(t) \tilde{\xi}_j(t')\rangle=\delta_{ij}\delta(t-t'),
\end{eqnarray}
where $\delta_{ij}$ is the Kronecker delta, and $\delta$ is the Dirac function.
Next, we introduce notation
\begin{eqnarray}\label{eq: l}
\tilde{{\mathbf v}}:=\begin{bmatrix}{\mathbf v}\\{\hat{\boldsymbol \omega}}\end{bmatrix},\qquad
{\mathbf L}:=\begin{bmatrix}-\frac{1}{m}{\mathbf Q}^T\hat{\bf R}^\tt {\mathbf Q} & -\frac{1}{m}{\mathbf Q}^T\hat{\bf R}^{\rm tr} \\- {\hat{\bf I}}^{-1}\hat{{\mathbf R}}^{{\rm tr}^T}{\mathbf Q}&- {\hat{\bf I}}^{-1}\hat{{\mathbf R}}^{\rm rr} \end{bmatrix}.
\end{eqnarray}
By \eqref{eq: l} and \eqref{eq: sigmatildeexpression}, we rewrite \eqref{eq: for velocity} in a compact form as \begin{eqnarray}\label{governingequation}
\dot{\tilde{{\mathbf v}}}={\mathbf L}\tilde{{\mathbf v}}+\tilde{\bm{\sigma}}\tilde{\bm{\xi}},
\end{eqnarray}
which may be recognized as a stochastic differential equation (SDE) for the six-dimensional generalized velocity $\tilde{{\mathbf v}}$.
To fix the fluctuation coefficients $\tilde{\bm{\sigma}}$, we consider the probability distribution function (PDF) $P=P(\tilde{{\mathbf v}}, t)$ in the generalized velocity space. Physically, the quantity $P(\tilde{{\mathbf v}}, t)\mathrm{d}\tilde{{\mathbf v}}$ is the probability of finding the particle with generalized velocity from the infinitesimal element $\mathrm{d}\tilde{{\mathbf v}}$ centered at $\tilde{{\mathbf v}}$.
From the master equation, it can be shown that the PDF $P(\tilde{{\mathbf v}}, t)$ associated with the stochastic process~\eqref{governingequation} satisfies the Fokker-Planck equation:~\citep{Kampen2007Stochastic}
\begin{equation} \label{Fokker-Planck for rtilde}
\frac{\partial P(\tilde{{\mathbf v}}, t)}{\partial t}=\nabla_{\tilde{{\mathbf v}}} \cdot \Big\{-\tilde{{\boldsymbol{ \alpha}}} P(\tilde{{\mathbf v}}, t)+\frac{1}{2}\tilde{{{\boldsymbol{\beta}}}}\nabla_{\tilde{{\mathbf v}}} P(\tilde{{\mathbf v}}, t)\Big\},
\end{equation}
where
\begin{eqnarray}\label{eq: alphabeta}
\label{alphabeta}
\begin{split}
&\boldsymbol{\tilde{\alpha}}=\lim_{d {\bf t} \rightarrow 0} \frac{\langle \tilde{{\mathbf v}}(t+d {\bf t} )-\tilde{{\mathbf v}}(t)\rangle }{d {\bf t} },\\
&\boldsymbol{\tilde{\beta}}=\lim_{d {\bf t} \rightarrow 0} \frac{\langle [\tilde{{\mathbf v}}(t+d {\bf t} )- \tilde{{\mathbf v}}(t)] \otimes [\tilde{{\mathbf v}}(t+d {\bf t} )- \tilde{{\mathbf v}}(t)]\rangle }{d {\bf t} }.
\end{split}
\end{eqnarray}
By \eqref{governingequation} and \eqref{eq: alphabeta}, it can be shown that~\cite{evans2012stochastic, Yuan2021}
\begin{eqnarray}\label{eq: alphabetatilde}
\begin{split}
&\boldsymbol{\tilde{\alpha}}={\mathbf L}\tilde{{\mathbf v}}\quad\mathrm{and}\quad\boldsymbol{\tilde{\beta}}=\tilde{\bm{\sigma}}\tilde{\bm{\sigma}}^T.
\end{split}
\end{eqnarray}
From the classical statistical physics, the Maxwell-Boltzmann distribution
\begin{eqnarray}\label{stationary}
P^s(\tilde{{\mathbf v}})\propto \exp{\left[-\left(\frac{m|{\mathbf v}|^2}{2k_BT}+\frac{{\hat{\boldsymbol \omega}}\cdot {\hat{\bf I}}{\hat{\boldsymbol \omega}}}{2k_BT}\right)\right]}
\end{eqnarray}
should be the stationary solution to \eqref{Fokker-Planck for rtilde}. Inserting \eqref{stationary} and \eqref{eq: alphabetatilde} into \eqref{Fokker-Planck for rtilde}, we find that
\begin{eqnarray}\label{fluctuationsdetermined}
\tilde{\bm{\sigma}}\tilde{\bm{\sigma}}^T=2k_BT\begin{bmatrix}\frac{1}{m^2}{\mathbf Q}^T\hat{\bf R}^\tt {\mathbf Q}&\frac{1}{m}{\mathbf Q}^T\hat{\bf R}^{\rm tr}{\hat{\bf I}}^{-1}\\\frac{1}{m}{\hat{\bf I}}^{-1}\hat{\bf R}^{{\rm tr}^T}{\mathbf Q}&{\hat{\bf I}}^{-1}\hat{\bf R}^{\rm rr}{\hat{\bf I}}^{-1}\end{bmatrix},
\end{eqnarray}
which essentially specify the random force ${\mathbf g}^B$ and torque $\hat{{{\bf \tau}}}^B$ in \eqref{eq:kinetic anisotropic}.\\
\section{The long-time diffusivity of microstructured particles} \label{sec:Laginvin}
We now consider the diffusion of particles in space. Since the Reynolds number is low, it is widely accepted that the inertia terms (i.e. $m{\dot{\bfv}}$ and ${\hat{\bf I}} \dot{{\hat{\boldsymbol \omega}}}$) in \eqref{eq:kinetic anisotropic} could be neglected for the evolution of particles in configurational space of spatial position and orientation, which is often referred to as the over-damped theory.~\citep{Travitzoverdamped} If ${\hat{\bf R}}^{\rm tr}\equiv 0$, it is clear that the first of~\eqref{eq:kinetic anisotropic} is independent of the second of \eqref{eq:kinetic anisotropic}, and hence the translational diffusivity can be determined without considering rotational motions.
To account for the case of ${\hat{\bf R}}^{\rm tr}\neq 0$, we first notice that the angular velocity ${\hat{\boldsymbol \omega}}$ and rotation matrix ${\mathbf Q}$ are kinematically related in \eqref{eq:kinetic anisotropic}. Specifically, the angular velocity ${\hat{\boldsymbol \omega}}$, as the (pseudo-)vector associated with the skew-symmetric matrix ${\mathbf Q}\dot{{\mathbf Q}}^T$, must satisfy
\begin{equation} \label{Qdot}
{\mathbf Q}\dot{{\mathbf Q}}^T{\mathbf a}={\hat{\boldsymbol \omega}}\times {\mathbf a},\quad\forall {\mathbf a}\in{\mathbb{R}}^3,
\end{equation}
where the rotation matrix ${\mathbf Q}$ represents the orientation of the particle with respect to the global frame $\{{\mathbf e}_i: i=1,2, 3\}$. For future calculations, it is necessary to introduce some parametrization $\boldsymbol{\Theta}\in\mathbb{R}^{m}$ for the rigid rotations in the continuous group ${\rm SO}(3)$, e.g., the quaternion ($m=4$) or the Euler angles (m=$3$). Once the parametrization $\boldsymbol{\Theta}$ is chosen, a transformation matrix ${\mathbf T}\in\mathbb{R}^{m\times 3}$ can be introduced to relate the angular velocity ${\hat{\boldsymbol \omega}}$ and the rate of the change of ${\bm{\Theta}}$, i.e.,
\begin{eqnarray}\label{omegaTrelation}
\dot{\boldsymbol{\Theta}}={\mathbf T}{\hat{\boldsymbol \omega}}.
\end{eqnarray}
Next, we introduce the position-orientation coordinates
\begin{eqnarray}\label{tildex}
\tilde{{\mathbf x}}=\begin{bmatrix}{\mathbf x}\\\boldsymbol{\Theta}\end{bmatrix}\in\mathbb{R}^{3+m}.
\end{eqnarray}
By \eqref{omegaTrelation} the position-orientation coordinates $\tilde{{\mathbf x}}$ and generalized velocities $\tilde{{\mathbf v}}$ in~\eqref{eq: l} are related by
\begin{eqnarray}\label{eq: Ptensor}
\dot{\tilde{{\mathbf x}}}={\mathbf P}\tilde{{\mathbf v}}
\end{eqnarray}
where
\begin{eqnarray}\label{eq: Ptensor2}
{\mathbf P}=\begin{bmatrix}{\rm Id}&0\\0&{\mathbf T}\end{bmatrix}
\end{eqnarray}
with ${\rm Id}\in\mathbb{R}^{3\times3}$ denoting the identity tensor.
Based on equations \eqref{tildex}-\eqref{eq: Ptensor2} and the over-damped postulate, we rewrite \eqref{eq:kinetic anisotropic} as
\begin{eqnarray}\label{governingequationnew2}
\dot{\tilde{{\mathbf x}}}=-\bm{\tilde{L}}\tilde{{\mathbf g}}-\bm{\tilde{L}}\tilde{\bm{\sigma}}^*\tilde{\bm{\xi}},
\end{eqnarray}
where
\begin{widetext}
\begin{eqnarray}\label{haha}
\begin{split}
\bm{\tilde{L}}:=&\begin{bmatrix}
-\left({\mathbf R}^\tt-{\mathbf R}^{\rm tr}{\mathbf R}^{{\rm rr}^{-1}}{\mathbf R}^{{\rm tr}^T}\right)^{-1}
&\left({\mathbf R}^\tt-{\mathbf R}^{\rm tr}{\mathbf R}^{{\rm rr}^{-1}}{\mathbf R}^{{\rm tr}^T}\right)^{-1}{\mathbf R}^{{\rm tr}}{\mathbf R}^{{\rm rr}^{-1}}{\mathbf Q}^{T}\\
{\mathbf T}{\mathbf Q}\left({\mathbf R}^{\rm rr}-{\mathbf R}^{{\rm tr}^T}{\mathbf R}^{\tt^{-1}}{\mathbf R}^{\rm tr}\right)^{-1}{\mathbf R}^{{\rm tr}^T}{\mathbf R}^{\tt^{-1}}
&-{\mathbf T}{\mathbf Q}\left({\mathbf R}^{\rm rr}-{\mathbf R}^{{\rm tr}^T}{\mathbf R}^{\tt^{-1}}{\mathbf R}^{\rm tr}\right)^{-1}{\mathbf Q}^T
\end{bmatrix}, \qquad \qquad \tilde{{\mathbf g}}:=\begin{bmatrix} {\mathbf g}^e\\ \hat{{{\bf \tau}}}^e\end{bmatrix},
\end{split}
\end{eqnarray}
\end{widetext}
and $\tilde{\bm{\sigma}}^*$ denotes the generalized fluctuation coefficients that satisfy
\begin{eqnarray}\label{sigmabar}
\tilde{\bm{\sigma}}^*\tilde{\bm{\sigma}}^{*^T}=2k_BT\begin{bmatrix}{\mathbf Q}^T\hat{\bf R}^\tt {\mathbf Q}&{\mathbf Q}^T\hat{\bf R}^{\rm tr} \\ \hat{\bf R}^{{\rm tr}^T}\bf Q& \hat{\bf R}^{\rm rr}\end{bmatrix}.
\end{eqnarray}
Suppose that the external force and torque is conservative, meaning that there exists a potential field $V=V({\tilde{\bfx}})$ such that the rate of work done by the external force and torque is equal to the decrease rate of potential energy, i.e.,
\begin{eqnarray}\label{solvegetaue}
\begin{split}
-\frac{\mathrm{d}}{\mathrm{d}t}V(\tilde{{\mathbf x}})&=-{\mathbf v}\cdot\nabla_{\mathbf x} V(\tilde{{\mathbf x}})-\dot{{\bm{\Theta}}}\cdot \nabla_\Theta V(\tilde{{\mathbf x}})\\
&={\mathbf g}^e\cdot{{\mathbf v}}+\hat{{{\bf \tau}}}^e\cdot{\hat{\boldsymbol \omega}}.
\end{split}
\end{eqnarray}
From \eqref{tildex} and \eqref{solvegetaue} , it follows that
\begin{eqnarray}\label{eq: forceg}
{\mathbf g}^e=-\nabla_{\mathbf x} V(\tilde{{\mathbf x}}); \; \qquad \;\hat{{{\bf \tau}}}^e=-{\mathbf T}^{T} \nabla_\Theta V(\tilde{{\mathbf x}}).
\end{eqnarray}
As for \eqref{governingequationnew2}, we consider the PDF $P=P(\tilde{{\mathbf x}}, t)\equiv P({\mathbf x}, {\bm{\Theta}}, t)$ in the position-orientation space. Based on \eqref{haha}, \eqref{sigmabar}, and \eqref{eq: forceg}, the Fokker-Planck equation for the PDF $P({\mathbf x}, {\bm{\Theta}}, t)$ of the stochastic process governed by \eqref{governingequationnew2} can be written as
\begin{widetext}
\begin{eqnarray} \label{eq:keyFP}
\begin{split}
\frac{\partial }{\partial t} P({\mathbf x}, {\bm{\Theta}}, t)=\nabla_{\mathbf x}\cdot &\Big[\Big(-\overline{{\boldsymbol{ \alpha}}}^\tt P({\mathbf x}, {\bm{\Theta}}, t)+\frac{1}{2}\overline{{{\boldsymbol{\beta}}}}^\tt\nabla_{\mathbf x} P({\mathbf x}, {\bm{\Theta}}, t)+\frac{1}{2}\overline{{{\boldsymbol{\beta}}}}^{\rm tr}\nabla_{\bm{\Theta}} P({\mathbf x}, {\bm{\Theta}}, t)\Big)\Big]\\
+\nabla_\Theta\cdot&\Big[\Big(-\overline{{\boldsymbol{ \alpha}}}^{\rm rr} P({\mathbf x}, {\bm{\Theta}}, t)+ \frac{1}{2}\overline{{{\boldsymbol{\beta}}}}^{\rm rr}\nabla_{\bm{\Theta}} P({\mathbf x}, {\bm{\Theta}}, t)+\frac{1}{2}\overline{{{\boldsymbol{\beta}}}}^{{\rm tr}^T}\nabla_{\mathbf x} P({\mathbf x}, {\bm{\Theta}}, t)\Big)\Big],
\end{split}
\end{eqnarray}
\end{widetext}
where
\begin{widetext}
\begin{eqnarray} \label{eq:alphabeta}
\begin{split}
\begin{bmatrix}
\overline{{\boldsymbol{ \alpha}}}^\tt\\
\overline{{\boldsymbol{ \alpha}}}^{\rm rr}
\end{bmatrix}=\begin{bmatrix}
\left({\mathbf R}^\tt-{\mathbf R}^{\rm tr}{\mathbf R}^{{\rm rr}^{-1}}{\mathbf R}^{{\rm tr}^T}\right)^{-1}\nabla_{\mathbf x} V- \left({\mathbf R}^\tt-{\mathbf R}^{\rm tr}{\mathbf R}^{{\rm rr}^{-1}}{\mathbf R}^{{\rm tr}^T}\right)^{-1}{\mathbf R}^{{\rm tr}}{\mathbf R}^{{\rm rr}^{-1}}{\mathbf Q}^{T}{\mathbf T}^{T} \nabla_\Theta V\\
{\mathbf T}{\mathbf Q}\left({\mathbf R}^{\rm rr}-{\mathbf R}^{{\rm tr}^T}{\mathbf R}^{\tt^{-1}}{\mathbf R}^{\rm tr}\right)^{-1}{\mathbf Q}^{T}{\mathbf T}^{T}\nabla_\Theta V-{\mathbf T}{\mathbf Q}\left({\mathbf R}^{\rm rr}-{\mathbf R}^{{\rm tr}^T}{\mathbf R}^{\tt^{-1}}{\mathbf R}^{\rm tr}\right)^{-1}{\mathbf R}^{{\rm tr}^T}{\mathbf R}^{\tt^{-1}} \nabla_{\mathbf x} V
\end{bmatrix}\\
\end{split}
\end{eqnarray}
\end{widetext}
and
\begin{widetext}
\begin{eqnarray}\label{key}
\begin{split}
\begin{bmatrix}
\overline{\bm{\beta}}^\tt
&\overline{\bm{\beta}}^{\rm tr}\\
\overline{\bm{\beta}}^{{\rm tr}^T}
&\overline{\bm{\beta}}^{\rm rr}
\end{bmatrix}&=\bm{\tilde{L}}\tilde{\bm{\sigma}}^*\tilde{\bm{\sigma}}^{*^T}\bm{\tilde{L}}^{T}=\begin{bmatrix}
2k_BT\left({\mathbf R}^\tt-{\mathbf R}^{\rm tr}{\mathbf R}^{{\rm rr}^{-1}}{\mathbf R}^{{\rm tr}^T}\right)^{-1}
&-2k_BT\left({\mathbf R}^{\rm rr}{\mathbf R}^{{\rm tr}^{-1}}{\mathbf R}^{\tt}-{\mathbf R}^{{\rm tr}^T}\right)^{-1}{\mathbf Q}^{T}{\mathbf T}^{T}\\
-2k_BT{\mathbf T}{\mathbf Q}\left({\mathbf R}^\tt{\mathbf R}^{{\rm tr}^{-T}}{\mathbf R}^{{\rm rr}}-{\mathbf R}^{{\rm tr}}\right)^{-1}
&2k_BT{\mathbf T}{\mathbf Q}\left({\mathbf R}^{\rm rr}-{\mathbf R}^{{\rm tr}^T}{\mathbf R}^{\tt^{-1}}{\mathbf R}^{\rm tr}\right)^{-1}{\mathbf Q}^{T}{\mathbf T}^{T}
\end{bmatrix}.
\end{split}
\end{eqnarray}
\end{widetext}
From \eqref{eq:keyFP}, we identify
$ \overline{\bm{\beta}}^\tt/2$ and $\overline{\bm{\beta}}^{\rm rr}/2$ as the macroscopic translational and rotational diffusivity, respectively. Typically, ${\bm{\Theta}}$ and $ {\bm{\Theta}}+2\pi$ refer to the same rotation. Then the crossover time scale $t_{\rm cross}$ can be estimated as
\begin{eqnarray} \label{timescale}
t_{\rm cross}\sim \frac{(2\pi)^2}{|{{\boldsymbol{\beta}}}^{\rm rr}|}.
\end{eqnarray}
Moreover, it is straightforward to verify that
\begin{eqnarray} \label{eq:PsTheta}
\begin{split}
&P^s({\mathbf x}, {\bm{\Theta}}, t)\propto \exp\Big({-\frac{V({\mathbf x}, {\bm{\Theta}})}{k_BT}}\Big) \\
\end{split}
\end{eqnarray}
is a stationary solution to \eqref{eq:keyFP}, which is consistent with the classical statistical mechanics.
We remark that both the diffusivity tensor $ \overline{\bm{\beta}}^\tt/2$ and $\overline{\bm{\beta}}^{\rm rr}/2$ in general depend on the current orientation ${\bm{\Theta}}$ of the particle (cf., \eqref{key}).
For a time scale that is much larger than the crossover time scale $t_{\rm cross}$ in \eqref{timescale}, it is anticipated that the orientation-dependence of translational diffusivity would be averaged out. If the external force ${\mathbf g}^e\equiv 0$, i.e., $V=V({\bm{\Theta}})$, we may
calculate the final effective translational diffusivity by considering trial solution $P=P(\tilde{{\mathbf x}}, t)$ to \eqref{eq:keyFP} of form:~\citep{Yuan2021}
\begin{eqnarray}\label{Ptranslation}
P({\mathbf x}, {\bm{\Theta}}, t) =\bar{P}({\mathbf x}, t)P^s({\bm{\Theta}}),
\end{eqnarray}
where $\bar{P}({\mathbf x}, t)$ denotes the PDF in spatial
position ${\mathbf x}$ and
\begin{eqnarray}\label{Ps}
P^s({\bm{\Theta}})\propto \exp\Big({-\frac{V({\bm{\Theta}})}{k_BT}}\Big)
\end{eqnarray}
is the stationary PDF in orientation ${\bm{\Theta}}$.
Inserting \eqref{Ptranslation} into \eqref{eq:keyFP} and integrating \eqref{eq:keyFP} over the ${\bm{\Theta}}$-space, we obtain the translational diffusion equation in the long-time limit:
\begin{eqnarray} \label{eq:keyFPx}
\begin{split}
\frac{\partial }{\partial t} \bar{P}({\mathbf x}, t)&= \nabla_{\mathbf x}\cdot \big[{\mathbf D}^{\rm eff} \nabla_{\mathbf x} \bar{P}({\mathbf x}, t)\big],
\end{split}
\end{eqnarray}
where
\begin{eqnarray} \label{eq:Deff general}
\begin{split}
{\mathbf D}^{\rm eff} &=\int_{{\rm SO}(3)} \frac{\overline{\bm{\beta}}^\tt}{2} P^s({\bm{\Theta}}) \\
&=k_BT\int_{{\rm SO}(3)}\left({\mathbf R}^\tt-{\mathbf R}^{\rm tr}{\mathbf R}^{{\rm rr}^{-1}}{\mathbf R}^{{\rm tr}^T}\right)^{-1}P^s({\bm{\Theta}})
\end{split}
\end{eqnarray}
is the effective diffusivity for the long-time diffusion.
The formula~\eqref{eq:Deff general} composes one of our main results concerning the diffusion of general microstructured particles in an external field. For brevity, we introduce the following mobility tensor in the body-frame:
\begin{eqnarray}\label{Met}
\hat{{\mathbf M}}=\left(\hat{{\mathbf R}}^\tt-\hat{{\mathbf R}}^{\rm tr}\hat{{\mathbf R}}^{{\rm rr}^{-1}}\hat{{\mathbf R}}^{{\rm tr}^T}\right)^{-1},
\end{eqnarray}
and recall that the drag tensors ${\hat{\bf R}}^{(\tt, {\rm tr}, {\rm rr})}$, and hence the mobility tensor $\hat{{\mathbf M}}$ in the body-frame are independent of orientation parameter ${\bm{\Theta}}$.
Then \eqref{eq:Deff general} can be rewritten as
\begin{eqnarray}\label{eq:Deff}
\begin{split}
{\mathbf D}^{\rm eff} = &k_BT\int_{{\rm SO}(3)}{\mathbf Q}^T({\bm{\Theta}})\hat{{\mathbf M}}{\mathbf Q}({\bm{\Theta}}) P^s({\bm{\Theta}}) .
\end{split}
\end{eqnarray}
Upon inspecting ~\eqref{eq:Deff general} or \eqref{eq:Deff}, we observe a few exact results which are listed below.
\noindent (i) Taking the trace of \eqref{eq:Deff} we find
\begin{eqnarray} \label{eq:TrD}
\begin{split}
{{\rm Tr}}({\mathbf D}^{\rm eff} ) &= k_BT\int_{{\rm SO}(3)} {{\rm Tr}} ({\mathbf Q}^T\hat{{\mathbf M}}{\mathbf Q}) P^s({\bm{\Theta}}) \\
&=k_BT {{\rm Tr}} (\hat{{\mathbf M}})\int_{{\rm SO}(3)} P^s({\bm{\Theta}}) =k_BT {{\rm Tr}} \hat{{\mathbf M}},
\end{split}
\end{eqnarray}
where the last equality follows from
\begin{eqnarray}\label{PSdefinition}
\int_{{\rm SO}(3)} P^s({\bm{\Theta}}) =1.
\end{eqnarray}
\noindent (ii) If ${\mathbf R}^{\rm tr}=0$, the translational motion is uncoupled with the rotational motions, \eqref{eq:Deff general} degenerates into
\begin{eqnarray}\label{hahaha}
{\mathbf D}^{\rm eff} =k_BT\int_{{\rm SO}(3)} {\mathbf R}^{\tt^{-1}} P^s({\bm{\Theta}}) ,
\end{eqnarray}
which was first derived in Yuan et al. \citep{Yuan2021} for homogenous axisymmetric particle.
\noindent (iii) At the absence of external field ($V\equiv0$), the long-time diffusivity of a free particle must be isotropic for an isotropic ambient fluid. Then by \eqref{eq:TrD}, we immediately find the isotropic long-time diffusivity of an arbitrary microstructured free particle is given by
\begin{eqnarray}\label{eq:Deff iso}
D^{\rm eff} =\frac{1}{3} k_BT{{\rm Tr}} (\hat{{\mathbf M}}) .
\end{eqnarray}
\noindent (iv) For spherical homogeneous particle of radius $r$ and at the absence of external fields,
the mobility tensor is isotropic and given by $\hat{M}=1/{6\pi\,\eta\,r}$ ($\eta$ is the viscosity). By \eqref{eq:Deff iso} we recover the celebrated Stokes-Einstein's relation:
\begin{eqnarray}\label{einstein relation}
D^{\rm eff} =\frac{k_\text{B} T}{6\pi\,\eta\,r}.
\end{eqnarray}
\section{Applications}\label{Sec:app}
In this section, we present some explicit formulas for the long-time diffusivity of heterogeneous spheroids which account for the effects of shape anisotropy and heterogeneities and may be regarded as generalizations of the classical Stokes-Einstein's formula. Based on these results, we study the diffusion of a pair of spheroids whose relative position and orientation are continuously relaxing from an initial non-equilibrium state to the final equilibrium state. The additional relaxation time-scales cause time-dependent diffusivity of the pair. Consequently, the mean squared displacement ($MSD$) scales differently with respect to time from the conventional Brownian motions because of the interplay between multiple time scales, i.e., the relaxation time-scales, crossover time-scale, and diffusion time-scale. We speculate that this may be a reasonable physical model to understand the anomalous diffusion observed in, e.g., the migration of macromolecules or cells in complex media.
\subsection{Effects of shape anisotropy and heterogeneity } \label{sec:Hetero}
Most of existing works on diffusions of microparticles assume uncoupled translational and rotational motions, i.e., the off-diagonal block ${\mathbf R}^{{\rm tr}}$ in the drag tensor \eqref{eq:bfRblock} vanishes. The presence of shape anisotropy and heterogeneity of the particle in general lead to the mismatch between the center of mass and the geometric centroid or hydrodynamic center of the particle and hence nonzero ${\mathbf R}^{\rm tr}$. In this section we consider heterogeneous spheroidal particles whose centers of mass deviate from their geometric centroids (or hydrodynamic centers). As will be shown shortly, the diffusivity of heterogeneous balls could be significantly different from the conventional Stokes-Einstein's formula if the mismatch between the center of mass and geometric center is large. This fundamental fact seems to be unnoticed in the literature and may find applications in, e.g., characterizing the uniformity of microparticles or separating microparticles of different uniformity.
To proceed, we recall that the equation of motion \eqref{eq:kinetic anisotropic000} for the particle is formulated with respect to the center of mass of the particle. On the other hand, the drag matrix is usually derived with respect to the centroid of the particle. To make a distinction, we denote by
\begin{eqnarray} \label{eq:Upsilon0}
\begin{split}
{\bm{\Upsilon}}:=\begin{bmatrix}{\bm{\Upsilon}}^\tt&{\bm{\Upsilon}}^{\rm tr}\\
({\bm{\Upsilon}}^{\rm tr})^T &{\bm{\Upsilon}}^{\rm rr}\end{bmatrix}
\end{split}
\end{eqnarray}
the drag matrix with respect to the centroid of the particle and use $\hat{{\bm{\Upsilon}}}$ (and sub-blocks $\hat{{\bm{\Upsilon}}}^{(\tt,{\rm rr},{\rm tr})}$) the drag matrices in the body-frame. For spheroids, the drag matrices $\hat{{\bm{\Upsilon}}}^{(\tt,{\rm rr},{\rm tr})}$ depends only on the shape of the particle and the viscosity of the ambient fluid and can be found in, e.g., the textbook of Kim and Karrila.~\citep{Kim1991Microhydrodynamics}
The drag matrices with respect to the center of mass $\hat{{\mathbf R}}^{(\tt,{\rm rr},{\rm tr})}$ in terms of $\hat{{\bm{\Upsilon}}}^{(\tt,{\rm tr},{\rm rr})}$ can then be obtained by a quick free-body-diagram analysis as (see, e.g., Brenner \cite{Brenner1965Coupling,Bernal1980})
{ \begin{eqnarray}\label{eq: R with c}
\begin{split}
&\hat{R}^\tt_{ij}={\hat{\Upsilon}}^\tt_{ij},\\
&\hat{R}^{\rm tr}_{ij}={\hat{\Upsilon}}^{\rm tr}_{ij}-\epsilon_{isk} c_k {\hat{\Upsilon}}^{\tt}_{sj},\\
&\hat{R}^{\rm rr}_{ij}={\hat{\Upsilon}}^{\rm rr}_{ij}-\epsilon_{ipq} \epsilon_{jks} c_q c_k {\hat{\Upsilon}}^{\tt}_{ps} +\epsilon_{jks} c_k {\hat{\Upsilon}}^{{\rm tr}}_{is}-\epsilon_{isk} c_k {\hat{\Upsilon}}^{{\rm tr}}_{sj},
\end{split}
\end{eqnarray}}
where
where ${\mathbf c}\in \mathbb{R}^3$ is the vector pointing from the centroid to the center of mass of the particle, and $\epsilon_{iks}$ the Levi-Civita symbol.%
\begin{figure}[H]
\centering
\includegraphics[width=8.7cm]{Configurationtwo.png}
\caption{Two typical configurations of heterogeneous micro-structured particles of prolate spheroid with axis ${\mathbf f}_1$: (a) the center of mass is on the axis of symmetry and (b) the center of mass is away from the axis of symmetry. The major semi-axis-length of the spheroidal particle is $a$, and the minor semi-axis-length of the spheroidal particle is $b$.}\label{configuration}
\end{figure}
As illustrated in Fig.~\ref{configuration}, we will consider two configurations of heterogeneous spheroidal particles. Denote by $a$ and $b$ the two principle semi-axis-lengths of the spheriod, and $e=a/b$ the aspect ratio. The body frame $\{{\mathbf f}_i: i=1,2,3\}$ is fixed at the centroid of the spheroidal particle with ${\mathbf f}_1$ being the axis of symmetry.
Because of the axis-symmetry, the off-diagonal block $\hat{\bm{\Upsilon}}^{\rm tr} $ vanishes whereas the diagonal components of $\hat{\bm{\Upsilon}}^\tt$ and $\hat{\bm{\Upsilon}}^{\rm rr}$ are given by ($\hat{\Upsilon}^\tt_{22}=\hat{\Upsilon}^\tt_{33}$, $\hat{\Upsilon}^{\rm rr}_{22}=\hat{\Upsilon}^{\rm rr}_{33}$ and all off-diagonal components vanish by symmetry) \cite{Kim1991Microhydrodynamics}
\begin{eqnarray}\label{eq: upsilon}
\begin{split}
&\hat{\Upsilon}^\tt_{11}=\frac{{8\pi \eta a{\left(e^2-1\right)^{\frac{3}{2}}}}}{e{\left[ {\left( 2e^2-1\right)\ln \left(e+\sqrt{e^2-1}\right) - e\sqrt{e^2-1}} \right]}},\\
&\hat{\Upsilon}^\tt_{22}= \frac{{16\pi \eta a{\left(e^2-1\right)^{\frac{3}{2}}}}}{e{\left[ {\left( 2e^2-3 \right)\ln \left(e+\sqrt{e^2-1}\right) +e\sqrt{e^2-1}} \right]}},\\
&\hat{\Upsilon}^{\rm rr}_{11}=\frac{{16\pi \eta a^3{\left(e^2-1\right)^{\frac{3}{2}}}}}{3e^3{\left[ {-\ln \left(e+\sqrt{e^2-1}\right) + e\sqrt{e^2-1}} \right]}},\\
&\hat{\Upsilon}^{\rm rr}_{22}= \frac{{16\pi \eta a^3{\left(e^2-1\right)^{\frac{3}{2}}}}(e^2+1)}{3e^3{\left[ {\left( 2e^2-1\right)\ln \left(e+\sqrt{e^2-1}\right) - e\sqrt{e^2-1}} \right]}}.
\end{split}
\end{eqnarray}
\subsubsection{Axisymmetric configuration} \label{sec:Aconfig}
In the first configuration illustrated in Fig.~\ref{configuration}(a), the center of mass is on the axis of symmetry with ${\mathbf c}=d{\mathbf f}_1$ for some $d\in[0,a)$.
Inserting \eqref{eq: R with c} into \eqref{Met} we find the mobility tensor is given by
\begin{eqnarray}\label{eq:Meffforf2}
\begin{split}
\hat{{\mathbf M}}=\mathrm{diag}\left(\frac{1}{{\hat{\Upsilon}}^\tt_{11}},\; \; \frac{1}{{\hat{\Upsilon}}^\tt_{22}}+\frac{d^2}{{\hat{\Upsilon}}^{\rm rr}_{22}},\;\;\frac{1}{{\hat{\Upsilon}}^\tt_{22}}+\frac{d^2}{{\hat{\Upsilon}}^{\rm rr}_{22}}\right).
\end{split}
\end{eqnarray}
Since the center of mass locates at the axis of symmetry,
it is sufficient to describe the orientation of the particle by the usual spherical coordinates ${\bm{\Theta}}=(\theta,\varphi)\in [0,\pi]\times[0,2\pi)$ for $S^2$ (unit sphere in ${\mathbb{R}}^3$), where $\theta$ is the angle between the symmetry axis ${\mathbf f}_1$ and ${\mathbf e}_1$, and $\varphi$ is the angle between ${\mathbf e}_2$ and the projected ray of ${\mathbf f}_1$ on the ${\mathbf e}_2$-${\mathbf e}_3$-plane.
For the diffusivity in the long-time limit, our goal is to evaluate the integral~\eqref{eq:Deff}. Being axis-symmetric, this integral over ${\rm SO}(3)$ reduces to an integral on $(\theta, \varphi)$ over $S^2$. Moreover, we find the rotation matrix ${\mathbf Q}$ as \citep{Yuan2021}
\begin{eqnarray}\label{Q electric2}
\begin{split}
{\mathbf Q}(\varphi ,\theta)
=&\begin{bmatrix}
\cos\theta &\sin\theta \cos\varphi & \sin\theta \sin\varphi \\
\sin\theta &-\cos\theta \cos\varphi & -\cos\theta \sin\varphi \\
0 & \sin\varphi & -\cos\varphi \\
\end{bmatrix},
\end{split}
\end{eqnarray}
and the stationary PDF $P^s(\theta,\varphi)$ in orientational space is given by (c.f., \eqref{Ps} and \eqref{PSdefinition})
\begin{eqnarray}\label{eq:psproperty2}
P^s(\theta,\varphi)= \frac{\sin\theta \exp\Big[-\frac{V(\theta,\varphi)}{k_BT}\Big]}{\int _0^{\pi }\!\!\!\int _0^{2 \pi } \sin\theta \exp\Big[-\frac{V(\theta,\varphi)}{k_BT}\Big]\;\mathrm{d}\theta\; \mathrm{d}\varphi}.
\end{eqnarray}
Inserting \eqref{eq:Meffforf2}
and \eqref{Q electric2} into \eqref{eq:Deff}, we find the diffusivity along the axes $\{{\mathbf e}_1, {\mathbf e}_2, {\mathbf e}_3\}$ of the global frame as:
\begin{eqnarray}
\label{ett electricinitial}
\begin{split}
D^{{\rm eff}}_{11}=&k_BT\big[\hat{M}_{11}- (\hat{M}_{11}-\hat{M}_{22}) \omega \big] ,\\
D^{{\rm eff}}_{22}=&k_BT\big[\hat{M}_{22}+ (\hat{M}_{11}-\hat{M}_{22})\omega' \big]\\
D^{{\rm eff}}_{33}=&k_BT\big[\hat{M}_{22}+ (\hat{M}_{11}-\hat{M}_{22})(\omega-\omega') \big],\\
\end{split}
\end{eqnarray}
where $\hat{M}_{11},\hat{M}_{22},\hat{M}_{33}$ are given by \eqref{eq:Meffforf2}, and
\begin{eqnarray} \label{eq:omega}
\begin{split}
&\omega=\int _0^{\pi }\!\!\!\int _0^{2 \pi }\!\!\!P^s(\theta,\varphi) \sin^2\theta\;\mathrm{d}\theta\; \mathrm{d}\varphi,\\
&\omega'=\int _0^{\pi }\!\!\!\int _0^{2 \pi }\!\!\!P^s(\theta,\varphi) \cos^2\varphi\sin^2\theta\;\mathrm{d}\theta\; \mathrm{d}\varphi,\\
\end{split}
\end{eqnarray}
Once the external potential $V(\theta,\varphi)$ is prescribed, one can simply evaluate the integrals in \eqref{eq:omega} to obtain the diffusivities in \eqref{ett electricinitial}. In general, we expect nontrivial off-diagonal components in the diffusivity tensor. For explicit results, we consider two special scenarios.
\noindent (i) The external field, e.g., a strong applied magnetic field along ${\mathbf e}_1$-direction, tends to align the spheroid axis ${\mathbf f}_1$ with ${\mathbf e}_1$.
The effect of this external field can be modeled by the external potential $V(\theta,\varphi)=E_0\sin^2\theta$. By symmetry it is easy to see all off-diagonal components of diffusivity tensor ${\mathbf D}^{\rm eff}$ vanish and $D^{{\rm eff}}_{22}=D^{{\rm eff}}_{33}$. Moreover, by \eqref{ett electricinitial} we find that
\begin{eqnarray}
\label{ett electric}
\begin{split}
D^{{\rm eff}}_{11}=&k_BT\big[\hat{M}_{11}- (\hat{M}_{11}-\hat{M}_{22}) \omega \big] ,\\
D^{{\rm eff}}_{22}=& k_BT\big[\hat{M}_{22}+ \frac{\hat{M}_{11}-\hat{M}_{22}}{2}\omega \big],
\end{split}
\end{eqnarray}
Let
\begin{eqnarray*}
\sigma^2={k_BT\over E_0}
\end{eqnarray*}
be the dimensionless parameter for the strength of alignment field. If
$ \sigma^2 \ll 1$, i.e., the spheroid is aligned along the field direction and weakly fluctuates, by \eqref{eq:psproperty2} the first integral in \eqref{eq:omega} is well-approximated by
\begin{eqnarray*}
\omega\approx{\int_0^\infty s^3 e^{-s^2/\sigma^2} ds\over \int_0^\infty s e^{-s^2/\sigma^2} ds}=\sigma^2.
\end{eqnarray*}
Therefore, the diffusivities in \eqref{ett electric} are approximately given by
\begin{eqnarray}
\label{ett electric1}
\begin{split}
&D^{{\rm eff}}_{11}\approx \frac{k_BT}{\hat{\Upsilon}^\tt_{11}} [1+(\frac{\hat{\Upsilon}^\tt_{11}}{{\hat{\Upsilon}^\tt_{22}}}-1)\sigma^2-\frac{\hat{\Upsilon}^\tt_{11}}{\hat{\Upsilon}^{\rm rr}_{22}}d^2\sigma^2],\\
& D^{{\rm eff}}_{22} \approx \frac{k_BT}{\hat{\Upsilon}^\tt_{22}} \big[1+\frac{\hat{\Upsilon}^\tt_{22}}{\hat{\Upsilon}^{\rm rr}_{22}}d^2-\frac{\sigma^2}{2}(1-\frac{\hat{\Upsilon}^\tt_{22}}{{\hat{\Upsilon}^\tt_{11}}})+\frac{\hat{\Upsilon}^\tt_{22}}{2\hat{\Upsilon}^{\rm rr}_{22}}d^2\sigma^2\big],
\end{split}
\end{eqnarray}
which may be compared with the results in Yuan {\em et al.}\citep{Yuan2021} In particular, we notice that the corrections in diffusivities because of the heterogeneities depend on $\sigma$. As the magnitude of the external field tends to infinity, we have $\sigma\to 0$, and hence
\begin{eqnarray} \label{eq:Daligned}
\begin{split}
&D^{{\rm eff}}_{11}=k_BT\hat{M}_{11}=\frac{k_BT}{\hat{\Upsilon}^\tt_{11}} ,\\
&D^{{\rm eff}}_{22} =\frac{ k_BT(\hat{M}_{22}+\hat{M}_{33})}{2}=\frac{k_BT}{\hat{\Upsilon}^\tt_{22}}(1+{\hat{\Upsilon}^\tt_{22} d^2\over \hat{\Upsilon}^{\rm rr}_{22}}).
\end{split}
\end{eqnarray}
From the above expressions, we see that the diffusivity along the axis-direction is independent of the heterogeneity parameter $d$ since the spheroid is always aligned with the (strong) external field direction, i.e., ${\mathbf f}_1\equiv{\mathbf e}_1$. In contrast, the translational motions on the transverse plane are coupled with the rotation around the axis, giving rise to $d$-dependent diffusivity $D^{{\rm eff}}_{22}$ in the transverse directions.
\noindent (ii) At the absence of external field, i.e., $V(\theta,\varphi)\equiv 0$, by directly evaluating the integrals in \eqref{eq:psproperty2} and \eqref{ett electric} we find that the diffusivity tensor is indeed isotropic, and the diffusivity (along any direction) is given by
\begin{eqnarray}\label{Deffiso3}
\begin{split}
{D}^{\rm eff}&=\frac{k_BT}{3}\frac{{\hat{\Upsilon}}^{\rm rr}_{22}(2{\hat{\Upsilon}}^\tt_{11}+{\hat{\Upsilon}}^\tt_{22})+2d^2{\hat{\Upsilon}}^\tt_{22}{\hat{\Upsilon}}^\tt_{11}}{{\hat{\Upsilon}}^\tt_{11}{\hat{\Upsilon}}^\tt_{22}{\hat{\Upsilon}}^{\rm rr}_{22}}\\
&={k_BT\over 6\pi a \eta} (\gamma_0+\gamma_d) ,
\end{split}
\end{eqnarray}
where
\begin{eqnarray} \label{eq:alphad}
\begin{split}
\gamma_0&=\frac{ e \ln (e+\sqrt{e^2-1})}{ \sqrt{e^2-1} },\\
\gamma_d&={3 e^4 \sqrt{e^2-1}- 3 e^2 \left(2 e^2-1\right) \ln (e+\sqrt{e^2-1})\over 4 \left(e^2-1\right)^{3/2} \left(e^2+1\right) } {d^2\over a^2} .
\end{split}
\end{eqnarray}
Compared with the Stokes-Einstein's formula~\eqref{einstein relation}, we recognize the dimensionless coefficients $\gamma_0$ reflects the effect of shape anisotropy whereas the dimensionless coefficients $\gamma_d$ signifies the importance of heterogeneity. In particular, if the particle is spherical with $e=1$, the diffusivity is given by ($r=a=b$)
\begin{eqnarray} \label{eq:Deffball}
{D}^{\rm eff}=\frac{k_\text{B} T}{6\pi\,\eta\,r}\left(1+\frac{1}{2}\frac{d^2}{r^2}\right),
\end{eqnarray}
which can be regarded as a generalization of the classic Stokes-Einstein's formula for heterogeneous spherical particles.
In Fig.~\ref{inhomo2}, we consider five different aspect ratios ($e=1,2,5,10,20$) at the absence of an external field and plot the normalized diffusivity $D^{\rm eff}/D^{\rm eff}_0=1+\gamma_d/\gamma_0$ versus $d/a$, where$D^{\rm eff}_0=D^{\rm eff}\big|_{d=0}$. It can be seen that the effects of heterogeneity is more pronounced for larger aspect ratio $e$. For a spherical particle the classic Stokes-Einstein's formula underestimates the diffusivity by $12.5\%$ if the heterogeneity gives rise to $d/r=1/2$ (c.f., \eqref{eq:Deffball}).
\begin{figure}[H]
\flushleft
\includegraphics[width=9.5cm]{inhomofive.png}
\caption{The nondimensionalized diffusivity $D^{\rm eff}/D^{\rm eff}_0=1+\gamma_d/\gamma_0$ of heterogeneous spheroids as illustrated in Fig.\protect\ref{configuration} (a) versus the normalized distance $\frac{d}{a}$, where $d$ is the distance between the center of mass and the centroid of the spheroid, $a$ is the major semi-axis-length of the particle, and $e=a/b$ is the aspect ratio.}\label{inhomo2}
\end{figure}
\subsubsection{Non-axisymmetric configuration}
In the second configuration illustrated in Fig.~\ref{configuration}(b), the vector pointing from the centroid of the particle to the center of mass of the particle is assumed to be ${\mathbf c}=d{\mathbf f}_2$ for some $d\in[0,b)$.
Substituting \eqref{eq: R with c} with ${\mathbf c}=d{\mathbf f}_2$ into \eqref{Met} yields
\begin{eqnarray}\label{eq:Meffforf1}
\begin{split}
\hat{{\mathbf M}}=\mathrm{diag}\left(\frac{1}{{\hat{\Upsilon}}^\tt_{11}}+\frac{d^2}{{\hat{\Upsilon}}^{\rm rr}_{22}},\; \; \frac{1}{{\hat{\Upsilon}}^\tt_{22}},\;\;\frac{1}{{\hat{\Upsilon}}^\tt_{22}}+\frac{d^2}{{\hat{\Upsilon}}^{\rm rr}_{11}}\right).
\end{split}
\end{eqnarray}
For this case, the center of mass deviates from the axis of symmetry and breaks the axis-symmetry and the integral \eqref{eq:Deff} over ${\rm SO}(3)$ can no longer be reduced to an integral over $S^2$. Nevertheless, as detailed in Appendix~\ref{Sec:AdxA} we recognize the homomorphism between rigid rotations in ${\rm SO}(3)$ and unit quaternions ${\mathbf q}\in S^3$ (unit sphere in ${\mathbb{R}}^4$), and then employ spherical coordinates ${\bm{\Theta}}=(\psi,\theta,\varphi)\in U\equiv [0,\pi]\times[0,\pi]\times[0,2\pi]$ for $S^3$ to parametrize unit quaternions and rotations. More precisely, a unit quaternion ${\mathbf q}=x_0+x_1{\mathbf i}+x_2{\mathbf j}+x_3{\mathbf k} $ is represented as
\begin{eqnarray}\label{parameterization}
\begin{split}
&x_0= \cos\psi,\\
&x_1= \sin\psi\cos\theta,\\
&x_2= \sin\psi\sin\theta\cos\varphi,\\
&x_3= \sin\psi\sin\theta\sin\varphi,
\end{split}
\end{eqnarray}
and the associated rigid rotation $
{\mathbf Q} $ in terms of ${\bm{\Theta}}=(\psi,\theta,\varphi)$ is given by \eqref{Q electric1}.
Moreover, the stationary PDF $P^s(\psi,\theta,\varphi)$ in orientational space is now given by (c.f., \eqref{Ps} and \eqref{PSdefinition})
\begin{eqnarray}\label{eq:psproperty0}
P^s\!(\!\psi,\!\theta,\!\varphi\!)\ \propto {\sin ^2\psi \sin\!\!\theta \exp\!\!\Big[-\frac{V(\psi,\theta,\varphi)}{k_BT}\Big]} .
\end{eqnarray}
Inserting \eqref{eq:Meffforf1}
and \eqref{Q electric1} into \eqref{eq:Deff}, we find the diffusivity along the axes $\{{\mathbf e}_1, {\mathbf e}_2, {\mathbf e}_3\}$ of the global frame as:
\begin{eqnarray}\label{deffpotential}
\begin{split}
D^{\rm eff}_{11}=&k_BT \Big(\lambda_{11}\hat{M}_{11} + \lambda_{12}\hat{M}_{22} + \lambda_{13}\hat{M}_{33} \Big),\\
D^{\rm eff}_{22}=&k_BT \Big(\lambda_{21}\hat{M}_{11} + \lambda_{22}\hat{M}_{22} + \lambda_{23}\hat{M}_{33} \Big),\\
D^{\rm eff}_{33}=&k_BT \Big(\lambda_{31}\hat{M}_{11} + \lambda_{32}\hat{M}_{22} + \lambda_{33}\hat{M}_{33}\Big),
\end{split}
\end{eqnarray}
where $\lambda_{ij}\;(i,j=1, 2,3)$, listed in \eqref{lambda} in Appendix~\ref{sec:AdxB}, are dimensionless parameters that depends on the PDF in \eqref{eq:psproperty0} and $\hat{M}_{11},\hat{M}_{22},\hat{M}_{33}$ listed in \eqref{eq:Meffforf1}.
Once the external potential $V(\psi, \theta,\varphi)$ is prescribed, we can evaluate the integrals in \eqref{lambda} for $\lambda_{ij}\;(i,j=1, 2,3)$ and obtain the diffusivities along each axis direction. In general, we expect nontrivial off-diagonal components in the diffusivity tensor. Below we consider two special scenarios for explicit results.
\noindent (i) The external field is a strong field along ${\mathbf e}_1$-direction that tends to align the spheroid axis ${\mathbf f}_1$ with ${\mathbf e}_1$. The effect of this external field can be modeled by the external potential $V(\psi, \theta,\varphi)=E_0[1-({\mathbf f}_1\cdot{\mathbf e}_1)^2]=E_0[1-Q_{11}^2]$ $(E_0\gg k_BT)$ where the expression of $Q_{11}$ is given in $\eqref{Q electric1}_1$. Then up to the order of $O(\sigma^2)$ the parameter matrix $[\lambda_{ij}]\;(i,j=1, 2,3)$ are given by ($\sigma^2=k_BT/E_0\ll1$)
\begin{eqnarray}\label{lambdaforspheroid}
\begin{bmatrix}
\lambda_{ij}
\end{bmatrix}= \frac{\;1}{\;2} \begin{bmatrix}
2-2\sigma^2&{\sigma^2 }& {\sigma^2 }\\
{\sigma^2 }& 1-\sigma^2 & 1-\sigma^2 \\
{\sigma^2 }&1-\sigma^2 &1-\sigma^2 \\
\end{bmatrix} +o(\sigma^2). \qquad
\end{eqnarray}
Therefore, to the leading order approximation the diffusivities in \eqref{deffpotential} are given by
\begin{eqnarray}
\label{ett electric2}
\begin{split}
&D^{{\rm eff}}_{11}{\approx} \frac{k_BT}{\hat{\Upsilon}^\tt_{11}} [1+\frac{\hat{\Upsilon}^\tt_{11}d^2}{\hat{\Upsilon}^{\rm rr}_{22}}+\frac{1}{2}(\frac{2\hat{\Upsilon}^\tt_{11}}{{\hat{\Upsilon}^\tt_{22}}}+\frac{\hat{\Upsilon}^\tt_{11}d^2}{\hat{\Upsilon}^{\rm rr}_{11}}-\frac{2\hat{\Upsilon}^\tt_{11}d^2}{\hat{\Upsilon}^{\rm rr}_{22}}-2)\sigma^2],\\
&D^{{\rm eff}}_{22}=D^{{\rm eff}}_{33}{\approx} \frac{k_BT}{\hat{\Upsilon}^\tt_{22}}\Big [1+\frac{\hat{\Upsilon}^\tt_{22}d^2}{2\hat{\Upsilon}^{\rm rr}_{11}}\\
&\quad\qquad\qquad+\frac{1}{2}(\frac{\hat{\Upsilon}^\tt_{22}}{{\hat{\Upsilon}^\tt_{11}}}+\frac{\hat{\Upsilon}^\tt_{22}d^2}{\hat{\Upsilon}^{\rm rr}_{22}}-\frac{\hat{\Upsilon}^\tt_{22}d^2}{{\hat{\Upsilon}^\tt_{11}}}-1)\sigma^2\Big].
\end{split}
\end{eqnarray}
As the magnitude of the external field tends to infinity, i.e., $\sigma\to 0$, by \eqref{ett electric2} we find that
\begin{eqnarray*}
\begin{split}
&D^{{\rm eff}}_{11}=k_BT\hat{M}_{11}=k_BT(\frac{1}{\hat{\Upsilon}^\tt_{11}}+ \frac{d^2}{\hat{\Upsilon}^{\rm rr}_{22}}),\\
&D^{{\rm eff}}_{22}= D^{{\rm eff}}_{33}= \frac{ k_BT(\hat{M}_{22}+\hat{M}_{33})}{2}=k_BT(\frac{1}{\hat{\Upsilon}^\tt_{22}}+ \frac{d^2}{2\hat{\Upsilon}^{\rm rr}_{11}}) .
\end{split}
\end{eqnarray*}
We remark that
unlike the axisymmetric configuration discussed earlier (c.f.,~\eqref{eq:Daligned}), the diffusivity along the axis-direction with center of mass deviating from axis depends on the deviation distance $d$ even if the spheroid is forced to align with the external field. This counterintuitive effect arises from the coupling between the rotational and translational motions (c.f., \eqref{eq:kinetic anisotropic}), causing the fluctuation in rotations increases the fluctuation in translations and hence the diffusivity along ${\mathbf e}_1$-direction.
\noindent (ii) At the absence of external field, i.e., $V(\psi, \theta,\varphi)\equiv 0$, from the discussion in Appendix~\ref{Sec:AdxA} the stationary PDF $P^s(\psi,\theta,\varphi)$ can be written as
\begin{eqnarray}\label{eq:psproperty1}
P^s\!(\!\psi,\!\theta,\!\varphi\!) ={1\over 2\pi^2} \sin ^2\psi \sin\!\!\theta .
\end{eqnarray}
Upon directly evaluating the dimensionless parameters $\lambda_i$ listed in \eqref{lambda}, we find that the diffusivity tensor is indeed isotropic, and the diffusivity (along any direction) is given by
\begin{eqnarray}\label{Deffiso2}
\begin{split}
{D}^{\rm eff}=&\frac{k_BT}{3}\frac{2{\hat{\Upsilon}}^{\rm rr}_{11}{\hat{\Upsilon}}^{\rm rr}_{22}({\hat{\Upsilon}}^\tt_{11}+{\hat{\Upsilon}}^\tt_{22})+d^2{\hat{\Upsilon}}^\tt_{11}{\hat{\Upsilon}}^\tt_{22}({\hat{\Upsilon}}^{\rm rr}_{11}+{\hat{\Upsilon}}^{\rm rr}_{22})}{{\hat{\Upsilon}}^\tt_{11}{\hat{\Upsilon}}^\tt_{22}{\hat{\Upsilon}}^{\rm rr}_{11}{\hat{\Upsilon}}^{\rm rr}_{22}}\\
=&{k_BT\over 6\pi a \eta} (\gamma_0+\gamma'_d),
\end{split}
\end{eqnarray}
where $\gamma_0$ is given by $\eqref{eq:alphad}_1$, and
\begin{eqnarray} \label{eq:alphad2}
\begin{split}
\gamma'_d&=\frac{3 e^3 \left(\left(e^2-2\right) \ln \left(\sqrt{e^2-1}+e\right)+\sqrt{e^2-1} e^3\right)}{8 \left(e^2-1\right)^{3/2} \left(e^2+1\right)} \frac{d^2}{a^2} .
\end{split}
\end{eqnarray}
We notice that $\gamma'_d$ is distinct from $\eqref{eq:alphad}_2$, meaning that the heterogeneity along different directions has different effects on the diffusivity for shape anisotropy.
In Fig.~\ref{inhomo1}, we consider five different aspect ratios ($e=1,2,5,10,20$) at the absence of an external field and plot the normalized diffusivity $D^{\rm eff}/D^{\rm eff}_0=1+\gamma'_d/\gamma_0$ versus $d/a$. In contrast to Fig.~\ref{inhomo2}, we see that the impacts of heterogeneity is more remarkable for smaller aspect ratio $e$. In summary, as demonstrated in Fig.~\ref{inhomo2} and Fig.~\ref{inhomo1} the heterogeneity could significantly increase the diffusivity of particles.
\begin{figure}[H]
\flushleft
\includegraphics[width=9.5cm]{inhomofive2.png}
\caption{The nondimensionalized diffusivity $D^{\rm eff}/D^{\rm eff}_0=1+\gamma'_d/\gamma_0$ of heterogeneous spheroids as illustrated in Fig. \protect\ref{configuration} (b) versus the normalized distance $\frac{d}{a}$, where $d$ is the distance between the center of mass and the centroid of the spheroid, $a $ is the major semi-axis-length of the particle, and $e=a/b$ is the aspect ratio.}\label{inhomo1}
\end{figure}
\subsection{Anomalous diffusion of a pair of elastically bonded spheroids} \label{sec:Pairs}
Based on results in Section~\ref{sec:Hetero}, in this section we propose a model for anomalous diffusions by considering a pair of spheroids in a relaxation process. As illustrated in Fig.~\ref{fig: Configuration pair degenerate}, suppose two identical homogeneous spheroids are bonded by some elastic ligaments. Suppose the axes of the two spheroids are on the same plane and form an angle $\vartheta$ and the distance between the centroids of two spheroids is given by $d$. We are interested in the long-time diffusivity of such a microstructured particle and how the diffusivity depends on the angle and distance $(\vartheta, d)$ and external fields.
\begin{figure}[H]
\centering
\includegraphics[width=6cm,height=5cm]{spheroidpairdegenerate.jpg
\caption{A simple geometric structure of the pair of spheroidal micro-structured particles with only 2 degrees of freedom. The semi-axis-length parallel to the axis of symmetry of the particle is $a$, and the semi-axis-length perpendicular to the axis of symmetry of the particle is $b$. The two spheroids are bonded by some ligaments (green line) }\label{fig: Configuration pair degenerate}
\end{figure}
For simplicity, we neglect the hydrodynamic interactions between the two spheroids in the sense that the force and torque on the centroid of each spheroid is given by \eqref{frictiontotal} with the drag tensor in the body frame specified by \eqref{eq: upsilon}. By a free-body-diagram analysis, we find the nonzero components of the blocks of the drag matrix for the pair with axis angle and separation distance $(\vartheta, d)$ can be written as
\begin{eqnarray}
\begin{split}
\hat{R}^\tt_{11}(d,\vartheta)=& \hat{\Upsilon}^\tt_{11} +\hat{\Upsilon}^\tt_{22}+(\hat{\Upsilon}^\tt_{11}-\hat{\Upsilon}^\tt_{22})\cos \vartheta,\\
\hat{R}^\tt_{22}(d,\vartheta)=& \hat{\Upsilon}^\tt_{11} +\hat{\Upsilon}^\tt_{22}-(\hat{\Upsilon}^\tt_{11}-\hat{\Upsilon}^\tt_{22})\cos \vartheta,\\
\hat{R}^\tt_{33}(d,\vartheta)=&2 \hat{\Upsilon}^\tt_{22},\\
\hat{R}^{\rm rr}_{11}(d,\vartheta)=& \frac{d^2 \hat{\Upsilon}^\tt_{22}}{2}+(\hat{\Upsilon}^{\rm rr}_{11}-\hat{\Upsilon}^{\rm rr}_{22}) \cos \vartheta+\hat{\Upsilon}^{\rm rr}_{11}+\hat{\Upsilon}^{\rm rr}_{22}\\
\hat{R}^{\rm rr}_{22}(d,\vartheta)=&(\hat{\Upsilon}^{\rm rr}_{22}-\hat{\Upsilon}^{\rm rr}_{11}) \cos \vartheta+\hat{\Upsilon}^{\rm rr}_{11}+\hat{\Upsilon}^{\rm rr}_{22},\\
\hat{R}^{\rm rr}_{33}(d,\vartheta)=&\frac{1}{4} \left(d^2 (\hat{\Upsilon}^\tt_{11}-\hat{\Upsilon}^\tt_{22}) \cos \vartheta+d^2 (\hat{\Upsilon}^\tt_{11}+\hat{\Upsilon}^\tt_{22})+8 \hat{\Upsilon}^{\rm rr}_{22}\right)\\
\hat{R}^{\rm tr}_{32}(d,\vartheta)=&\frac{1}{2} d (\hat{\Upsilon}^\tt_{11}-\hat{\Upsilon}^\tt_{22}) \sin \vartheta,\\
\end{split}
\end{eqnarray}
where $\hat{{\Upsilon}}^\tt_{11}, \hat{{\Upsilon}}^\tt_{22}$ and $\hat{{\Upsilon}}^{\rm rr}_{11}, \hat{{\Upsilon}}^{\rm rr}_{22}$, given by \eqref{eq: upsilon}, are the components of the diagonal blocks of the drag matrix for a single spheroid.
We first consider the diffusion of the pair with fixed $(d,\vartheta)$. At the absence of external fields, the stationary PDF $P^s(\psi,\theta,\varphi)$ is given by \eqref{eq:psproperty1}. Upon repeating the procedure from \eqref{eq:Meffforf1} to \eqref{Deffiso2},
we find that the diffusivity tensor is indeed isotropic, and the diffusivity (along any direction) is given by
\begin{eqnarray}\label{deffforthepair}
\begin{split}
&D^{\rm eff} (\vartheta, d)=\!\!\!\frac{k_BT}{3}\!\Big[\frac{1}{2\hat{\Upsilon}^\tt_{22}}+\frac{1}{ 2\hat{\Upsilon}^\tt_{11} \cos^2 \frac{\vartheta}{2} +2\hat{\Upsilon}^\tt_{22} \sin^2 \frac{\vartheta}{2} }+\\
&\quad+\frac{4\hat{\Upsilon}^{\rm rr}_{22}+d^2( \hat{\Upsilon}^\tt_{11} \cos^2 \frac{\vartheta}{2} +\hat{\Upsilon}^\tt_{22} \sin^2 \frac{\vartheta}{2})}{8\hat{\Upsilon}^{\rm rr}_{22}(\hat{\Upsilon}^\tt_{11} \sin^2 \frac{\vartheta}{2} +\hat{\Upsilon}^\tt_{22} \cos^2 \frac{\vartheta}{2})+2d^2\hat{\Upsilon}^\tt_{11}\hat{\Upsilon}^\tt_{22}}\Big].
\end{split}
\end{eqnarray}
{As $d\to +\infty$, i.e., the two spheroids are far apart, we have
\begin{eqnarray}\label{converge}
\begin{split}
D^{\rm eff} \to
&\frac{k_BT}{3}\!\Big[\frac{1}{ 2\hat{\Upsilon}^\tt_{22}}+\frac{(\hat{\Upsilon}^\tt_{11}-\hat{\Upsilon}^\tt_{22}) \cos \vartheta+\hat{\Upsilon}^\tt_{11}+\hat{\Upsilon}^\tt_{22}}{4 \hat{\Upsilon}^\tt_{11} \hat{\Upsilon}^\tt_{22}}\\
&+\frac{1}{(\hat{\Upsilon}^\tt_{11}-\hat{\Upsilon}^\tt_{22}) \cos \vartheta+\hat{\Upsilon}^\tt_{11}+\hat{\Upsilon}^\tt_{22}}\Big].
\end{split}
\end{eqnarray}
}
{In particular, if $\vartheta=0$ or $\vartheta=\pi$, i.e., the two spheroids are aligned, we have
\begin{eqnarray}\label{deffforthepairal}
\begin{split}
D^{\rm eff} = \frac{k_BT}{6}\Big[ \frac{ 1}{ \hat{\Upsilon}^\tt_{11} }+ \frac{ 2}{ \hat{\Upsilon}^\tt_{22} } \Big],
\end{split}
\end{eqnarray}
which can also be derived from \eqref{hahaha} since ${\mathbf R}^{\rm tr}={\bf 0}$. }
If, in particular, the spheroid is a ball such that $\hat{\Upsilon}^\tt_{22}=\hat{\Upsilon}^\tt_{11}$,
\begin{eqnarray*}
D^{\rm eff}
= \frac{ k_BT}{2 \hat{\Upsilon}^\tt_{11}} ,
\end{eqnarray*}
which is precisely half of the diffusivity of a single spherical particle.
To see the influence of structural parameters $(\vartheta, d)$ of the pair on the long-time diffusivity, we consider two cases: a pair of prolates with major and minor semi-axis-length $1 \mu m$ and $0.1 \mu m$, respectively, and a pair of oblates with the same volume and minor axis-length. In Fig.~\ref{pairtheta} we plot the normalized diffusivity $D^{\rm eff}/D^{\rm eff}_0$ versus $\vartheta$ for { $d=2\mu m$}, where $D^0=\frac{k_BT}{3}{{\rm Tr}} \;[\hat{{\bm{\Upsilon}}}^{\tt^-1}]$ denotes the diffusivity of a single spheroid. Meanwhile, the normalized diffusivity $D^{\rm eff}/D^{\rm eff}_0$ is plotted against $d$ in Fig.~\ref{paird} for $\vartheta=\pi/2$.
\begin{figure}[H]
\centering
\includegraphics[width=9.3cm]{spheroidpairfortheta.png
\caption{
The nondimensionalized long-time diffusivity $\frac{{D}^{\rm eff}}{D^0}$ of the pair of spheroids (as illustrated in Fig.\protect\ref{fig: Configuration pair degenerate}) versus the angle $\vartheta$ between two axes of spheroids. The major and minor semi-axis-length of spheroid are $1 \mu m$ and $0.1 \mu m$ for prolate spheroids (blue), respectively. The oblate spheroids (red) have the same volume and minor-axis length. The distance $d$ between the centroids of the two particles is fixed at $2\mu m$.
}\label{pairtheta}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=9.3cm]{spheroidpairford.png
\caption{
The nondimensionalized long-time diffusivity $\frac{{D}^{\rm eff}}{D^0}$ of the pair of spheroids (as illustrated in Fig.\protect\ref{fig: Configuration pair degenerate}) versus the distance $d$ between two spheroids. The major and minor semi-axis-length of spheroid are $1 \mu m$ and $0.1 \mu m$ for prolate spheroids (blue), respectively. The oblate spheroids (red) have the same volume and minor-axis length. The angle $\vartheta$ between the two axes of symmetry is fixed at $\frac{\pi}{2}$.
}\label{paird}
\end{figure}
From Fig.~\ref{pairtheta} we see that the diffusivity ${D}^{\rm eff}$ of the pair is unsurprisingly lower than that of a single particle since the size of the pair is larger (c.f., the Stokes-Einstein relation~\eqref{einstein relation}). On the other hand, there exist some optimal angles $\vartheta\in [0,\pi]$ for which the diffusivity ${D}^{\rm eff}$ of the pairs is either minimized or maximized.
From Fig.~\ref{paird} we see that the long-time diffusivity monotonically increases with $d$ until the curves flatten out, as is expected by \eqref{deffforthepair} and \eqref{converge}.
Based on the explicit formula~\eqref{deffforthepair}, we next consider the model of a pair of spheroids whose relative angle and distance depends on time. Suppose the relative angle and distance pair are initially given by
$(\vartheta_0, d_0)$ and the elastic ligament is not fully relaxed. Because of the elastic energy in the ligament, we anticipate the relaxation of angle and distance can be characterized by two relaxation time scales (${\tau^*_1}, \tau^\ast_2$) in the sense that the time-dependent angle and distance are given by
\begin{eqnarray}\label{springd}
\begin{split}
\vartheta(t)&=\vartheta_f+(\vartheta_0- \vartheta_f)e^{-\frac{t}{\tau^*_1}},\\
d(t)&=\mathrm{d}_f+(\mathrm{d}_0-\mathrm{d}_f)e^{-\frac{t}{\tau^*_2}},\\
\end{split}
\end{eqnarray}
where $(\vartheta_f, d_f)$ denote the angle and distance between the pair in the final equilibrium state.
In typical experimental measurements, the diffusivity or anomalous diffusion is characterized by the Mean Square Displacement ($MSD$) defined by \citep{Metzler2014}
\begin{eqnarray}\label{MSD0}
MSD(t):=\frac{1}{t_m}\int_0^{t_m} |{\mathbf x}(t'+t)-{\mathbf x}(t')|^2\;\mathrm{d}t',
\end{eqnarray}
where ${\mathbf x}(t)$ is the position of the center of mass of the pair, and $t_m$ is the total measure time. For normal diffusions with a single time scale, e.g., diffusion of a homogeneous rigid spherical particle, the scaling of $MSD$ in \eqref{MSD0} with respect to time satisfies
\begin{eqnarray}\label{MSD00}
{MSD(t) \over 2t}\to {D} \qquad {{\rm\; as\;}}\;t\to +\infty ,
\end{eqnarray}
where $D$ is the macroscopic diffusivity. Now we consider the scaling of $MSD$ with respect to time for the pair of spheroids that relax from an initial non-equilibrium state. From the prior discussions, this process involves at least four time scales: cross-over time scale $t_{cross}$ (c.f., \eqref{timescale}), the two relaxation time scales $\tau^*_1, \tau^*_2$ for evolution of the relative angle and distance $(\vartheta, d)$ between the pair, and the translation diffusion time-scale of the pair. Therefore, the $MSD$ of the pair should exhibit much more complicated scaling behaviors with respect to $t$.
Nevertheless, if there is separation of time scales in the sense that
\begin{eqnarray*}
t\gg \tau^*_1, \tau^*_2\gg t_{cross},
\end{eqnarray*}
by \eqref{MSD00} we expect that $MSD$ should approximately behave as
\begin{eqnarray}\label{MSD}
MSD(t) \approx 2t {D}^{\rm eff}(\vartheta(t), d(t)),
\end{eqnarray}
where the diffusivity ${D}^{\rm eff}(\vartheta(t), d(t))$ is given by \eqref{deffforthepair} and $(\vartheta(t), d(t))$ is specified by \eqref{springd}.
{For comparison, let
\begin{eqnarray}\label{MSD0}
MSD_0(t) =2t {D}^{\rm eff}(\vartheta_0, d_0),
\end{eqnarray}
denote the Mean Square Displacement for the normal diffusion of the pair with $\vartheta(t)\equiv\vartheta_0$ and $d(t)\equiv d_0$.
Then, in Fig.~\ref{superdiffusion}, we plot the difference $\Delta MSD=MSD-MSD_0$ between the MSDs of the center of mass of the pair against $t$ for two cases.}
Specifically, the distance $d_f$ between the pair in the final equilibrium state is set to be $4a$ for case~1, whereas the distance $d_f$ between the pair in the final equilibrium state is set to be $2a$ for case~2. The initial distance $d_0$ between the pair in the initial non-equilibrium state is set to be $3a$ for both cases.
\begin{figure}[H]
\centering
\includegraphics[width=9.3cm]{anomalousdiffusion.png}
\caption{
{The difference $\Delta MSD=MSD-MSD_0$ between the MSDs} of the center of mass of the pair against time $t$ for two cases. The blue line refers to case 1 for which the final equilibrium distance $d_f=4a$, whereas the red line refers to case 2 with the final equilibrium distance $d_f=2a$. The distance between the pair in the initial non-equilibrium is $d_0=3a$. The major and minor semi-axis-lengths of the particle are $1 \mu m$ and {$0.1 \mu m$}, respectively.
The angle $\vartheta$ between the axes of symmetry of the pair of the particles is fixed at $\frac{\pi}{2}$. The relaxing time-scale is assumed to be $\tau^*_2=100s$. The ambient fluid is the mineral oil with viscosity $\eta=12.7\times 10^{-3}\;N\cdot s/ m^2$.
}\label{superdiffusion}
\end{figure}
From Fig.~\ref{superdiffusion} we observe two different types of scalings fo MSD versus $t$, which may be interpreted as ``superdiffusion'' and ``subdiffusion'', respectively. From this viewpoint, we expect physical models like ours could shed light on many anomalous diffusions observed in biological systems and complex media.
\section{Conclusion remarks} \label{sec:con}
In summary, we have conducted a systematic analysis on the long-time diffusion of microstructured particles of arbitrary shape and heterogeneity in a Newtonian fluid. The microscopic Brownian dynamics in position-orientation space is coarse-grained into a Fokker-Planck equation governing the evolution of the PDF on ${\mathbb{R}}^3\times {\rm SO}(3)$. By analyzing the Fokker-Planck equation, we identify a formula for the long-time diffusivity of the microstructured particle in an alignment field (c.f.,~\eqref{eq:Deff general} or \eqref{eq:Deff}). Applied to heterogeneous spheroids, we discover generalizations of the classical Stokes-Einstein relation which assert that the diffusivity of a rigid heterogeneous particle depends on the deviation of the center of mass from the geometric centroid (c.f.,~\eqref{Deffiso3}, \eqref{eq:Deffball} and \eqref{Deffiso2}). We have also addressed the effects of an external alignment field and achieved leading-order corrections on the diffusivities for large alignment fields (c.f.,~\eqref{ett electric1} and \eqref{ett electric2}). Based on these results, we consider diffusion of a pair of spheroids bonded by elastic ligaments. If the pair are initially in a non-equilibrium state, the Brownian motion superimposed with the relaxation process could be characterized as apparent ``superdiffusion'' or ``subdiffusion'', entailing a mechanistic perspective on anomalous diffusions. We believe that our method and results lay a solid foundation for understanding diffusions in complex media and may inspire new applications of controlled diffusions in biophysics and materials science.
| 2f1eba0c182bc01198b78fe9a14787494d3bfcfa | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\let\thefootnote\relax\footnotetext{
$^\ast$ Co-first authors
}
In the clinical diagnosis of vascular diseases, vessel segmentation is necessary to analyze the vessel structures and therapy planning. In particular, when diagnosing coronary artery disease, X-ray angiography is taken to enhance vessel visualization by injecting a contrast agent into the blood vessels \citep{cong2015quantitative}. However, it is challenging to extract vessels accurately due to low contrast, motion artifacts, many tiny branches, structural interference in the backgrounds, etc \citep{xia2019vessel, chen2014artifact}.
To segment vascular structures, various segmentation methods have been explored. Traditional optimization models \citep{law2009efficient, taghizadeh2014local} typically require complicated preprocessing steps and manual tuning. Furthermore, they are computationally expensive to process many images. On the other hand, learning-based methods \citep{nasr2016vessel, fan2018multichannel, chen2019unsupervised} generate segmentation maps in real-time once the models are trained. However, supervised methods require a huge amount of labeled data for training, which complicates their use in practical applications. Also, existing unsupervised methods designed on natural images are difficult to apply to medical vessel images due to low contrast subtle branches and confusing background structures. Although a recent self-supervised method \citep{ma2021self} is presented to learn vessel representations, this requires two different adversarial networks to segment vessels, which leads to increasing training complexity.
Recently, diffusion models such as denoising diffusion probabilistic model (DDPM) \citep{ho2020denoising} has become one of the main research topics in modeling data distribution and sampling diverse images. By learning the Markov transformation of the reverse diffusion process from Gaussian noise to data, DDPM is successfully applied to many low-level computer vision tasks such as super-resolution \citep{chung2022come}, inpainting \citep{lugmayr2022repaint}, and colorization \citep{song2020score}. For high-level vision tasks, while a recent study \citep{baranchuk2021label} shows that DDPM can capture semantic information and be used as image representations, methods applying DDPM in learning semantic segmentation without labeled data have so far not been developed.
In this paper, we introduce a novel concept of diffusion adversarial representation learning (DARL), which is a non-iterative version of diffusion-based generative model that can be successfully applied to self-supervised vessel segmentation. As illustrated in Figure~\ref{fig:method}, our model is composed of a diffusion module and a generation module, which learns semantic information of vessels via adversarial learning without ground-truth labels. In particular, given unpaired data of background images and angiography images that are taken before and after injection of the contrast agent, the diffusion module is designed to learn the background image distribution. Also, inspired by the spatially-adaptive denormalization (SPADE) layer \citep{park2019semantic} that is effective in image synthesis given semantic masks, we design a {\em switchable} version of SPADE as the generation module to estimate vessel segmentation masks as well as fake angiograms using the synthetic fractal masks \citep{ma2021self}.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{figs/method.png}
\caption{Our proposed diffusion adversarial representation model for self-supervised vessel segmentation. In path (A), given a real noisy angiography image ${\boldsymbol x}^{a}_{t_a}$, our model estimates vessel segmentation masks $\hat{{\boldsymbol s}}^{v}$. In path (B), given a real noisy background image ${\boldsymbol x}^{b}_{t_b}$ and a vessel-like fractal mask ${\boldsymbol s}^{f}$, our model generates a synthetic angiography image $\hat{{\boldsymbol x}}^{a}$. }
\label{fig:method}
\end{figure}
More specifically, as shown in Figure~\ref{fig:method}, there are two paths for feeding the inputs into our proposed model: (A) when the real angiography images are given, our model without the SPADE estimates vessel segmentation masks; (B) when the background images are given, our model with the SPADE generates synthetic angiograms that composite fake vessel masks such as fractal masks with the input backgrounds.
Also, by feeding the generated angiograms into the (A) path, we apply the cycle consistency between the segmentation results of the synthetic angiograms and the inputted fake vessel masks to capture semantic information of vessels. Since the diffusion module learns the background distribution, the foreground vessel structures of angiograms are considered as outlier noise so that latent features from the diffusion module can isolate the vessel representations. This enables the generation module to effectively segment the vessels.
We build our model on X-ray coronary angiography using XCAD dataset \citep{ma2021self} and apply to several different blood vessel datasets, including retinal images. Experimental results show that our method outperforms existing unsupervised and self-supervised learning methods in the absence of labeled data for training. The main contributions are summarized as:
\begin{enumerate}
\item We propose a diffusion adversarial representation model, a non-iterative version of diffusion model for image generation, and apply it for self-supervised vessel segmentation. Specifically, the latent features of our diffusion module provide vessel information and thus improve the segmentation performance.
\item Through the proposed generation module with switchable SPADE layers, our model not only generates synthetic angiography images but also segments vessel structures.
\item Experimental results verify that our model achieves superior segmentation performance by learning vessel representations. In particular, although the model is trained using X-ray coronary angiograms, it provides the state-of-the-art performance for un-/self-supervised retinal vessel segmentation as well, confirming the generalization capability of the model.
\end{enumerate}
\section{Backgrounds and related works}
\paragraph{Denoising diffusion probabilistic model}
Diffusion model \citep{sohl2015deep, ho2020denoising,song2019generative} is one of generative models that sample realistic data by learning the distribution of real images. In particular, the denoising diffusion probabilistic model (DDPM) \citep{ho2020denoising} with a score matching has been shown superior performance in image generation. Specifically, DDPM learns the Markov chain to convert the Gaussian noise distribution ${\boldsymbol x}_T \sim \mathcal{N}(\bm 0, \bm I)$ into the target distribution ${\boldsymbol x}_0$. In the forward diffusion process, the noise is gradually added the noise to the data by:
\begin{align}
q({\boldsymbol x}_t|{\boldsymbol x}_{t-1}) = \mathcal{N}({\boldsymbol x}_t;\sqrt{1-\beta_t}{\boldsymbol x}_{t-1}, \beta_t \bm I),
\end{align}
where $\beta_t\in[0,1]$ is a fixed variance. Accordingly, a noisy target ${\boldsymbol x}_t$ distribution from the data ${\boldsymbol x}_0$ is represented as:
\begin{align}
\label{eq:diff_forward}
q({\boldsymbol x}_t|{\boldsymbol x}_0) = \mathcal{N}({\boldsymbol x}_t; \sqrt{\alpha_t}{\boldsymbol x}_0, (1-\alpha_t)\bm I),
\end{align}
where $\alpha_t=\Pi_{s=1}^t(1-\beta_s)$.
Then, DDPM is trained to approximate reverse diffusion process:
\begin{align}
p_\theta ({\boldsymbol x}_{t-1}|{\boldsymbol x}_t) = \mathcal{N}({\boldsymbol x}_{t-1}; \bm\mu_\theta ({\boldsymbol x}_t, t), \sigma_t^2 \bm I),
\end{align}
where $\sigma_t$ is a fixed variance, and $\bm\mu_\theta$ is a parameterized mean with the noise predictor $\bm\epsilon_\theta$:
\begin{align}
\bm\mu_\theta ({\boldsymbol x}_t, t) = \frac{1}{\sqrt{1-\beta_t}}\left({\boldsymbol x}_t-\frac{\beta_t}{\sqrt{1-\alpha_t}}\bm\epsilon_\theta({\boldsymbol x}_t,t) \right).
\end{align}
Thus, in the generative process, the sample can be obtained from the Gaussian noise by the iterative denoising steps: ${\boldsymbol x}_{t-1}=\bm\mu_\theta ({\boldsymbol x}_t, t)+\sigma_t {\boldsymbol z}$, where ${\boldsymbol z}\sim\mathcal{N}(0, I)$.
Through this stochastic process, DDPM provides diverse realistic samples and has been exploited in many applications, including super-resolution \citep{chung2022come, saharia2021image}, inpainting \citep{lugmayr2022repaint}, and colorization \citep{song2020score, saharia2022palette}. However, the application study of semantic segmentation is limited. Although several works \citep{baranchuk2021label, amit2021segdiff} are recently presented to solve high-level vision problems, they require annotated data to train the models.
\paragraph{Self-supervised vessel segmentation}
For the vessel segmentation task, it is difficult to obtain fine-grained labels for supervised learning, since the vessel has complex structures with numerous tiny branches. While this label scarcity issue can be alleviated by semi- or unsupervised learning, fully unsupervised methods to segment the tiny vessels with reasonable performance are relatively scarce. In fact, recent unsupervised learning methods trained with natural images have great generalization capability on unseen datasets \citep{ahn2021spatial, chen2019unsupervised, melas2022deep}, thus they can be easily adapted to medical image segmentation tasks. However, due to the unique characteristics of angiography, e.g. confusing background factors and sophisticated vessel structures, most unsupervised methods fail to endure drastic performance degradation when they are applied to vessel segmentation tasks.
As a type of unsupervised learning, self-supervised learning also has been introduced to utilize self-generated supervisory labels from data themselves to efficiently learn target representations in various medical image segmentation tasks and has demonstrated its potential \citep{mahmood2019deep, ma2021self, oh2021unifying}. Specifically,
\cite{ma2021self} introduces an end-to-end adversarial learning framework for vessel segmentation with the CycleGAN \citep{zhu2017unpaired} structure, which learns realistic angiogram generation that adds fractal-guided pseudo labels to the background images. However, the simple arithmetic operation for synthetic vessel generation often fails to yield realistic pseudo-vessel images, thus training the adversarial networks using unrealistic synthetic images is difficult to produce optimal segmentation performance.
\section{Diffusion adversarial representation learning}
In this section, we describe our novel diffusion adversarial representation learning (DARL) model, tailored for self-supervised vessel segmentation method. As shown in Figure~\ref{fig:method}, DARL model is trained on unpaired real angiography images ${\boldsymbol x}_0^a$ and real background images ${\boldsymbol x}_0^b$.
Specifically, our model is comprised of a diffusion module $\bm\epsilon_{\bm\theta}$ to estimate latent features, a generation module $G$ to estimate both the vessel segmentation masks $\hat{{\boldsymbol s}}^v$ and the synthetic angiograms $\hat{{\boldsymbol x}}^a$, and two discriminators ($D_s$, $D_a$) to distinguish real and fake images of the vessel masks and the angiograms, respectively.
Here, to generate fake angiograms, we design the generation module to synthesize the fake angiograms by combining fake vessel masks ${\boldsymbol s}^{f}$ with the background images, where the fake vessels are random fractal masks presented in \cite{ma2021self}. In particular, we utilize the SPADE normalization layer \citep{park2019semantic} that facilitates realistic image generation while preserving the semantic information of a given input mask. In the following, we explain the proposed generation module in detail, and then describe the model training and inference.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{figs/method_detail.png}
\caption{Training flow of our model. The generation module with the switchable SPADE layers takes $\bm\epsilon_{\bm\theta}$ with the noisy images and generates desired outputs corresponding to the paths, i.e. (A) path for the angiography data to generate segmentation masks, and (B) path for the background images to generate the fake angiography images. The cycle path enables capturing the vessel information. }
\label{fig:method_detail}
\end{figure}
\paragraph{Generation module with switchable SPADE layer}
As illustrated in Figure~\ref{fig:method}, the proposed generation module consists of $N$ residual blocks (ResnetBlock) that have switchable SPADE (S-SPADE) layers. Let $\v\in \mathbb{R}^{M\times C\times H\times W}$ be the feature map in the ResnetBlock, where $M$, $C$, $H$, and $W$ are the size of batch, channel, height, and width, respectively. Then, the switchable SPADE layer normalizes feature maps differently depending on the existence of an input mask ${\boldsymbol s}$:
\begin{align}
\v=
\begin{cases}
\mathrm{SPADE}(\v, {\boldsymbol s}), & \text{if mask ${\boldsymbol s}$ is given,} \\
\mathrm{IN}(\v), & \text{otherwise,}
\end{cases}
\end{align}
where IN is the instance normalization \citep{ulyanov2017improved}. So, when our model is given the fake vessel mask ${\boldsymbol s}^{f}$, the SPADE is computed by:
\begin{align}
x_{m,c,h,w} = \gamma_{c,h,w}({\boldsymbol s}^{f}) \frac{x_{m,c,h,w}-\mu_c}{\sigma_c} + \beta_{c,h,w}({\boldsymbol s}^{f}),
\end{align}
where $x_{m,c,h,w}$ denotes the $(m,c,h,w)$-th element of the feature tensor $\v$, ($\mu_c, \sigma_c$) are the mean and standard deviation of the feature map in channel $c$, and ($\gamma_{c,h,w}, \beta_{c,h,w}$) are learned modulation parameters during training.
Thus, in the (A) path, given the noisy angiogram ${\boldsymbol x}_{t}^{a}$ and the latent feature $\bm\epsilon_{\bm\theta}({\boldsymbol x}_t^a,t)$, the generation module $G$ estimates the vessel segmentation masks $\hat{{\boldsymbol s}}^{v}$ without the SPADE:
\begin{align}
\hat{\boldsymbol s}^v = G(\bm\epsilon_{\bm\theta}({\boldsymbol x}_t^a,t);\bm 0).
\end{align}
On the other hand, in the (B) path that provides the fractal mask ${\boldsymbol s}^{f}$, the generation module taking the noisy background ${\boldsymbol x}_{t}^{b}$ and its latent feature $G(\bm\epsilon_{\bm\theta}({\boldsymbol x}_t^b,t))$ synthesizes the fake angiograms $\hat{{\boldsymbol x}}^{a}$:
\begin{align}
\hat{\boldsymbol x}^a = G(\bm\epsilon_{\bm\theta}({\boldsymbol x}_t^b,t);{\boldsymbol s}^f).
\end{align}
\subsection{Network Training}
\label{sec:loss}
In contrast to the DDPM that pretrains the diffusion model, our method trains the diffusion module, generation module, and discriminators simultaneously using an adversarial learning framework.
Figure~\ref{fig:method_detail} depicts the detailed training flow of our model. There are two distinct paths: (A) one feeds the real angiograms ${\boldsymbol x}_0^{a}$ into the model to provide vessel masks $\hat{{\boldsymbol s}}^{v}$, and (B) the other takes the real backgrounds ${\boldsymbol x}_0^{b}$ and the fractal masks ${\boldsymbol s}^f$ for the model to generate fake angiograms $\hat{{\boldsymbol x}}^{a}$.
Here, as shown in Figure~\ref{fig:method_detail}(B), since the input fractal masks can be regarded as vessel segmentation labels of the fake angiograms, we forward the fake angiograms generated through the (B) path to the (A) path, and apply cycle consistency between the estimated segmentation masks and the fractal masks to capture the vessel information.
\subsubsection{Loss function}
To train the model, we employ LSGAN \citep{mao2017least} framework, which leads to the alternating application of the following two optimization problems:
\begin{align}
\min_{\bm\theta,G} \mathcal{L}^G(\bm\epsilon_{\bm\theta}, G, D_s, D_a), \quad \min_{D_s, D_a} \mathcal{L}^D(\bm\epsilon_{\bm\theta}, G, D_s, D_a),
\end{align}
where $\mathcal{L}^G$, and $\mathcal{L}^D$ denotes the losses for the diffusion/generator and discriminator, respectively,
which are given by:
\begin{align}
\mathcal{L}^G(\bm\epsilon_{\bm\theta}, G, D_s, D_a)& = \mathcal{L}_{diff}(\bm\epsilon_{\bm\theta})
+ \alpha \mathcal{L}_{adv}^G(\bm\epsilon_{\bm\theta},G, D_s, D_a)
+ \beta \mathcal{L}_{cyc} (\bm\epsilon_{\bm\theta},G), \\
\mathcal{L}^{D}(\bm\epsilon_{\bm\theta}, G, D_s, D_a) &= \mathcal{L}_{adv}^{D_s}(\bm\epsilon_{\bm\theta},G, D_s)+ \mathcal{L}_{adv}^{D_a}(\bm\epsilon_{\bm\theta},G, D_a),
\end{align}
where $\alpha$ and $\beta$ are hyperparameters, $\mathcal{L}_{diff}$ is the diffusion loss, $\mathcal{L}_{adv}$ is adversarial loss, and $\mathcal{L}_{cyc}$ is cyclic reconstruction loss. The detailed description of each loss function is as follows.
\paragraph{Diffusion loss}
Recall that the diffusion module learns the distribution of images to estimate meaningful latent features of the inputs. We follow the standard loss for DDPM training \citep{ho2020denoising}:
\begin{align}
\label{eq:objective}
\mathcal{L}_{diff}(\bm\epsilon_{\bm\theta}):=\mathbb{E}_{t,{\boldsymbol x}_0,\boldsymbol{\epsilon}}\Big{[}\|\boldsymbol{\epsilon}-\boldsymbol{\epsilon}_\theta(\sqrt{{\alpha}_t}{\boldsymbol x}_0+\sqrt{1-{\alpha}_t}\boldsymbol{\epsilon},t)\|^2\Big{]}.
\end{align}
where $\bm\epsilon\sim\mathcal{N}(\bm 0, \bm I)$.
In particular, to let the diffusion module represent the vessels of angiograms effectively, we define the diffusion loss on the background images, i.e. ${\boldsymbol x}_0 = {\boldsymbol x}_0^{b}$ in the (B) path and set the sampling schedule in $t\in [0,T]$. Accordingly, the diffusion module is trained intensively to learn the background image distribution, allowing the module in the (A) path to regard the vessel structures of the angiograms as outlier noise and represent vessels in the latent features.
\paragraph{Adversarial loss}
To generate both vessel segmentation masks and synthetic angiograms without the ground-truth labels, the proposed model is trained by adversarial learning using the two discriminators $D_s$ and $D_a$. As shown in Figure~\ref{fig:method}, the discriminator $D_s$ attempts to distinguish the estimated segmentation masks $\hat{{\boldsymbol s}}^{v}$ from the real fractal mask ${\boldsymbol s}^{f}$ (in the (A) path), while the discriminator $D_a$ tries to discriminate between the generated angiograms $\hat{{\boldsymbol x}}^a$ and the real aniography images ${\boldsymbol x}_0^{a}$ (in the (B) path). As we employ LSGAN \citep{mao2017least}, the adversarial loss of generator $\mathcal{L}_{adv}^G$ can be formulated by:
\begin{align}
\mathcal{L}_{adv}^G(\bm\epsilon_{\bm\theta},G,\!D_s,\!D_a)
\!=\!\mathbb{E}_{{\boldsymbol x}^{a}}[(D_s(G(\bm\epsilon_{\bm\theta}({\boldsymbol x}^{a});\bm 0))-1)^2] + \mathbb{E}_{{\boldsymbol x}^{a},{\boldsymbol s}^f}[(D_a(G(\bm\epsilon_{\bm\theta}({\boldsymbol x}^{b}); {\boldsymbol s}^f))-1)^2].
\end{align}
On the other hand, the discriminators are trained to compete against the generator with the adversarial loss functions, $\mathcal{L}_{adv}^{D_s}$ and $\mathcal{L}_{adv}^{D_a}$, which are defined by:
\begin{align}
\mathcal{L}_{adv}^{D_s}(\bm\epsilon_{\bm\theta},G, D_s) &= \frac{1}{2}\mathbb{E}_{{\boldsymbol s}^{f}}[(D_s({\boldsymbol s}^{f})-1)^2] + \frac{1}{2}\mathbb{E}_{{\boldsymbol x}^{a}}[(D_s(G(\bm\epsilon_{\bm\theta}({\boldsymbol x}^{a});\bm 0)))^2], \\
\mathcal{L}_{adv}^{D_a}(\bm\epsilon_{\bm\theta},G, D_a)&= \frac{1}{2}\mathbb{E}_{{\boldsymbol x}_0^{a}}[(D_a({\boldsymbol x}_0^{a})-1)^2] + \frac{1}{2}\mathbb{E}_{{\boldsymbol x}^{a},{\boldsymbol s}^f}[(D_a(G(\bm\epsilon_{\bm\theta}({\boldsymbol x}^{b}); {\boldsymbol s}^f)))^2].
\end{align}
This adversarial loss enables the single generator $G$ to fool the discriminator $D_s$ and $D_a$, by generating realistic segmentation masks
$\hat{{\boldsymbol s}}^{v}=G(\bm\epsilon_{\bm\theta}({\boldsymbol x}^{a});\bm 0)$ and angiograms $\hat{{\boldsymbol x}}^{a}=G(\bm\epsilon_{\bm\theta}({\boldsymbol x}^{b}); {\boldsymbol s}^f)$. In contrast, the discriminators attempt to distinguish these generated images being fake and the real images of ${\boldsymbol s}^{f}$ and ${\boldsymbol x}_0^{a}$ being real.
\paragraph{Cyclic reconstruction loss}
For the generator $G$ to capture the semantic information of the vessels, we also constrain our model with the cyclic reconstruction loss on the fractal masks. Specifically, as the vessel-like fractal masks ${\boldsymbol s}^{f}$ can be labels for the synthetic angiograms $\hat{\boldsymbol x}^{a}$ generated in the (B) path, we feed the $\hat {\boldsymbol x}^{a}$ into our model and reconstruct the fractal masks by the (A) path. Therefore, the cyclic reconstruction loss is computed between the reconstructed segmentation masks and the real fractal masks, which can be written by:
\begin{align}
\mathcal{L}_{cyc} (\bm\epsilon_{\bm\theta},G) = \mathbb{E}_{{\boldsymbol x}^{b}, {\boldsymbol s}^{f}}[||G(\bm\epsilon_{\bm\theta}(G(\bm\epsilon_{\bm\theta}({\boldsymbol x}^{b}); {\boldsymbol s}^f));\bm 0)-{\boldsymbol s}^{f} ||_1].
\end{align}
Here, we solve the segmentation problem as a vessel mask image generation, which is why we use L1 loss in the cyclic loss.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{figs/inference.png}
\caption{Vessel segmentation according to the noise level $t_a$. Our model estimates the segmentation masks $\hat{{\boldsymbol s}}^{v}$ using the latent features $\bm\epsilon_{\bm \theta}$ for the noisy angiograms ${\boldsymbol x}_{t_a}^a$. ${\boldsymbol s}^{v}$ is the ground-truth label.}
\label{fig:inference}
\vspace{-0.2cm}
\end{figure}
\subsubsection{Image perturbation for the model input}
Given real images of ${\boldsymbol x}_0^{a}$ and ${\boldsymbol x}_0^{b}$, our diffusion module takes noisy angiograms ${\boldsymbol x}_{t_a}^{a}$ in the (A) path and noisy background images ${\boldsymbol x}_{t_b}^{b}$ in the (B) path {as the input}, in which each noisy image is sampled based on the forward diffusion process (\ref{eq:diff_forward}):
\begin{align}
\label{eq:diff_samA}
{\boldsymbol x}_t=\sqrt{{\alpha}_t}{\boldsymbol x}_0+\sqrt{1-{\alpha}_t}\boldsymbol{\epsilon},
\end{align}
where $\bm \epsilon\sim\mathcal{N}(\bm 0, \bm I)$, and both of $t_a$ and $t_b$ are uniformly sampled time step in $[0, T]$. Here, for the diffusion module not only to learn the background image distribution in the (B) path but also to provide useful information for the generation module to segment the vessel structures under even certain noisy angiogram images in the (A) path, we sample $t_a$ in the range of $[0, T_a]$ where $T_a<T$. Empirically, we found that this makes our model learn vessel representations robust to the noise.
\vspace{-0.1cm}
\subsection{Inference of vessel segmentation}
The inference phase of DARL is different from the conventional diffusion model in that our model do not require iterative reverse process, similar to the recent diffusion-based unsupervised learning method called DiffuseMorph \citep{kim2021diffusemorph}. Specifically, once the proposed DARL is trained, in the inference, we can obtain the vessel segmentation masks of angiograms from the (A) path by one step. For the noisy angiograms ${\boldsymbol x}_{t_a}^a$ given by the forward diffusion process (\ref{eq:diff_samA}), our model provides the vessel segmentation masks using the latent features $\bm\epsilon_{\bm\theta}({\boldsymbol x}_{t_a}^a, t_a)$ estimated from the diffusion module. As shown in Figure~\ref{fig:inference}, our model can generate the segmentation masks for any noise level $t_a$ within a certain range (i.e. $[0, T_a]$). Nevertheless, since the angiography image ${\boldsymbol x}_0^{a}$ can be considered as one of the clean target images, the closer $t_a$ is to zero, the better the vessel segmentation performance. Therefore, we test our model by setting $t_a=0$.
\vspace{-0.1cm}
\section{Experiments}
\vspace{-0.1cm}
In this section, we thoroughly evaluate the vessel segmentation performance of our method. We firstly compare the proposed DARL to existing unsupervised and self-supervised baseline models on various angiography datasets, including X-ray coronary angiography and retinal images. Also, we study the noise robustness of our model. Then, we analyze the success of our model in vessel representation and conduct an ablation study.
\vspace{-0.2cm}
\paragraph{Datasets}
To realize the self-supervised learning framework, we train our model with the publicly available X-ray coronary angiography disease (XCAD) dataset obtained during stent placement surgery \citep{ma2021self}. A total of 1,621 angiography frames and their corresponding first frames, taken before injecting the contrast agents, are used as the real angiography and background images, respectively. Also, we generate 1,621 synthetic fractal masks by \cite{ma2021self}. Additional 126 angiography images, along with the ground-truth vessel masks annotated by experienced radiologists, are divided into validation and test sets by 10\% and 90\%, respectively. We subsample all data into 256$\times$256. Also, in testing, we utilize two external X-ray coronary angiography (XCA) datasets acquired from different machines. 134 XCA dataset is composed of 134 angiography images with the vessel masks labeled by an expert cardiologist \citep{app9245507}. 30 XCA datset is composed of 30 sequences of angiography images \citep{HAO2020172}. We utilize one angiography image from each sequence, along with its corresponding ground-truth vessel mask labeled by experts. All the test images are resized to 512$\times$512. Furthermore, we evaluate cross-organ generalization capability on retinal imaging datasets. We use DRIVE \citep{staal2004ridge} and STARE \citep{hoover2003locating} datasets, each of which is composed of 20 retinal images and the corresponding expert-labeled vessel masks. Since retinal imaging is taken under high-resolution, we resize the image into 768$\times$768 and split into 9 patches with 256$\times$256.
\begin{figure}[b!]
\vspace{-0.2cm}
\centering
\includegraphics[width=\linewidth]{figs/results_extended.jpg}
\caption{Visual comparison results on the vessel segmentation of various angiography images.}
\label{fig:main}
\end{figure}
\paragraph{Implementation details}
Our model is implemented by employing the network architectures proposed in DDPM \citep{ho2020denoising} and SPADE \citep{park2019semantic} for the diffusion module and the generation module, respectively. Also, for the discriminators, we use the network of PatchGAN \citep{isola2017image}. To train the model, we set the number of time steps as $T=2000$ with the linearly scheduled noise levels from $10^{-6}$ to $10^{-2}$. Within this range, we sample the noisy angiograms by setting $T_a$ to 200. Also, we set the hyperparameters of loss function as $\alpha=0.2$ and $\beta=5$. Our model is optimized by using the Adam algorithm \citep{kingma2014adam} with a learning rate of $5\times10^{-6}$ on a single GPU card of Nvidia Quadro RTX 6000. We train the model for 150 epochs, and the model in the epoch with the best performance on the validation set is used for test data. All the implementations are done using the library of PyTorch \citep{paszke2019pytorch} in Python. The details of network structures and hyperparameter setting can be found in Appendix.
\paragraph{Baseline methods and metrics}
We compare our model to several baseline methods of un-/self-supervised learning, which do not require ground-truth vessel labels. For unsupervised learning methods, we utilize Spatial-Guided Clustering (SGC) \citep{ahn2021spatial}, Redrawing \citep{chen2019unsupervised}, and Deep Spectral (DS) \citep{melas2022deep}. For self-supervised learning methods, we employ Self-supervised Transformer with Energy-based Graph Optimization (STEGO) \citep{hamilton2022unsupervised}, Deep Adversarial (DA) \citep{mahmood2019deep}, and Self-Supervised Vessel Segmentation (SSVS) \citep{ma2021self}. All these methods are implemented under identical training conditions to our model, unless the method needs no training procedure. For baseline methods that require heuristic thresholds, optimal performance is achieved by selecting data-specific thresholds within the range from 0.2 to 0.8 in increments of 0.1. To quantitatively evaluate the segmentation performance, we compute Intersection over Union (IoU), Dice similarity coefficient, and Precision.
\vspace{-0.1cm}
\subsection{Experimental results}
Figure~\ref{fig:main} shows the vessel segmentation masks from the baseline methods and our proposed method on three different coronary angiography datasets and two retinal imaging datasets. Quantitative evaluation results of the methods are presented in Table \ref{tab:main}. The analysis of the results is as follows.
\vspace{-0.2cm}
\paragraph{Comparison of ours to baselines}
When we compare the proposed method to the baselines, our model segments vessel structures including tiny branches more accurately. Also, as shown in Table~\ref{tab:main}, our model consistently achieves the SOTA performance by large margin compared to existing unsupervised and self-supervised methods. In specific, our network shows significantly improved precision scores, which demonstrates advantages of our DARL that effectively differentiates foreground vessel structure and eliminates false positive signals from the noisy backgrounds.
\begin{table}[t!]
\caption{Quantitative evaluation results on the vessel segmentation of various angiography images.}
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{clccccccc}
\toprule
\multirow{2}{*}{Data} & \multirow{2}{*}{Metric} & \multicolumn{3}{c}{\textbf{Unsupervised} } & \multicolumn{4}{c}{\bf{Self-supervised}} \\
\cmidrule(r){3-5}\cmidrule(l){6-9}
& & SGC & Redrawing & DS & STEGO & DA & SSVS & \textbf{Ours} \\
\midrule
\multicolumn{9}{l}{\bf{Coronary angiography}} \\
\midrule
\multirow{3}{*}{XCAD}
& IoU & $0.060_{\pm 0.034}$ & $0.059_{\pm 0.032}$ & $0.366_{\pm 0.105}$ & $0.146_{\pm 0.070}$ & $0.375_{\pm 0.066}$ & $0.410_{\pm 0.087}$ & $\bf0.471_{\pm 0.076}$ \\
& Dice & $0.111_{\pm 0.060}$ & $0.109_{\pm 0.056}$ & $0.526_{\pm 0.131}$ & $0.249_{\pm 0.103}$ & $0.542_{\pm 0.073}$ & $0.575_{\pm 0.091}$ & $\bf0.636_{\pm 0.072}$ \\
& Precision & $0.062_{\pm 0.034}$ & $0.139_{\pm 0.081}$ & $0.469_{\pm 0.127}$ & $0.152_{\pm 0.077}$ & $0.557_{\pm 0.115}$ & $0.590_{\pm 0.119}$ & $\bf0.701_{\pm 0.115}$ \\
\midrule
\multirow{3}{*}{134 XCA}
& IoU & $0.045_{\pm 0.035}$ & $0.056_{\pm 0.018}$ & $0.256_{\pm 0.110}$ & $0.134_{\pm 0.081}$ & $0.190_{\pm 0.155}$ & $0.318_{\pm 0.128}$ & $\bf0.426_{\pm 0.059}$ \\
& Dice & $0.085_{\pm 0.063}$ & $0.105_{\pm 0.033}$ & $0.394_{\pm 0.159}$ & $0.228_{\pm 0.109}$ & $0.291_{\pm 0.217}$ & $0.468_{\pm 0.156}$ & $\bf0.595_{\pm 0.058}$ \\
& Precision & $0.047_{\pm 0.036}$ & $0.058_{\pm 0.019}$ & $0.280_{\pm 0.123}$ & $0.136_{\pm 0.088}$ & $0.506_{\pm 0.201}$ & $0.592_{\pm 0.125}$ & $\bf0.781_{\pm 0.118}$ \\
\midrule
\multirow{3}{*}{30 XCA}
& IoU & $0.083_{\pm 0.039}$ & $0.048_{\pm 0.022}$ & $0.339_{\pm 0.086}$ & $0.191_{\pm 0.072}$ & $0.298_{\pm 0.109}$ & $0.324_{\pm 0.146}$ & $\bf0.427_{\pm 0.184}$ \\
& Dice & $0.150_{\pm 0.064}$ & $0.091_{\pm 0.040}$ & $0.499_{\pm 0.113}$ & $0.314_{\pm 0.100}$ & $0.447_{\pm 0.148}$ & $0.468_{\pm 0.193}$ & $\bf0.572_{\pm 0.205}$ \\
& Precision & $0.090_{\pm 0.041}$ & $0.144_{\pm 0.074}$ & $0.525_{\pm 0.130}$ & $0.200_{\pm 0.081}$ & $0.612_{\pm 0.174}$ & $0.613_{\pm 0.212}$ & $\bf0.729_{\pm 0.152}$ \\
\midrule
\multicolumn{9}{l}{\bf{Retinal imaging}} \\
\midrule
\multirow{3}{*}{DRIVE} & IoU & $0.063_{\pm 0.055}$ & $0.057_{\pm 0.033}$ & $0.217_{\pm 0.143}$ & $0.152_{\pm 0.073}$ & $0.245_{\pm 0.090}$ & $0.314_{\pm 0.101}$ & $\bf0.372_{\pm 0.148}$ \\
& Dice & $0.115_{\pm 0.093}$ & $0.105_{\pm 0.059}$ & $0.333_{\pm 0.201}$ & $0.257_{\pm 0.106}$ & $0.386_{\pm 0.117}$ & $0.469_{\pm 0.119}$ & $\bf0.525_{\pm 0.161}$ \\
& Precision & $0.069_{\pm 0.061}$ & $0.199_{\pm 0.155}$ & $0.243_{\pm 0.175}$ & $0.169_{\pm 0.100}$ & $0.503_{\pm 0.218}$ & $0.549_{\pm 0.216}$ & $\bf0.617_{\pm 0.271}$ \\
\midrule
\multirow{3}{*}{STARE} & IoU & $0.055_{\pm 0.045}$ & $0.074_{\pm 0.048}$ & $0.180_{\pm 0.141}$ & $0.125_{\pm 0.076}$ & $0.237_{\pm 0.122}$ & $0.311_{\pm 0.148}$ & $\bf0.368_{\pm 0.191}$ \\
& Dice & $0.101_{\pm 0.077}$ & $0.134_{\pm 0.080}$ & $0.281_{\pm 0.201}$ & $0.216_{\pm 0.109}$ & $0.367_{\pm 0.167}$ & $0.454_{\pm 0.185}$ & $\bf0.508_{\pm 0.216}$ \\
& Precision & $0.058_{\pm 0.047}$ & $0.227_{\pm 0.157}$ & $0.205_{\pm 0.172}$ & $0.135_{\pm 0.092}$ & $0.427_{\pm 0.233}$ & $0.490_{\pm 0.230}$ & $\bf0.537_{\pm 0.280}$ \\
\bottomrule
\end{tabular}
}
\label{tab:main}
\vspace{-0.3cm}
\end{table}
\vspace{-0.2cm}
\paragraph{Generalization capability}
To verify that our trained DARL can be generally applied to various vessel image taken from different machines or different anatomic region-of-interests (ROI), we also study the generalization capability of our method. First, when we test our model on external 134 XCA and 30 XCA datasets directly, which have different resolutions and noise distributions to those of the XCAD dataset, {as shown in Figure~\ref{fig:main} and Table~\ref{tab:main}}, our model achieves higher performance than the others. Moreover, when we evaluate the model on DRIVE and STARE retinal datasets that have unseen data distribution of different modalities and ROIs from the XCAD, our DARL shows the most promising cross-organ generalization performance. This may come from the proposed adversarial framework that reuses the generated angiography images for the segmentation process through the cycle path, diversifying the input data distribution. Also, the diffusion module learning the stochastic diffusion process enables our model to be used in general for vessel segmentation.
\vspace{-0.2cm}
\paragraph{Robustness to noises}
\label{sec:noise}
As X-ray images are often acquired under low-dose radiation exposure to reduce potential risks, we further evaluate the performance of our model on simulated noisy angiograms. Using the XCAD dataset, we add Gaussian noise to the angiogram with different levels of $\sigma=$10, 25, and 50. We show the segmentation results according to the noise levels in Figure~\ref{fig:main}. Also, we report the quantitative evaluation results in Table~\ref{tab:noise}. It is noteworthy that our DARL is the only method to segment vessel structures with reasonable performance under noise corruption. Since the proposed segmentation method is trained through the diffusion module that perturbs the input images,
the model is highly robust to segment vessel structure even from the noisy data.
\begin{table}[h!]
\caption{Results of noise robustness test according to the Gaussian noise with $\sigma$.}
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{M{0.8cm}lccccccc}
\toprule
\multirow{2}{*}{$\sigma$} & \multirow{2}{*}{Metric} & \multicolumn{3}{c}{\textbf{Unsupervised} } & \multicolumn{4}{c}{\bf{Self-supervised}} \\
\cmidrule(r){3-5}\cmidrule(l){6-9}
& & SGC & Redrawing & DS & STEGO & DA & SSVS & \textbf{Ours} \\
\midrule
\multirow{3}{*}{10} & IoU & $0.066_{\pm 0.033}$ & $0.052_{\pm 0.031}$ & $0.331_{\pm 0.104}$ & $0.144_{\pm 0.073}$ & $0.353_{\pm 0.065}$ & $0.258_{\pm 0.079}$ & $\bf0.451_{\pm 0.080}$ \\
& Dice & $0.122_{\pm 0.059}$ & $0.096_{\pm 0.053}$ & $0.487_{\pm 0.133}$ & $0.245_{\pm 0.107}$ & $0.519_{\pm 0.073}$ & $0.404_{\pm 0.099}$ & $\bf0.617_{\pm 0.076}$ \\
& Precision & $0.069_{\pm 0.035}$ & $0.126_{\pm 0.077}$ & $0.480_{\pm 0.135}$ & $0.157_{\pm 0.091}$ & $0.481_{\pm 0.104}$ & $0.477_{\pm 0.117}$ & $\bf0.710_{\pm 0.115}$ \\
\midrule
\multirow{3}{*}{25} & IoU & $0.069_{\pm 0.035}$ & $0.036_{\pm 0.021}$ & $0.232_{\pm 0.094}$ & $0.118_{\pm 0.064}$ & $0.247_{\pm 0.072}$ & $0.059_{\pm 0.033}$ & $\bf0.389_{\pm 0.088}$ \\
& Dice & $0.128_{\pm 0.061}$ & $0.069_{\pm 0.039}$ & $0.366_{\pm 0.132}$ & $0.206_{\pm 0.095}$ & $0.391_{\pm 0.092}$ & $0.109_{\pm 0.058}$ & $\bf0.554_{\pm 0.092}$ \\
& Precision & $0.072_{\pm 0.036}$ & $0.095_{\pm 0.058}$ & $0.446_{\pm 0.159}$ & $0.144_{\pm 0.115}$ & $0.371_{\pm 0.106}$ & $0.149_{\pm 0.082}$ & $\bf0.727_{\pm 0.119}$ \\
\midrule
\multirow{3}{*}{50} & IoU & $0.070_{\pm 0.025}$ & $0.020_{\pm 0.012}$ & $0.077_{\pm 0.065}$ & $0.060_{\pm 0.050}$ & $0.102_{\pm 0.056}$ & $0.021_{\pm 0.013}$ & $\bf0.269_{\pm 0.081}$ \\
& Dice & $0.130_{\pm 0.045}$ & $0.040_{\pm 0.022}$ & $0.136_{\pm 0.109}$ & $0.108_{\pm 0.088}$ & $0.180_{\pm 0.091}$ & $0.041_{\pm 0.025}$ & $\bf0.417_{\pm 0.101}$ \\
& Precision & $0.072_{\pm 0.026}$ & $0.061_{\pm 0.038}$ & $0.221_{\pm 0.168}$ & $0.076_{\pm 0.067}$ & $0.169_{\pm 0.094}$ & $0.060_{\pm 0.038}$ & $\bf0.716_{\pm 0.147}$ \\
\bottomrule
\end{tabular}
}
\label{tab:noise}
\vspace{-0.3cm}
\end{table}
\vspace{-0.2cm}
\paragraph{Latent representation}
To study the origin of the performance improvement, in Figure~\ref{fig:discuss_diff}, we show the latent features $\bm\epsilon_{\bm\theta}({\boldsymbol x}_t, t)$ given ${\boldsymbol x}_0$ for (A) the angiography ${\boldsymbol x}_0={\boldsymbol x}_0^a$ and (B) the
\begin{wrapfigure}[11]{r}{0.43\textwidth}
\centering
\includegraphics[width=0.43\textwidth]{figs/discuss_diff.png}
\caption{Estimated latent features $\bm\epsilon_{\bm\theta}$ in the (A) and (B) paths of our model.}
\label{fig:discuss_diff}
\end{wrapfigure}
backgrounds ${\boldsymbol x}_0={\boldsymbol x}_0^b$ with $t=100$, respectively. In contrast to the (B) path, the latent representation in the (A) path emphasizes the vessel structures.
This implies that although there are no ground-truth labels, our model learns the background image representation so that the vessel structure can be captured as outlier noise, leading to improved segmentation performance.
\vspace{-0.2cm}
\paragraph{Ablation study}
Table~\ref{tab:ablation} shows the quantitative evaluation results of several ablation studies, and visual results are in Appendix.
(a) Our model without the diffusion module and $\mathcal{L}_{diff}$ shows lower performance by about 5\% for all metrics compared to our model, which suggests that the diffusion module guides the generation module to extract vessel representation accurately.
(b) The generation module without the proposed S-SPADE layers is degraded by more than 5\% over (a) for all metrics, verifying that our SPADE-based unified generator effectively captures vessel semantic information through the synergy of learning both image segmentation and generation.
(c) Through the implementation of our model without the proposed cyclic loss $\mathcal{L}_{cyc}$, we verify that $\mathcal{L}_{cyc}$ allows our model to segment proper vessel regions.
(d) When training our model by converting the L1 loss for $\mathcal{L}_{cyc}$ to the cross-entropy (CE) loss, the performance is much worse than ours in all metrics, which implies that our approach using L1 loss for the cycle path is proper to obtain the vessel masks.
\begin{table}[!ht]
\caption{Results of ablation study on the proposed model and loss function.}
\centering
\resizebox{0.96\linewidth}{!}{
\begin{tabular}{ccccccccc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{2}{c}{Module} & \multicolumn{3}{c}{Loss function} & \multicolumn{3}{c}{Metric} \\
\cmidrule(r){2-3}\cmidrule(r){4-6}\cmidrule(r){7-9}
& Diffusion & Generation & $\mathcal{L}_{diff}$ & $\mathcal{L}_{adv}$ & $\mathcal{L}_{cyc}$ & IoU & Dice & Precision \\
\midrule
Ours & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark
& $\bf0.471_{\pm 0.076}$ & $\bf0.636_{\pm 0.072}$ & $\bf0.701_{\pm 0.115}$ \\
\midrule
(a) & & \checkmark & & \checkmark & \checkmark
& $0.420_{\pm 0.081}$ & $0.587_{\pm 0.084}$ & $0.606_{\pm 0.101}$ \\
(b) & & {w/o S-SPADE} & & \checkmark & \checkmark &
$0.362_{\pm 0.065}$ & $0.528_{\pm 0.072}$ & $0.483_{\pm 0.102}$ \\
\midrule
(c) & \checkmark & \checkmark & \checkmark & \checkmark &
& $0.322_{\pm 0.055}$ & $0.485_{\pm 0.064}$ & $0.580_{\pm 0.112}$ \\
(d) & \checkmark & \checkmark & \checkmark & \checkmark & {L1$\rightarrow$CE}
& $0.346_{\pm 0.084}$ & $0.508_{\pm 0.094}$ & $0.672_{\pm 0.147}$ \\
\bottomrule
\end{tabular}
}
\label{tab:ablation}
\vspace{-0.2cm}
\end{table}
\section{Conclusion}
We present a non-iterative diffusion model called DARL for self-supervised vessel segmentation. Our model composed of the diffusion and generation modules learns vessel representation without labels via adversarial learning, in the guidance of latent features estimated from the diffusion module. Also, through the proposed switchable SPADE layer, we generate synthetic angiograms as well as vessel segmentation masks, leading to learning semantic information about vessels more effectively. Although the diffusion module training is combined with other loss functions, the inference is not iterative but only done in one step, which makes it faster and unique compared to the existing diffusion models. Using various medical vessel datasets, we verify that our model is much superior to existing un-/self-supervised learning methods. Moreover, thanks to the diffusion module, our model is robust to image diversity and noise, suggesting that our model can be an important platform for designing a general vessel segmentation model.
| 4601d160d7bc8e567605102ca63101b306713c7e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{introduction and main results}
A central topic in differential geometry is how the curvature conditions restrict the topology. In the case of Ricci curvature (abbreviated by ``Ric") in K\"{a}hler geometry this principle goes back to Bocher, who proved that (\cite{Bo1}, \cite{Bo2}, \cite{YB}) the condition of $\text{Ric}>0$ or $\text{Ric}<0$ on compact K\"{a}hler manifolds or the existence of K\"{a}hler-Einstein metrics impose heavy restrictions on
their holomorphic tensor fields. These conditions can now be reformulated in terms of the first Chern class, thanks to the celebrated Calabi-Yau theorem and the Aubin-Yau theorem (\cite{Yau77}). In order to precisely state Bochner's results, let (throughout this article) $TM$ and $T^{\ast}M$ be respectively the (\emph{holomorphic}) tangent and cotangent bundle of a compact complex manifold $M$, and
\be\label{holomorphic tensor field}\Gamma^p_q(M):=H^0\big(M,(TM)^{\otimes p}\otimes(T^{\ast}M)^{\otimes q}\big),\qquad (p, q\in\mathbb{Z}_{\geq 0})\ee
the space of \emph{$(p,q)$-type holomorphic tensor fields} on $M$. Bochner's results can be stated as follows (\cite{Ko80-1}, \cite{Ko80-2}, \cite[p. 57]{KH83}).
\begin{theorem}[Bochner, Calabi-Yau, Aubin-Yau]\label{Bochner}
Let $M$ be an $n$-dimensional compact K\"{a}hler manifold.
\begin{enumerate}
\item
If $c_1(M)$ is quasi-positive, then there exists a positive constant $C=C(M)$ such that $\Gamma^p_q(M)=0$ when $q>C\cdot p$, and consequently $\Gamma^0_q(M)=0$ when $q\geq 1$. In particular the Hodge numbers $h^{q,0}(M)=0$ when $1\leq q\leq n$.
\item
If $c_1(M)$ is quasi-negative, then there exists a positive constant $C=C(M)$ such that $\Gamma^p_q(M)=0$ when $p>C\cdot q$, and consequently $\Gamma^p_0(M)=0$ when $p\geq 1$.
\item
If $c_1(M)<0$, then $\Gamma^p_q(M)=0$ when $p>q$.
\end{enumerate}
\end{theorem}
\begin{remark}
The condition quasi-positivity (resp. quasi-negativity) of $c_1(M)$ means that there exists a closed $(1,1)$-form representing $c_1(M)$ which is nonnegative (resp. non-positive) everywhere and positive (resp. negative) somewhere.
\end{remark}
It is well-known that the condition $\text{Ric}>0$, which is equivalent to be Fano due to the Calabi-Yau theorem, implies simple-connectedness (\cite{Ko61}). This was further strengthened to be rational connected by Campana (\cite{Ca}) and Koll\'{a}r-Miyaoka-Mori (\cite{KMM}) independently. Recall that a complex manifold is called \emph{rationally connected} if any two points on it can be joined by a chain of rational curves. Rational connetedness is an important tool/notion in algebraic geometry.
In view of the subtle relationship between Ric and the holomorphic sectional curvature $H$ (\cite[p. 181]{Zh}), one may expect that the above-mentioned conclusions still hold when Ric is replaced by $H$. The simple-connectedness under the condition $H>0$ is well-known (\cite{Ts}).
A recent important result due to Wu-Yau, Tosatti-Yang and Diverio-Trapani (\cite{WY16-1}, \cite{TY17}, \cite{DT19}, \cite{WY16-2}) implies that the quasi-negativity of $H$ implies $c_1(M)<0$, and hence the conclusions of parts (2)-(3) in Theorem \ref{Bochner} remain true when $H$ is quasi-negative. Yau conjectured that (\cite[Problem 47]{Yau82}) a compact K\"{a}hler manifold with $H>0$ is projective and rational connected. Assuming the projectivity this was proved by Heier-Wong (\cite{HW}). Shortly afterwards the projectivity was proved by Yang in \cite{Yang18} by showing that the Hodge number $h^{2,0}=0$ (\cite[p. 143]{MK71}), who also provided an alternative proof of the rational connectedness therein. Recently Yang introduced in \cite{Yang20} the notion of \emph{uniform RC-positivity} on Hermitian holomorphic vector bundles, which is satisfied by the tangent bundle of a compact K\"{a}hler manifold with $H>0$, and deduced that a compact K\"{a}hler manifold whose tangent bundle is uniform RC-positive with respect to a (possibly different) Hermitian metric is projective and rational connected (\cite[Thm 1.3]{Yang20}).
In contrast to the negative case, in general the condition $H>0$ is \emph{not} able to yield $c_1(M)>0$, as exhibited by Hitchin (\cite{Hi75}) that the Hirzebruch surfaces \be\label{Hitchin example}F:=\mathbb{P}\big(\mathcal{O}_{\mathbb{P}^1}(-k)
\oplus\mathcal{O}_{\mathbb{P}^1}\big)\qquad(k\geq 2)\ee
admit K\"{a}hler metrics with $H>0$ but $c_1(F)$ are \emph{not} positive (cf. \cite[p. 949]{Yang16}). More related examples can be found in \cite{AHZ} and \cite{NZ0}. Even so, we can still ask whether the conclusions of part (1) in Theorem \ref{Bochner} hold when $H>0$. In fact, Yang showed that (\cite[Thm 1.7]{Yang18}) the condition $H>0$ leads to the Hodge numbers $h^{q,0}=0$ for $1\leq q\leq n$ and thus provides positive evidence towards this validity. Recently the author showed that the conclusions of part (1) in Theorem \ref{Bochner} are true when $H>0$ (\cite{Li21}). Indeed what we proved in \cite[Thm 1.5]{Li21} is in a more general setting, i.e., for those Hermitian metrics whose Chern curvature tensors behave like the usual K\"{a}hler curvature tensors.
The \emph{$k$-Ricci curvatures} $\text{Ric}_k$ for $1\leq k\leq n$ was introduced by Ni (\cite{Ni21-1}) in his study of the $k$-hyperbolicity of a compact K\"{a}hler manifold. $\text{Ric}_k$ coincides with $H$ when $k=1$ and with Ric when $k=n$ and so they interpolate between $H$ and $\text{Ric}$. Ni showed that the condition $\text{Ric}_k>0$ for some $k\in\{1,\ldots,n\}$ yields rational connectedness and $h^{q,0}=0$ for $1\leq q\leq n$, and hence the simple-connectedness and projectivity (\cite[Thms 1.1, 4.2]{Ni21-2}). Chu-Lee-Tam showed that (\cite[Thm 1.1]{CLT}) the condition $\text{Ric}_k<0$ for some $k\in\{1,\ldots,n\}$ also implies $c_1(M)<0$. (see \cite{LNZ} by Li-Ni-Zhu for an alternative proof). Hence the conclusions of parts (2) and (3) in Theorem \ref{Bochner} hold true under the condition $\text{Ric}_k<0$ for some $k\in\{1,\ldots,n\}$.
In view of the discussions above, we may wonder whether the conclusions of part (1) in Theorem \ref{Bochner} are true when $\text{Ric}_k>0$ for some $k\in\{1,\ldots,n\}$. One main motivation in this article is, as an application of the main results, to affirmatively prove it. Another motivation is to extend Yang's differential-geometric criterion for rational connectedness (\cite[Thm 1.3]{Yang20}) from the case of uniform RC-positivity to that of uniform RC $k$-positivity for arbitrary $k$.
\section{Main results}\label{main results}
Our main results and applications are stated in this section. The following concept is inspired by that of uniform RC-positivity introduced by Yang (\cite{Yang20}) as well as BC $k$-positivity by Ni (\cite{Ni21-2}).
\begin{definition}\label{uniform rc k positivity}
A Hermitian holomorphic vector bundle $(E,h)$ over an $n$-dimensional Hermitian manifold $(M,\omega)$ is called \emph{uniformly RC $k$-positive} (resp. \emph{uniformly RC $k$-negative}) $(1\leq k\leq n)$ at $x\in M$ if there exists a $k$-dimensional subspace $\Sigma\subset T_xM$ such that for every nonzero vector $u\in E_x$ (the fiber of $E$ at $x$),
\be\label{RC-positivity}
R^{(E,h)}(\Sigma;u,\overline{u}):=
\sum_{i=1}^kR^{(E,h)}(E_i,\overline{E_i},u,\overline{u})>0~(\text{resp. $<0$}),\ee
where $R^{(E,h)}$ is the Chern curvature tensor of $(E,h)$ and $\{E_1,\ldots,E_k\}$ a unitary basis of $\Sigma$ with respect to $\omega$. If this holds for any $x\in M$, then it is called \emph{uniformly RC $k$-positive} (resp. \emph{uniformly RC $k$-negative}).
\end{definition}
\begin{remark}
\begin{enumerate}
\item
The uniform RC-positivity in \cite{Yang20} is exactly the uniform RC $1$-positivity in our notion, whose definition is indeed irrelevant to the metric $\omega$. Nevertheless, the definition of uniform RC $k$-positivity for $k\geq2$ relies on the metric $\omega$. A different but closely related notion called \emph{BC $k$-positivity} was introduced by Ni in \cite[p. 282]{Ni21-2}. We refer to Definition \ref{uniform RC kl definition} and Remark \ref{uniform RC remark} for more details on various positivity concepts.
\item
It is obvious that $(E,h)$ is uniformly RC $k$-positive if and only if the dual bundle $(E^{\ast},h^{\ast})$ is uniformly RC $k$-negative, where $h^{\ast}$ is the metric induced from $h$.
\end{enumerate}
\end{remark}
The following vanishing theorem for uniformly RC $k$-positive vector bundles is a key technical tool.
\begin{theorem}\label{RC vanishing theorem}
Let $(E,h)$ be a uniformly RC $k$-positive holomorphic vector bundle over an $n$-dimensional compact Hermitian manifold $(M,\omega)$ for some $k\in\{1,\ldots,n\}$, and $F$ any holomorphic vector bundle over $M$. Then there exist positive constants $C_1=C_1(h,\omega)$ and $C_2=C_2(F,h,\omega)$ such that
\be\label{RC vanishing formula}H^0\big(M,E^{\otimes p}\otimes (E^{\ast})^{\otimes q}\otimes F^{\otimes m}\big)=0\ee
for all $p,q,m\in\mathbb{Z}_{\geq 0}$ with $q> C_1\cdot p+C_2\cdot m$.
\end{theorem}
The main motivation to introduce this concept is that the tangent bundle of a compact K\"{a}her manifold $(M,\omega)$, or more generally a compact K\"{a}hler-like Hermitian manifold $(M,\omega)$, with the condition $\text{Ric}_k(\omega)>0$ (resp. $\text{Ric}_k(\omega)<0$) is uniformly RC $k$-positive (resp. negative). To this end, let us recall the following notion, which was first proposed and investigated in detail by B. Yang and Zheng in \cite{YZ}.
\begin{definition}
Let $(M,\omega)$ be a Hermitian manifold and $R$ the \emph{Chern curvature tensor} of $\omega$. The Hermitian metric $\omega$ is called \emph{Chern-K\"{a}hler-like} (abbreviated by CKL) if \be\label{CKL}
R(X,\overline{Y},Z,\overline{W})=R(Z,\overline{Y},X,\overline{W})\ee for any $(1,0)$-type tangent vectors $X$, $Y$, $Z$ and $W$.
\end{definition}
\begin{remark}\label{remark}
\begin{enumerate}
\item
When $\omega$ is K\"{a}hler, $R$ is the usual curvature tensor, which clearly satisfies (\ref{CKL}).
By taking the complex conjugation, $(\ref{CKL})$ implies that
$$R(X,\overline{Y},Z,\overline{W})=
\overline{R(Y,\overline{X},W,\overline{Z})}=
\overline{R(W,\overline{X},Y,\overline{Z})}=
R(X,\overline{W},Z,\overline{Y})
.$$
Therefore the condition (\ref{CKL}) ensures that $R$ obeys \emph{all} the symmetries satisfied by a K\"{a}hler metric and thus the term CKL is justified.
\item
There are many \emph{non-K\"{a}hler} Hermitian metrics which are CKL (see \cite[p. 1197]{YZ}). On the other hand, Yang-Zheng showed that a CKL Hermitian metric $\omega$ must be balanced, i.e., $d\omega^{n-1}=0$ (\cite[Thm 3]{YZ}). Hence CKL Hermitian metrics interpolate between K\"{a}hler and balanced metrics.
\end{enumerate}
\end{remark}
With this notion understood, our second main result is the following
\begin{theorem}\label{rick lead to uniform}
Let $(M,\omega)$ be an $n$-dimensional compact Chern-K\"{a}hler-like Hermitian manifold and $\text{Ric}_{k}(\omega)>0$ (resp. $\text{Ric}_{k}(\omega)<0$) for some $k\in\{1,\ldots,n\}$. Then $(TM,\omega)$ is uniformly RC $k$-positive (resp. uniformly RC $k$-negative) over $(M,\omega)$.
\end{theorem}
Apply Theorem \ref{RC vanishing theorem} to Theorem \ref{rick lead to uniform} by taking the holomorphic vector bundle $F$ trivial, it yields the following desired vanishing theorem, which extends \cite[Thm 1.5]{Li21} from the case $\text{Ric}_1$ to arbitrary $\text{Ric}_k$.
\begin{theorem}\label{CKL result}
Let $(M,\omega)$ be an $n$-dimensional compact Chern-K\"{a}hler-like Hermitian manifold.
\begin{enumerate}
\item
If $\text{Ric}_k(\omega)>0$ for some $k\in\{1,\ldots,n\}$, then there exists a positive constant $C=C(\omega)$ such that $\Gamma^p_q(M)=0$ when $q>C\cdot p$, and consequently $\Gamma^0_q(M)=0$ when $q\geq 1$. In particular the Hodge numbers $h^{q,0}(M)=0$ when $1\leq q\leq n$.
\item
If $\text{Ric}_k(\omega)<0$ for some $k\in\{1,\ldots,n\}$, then then there exists a positive constant $C=C(\omega)$ such that $\Gamma^p_q(M)=0$ when $p>C\cdot q$, and consequently $\Gamma^p_0(M)=0$ when $p\geq 1$.
\end{enumerate}
\end{theorem}
\begin{remark}
When $\omega$ is K\"{a}hler, part $(2)$ in Theorem \ref{CKL result} follows from Theorem \ref{Bochner} and the results due to Wu-Yau et al. and Chu-Lee-Tam, as mentioned above. Nevertheless, if the CKL metric $\omega$ is \emph{non-K\"{a}hler}, the results in part $(2)$ are still new.
\end{remark}
Our second major application is a differential-geometric criterion for the rational connectedness of compact K\"{a}hler manifolds. The following nice criterion for rational connectedness was established in \cite[Thm 1.1]{CDP}, on which both \cite[Thm 1.3]{Yang20} and \cite[Thm 1.1]{Ni21-2} are based.
\begin{theorem}[Campana-Demailly-Peternell]\label{CDP}
Let $M$ be an projective manifold. Then $M$ is rational connected if and only if for any ample line bundle $L$ on $M$, there exists a positive constant $C=C(L)$ such that
$$H^0\big(M,(T^{\ast}M)^{\otimes q}\otimes L^{\otimes m}\big)=0$$
for $q,m\in\mathbb{Z}_{\geq 0}$ with $q> C\cdot m$.
\end{theorem}
Combining Theorems \ref{RC vanishing theorem}, \ref{rick lead to uniform}, \ref{CKL result} and \ref{CDP} we have the following criterion for projectivity and the rational connectedness.
\begin{theorem}\label{rational connectedness}
Let $M$ be an $n$-dimensional compact K\"{a}hler manifold. Then it is projective and rational connected (and hence simply-connected) provided one of the following two conditions holds true.
\begin{enumerate}
\item
There exists a (possibly different) Hermitian metric $\omega$ such that
$(TM,\omega)$ is uniformly RC $k$-positive over $(M,\omega)$ for some $k\in\{1,\ldots,n\}$.
\item
$M$ has a Chern-K\"{a}hler-like Hermitian metric $\omega$ with $\text{Ric}_k(\omega)>0$ for some $k\in\{1,\ldots,n\}$.
\end{enumerate}
\end{theorem}
\begin{remark}
\begin{enumerate}
\item
Part $(1)$ extends \cite[Thm 1.3]{Yang20} from the case of uniform RC $1$-positivity to that of uniform RC $k$-positivity for arbitrary $k$. Part $(2)$ technically improves on \cite[Thm 1.1]{Ni21-2} from K\"{a}hler metrics to Chern-K\"{a}hler-like Hermitian metrics.
\item
The Hodge number $h^{2,0}=0$ and hence projectivity can be derived via a somewhat weaker \emph{BC $2$-positivity} by Ni (\cite[p. 280-281]{Ni21-2} (see Definition \ref{uniform RC kl definition} and Remark \ref{uniform RC remark} for more details), and the proof in \cite[Thm 4.6]{Ni21-2} can be adopted to give an alternative one of \cite[Thm 1.3]{Yang20} (cf. \cite[p. 285]{Ni21-2}).
\end{enumerate}
\end{remark}
The rest of this article is organized as follows. Some necessary background materials are collected in Section \ref{Preliminaries}. Sections \ref{proof of rc vanishing} and \ref{proof of theorem rick yield rc} are devoted to the proofs of Theorems \ref{RC vanishing theorem} and \ref{rick lead to uniform} respectively. In Section \ref{further remarks} some related questions and remarks shall be discussed, in which various positivity notions are proposed and their relations are briefly discussed for possible further study in the future.
\section{Background materials}\label{Preliminaries}
We collect in this section some basic facts on Hermitian holomorphic vector bundles and Hermitian and K\"{a}hler manifolds in the form we shall use to prove our main results. A thorough treatment can be found in \cite{Ko87} and \cite{Zh}.
\subsection{Hermitian holomorphic vector bundles}
Let $(E,h)\rightarrow M$ be a Hermitian holomorphic vector bundle of rank $r$ over an $n$-dimensional complex manifold $M$ with canonical Chern connection $\nabla$. The Chern curvature tensor $$R=R^{(E,h)}:=\nabla^2\in\Gamma(\Lambda^{1,1}M\otimes E^{\ast}\otimes E)$$
Here and throughout this article $\Gamma(\cdot)$ is used to denote the space of \emph{smooth} sections for holomorphic vector bundles and the notation $\Gamma^p_q(M)$ in (\ref{holomorphic tensor field}) is reserved throughout this article to denote the space of $(p,q)$-type \emph{holomorphic} tensor fields on $M$.
Take a local frame field $\{s_1,\ldots,s_r\}$ of $E$, whose dual coframe field is denoted by $\{s_1^{\ast},\ldots,s_r^{\ast}\}$, and a local holomorphic coordinates $\{z^1,\ldots,z^n\}$ on $M$. With the Einstein summation convention adopted here and in what follows, the Chern curvature tensor $R$ and the Hermitian metric $h$ can be written as
\begin{eqnarray}\label{curvature tensor}
\left\{ \begin{array}{ll}
\displaystyle R=:\Omega^{\beta}_{\alpha}s^{\ast}_{\alpha}\otimes s_{\beta}=:R^{\beta}_{i\bar{j}\alpha}dz^i\wedge d\bar{z}^j\otimes s^{\ast}_{\alpha}\otimes s_{\beta},\\
~\\
\displaystyle h=(h_{\alpha\bar{\beta}}):=\big(h(s_{\alpha},s_{\beta})\big),\\
~\\
\displaystyle R_{ij\alpha\bar{\beta}}:=R_{ij\alpha}^{\gamma}h_{\gamma\bar{\beta}}.
\end{array} \right.\nonumber
\end{eqnarray}
The Hermitian metric $h(\cdot,\cdot)$ and the induced metrics on various vector bundles arising naturally from $E$ are sometimes denoted by $<\cdot,\cdot>$.
For $u=u^{\alpha}s_{\alpha}\in\Gamma(E)$, $v=v^{\alpha}s_{\alpha}\in\Gamma(E)$, $X=X^i\frac{\partial}{\partial z^i}$, and $Y= Y^i\frac{\partial}{\partial z^i}$, we have
$$R(u)=(\Omega^{\beta}_{\alpha}s^{\ast}_{\alpha}\otimes s_{\beta})(u^{\gamma}s_{\gamma})=
\Omega^{\beta}_{\alpha}u^{\alpha}s_{\beta}
\in\Gamma(\Lambda^{1,1}M\otimes E),$$
$$R_{X\overline{Y}}(u)=\Omega^{\beta}_{\alpha}
(X,\overline{Y})u^{\alpha}s_{\beta}=
R^{\beta}_{i\bar{j}\alpha}X^i\overline{Y^j}u^{\alpha}s_{\beta}
\in\Gamma(E),$$
and therefore,
\be\label{Hermitian formula}
\begin{split}
R(X,\overline{Y},u,\overline{v}):=&
<R_{X\overline{Y}}(u),v>\\
=&
<R^{\beta}_{i\bar{j}\alpha}X^i\overline{Y^j}u^{\alpha}s_{\beta},
v^{\gamma}s_{\gamma}>\\
=&R^{\beta}_{i\bar{j}\alpha}X^i\overline{Y^j}u^{\alpha}\overline{v^{\gamma}}h_{\beta\bar{\gamma}}\\
=&R_{i\bar{j}\alpha\bar{\beta}}X^i\overline{Y^j}u^{\alpha}
\overline{v^{\beta}}.\end{split}
\ee
Here and in what follows we always use capital letters to denote vectors in $TM$ and lowercase letters to denote vectors in $E$, so as to distinguish them.
The formula (\ref{Hermitian formula}) yields an easy but useful fact, which we record in the next lemma for our later reference.
\begin{lemma}\label{diagonal and real lemma}
The linear map $R_{X\overline{X}}(\cdot)$ is a Hermitian transformation: \be\label{Hermitan transformation}<R_{X\overline{X}}(u),v>=<u,R_{X\overline{X}}(v)>,\qquad\text{for any $u,v\in\Gamma(E)$}.\ee
Hence $R_{X\overline{X}}(\cdot)$ is diagonalizable and its eigenvalues are all real at any point in $M$.
\end{lemma}
\subsection{The Ricci and scalar $k$-curvatures}
Let $(M,\omega)$ be an $n$-dimensional Hermitian manifold, $R=R^{(TM,\omega)}$ the Chern curvature tensor.
Denote by $T_xM$ the \emph{$(1,0)$-type} tangent space at $x\in M$ and $\{E_1,\ldots,E_n\}$ a unitary basis of $T_xM$. For $X\in T_xM$, let $H(X)$, $\text{Ric}(X,\overline{X})$ and $S(x)$ be respectively the holomorphic sectional curvature, Chern-Ricci curvature and Chern scalar curvature:
\begin{eqnarray}
\left\{ \begin{array}{ll}
\displaystyle H(X):=\frac{R(X,\overline{X},X,\overline{X})}{|X|^4},\\
~\\
\displaystyle\text{Ric}(X,\overline{X}):=\sum\limits_{i=1}^nR(X,\overline{X},E_i,\overline{E_i}),\\
~\\
\displaystyle S(x):=\sum\limits_{i,j=1}^nR(E_i,\overline{E_i},E_j,\overline{E_j}).
\end{array} \right.\nonumber
\end{eqnarray}
When $\omega$ is K\"{a}hler, $H(X)$, $\text{Ric}(X,\overline{X})$ and $S(x)$ are the usual notions of holomorphic sectional curvature, Ricci curvature and scalar curvature in K\"{a}hler geometry.
For $k\in\{1,\ldots,n\}$, pick a $k$-dimensional subspace $\Sigma\subset T_xM$ and choose a unitary basis $\{E_1,\ldots,E_k\}$ of $\Sigma$. Define
\be\label{def riccik}\text{Ric}_k(x,\Sigma)(X,\overline{X}):=\sum_{i=1}^kR(X,
\overline{X},E_i,\overline{E_i}),\qquad\text{for any $X\in\Sigma$,}\ee
and call it the \emph{$k$-Ricci curvature} of $\omega$, which was introduced by Ni in \cite{Ni21-1}. The $k$-Ricci curvatures $\text{Ric}_k$ interpolate $H$ and $\text{Ric}$ in the sense that
$$\text{Ric}_1(x,\mathbb{C}X)(X,\overline{X})=|X|^2H(X)\quad \text{and} \quad\text{Ric}_n=\text{Ric}.$$
We say that $\text{Ric}_k(x)>0$ if $\text{Ric}_k(x,\Sigma)(X,\overline{X})>0$ for any $k$-dimensional subspace $\Sigma$ in $T_xM$ and any nonzero $X\in\Sigma$, and $\text{Ric}_k=\text{Ric}_k(\omega)>0$ if $\text{Ric}_k(x)>0$ for any $x\in M$.
Closely related to $\text{Ric}_k$ are
the \emph{$k$-scalar curvatures} $S_k=S_k(\omega)$ ($1\leq k\leq n$) introduced in \cite{NZ} by Ni-Zheng for \emph{K\"{a}hler} metrics, which is the average of $H$ on $k$-dimensional subspaces of $(1,0)$-type tangent spaces. These $\{S_k\}$ interpolate between $H$ ($k=1$) and the usual scalar curvature ($k=n$). They extended Yang's results by showing that the condition $S_k>0$ for a \emph{K\"{a}hler} metric implies that the Hodge numbers $h^{q,0}=0$ for $k\leq q\leq n$ (\cite[Thm 1.3]{NZ}). In particular, the condition $S_2>0$ is enough to gurantee the projectivity.
Here we shall define $S_k=S_k(\omega)$ for $k\in\{1,\ldots,n\}$ in the Hermitian situation, which turns out to be the same as the original one in \cite{NZ} when $\omega$ is K\"{a}hler (see Lemma \ref{relations riccik and sk}). As above, for $x\in M$, a $k$-dimensional subspace $\Sigma\subset T_xM$ and a unitary basis $\{E_1,\ldots,E_k\}$ of $\Sigma$, define
\be\label{def sk}S_k(x,\Sigma):=\sum_{i,j=1}^k
R(E_i,\overline{E_i},E_j,\overline{E_j}).\ee
It is obvious that in the Hermitian case these $\{S_1,\ldots,S_n\}$ interpolate between $H$ ($k=1$) and the Chern scalar curvature $S(x)$ ($k=n$). Similarly we say that $S_k(x)>0$ if $S_k(x,\Sigma)>0$ for any $k$-dimensional subspace $\Sigma$ in $T_xM$, and $S_k=S_k(\omega)>0$ if $S_k(x)>0$ for any $x\in M$.
\subsection{Integral formulas}
The trick of the proof in the next lemma is usually attributed to Berger, who first applied it to show that for \emph{K\"{a}hler} metrics the sign of $H_x(\cdot)$ determines that of $S(x)$.
\begin{lemma}\label{integral formula lemma}
Let the notation be as above and
$$f(\cdot,\cdot):~T_xM\times T_xM\rightarrow\mathbb{C},\qquad
g(\cdot,\cdot,\cdot,\cdot):~T_xM\times T_xM\times T_xM\times T_xM\rightarrow\mathbb{C}$$
be two smooth maps such that the first variable of $f$ (resp. the first and third variables of $g$) is (resp. are) linear and the second variable of $f$ (resp. the second and fourth variables of $g$) is (resp. are) conjugate-linear.
Then
\be\label{integral formula1}\int_{Y\in\Sigma,~|Y|=1}f(Y,Y)d\theta(Y)=
\frac{\text{V}(\mathbb{S}^{2k-1})}{k}\sum_{i=1}^kf(E_i,E_i),\ee
and
\be\label{integral formula2}\begin{split}&\int_{Y\in\Sigma,~|Y|=1}g(Y,Y,Y,Y)d\theta(Y)\\
=&\frac{\text{V}(\mathbb{S}^{2k-1})}{k(k+1)}\sum_{i,j=1}^k
\Big[g(E_i,E_i,E_j,E_j)+g(E_i,E_j,E_j,E_i)\Big],\end{split}\ee
where $d\theta(Y)$ is the spherical measure on $\{Y\in\Sigma,~|Y|=1\}\cong\mathbb{S}^{2k-1}$ and $\text{V}(\mathbb{S}^{2k-1})$ the volume with respect to it.
\end{lemma}
\begin{proof}
Let $Y=\sum_{i=1}^kY^iE_i\in\Sigma$ and recall the following two classical identities on spherical measure:
\be\label{integral over sphere1}\int_{Y\in\Sigma,~|Y|=1}Y^i\overline{Y^j}d\theta(Y)=
\frac{\delta_{ij}}{k}\cdot\text{V}(\mathbb{S}^{2k-1}),\ee
and
\be\label{integral over sphere2}\int_{Y\in\Sigma,~|Y|=1}Y^i\overline{Y^j}Y^r\overline{Y^s}d\theta(Y)
=\frac{\delta_{ij}\delta_{rs}+\delta_{is}\delta_{rj}}{k(k+1)}
\cdot\text{V}(\mathbb{S}^{2k-1}),
\ee
where $\delta_{ij}$ is the Kronecker delta.
Then
\be\begin{split}\int_{Y\in\Sigma,~|Y|=1}f(Y,Y)d\theta(Y)&=
\int_{Y\in\Sigma,~|Y|=1}f(Y^iE_i,Y^jE_j)d\theta(Y)\\
&=f(E_i,E_j)
\int_{Y\in\Sigma,~|Y|=1}Y^i\overline{Y^j}d\theta(Y)\\
&=\frac{\text{V}(\mathbb{S}^{2k-1})}{k}
\sum_{i=1}^kf(E_i,E_i),\qquad\big(\text{by (\ref{integral over sphere1})}\big)\end{split}\nonumber\ee
and
\be\begin{split}&\int_{Y\in\Sigma,~|Y|=1}
g(Y,Y,Y,Y)d\theta(Y)\\
=&
\int_{Y\in\Sigma,~|Y|=1}
g(Y^iE_i,Y^jE_j,Y^rE_r,Y^sE_s)d\theta(Y)\\
=&g(E_i,E_j,E_r,E_s)
\int_{Y\in\Sigma,~|Y|=1}Y^i\overline{Y^j}Y^r\overline{Y^s}d\theta(Y)\\
=&\frac{\text{V}(\mathbb{S}^{2k-1})}{k(k+1)}
\sum_{i,j=1}^k\Big[g(E_i,E_i,E_j,E_j)
+g(E_i,E_j,E_j,E_i)\Big].\qquad\big(\text{by (\ref{integral over sphere2})}\big)
\end{split}\nonumber\ee
\end{proof}
Applying Lemma \ref{integral formula lemma} to (\ref{RC-positivity}), (\ref{def riccik}) and (\ref{def sk}) produces the following alternative definitions as integrals over the unit sphere in $\Sigma$.
\begin{corollary}\label{relations riccik and sk}
Let the notation be as above. We have
\be\label{def integral uniform}R^{(E,h)}(\Sigma;u,\overline{u})=
\frac{k}{\text{V}(\mathbb{S}^{2k-1})}
\int_{Y\in\Sigma,~|Y|=1}R^{(E,h)}(Y,\overline{Y},u,\overline{u})d\theta(Y),\ee
\be\label{def integral riccik}\text{Ric}_k(x,\Sigma)(X,\overline{X})=
\frac{k}{\text{V}(\mathbb{S}^{2k-1})}\int_{Y\in\Sigma,~|Y|=1}R(X,\overline{X},Y,\overline{Y})d\theta(Y),\ee
and
\be\label{def integral sk} S_k(x,\Sigma)=\frac{k(k+1)}{2\text{V}(\mathbb{S}^{2k-1})}\int_{Y\in\Sigma,~|Y|=1}H(Y)d\theta(Y),\qquad\text{when $\omega$ is CKL}.\ee
\end{corollary}
\begin{remark}\label{remark2}
In \cite{NZ} the identity (\ref{def integral sk}) was taken as the definition of $S_k$ for K\"{a}hler metrics. Either formula of $\text{Ric}_k$ or $S_k$ has its own advantage. For instance, from (\ref{def riccik}) and (\ref{def sk}) the condition $\text{Ric}_k(x)>0$ (resp. $\text{Ric}_k>0$) implies $S_k(x)>0$ (resp. $S_k>0$). On the other hand, (\ref{def integral sk}) tells us that $S_k>0$ (resp. $S_k<0$) implies $S_{k+1}>0$ (resp. $S_{k+1}<0$). In contrast to it, in general the sign of $\text{Ric}_k$ is independent from that of $\text{Ric}_l$ when $k\neq l$, as illustrated by Hitchin's examples (\ref{Hitchin example}).
\end{remark}
\section{Proof of Theorem \ref{RC vanishing theorem}}\label{proof of rc vanishing}
\subsection{Two lemmas}
We prepare two crucial lemmas in order to establish Theorem \ref{RC vanishing theorem}. The following result was obtained in \cite[Lemma 2.1]{Li21}, whose proof is to apply a $\partial\bar{\partial}$-Bochner formula and the maximum principle to part of directions.
\begin{lemma}\label{Bochner fromula0 lemma}
Let $u\in\Gamma(E)$ be a holomorphic section of the Hermitian holomorphic vector bundle $(E,h)$ over a compact complex manifold $M$, and the maximum of $|u|:=<u,u>^{\frac12}$ is attained at $x\in M$. Then
\be\label{Bochner formula0}<R_{X\overline{X}}(u),u>\Big|_x\geq0,\qquad\text{for all $X\in T_xM.$}\ee
\end{lemma}
\begin{remark}
The use of a $\partial\bar{\partial}$-Bochner formula, together with the maximum principle to part of directions, was recently revived by some works (\cite{An}, \cite{AC}, \cite{Liu}, \cite{Ni13}, \cite{NZ}, \cite{Yang18}).
\end{remark}
The next lemma is parallel to \cite[Prop. 2.9]{Yang20}.
\begin{lemma}\label{equivalent relation}
Let $(E,h)$ be a uniformly RC $k$-positive Hermitian holomorphic vector bundle over a compact Hermitian manifold $(M,\omega)$. Then there exists a constant $C=C(h,\omega)>0$ such that for any $x\in M$, there exists a $k$-dimensional subspace $\Sigma_x\subset T_xM$ such that \be\label{inequality}R^{(E,h)}(\Sigma_x;u,\overline{u})\geq C,\qquad\text{for any unit vector $u\in E_x$}.\ee
\end{lemma}
\begin{proof}
For $x\in M$, let
$$\mathbb{G}_k(T_xM):=\Big\{\Sigma~\big|~\text{$\Sigma$ are $k$-dimensional subpaces in $T_xM$}\Big\}$$
be the associated complex Grassmannian of $T_xM$ and $$\mathbb{S}E_x:=\big\{u\in E_x,~|u|=1\big\}$$ the unit sphere of $E_x$. Let
\be\label{def C(h,w)}C=C(h,\omega):=\min_{x\in M}\max_{\Sigma\in\mathbb{G}_k(T_xM)}
\min_{u\in\mathbb{S}E_x}R^{(E,h)}(\Sigma;u,\overline{u}),\ee
which is well-defined due to the compactness of $M$, $\mathbb{G}_k(T_xM)$ and $\mathbb{S}E_x$.
First note that such defined $C$ satisfies $(\ref{inequality})$. In fact, by definition (\ref{def C(h,w)}) implies that for any $x\in M$, $$\max_{\Sigma\in\mathbb{G}_k(T_xM)}
\min_{u\in\mathbb{S}E_x}R^{(E,h)}(\Sigma;u,\overline{u})\geq C.$$ This in turn yields that there exists a $\Sigma_x\in\mathbb{G}_k(T_xM)$ such that $$\min_{u\in\mathbb{S}E_x}R^{(E,h)}(\Sigma_x;u,\overline{u})\geq C,$$ which leads to (\ref{inequality}).
It suffices to show $C>0$. Suppose on the contrary that $C\leq0$. Then (\ref{def C(h,w)}) implies that there exists an $x_0\in M$ such that
$$\max_{\Sigma\in\mathbb{G}_k(T_{x_0}M)}
\min_{u\in\mathbb{S}E_{x_0}}R^{(E,h)}(\Sigma;u,\overline{u})=C\leq0.$$
This leads to $$\min_{u\in\mathbb{S}E_{x_0}}R^{(E,h)}(\Sigma;u,\overline{u})\leq0$$ for \emph{any} $\Sigma\in \mathbb{G}_k(T_{x_0}M)$, which in turn yields that for each $\Sigma\in\mathbb{G}_k(T_{x_0}M)$ there exists a $u(\Sigma)\in\mathbb{S}E_{x_0}$ such that $R^{(E,h)}\big(\Sigma;u(\Sigma),\overline{u(\Sigma)}\big)\leq0$.
In summary, under the assumption of $C\leq0$, we derive a conclusion that there exists an $x_0\in M$ such that, for every $\Sigma\in\mathbb{G}_k(T_{x_0}M)$, there exists $u(\Sigma)\in\mathbb{S}E_{x_0}$ such that $R^{(E,h)}\big(\Sigma;u(\Sigma),\overline{u(\Sigma)}\big)\leq0$, which exactly contradicts to the condition of uniform RC $k$-positivity of $(E,h)\rightarrow(M,\omega)$!
\end{proof}
\subsection{Proof of Theorem \ref{RC vanishing theorem}}
Let $T\in H^0\big(M,E^{\otimes p}\otimes (E^{\ast})^{\otimes q}\otimes F^{\otimes m}\big)$, $x\in M$ and a \emph{unit} vector $X\in T_xM$. Arbitrarily choose a Hermitian metric $h_F$ on $F$ and simply write $R$ to denote the Chern curvature tensors of various Hermitian holomorphic vector bundles involved in the proof.
Due to Lemma \ref{diagonal and real lemma} denote by the real eigenvalues of the Hermitian transformations $R_{X\overline{X}}(\cdot)$ on $\big(E_x,h(x)\big)$ and $\big(F_x,h_F(x)\big)$ by $\lambda_i=\lambda_i(x,X)$ ($1\leq i\leq r_1$) and $\mu_j=\mu_j(x,X)$ ($1\leq j\leq r_2$) respectively, where $r_1$ and $r_2$ are the ranks of $E$ and $F$. Let $\{e_1,\ldots,e_{r_1}\}$ and $\{s_1,\ldots,s_{r_2}\}$ be unitary bases of $E_x$ and $F_x$ such that
\begin{eqnarray}\label{0.1}
\left\{ \begin{array}{ll}
\displaystyle R_{X\overline{X}}(e_i)=\lambda_ie_i, &(1\leq i\leq r_1)\\
~\\
R_{X\overline{X}}(s_j)=\mu_js_j. &(1\leq j\leq r_2)
\end{array} \right.
\end{eqnarray}
Let $\{\theta^1,\ldots,\theta^{r_1}\}$ be the unitary basis of $E^{\ast}_x$ dual to $\{e_i\}$. Then the induced action of $R_{X\overline{X}}(\cdot)$ on $\{\theta^i\}$ is given by
\be\label{0.2}R_{X\overline{X}}(\theta^i)=-\lambda_i\theta^i,\qquad 1\leq i\leq r_1.\ee
Write
$$T\Big|_x=T^{\alpha_1\cdots\alpha_p\gamma_1\cdots\gamma_m}_{\beta_1\cdots\beta_q}e_{\alpha_1}
\otimes\cdots\otimes e_{\alpha_p}\otimes\theta^{\beta_1}\otimes\cdots\otimes\theta^{\beta_q}\otimes s_{\gamma_1}\otimes\cdots\otimes s_{\gamma_m}.$$
Then (\ref{0.1}) and (\ref{0.2}) imply
\be
\begin{split}R_{X\overline{X}}(T)\Big|_x
=\sum_{\begin{subarray}{1}
\alpha_1,\ldots,\alpha_p\\ \beta_1,\ldots,\beta_q\\ \gamma_1,\ldots,\gamma_m\end{subarray}}
\bigg[&\Big(\sum_{i=1}^p\lambda_{\alpha_i}-\sum_{j=1}^q
\lambda_{\beta_j}+\sum_{l=1}^m\mu_{\gamma_l}\Big)
T^{\alpha_1\cdots\alpha_p\gamma_1\cdots\gamma_m}_{\beta_1\cdots\beta_q}\\
&e_{\alpha_1}
\otimes\cdots\otimes e_{\alpha_p}\otimes
\theta^{\beta_1}\otimes\cdots\otimes\theta^{\beta_q}\otimes s_{\gamma_1}\otimes\cdots\otimes s_{\gamma_m}\bigg],\end{split}\nonumber\ee
which yields that
\be\label{T-formula}
\begin{split}&<R_{X\overline{X}}(T)\Big|_x,T\Big|_x>\\
=&
\sum_{\begin{subarray}{1}
\alpha_1,\ldots,\alpha_p\\ \beta_1,\ldots,\beta_q\\ \gamma_1,\ldots,\gamma_m\end{subarray}}
\bigg[\Big|T^{\alpha_1\cdots\alpha_p\gamma_1\cdots\gamma_m}_{\beta_1\cdots\beta_q}\Big|^2
\Big(\sum_{i=1}^p\lambda_{\alpha_i}-\sum_{j=1}^q\lambda_{\beta_j}+\sum_{l=1}^m\mu_{\gamma_l}\Big)\bigg].
\end{split}\ee
Let
\begin{eqnarray}\label{maxmin}
\left\{ \begin{array}{ll}
\lambda_{\max}(x,X):=\max\limits_{1\leq i\leq r_1}\big\{\lambda_i(x,X)\big\}\\
~\\
\lambda_{\min}(x,X):=\min\limits_{1\leq i\leq r_1}\big\{\lambda_i(x,X)\big\}\\
~\\
\lambda_{\max}:=\max\limits_{x\in M, X\in\mathbb{S}T_xM}\lambda_{\max}(x,X)\\
~\\
\lambda_{\min}:=\min\limits_{x\in M, X\in\mathbb{S}T_xM}\lambda_{\min}(x,X)
\end{array} \right.
\end{eqnarray}
and similarly define $\mu_{\max}(x,X)$, $\mu_{\min}(x,X)$, $\mu_{\min}$ and $\mu_{\max}$.
We remark that in general $\lambda_{\max}(x,X)$ and $\lambda_{\min}(x,X)$ are only \emph{continuous} functions and may not be smooth. Nevertheless, continuity is enough to guarantee that $\lambda_{\max}$ and $\lambda_{\min}$ are both well-defined real numbers as the maximum and minimum in (\ref{maxmin}) are over the unit sphere bundle of $TM$, which is compact. Clearly $\lambda_{\max}$ and $\lambda_{\min}$ depends only on $h$ and $\omega$. Moreover, continuity is also enough to do integration as we shall see in the following lemma. Accordingly, for various $\mu$'s similar properties hold.
An efficient upper bound estimate for (\ref{T-formula}) is exhibited by the next result.
\begin{lemma}\label{estimate of RXXTT}
Let $(E,h)\rightarrow(M,\omega)$ be a uniformly RC $k$-positive Hermitian holomorphic vector bundle. Let $C(h,\omega)$ be the positive constant and $\Sigma_x$ the desired $k$-dimensional subspace in $T_xM$ in Lemma \ref{equivalent relation}. Then for any $x\in M$, we have
\be\begin{split}\label{estimate formula}&\int_{X\in\Sigma_x,|X|=1}<R_{X\overline{X}}(T)\Big|_x,T\Big|_x>d\theta(X)\\
\leq&
\Big[\lambda_{\max}\cdot p-\frac{C(h,\omega)}{k}\cdot q+\mu_{\max}\cdot m\Big]\cdot\text{V}(\mathbb{S}^{2k-1})\cdot\big|T\big|^2\Big|_x.\end{split}\ee
\end{lemma}
\begin{proof}
To simplify the notation, denote by $\oint f(X)$ the integral of the function $f(X)$ over $\{X\in\Sigma_x~|~|X|=1\}\cong\mathbb{S}^{2k-1}$ with respect to the spherical measure $d\theta(X)$.
First we claim that
\be\label{0.5}\oint\mu_i(x,X)\leq
\mu_{\max}\cdot\text{V}(\mathbb{S}^{2k-1})\ee
and
\be\label{1}0<\frac{C(h,\omega)}{k}\cdot\text{V}(\mathbb{S}^{2k-1})
\leq\oint\lambda_i(x,X)\leq\lambda_{\max}\cdot\text{V}(\mathbb{S}^{2k-1}).\ee
In fact, (\ref{0.5}) and the last inequality in (\ref{1}) are obvious as $\mu_i(x,X)\leq\mu_{\max}$ and $\lambda_i(x,X)\leq\lambda_{\max}$. On the other hand, by definition $$\lambda_i(x,X)\overset{(\ref{0.1})}{=}<R_{X\overline{X}}(e_i),e_i>=R(X,\overline{X},e_i,\overline{e_i}).$$ Thus
\be\begin{split}
\oint\lambda_i(x,X)&=\oint R(X,\overline{X},e_i,\overline{e_i})\\
&=\frac{\text{V}(\mathbb{S}^{2k-1})}{k}R(\Sigma_x;e_i,\overline{e_i})
\qquad\big(\text{by (\ref{def integral uniform})}\big)\\
&\geq\frac{\text{V}(\mathbb{S}^{2k-1})}{k}\cdot C(h,\omega),
\qquad\big(\text{by (\ref{inequality})}\big)
\end{split}\nonumber\ee
which gives the desired one in (\ref{1}).
Taking integrals $\oint(\cdot)$ on both sides of (\ref{T-formula}) yields
\be\begin{split}&\oint<R_{X\overline{X}}(T)\Big|_x,T\Big|_x>\\
=&\sum_{\begin{subarray}{1}
\alpha_1,\ldots,\alpha_p\\ \beta_1,\ldots,\beta_q\\ \gamma_1,\ldots,\gamma_m\end{subarray}}
\Big|T^{\alpha_1\cdots\alpha_p\gamma_1\cdots\gamma_m}_{\beta_1\cdots\beta_q}\Big|^2
\oint\Big(\sum_{i=1}^p\lambda_{\alpha_i}-\sum_{j=1}^q\lambda_{\beta_j}+\sum_{l=1}^m\mu_{\gamma_l}\Big)\\
\leq&\sum_{\begin{subarray}{1}
\alpha_1,\ldots,\alpha_p\\ \beta_1,\ldots,\beta_q\\ \gamma_1,\ldots,\gamma_m\end{subarray}}
\Big|T^{\alpha_1\cdots\alpha_p\gamma_1\cdots\gamma_m}_{\beta_1\cdots\beta_q}\Big|^2
\Big[\lambda_{\max}\cdot p-
\frac{C(h,\omega)}{k}\cdot q+\mu_{\max}\cdot m\Big]\cdot\text{V}(\mathbb{S}^{2k-1})
\quad\big(\text{by (\ref{0.5}) and (\ref{1})}\big)\\
=&\Big[\lambda_{\max}\cdot p-\frac{C(h,\omega)}{k}\cdot q+\mu_{\max}\cdot m\Big]\cdot\text{V}(\mathbb{S}^{2k-1})\cdot\big|T\big|^2\Big|_x.
\end{split}\nonumber\ee
This gives the desired estimate (\ref{estimate formula}).
\end{proof}
We are ready now to complete the proof of Theorem \ref{RC vanishing theorem} in the following lemma.
\begin{lemma}
Let $(E,h)\rightarrow(M,\omega)$ be a uniformly RC $k$-positive Hermitian holomorphic vector bundle and $F$ any holomorphic vector bundle over $M$. With the notation above understood and set two positive constants by
\begin{eqnarray}\label{def of c1 c2}
C_1=C_1(h,\omega):=\displaystyle\frac{k\cdot\lambda_{\max}}{C(h,\omega)},\quad C_2=C_2(F,h,\omega):=\left\{ \begin{array}{ll}
\displaystyle\frac{k\cdot\mu_{\max}}{C(h,\omega)},&\text{if $\mu_{\max}>0$}\\
~\\
\displaystyle 1.&\text{if $\mu_{\max}\leq0$}
\end{array} \right.
\end{eqnarray}
Then for all $p,q,m\in\mathbb{Z}_{\geq 0}$ with $q> C_1\cdot p+C_2\cdot m$, we have
\be H^0\big(M,E^{\otimes p}\otimes (E^{\ast})^{\otimes q}\otimes F^{\otimes m}\big)=0\nonumber\ee
and hence Theorem \ref{RC vanishing theorem} holds.
\end{lemma}
\begin{proof}
First note that the positivity of $\lambda_{\max}$ and hence $C_1$ follows from the inequality (\ref{1}).
Let $T\in H^0\big(M,E^{\otimes p}\otimes (E^{\ast})^{\otimes q}\otimes F^{\otimes m}\big)$ and $\big|T\big|$ attains its \emph{maximum} at $x_0\in M$. On the one hand, Lemma \ref{Bochner fromula0 lemma} yields
\be\label{2}\int_{X\in \Sigma_{x_0},~|X|=1}<R_{X\overline{X}}(T)\Big|_{x_0},T\Big|_{x_0}>
d\theta(X)\geq0.\ee
On the other hand, applying Lemma \ref{estimate of RXXTT} to $x_0$ leads to
\be\label{3}\begin{split}&\int_{X\in \Sigma_{x_0},~|X|=1}<R_{X\overline{X}}(T)\Big|_{x_0},T\Big|_{x_0}>d\theta(X)\\
\leq&-\frac{C(h,\omega)}{k}\cdot
\Big[q-\big(\frac{k\lambda_{\max}}{C(h,\omega)}\cdot p+\frac{k\mu_{\max}}{C(h,\omega)}\cdot m\big)\Big]\cdot\text{V}(\mathbb{S}^{2k-1})\cdot\big|T\big|^2\Big|_{x_0}\\
\leq&
-\frac{C(h,\omega)}{k}\cdot
\Big[q-\big(C_1\cdot p+C_2\cdot m\big)\Big]\cdot
\text{V}(\mathbb{S}^{2k-1})\cdot\big|T\big|^2\Big|_{x_0}.
\quad\big(\text{by (\ref{def of c1 c2})}\big)
\end{split}\ee
If $q> C_1\cdot p+C_2\cdot m$, (\ref{2}) and (\ref{3}) together imply
$$\int_{X\in \Sigma_{x_0},~|X|=1}<R_{X\overline{X}}(T)\Big|_{x_0},T\Big|_{x_0}>d\theta(X)=0,$$
which, in turn via (\ref{3}) tells us that the only possibility is $\big|T\big|^2\Big|_{x_0}=0$. The maximum of $\big|T\big|$ at $x_0$ then implies $T\equiv0$. This completes the proof of this lemma and hence that of Theorem \ref{RC vanishing theorem}.
\end{proof}
\section{Proof of Theorem \ref{rick lead to uniform}}\label{proof of theorem rick yield rc}
\subsection{A lemma and the proof of Theorem \ref{rick lead to uniform}}
The proof of Theorem \ref{rick lead to uniform} follows from the
next result, which was essentially obtained by Ni-Zheng in \cite[Prop. 3.1]{NZ}, although the conclusion there was stated for K\"{a}hler manifolds.
\begin{lemma}[Ni-Zheng]\label{NZ estimate lemma}
Let $(M,\omega)$ be a compact Chern-K\"{a}hler-like Hermitian manifold, $x\in M$, and the $k$-dimensional subspace $\Sigma\subset T_xM$ minimize (resp. maximize) the $k$-scalar curvature $S_k(x,\cdot)$ at $x$. Then for any $Y\in\Sigma$ and $Z\in\Sigma^{\bot}:=\Big\{W\in T_xM~\big|~W\bot\Sigma\Big\}$, we have
\be\label{NZ estimate1}
\int_{X\in\Sigma,~|X|=1}R(X,\overline{X},Y,\overline{Z})d\theta(X)=
\int_{X\in\Sigma,~|X|=1}R(X,\overline{X},Z,\overline{Y})d\theta(X)=0,\ee
and
\be\label{NZ estimate2}
\int_{X\in\Sigma,~|X|=1}R(X,\overline{X},Z,\overline{Z})d\theta(X)\underset{(\leq)}
{\geq}\frac{\text{V}(\mathbb{S}^{2k-1})}{k(k+1)}\cdot S_k(x,\Sigma)\cdot|Z|^2.\ee
\end{lemma}
\begin{remark}
Lemma \ref{NZ estimate lemma} can be viewed as a $k$-dimensional generalization of \cite[Lemma 6.1]{Yang18} for holomorphic sectional curvature in the K\"{a}hler situation, which still holds true for CKL Hermitian metrics (\cite[Lemma 2.4]{Li21}). The proof in \cite[Prop. 3.1]{NZ}, in spite of the similar principle to the one in \cite[Lemma 6.1]{Yang18}, is skillful. For the reader's convenience as well as for completeness, we include a detailed proof below. We shall see in the process of the proof that what we really need is various K\"{a}hler-type symmetries of the Chern curvature tensor, which, as explained in Remark \ref{remark}, is satisfied by the CKL Hermitian metrics.
\end{remark}
We first explain how Lemma \ref{NZ estimate lemma}, together with the materials in Section \ref{Preliminaries}, leads to Theorem \ref{rick lead to uniform} and postpone its proof to the next subsection.
\emph{Proof of Theorem \ref{rick lead to uniform}}.
\begin{proof}
Assume that $(M,\omega)$ is an $n$-dimensional compact \emph{Chern-K\"{a}hler-like} Hermitian manifold and $\text{Ric}_{k}(\omega)>0$ (resp. $\text{Ric}_{k}(\omega)<0$) for some $k\in\{1,\ldots,n\}$. Let
\be\label{lower bound for riccik} D:=\min_{\begin{subarray}{1}x\in M\\
\Sigma\in\mathbb{G}_k(T_xM)\\
X\in\Sigma,~|X|=1\end{subarray}}
\text{Ric}_k(x,\Sigma)(X,\overline{X})>0\qquad(\text{resp. $D:=\max\cdots<0$}).\ee
The definitions of $\text{Ric}_k(\omega)$ and $S_k(\omega)$ in (\ref{def riccik}) and (\ref{def sk}) then imply that
\be\label{lower bound for sk}\min_{\overset{x\in M}{\Sigma\in\mathbb{G}_k(T_xM)}}
S_k(x,\Sigma)\geq kD>0\qquad(\text{resp. $\max\cdots\leq kD<0$}).\ee
For any $x\in M$, let the $k$-dimensional subspace $\Sigma\subset T_xM$ \emph{minimize} (resp. \emph{maximize}) the $k$-scalar curvature $S_k(x,\cdot)$ at $x$, as required in Lemma \ref{NZ estimate lemma}. We shall show that such $\Sigma$ satisfies the inequality (\ref{RC-positivity}) and hence $(TM,\omega)$ is uniformly RC $k$-positive (resp. uniformly RC $k$-negative) over $(M,\omega)$. In fact, for any $X\in T_xM$, decompose $X$ as $X=X_1+X_2$, where $X_1\in\Sigma$ and $X_2\in\Sigma^{\bot}$. Like before, denote by $\oint f(Y)$ the integral of $f(Y)$ over $\{Y\in\Sigma~|~|Y|=1\}$ with respect to the spherical measure. Then
\be
\begin{split}
&R^{(TM,\omega)}(\Sigma;X,\overline{X})\\
=&\frac{k}{\text{V}(\mathbb{S}^{2k-1})}
\oint R(Y,\overline{Y},X,\overline{X})
\qquad\big(\text{by (\ref{def integral uniform})}\big)\\
=&\frac{k}{\text{V}(\mathbb{S}^{2k-1})}
\oint
R\big(Y,\overline{Y},X_1+X_2,
\overline{X_1}+\overline{X_2}\big)\\
=&\frac{k}{\text{V}(\mathbb{S}^{2k-1})}
\oint\Big[R\big(Y,\overline{Y},X_1,
\overline{X_1}\big)+
R\big(Y,\overline{Y},X_2,
\overline{X_2}\big)\Big]
\qquad\big(\text{by (\ref{NZ estimate1})}\big)\\
=&\text{Ric}_k(x,\Sigma)(X_1,\overline{X_1})+
\frac{k}{\text{V}(\mathbb{S}^{2k-1})}
\oint R\big(Y,\overline{Y},X_2,
\overline{X_2}\big)
\qquad\big(\text{by (\ref{def integral riccik}) and $\omega$ CKL}\big)\\
\underset{(\leq)}{\geq}&\text{Ric}_k(x,\Sigma)(X_1,\overline{X_1})+
\frac{S_k(x,\Sigma)}{k+1}|X_2|^2\qquad\big(\text{by (\ref{NZ estimate2})}\big)\\
\underset{(\leq)}{\geq}&D|X_1|^2+\frac{kD}{k+1}|X_2|^2\qquad
\big(\text{by (\ref{lower bound for riccik}) and (\ref{lower bound for sk})}\big)\\
\underset{(\leq)}{\geq}&\frac{kD}{k+1}|X|^2.
\end{split}
\ee
This completes the proof of Theorem \ref{rick lead to uniform}.
\end{proof}
\subsection{Proof of Lemma \ref{NZ estimate lemma}}
The proof here basically follows the strategy in \cite[Prop. 3.1]{NZ}, but with more clear presentation in places.
Let $U(n)$ be the isometry group of $(T_xM,\omega_x=<,>)$ and
$$\mathfrak{u}(n):=
\Big\{a\in\text{Hom}(T_xM)\big|<a(W_1),W_2>+<W_1,a(W_2)>=0,~\forall~ W_1,W_2\in T_xM\Big\}$$
the (real) Lie algebra of $U(n)$. It is well-known that
$$e^{ta}:=\exp(ta):=\text{id}+\sum_{i=1}^{\infty}\frac{(ta)^i}{i!}\in U(n)\quad\text{and}\quad \frac{\text{d}^i}{\text{d}t^i}(e^{ta})=a^ie^{ta},\quad\text{
for any $t\in\mathbb{R}$}.$$
As before denote by $\oint f(X)d\theta(X)$ the integral of $f$ over $\{X\in\Sigma,|X|=1\}$, where $\Sigma$ is the $k$-dimensional subspace in $T_xM$ \emph{minimizing} (resp. \emph{maximizing}) $S_k(x,\cdot)$ at $x$.
For any $a\in\mathfrak{u}(n)$ consider the following \emph{real}-valued function
$$f(t):=\oint H(e^{ta}X)d\theta(X),\qquad t\in\mathbb{R}.$$
The minimum (resp. maximum) of $f(t)$ at $t=0$ implies that $f'(t)=0$ and $f''(t)\geq0$ (resp. $\leq0$).
Direct calculations, together with the fact that the Chern curvature tensor $R$ satisfy K\"{a}hler-like symmetries ensured by the condition of CKL, lead to
\be\label{5}\oint\Big[R\big(a(X),\overline{X},X,\overline{X}\big)+
R\big(X,\overline{a(X)},X,\overline{X}\big)\Big]d\theta(X)=0\ee
and
\be\label{6}\begin{split}
\oint\Big[&R\big(a^2(X),\overline{X},X,\overline{X}\big)+
R\big(X,\overline{a^2(X)},X,\overline{X}\big)+
4R\big(a(X),\overline{a(X)},X,\overline{X}\big)\\
+&R\big(a(X),\overline{X},a(X),\overline{X}\big)+
R\big(X,\overline{a(X)},X,\overline{a(X)}\big)
\Big]d\theta(X)\underset{(\leq)}{\geq}0.
\end{split}\ee
We may assume that the two vectors $Y\in\Sigma$ and $Z\in\Sigma^{\bot}$ are both \emph{nonzero} as otherwise (\ref{NZ estimate1}) and (\ref{NZ estimate2}) trivially hold. Let
$$a(\cdot):=\sqrt{-1}\big(<\cdot,Y>Z+<\cdot,Z>Y\big)$$
and it can be easily checked that $a(\cdot)\in\mathfrak{u}(n)$.
Then
\be\label{a formula}a(X)=\sqrt{-1}<X,Y>Z,\quad a^2(X)=-<X,Y>|Z|^2Y.\ee
Apply (\ref{5}) to (\ref{a formula}) and also the one with $Z$ being replaced by $\sqrt{-1}Z$, and sum the two we have
\be\label{8}\oint<X,Y>R(Z,\overline{X},X,\overline{X})d\theta(X)=0.\ee
We take a unitary basis of $\Sigma$ as follows \be\label{7}\big\{E_1=Y/|Y|,E_2,\ldots,E_k\big\},\ee
and simply write $V:=\text{V}(\mathbb{S}^{2k-1})$. Then
\be\begin{split}
0=&\oint<X,Y>R(Z,\overline{X},X,\overline{X})d\theta(X)\qquad\big(\text{by (\ref{8})}\big)\\
=&\frac{V}{k(k+1)}
\sum_{i,j=1}^k\Big[<E_i,Y>R(Z,\overline{E_i},E_j,\overline{E_j})+
<E_i,Y>R(Z,\overline{E_j},E_j,\overline{E_i})\Big]~\big(\text{by (\ref{integral formula2})}\big)\\
=&\frac{2V}{k(k+1)}\sum_{j=1}^k
R(Z,\overline{Y},E_j,\overline{E_j})\quad\big(\text{by (\ref{7}) and $\omega$ CKL}\big)\\
=&\frac{2}{k+1}\oint R(Z,\overline{Y},X,\overline{X})d\theta(X).
\quad\big(\text{by (\ref{integral formula1})}\big)
\end{split}
\ee
This and its conjugate lead to (\ref{NZ estimate1}).
We now derive (\ref{NZ estimate2}). Apply (\ref{6}) to the above (\ref{a formula}) and also the one being replaced by $\sqrt{-1}Z$, and sum the two we have
\be\label{9}\begin{split}
&4\oint|<X,Y>|^2R(Z,\overline{Z},X,\overline{X})d\theta(X)\\
\underset{(\leq)}{\geq}&
\oint\Big[<X,Y>R(Y,\overline{X},X,\overline{X})+
<Y,X>R(X,\overline{Y},X,\overline{X})\Big]|Z|^2d\theta(X).
\end{split}\ee
Arbitrarily choose a unitary basis $\{E_1,\ldots,E_k\}$ of $\Sigma$, which is irrelevant to (\ref{7}). Then
\be\label{10}\begin{split}
&\oint\Big[\text{LHS of (\ref{9})}\Big]d\theta(Y)\\
=&\frac{4V}{k}\sum_{i=1}^k\oint|<X,E_i>|^2R(Z,\overline{Z},X,\overline{X})d\theta(X)
\qquad\big(\text{by (\ref{integral formula1})}\big)\\
=&\frac{4V^2}{k^2(k+1)}\sum_{i,j,t=1}^k
\Big[|<E_j,E_i>|^2R(Z,\overline{Z},E_t,\overline{E_t})\\ &+<E_j,E_i>\overline{<E_t,E_i>}R(Z,\overline{Z},E_t,\overline{E_j})\Big]
\qquad\big(\text{by (\ref{integral formula2})}\big)\\
=&\frac{4V^2}{k^2}\sum_{i=1}^kR(Z,\overline{Z},E_i,\overline{E_i})\\
=&\frac{4V}{k}\oint R(Z,\overline{Z},X,\overline{X})d\theta(X).
\qquad\big(\text{by (\ref{integral formula1})}\big)
\end{split}\ee
Analogous to (\ref{10}) it can be derived that
\be\label{11}
\oint\Big[\text{RHS of (\ref{9})}\Big]d\theta(Y)=
\frac{4V^2}{k^2(k+1)}S_k(x,\Sigma)|Z|^2.
\ee
Putting (\ref{9}), (\ref{10}) and (\ref{11}) together yields the desired
(\ref{NZ estimate2}).
\section{Further questions and remarks}\label{further remarks}
In the process of proofs we have seen that the \emph{strict} positivity or negativity in Theorems \ref{RC vanishing theorem} and \ref{rick lead to uniform} is needed.
In view of Theorem \ref{Bochner}, it is natural to
wonder whether the conclusions remain true if only assuming
quasi-positivity or quasi-negativity of various curvature conditions.
Therefore the following question can be proposed (see \cite[Conj. 1.9]{Yang20}).
\begin{question}\label{question}
Whether the conclusions in Theorems \ref{RC vanishing theorem},
\ref{rick lead to uniform}, \ref{CKL result} and
\ref{rational connectedness} hold true if various positivity or negativity conditions
are weakened respectively to quasi-positivity or quasi-negativity?
\end{question}
\begin{remark}
Matsumura showed in \cite{Ma} that
a projective K\"{a}hler manifold with quasi-positive
holomorphic sectional curvature is rationally connected, which indeed provides some positive evidence towards Question \ref{question} as well as \cite[Conj. 1.9]{Yang20}.
\end{remark}
Among various geometric positivity concepts for Hermitian holomorphic vector bundles, the \emph{Griffiths positivity} (\cite{Gr})
may be the best-known. In view of the results proved in this article, it may be interesting to propose the
following notion.
\begin{definition}\label{Griffith kl definition}
A rank $r$ Hermitian holomorphic vector bundle $(E,h)$ over an $n$-dimensional Hermitian manifold $(M,\omega)$
is called \emph{Griffiths $(k,l)$-positive} $(1\leq k\leq n,~1\leq l\leq r)$ if at each point $x\in M$, and
for any $\Sigma\in\mathbb{G}_k(T_xM)$ and any
$\sigma\in\mathbb{G}_l(E_x)$, we have
\be\label{Griffith kl formula}
R^{(E,h)}(\Sigma;\sigma):=
\sum_{\underset{1\leq j\leq l}{1\leq i\leq k}}
R^{(E,h)}(E_i,\overline{E_i},e_j,\overline{e_j})>0,\ee
where $\{E_1,\ldots E_k\}$ (resp. $\{e_1,\ldots,e_l\}$) is a unitary basis of
$\Sigma$ (resp. $\sigma$). \emph{Griffiths $(k,l)$-negativity} and various quasi-versions can be similarly defined.
\end{definition}
\begin{remark}
\begin{enumerate}
\item
By (\ref{integral formula1}), an alternative definition for $R^{(E,h)}(\Sigma;\sigma)$ is
\be\begin{split}
&R^{(E,h)}(\Sigma;\sigma)\\
=&
\frac{k^2}{\text{V}(S^{2k-1})\text{V}(S^{2l-1})}
\iint_{\underset{\{X\in\Sigma,~|X|=1\}}{\{u\in\Gamma,~|u|=1\}}}
R^{(E,h)}(X,\overline{X},u,\bar{u})d\theta(X)d\theta(u).
\end{split}\nonumber\ee
\item
Griffiths $(1,1)$-positivity is the original Griffiths positivity, in which case the metric $\omega$ is irrelevant. By definition Griffiths $(k,l)$-positivity (resp. negativity) implies Griffiths $(k+1,l)$-positivity and $(k,l+1)$-positivity (resp. negativity). Thus the condition Grffiths $(k,l)$-positivity becomes weaker as $k$ or $l$ increases (compare to the $k$-scalar curvatures in Remark \ref{remark2}).
\item
The subject of cohomology vanishing theorems for Griffiths positive holomorphic vector bundles occupies a central role in several complex variables and algebraic geometry (\cite[\S 7]{De}, \cite[\S 6]{SS}). It seems to be possible to generalize these to Griffiths $(k,l)$-positive vector bundles. The author hopes to discuss it in a separate article in the near future.
\end{enumerate}
\end{remark}
Another two versions of $(k,l)$-positivity (resp. negativity) related to Definition \ref{uniform rc k positivity} are as follows.
\begin{definition}\label{uniform RC kl definition}
Let $(E,h)$ be a rank $r$ Hermitian holomorphic vector bundle over an $n$-dimensional Hermitian manifold $(M,\omega)$,
and $k\in\{1,\ldots,n\}$ and $l\in\{1,\ldots,r\}$.
\begin{enumerate}
\item
It is called \emph{RC $(k,l)$-positive} \big(resp. \emph{BC $(k,l)$-positive}\big) at $x\in M$ if for any $\sigma\in\mathbb{G}_l(E_x)$ \big(resp. $\Sigma\in\mathbb{G}_k(T_xM)$\big), there exists $\Sigma\in \mathbb{G}_k(T_xM)$ \big(resp. $\sigma\in\mathbb{G}_l(E_x)$\big) such that $R^{(E,h)}(\Sigma;\sigma)>0$. If this holds at each $x\in M$, then it is called RC $(k,l)$-positive \big(resp. BC $(k,l)$-positive\big).
\item
It is called \emph{uniformly RC $(k,l)$-positive} \big(resp. \emph{uniformly BC $(k,l)$-positive}\big) at $x\in M$ if there exists $\Sigma\in\mathbb{G}_k(T_xM)$ \big(resp. $\sigma\in\mathbb{G}_l(E_x)$\big) such that for every $\sigma\in\mathbb{G}_l(E_x)$ \big(resp. $\Sigma\in\mathbb{G}_k(T_xM)$\big), $R^{(E,h)}(\Sigma;\sigma)>0.$ If this holds at each $x\in M$, then it is called uniformly RC $(k,l)$-positive \big(resp. uniformly BC $(k,l)$-positive\big).
\end{enumerate}
(Uniform) RC $(k,l)$-negativity or BC $(k,l)$-negativity and various quasi-versions can be similarly defined.
\end{definition}
\begin{remark}\label{uniform RC remark}
\begin{enumerate}
\item
Uniform RC $(k,1)$-positivity is the one in Definition \ref{RC-positivity}, RC $(1,1)$-positivity is the RC-positivity in \cite{Yang18}, and BC $(1,l)$-positivity is the BC $l$-positivity in \cite[p. 280]{Ni21-2}. For CKL Hermitian manifold $(M,\omega)$, $(TM,\omega)$ is (uniformly) RC $(k,l)$-positive (resp. negative) if and only if it is (uniformly) BC $(l,k)$-positive (resp. negative).
\item
It turns out that BC $k$-positivity of the tangent bundle of a Hermitian metric for some $k$ implies $h^{k,0}=0$ (\cite[Coro. 4.4]{Ni21-2}). It may be interesting to explore possible consequences of geometric importance for general uniform RC (resp. BC) $(k,l)$-positivity.
\item
By the proof of Lemma \ref{equivalent relation}, it is easy to see that RC $(k,l)$-positivity or BC $(k,l)$-positivity on a \emph{compact} Hermitian manifold amounts to
$$\min_{x\in M}
\min_{\sigma\in\mathbb{G}_l(E_x)}
\max_{\Sigma\in\mathbb{G}_k(T_xM)}R^{(E,h)}(\Sigma;\sigma)>0$$
or
$$\min_{x\in M}\min_{\Sigma\in\mathbb{G}_k(T_xM)}
\max_{\sigma\in\mathbb{G}_l(E_x)}
R^{(E,h)}(\Sigma;\sigma)>0,$$
and uniform RC $(k,l)$-positivity or uniform BC $(k,l)$-positivity amounts to
$$\min_{x\in M}\max_{\Sigma\in\mathbb{G}_k(T_xM)}
\min_{\sigma\in\mathbb{G}_l(E_x)}R^{(E,h)}(\Sigma;\sigma)>0$$
or
$$\min_{x\in M}\max_{\sigma\in\mathbb{G}_l(E_x)}
\min_{\Sigma\in\mathbb{G}_k(T_xM)}R^{(E,h)}(\Sigma;\sigma)>0$$
respectively.
\item
Besides the curvatures mentioned in this article, some other interesting curvature notions were also introduced by Ni and Ni-Zheng. We refer the reader to their survey article \cite{NZ19} for more details.
\end{enumerate}
\end{remark}
| 5a46aff636de48d0fd9f1438096ba36f13828670 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $X$ be an integral algebraic variety over a field $F$, i.e. a separated scheme and of finite type over $F$.
The $i$-th unramified cohomology group of $X$ denoted by $H_{nr}^{i}(F(X)/F,\mathbb{Q/Z}(j))$ (see section \S \ref{sec:pre} ) is defined by a subgroup of $H^{i}(F(X),\mathbb{Q/Z}(j))$,
which were first introduced by Colliot-Thélène and Ojanguren \cite{colliot1989varietes}.
We will write $\bar{H}_{nr}^{i}(F(X)/F,\mathbb{Q/Z}(j))$ as $H_{nr}^{i}(F(X)/F,\mathbb{Q/Z}(j))/H^{i}(F,\mathbb{Q/Z}(j))$.
The unramified cohomology groups are homotopy invariants
and can be used as an invariant to disprove the rationality of an algebraic variety.
The non-triviality of the group $\bar{H}_{nr}^{i}(F(X)/F,\mathbb{Q/Z}(j))$ implies the stably irrational property of $X$.
Peyre \cite{peyre1993unramified}, Colliot-Thélène and Pirutka \cite{colliot2016hypersurfaces}, Schreieder \cite{schreieder2019stably} have given many examples of varieties which are not stably rational but unirational.
The unramified cohomology groups also play a role on rational points. In \cite{harari2015weak,harari2013local}, a certain
subgroup of the third unramified cohomology group can control the defect of weak approximation for tori over $p$-adic function fields.
There are no general methods to estimate the unramified cohomology groups for general variety $X$. Let $W$ be a torus over $F$.
In 1995, Colliot-Th\'el\`ene \cite[p. 39]{ct1995} raised the problem: for $n$ prime to $\mathrm{char}(F)$ and $i>0$ , determine the group $\bar{H}_{nr}^{i}(F(W)/F,\mu_n^{\otimes(i-1)})$.
The first unramified cohomology group $\bar{H}_{nr}^{1}(F(W)/F,\Z/n\Z)$ is trivial.
The second unramified cohomology group $\bar{H}_{nr}^{2}(F(W)/F,\mathbb{Q/Z}(1))$ coincides with the unramified Brauer group $\mathrm{Br}_{nr}(W)/\mathrm{Br}(F)$ which is isomorphic to the group $\Sha_\omega^2(\hat{W})$ and has been extensively studied (see \cite{colliot1987principal} and \cite{karpilovsky1987schur}).
In 2013, Blinstein and Merkurjev investigated Colliot-Th\'el\`ene's problem when $i=3$ and obtained an exact sequence for the group $\bar{H}_{nr}^{3}(F(W)/F,\mathbb{Q/Z}(2))$, see \cite[Proposition 5.9]{m2013}.
Let $K/F$ be a finite abelian extension and $W$ the norm one torus $R_{K/F}^{(1)}\mathbb{G}_{m,K}$. Based on Blinstein and Merkurjev's work, we explicitly determine the $p$-primary component of $\bar{H}_{nr}^{3}(F(W)/F,\mathbb Q/\mathbb Z(2))$ which we denote by $\bar{H}_{nr}^{3}(F(W)/F,\mathbb Q/\mathbb Z(2))\{p\}$ for any odd prime~$p$; furthermore, if $K/F$ has odd degree, we explicitly determine $\bar{H}_{nr}^{3}(F(W)/F,\mathbb Q/\mathbb Z(2))$.
\begin{thm*}
Let $W=R_{K/F}^{(1)}(\mathbb G_{m,K})$ and $K/F$ an abelian extension with $\mathrm{Gal}(K/F)=G\simeq\oplus_{i=1}^mC_i$ such that each $C_i$ is cyclic and $\#C_{i-1}|\#C_i$ $(1<i\leq m)$. Denote $d_i=(m-i)(m-i-1)/2$.
Then for any odd prime $p$,
$$\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))\{p\}\simeq H^3(G,K^*)\{p\} \oplus\oplus_{i=1}^{m-2}C_i\{p\}^{d_{i}}.
$$
Moreover, if $K/F$ has odd degree, then
$$
\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))\simeq H^3(G,K^*) \oplus \oplus_{i=1}^{m-2}C_i^{d_{i}}.
$$
\end{thm*}
If $G=\Z/n\oplus \Z/n$ and $n$ is odd, then $\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2)) \simeq H^3(G,K^*)$.
In particular, if $F$ is a local field or some special number field (see Examples \ref{exa:un}), then $\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))=0$. However, $\bar{H}^2_{nr}(F(W),\mathbb Q/\mathbb Z(1))=\Z/n$ is always nontrivial.
\bigskip
{\bf Acknowledgements. }
The second author is supported by National Key R$\&$D Program of China.
\section{Preliminaries}\label{sec:pre}
In this section, we will introduce the unramified cohomology groups.
The notation and definition coincide with \cite{m2013,ct1995,ct1977}.
Let $F$ be the base field (of arbitrary characteristic) and $\Gamma=\Gal(F^{sep}/F)$ the absolute Galois
group of $F$.
Given an integer $j>0$,
we denote $\Q/\Z(j)$ as the direct sum of
$$
\Q_p/\Z_p(j)=
\begin{cases}
\underset{n}{\mathrm{colim}}\ \mu_{p^n}^{\otimes j}, \quad p\neq \mathrm{char}(F)\\
\underset{n}{\mathrm{colim}}\ W_n\Omega_{\log}^j[-j], \quad p= \mathrm{char}(F)
\end{cases}
$$
in the derived category of sheaves of abelian groups on the big \'etale site of $\mathrm{Spec}\ F$ over all primes $p$, where $ W_n\Omega_{\log}^j[-j]$ is the sheaf of logarithmic de Rham–Witt differentials (see \cite{ASENS_1979_4_12_4_501_0} and \cite{m2013}).
Given a field extension $L/F$, let $A$ be a rank one discrete valuation ring with fraction field $L$
which contains $F$.
The $i$-th unramified cohomology group of $L/F$ is defined by the group
$$
H_{nr}^i(L/F,\Q/\Z(j))):=\bigcap_{A\in P(L)}\mathrm{im}\ (H^i(A,\Q/\Z(j))\longrightarrow H^{i}(L,\Q/\Z(j)))
$$
where $P(L)$ is the set of all rank one discrete valuation rings which contains $F$ and has quotient field $L$.
They can also be defined by the intersection of the kernel of residue map $\partial_A$ for $A\in P(L)$ when $\mathrm{char}(F)=0$ \cite{ct1995}.
If $X$ is a smooth integral variety over $F$, the $i$-th unramifiel cohomology group of $X$ is defined by $H_{nr}^i(F(X)/F,\Q/\Z(j))$.
These groups are only dependent on the function field of $X$, and when $X$ is proper, we can replace $P(L)$ by the set of local rings of codimension one points in $X$ \cite{ct1995}.
If $i=2$, the unramified chomology group coincides with the unramified Brauer group of $X$.
We will denote $\bar{H}_{nr}^{i}(F(X)/F,\mathbb{Q/Z}(j))$ by $H_{nr}^{i}(F(X)/F,\mathbb{Q/Z}(j))/H^{i}(F,\mathbb{Q/Z}(j))$.
Recall that an algebraic torus of dimension $n$ over $F$ is an algebraic group $T$ such that $T_{sep}$ is isomorphic to $\mathbb{G}_m^{n}$.
For an algebraic torus $T$, we write $\hat{T}_{sep}$ as the $\Gamma$-module $\mathrm{Hom}(T_{sep},\mathbb{G}_m)$ and call it character lattice of $T$.
The contravariant functor $T \mapsto \hat{T}_{sep}$ is an anti-equivalence between the
category of algebraic torus over $F$ and the category of $\Gamma$-lattices.
We write $\hat{T}$ as the character group $\mathrm{Hom}_F(T,\mathbb G_m)$ and $T^\circ$ as the dual torus of $T$ such that $(\hat{T}^\circ)_{sep}=(\hat{T}_{sep})^\circ$.
Let $K/F$ be a finite Galois extension with $G=\mathrm{Gal}(K/F)$.
We embed $K$ to $F_{sep}$, the separable closure of $F$, such that the absolute Galois group $\Gamma_K$ is a subgroup of $\Gamma$. We mainly consider the unramified cohomology of the norm one torus $W=R_{K/F}^{(1)}(\mathbb G_{m,K})$. It is defined by the equation $N_{K/F}(\Xi)=1$ and can be constructed by the exact sequence
$$
1\longrightarrow W\longrightarrow R_{K/F}(\mathbb G_{m,K})\longrightarrow\mathbb G_{m}\longrightarrow 1
$$
and for its character lattice, we have
$$
0\longrightarrow\mathbb Z\stackrel{N_G}{\longrightarrow} \mathbb Z[G]\longrightarrow\hat{W}_{sep}\longrightarrow 0
$$
where $N_G$ maps $1$ to $\sum_{g\in G}g$.
A torus $T$ is called quasisplit if its character lattice $\hat{T}_{sep}$ is a permutation $\Gamma$-module, i.e. isomorphic to the finite direct-sum of $\mathbb Z[\Gamma/\Gamma_i]$ for open subgroups $\Gamma_i\leq \Gamma$.
And $T$ is called flasque (resp. coflasque) if $H^1(L,\hat{T}_{sep}^\circ)=0$ (resp. $H^1(L,\hat{T}_{sep})=0$) for all finite field extension $L/F$.
A flasque resolution (resp. coflasque resolution) of a torus $T$ is an exact sequence of torus
$$
1\longrightarrow S\longrightarrow P\longrightarrow T\longrightarrow 1
$$
(resp. $1\longrightarrow T\longrightarrow P\longrightarrow M\longrightarrow 1$ ) where $P$ is quasisplit and $S$ is flasque (resp. $M$ is coflasque).
If $T$ is the norm one torus of a finite Galois extension, then the flasque resolution (resp. coflasque resolution) of $T$ can be constructed as follows.
\begin{prop}(\cite{ct1977} , Proposition 15)\label{flasqueresolution}
Let $K/F$ be a finite Galois extension with $G=\mathrm{Gal}(K/F)$, generated by $\{s_i\}_{1\leq i\leq r}$. The epimorphism
$$
R_{K/F}(\mathbb{G}_{m,K})^r\rightarrow R^{(1)}_{K/F}(\mathbb{G}_{m,K}),
$$
defined by $x\mapsto s_i(x)/x$, provides a flasque resolution of $T$.
\end{prop}
For a torus $S$ over $F$, Blinstein and Merkurjev \cite{m2013} give an estimate of $\bar{H}^3_{nr}(F(S),\mathbb Q/\mathbb Z(2))$:
\begin{thm}(\cite{m2013}, Proposition 5.9)\label{mainlem}
Let $S$ be a torus over $F$ and let
$$
1\longrightarrow T\longrightarrow P\longrightarrow S\longrightarrow 1
$$
be a flasque resolution of $S$. Then we have an exact sequence
\begin{align*}
0\longrightarrow CH^2(BT)_{tors}\longrightarrow H^1(F,T^\circ)\longrightarrow&\bar{H}^3_{nr}(F(S),\mathbb Q/\mathbb Z(2))\\
\longrightarrow& H^0(F,S^2(\hat{T}_{sep}))/\Dec\longrightarrow H^2(F,T^\circ).
\end{align*}
For an odd prime $p$, there is a canonical direct sum decomposition
$$
\bar{H}^3_{nr}(F(S),\mathbb Q/\mathbb Z(2))\{p\}\simeq H^1(F,T^\circ)\{p\}\oplus (H^0(F,S^2(\hat{T}_{sep}))/\Dec)\{p\}.
$$
\end{thm}
The first term is the torsion subgroup of the second Chow group of the classify variety of $T$.
The symbol $\Dec$ in the fourth term is the subgroup $\Dec(H^0(F,S^2(\hat{T}_{sep})))$ of $H^0(F,S^2(\hat{T}_{sep}))$, defined as follows.
Given a $\Gamma$-module $A$, $\Dec(A)\subset S^2(A)$ is generated by $(A^{\Gamma})^2$ and $\Qtr_{\Gamma'}(A^{\Gamma'})$ for all open subgroups $\Gamma'\subset\Gamma$, with the following definition
\begin{align*}
\Qtr_{\Gamma'}:\ A^{\Gamma'}\to&\ S^2(A)^{\Gamma}\\
a\mapsto&\ \sum_{i<j}\gamma_i(a)\cdot \gamma_j(a)
\end{align*}
where $\{\gamma_i\}$ is a representative for the left cosets of $\Gamma'$ in $\Gamma$.
\section{Main results}
We continue to use the notations in \S \ref{sec:pre}.
For the norm one torus $W=R^{(1)}_{K/F}(\mathbb{G}_{m,K})$ of Galois extension $K/F$,
we fix a flasque resolution of $W$ constructed from Lemma \ref{flasqueresolution}:
\begin{equation}\label{seq:flasque}
1\longrightarrow T\longrightarrow P\longrightarrow W\longrightarrow 1
\end{equation}
where $P=R_{K/k}(\mathbb{G}_{m,K})^r$ and $r$ is the number of generators of $G=\Gal(K/F)$.
We mainly consider the $3$-th unramified cohomology group
$\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))$ and it can be expressed by $T$ from Theorem \ref{mainlem}.
Building on this work, we will reduce $\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))$ to a form determined only by the field extension $K/F$ and the Galois group $\mathrm{Gal}(K/F)$.
\begin{lem}\label{lemma2}
If $M$ is an algebraic torus over $F$ and is split by a finite Galois extension $K/F$ with $G=\mathrm{Gal}(K/F)$, then $H^1(K,M)=0$ and hence $H^1(F,M)\simeq H^1(G,M(K))$.
\end{lem}
\begin{proof}
$M$ is split by $K/F$, therefore $M_K\simeq\mathbb G_{m,K}^{s}$ for some $s\in \mathbb Z^{\ge1}$.
This implies $H^1(K,M)\simeq H^1(K,M_K)\simeq H^1(K,\mathbb G_{m,K}^{s})=0$ by Hilbert's theorem 90.
By Hochschild-Serre's spectral sequence, we have:
$$
0\longrightarrow H^1(G,M(K))\longrightarrow H^1(F,M)\longrightarrow H^1(K,M),
$$
this induces an isomorphism $H^1(G,M(K))\simeq H^1(F,M)$.
\end{proof}
\begin{prop}\label{firstterm}
Let $W=R_{K/F}^{(1)}(\mathbb G_{m,K})$ be a norm one torus of the finite Galois extension $K/F$ with $G=\mathrm{Gal}(K/k)$ and let
$T$ be the torus in the flasque resolution (\ref{seq:flasque}) of $W$. Then we have $H^1(F,T^\circ)\simeq H^3(G,K^*)$.
\end{prop}
\begin{proof}
By Lemma \ref{lemma2}, we have an isomorphism $H^1(F,T^\circ)\simeq H^1(G,T^\circ(K))$.
The sequence (\ref{seq:flasque}) derives the following exact sequence
$$
1\longrightarrow W^\circ(K)\longrightarrow P^\circ(K)\longrightarrow T^\circ(K)\longrightarrow 1.
$$
One has $H^i(G,P^\circ(K))=0$ for $i>0$ since $$R_{K/F}(\mathbb G_{m,K})(K)\simeq \mathbb G_{m,K}(K\otimes_F K)\simeq K^*\otimes_{\mathbb Z}\mathbb Z[G]$$ is an induced module.
Hence we get an isomorphism $ H^1(G,T^0(K))\simeq H^2(G,W^\circ(K))$.
On the other hand, the following sequence is exact
$$
1\longrightarrow\mathbb G_{m}(K)\longrightarrow R_{K/F}(\mathbb G_{m,K})(K)\longrightarrow W^\circ(K)\longrightarrow 1
$$
by Hilbert's theorem 90.
The proof then follows from $$H^1(F,T^\circ)\simeq H^1(G,T^\circ(K))\simeq H^2(G,W^\circ(K))\simeq H^3(G,K^*).\qedhere$$
\end{proof}
\begin{rem}\label{rem: K}
If $K/F$ is a Galois extension of local fields, then $H^3(G,K^*)=0$ by \cite[Corollary 7.2.2]{neukirch2013}.
If $K/F$ is a Galois extension of global fields, we consider the exact sequence
$$
1\longrightarrow K^*\longrightarrow I_K\longrightarrow C_K\longrightarrow 1
$$
where $I_K$ is the idèle group and $C_K$ is the idèle class group.
This induces the following commutative diagram with the first row exact \cite[Chapter VIII]{neukirch2013}
\[
\begin{tikzcd}
H^2(G,I_K)\arrow[r,"\psi"]\arrow[d,"\simeq"]&H^2(G,C_K)\arrow[r]\arrow[dd,"\simeq","inv_{K|F}"']&H^3(G,K^*)\arrow[r]& H^3(G,I_K)\arrow[d, equal]\\
\oplus_{\mathfrak p}H^2(G_{\mathfrak p},K^*_{\mathfrak p})\arrow[d,"\simeq","inv_{\mathfrak p}"']&&& \oplus_{\mathfrak p}H^3(G_{\mathfrak p},K_{\mathfrak p}^*)\arrow[d, equal]\\
\oplus_{\mathfrak p}{1\over [K_{\mathfrak p}:F_{\mathfrak p}] }\mathbb Z/\mathbb Z\arrow[r]&{1\over [K:F] }\mathbb Z/\mathbb Z&& 0
\end{tikzcd}
\]
where the sum run all prime $\mathfrak p$ over $F$ and $G_{\mathfrak p}$ is the decomposition subgroup of $G$ with some prime $\mathfrak B$ above $\mathfrak p$.
Therefore $H^3(G,K^*)\simeq \mathrm{coker}\ \psi$.
In general $\psi$ is not surjective, such as $F=\Q$ and $K=\Q(\sqrt{-1},\sqrt{17})$.
\end{rem}
The flasque resolution (\ref{seq:flasque}) of $W=R^{(1)}_{K/F}(\mathbb G_{m,K})$ can derive the following exact sequence of dual $\Gamma$-modules
\begin{equation}
1\longrightarrow \hat{W}_{sep}\longrightarrow \hat{P}_{sep}\longrightarrow \hat{T}_{sep}\longrightarrow 1.\label{exactseq}
\end{equation}
The part $H^0(F,S^2(\hat{T}_{sep}))/\Dec$ in Theorem \ref{mainlem} is actually only dependent on the group $\mathrm{Gal}(K/F)$.
To prove this, we need several lemmas.
\begin{lem}
Let $N$ be the kernel of the natural surjective homomorphism $S^2(\hat{P}_{sep})\to S^2(\hat{T}_{sep})$, then the sequence
$$
0\rightarrow\mathbb\wedge^2\hat{W}_{sep} \stackrel{f}{\rightarrow} \hat{W}_{sep}\otimes\hat{P}_{sep}\stackrel{f'}{\rightarrow} N\rightarrow 0
$$
is exact where $f(a\wedge b)=a\otimes b-b\otimes a$ and $f'$ is induced by $\hat{W}_{sep}\otimes\hat{P}_{sep}\longrightarrow S^2(\hat{P}_{sep})$.
\end{lem}
\begin{proof}
The exactness at $\wedge^2\hat{W}_{sep}$ and $\hat{W}_{sep}\otimes\hat{P}_{sep}$ is obvious. We only need to show that $f'$ is surjective.
Assume that $\hat{T}_{sep}=<\beta_1,\cdots,\beta_t>$, and $\beta_i$ is the image of $\alpha_i\in\hat{P}_{sep}$, so $\alpha_1,\cdots,\alpha_t$ is $\mathbb Z$-linear independent in $\hat{P}_{sep}$.
Consider the following commutative diagram:
\[
\begin{tikzcd}
\hat{P}_{sep}\otimes\hat{P}_{sep}\arrow[r,"g"]\arrow[d,"h"]&\hat{T}_{sep}\otimes\hat{T}_{sep}\arrow[d,"h'"]\\
S^2(\hat{P}_{sep})\arrow[r,"g'"]&S^2(\hat{T}_{sep})
\end{tikzcd}
\]
where $h$, $h'$ are the natural homomorphisms and $g$, $g'$ are induced by the exact sequence (\ref{exactseq}).
It is not difficult to see that $g$, $g'$, $h$, $h'$ are surjective.
If $\alpha\in \hat{P}_{sep}\otimes\hat{P}_{sep}$ such that $g'\circ h(\alpha)=0\in S^2(\hat{T}_{sep}),$ this implies $$ g(\alpha)=\sum_{i<j}b_{ij}(\beta_i\otimes\beta_j-\beta_j\otimes\beta_i)\in \hat{T}_{sep}\otimes\hat{T}_{sep}, b_{ij}\in \mathbb Z.$$
Hence $$\alpha=\sum_{i<j}b_{ij}((\alpha_i+\omega_i)\otimes(\alpha_j+\omega_j)-(\alpha_j+\omega_j')\otimes(\alpha_i+\omega_i')), \ b_{ij}\in \mathbb Z,$$ where $\omega_i, \omega_i', \omega_j, \omega_j'\in \hat{W}_{sep}$. Then we can deduce $$\alpha=\alpha'+\alpha'',\alpha'\in \{a\otimes b-b\otimes a|a,b\in \hat{P}_{sep}\},\alpha''\in \hat{W}_{sep}\otimes\hat{P}_{sep}+\hat{P}_{sep}\otimes\hat{W}_{sep}.$$ Therefore, the image of $\alpha$ in $S^2(\hat{P}_{sep})$ comes from $\hat{W}_{sep}\otimes\hat{P}_{sep}$, thus $f'$ is surjective.
\end{proof}
Because $\Gamma_K$ acts trivially on $S^2(\hat{P}_{sep})$, we can only consider the $G$-action: $S^2(\hat{T}_{sep})^{\Gamma} = S^2(\hat{T}_{sep})^G$.
The short exact sequence
$
0\longrightarrow N\longrightarrow S^2(\hat{P}_{sep}) \stackrel{g'}{\longrightarrow} S^2(\hat{T}_{sep})\longrightarrow 0
$ induce an exact sequence:
\begin{equation}
S^2(\hat{P}_{sep})^G \stackrel{g'}{\longrightarrow} S^2(\hat{T}_{sep})^G\stackrel{\delta}{\longrightarrow} H^1(G,N)\longrightarrow H^1(G,S^2(\hat{P}_{sep})).\label{dec}
\end{equation}
\begin{lem}\label{duichen}
If $G$ is a finite group and $p$ is an odd prime number, then $$H^i(G,S^2(\mathbb Z[G]))\{p\}=H^i(G,\wedge^2\mathbb Z[G])\{p\}= 0$$ for $i\ge1$. Furthermore, if $G$ has odd order, then $$H^i(G,S^2(\mathbb Z[G]))=H^i(G,\wedge^2\mathbb Z[G])= 0$$ for $i\ge1$.
\end{lem}
\begin{proof}
We have a commutative diagram with the row exact:
\[
\begin{tikzcd}
&&S^2(\mathbb Z[G])\arrow[r,equal]\arrow[d,"i_1"]&S^2(\mathbb Z[G])\arrow[d,"\times 2"]&\\
0\arrow[r]&\wedge^2\mathbb Z[G]\arrow[r,"i_2"]\arrow[d,"\times 2"]&\mathbb Z[G]^{\otimes 2}\arrow[r,"p_1"]\arrow[d,"p_2"]&S^2(\mathbb Z[G])\arrow[r]&0\\
&\wedge^2\mathbb Z[G]\arrow[r,equal]&\wedge^2\mathbb Z[G]&&
\end{tikzcd}
\]
where $i_1(a\cdot b)=a\otimes b+b\otimes a$, $i_2(a\wedge b)=a\otimes b-b\otimes a$, $p_1(a\otimes b)=a\cdot b$ and $p_2(a\otimes b)=a\wedge b$.
Therefore, this induces a commutative diagram with the row exact for $i\ge1$:
\[
\begin{tikzcd}
&H^i(G,S^2(\mathbb Z[G]))\arrow[r,equal]\arrow[d,"i_1^*"]&H^i(G,S^2(\mathbb Z[G]))\arrow[d,"\times2"]\\
H^i(G,\wedge^2\mathbb Z[G])\arrow[r,"i_2^*"]\arrow[d,"\times2"]&H^i(G,\mathbb Z[G]^{\otimes 2})\arrow[r,"p_1^*"]\arrow[d,"p_2^*"]&H^i(G,S^2(\mathbb Z[G]))\\
H^i(G,\wedge^2\mathbb Z[G])\arrow[r,equal]&H^i(G,\wedge^2\mathbb Z[G])&
\end{tikzcd}
\]
Since $H^i(G,\mathbb Z[G]^{\otimes 2})=0$, one obtains $2\cdot H^i(G,S^2(\mathbb Z[G])) =2\cdot H^i(G,\wedge^2\mathbb Z[G])= 0.$
This implies $H^i(G,S^2(\mathbb Z[G]))\{p\}=H^i(G,\wedge^2\mathbb Z[G])\{p\}= 0$ for $i\ge 1$ and odd prime $p$. If $G$ has odd order, then $H^i(G,S^2(\mathbb Z[G]))\{p\}$ and $H^i(G,\wedge^2\mathbb Z[G])\{p\}$ have odd order, the proof then follows.
\end{proof}
\begin{lem}\label{invertibletorus}
Let $P$ be the quasisplit torus, then $S^2(\hat{P}_{sep})^G/\Dec=0$.
\end{lem}
\begin{proof}
By the definition of $\Dec$, we can deduce (see \cite[A-II]{m2013}):
\begin{align*}S^2(A\otimes B)^\Gamma&\simeq S^2(A)^\Gamma\oplus (A\otimes B)^\Gamma\oplus S^2(B)^\Gamma,\\
\Dec(A\oplus B)&\simeq \Dec(A)\oplus \Dec(A,B)\oplus \Dec(B),
\end{align*}
where $\Dec(A,B)$ is the subgroup of $(A\otimes B)^{\Gamma}$ generated by elements $\Tr(a\otimes b)=\sum_{i}\sigma_i a\otimes\sigma_i b$ for all open subgroups $\Gamma'$ of $\Gamma$ and all $a\in A^{\Gamma'}$ and $b\in B^{\Gamma'}$, where $\{\gamma_i\}$ is a representative of $\Gamma/\Gamma'$. Assume that $A$ and $B$ are permutation modules, it is clear that $(A\otimes B)^{\Gamma}=\Dec(A, B)$.
So we can reduce to the case $\hat{P}_{sep}=\mathbb Z[\Gamma/\Gamma_L]$ by induction, where $L/F$ is a finite field extension and $\Gamma_L$ is the absolute Galois group of $L$.
We denote $\bar \sigma_1\cdot \bar \sigma_2$ to be the image of $\bar \sigma_1\otimes \bar \sigma_2$ in $S^2(\mathbb Z[\Gamma/\Gamma_L])$, where $\bar \sigma_i \in \Gamma/\Gamma_L$ and $i=1,2$. Let $<\bar \sigma_1\cdot \bar \sigma_2>$ be the orbit of $\Gamma$-action on $\bar \sigma_1\cdot \bar \sigma_2$. Let $N_{\bar \sigma}:=\sum_{a\in <\bar e\cdot \bar \sigma>}a$. It is clear that $S^2(\mathbb Z[\Gamma/\Gamma_L])^{\Gamma}$ is generated by all $N_{\bar \sigma}, \bar \sigma\in \Gamma/\Gamma_L$. It suffices to show $N_{\bar \sigma}\in \Dec(\mathbb Z[\Gamma/\Gamma_L])$.
Let $\Gamma'=\Gamma_L\cap \Gamma_{\sigma(L)}$ and $\Gamma_{\bar \sigma}$ the stabilizer of $\bar e\cdot \bar \sigma $, which is an open subgroup of $\Gamma$ containing $\Gamma'$.
\begin{lem}\label{lem:stabilizer}
Either $\Gamma_{\bar \sigma}=\Gamma'$ or $\Gamma_{\bar \sigma}=\Gamma'\cup \sigma'\Gamma' $, $\sigma'\Gamma_L= \sigma\Gamma_L$ and $\sigma'\not\in \Gamma'$.
\end{lem}
\begin{proof} We only need to prove the case $\Gamma_{\bar \sigma}\neq\Gamma'$.
Assume that $\sigma',\sigma''\in \Gamma_{\bar \sigma}\setminus\Gamma'$. So ${\sigma'\bar\sigma}={\sigma''\bar\sigma}=\bar e$ and $\bar \sigma'=\bar \sigma''=\bar \sigma$.
Then $\sigma'^{-1}\sigma''\in \Gamma_{\sigma(L)}\cap \Gamma_L=\Gamma'$, hence $\sigma'\Gamma'=\sigma''\Gamma'$.\qedhere
\end{proof}
Let $\bar \sigma\in \Gamma/ \Gamma_L$ and $\sigma_1,\cdots,\sigma_m$ representatives of $\Gamma/\Gamma'$.
We have \begin{align*}
\Qtr_{\Gamma'}(\bar e+\bar \sigma)&=\sum_{i<j}\sigma_i(\bar e+\bar \sigma)\cdot\sigma_i(\bar e+\bar \sigma)\\
&=\sum_{i<j}\bar \sigma_i\cdot \bar \sigma_j+\sum_{i<j}{\sigma_i\bar \sigma}\cdot {\sigma_j\bar \sigma}+\sum_{i<j}\bar \sigma_i\cdot {\sigma_j\bar\sigma}+\sum_{i>j}\bar \sigma_i\cdot {\sigma_j\bar\sigma}\\
&=\Qtr_{\Gamma'}(\bar e)+\Qtr_{\Gamma'}(\bar \sigma)+\mathrm {Tr}(\bar e)\cdot \mathrm {Tr}(\bar e)-\sum_{i}\bar\sigma_i\cdot\sigma_i\bar\sigma.
\end{align*}
Therefore, $\sum_{i}\bar\sigma_i\cdot\sigma_i\bar\sigma\in \Dec(\Bbb Z[\Gamma/\Gamma_L])$. If $\Gamma_{\bar \sigma}=\Gamma'$, then
$N_{\bar \sigma}=\sum_{i}\bar\sigma_i\cdot\sigma_i\bar\sigma\in \Dec(\mathbb Z[\Gamma/\Gamma_L])$.
Suppose $[\Gamma_{\bar \sigma}:\Gamma']=2$. Let $\sigma'\in \Gamma_{\bar \sigma}\setminus \Gamma'$ and $\sigma_1,\cdots,\sigma_m$ representatives of $\Gamma/\Gamma_{\bar \sigma}$, hence $\sigma_1,\cdots,\sigma_m,\sigma_1\sigma',\cdots,\sigma_m\sigma'$ are representatives of $\Gamma/\Gamma_{\bar \sigma}$. Since $\bar \sigma'=\bar\sigma\in \Gamma/\Gamma_L$ by Lemma~\ref{lem:stabilizer}, one has $\bar e+\bar \sigma'=\bar e+\bar \sigma\in \mathbb Z[\Gamma/\Gamma_L]^{\Gamma_{\bar\sigma}}$, hence
\begin{align*}
\Qtr_{\Gamma_{\bar \sigma}}(\bar e+\bar \sigma')&=\sum_{i<j}\sigma_i(\bar e+\bar \sigma')\cdot \sigma_j(\bar e+\bar \sigma')\\
&=\sum_{i<j}\bar \sigma_i\cdot \bar \sigma_j+\sum_{i<j}{\sigma_i\bar \sigma'}\cdot {\sigma_j\bar \sigma'}+\sum_{i<j}\bar \sigma_i\cdot {\sigma_j\bar\sigma'}+\sum_{i>j}\bar \sigma_i\cdot {\sigma_j\bar\sigma'}\\
&=\Qtr_{\Gamma'}(\bar e)-\sum_{i}\bar\sigma_i\cdot{\sigma_i\bar \sigma},
\end{align*}
which implies $N_{\bar \sigma}=\sum_{i}\bar\sigma_i\cdot{\sigma_i\bar \sigma}\in \Dec(\mathbb Z[\Gamma/\Gamma_L])$, the proof of Lemma \ref{lem:stabilizer} then follows.
\end{proof}
\begin{rem}
Let $S=\mathrm{Spec}\ F$ be the trivial torus.
Then $
1\longrightarrow P\longrightarrow P\longrightarrow S\longrightarrow 1$
is a flasque resolution of $S$. Let $p$ be an odd prime. Therefore
$$
0=\bar{H}^3_{nr}(F(S),\mathbb Q/\mathbb Z(2))\{p\}\simeq H^1(F,P^\circ)\{p\}\oplus (H^0(F,S^2(\hat{P}_{sep}))/\Dec)\{p\}
$$
by Theorem \ref{mainlem}.
Hence $(S^2(\hat{P}_{sep})^G/\Dec)\{p\}=0$.
However, the $2$-primary part cannot be determined by this method.
\end{rem}
\begin{lem}\label{commutator}
If $\#[G,G]$ is a power of $2$, then for an odd prime number $p$
$$(H^0(F,S^2(\hat{T}_{sep}))/\Dec)\{p\}\simeq (H^1(G,N)/\delta((\hat{T}_{sep}^G)^{\otimes2}))\{p\},$$ where $N,\delta$ are defined in (\ref{dec}).
\end{lem}
\begin{proof}
Let $H$ be a subgroup of $G$, $G^{ab}=G/[G,G]$ and $H^{ab}=H/[H,H]$.
Since $\#[G,G]$ is a power of $2$, the exact sequence $
0\longrightarrow\mathbb Z\stackrel{N}{\longrightarrow} \mathbb Z[G]\longrightarrow\hat{W}_{sep}\longrightarrow 0
$ derives the following commutative diagram.
\[
\begin{tikzcd}
H^1(G,\hat{W}_{sep})\arrow[r,"\simeq"]\arrow[d,"j"]&H^2(G,\mathbb Z)\arrow[r,"\simeq"]\arrow[d]&H^1(G,\mathbb Q/\mathbb Z)\arrow[d]\arrow[r,"\simeq"]&\check{G^{ab}}\arrow[d,"j'"]\\
H^1(H,\hat{W}_{sep})\arrow[r,"\simeq"]&H^2(H,\mathbb Z)\arrow[r,"\simeq"]&H^1(H,\mathbb Q/\mathbb Z)\arrow[r,"\simeq"]&\check{H^{ab}}
\end{tikzcd}
\]
Therefore, the cokernel $\mathrm{Coker}(j)\simeq \mathrm{Coker}(j')$ is a $2$-group.
The sequence (\ref{exactseq}) derives the following commutative diagram with the row exact
\[
\begin{tikzcd}
0\arrow[r]&\hat{W}_{sep}^G\arrow[r]\arrow[d]&\hat{P}_{sep}^G\arrow[r]\arrow[d,"r"]&\hat{T}_{sep}^G\arrow[d,"r'"]\arrow[r]&H^1(G,\hat{W}_{sep})\arrow[r]\arrow[d,"j"]&H^1(G,\hat{P}_{sep})=0\\
0\arrow[r]&\hat{W}_{sep}^H\arrow[r]&\hat{P}_{sep}^H\arrow[r]&\hat{T}_{sep}^H\arrow[r]&H^1(H,\hat{W}_{sep})\arrow[r]&H^1(H,\hat{P}_{sep})=0.
\end{tikzcd}
\]
Since the $\mathrm{Coker}(j)$ is a $2$-group and the sequence $ \mathrm{Coker}(r)\to \mathrm{Coker}(r') \to\mathrm{Coker} (j)\to 0$ is exact,
it shows that
$2^m\hat{T}_{sep}^H\subset \mathrm{im}(\hat{P}_{sep}^H)+\hat{T}_{sep}^G\subset\hat{T}_{sep}^H$ for some integer $m$.
Let $\{\gamma_i\}$ be a representative for the left cosets of $H$ in $G$.
If $a\in \hat{T}_{sep}^H$, then $2^ma=b+c\in \hat{T}_{sep}^H, b\in \mathrm{im}(\hat{P}_{sep}^H),c\in\hat{T}_{sep}^G$. By the commutative diagram
\[
\begin{tikzcd}
\hat{P}_{sep}^H\arrow[r]\arrow[d,"\Qtr"]&\hat{T}_{sep}^H\arrow[d,"\Qtr"]\\
S^2(\hat{P}_{sep})^G\arrow[r,"g"]&S^2(\hat{T}_{sep})^G,
\end{tikzcd}
\]
one obtains
\begin{align*}
\Qtr(2^m a)&=\sum_{i<j}\gamma_i(b+c)\cdot\gamma_j(b+c)\\
&=\sum_{i<j}\gamma_i b\cdot\gamma_j b+\sum_{i<j}c\cdot c+lc\cdot\sum_{i}\gamma_ib,
\end{align*}
where $l=[G:H]-1$.
So $$\Qtr(2^m\hat{T}_{sep}^H)=2^{2m}\Qtr(\hat{T}_{sep}^H)\subset g(\Dec(\hat{P}_{sep}))+(\hat{T}_{sep}^G)^{2}.$$
Let $H$ run through all subgroups of $G$, since $G$ is finite, there exists a large enough integer $m'$ such that $2^{m'}\Dec(\hat{T}_{sep})\subset g(\Dec(\hat{P}_{sep}))+(\hat{T}_{sep}^G)^{2}$, hence
\begin{align*}
(S^2(\hat{T}_{sep})^G/2^{2m'}\Dec(\hat{T}_{sep}))\{p\}&= [S^2(\hat{T}_{sep})^G/(g'(\Dec(\hat{P}_{sep}))+(\hat{T}_{sep}^G)^{2})]\{p\}\\
&= (S^2(\hat{T}_{sep})^G/\Dec(\hat{T}_{sep}))\{p\}.
\end{align*}
On the other hand, by Lemma \ref{invertibletorus} and Lemma \ref{duichen},
the sequence \eqref{dec} derives the isomorphism
$(S^2(\hat{T}_{sep})^G/g(\Dec(\hat{P}_{sep})))\{p\}\stackrel{\delta}{\longrightarrow} H^1(G,N)\{p\}
$
thus $$(H^0(F,S^2(\hat{T}_{sep}))/\Dec(\hat{T}_{sep}))\{p\}\simeq (H^1(G,N)/\delta((\hat{T}_{sep}^G)^{2}))\{p\}.\qedhere$$
\end{proof}
\begin{rem}
Assume that $G$ is abelian.
Then $\Dec(\hat{T}_{sep})= g'(\Dec(\hat{P}_{sep}))+(\hat{T}_{sep}^G)^{2}$
from the proof above. By Lemma \ref{invertibletorus} and the sequence (\ref{dec}), we deduce the following isomorphism
\begin{equation}
S^2(\hat{T}_{sep})^G/\Dec\stackrel{\simeq}{\longrightarrow} H^1(G,N)/\delta((\hat{T}_{sep}^G)^{\otimes2}).\label{abelian}
\end{equation}
\end{rem}
Let $N'$ be the kernel of $\hat{P}_{sep}\otimes\hat{T}_{sep}\to S^2(\hat{T}_{sep})$ and $N''$ the kernel of $\hat{P}_{sep}\otimes\hat{P}_{sep}\to S^2(\hat{T}_{sep})$.
Then we have the following commutative diagram with the exact rows:
\[
\begin{tikzcd}
0\arrow[r]&N\arrow[r]& S^2(\hat{P}_{sep})\arrow[r]& S^2(\hat{T}_{sep})\arrow[r]&0\\
0\arrow[r]&N''\arrow[r]\arrow[u]\arrow[d]& \hat{P}_{sep}\otimes\hat{P}_{sep}\arrow[r]\arrow[u]\arrow[d]& S^2(\hat{T}_{sep})\arrow[r]\arrow[u,equal]\arrow[d,equal]&0\\
0\arrow[r]&N'\arrow[r]& \hat{P}_{sep}\otimes\hat{T}_{sep}\arrow[r]& S^2(\hat{T}_{sep})\arrow[r]&0
\end{tikzcd}
\]
By construction $\hat{P}_{sep}\simeq \mathbb{Z}[G]^{\oplus r}$, for $i\ge 1$ and an odd prime $p$, $\hat{H}^i(G,\hat{P}_{sep}\otimes\hat{P}_{sep})=\hat{H}^i(G, \hat{P}_{sep}\otimes\hat{T}_{sep})=0$ by Shapiro's Lemma, $\hat{H}^i(G,S^2(\hat{P}_{sep}))\{p\}=0$ by Lemma \ref{duichen}.
Therefore the following diagram is commutative
$$
\begin{tikzcd}
&\hat{H}^1(G,N)\{p\}\\
\hat{H}^0(G,S^2(\hat{T}_{sep}))\{p\}\arrow[dr, "\delta'", "\simeq"']\arrow[r, "\delta''", "\simeq"']\arrow[ur,"\delta", "\simeq"']&\hat{H}^1(G,N'')\{p\}\arrow[u, "\simeq"]\arrow[d, "\simeq"]\\
&\hat{H}^1(G,N')\{p\}
\end{tikzcd}
$$
and we deduce
\begin{align*}
(H^1(G,N)/\delta((\hat{T}_{sep}^G)^{\otimes2}))\{p\}&\simeq(H^1(G,N')/\delta'((\hat{T}_{sep}^G)^{\otimes2}))\{p\}\\
&\simeq(H^1(G,N'')/\delta''((\hat{T}_{sep}^G)^{\otimes2}))\{p\}.
\end{align*}
\begin{prop}\label{prop:cup}
Let $W=R_{K/F}^{(1)}(\mathbb G_{m,K})$ and $K/F$ the finite Galois extension with $G=\mathrm{Gal}(K/F)$. Let
$$
1\longrightarrow T\longrightarrow P\longrightarrow W\longrightarrow 1
$$
be the flasque resolution (\ref{seq:flasque}) of $W$. Suppose the commutator $[G,G]$ is a $2$-group. Then
$$
(H^0(F,S^2(\hat{T}_{sep}))/\Dec)\{p\}\simeq \mathrm{Coker}(H^2(G,\mathbb Z)\times H^2(G,\mathbb Z)\stackrel{\cup}{\longrightarrow}H^4(G,\mathbb Z))\{p\}.
$$
\end{prop}
\begin{proof}
By \cite[Chapter I, \S 4]{neukirch2013}, we have the following commutative diagram:
$$
\begin{tikzcd}
\hat{H}^0(G,\hat{T}_{sep})\arrow[d,"\simeq"]\arrow[phantom, r, description, "\times"]&\hat{H}^0(G,\hat{T}_{sep})\arrow[d,equal]\arrow[r,"\cup"]&\hat{H}^{0}(G,S^2(\hat{T}_{sep}))\arrow[d, "\delta'"]\arrow[dr, "\delta''"]\arrow[drr, "\delta"]\\
\hat{H}^{1}(G,\hat{W}_{sep})\arrow[d,equal]\arrow[phantom, r, description, "\times"]&\hat{H}^0(G,\hat{T}_{sep})\arrow[d,"\simeq"]\arrow[r,"\cup"]&\hat{H}^{1}(G,N')&\hat{H}^{1}(G,N'')\arrow[l,"\simeq"]\arrow[r,"\simeq"]&\hat{H}^{1}(G,N)\arrow[d,"\simeq"]\\
\hat{H}^1(G,\hat{W}_{sep})\arrow[d,"\simeq"]\arrow[phantom, r, description, "\times"]&\hat{H}^1(G,\hat{W}_{sep})\arrow[d,equal]\arrow[rrr,"\cup"]&&&\hat{H}^{2}(G,\wedge^2\hat{W}_{sep})\arrow[d,""]\\
\hat{H}^2(G,\mathbb Z)\arrow[d,equal]\arrow[phantom, r, description, "\times"]&\hat{H}^1(G,\hat{W}_{sep})\arrow[d,"\simeq"]\arrow[rrr,"\cup"]&&&\hat{H}^{3}(G,\hat{W}_{sep})\arrow[d,"\simeq"]\\
\hat{H}^2(G,\mathbb Z)\arrow[phantom, r, description, "\times"]&\hat{H}^2(G,\mathbb Z)\arrow[rrr,"\cup"]&&&\hat{H}^{4}(G,\mathbb Z)
\end{tikzcd}
$$
a. The first square is induced by the exact sequences:
$
1\longrightarrow \hat{W}_{sep}\longrightarrow \hat{P}_{sep}\longrightarrow \hat{T}_{sep}\longrightarrow 1
$
and
$
0\longrightarrow N'\longrightarrow\hat{P}_{sep}\otimes\hat{T}_{sep}\longrightarrow S^2(\hat{T}_{sep})\longrightarrow0
$.
Note that $\hat{H}^0(G,\hat{P}_{sep})=\hat{H}^1(G,\hat{P}_{sep})=0$
thus the vertical arrows at left is an isomorphism.
b. The second square is induced by the exact sequences:
$
1\longrightarrow \hat{W}_{sep}\longrightarrow \hat{P}_{sep}\longrightarrow \hat{T}_{sep}\longrightarrow 1
$
and
$
0\longrightarrow\mathbb\wedge^2\hat{W}_{sep} \stackrel{f}{\longrightarrow} \hat{W}_{sep}\otimes\hat{P}_{sep}\longrightarrow N\longrightarrow 0
$.
Note that $\hat{H}^1(G,\hat{W}_{sep}\otimes\hat{P}_{sep})=\hat{H}^2(G,\hat{W}_{sep}\otimes\hat{P}_{sep})=0$.
c. The third square is induced by the exact sequences:
$
0\longrightarrow\mathbb Z\stackrel{N_G}{\longrightarrow} \mathbb Z[G]\longrightarrow\hat{W}_{sep}\longrightarrow 0
$
and
$
0\longrightarrow\hat{W}_{sep} \stackrel{\varphi}{\longrightarrow}\wedge^2\mathbb Z[G] \longrightarrow \wedge^2\hat{W}_{sep} \longrightarrow 0
$, where $\wedge^2\mathbb Z[G] \longrightarrow \wedge^2\hat{W}_{sep}$ are the natural surjective homomorphism.
We now define the hommomorphism $\varphi$ and prove the exactness at $\wedge^2\mathbb Z[G]$.
We write $\bar{g}\in\hat{W}_{sep}$ as the image of $g\in \Z[G]$ and $\bar{g}\wedge\bar{h}\in\wedge^2\hat{W}_{sep}$ as the image of $g\wedge h\in \wedge^2\Z[G]$.
Let $\bar{b}\in \hat{W}_{sep}$ and $b$ be an arbitrary lifting of $\bar{b}$.
We define $\varphi(\bar{b})=b\wedge N_G(1)$, and it is easy to see that the morphism $\varphi$ does not depend on the choice of the lifting.
Assume that $G=\{g_1,\cdots,g_n\}$, therefore
$\wedge^2\hat{W}_{sep}$ has a $\mathbb Z$-basis $\{\bar{g}_i\wedge \bar{g}_j\}_{1\leq i<j\leq n-1}$, since $\bar{g}_n=-\sum_{i=1}^{n-1}\bar{g}_i$ in $\hat{W}_{sep}$.
Suppose $a=\sum_{1\leq i< j\leq n}a_{ij}g_i\wedge g_j\in \wedge^2\mathbb Z[G]$ but $\bar{a}=0\in \wedge^2\hat{W}_{sep}$.
We obtain $\bar{a}=\sum_{1\leq i< j\leq n-1}(a_{ij}-a_{in}+a_{jn})\bar{g}_i\wedge \bar{g}_j$, since $\bar{g}_n=-\sum_{i=1}^{n-1}\bar{g}_i$ in $\hat{W}_{sep}$, therefore $a_{ij}=a_{in}-a_{jn}$, and $a=\sum_{1\leq i< j\leq n}(a_{in}-a_{jn})g_i\wedge g_j$ in $\wedge ^2\mathbb Z[G]$.
Let $b=\sum_{i=1}^{n}a_{in}g_i\in \Z[G]$, then it is clear that $\varphi(\bar{b})=(\sum_{i=1}^{n}a_{in}g_i)\wedge N_G(1)=a$ and this proves the exactness.
d. The forth square is induced by the exact sequences:
$
0\rightarrow\mathbb Z\stackrel{N}{\rightarrow} \mathbb Z[G]\rightarrow\hat{W}_{sep}\rightarrow 0
$.
Note that
$\delta$, $\delta'$, $\delta''$ and the right vertical arrow in the third square are isomorphisms when we take $p$-primary component.
By Lemma \ref{commutator} and the above commutative diagram, one obtains $$(H^0(F,S^2(\hat{T}_{sep}))/\Dec)\{p\}\simeq \mathrm{Coker}(H^2(G,\mathbb Z)\times H^2(G,\mathbb Z)\stackrel{\cup}{\longrightarrow}H^4(G,\mathbb Z))\{p\}.\qedhere$$
\end{proof}
Combining Theorem \ref{mainlem} with Proposition \ref{prop:cup}, we immediately have the following result.
\begin{cor}
Let $K/F$ be a Galois extension with $\mathrm{Gal}(K/F)=G$ and $W=R_{K/F}^{(1)}(\mathbb G_{m,K})$. Suppose $\#[G,G]$ is a power of $2$. Then for an odd prime $p$,
$$
\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))\{p\}\simeq H^3(G,K^*)\{p\} \oplus\mathrm{Coker}(H^2(G,\mathbb Z)\times H^2(G,\mathbb Z)\stackrel{\cup}{\longrightarrow}H^4(G,\mathbb Z))\{p\}.
$$
\end{cor}
If $G$ is abelian, then $\#[G,G]=1$,
this satisfies the condition of the above Corollary,
and we can give a more precise expression for this cokernel by induction and Hochschild-Serre spectral sequence.
If $G=G'\times G''$ and acts trivially on $\mathbb Z$, then the Hochschild-Serre spectral sequence
$$
E_2^{p,q} = H^p(G',H^q(G'',\mathbb Z))\implies H^{p+q}(G,\mathbb Z)
$$
degenerates, and there is the decomposition by the main theorem of \cite{Jannsen1990TheSO}
$$
H^{n}(G,\mathbb Z)\simeq \oplus_{p+q=n}H^p(G',H^q(G'',\mathbb Z)).
$$
Moreover, the following diagram is commutative:
\begin{equation}
\begin{tikzcd}
H^p(G',\mathbb Z)\arrow[d,"\simeq"]\arrow[phantom, r, description, "\times"]&H^q(G'',\mathbb Z)\arrow[d,"\simeq"]\arrow[r,"\cup"]&H^{p+ q}(G,\mathbb Z)\arrow[d,"projection"]\\
E_2^{p,0}\arrow[phantom, r, description, "\times"]&E_2^{0,q}\arrow[r,"\cup"]&E_2^{p,q}
\end{tikzcd}\label{cup}
\end{equation}
Since $G$ is finite and abelian, we can represent $G$ as the direct sum of cyclic groups: $G=\oplus_{i=1}^mC_i$ such that $\#C_{i-1}|\#C_i$ ($1<i\leq m$). We set $G_i=\oplus_{j=1}^{i}C_j$, $G=G_m$, $G_{i+1}=G_i\oplus C_{i+1}$.
\begin{lem}
Notations as above, let $d_{i}=(m-i)(m-i-1)/2$, then
$$\mathrm{Coker}(H^2(G,\mathbb Z)\times H^2(G,\mathbb Z)\stackrel{\cup}{\longrightarrow}H^4(G,\mathbb Z))\simeq \oplus_{i=1}^{m-2}G_i^{m-i-1} \simeq \oplus_{i=1}^{m-2}C_i^{d_{i}}.
$$
\end{lem}
\begin{proof}
Firstly, $H^2(G,\mathbb Z)\simeq E_2^{2,0}\oplus E_2^{0,2}$ by the Hochschild-Serre spectral sequence with $G=G_{m-1}\times C_m$ and $E_2^{r,s}=0$ when $s$ is even.
Similarly, $H^4(G,\mathbb Z)\simeq E_2^{4,0}\oplus E_2^{2,2}\oplus E_2^{0,4}$. Therefore, from \eqref{cup}, we can divide into three steps:
a) $\ \mathrm{Coker}(E_2^{0,2}\times E_2^{0,2}\stackrel{\cup}{\longrightarrow}E_2^{0,4})$:
This is trivial by \cite[Proposition 1.6.12]{neukirch2013} since $C_m$ is cyclic .
b) $ \mathrm{Coker}(E_2^{2,0}\times E_2^{0,2}\stackrel{\cup}{\longrightarrow}E_2^{2,2})$:
Assume that $C_m=\mathbb Z/n_m\mathbb Z$. Since $E_2^{0,2}\simeq C_m=\mathbb Z/n_m\mathbb Z$, the cokernel of this cup product is equal to the cokernel of the map $H^2(G_{m-1},\mathbb Z)\stackrel{}{\longrightarrow}H^2(G_{m-1},\mathbb Z/n_m\mathbb Z)$ induced by $\mathbb Z\to \mathbb Z/n_m\mathbb Z$.
The exact sequence
$$
0\longrightarrow \Z \stackrel{\times n_m}{\longrightarrow} \Z\longrightarrow \Z/m\longrightarrow 0
$$ derives the exact sequence
$$
H^2(G_{m-1},\mathbb Z)\longrightarrow H^2(G_{m-1},\mathbb Z/n_m\mathbb Z)\longrightarrow H^3(G_{m-1},\mathbb Z)\stackrel{\times n_m}{\longrightarrow} H^3(G_{m-1},\mathbb Z).
$$
The last homomorphism is 0 because of the decomposition of $G$.
Thus, by an easy induction, one has $ \mathrm{Coker}(E_2^{2,0}\times E_2^{0,2}\stackrel{\cup}{\longrightarrow}E_2^{2,2})\simeq H^3(G_{m-1},\mathbb Z)\simeq \oplus_{i=1}^{m-2}G_i.$
c) $ \mathrm{Coker}(E_2^{2,0}\times E_2^{2,0}\stackrel{\cup}{\longrightarrow}E_2^{4,0})$:
This can be obtained by an induction from the steps a) and b).
Finally, we obtain the result by induction:
\begin{align*}
\mathrm{Coker}(H^2(G,\mathbb Z)&\times H^2(G,\mathbb Z)\stackrel{\cup}{\longrightarrow}H^4(G,\mathbb Z))
\simeq \oplus_{i=1}^{m-1}H^3(G_{i},\mathbb Z)\\
&\simeq \oplus_{i=1}^{m-2}G_i^{m-i-1} \simeq \oplus_{i=1}^{m-2}C_i^{d_{i}}. \qedhere
\end{align*}
\end{proof}
We immediately obtain the following theorem by the previous lemmas.
\begin{thm}
Let $W=R_{K/F}^{(1)}(\mathbb G_{m,K})$ and $K/F$ an abelian extension with $\mathrm{Gal}(K/F)= G\simeq\oplus_{i=1}^mC_i$ such that each $C_i$ is cyclic and $\#C_{i-1}|\#C_i$ $(1<i\leq m)$. Denote $d_{i}=(m-i)(m-i-1)/2$. Then for any odd prime $p$,
$$
\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))\{p\}\simeq H^3(G,K^*)\{p\} \oplus\oplus_{i=1}^{m-2}C_i\{p\}^{d_{i}}.
$$
Moreover, if $K/F$ has odd degree, then
$$
\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))\simeq H^3(G,K^*) \oplus \oplus_{i=1}^{m-2}C_i^{d_{i}}.
$$
\end{thm}
\begin{proof}
By (\ref{abelian}), if $\Gal(K/F)$ has odd order, $2$-primary part of $\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))$ is trivial, the proof then follows.
\end{proof}
In particular, if $m=2$, we deduce:
\begin{cor}
Assume $W=R_{K/F}^{(1)}(\mathbb G_{m,K})$ and $K/F$ an abelian extension with $\mathrm{Gal}(K/F)=G\simeq C_1\oplus C_2$ such that $C_1$ and $C_2$ are cyclic group, then $$\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))\{p\}= H^3(G,K^*)\{p\}$$ for any odd prime $p$. In particular, if $G$ has odd order, then $$\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))= H^3(G,K^*).$$
\end{cor}
\begin{exa} \label{exa:un}
Let $F=\mathbb{Q}(\sqrt{-3})$, $K/F$ an abelian field and $W=R_{K/F}^{(1)}(\mathbb G_{m,K})$.
(1) Let $K=F(\sqrt[3]{2}, \sqrt[3]{7})$,
then $\mathrm{Gal}(K/F)=G\simeq\mathbb{ Z}/3 \oplus \mathbb{ Z}/3$. By Remark \ref{rem: K}, one obtains $H^3(G,K^*)=0$ since $K_{\frak \nu}/F_{\frak p}$ has degree $9$, where $\nu\mid \frak p$ are places over $7$. So $$\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2)) = 0.$$
(2) Let $K=F(\sqrt[3]{2}, \sqrt[3]{3})$,
then $\mathrm{Gal}(K/F)=G\simeq\mathbb{ Z}/3\oplus \mathbb{ Z}/3$. Its local Galois groups are cyclic, which implies $H^3(G,K^*)=\mathbb{ Z}/3$ by Remark \ref{rem: K}. So $$\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))= \Z/3.$$
(3) Let $K=F(\sqrt[3]{2},\sqrt[3]{3}, \sqrt[3]{7})$,
then $\mathrm{Gal}(K/F)=G\simeq(\mathbb{ Z}/3)^3$. Its local Galois groups are $0, \mathbb{ Z}/3 \text{ or }(\mathbb{ Z}/3)^2$, which implies $H^3(G,K^*)=\mathbb{ Z}/3$ by Remark \ref{rem: K}. So $$\bar{H}^3_{nr}(F(W),\mathbb Q/\mathbb Z(2))= (\Z/3)^2.$$
\end{exa}
| bb50d72d4d7e43e97969326359e68f020e4594d5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\section{Minimum message length}
\label{sec:mml}
Introduced in the late 1960s by Wallace and Boulton, the minimum message length (MML) principle~\cite{WallaceBoulton68, WallaceFreeman87, WallaceDowe99a, Wallace05} is a framework for inductive inference rooted in information theory. The key idea underlying MML is that both parameter estimation and model selection problems can be interpreted as examples of data compression. It is well known that a random data string is not compressible. Conversely, if we have managed to compress a string of data we have (with high probability) learned something about the underlying structure of the data. Given data ${\bf y} \in \mathbb{R}^n$, the MML principle provides a method for computing the minimum length of a message that describes this data. In the MML approach, this message length is the single necessary inferential quantity. For the MML message to be decodable by a receiver, we require that the message encodes both a model for the data as well as the data itself. Formally, a MML message comprises two parts:
\begin{enumerate}
\item the \emph{assertion}: describes the structure of the model, including all model parameters $\bm{\theta} \in \bm{\Theta} \in \mathbb{R}^p$. Let $I(\bm{\theta})$ denote the codelength of the assertion.
\item the \emph{detail}: describes the data ${\bf y}$ using the model $p({\bf y} | \bm{\theta})$ nominated in the assertion. Let $I({\bf y} | \bm{\theta})$ denote the codelength of the detail.
\end{enumerate}
The total length of the MML message, $I({\bf y}, \bm{\theta})$, measured in units of information (for example, bits) is the sum of the lengths of the assertion and the detail:
\begin{equation}
\label{eqn:mml:codelength}
I(D, \bm{\theta}) = \underbrace{ I(\bm{\theta}) }_{\rm assertion} + \underbrace{I({\bf y} | \bm{\theta})}_{\rm detail} .
\end{equation}
The length of the assertion measures the complexity of the model, with longer assertions able to state more parameters with high accuracy or describe more complicated model structures. In contrast, a short assertion may encode the model parameters imprecisely and describe only simple models. The length of the detail tells us how well the model stated in the assertion is able to fit or describe the data. A complex model, that is one with a long assertion, will have lots of explanatory power and be able to encode more data strings using fewer bits compared to a simple model. MML seeks the model
\begin{equation}
\hat{\bm{\theta}}({\bf y}) = \argmin_{\bm{\theta} \in \bm{\Theta}} \left\{ I({\bf y}, \bm{\theta}) \right\}
\label{eqn:mml}
\end{equation}
that minimises the length of the two-part message message. The key point is that minimising the two part message requires balancing the complexity of the model (assertion) with how well the model describes the data (detail). Ideally, we wish to find the simplest model that fits the observed data well enough; essentially, a formalisation of the famous razor of Occam. An advantage of MML is that the message length, measured in (say) bits, is a universal gauge that allows comparison across models with different model structures and numbers of parameters. As long as we can compute the MML codelengths of models, we can compare them. In this fashion, an MML practitioner is able to compare, for example, a linear regression model~\cite{SchmidtMakalic09c}, to a finite mixture model~\cite{WallaceDowe00} to a decision tree~\cite{WallacePatrick93} via their codelengths for some observed data set.
The aim of this article is to introduce the MML principle of inductive inference to readers that may not be overly familiar with information theoretic statistics. We will examine two key MML codelength solutions to (\ref{eqn:mml}): the exact, or the so called Strict MML codelength (see Section~\ref{sec:smml}), and a practical approximation to the exact codelength (see Section~\ref{sec:mml87}). We will demonstrate how to perform MML analysis of two common statistical problems: the two-sample $t$-test problem (see Section~\ref{sec:mml:ttest}) and testing of the correlation coefficient (see Section~\ref{sec:mml:corr}). Lastly, we compare MML to minimum description length (MDL), an alternative approach to inductive inference developed independently by Rissanen and colleagues~\cite{Rissanen78,Rissanen96,RissanenRoos07a,Rissanen09b}.
\subsection{Strict Minimum Message Length}
\label{sec:smml}
The exact solution to (\ref{eqn:mml}) is usually called Strict MML~\cite{WallaceBoulton75, Wallace05}, and may be viewed as a ``gold standard' which more tractable formulae seek to approximate. To describe Strict MML we shall closely follow the notation in~\cite{Wallace05} (pp. 154--155). Let
\begin{equation}
r({\bf y}) = \int \pi(\bm{\theta}) p({\bf y} | \bm{\theta}) \, d\bm{\theta} ,
\end{equation}
denote the marginal distribution of the $n$ data points ${\bf y}$ given a prior distribution $\pi(\bm{\theta})$ over the model parameters $\bm{\theta} \in \mathbb{R}^p$. We note that MML is strictly Bayesian and requires a prior distribution for the model parameters in order to perform inference. Next we define the subset $\Theta^* = \{\bm{\theta}^*_1, \bm{\theta}^*_2, \ldots\} \subset \Theta$ as the set of parameters that can be used to encode the data ${\bf y}$, where each $\bm{\theta}_j^*$ may be encoded with different codelength. Recall that the assertion describes the model structure and all model parameters $\bm{\theta} \in \Theta$. To encode the assertion we require that $\Theta^*$ be a countable set as the assertion is finite and can encode only a countable subset of $\Theta$. Clearly, in many inference problems the parameter set $\Theta$ is continuous and not countable and the question of how to quantize the parameter space $\Theta$ into a countable set $\Theta^*$ is a key issue in MML inference. Assuming we have somehow obtained the quantized parameter set $\Theta^*$, the code for $\Theta^*$ implies a probability distribution over the set $\Theta^*$ since
\begin{equation*}
p(\bm{\theta}^*_j) = q_j > 0, \quad \sum_{j: \bm{\theta}_j^* \in \Theta^*} q_j = 1, \quad (j = 1,2,\ldots).
\end{equation*}
The assertion is completely determine by the set $\Theta^*$ and a distribution (code) over the set implying that a $\bm{\theta}_j^*$ should be optimally encoded with length $-\log q_j$.
The second part of the message, the detail, is the encoding of the data given the model $\hat{\bm{\theta}} (\equiv \bm{\theta}_j^*)$ specified in the assertion. Once we have chosen the model $\hat{\bm{\theta}}$, the data ${\bf y}$ can be encoded with length $-\log p({\bf y} | \hat{\bm{\theta}})$. Let $t_j = \{{\bf y} \in \mathbb{R}^n: m({\bf y}) = \bm{\theta}^*_j\}$ denote the set of data values ${\bf y}$ that will be encoded using some parameter $\bm{\theta}^*_j \in \Theta^*$ in the detail component of the message. Similarly, denote by $q_j = q(\bm{\theta}^*_j)$ the probability mass associated with parameter $\bm{\theta}^*_j \in \Theta^*$ used to construct the assertion codeword for $\bm{\theta}^*_j$. Strict MML seeks the mapping $m(\cdot)$ which minimizes the expected codelength of the message~\cite{Wallace05} (pp. 155)
\begin{align}
I_S &= \sum_{{\bf y} \in \mathbb{R}^n} r({\bf y}) \left[ - \log q(m({\bf y})) - \log p({\bf y} | m({\bf y}) \right],
\label{eqn:smml:codelength1}
\end{align}
where the expectation is taken with respect to the marginal distribution of the data. The first term in the expectation, $-\log q(m({\bf y}))$, is the length of the codeword stating parameter $\bm{\theta}^*_j \in \Theta^*$ in the assertion, while the second term, $-\log p({\bf y} | m({\bf y}))$, denotes the codelength of a data set ${\bf y}$ encoded using parameter $\bm{\theta}^*_j$.
\begin{table*}[tb]
\footnotesize
\begin{center}
\begin{tabular}{cllc}
\toprule[1pt]
$n$ & SMML PARTITION & $\Theta^*$ & $I_S$ (bits) \\
\cmidrule{1-4}
1 &\{0..1\} &\{0.500\} & 1.000\\
2 &\{0..0, 1..2\} &\{0.000, 0.750\} & 1.667\\
3 &\{0..0, 1..3\} &\{0.000, 0.667\} & 2.085\\
4 &\{0..0, 1..3, 4..4\} &\{0.000, 0.500, 1.000\} & 2.454\\
5 &\{0..0, 1..4, 5..5\} &\{0.000, 0.500, 1.000\} & 2.704\\
6 &\{0..0, 1..5, 6..6\} &\{0.000, 0.500, 1.000\} & 2.962\\
7 &\{0..3, 4..7\} &\{0.214, 0.786\} & 3.165\\
8 &\{0..2, 3..7, 8..8\} &\{0.125, 0.625, 1.000\} & 3.337\\
9 &\{0..0, 1..5, 6..9\} &\{0.000, 0.333, 0.833\} & 3.491\\
10 &\{0..0, 1..4, 5..9, 10..10\} &\{0.000, 0.250, 0.700, 1.000\} & 3.647\\
11 &\{0..0, 1..5, 6..10, 11..11\} &\{0.000, 0.273, 0.727, 1.000\} & 3.762\\
12 &\{0..0, 1..6, 7..11, 12..12\} &\{0.000, 0.292, 0.750, 1.000\} & 3.887\\
13 &\{0..0, 1..6, 7..12, 13..13\} &\{0.000, 0.269, 0.731, 1.000\} & 3.998\\
14 &\{0..3, 4..10, 11..14\} &\{0.107, 0.500, 0.893\} & 4.107\\
15 &\{0..0, 1..5, 6..12, 13..15\} &\{0.000, 0.200, 0.600, 0.933\} & 4.204\\
16 &\{0..0, 1..5, 6..12, 13..16\} &\{0.000, 0.188, 0.563, 0.906\} & 4.289\\
17 &\{0..0, 1..6, 7..13, 14..17\} &\{0.000, 0.206, 0.588, 0.912\} & 4.372\\
18 &\{0..0, 1..5, 6..12, 13..17, 18..18\} &\{0.000, 0.167, 0.500, 0.833, 1.000\} & 4.457\\
19 &\{0..0, 1..5, 6..12, 13..18, 19..19\} &\{0.000, 0.158, 0.474, 0.816, 1.000\} & 4.531\\
20 &\{0..0, 1..6, 7..14, 15..19, 20..20\} &\{0.000, 0.175, 0.525, 0.850, 1.000\} & 4.601\\
21 &\{0..0, 1..6, 7..14, 15..20, 21..21\} &\{0.000, 0.167, 0.500, 0.833, 1.000\} & 4.665\\
22 &\{0..0, 1..6, 7..15, 16..21, 22..22\} &\{0.000, 0.159, 0.500, 0.841, 1.000\} & 4.737\\
23 &\{0..3, 4..11, 12..19, 20..23\} &\{0.065, 0.326, 0.674, 0.935\} & 4.801\\
24 &\{0..2, 3..9, 10..17, 18..23, 24..24\} &\{0.042, 0.250, 0.563, 0.854, 1.000\} & 4.863\\
25 &\{0..0, 1..6, 7..15, 16..22, 23..25\} &\{0.000, 0.140, 0.440, 0.760, 0.960\} & 4.920\\
26 &\{0..0, 1..6, 7..14, 15..22, 23..26\} &\{0.000, 0.135, 0.404, 0.712, 0.942\} & 4.974\\
27 &\{0..0, 1..6, 7..15, 16..23, 24..27\} &\{0.000, 0.130, 0.407, 0.722, 0.944\} & 5.025\\
28 &\{0..3, 4..12, 13..21, 22..27, 28..28\} &\{0.054, 0.286, 0.607, 0.875, 1.000\} & 5.080\\
29 &\{0..4, 5..13, 14..22, 23..28, 29..29\} &\{0.069, 0.310, 0.621, 0.879, 1.000\} & 5.129\\
30 &\{0..0, 1..5, 6..14, 15..23, 24..29, 30..30\} &\{0.000, 0.100, 0.333, 0.633, 0.883, 1.000\} & 5.176\\
\vspace{-3mm} \\
\bottomrule[1pt]
\vspace{+1mm}
\end{tabular}
\caption{SMML partitions for a binomial distribution under a uniform prior on the probability of success. The countable set $\Theta^*$ denotes the parameters that would be used to code the data.\label{tab:smml:binomial}}
\end{center}
\end{table*}
First, note that (\ref{eqn:smml:codelength1}) minimizes the expected codelength over all possible data values because the map $m(\cdot)$ depends on unobserved data values that could be generated by $r({\bf y})$, and not just on the particular data set that has been observed. Second, the expression for the expected codelength includes a sum over all possible data sets of size $n$ as SMML considers the space of data to be countable. This is reasonable since we are interested in computing the length of a message that transmits the observed data and this message must be finite. The expected codelength can be written as
\begin{align}
\nonumber
I_S &= -\sum_{\bm{\theta}^*_j \in \Theta^*} \sum_{{\bf y}_i \in t_j} r({\bf y}_i) \left( \log q(\bm{\theta}^*_j) + \log p({\bf y}_i | \bm{\theta}^*_j) \right),\\
\label{eq:smml:optim}
&= \underbrace{-\sum_{\bm{\theta}^*_j \in \Theta^*} q(\bm{\theta}_j) \log q(\bm{\theta}^*_j) }_{\text{assertion}} \underbrace{-\sum_{\bm{\theta}^*_j \in \Theta^*} \sum_{y_i \in t_j} r({\bf y}_i) \log p({\bf y}_i | \bm{\theta}^*_j) }_{\text{detail}},
\end{align}
where the coding probability of $\bm{\theta}^*_j \in \Theta^*$ is
\begin{equation}
q(\bm{\theta}^*_j) = \sum_{{\bf y}_i \in t_j} r({\bf y}_i) ,
\end{equation}
which is the sum of the marginal probabilities of the data values that result in the estimate $\bm{\theta}^*_j\in \Theta^*$ being used in the detail. The first term in the expected Strict MML codelength is the \emph{expected} length of the assertion, while the second term gives the expected length of the detail. Strict MML seeks to minimize the expected codelength since the map $m(\cdot)$ depends on all data values ${\bf y} \in \mathbb{R}^n$ and not just the observed data.
By seeking the optimal map $m(\cdot)$, Strict MML partitions the data space into disjoint regions $t_j$, such that all data sets that fall within a particular region $t_j$ are encoded with the same parameter value $\bm{\theta}^*_j \in \Theta^*$. This partitioning of the data space yields the quantized parameter set $\Theta^*$ where members of $\Theta^*$ are the feasible parameter estimates, with $m({\bf y})$ being the point estimate associated with data string ${\bf y}$.
The countable set $\Theta^*$ that minimises (\ref{eq:smml:optim}) is not necessarily unique; however this is of no practical consequence as all optimal $\Theta^*$ result in an efficient encoding, and hence feasible parameter values~\cite{Wallace05} (pp. 161). The Strict MML codelength is invariant to data and model representation and represents the shortest two-part codelength for a message that conveys the data and a model for the data. Strict MML has very close links to fundamental concepts such as Turing machines and Kolmolgorov complexity~\cite{WallaceDowe99c}. We note that, while as have assumed the data space to be discrete in our exposition of Strict MML, this is not necessary and the Strict MML code for continuous data is a straightforward extension of the codelength (\ref{eq:smml:optim}); see~\cite{Wallace05}, pp. 166--169.
Unfortunately, the computational complexity of partitioning the data space and obtaining $\Theta^*$ is NP complete in general and only practically computable for simple models~\cite{FarrWallace02}. To see this, we note that the Strict MML data space partitioning problem is an example of the well-known set partitioning problem (also known as the coalition structure generation problem in the artificial intelligence research community) provided the data space is countably finite~\cite{LamarchePerrin14,RahwanEtAl15}. The set partitioning problem is known to be NP complete in general and cannot be solved by brute-force search as the total number of possible partitions, assuming a finite data space of size $n$, is the $n$-th Bell number $B_n$~\cite{Bell34} and satisfies $\alpha n^n \leq B_n \leq n^n$ for some $\alpha > 0$~\cite{SandholmEtAl99}. In some cases, constraints that impose structure and reduce the search space can be assumed. For example, if the data space is one-dimensional and ordered like in the example below, this is know as the ordered set partitioning problem and is identical to the shortest path problem that can be solved in $O(n^2)$ polynomial time. \\
{\noindent \bf Example: } Consider data $y$ representing the number of successes in $n$ trials which follows the binomial distribution with unknown success probability $\theta \in \Theta = [0,1]$; that is,
\begin{equation}
\label{eqn:pdf:binomial}
p(y | \theta) = \binom{n}{y} \theta^y (1 - \theta)^{n - y} .
\end{equation}
Suppose that the success probability $\theta$ is assumed equally likely to take on any value in the parameter space $\Theta$ a priori; in other words, we assume the uninformative prior distribution $\pi(\theta) = 1$. The marginal distribution of the data is
\begin{equation*}
r(y) = \int_0^1 p(y | \theta) \pi(\theta) d\theta = \int_0^1 \binom{n}{y} \theta^y (1 - \theta)^{n - y} d\theta = \frac{1}{n+1} .
\end{equation*}
To construct a Strict MML code for the binomial distribution we require the map $m(\cdot)$ that minimizes the expected codelength (\ref{eq:smml:optim}) over all possible data sets $y \in \mathbb{N} = \{0,1,\ldots,n\}$, where the expectation is taken with respect to $r(y)$. This optimal codelength will be based on an optimal partitioning of the data space which implicitly yields the quantized parameter set $\Theta^*$. The expected Strict MML codelength for the binomial distribution is
\begin{equation*}
I_S = -\sum_{\theta^*_j \in \Theta^*} q(\theta_j^*) \log q(\theta^*_j) -\sum_{\theta^*_j \in \Theta^*} \sum_{y_i \in t_j} \left(\frac{1}{n+1} \right) \log \binom{n}{y_i} (\theta_j^*)^{y_i} (1 - \theta_j^*)^{n - y_i} , \quad
%
q(\theta^*_j) = \sum_{y_i \in t_j} \frac{1}{n+1},
\end{equation*}
and $t_j$ is the set of data values $y_i$ that will be encoded using $\theta_j^* \in \Theta^*$. For a given partition $\Theta^*$, the contribution each estimate $\theta_j^*$ makes to the expected codelength is part of the detail only and does not influence the codelength of the assertion. To minimize the expected codelength for a particular $\theta_j^*$, we therefore only need to minimize the second part of the message
\begin{equation*}
- \sum_{y_i \in t_j} r(y_i) \log \binom{n}{y_i} (\theta_j^*)^{y_i} (1 - \theta_j^*)^{n - y_i} .
\end{equation*}
Clearly, the $\hat{\theta}^*_j$ that minimizes the above is
\begin{equation*}
\hat{\theta}_j^* = \frac{\sum_{y_i \in t_j} r(y_i) y_i}{n \sum_{y_i \in t_j} r(y_i)} = \frac{1}{n |t_j|} \sum_{y_i \in t_j} y_i, \quad j = 1,2,\ldots ,
\end{equation*}
where $|t_j|$ denotes the size of the set $t_j$; i.e., the number of data values that are encoded with parameter $\theta_j^*$.
It remains to determine the optimal data space partitions $t_j$ and hence the quantized set $\Theta^*$. Before we look at the optimal Strict MML partitioning, we shall look at two examples with suboptimal codelengths. The simplest possible partitioning of the data space $\{0,1,\ldots, n\}$ assumes that all data values are in the same partition, say, $t_1 = \{0, 1, \ldots, n\}$, where the size of the partition is $|t_1| = (n+1)$. This implies that the quantized parameter set $\Theta^* = \{ \theta^*\}$ consists of a single member
\begin{equation*}
\hat{\theta}^* = \frac{1}{n |t_1|} \sum_{y_i \in t_1} y_i = \frac{1}{2} .
\end{equation*}
which has probability $q(\theta^*) = 1$. The expected codelength of the assertion is then 0 bits and the expected codelength of the detail, and hence the total codelength, is
\begin{equation*}
I_S = n \log 2 - \left(\frac{1}{n+1}\right)\sum_{i=0}^n \log \binom{n}{i} .
\end{equation*}
For example, for $n = 10$ and assuming a single partition that contains all possible data of size $n$, the Strict MML codelength for some observed data $y$ is approximately 5.01 bits; here, the estimate of the success probability is $\hat{\theta} = 0.5$.
Alternatively, consider the more complex partitioning where the data space is divided into $(n+1)$ partitions with each datum $y_i \in \mathbb{N}$ belonging to its own partition. In this case, there are $(n+1)$ possible estimates and $\Theta^* = \{ \theta^*_j: \theta^*_j = j / n, j=0,1,\ldots,n\}$. The probability of each estimate is
\begin{equation*}
q_j = q(\theta^*_j) = \sum_{y_i \in t_j} r(y_i) = \frac{1}{n+1}, \quad j = 0, \ldots, n.
\end{equation*}
The expected Strict MML codelength is now
\begin{equation*}
I_S = \log(n+1) -\frac{1}{n+1}\sum _{i=0}^n \log \binom{n}{i} \left(\frac{i}{n}\right)^i \left(1-\frac{i}{n}\right)^{n-i} .
\end{equation*}
For the $n=10$ example, the total codelength of this partitioning is approximately 9.84 bits. This codelength is significantly worse than the single partition Strict MML code.
It is possible to compute the optimal Strict MML solution for the binomial distribution using the dynamic programming algorithm of Farr and Wallace~\cite{FarrWallace02}. The results of the algorithm are shown in Table~\ref{tab:smml:binomial} for all $1 \leq n \leq 30$; a MATLAB implementation of this algorithm can be downloaded from the Mathworks File Exchange website (ID: 80167). When $n=10$, for example, Strict MML partitions the data space into four partitions $\{0, 1..4, 5..9, 10\}$ yielding the parameter set $\Theta^* = \{ 0.00, 0.25, 0.70, 1.00 \}$. This implies that if we observe between 5 and 9 success counts, the estimated probability of success is 0.7. In contrast, if $0$ or $10$ success counts were observed, we would estimate the probability of success as 0 and 1, respectively. The Farr and Wallace algorithm is applicable to any one dimensional Strict MML problem with discrete data and is of polynomial complexity in terms of the sample size $n$. \\
An algorithm to obtain optimal Strict MML codelengths for one dimensional continuous data from exponential families with continuous sufficient statistics was developed by Dowty~\cite{Dowty15a}. Unfortunately, there exists no general algorithm for computing optimal Stict MML partitions outside of the one dimensional (continuous or discrete) data setting. In the case of the $k$-nomial distribution with $k>2$, the Strict MML data partitioning problem is an instance of the set partitioning (coalition structure generation) over graphs. While a dynamic programming algorithm exists to solve this problem, it runs in exponential time and is only practical for small sample sizes~\cite{MichalakEtAl16}. We note that many anytime exact algorithms and heuristic algorithms have also been proposed for this setting; see Rahwan et. al for a recent survey~\cite{RahwanEtAl15}.
\subsection{A practical approximation to the Strict MML codelength}
\label{sec:mml87}
Due to the high computational complexity of deriving the exact Strict MML codelength, Strict MML is mostly of interest from a theoretical standpoint. There exist several approximations to the Strict MML codelength with the MML87 approximation~\cite{WallaceFreeman87,Wallace05} being the most popular. Under some regularity conditions~\cite{Wallace05}) (pp. 226), the MML87 codelength for data $D$ is
\begin{equation}
\label{eqn:mml87:codelength}
\mathcal{I}_{87}({\bf y}, \bm{\theta}) = \underbrace{-\log \pi(\bm{\theta}) + \frac{1}{2} \log \abs{J_{\bm{\theta}}(\bm{\theta})} + \frac{p}{2} \log \kappa_p}_{\rm assertion} + \underbrace{\frac{p}{2} - \log p({\bf y}|\bm{\theta})}_{\rm detail}
\end{equation}
where $\pi_{\bm{\theta}}(\bm{\theta})$ is the prior distribution for the parameters $\bm{\theta}$, $\abs{J_{\bm{\theta}}(\bm{\theta})}$ is the determinant of the expected Fisher information matrix, $p({\bf y}|\bm{\theta})$ is the likelihood function of the model and $\kappa_p$ is a quantization constant~\cite{ConwaySloane98,AgrellEriksson98}; for small $p$ we have
\begin{equation}
\kappa_1 = \frac{1}{12}, \quad \kappa_2 = \frac{5}{36 \sqrt{3}}, \quad \kappa_3 = \frac{19}{192 \times 2^{1/3}},
\end{equation}
while, for large $p$, $\kappa_p$ is well-approximated by~\cite{Wallace05}:
\begin{equation}
\frac{p}{2} (\log \kappa_p + 1) \approx -\frac{p}{2} \log 2\pi + \frac{1}{2} \log p \pi - \gamma,
\end{equation}
where $\gamma \approx 0.5772$ is the Euler--Mascheroni constant.
In contrast to SMML, which quantizes the parameter space by associating a single parameter estimate with a number of different data strings, the MML87 approximation achieves computational tractability by instead associating a set of models, deemed ``indistinguishable'' in an information sense, with the observed data string ${\bf y}$. This set of parameter estimates can be thought of as the range of possible parameter estimates that a plausible choice of $m(\cdot)$ would associate with the data string ${\bf y}$. The volume of this set of models, which is called the \emph{uncertainty region} in MML parlance, explicitly determines the accuracy to which the continuous model parameters should be stated (i.e., the degree of quantization). We can re-write the codelength of the assertion in terms of the volume of the uncertainty region $w(\bm{\theta})$ as follows
\begin{equation}
\underbrace{-\log \pi(\bm{\theta}) w(\bm{\theta})}_{\rm assertion}, \quad w(\bm{\theta}) = \left( \abs{J_{\bm{\theta}}(\bm{\theta})} \kappa_p^p \right)^{-1/2} .
\end{equation}
It is clear that the volume of the uncertainty region depends on the variation of the negative log-likelihood function. If a small change in $\bm{\theta}$ results in a large change in the log-likelihood, the volume of the uncertainty region will be relatively small, implying that the parameters must be stated with higher precision. In contrast, if the negative log-likelihood is not particularly sensitive to small changes in $\bm{\theta}$, the volume of the uncertainty region will be fairly large. Under suitable regularity conditions, the volume of the uncertainty region is $O(n^{-p/2})$.
For many sufficiently well-behaved models, the MML87 codelength is virtually identical to the Strict MML codelength while being simpler to compute, requiring only the prior distribution for the model parameters and the determinant of the expected Fisher information matrix. Unlike the Strict MML codelength $I_S$ (\ref{eqn:smml:codelength1}) which must be found via minimization of an expectation, taken over all possible data with respect to the marginal distribution $r(\cdot)$, the MML87 codelength is computed using the observed data only. For large sample sizes $n \to \infty$, it is easy to show that the MML87 codelength is asymptotically equivalent to the well-known Bayesian information criterion (BIC)~\cite{Schwarz78}
\begin{equation}
\label{eqn:mml87:bic}
\mathcal{I}_{87}({\bf y}, \bm{\theta}) = - \log p({\bf y}|\bm{\theta}) + \frac{p}{2} + O(1) ,
\end{equation}
where the $O(1)$ term depends on the prior distribution, the Fisher information and the number of parameters $p$. In fact, as the MML87 codelength $\mathcal{I}_{87}({\bf y}, \bm{\theta})$ can be interpreted as the negative logarithm of the posterior probability mass attached to a dataset ${\bf y}$ and model $\bm{\theta}$, the difference in message lengths admits the interpretation as the logarithm of the posterior-odds of two models, e.g.,
\[
\mathcal{I}_{87}({\bf y}, \bm{\theta}_0) - \mathcal{I}_{87}({\bf y}, \bm{\theta}_1)
\]
can be interpretated as the posterior log-odds in favour of model $\bm{\theta}_1$ against $\bm{\theta}_0$. The MML87 codelength results in estimates that are invariant under (smooth) one-to-one reparametarization, just like the maximum likelihood estimate. MML87 has been applied to a wide range of statistical models including decision tress~\cite{WallacePatrick93}, factor analysis~\cite{WallaceFreeman92} and mixture models~\cite{WallaceDowe00}. \\
\noindent {\bf Example:} We will compute the MML87 codelength for the binomial distribution and compare it to the Strict MML code~\cite{Wallace05} (pp. 246). Let $y \in [0, n]$ be the count of successes which follows a binomial distribution with probability density function (\ref{eqn:pdf:binomial}) where $0 < \theta < 1$ is the probability of observing a success. Suppose we assume a uniform prior distribution for the success probability $\pi(\theta) = 1$. The determinant of the expected Fisher information matrix for a binomial distribution is $J(\theta) = n/(\theta(1-\theta))$. The volume of the uncertainty region is
\begin{equation*}
w(\theta) = \left(\frac{12 (1-\theta ) \theta }{n}\right)^{1/2}
\end{equation*}
which gets smaller as $n \to \infty$, or as the success probability $\theta$ approaches $\theta \to 0$ or $\theta \to 1$.
Substituting into the MML87 codelength (\ref{eqn:mml87:codelength}) we obtain
\begin{align}
\mathcal{I}_{87}(y, \theta) &= \frac{1}{2} \log \left( \frac{n}{\theta(1-\theta)} \right) + \frac{1}{2} \log \frac{1}{12} - \log \binom{n}{y} \theta^y (1 - \theta)^{n - y} + \frac{1}{2} \nonumber \\
&= -\left(y + \frac{1}{2}\right) \log \theta - \left(n - y + \frac{1}{2}\right) \log (1 - \theta) + \frac{1}{2} \left(1 + \log \frac{n}{12 \binom{n}{y}^2}\right)
\end{align}
This codelength is minimized by choosing $\theta$ to be the MML87 estimate
\begin{equation}
\hat{\theta}(y) = \frac{y + 1/2}{n + 1} ,
\end{equation}
which results in the optimal MML87 codelength for data $y$ being
\begin{align*}
\mathcal{I}_{87}(y)= -\left(y + \frac{1}{2}\right) \log \hat{\theta}(y) - \left(n - y + \frac{1}{2}\right) \log (1 - \hat{\theta}(y)) + \frac{1}{2} \left(1 + \log \frac{n}{12 \binom{n}{y}^2}\right) .
\end{align*}
As an example, if $n = 10$ and $y = 3$, the MML87 estimate of the success probability is $\hat{\theta}(y) = 7/22 \approx 0.32$, the volume of the uncertainty region is $w(\theta) \approx 0.51$ and the MML87 codelength is approximately 3.61 bits. In contrast, the SMML estimate of the success probability is $\hat{\theta}(y) = 0.25$ resulting in a codelength of 3.647 bits. We can empirically show that the expected MML87 codelength with respect to the marginal distribution $r(y) = 1/(n+1)$ falls within 0.1 bits of the expected Strict MML codelength for all $n \geq 5$, while being significantly simpler to compute.
\section{Minimum message length $t$-test}
\label{sec:mml:ttest}
Perhaps the most popular hypothesis test in practice is the frequentist two-sample $t$-test for comparison of two means. As an example of the $t$-test popularity, a search for the term `paired $t$-tests' returns approximately 12,000 articles from 1990 onwards in PubMed. Similarly, articles published in peer reviewed psychology journals feature, on average, more than three $t$-tests per article~\cite{WetzelsEtAl11}. Formally, the $t$-test assumes that we observe normally distributed data from two, possibly different, populations:
\begin{equation}
Y_{1i} \sim \mathcal{N}\left(\mu + \frac{\sigma \delta}{2}, \sigma^2\right), \quad Y_{2j} \sim \mathcal{N}\left(\mu - \frac{\sigma \delta}{2}, \sigma^2\right) ,
\end{equation}
for $i=1,\ldots,n_1$ and $j = 1, \ldots, n_2$. Here, the unknown parameter $\mu \in \mathbb{R}$ is the overall (grand) mean, $\sigma > 0$ is a standard deviation that is common to both groups, $\delta>0$ is the standardised effect size and $(n_1, n_2)$ are the sample sizes of the two groups. This formulation of the $t$-test in terms of the overall mean and standardised effect size is due to \cite{GonenEtAl05}. The usual null hypothesis says that the mean is the same in both groups while the alternative hypothesis specifies a different mean for each group; formally, we have
\begin{equation}
\mathcal{H}_0: \delta = 0, \quad \mathcal{H}_1: \delta \neq 0 .
\end{equation}
Let $\bm{\theta}_0 = \{\mu, \sigma\} \in (\mathbb{R} \times \mathbb{R}_+) \equiv \Theta_0$ and $\bm{\theta}_1 = \{\mu, \sigma, \delta\} \in (\mathbb{R}^2 \times \mathbb{R}_+) \equiv \Theta_1$ denote the parameter set under the two models. Given data ${\bf y} = \{{\bf y}_1, {\bf y}_2\}$ where ${\bf y}_1 = (y_{1,1}, \ldots, y_{n_1,1})^\prime$ and ${\bf y}_2 = (y_{2,1}, \ldots, y_{n_2,1})^\prime$, the task is estimate the effect size $\delta > 0$ and thus determine whether the population means of the two groups are equal ($\delta = 0$).
Recently, there has been a great deal of interest in Bayesian alternatives to the frequentist $t$-test~\cite{GonenEtAl05,RouderEtAl09,WangLiu16,GronauEtAl19,Kelter21}. A key quantity in Bayesian analysis is the Bayes factor
\begin{equation}
\underbrace{ \frac{P(\mathcal{H}_1 | {\bf y})}{P(\mathcal{H}_0 | {\bf y})} }_{\text{Posterior odds}} = \underbrace{ \frac{p({\bf y} | \mathcal{H}_1)}{p({\bf y} | \mathcal{H}_0)} }_{\text{Bayes factor}} \underbrace{ \frac{P(\mathcal{H}_1)}{P(\mathcal{H}_0} }_{\text{Prior odds}} ,
\end{equation}
which is the ratio of the marginal likelihoods of the data obtained by integrating the parameters out with respect to their prior distribution; that is,
\begin{equation}
p({\bf y} | \mathcal{H}_i) = \int p({\bf y} | \bm{\theta}_i) \pi_i(\bm{\theta}_i) \, d\bm{\theta}, \quad i = 0, 1 ,
\end{equation}
where $\pi_i(\bm{\theta}_i)$ is the prior distribution of the parameters under model $i$. The Bayes factor, henceforth denoted by BF$_{10}$, measures the amount of evidence in support of hypothesis $\mathcal{H}_1$ over hypothesis $\mathcal{H}_0$, where BF$_{10} > 3$ indicates substantial evidence in favour of $\mathcal{H}_1$~\cite{KassRaftery95}. Given a proper prior $\pi(\delta)$ on the effect size $\delta$, Gronau et al.~\cite{GronauEtAl19}, generalizing the work of \cite{GonenEtAl05,RouderEtAl09}, showed that the Bayes factor for the $t$-test problem can be expressed as
\begin{equation}
\text{BF}_{10} = \frac{\int T_\nu(t | \sqrt{n_\delta} \delta ) \pi(\delta) \, d\delta }{T_{\nu}(t)}
\end{equation}
where $t$ is the observed $t$-statistic
\begin{equation}
t = \sqrt{n_\delta} (\bar{\bf y}_1 - \bar{\bf y}_2) / s_p, \quad \nu s_p = (n_1 - 1) s_1^2 + (n_2 - 1) s_2^2,
\end{equation}
and $s_j$ ($j=1,2)$ are the usual unbiased sample variance estimates. Here, $\nu = n_1 + n_2 - 2$ denotes the degrees of freedom, $n_\delta = (1/n_1+1/n_2)^{-1}$ is the effective sample size, and $T_\nu(\cdot|a)$ is the non-central Student $t$ distribution with degrees of freedom $\nu$ and non-centrality parameter $a$. For most prior distributions $\pi(\delta)$, the integral in the Bayes factor must be computed by numerical integration.
In order to perform the two-sample $t$ test within the MML framework, we require the codelengths of the data under the null distribution $(\delta = 0)$ and under the alternative hypothesis $(\delta \neq 0)$. Under the null model, the negative log-likelihood of the data ${\bf y}$ is
\begin{eqnarray}
-\log \ell_0(\bm{\theta}_0) &=& \frac{n_1}{2} \log(2 \pi \sigma^2) + \frac{1}{2 \sigma^2} \sum_{i=1}^{n_1} \left(y_{1i} - \mu\right)^2 + \frac{n_2}{2} \log(2 \pi \sigma^2) + \frac{1}{2 \sigma^2} \sum_{j=1}^{n_2} \left(y_{2j} - \mu\right)^2, \nonumber \\
&=& \frac{n}{2} \log(2 \pi \sigma^2) + \frac{1}{2 \sigma^2} \sum_{i=1}^{n} \left(y_{i} - \mu\right)^2, \nonumber
\end{eqnarray}
where ${\bf y} = (y_1, \ldots, y_n)$ denotes the two-samples stacked into a single data vector of length $n = (n_1 + n_2)$. Let $\pi_0(\bm{\theta}_0)$ denote the prior distributions for the parameters under the null hypothesis. We use the usual right Haar prior for the mean and standard deviation:
\begin{equation}
\pi_0(\bm{\theta}_0) = (\Omega \sigma)^{-1}
\end{equation}
defined over some suitable parameter range $\Omega$ implying a location and scale invariant distribution of the mean and standard deviation, respectively. The determinant of the expected Fisher information matrix for the normal distribution is $|J(\bm{\theta}_0)| = 2n^2/\sigma^4$ leading to the MML87 codelength:
\begin{align*}
\mathcal{I}_0({\bf y}, \bm{\theta}_0) &= -\log \ell_0(\bm{\theta}_0) + \frac{1}{2} \log |J(\bm{\theta}_0)| - \log \pi_0(\bm{\theta}_0) + \log \kappa_2 + 1 \\
&= \frac{1}{2 \sigma^2} \sum_{i=1}^{n} \left(y_{i} - \mu\right)^2 + (n-1) \log (\sigma ) + \frac{1}{2} \log \left( 2^{n+1} \pi^n \left(n e \kappa _2 \Omega \right)^2\right)
\end{align*}
where $\kappa_2 = 5/36/\sqrt{3} \approx 0.0802$. MML estimates that minimize the codelength $\mathcal{I}_0({\bf y}, \bm{\theta}_0)$ correspond to the usual sample mean and the unbiased sample variance estimates
\begin{equation}
\hat{\mu}({\bf y}) = \frac{1}{n} \sum_{i=1}^n y_i = \bar{\bf y}, \quad \hat{\sigma}^2({\bf y}) = \frac{1}{n-1} \sum_{i=1}^n (y_i - \bar{\bf y})^2 .
\end{equation}
The minimum MML87 codelength for the data under the null hypothesis is then
\begin{equation}
\mathcal{I}_0({\bf y}) = \frac{n-1}{2} \left( 1 + \log \hat{\sigma}^2({\bf y}) \right) + \frac{1}{2} \log \left( 2^{n+1} \pi^n \left(n e \kappa _2 \Omega \right)^2\right)
\end{equation}
We observe that the MML principle allows for automatic parameter estimation and model selection, as long as we can compute codelengths of the candidate models. Next, we derive the MML87 codelength for the alternative hypothesis where the two means are different at population level ($\delta \neq 0$).
Recall that $\bm{\theta}_1 = \{\mu, \sigma, \delta\} \in (\mathbb{R}^2 \times \mathbb{R}_+) \equiv \Theta_1$ denotes the parameter set of the alternative hypothesis, which compared to the null model also includes the effect size parameter $\delta$. The negative log-likelihood of the data ${\bf y}$ under the alternative hypothesis, $-\log \ell_1({\bf y} | \bm{\theta}_1)$, is
\begin{equation}
\label{eqn:h1:nll}
\frac{n}{2} \log(2 \pi \sigma^2) + \frac{1}{2 \sigma^2} \sum_{i=1}^{n_1} \left(y_{1i} - \mu - \frac{\sigma \delta}{2}\right)^2 + \frac{1}{2 \sigma^2} \sum_{j=1}^{n_2} \left(y_{2j} - \mu + \frac{\sigma \delta}{2}\right)^2
\end{equation}
where, as before, $n = (n_1+n_2)$ is the total sample size. Let
\begin{equation}
S_j = \sum_{i=1}^{n_j} y_{ji}, \quad S^2 = \sum_{j=1}^2 \sum_{i=1}^{n_j} y_{ji}^2 = {\bf y}^\prime {\bf y}, \quad j = 1,2
\end{equation}
denote the sufficient statistics. For reference, the maximum likelihood estimates of the model parameters are easily shown to be
\begin{align*}
\hat{\mu}_{\rm ML}({\bf y}) &= \frac{1}{2} \left(\frac{S_1}{n_1}+\frac{S_2}{n_2}\right) = \frac{1}{2} \left(\bar{\bf y}_1 + \bar{\bf y}_2\right), \\
%
\hat{\sigma}_{\rm ML}^2({\bf y}) &= \frac{1}{n} \left( S^2 - \left(\frac{S_1^2}{n_1}+\frac{S_2^2}{n_2} \right) \right) = \frac{1}{n} \left( S^2 - \left( n_1 \bar{\bf y}_1^2 + n_2 \bar{\bf y}_2^2\right) \right), \\
%
\hat{\delta}_{\rm ML}({\bf y}) &= \frac{1}{\hat{\sigma}_{\rm ML}} \left(\frac{S_1}{n_1}-\frac{S_2}{n_2}\right) = \frac{1}{\hat{\sigma}_{\rm ML}} \left(\bar{\bf y}_1 - \bar{\bf y}_2\right).
\end{align*}
Next we derive the MML87 codelength for the model and the corresponding MML87 estimates. For the MML codelength of the data under the alternative hypothesis, we again require a prior distribution for all the model parameters and the determinant of the expected Fisher information matrix. Following~\cite{GronauEtAl19}, the prior distribution for the parameters $\bm{\theta}_1$ is chosen to be
\begin{equation}
\label{eqn:h1:prior}
\pi_1(\bm{\theta}_1) = \pi_0(\bm{\theta}_0) \pi(\delta), \quad \pi(\delta) = \frac{1}{\gamma_{\delta}} T_{\kappa} \left( \frac{\delta - \mu_{\delta}}{\gamma_{\delta}}\right)
\end{equation}
where $\pi(\delta)$ is a student $t$ distribution with $\kappa > 0$ degrees of freedom, and location and scale hyperparameters $\mu_{\delta}$ and $\gamma_{\delta} > 0$, respectively. This prior density for the effect size is a symmetric, heavy-tailed flexible distribution allowing experts to incorporate prior knowledge via the hyperparameters $(\mu_{\delta}, \gamma_{\delta})$. Note that the prior distribution for the mean and standard deviation parameters is the same under the null and alternative hypotheses, which implies that the choice of the range of parameters $\Omega$ will have no effect on MML inference as it contributes equally to the codelengths of both hypotheses. The determinant of the expected Fisher information matrix under the alternative hypothesis is
\begin{equation}
\label{eqn:h1:fisher}
| J(\bm{\theta}_1) | = \frac{2 n_1 n_2 (n_1+n_2)}{\sigma^4} .
\end{equation}
Substituting (\ref{eqn:h1:nll}), (\ref{eqn:h1:prior}) and (\ref{eqn:h1:fisher}) into (\ref{eqn:mml87:codelength}) yields the MML87 codelength $\mathcal{I}({\bf y}, \bm{\theta}_1)$ of the data under the alternative hypothesis model. Under this choice of the prior distribution, MML87 estimates of the parameters $\bm{\theta}_1$ are unavailable in closed form and must be obtained via numerical optimisation
\begin{equation}
\hat{ \bm{\theta} }_1({\bf y}) = \argmin_{\bm{\theta}_1 \in \Theta_1} \mathcal{I}({\bf y}, \bm{\theta}_1) .
\end{equation}
To perform a MML $t$-test, we compute the codelength of the null model $\mathcal{I}({\bf y}, \hat{\bm{\theta}}_0)$ and the alternative model $\mathcal{I}({\bf y}, \hat{\bm{\theta}}_1)$, with the preferred model being the one with the shorter codelength. As with standard Bayesian analysis, a codelength difference of about $2.3$ nits or more indicates substantial preference of the model with the smaller codelength. Recall that we can interpret the difference in MML codelengths as the log posterior probability in favour of the model with the shorter codelength.
A brief simulation was performed to compare the behaviour of the MML87 estimate of the standardised effect size $\delta$ to the ML estimate. Data with sample sizes $(n_1, n_2) \in \{5, 10, 20, 50\}$ was generated from the model $(\mu = 0, \sigma = 1)$ with $0.1 \leq \delta \leq 5$ and both ML and MML87 were asked to nominate an estimate of $\delta$. The experiment was repeated for $10^5$ iterations and the estimates were compared using the normalized mean squared error metric
\begin{equation}
\text{NMSE} = \frac{ (\delta - \hat{\delta}({\bf y}))^2 }{\delta} ,
\end{equation}
averaged over all $10^5$ iterations for each experiment. Results for sample sizes $n_1 = n_2 = 5$ and $n_1 = n_2 = 10$ are shown in Figure~\ref{fig:delta:nmse}. Behaviour of the ML and MML87 estimates for moderate to large sample sizes was indistinguishable, as predicted by asymptotic analysis. When the sample size was small, ML tends to overestimate $\delta$ and underestimate $\sigma$. This is in contrast to MML87 which tends to underestimate $\delta$ and is more conservative.
\begin{figure*}[t]
\begin{center}
\subfigure[$n_1=n_2=5$]{
\includegraphics[width=6.0cm]{figs/lognmsen5.pdf}
}%
\subfigure[$n_1=n_2=10$]{
\includegraphics[width=6.0cm]{figs/lognmsen10.pdf}
} \\
\end{center}
\caption{Normalised mean square error for the maximum likelihood and MML estimates of $\delta$ with $n_1=n_2=5$ and $n_1=n_2=10$ data samples averaged over $10^5$ simulations. In all cases, the mean and standard deviation of the data generating model were $\mu=0, \sigma = 1$. \label{fig:delta:nmse}}
\end{figure*}
In terms of model (hypothesis) selection, unlike the frequentist methodology, neither the Bayes factor nor MML approach are designed to explicitly control type I error. Empirically, we observe that both the Bayes factor derived by Gronau et al~\cite{GronauEtAl19} and the MML87 $t$-test derived in Section~\ref{sec:mml:ttest} control type I error rate at approximately 0.10 for all $0.1 < \delta < 2$, if the alternative model is chosen whenever $\mathcal{I}_1(y) < \mathcal{I}_0(y)$, or equivalently, whenever the Bayes factor BF$_{10} > 1$. To control the type I error rate at $0.05$, we would need to select the alternative model for all BF$_{10} > 1.87$ or whenever $(\mathcal{I}_1(y) + \log(1.65)) < \mathcal{I}_0(y)$. Under this setup, all three methods were virtually indistinguishable in terms of type I error rate for all $0.1 < \delta < 2$.
\section{Minimum message length testing of the correlation coefficient}
\label{sec:mml:corr}
Suppose that a pair of random variables $(Y_1, Y_2)$ follows a bivariate normal distribution with means $(\mu_1, \mu_2)$, variances $(\sigma^2_1, \sigma^2_2)$ and correlation coefficient $\rho \in (0,1)$.
We have $n$ observations ${\bf y} = \{(y_{1,i}, y_{2,i})\}$ ($i=1,\ldots,n$) and wish to test the hypothesis
\begin{equation*}
H_0: \rho = \rho_0 \quad \text{versus} \quad H_1: \rho \neq \rho_0
\end{equation*}
where $\rho_0 \in (0,1)$ is a fixed, user-specified value. Let $\bm{\theta}_0 = \{\mu_1,\mu_2,\sigma_1,\sigma_2\} \in \mathbb{R}^2 \times \mathbb{R}^2_+ \equiv \Theta_0$ and $\bm{\theta}_1 = \{\mu_1,\mu_2,\sigma_1,\sigma_2,\rho\} \in \mathbb{R}^2 \times \mathbb{R}^2_+ \times (0,1) \equiv \Theta_1$ denote the parameter sets under the null and alternative hypothesis models, respectively. A large number of frequentist as well as Bayesian procedures exist for this hypothesis testing problem. A good summary of the recent literature is in \cite{PengWang21}, which also introduces a new Bayesian procedure for testing the correlation coefficient under divergence-based priors~\cite{BayarriGarciaDonato08}.
Under the null model, the negative log-likelihood of the data ${\bf y}$ is
\begin{equation}
-\log \ell_0(\bm{\theta}_0) = n \log(2\pi) + n \log(\sigma_1 \sigma_2) + \frac{n}{2} \log(1 - \rho_0^2) + \frac{ 1 }{2 \left(1-\rho_0^2\right)} \sum_{i=1}^n Q_i({\bf y},\bm{\theta}_0)
\end{equation}
where
\begin{equation}
Q_i({\bf y},\bm{\theta}) = \frac{\left(y_{1,i}-\mu _1\right){}^2}{\sigma _1^2}-\frac{2 \rho \left(y_{1,i}-\mu _1\right) \left(y_{2,i}-\mu _2\right)}{\sigma _1 \sigma _2}+\frac{\left(y_{2,i}-\mu _2\right){}^2}{\sigma _2^2} .
\end{equation}
Let
\begin{equation}
\bar{y}_j = \frac{1}{n} \sum_{i=1}^n y_{j,i},\quad s_j^2 = \frac{1}{n} \sum_{i=1}^n (y_{j,i} - \bar{y}_j)^2, \quad r = \frac{1}{n s_1 s_2} \sum_{i=1}^n (y_{1,i} - \bar{y}_1)(y_{2,i} - \bar{y}_2),
\end{equation}
for j = 1,2 denote the usual sufficient statistics. Maximum likelihood estimates of the model parameters are
\begin{equation}
\hat{\mu}_j({\bf y})_{\text{ML}} = \bar{y}_j, \quad \hat{\sigma}_j^2({\bf y})_{\text{ML}} = s_j^2 \left(\frac{ 1-\rho_0 \, r}{1-\rho_0 ^2} \right), \quad (j=1,2) .
\end{equation}
For the MML codelength, we again use the right Haar prior for the means and standard deviations
\begin{equation}
\pi_0(\bm{\theta}_0) = (\Omega_0 \sigma_1 \sigma_2)^{-1}
\end{equation}
where $\Omega_0$ is a normalisation constant. The determinant of the expected Fisher information matrix for a bivariate normal distribution with a fixed (known) correlation coefficient $\rho_0$ is
\begin{equation}
|J(\bm{\theta}_0)| = \frac{4 n^4}{\left(1 - \rho _0^2\right){}^2 \sigma _1^4 \sigma _2^4}
\end{equation}
and the MML87 codelength for the data ${\bf y}$ is
\begin{align*}
\mathcal{I}_0({\bf y}, \bm{\theta}_0) &= -\log \ell_0(\bm{\theta}_0) + \frac{1}{2} \log |J(\bm{\theta}_0)| - \log \pi_0(\bm{\theta}_0) + 2 (1 + \log \kappa_4) \\
&= (n-1) \log(\sigma_1 \sigma_2) + \frac{n-2}{2} \log(1 - \rho_0^2) + \frac{ 1 }{2 \left(1-\rho_0^2\right)} \sum_{i=1}^n Q_i({\bf y},\bm{\theta}_0) \\
& + \log \left(\Omega _0 \pi ^n 2^{n+1} (\kappa \, n)^2\right) + 2
\end{align*}
where $\kappa_4 \approx 0.076603$. MML estimates that minimise the codelength under the null hypothesis are
\begin{equation}
\hat{\mu}_j({\bf y}) = \bar{y}_j, \quad \hat{\sigma}_j^2({\bf y}) = s_j^2 \left(\frac{ n (1-\rho_0 \, r)}{(n-1) \left(1-\rho_0 ^2\right)} \right), \quad (j=1,2) .
\end{equation}
We observe that the MML estimates of the mean parameters are the same as the usual maximum likelihood estimates. However, MML estimates of the variances include the extra term $n/(n-1)$. If the null hypothesis corresponds to the no correlation model ($\rho_0 = 0)$, MML estimates of the variances are unbiased unlike the corresponding maximum likelihood estimates.
Under the alternative hypothesis, the correlation coefficient $\rho$ is unknown and must be estimated from the data along with the means and variances. The negative log-likelihood of the data ${\bf y}$ is now
\begin{equation}
-\log \ell_1(\bm{\theta}_1) = n \log(2\pi) + n \log(\sigma_1 \sigma_2) + \frac{n}{2} \log(1 - \rho^2) + \frac{ 1 }{2 \left(1-\rho^2\right)} \sum_{i=1}^n Q_i({\bf y},\bm{\theta}_1)
\end{equation}
Maximum likelihood estimates of the parameters are
\begin{equation}
\hat{\mu}_j({\bf y})_{\text{ML}} = \bar{y}_j, \quad \hat{\sigma}_j^2({\bf y})_{\text{ML}} = s_j^2, \quad \hat{\rho}({\bf y})_{\text{ML}} = r, \quad (j=1,2) .
\end{equation}
We use the same right Haar prior for the means and standard deviations and opt for the uniform prior distribution over the correlation coefficient leading to
\begin{equation}
\pi_1(\bm{\theta}_1) = (\Omega_1 \sigma_1 \sigma_2)^{-1}
\end{equation}
where $\Omega_1$ is a normalisation constant. The determinant of the expected Fisher information matrix for a bivariate normal distribution with an unknown correlation coefficient is
\begin{equation}
|J(\bm{\theta}_1)| = \frac{4 n^5}{\left(1 - \rho ^2\right)^4 \sigma _1^4 \sigma _2^4} .
\end{equation}
MML codelength of the alternative hypothesis is
\begin{align*}
\mathcal{I}_1({\bf y}, \bm{\theta}_1) &= -\log \ell_1(\bm{\theta}_1) + \frac{1}{2} \log |J(\bm{\theta}_1)| - \log \pi_1(\bm{\theta}_1) + \frac{5}{2} (1 + \log \kappa_5) \\
&= (n-1) \log(\sigma_1 \sigma_2) + \frac{n-4}{2} \log(1 - \rho^2) + \frac{ 1 }{2 \left(1-\rho^2\right)} \sum_{i=1}^n Q_i({\bf y},\bm{\theta}_1) \\
& + \log \left( \Omega _1 \pi ^n 2^{n+1} (n \, \kappa _5)^{5/2} \right) + \frac{5}{2}
\end{align*}
where $\kappa_5 \approx 0.075625$. MML estimates that minimise the codelength are
\begin{align*}
\hat{\mu}_j({\bf y}) &= \bar{y}_j, \\
\hat{\sigma}^2_j ({\bf y}) &= s_j^2 \left(\frac{n (n + 2 + \sqrt{(n+2)^2 - 12 r^2 (n-1)} )} {2(n+2)(n-1)} \right) = s_j^2 \left( \frac{n (n-3 \hat{\rho} r+2)}{(n-1) (n+2)} \right) \\
\hat{\rho}({\bf y}) &= \frac{n + 2 - \sqrt{(n+2)^2 - 12 r^2 (n-1)} }{6 r} = \frac{(n+2)}{3 r} \left(1-\left(1-\frac{1}{n}\right) \left(\frac{\hat{\sigma}_j}{s_j}\right)^2 \right) .
\\
&= r \left( \frac{ n-1 } {n+2} \right) + O\left(n^{-1}\right)
\end{align*}
for $j= 1,2$. It is easy to show that the MML estimate of the correlation parameter is always smaller in magnitude compared to the sample correlation coefficient; ie, $|\hat{\rho}({\bf y})| < |r|$. Figure~\ref{fig:mserisk:rho} shows the mean squared error (MSE) risk for $n \in \{20, 50\}$ for the MML, maximum likelihood and the unique minimum variance unbiased estimator~\cite{OlkinPratt58}
\begin{equation*}
\hat{\rho}({\bf y})_{\text{Unbiased}} = r \, _2F_1\left(\frac{1}{2},\frac{1}{2};\frac{n-1}{2};1-r^2\right)
\end{equation*}
of the correlation parameter $\rho$; here, $_2 F_1(\cdot)$ is Gaussian hypergeometric function. When $n=20$, the MML estimate dominates the maximum likelihood estimate in terms of MSE risk for all $|\rho| < 0.64$. As $|\rho| \to 1$, all three estimates are virtually indistinguishable. When $n=50$, the MML estimate again dominates the maximum likelihood estimate for all $|\rho| < 0.6$.
\begin{figure*}[tbh]
\begin{center}
\subfigure[$n=20$]{
\includegraphics[height=6cm]{figs/mseriskn20.pdf}
}%
\subfigure[$n=50$]{
\includegraphics[height=6cm]{figs/mseriskn50.pdf}
} \\
\end{center}
\caption{Expected mean squared error for the MML, maximum likelihood and the unbiased minimum variance estimators of the correlation parameter $\rho$. \label{fig:mserisk:rho}}
\end{figure*}
\begin{table*}[tbph]
\scriptsize
\begin{center}
\begin{tabular}{ccccccccccccc}
\toprule
$\rho_0$ & $\rho$ & $n$ & \multicolumn{7}{c}{Relative frequency of rejecting $H_0: \rho = \rho_0$} & & \multicolumn{2}{c}{KL Divergence} \\
& & & $\text{DB}^S$ & $\text{DB}^M$ & Fisher Z & HP & Mud & Rud & MML & ~ & MLE & MML \\
\cmidrule{1-13}
\multirow{8}{*}{-0.30} & \multirow{2}{*}{-0.75} & 15 & 0.730 & {\bf 0.812} & 0.704 & 0.730 & 0.702 & 0.679 & 0.570 & & {\bf 0.408} & 0.414 \\
& & 30 & 0.921 & {\bf 0.950} & 0.936 & 0.938 & 0.936 & 0.932 & 0.890 & & {\bf 0.129} & 0.145 \\
& \multirow{2}{*}{-0.30} & 15 & 0.073 & 0.124 & 0.060 & 0.068 & 0.059 & 0.058 & {\bf 0.050} & & 0.262 & {\bf 0.230} \\
& & 30 & 0.043 & 0.081 & 0.046 & 0.050 & 0.046 & 0.046 & {\bf 0.035} & & 0.132 & {\bf 0.124} \\
& \multirow{2}{*}{0.00} & 15 & 0.236 & {\bf 0.329} & 0.184 & 0.206 & 0.178 & 0.193 & 0.245 & & 0.229 & {\bf 0.190} \\
& & 30 & 0.354 & {\bf 0.465} & 0.371 & 0.382 & 0.371 & 0.382 & 0.358 & & 0.092 & {\bf 0.083} \\
& \multirow{2}{*}{0.50} & 15 & 0.908 & {\bf 0.934} & 0.883 & 0.893 & 0.881 & 0.887 & 0.904 & & 0.278 & {\bf 0.223} \\
& & 30 & 0.995 & {\bf 0.998} & 0.995 & 0.996 & 0.995 & 0.995 & 0.994 & & 0.105 & {\bf 0.097} \\
\vspace{-2mm} \\
\cmidrule{2-13}
\vspace{-2mm} \\
\multirow{8}{*}{0.00} & \multirow{2}{*}{-0.75} & 15 & 0.969 & {\bf 0.988} & 0.953 & 0.960 & 0.950 & 0.950 & 0.938 & & 0.298 & {\bf 0.252} \\
& & 30 & 0.998 & 0.998 & 0.998 & 0.998 & 0.998 & 0.998 & {\bf 0.999} & & 0.106 & {\bf 0.099} \\
& \multirow{2}{*}{-0.30} & 15 & 0.262 & {\bf 0.362} & 0.207 & 0.224 & 0.204 & 0.204 & 0.198 & & 0.277 & {\bf 0.237} \\
& & 30 & 0.352 & {\bf 0.476} & 0.356 & 0.368 & 0.356 & 0.356 & 0.322 & & 0.126 & {\bf 0.118} \\
& \multirow{2}{*}{0.00} & 15 & 0.074 & 0.130 & 0.055 & 0.060 & 0.055 & 0.055 & {\bf 0.051} & & 0.217 & {\bf 0.189} \\
& & 30 & 0.050 & 0.106 & 0.053 & 0.057 & 0.053 & 0.053 & {\bf 0.037} & & 0.087 & {\bf 0.080} \\
& \multirow{2}{*}{0.50} & 15 & 0.613 & {\bf 0.700} & 0.559 & 0.580 & 0.554 & 0.548 & 0.508 & & 0.330 & {\bf 0.277} \\
& & 30 & 0.809 & {\bf 0.893} & 0.820 & 0.057 & 0.053 & 0.053 & 0.791 & & 0.126 & {\bf 0.120} \\
\vspace{-2mm} \\
\cmidrule{2-13}
\vspace{-2mm} \\
\multirow{8}{*}{0.70} & \multirow{2}{*}{-0.75} & 15 & {\bf 1.000} & 1.000 & 1.000 & 1.000 & 1.000 & 1.000 & 1.000 & & 0.273 & {\bf 0.232} \\
& & 30 & {\bf 1.000} & 1.000 & 1.000 & 1.000 & 1.000 & 1.000 & 1.000 & & 0.106 & {\bf 0.098} \\
& \multirow{2}{*}{-0.30} & 15 & 0.984 & 0.987 & 0.985 & 0.986 & 0.985 & 0.986 & {\bf 0.992} & & 0.264 & {\bf 0.221} \\
& & 30 & {\bf 1.000} & 1.000 & 1.000 & 1.000 & 1.000 & 1.000 & 1.000 & & 0.105 & {\bf 0.096} \\
& \multirow{2}{*}{0.00} & 15 & 0.834 & 0.875 & 0.850 & 0.865 & 0.849 & 0.867 & {\bf 0.910} & & 0.242 & {\bf 0.207} \\
& & 30 & 0.992 & 0.998 & 0.998 & 0.998 & 0.998 & {\bf 0.999} & 0.993 & & 0.104 & {\bf 0.095} \\
& \multirow{2}{*}{0.50} & 15 & 0.131 & 0.182 & 0.150 & 0.167 & 0.149 & 0.173 & {\bf 0.258} & & 0.328 & {\bf 0.287} \\
& & 30 & 0.275 & 0.397 & 0.403 & 0.412 & 0.401 & {\bf 0.417} & 0.354 & & {\bf 0.176} & 0.178 \\
\vspace{-3mm} \\
\bottomrule
\vspace{+1mm}
\end{tabular}
\caption{Relative frequency of rejecting the null hypothesis and Kullback--Leibler (KL) divergence from the data generating model for maximum likelihood (MLE) and minimum message length (MML) estimates. \label{tab:results:corr}}
\end{center}
\end{table*}
\section{Discussion}
\label{sec:sim}
An advantage of MML over alternative approaches is that it allows conceptually straightforward comparison of models with different structures and parameters as long as we can compute their codelengths. The codelength is measured in the unit of information (e.g., bits) and is therefore a universal yardstick for model comparison. It is then easy to extend the MML $t$-test to, for example, handle the case of unequal variances in the two groups, a version of the so-called Behrens--Fisher problem. The setup is now
\begin{equation}
Y_{1i} \sim \mathcal{N}\left(\mu_1, \sigma_1^2\right), \quad Y_{2j} \sim \mathcal{N}\left(\mu_2, \sigma_2^2\right) ,
\end{equation}
for $i=1,\ldots,n_1$ and $j = 1, \ldots, n_2$ and the task is, given data ${\bf y}$, to determine whether the two means are equal $\mu_1 = \mu_2$. MML87 codelengths for this setup were derived in \cite{MakalicSchmidt13} and can be compared to the codelengths of the null and alternative hypotheses under the equal variance assumption derived in Section~\ref{sec:mml:ttest} to discriminate between the four candidate models. Of course, we can also include other models in the comparison, as long as we can compute the corresponding codelengths.
\subsection{Minimum description length}
Minimum description length (MDL)~\cite{Rissanen78,Rissanen96,RissanenRoos07a,Rissanen09b,GrunwaldRoos19,BruniEtAl22} is a closely related approach to inductive inference that was developed around the same time as MML by Rissanen and colleagues and has gone through various refinements. MDL seeks a model class $\mathcal{M}$ that results in the shortest codelength of the data (ie, a model that best compresses the data), where a model class is defined as a set of parametric distributions indexed by parameter $\bm{\theta}$:
\begin{equation*}
\mathcal{M} = \{ p(\cdot | \bm{\theta}): \bm{\theta} \in \Theta\} .
\end{equation*}
A popular version of MDL is the general normalized maximum likelihood (NML) code whose codelength is given by
\begin{equation}
-\log p_{\text{NML}}({\bf y} | v, \mathcal{M}) = \underbrace{-\log \left[\max_{\bm{\theta} \in \Theta} p({\bf y} | \bm{\theta}({\bf y}), \mathcal{M}) \, v(\bm{\theta}) \right] }_{\rm model\; fit} + \underbrace{\log \left[\sum_{{\bf x}} \max_{\bm{\theta} \in \Theta} p({\bf x} | \bm{\theta}({\rm x}), \mathcal{M}) \, v(\bm{\theta}) \right]}_{\rm parametric\; complexity}
\end{equation}
where $v: \Theta \to [0, \infty)$ is a user-specified function, referred to as the \emph{luckiness} function in MDL terminology, and the sum (integral) is over the entire data space in the case of discrete (continuous) data. This is the general NML codelength for data ${\bf y}$ with respect to some model class $\mathcal{M}$ parameterised by models $\bm{\theta} \in \Theta \in \mathcal{M}$. We observe that the NML codelength is a sum of two terms:
\begin{enumerate}
\item the (penalized) negative log-likelihood of the data evaluated at the maximum, and
%
\item the so-called parametric complexity of the model class $\mathcal{M}$ which measures how well models $\bm{\theta} \in \mathcal{M}$ approximate random data sequences.
\end{enumerate}
The first term in the NML codelength represents how well the model $\hat{\theta} \in \mathcal{M}$ fits the observed data with smaller codelengths indicating better data fit. In contrast, a low (high) parametric complexity implies that a few (many) data sequences can be modelled sufficiently well by models in class $\mathcal{M}$. The parametric complexity denotes the logarithm of the number of distinguishable distributions in the model class~\cite{Balasubramanian05} and is similar in spirit to the assertion in the MML code, accounting for the complexity of the model class from which the maximum likelihood estimate is derived. A detailed discussion of the NML distribution and the corresponding optimality properties is available in~\cite{Rissanen01,Rissanen07,GrunwaldRoos19}.
Originally, Risannen advocated the use of the NML distribution for $v(\theta) = 1$ only. Unfortunately, the parametric complexity in this case is easily shown to be infinite for many model classes rendering a straightforward application of the NML codelength impossible. Subsequently, Rissanen and colleagues suggested other forms of the luckiness function as well as variations of the NML code such as the restricted approximate normalized maximum likelihood (ANML), the two-part ANML or the objective Bayesian code, among others~\cite{RooijGrunwald06,GrunwaldRoos19}. \\
\noindent {\bf Example:} Assuming $v(\theta) = 1$, the NML codelength for the binomial distribution is
\begin{equation}
-\log p_{\text{NML}}(y | \mathcal{M}) = -\log \binom{n}{y} \hat{\theta}^y (1-\hat{\theta})^{n-y}+ \log \mathcal{C}(n)
\end{equation}
where
\begin{equation}
\mathcal{C}(n) = \sum_{x = 0}^n \binom{n}{x} \left( \frac{x}{n}\right)^{x} \left(\frac{n - x}{n}\right)^{n - x}
\end{equation}
and $\hat{\theta}(y) = (y/n)$ is the maximum likelihood estimate of the success probability. An efficient linear-time algorithm to compute the parametric complexity of a general multinomial distribution is available in~\cite{KontkanenMyllymaki07b}. As an example, if $n = 10$ and $y = 3$, the maximum likelihood estimate is $\hat{\theta}(y) = 0.3$, the parametric complexity is $\log \mathcal{C}(n) \approx 2.22$ bits and the total NML codelength is approximately 4.13 bits.
For the original NML distribution with $v(\theta) = 1$, Rissanen~\cite{Rissanen96} showed that
\begin{equation}
-\log p_{\text{NML}}({\bf y} | \mathcal{M}) = -\log p({\bf y} | \hat{\bm{\theta}}({\bf y}), \mathcal{M}) + \log \int_\Theta \sqrt{|J_1(\bm{\theta}) | } + \frac{p}{2} \log \left( \frac{n}{2\pi} \right) + o(1)
\end{equation}
where $J_1(\cdot)$ is the per-sample Fisher information matrix. Additionally, Rissanen proved that the NML codelength reduces to the well-known Bayesian information criterion (BIC) in the limit as the sample size $n \to \infty$ and demonstrated that the asymptotic approximation is accurate for large sample sizes. Other, somewhat shaper, approximations to this NML codelength exist, see for example, Mera et al.~\cite{MeraEtAl2020}. Rissanen's asymptotic expansion was recently generalized to include arbitrary luckiness functions and a novel methodology for evaluation of the general, non-asymptotic NML codelength for exponential family models~\cite{SuzukiYamanishi18}.
Although there are many similarities between MML and MDL, there exist some important differences in their inherent inference philosophies. Unlike MML, MDL is non-Bayesian and does not allow for the use of prior information in inference. While MML provides new means of parameter estimation, MDL was originally based on the maximum likelihood estimator and aimed to discover the best model class, rather than the best single model, for a given data set. However, the recent inclusion of luckiness functions in the NML code allows for new MDL-based penalized maximum likelihood estimation procedures. Lastly, the NML codelength is not a two-part code as Strict MML and is instead derived to minimize the worst-case excess codelength relative to an ideal code. For a more detailed discussion of MML and MDL, we recommend~\cite{BaxterOliver94} and \cite{Wallace05} (pp. 413--415).
\bibliographystyle{amsplain}
| ec550a2e6e012440d85c30d6a93c7cca35ba7970 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Literature we should mention/cite}
Quantum computation is a disruptive technology that has irrevocably changed the way that computation is envisioned. It holds the potential for addressing a wide range of computational challenges~\cite{Fedorov2022} from factoring~\cite{shor1994algorithms} and database search~\cite{grover1996fast} to finance~\cite{Herman2022} and machine learning~\cite{Cong2019} problems much more efficiently than what can be achieved by classical machines. Yet, current noisy intermediate-scale quantum (NISQ) devices~\cite{Preskill2018} are still far from addressing these applications at a practically relevant scale and early demonstrations of quantum advantages remain confined to algorithms that do not yet have clearly identifiable broad applications~\cite{arute2019quantumshort, zhu2021quantumshort}. The primary obstacle for today's quantum computers, which now feature 10s to 100s of qubits \cite{Bernien-Lukin2017, Zhang-Monroe2017, Otterbach-Rigetti2017, WangEtAl2018, FriisMartyEtal2018, BradleyEtAl2019, PogorelovEtAl2021, MooneyWhiteHillHollenberg2021b} remains noise and decoherence, which limits the number of entangling operations and therefore the achievable circuit depth.
Quantum error correction (QEC) promises to overcome these obstacles and enable fault-tolerant quantum information processing, yet only at the cost of significant overheads in experimental complexity. Despite impressive progress~\cite{NiggEtAl2014, Barends-Martinis2014, ChouEtal2018, HarperFlammia2019, HuEtAl2019, ErhardPoulsenNautrup2021,Bermudez2017,Postler2021,Ryan-Anderson21}, QEC proves to be an extremely challenging goal to realise and its cost is highly dependent on the performance of the underlying hardware and implementation. Hence, it is imperative, not just for near-term, but also for future large-scale quantum devices, to maximise the efficiency of quantum information processing and use available resources to their full extent.
Fortunately, there is a lot of unused potential in current quantum devices, which tend to use only a small fraction of the available Hilbert space. Indeed, control over the inherently high-dimensional Hilbert space has been demonstrated in all major quantum technology platforms~\cite{Wang2018,Morvan2020,Ringbauer2021,Chi2022}, motivating the exploitation of a new paradigm of quantum computing based on $d$-dimensional \emph{qudits}, rather than qubits. Compared to their two-level counterpart, qudit architectures offer much richer coherence~\cite{Ringbauer2017} and entanglement structures~\cite{Kraft2018}, which can be exploited for efficient quantum information processing~\cite{lanyon2009simplifying,nikolaeva2021efficient,KiktenkoNikolaevaXuShlyapnikovFedorov2020,WangHuSandersKais2020} and improved QEC~\cite{Watson2012,Campbell2014}. Since entangling operations tend to be the bottleneck in current quantum devices, the efficiency of a quantum computation, or the complexity of a quantum circuit, is traditionally measured by counting the number of entangling operations. While this is clearly an incomplete picture, it serves as a good hardware-agnostic approximation, because the optimal circuit for a given quantum operation is elusive and highly dependent on the available gate set.
Here, we investigate \emph{qudit circuit compression}, as a way to simplify a given qubit circuit by rephrasing it as a qudit circuit. We achieve this in two steps: First, the qubits are partitioned into groups of equal size such that the number of gates within the groups is maximized. Each group is then interpreted as a qudit, turning entangling gates into local gates and thus reducing the overall entangling gate count by a combinatorial factor for which we find lower and upper bounds using a graph-based approach. Second, by considering an extended gate set including not only qubit-entangling gates, but also genuine qudit-entangling gates, the number of entangling gates in the resulting qudit circuit can be further reduced, even saturating the combinatorial lower bound. To showcase these two effects, we study the compression of exemplary qubit circuits under different gate sets.
We illustrate this gate compression with experimental details for two contemporary quantum technologies using qudits: photonic qudits encoded in orbital angular momentum and trapped ions with multiple addressable levels, showing that, already today, qubit circuits can be more efficiently compiled on qudit architectures.
Aside from algorithmic improvements, reducing the number of quantum information carriers tends to make the system experimentally easier to control, leading to improved performance. Furthermore, any reduction in entangling-gate count for a given application directly lowers the resource requirements for quantum error correction, thus greatly benefiting future large-scale quantum devices.
We provide general design principles, as well as upper and lower bounds for the possible reduction in the number of gates, exemplified with explicit constructions for photonic systems and trapped ions.
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth,trim={0cm 8cm 0cm 0cm},clip]{Combined_figure.pdf}
\caption{\textbf{Gate compression}.
(a) \& (b)
Compression of quantum circuits of $N$-qubit controlled phase-flip (CPF) gates $C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(N)}$. (a) The best known \cite{barenco1995elementary} decomposition of a $4$-qubit CPF gate $C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(4)}$ in terms of two-qubit gates requires $13$ entangling gates, i.e., $6$ CNOTs and $7$ controlled $T$ (or $T^{\dagger}$) gates (CT) with $T=\operatorname{diag}\{1,\exp(i\pi/4)\}$. By compressing the circuit to two qudits of dimension $4$, two quarts, only $7$ entangling gates are required in terms of embedded two-qubit gates, while $6$ previously nonlocal gates become local. However, the same operation $C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(4)}$ could also be realized by a single controlled-$C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(2)}=$CZ$=\operatorname{diag}\{1,1,1,-1\}$ gate on two ququarts, carried out if ququart $1$ is in the state $\ket{d-1}=\ket{3}$ corresponding to the two-qubit state $\ket{11}$.
(b) Similarly, the most efficient decomposition of $C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(6)}$ for $6$ qubits requires $61$ two-qubit gates~\cite{barenco1995elementary}, whereas the compression to $3$ ququarts ($d=4$) requires (at most) $9$ (native) two-ququart gates, including adjoints and powers of controlled CT gates as well as controlled $\Pi$ gates, where $\Pi=\sum_{n=0}^{3}\ket{(n-1)_{\mathrm{mod}(3)}}\!\bra{n}$ is a permutation. Further compressing to two quocts ($d=8$), as little as $1$ controlled-$C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(3)}$ gate, conditioned on the quoct computational-basis state $\ket{d-1}=\ket{7}$ may be required. In both (a) and (b) the final decomposition of the $N$-qubit CPF gate into a single controlled-$C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(N/2)}$ gate requires a non-factorizable two-level entangling gate.
(c) \& (d)
Illustration of the effect of gate compression for the creation of a (quadratic) 2D cluster state consisting of four (c) and nine qubits (d), respectively. Cluster states are a particular type of graph state that can be schematically represented by vertices (here shown as circles carrying the local subsystem dimension $d$ as labels, 2 for qubits, 4 for ququarts, etc.) connected by lines representing entangling gates (controlled-$Z$ gates for qubits). By grouping (blue boxes) the qubits into pairs (c) or triples (d) the subsystem dimension becomes $d=4$ (ququart) and $d=8$ (quoct), respectively, but the number of non-local entangling gates can be reduced (c) from $4$ two-qubit gates to either $2$ embedded two-qubit gates or even $1$ single-qudit gate, and (d) from $12$ two-qubit gates to $6$ embedded two-qubit gates, $4$ two-qudit gates, or $2$ two-quoct gates, as is explained below.
}
\label{fig1:combinedfigure}
\vspace{-0.2cm}
\end{figure*}
\emph{Circuit compression.-- }
Quantum circuits are built from a sequence of gate operations. At the lowest level of abstraction, circuit compression is hence concerned with compiling an $N$-qubit unitary to an $M$-qudit architecture ($M<N$), which we call \emph{gate compression}, see Fig.~\ref{fig1:combinedfigure}.
The task is to encode the qubit circuit into the qudit architecture in such a way that the maximal number of entangling gates in the qubit circuit manifests as local gates in the resulting qudit circuit. Importantly, the remaining entangling gates retain their qubit-entangling structure, in terms of maximally generated entanglement entropy, when embedded in the qudit Hilbert space. Consequently, this procedure always reduces the amount of entanglement needed (irrespective of the available gate set for qudits), by compressing non-local gates into local ones. Curiously, however, while embedded qubit gates still create two-level entanglement (i.e., equivalent to a two-qubit gate in terms of entanglement entropy), they are not necessarily easily implementable in a qudit architecture. This is easy to see for the example of a CNOT gate $U\subtiny{0}{0}{\mathrm{CNOT}}\suptiny{1}{0}{(c,t)}$ applied to qubits 2 (control) and 3 (target) in a four-qubit register, taking the form $\mathds{1}_1 \otimes\, U\subtiny{0}{0}{\mathrm{CNOT}}\suptiny{1}{0}{(2,3)} \otimes \mathds{1}_4$. When qubits 1 and 2 are encoded in the first qudit, and 3 and 4 in the second, the resulting embedded version of the original gate would have to be a subspace-agnostic operation (i.e., applying the same operation on both subspaces pertaining to the encoded qubits 1 and 4). While such operations still only create two-level entanglement, they are often not available natively. On the other hand, the qudit encoding enables new kinds of two-level entangling operations, which do not admit a tensor-product structure in the corresponding qubit circuit. Such gates provide new powerful tools for circuit compression, as discussed below, while also emphasising the importance of the available gate set for qudit architectures.
While finding the ideal embedding is hard in general, it is easy to visualise using graphs:
Given a qubit circuit, we want to find the optimal partition into qudits by encoding the non-local gates of the circuit into a weighted graph and applying graph partitioning.
A weighted graph $G=(V,E)$ is a pair of a set of vertices $V=\{v_1,\dots,v_n\}$ and a set of edges $E=\{e_1,\dots,e_k\}$, where each edge $e_i= ((v_l,v_m),w_i)$ is a pair of vertices together with a weight $w_i$.
We can encode the non-local properties by identifying each vertex with a qubit and drawing an edge whenever a non-local gate is present.
The weight is given by the number of non-local gates between each pair of qubits.
\emph{Graph partitioning} describes a family of several different partitioning problems.
However, we are interested in the following variation of the problem:
Given a weighted graph $G$, find a partition of $V=\bigcup_i V_i$, under the condition that each subset $V_i$ has the same cardinality such that the sum of weights of the edges connecting these subsets is minimal.
Graph partitioning is also a hard problem, however, there exist several heuristic algorithms which find solutions at least for a small number of vertices.
\emph{Gate set optimisation.-- }
After we have reduced the width of the circuit we can further improve upon the circuit by considering the extended gate set we can access for qudits. After the partitioning, previously non-local gates within a partition now act as local gates for individual qudits, while non-local gates between partitions still generate two-level entanglement between different qudits. Depending on the resulting entanglement structure, the experimental complexity can be further reduced by considering more powerful entangling gates than just embedded qubit gates.
\emph{Upper and lower bounds.-- }
Using the above principles, we can derive upper and lower bounds on the \emph{compression ratio} $C_d$ as the ratio of the number of entangling gates in the compressed circuit to the number of gates in the original circuit. Consider the vector $\vec{\tilde w} \in \mathds{Z}^{M\choose 2}$ of weights $\vec{\tilde w}=(\tilde w_1,\dots, \tilde w_{M\choose 2})^T$ corresponding to the qudit graph for the compressed circuit. This vector is obtained from the weights $w_i$ of the qubit circuit by summing those that correspond to qubits encoded in the same qudit. The upper bound is then given by the sum of compressed weights, representing the total number of entangling gates remaining after compression. The lower bound, on the other hand, is given by the number of non-zero weights, representing the number of entangling links between the qudits.
\begin{equation}
\frac{\norm{\vec{\tilde w}}_0}{\norm{\vec{w}}_1} \leq C_d \leq \frac{\norm{\vec{\tilde w}}_1}{\norm{\vec{w}}_1}
\end{equation}
The span between the lower and upper bounds indicates potential improvements based on an appropriate gate set.
In the following, we will provide examples saturating both the lower and the upper bound.
\emph{Controlled phase gates.-- }
As a key example, we will now study $N$-qubit controlled phase-flip gates $C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(N)}$ (i.e., phase gates with a $\pi$ phase shift). These gates are central elements in quantum computing, for example, as a key part of Grover's search algorithm \cite{grover1996fast}.
It can be written as follows,
\begin{equation}
\begin{aligned}
C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(N)} = \mathds{1}^{\otimes N}+\ket{1}\!\bra{1}^{\otimes (N-1)}\otimes(Z-\mathds{1}),
\label{eq:equation}
\end{aligned}
\end{equation}
where $\mathds{1}$ is the identity operator and $Z$ denotes the single-qubit Pauli-$z$ gate.
Due to its central role, the decomposition of $C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-2}{(N)}$ is often used as a benchmark for comparing circuit decompositions and gate sets. Importantly, provably optimal decompositions into two-qubit gates and local operations are known for 3 to 8 qubits~\cite{barenco1995elementary}. As we noted above, the optimality of a quantum circuit depends on the metric used. However, even when the metric is fixed, finding an efficient way to realize the circuit is highly non-trivial.
For example, for a four-qubit controlled phase flip $C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(4)}$, the most efficient decomposition known, believed to be optimal, requires 13 two-qubit gates~\cite{barenco1995elementary}, as shown in Fig.~\ref{fig1:combinedfigure}~(a). The same circuit can be realized in a qudit system by embedding two qubits each into two ququarts. There are 3 ways to achieve such a partition, and the best choice turns 6 entangling gates into local gates, leaving 7 embedded two-qubit gates between the two ququarts. If we consider also ququart gates in our gate set, this can further reduce the nonlocal-gate count to just a single gate between the two ququarts, which turn out to generate only two-level entanglement, as shown in Fig.~\ref{fig1:combinedfigure}~(a).
In case $N>4$, we can cut $N$ qubits into $k$ equal parts by using a $2^{N/k}$-dimensional qudit in each part and achieve similar improvements.
For the six-qubit controlled phase-flip gate ($C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(6)}$), for example, 61 two-qubit gates are needed in the best known (again, believed to be optimal) decomposition~\cite{barenco1995elementary}. Encoding the same circuit into 3 ququarts reduces the requirement to 9 two-ququart gates, including four two-ququart controlled-shift gates and five two-ququart controlled phase gates \cite{bullock2006efficient}, see Fig.~\ref{fig1:combinedfigure}~(b). Going further by encoding the gate into two qudits of dimension $d=8$ (quocts), the gate can again be realized with a single two-level entangling qudit gate.\\[-1mm]
\emph{Graph states and the saturation of graph partitioning bounds.--}
Graph states are another highly relevant example. For a set of edges $E=\{e\}$, they can always be created by applying controlled-$Z$ gates across all edges on computational-superposition product states, i.e., $|G\rangle:=\prod_{e\in E} CZ_e|+\rangle^{\otimes n}$ with $|+\rangle=\bigl(|0\rangle+|1\rangle\bigr)/\sqrt{2}$. As all $CZ_e$ commute, it directly follows that, with an unrestricted entangling power of the gate set, the lower bound $\norm{\vec{\tilde w}}_0$ can be saturated. If the gate set is restricted to two-level entangling gates, however, the fact that a high amount of entanglement may be generated if multiple edges cut across a bipartition directly implies that the upper bound needs to be observed with $\norm{\vec{\tilde w}}_1 $ embedded qubit gates.
Consider, for example, a 4-qubit quadratic 2D cluster state as illustrated in Fig.~\ref{fig1:combinedfigure}~(c). Due to the symmetry of the state the partition is irrelevant and we can thus combine the first and second qubit as well as the third and fourth qubit to a qudit each without loss of generality. This reduces the number of non-local gates from $4$ to $2$, which saturates the upper bound $\norm{\vec{\tilde w}}_1$. Now, we can use gate optimization to reduce the number to $1$ by combining the two non-local gates between the qudits to a single genuine qudit gate, saturating the lower bound $\norm{\vec{\tilde w}}_0$. This example demonstrates that the commutativity of the entangling gates across compressed partitions as well as the maximum amount of entanglement generated across these partitions are crucial parameters that determine how well a qubit circuit can be compressed when compiled on a qudit architecture and which gate set would be required for the qudits.
\emph{Photonic implementation.--}
Photonic systems are excellent candidates for gate compression. Current developments enable increased control over higher-dimensional degrees of freedom by manipulating a photon's polarisation, spatial profile, temporal profile, or frequency, either separately or simultaneously. This enables the encoding of multiple bits into a single photon, as is routinely done in entanglement-based quantum communication \cite{Sit:17,AchatzEtAl2022,Herrera-ValenciaSrivastavPivoluskaHuberFriisMcCutcheonMalik2020}. Local unitary operations are easily done within a certain degree of freedom, such as spatial manipulation through multi-plane light conversion \cite{Brandt:20}, frequency manipulation \cite{kues2017chip,CabrejoPonceMarquesMunizHuberSteinlechner2022}, or between different degrees of freedom, for example, using polarizing beam splitters to couple path and polarization. For instance, high-dimensional Pauli $X$- and $Z$-gates, which are parts of higher-dimensional universal gate sets, have recently been implemented in a number of ways \cite{asadian2016heisenberg, babazadeh2017high, wang2015quantum, gao2019arbitrary, Brandt:20}. While local gates can often be performed with near unit efficiency and fidelity, entangling gates remain the Achilles heel of photonic information processing, as entangling two photons can be achieved probabilistically at best, leading to an exponential decrease in success probability with the number of entangling gates. However, recent developments show promising results \cite{muthukrishnan2000multivalued, erhard2018twisted}, including the implementation of high-dimensional multi-partite quantum gates \cite{gao2019computer} and an implementation of the SUM gate (a high-dimensional controlled-$X$ gate) in the time and frequency degrees of freedom of photons \cite{imany2019high}.
\emph{Trapped-ion implementation.--} Trapped ions are among the leading platforms for quantum information processing~\cite{Bermudez2017}, where the electronic energy levels of each ion naturally provide a high-dimensional Hilbert space. Recently it was shown that such a system can be operated as a universal qudit quantum processor up to dimension 7~\cite{Ringbauer2021}.
The qudit-gate set used in this demonstrations consisted of arbitrary local gates and two-qubit CNOT gates embedded in a qudit Hilbert space. Compared to a standard qubit CNOT, the embedded version exhibits error rates larger by roughly a factor of 2, independent of the Hilbert-space dimension. Beyond this basic gate set, it has been shown that both dominant gate mechanisms in this platform, the M{\o}lmer-S{\o}rensen gates~\cite{Low2020} and light-shift gates~\cite{Hrmo2022}, can be generalized to achieve genuine qudit entanglement. A first experimental realization of the latter demonstrated the generation of genuine qudit entanglement in a scalable fashion and with highly competitive error rates~\cite{Hrmo2022}.
Compiling the example of Fig.~\ref{fig1:combinedfigure}~(a) and using state-of-the-art error rates for trapped-ion quantum processors of about 0.01 per qubit CNOT gate~\cite{Bermudez2017}, a rough estimate suggests that the implementation of the four-qubit $C\subtiny{0}{0}{\mathrm{PF}}\suptiny{1}{-1}{(4)}$ gate could achieve an error rate on the order of 0.12 with the standard two-level decomposition using 13 two-qubit entangling gates. Curiously, while it is known on the one hand that enlarging the Hilbert space locally (i.e., encoding 4 qubits into 4 qudits) can reduce the required number of gates quadratically~\cite{lanyon2009simplifying}, this gain is offset almost exactly in the 4-qubit case by the factor of 2 increased error rates incurred in the experimental implementation~\cite{Ringbauer2021}. On the other hand, when two qubits each are encoded in a qudit of dimension 5 (4 computational levels and 1 auxiliary level), the required number of two-level entangling gates drops to 1 (albeit with a rotation angle equivalent to 4 qubit gates), achieving an estimated error rate of 0.04, see Appendix~\ref{sec:appendix B} for the circuit. Curiously, this is an example where it is optimal to use a two-level entangling gate in the qudit circuit, which is not an embedded qubit gate.
\emph{Conclusion.--}
We have explored two ways in which higher-dimensional architectures are universally beneficial to quantum computing. First, using gate compression, we can cut down on the amount of entanglement needed for a specific $N$-qubit circuit, and second, by exploiting the added capabilities of qudit systems in the form of richer gate sets, we can further reduce the number of non-local gates required to generate that entanglement. This highlights the key role played by entanglement and the available gate set in efficient qudit QIP.
As every higher-dimensional gate can be achieved by any universal gate set applied a number of times, the advantage from larger gate sets is constant in the number of qubits. Similarly, even the most efficient partitioning of the qubit circuit leads to only a constant advantage. Such constant improvements, however, can make the difference between feasibility and failure.
Another important aspect is the potential breakdown of conventional wisdom regarding easily implementable local gates versus hard non-local gates. Once qudit dimensions get sufficiently large, the performance gap between local and non-local gates might change. Moreover, gate performance typically degrades somewhat with system size, providing another motivation for reducing the number of quantum-information carriers and making more efficient use of available resources. Finally, we emphasize that a case-by-case evaluation of the actual trade-offs is critical to finding the optimal dimensionality for a given problem and hardware platform.
Our examples provide a promising first step, showing that qudit encodings can lead to a significant reduction in gate count. Hence, this approach can greatly increase the utility of current and future quantum hardware, using only degrees of freedom that are already present in today's quantum technology. Indeed, various quantum-computing platforms have demonstrated qudit control with ever-increasing performance.
Both gate compression and gate-set optimisation will be central tools for making the most of the next generation of high-dimensional quantum processors, harnessing the full potential of physical quantum information carriers.
\emph{Acknowledgements.--}
We are indebted to Mateus Ara\'ujo for discussion and input during early stages of this work, and we thank him for his Krapfen.
We acknowledge support from the Austrian Science Fund (FWF) through the START project Y879-N27 and the project P~31339-N27. This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sk{\l}odowska-Curie grant agreement No 840450. X.G. acknowledges the support of Joint Centre for Extreme Photonics (JCEP). M.H. acknowledges funding from the European Research Council (Consolidator grant ’Cocoquest’ 101043705).
\bibliographystyle{apsrev4-1fixed_with_article_titles_full_names_new}
| 8df772e11250df6fa084fc7d2ed7a5219f5f8577 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The detection of gravitational waves (GWs) from black hole mergers gave birth to GW astronomy \citep{abbot16PRL}, while the observation of neutron star merger in gravitational and electromagnetic waves is the prime example of the multi-messenger astronomy \citep{abbott17PRL}. One of the most promising multi-messenger sources that are yet to be detected in GWs are core-collapse supernovae (CCSNe).
CCSNe are the powerful explosions of massive stars at the end of their life. The collapse of the stellar core releases $\sim\! 10^{53}$ erg of its gravitational binding energy. While most of this escapes in the form of neutrinos, the rest powers the explosion \citep{Bethe:1990mw}. Supernovae involve powerful aspherical flows that generate GWs with energies up to $\sim\! 10^{47}$ erg \citep[e.g.,][]{ott:09review, kotake:13review, kotake:17}. The explosion front breaks out of the stellar surface hours later \citep{waxman17}, producing a blast of photons across the electromagnetic spectrum \citep[e.g,][]{Nakamura:2016kkl}.
The supernova dynamics, and thus the GW signal, strongly depends on rotation. In non- or slowly rotating models, which represent the majority of CCSNe \cite[e.g.,][]{heger:05}, the explosion is governed by the \emph{neutrino mechanism} \citep[e.g.,][for recent reviews]{kotake:12snreview, janka:12a,mueller:20review, Burrows21review, mezzacappa22}. A fraction of neutrinos emitted by the protoneutron star (PNS) heats the post-shock material, giving rise to neutrino-driven convection \citep{herant:92, bhf:95, janka:95} and standing accretion shock instability (SASI) \citep{blondin:03, foglizzo:06, mueller:12b}. These flows perturb the PNS and excite its oscillations, generating powerful GWs \cite[e.g.,][]{murphy:09, mueller:13,cerda:13, yakunin15, andresen:17, Hayama18, radice:19gw, mezzacappa:20a, Raynaud22}. Asymmetries in neutrino emission also contributes to the GW signal \cite[e.g.,][]{Kotake07, takiwaki:18, Vartanyan20}. The GWs are detectable for sources within our Galaxy with current generation detectors \citep{gossan:16, abbott:20ccsn, Lopez21, Szczepanczyk21, Antelis22}, while the future generation detectors will enable a more detailed observation \citep{Srivastava:2019fcb}.
In rare rapidly rotating stars, the rotational kinetic energy powers the explosion via the \emph{magnetorotational mechanism} \citep{leblanc:70, bisno:76, akiyama:03, burrows:07a, moesta:14a, kuroda:20, obergaulinger:20, Raynaud20}. Due to the centrifugal force, the PNS forms with an oblate perturbation, triggering PNS ring-down oscillations that last for $\sim\! 10 $ ms \citep{ott:12, fuller:15}. In some cases, the PNS may be subject to non-axisymmetric instabilities \citep[e.g.,][]{ott:07cqg, scheidegger:10cqg, shibagaki:20, Takiwaki21}. When this happens, the non-axisymetrically-deformed PNS emits GW for many rotation periods, significantly enhancing the detectability \citep{abbott:20ccsn}. For moderate rotation, both the rotational bounce and convection contribute to the GW signal \citep{andresen19, Pan21, Jardine22}.
Once detected, it is possible to estimate the parameters of source using the GWs emitted by the source \citep[e.g.,][for a recent review]{Christensen22}. Since slowly and rapidly rotating models have significantly different dynamics, their GW signals can be confidently distinguished \citep{Logue12, Powell16, Chan20, Szczepanczyk21, Saiz-Perez22}. Both convection and SASI develop from stochastic perturbations, so the GWs coming from slowly rotating models contain stochastic components. In contrast, the bounce GW signal in rapidly rotating stars can be determined precisely for a given (physical and computational) model parameters \cite[e.g.,][]{zwerger:97, dimmelmeier:08}. Despite the presence of the stochastic contributions, the GW spectra contain the frequencies of the physical processes happening in the central regions \citep[e.g.,][]{Kotake11, mueller:13, Astone18, Roma19, Srivastava:2019fcb, Powell22}. In particular, the PNS oscillations is the dominant component of the signal \citep{murphy:09, cerda:13, morozova:18, radice:19gw, Warren20}, the frequencies of which can be related to the physical parameters of the system, such as the mass and radius of the PNS \cite[e.g.,][]{dimmelmeier:06, mueller:13, Torres-Forne19, Sotani21, Bizouard21}.
For rapidly rotating stars, \citet{abdikamalov:14} explored the possibility to infer the rotation and its distribution in the supernova core. They considered $\sim\!100$ different rotational configurations of a stellar model with 5 different degrees of differential rotation. They found that, for a source at $10$~kpc distance, it is possible to measure rotation with $\sim\! 20\% $ accuracy for rapidly rotating models, in which the rotational kinetic energy exceed $\sim\! 8\%$ of the potential binding energy. In slowly rotating models, due to smaller GW amplitudes and higher stochastic contribution, the error becomes larger. These estimates were further improved by \citet{Engels14, Edwards14}, and \citet{Afle21}. \citet{Hayama16} showed that signs of rapid rotation can be found in the circular polarization of the GW signal. \citet{Yokozawa15} proposed to combine GW and neutrino observations to infer the rotation from the time delays between the bursts of these two signals. \citet{pajkos19, pajkos21} proposed to combine the the core-bounce signal with the dominant frequency mode of the PNS in the pre-explosion to constrain the structure of the progenitor star.
\citet{richers:17} studied the dependence of the GW signal on the parameters of the equation of state (EOS) of high-density matter in protoneutron stars and treatment the electron capture rate during collapse. They find a modest impact of these parameters to the bounce and the early $\sim\! 20$ ms post-bounce GW signal. This emphasizes the importance of accurate modeling \citep[e.g.,][]{lentz:12a, Kotake18, Pan18, Pan19, mezzacappa20, Andresen21}. Using deep convolutional neural networks, \citet{Edwards21} classified these EOSs $72\%$ correctly, while their most probable five EOSs were found with $97\%$ accuracy \cite[see also][]{Chao22}.
In this work, we study if it is possible to probe the iron core mass of CCSN progenitor from the bounce and early ring-down GW signal alone. \citet{ott:12} showed that progenitors with different masses will produce similar GWs if they have a similar angular momentum distributions at a given mass coordinate in the stellar core. This suggest that the progenitor mass at most has a subtle effect on the bounce dynamics. We extend this work further by using machine learning (ML) for signal classification and studying a wider ranges of progenitors and rotational configurations. We look at idealized optimistic scenario for measuring mass from the bounce and early ring-down GW signal alone. First, we consider rapidly rotating models, which produce strong GWs. Second, we use the deleptonization method that is known to artificially amplify the differences between progenitors. Third, we limit ourselves to model progenitors with four different iron core masses only, which simplifies the selection process. Despite these ideally favorable conditions, we show that the iron core mass cannot be accurately measured from the GW bounce signal alone. This suggests that, to measure the mass, one has to incorporate additional information such as longer post-bounce signal or neutrino observations, or both.
\section{Methodology}
\subsection{Gravitational waveforms}
\label{sec:gw_data}
We consider four progenitor models with zero-age main sequence masses ranging from $12$ to $40$ $M_\odot$. At the pre-collapse stage, these models develop iron cores with masses ranging from $1.3M_\odot$ to $1.8M_\odot$, as shown in Table ~\ref{tab:progenitor_param}, which also provides the central densities and entropies. The $12M_\odot$ and $40M_\odot$ models are produced by \cite{woosley:07}, while the $15M_\odot$ model is evolved by \cite{heger:05} with magnetic field prescription by \cite{spruit:02}. The $27M_\odot$ model is produced by \cite{whw:02}. All models have solar metallicity.
To obtain the GW signals, we perform simulations using the {\tt CoCoNuT} code \citep{Dimmelmeier02a, Dimmelmeier02b, dimmelmeier:05MdM}. The simulations are continued until 25 ms after bounce. This time interval contains full bounce and ring-down signal from a~CCSN event \citep[][]{abdikamalov:22review}. We use the same computational setup as in \citet{richers:17} with the SFHo nuclear equation of state \citep{steiner:13b}. During collapse, our code uses the $Y_\mathrm{e}(\rho)$ deleptonization method of \cite{liebendoerfer:05}. Following \citet{richers:17}, we use $Y_\mathrm{e}(\rho)$ profile obtained from spherically symmetric radiation hydrodynamics simulations using the {\tt GR1D} code \citep{oconnor:15a}. This method is known to amplify the contrast between progenitors with different masses \citep{mueller:09phd, pajkos21}: it produces variations in the inner core mass at bounce of $\simeq 10\%$ between progenitors with different masses, whereas full neutrino-transport simulations yields inner core masses that are practically independent of the progenitor \citep{mueller:09phd}. However, this weakness of the method represents the strength of our work: as we show below, even with the artificial amplification, the difference in the GW signals is too small to be distinguishable. In more realistic models, the distinction should be even smaller.
Following \cite{abdikamalov:14}, for each progenitor we consider about $100$ different rotational profiles ranging from slow to rapid rotation, where rotation has little to large effect on the core dynamics, respectively. At distance $\varpi$ from the rotation axis, the angular velocity of pre-collapse models is given by
\begin{equation}
\label{eq:rot_law}
\Omega(\varpi) = \Omega_0 \left[ 1 + \left(\frac{\varpi}{A}\right)^2\right]^{-1},
\end{equation}
where $A$ is a measure of degree of differential rotation, $\Omega_0$ is the central angular velocity \citep{komatsu:89a}. As in \cite{abdikamalov:14}, we consider five different values of $A$, ranging from $300$ to $10,\!000$ km, which correspond to the limits of, respectively, extreme differential and uniform rotations in the stellar core. For a given $\Omega_0$ and $A$, we impose the rotation law (\ref{eq:rot_law}) to the $12M_\odot$ model, as described in \cite{abdikamalov:14}. We then map the specific angular momentum distribution homologously to the other progenitors with different masses. This ensures that the specific angular momentum at a given enclosed mass coordinate is similar in all four progenitors. As summarized in Table~\ref{tab:T1}, the total number of rotational configurations we consider is 97, 99, 102, and 104 for s12, s15, s27, and s40 models, respectively. The difference in this number is caused by the fact that at the extreme rapid rotation limit, due to a fine balance between gravity and centrifugal force in the pre-collapse stage, some of the models do not collapse for some progenitors and collapse for other.
We extract the GW waveforms from the simulations using the Newtonian quadrupole formula in the first moment of momentum density formulation \citep{Dimmelmeier02a, Dimmelmeier02b, dimmelmeier:05MdM}, which yields accurate waveforms for stellar core collapse simulations \citep{reisswig:11ccwave}.
\begin{table}
\setlength\extrarowheight{2.09pt}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
ZAMS mass & $M_\mathrm{IC}$ & $s_\mathrm{c}$ & $\rho_\mathrm{c}$ \\ [.01em]
[$M_\odot$] & [$M_\odot$] & [$k_\mathrm{B}/\mathrm{baryon}$] & [$10^9\,\mathrm{g/cm}^3$] \\ [.3em]
\hline
12 & 1.3 & $0.60$ & $12.0$ \\
15 & 1.4 & $0.62$ & $8.74$ \\
27 & 1.5 & $0.77$ & $4.96$ \\
40 & 1.8 & $1.06$ & $2.17$ \\
\hline
\end{tabular}
\end{center}
\caption{Progenitor model parameters used in our work. The ZAMS mass is the initial (the so-called zero-age) mass of the star. Parameter $M_\mathrm{IC}$ is the inner core mass, while $s_\mathrm{c}$ and $\rho_\mathrm{c}$ are the central entropy and density before collapse.}
\label{tab:progenitor_param}
\end{table}
\begin{table}
\setlength\extrarowheight{2.09pt}
\begin{center}
\begin{tabular}{c c}
\hline
Progenitor Mass & No. of Events \\ [.3em]
[$M_\odot$] & \\ [.3em]
\hline
12 & 97 \\
15 & 99 \\
27 & 102 \\
40 & 104 \\
\hline
Total & $402$\\
\hline
\end{tabular}
\end{center}
\caption{Number of the generated GW events grouped by their progenitor mass types.}
\label{tab:T1}
\end{table}
\subsection{Machine learning}
\begin{figure*}
\centering
\makebox[1.\textwidth]{\includegraphics[width=1.0\textwidth]{Plots/flowchart.pdf}}
\caption{Flowchart of the analysis pipeline of our study.}
\label{fig:flowchart}
\end{figure*}
Our analysis pipeline can be divided in the following steps:
\begin{itemize}
\item Preprocessing: Customize the waveforms for ML input.
\item Noise: Add noise to the signals to model the detector noise.
\item ML: Optimize hyperparameters\footnote{Hyperparameters are parameters of the algorithm, which are tuned to optimize the learning. Their values are chosen before initialising learning. They should not be confused with parameters of the data, which are obtained during the learning. The choice of hyperparameter has a direct impact on the performance of the ML model and hence they are optimized to produce the best performance.} and apply ML algorithm.
\item Analysis: Explore confusion matrix and other metrics.
\end{itemize}
A flowchart depiction of the pipeline is illustrated in Fig.~\ref{fig:flowchart}.
To make the waveforms suitable to ML analysis, we condense all candidate waveforms into a single data cube with the progenitor mass of each candidate treated as target class. We map each candidates to the same cadence along the time axis with uniform starting and ending points of $-5$ and $+15$ ms around the bounce time. The time of bounce corresponds to time equal to zero. We generate $10,000$ interpolated points along the time axis per candidate, resulting in a data cube of dimensions $($402$\times10000\times2)$.
\subsubsection{Random Forest}
\label{sec:rf}
\begin{table}
\setlength\extrarowheight{2.09pt}
\begin{center}
\begin{tabular}{|c | c| }
\hline
hyperparameters & Range \\ [.3em]
\hline \hline
No. of estimators & $[100,400]$ \\
Max. features & [auto, sqrt, log2] \\
Max. depth & $[10,250]$\\
Min. samples split & $[2,100]$ \\
min. samples leaf & $[2,100]$\\
Bootstrap & [True, False]\\
OOB score & [True, False]\\
Warm start & [True, False] \\
\hline
\hline
\end{tabular}
\end{center}
\caption{Hyperparameters and their ranges (or states) used during random grid search.}
\label{tab:rf_hyper}
\end{table}
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/feature.pdf}
\caption{Mean test score of grid search results for $\mathrm{SNR}=100$ plotted against individual hyperparameters listed in Tab.~\ref{tab:rf_hyper}.}
\label{fig:hyper}
\end{figure}
Random forest (RF) is a learning method that operates by constructing an ensemble of decision trees \citep{tree}. It can be applied to classification and regression problems. For classification problems, the RF output is the class selected by the largest number trees, while for regression problems, the result is the average prediction of all trees. RF classifies samples using a forking path of decision points. We move from one decision point to another and at each point we apply a rule that decides which branch to follow \citep{Rf-advntg1}. At the end, we arrive at a leaf, which has a corresponding class label. We conclude the selection process by assigning the data to the class \citep{Denisko18}.
Tree-based approaches such as the RF have generated significant interest recently \citep{kennedy,hernandez,bonjean} because they are robust, difficult to overfit, requires minimal computational resources, and has availability of methods to assist in interpreting the results \citep{RF-Hastie2010,RF-breiman2001}, while being quick to implement. It is also efficient in learning highly non-linear relations between the input and the labels, making them suitable for large training data set and large number of input variables.
We use the accuracy score as the performance metric for optimizing our RF classifier network and to evaluate their performance. To arrive at the final classifier settings for running the pipeline, we split our train-test data into $90-10\%$ ratio and tune several hyperparameters using a randomised grid search algorithm \citep{sklearn_api}. Table~\ref{tab:rf_hyper} lists all the hyperparameters used in the pipeline and their values (or ranges) considered during the grid search. We perform a $100$ fold randomized cross validation for tuning the hyperparameters. Since cross validation can produce an unbiased estimator of a model on unseen data \citep{kuhn2013applied}, we do not use a separate validation dataset. For comparison and consistency of our results, we test additional algorithms, as discussed in Appendix \ref{appendix-1}).
\subsubsection{Detector Noise}
\label{sec:noise}
The GW detectors can be impacted by a wide variety of noises, stemming from, e.g., quantum sensing, seismic disturbances, suspension and mirror coating thermal vibrations, and transient perturbations such as anthropogenic interventions, weather, equipment malfunctions as well as occasional transient noise of unknown origin \citep{LIGOplus, noise-1, noise-2, Driggers19, noise-3, Ormiston20}. In this analysis, we model the combination of all these possible sources of noise as additive white Gaussian noise (AWGN) \citep{awgn1,awgn0,awgn2}. While AWGN cannot model instrumental noise transients and loud GW bursts contributing to non-Gaussian and non-stationary features \citep{Abbot20ligonoise}, as we show below, this approximation is adequate to establish the upper limit for the classification accuracy, which is the main goal of this work.
For a discrete time sequence signal, $x(t_i) \equiv x_i$, the continuum signal $G(t)$ can be defined with the help of a linear correlator as
\begin{equation}
G \equiv \int q(t)x(t)dt,
\label{eq:noise}
\end{equation}
where $q(t)$ encapsulates the information from the inverse noise covariance matrix. This signal can be matched to a known waveform from a linear matched filter. For zero signal, the expectation value $E$ of $G(t)$ becomes zero. One can define the signal to noise ratio (SNR) as
\begin{equation}
\frac{S}{N}=\frac{ E\{G\} }{\sqrt{{\left(\sigma\left\{G\right\}\right)}^2}},
\label{eq:noise2}
\end{equation}
where the numerator is the expectation\footnote{It is derived from the power spectral density of the signal noise \citep{Flanagan:98}.} and denominator is the variance of the signal. In other words, it can be defined as the ratio of signal power to variance. The SNR is a measure of the strength of a signal observed by a detector with a given level of noise. It is directly proportional to the amplitude of the signal buried in the noise.
In our analysis, we add AWGN noise $h_\mathrm{noise}$ to GW strain $h$ in the time domain before the application of ML analysis, for every single chosen value of SNR. The mean value of noise is assumed to be 0. The standard deviation $\sigma_\mathrm{noise}$ of $h_\mathrm{noise}$ is governed by the value of SNR. A consistency check on the stationarity and Gaussianity of the waveform distribution can be done by transforming them in Fourier space and studying their real and imaginary parts individually \citep{Abbot20ligonoise}. If noise distributions for both these components yield mean value of $0$, then white Gaussianity can be assumed. In this analysis, we have performed this check on our noise data and found them to be consistent with this requirement.
As the final step before applying ML, we standardize the dataset using the Z-score normalization \citep{zscore,zscore2}. Roughly speaking, this normalization moves the mean of the signal to $0$ and adjusts the average strength to $1$.
\section{Results}
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/TW1_A3_v4.pdf}
\caption{GW strain as a function of time for different progenitors. The top, center, and bottom panels show slowly, moderately rapidly, and rapidly rotating models, respectively. The time value of $0$ ms corresponds to the time of bounce.}
\label{fig:GW_strain}
\end{figure}
The top, center, and bottom panels of Fig.~\ref{fig:GW_strain} show the GW strain as a function of time for three sets of models with slow, moderate and rapid rotations, respectively. The models in each set have similar distributions of angular momentum in their cores. The plot shows that they produce GW signals that are similar to each other. This supports the observation that different progenitors produce similar bounce GW signal for the same distribution of rotation across mass coordinate \citep{ott:12}. Slowly rotating models exhibit contribution from prompt convection $\gtrsim 6 \, \mathrm{ms}$ after bounce. Since convection is a stochastic process, there is no obvious correlation with the progenitor mass within the time scale considered in this work.
The minor difference in the GW signal between different progenitor masses is caused by the differences in the specific entropy of the iron core. The latter tend to increase with the progenitor mass, but the dependence can be non-monotonic \citep[e.g.,][]{whw:02}. The variation of the inner core mass is amplified by our treatment of deleptonization \citep{mueller:09phd}. As discussed above, our deleptonization scheme produces inner core masses bounce that are $\simeq 10\%$ larger in progenitors with large masses than those in with small masses. In full neutrino transport simulations, the inner core mass is practically independent of the progenitor mass \citep{mueller:09phd}. For this reason, we can only put the upper limit to the accuracy of the classification accuracy, which, as we show below, is too low for any reliable identification of the progenitor mass from GW bounce signal along.
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Rolling_window.pdf}
\caption{The best accuracy score returned by the RF classifier pipeline for classifying the progenitor mass for different input time ranges using the rolling window of fixed width of $8$ ms. The time value of $0$ ms corresponds to the time of bounce.}
\label{fig:rolling}
\end{figure}
Figure~\ref{fig:rolling} shows the best accuracy score for classifying the progenitor mass for different ranges of time windows computed at maximum SNR of $100$. This score measures how many labels the model gets right out of the total number of predictions. The highest accuracy is achieved for range $[-2,\ +6]$ ms, which is also the range used in the analysis of \citet{richers:17}. This is expected since most of the bounce and ring-down GW signal is emitted in this time interval \citep[e.g.,][]{abdikamalov:22review}. Moreover, the prompt convection develops $\gtrsim\! 6$ ms after bounce, which, due to its stochastic nature, makes it harder to identify the source parameters from the signal. Consequently, we carried out the rest of the analysis with the wave spectra in the range $[-2,\ +6]$ ms.
We run the above pipeline for six different SNR values $1, 10, 25, 50, 70$, and $100$. We then compute the best accuracy score for each run. This score measures how many labels the model gets right out of the total number of predictions. Fig.~\ref{fig:accuracy} shows the accuracy rate of the classifier as a function of the SNR. The accuracy curve is obtained using cubic interpolation between the discrete SNR values. At the highest SNR of $100$, we see that there is $\sim70\%$ probability of correct classification of the progenitor mass.
This finding is mirrored in the behavior of the confusion matrix used to asses the classifier performance shown in Fig.~\ref{fig:Confusion}. Along each rows, the fraction of correct prediction is shown in percentage. The sum of these fractions is $1$. The correct predictions are along the diagonal. All classes have more than $50\%$ correct predictions. The s40 model is the most successfully predicted class with $10$ out of $11$ cases accurately predicted.
Apart from studying the accuracy metric and the confusion matrix, we also check the classification performance using the recall (also known as completeness) and precision (also known as correctness) of each progenitor mass class. The recall is defined as
\begin{equation}
\mathrm{recall = \frac{TP}{TP + FN}},
\label{eq:recall}
\end{equation}
while precision is defined as
\begin{equation}
\mathrm{precision = \frac{TP}{TP + FP}},
\label{eq:precision}
\end{equation}
where TP, FN and FP stand for the number of true positives, false negatives, and false positives for classification of target labels, respectively. Precision measures the quality of positive predictions made by the machine, i.e., the fraction of results which are relevant. Recall reflects the percentage of total true positive results correctly classified by the algorithm (hence also called as true positive rate or sensitivity). However, there is always a trade-off between recall and precision, and one can not maximize both these metrics at the same time \citep{recallprecision}.
The recall and precision for our data are shown as a function of SNR in Fig.~\ref{fig:recall} and \ref{fig:precision}, respectively. We see from Fig.~\ref{fig:recall} that the s40 model achieves the highest recall. For $\mathrm{SNR}\gtrsim 15$, we see that recall value for s40 peaks off around $1$, implying zero or low false negatives for $\mathrm{SNR} \gtrsim 15$ (see Eq.~\ref{eq:recall}), making it the most successfully classified progenitor. This is also consistent with the values of the confusion matrix shown in Fig.~\ref{fig:Confusion}, which has an accuracy of $\sim 91\%$ for s40. The s15 model is seen to be the worst performing model. s12 and s27 show similar trends in their recall and precision curves. Except in model s27, where recall improves in high SNR range and correspondingly the precision falls, thus negating any additional gain in classification results, achieved by the improved recall at high SNR values. In conclusion, we find that the recall and precision plots in combination, show overall poor classification performance and also beyond SNR $=40$ the improvement in performance levels off and minor gain is achieved.
To understand the classification mechanism and which signal features the classifier labelled as most dominating, we use the Gini importance score or the Gini index \citep{Gini,Gini2}. The Gini index is commonly used as the splitting criterion in classification trees, the corresponding impurity importance is often called Gini importance or mean decrease in impurity. While it is debated that correlation between data features can hinder importance interpretation, it is still a robust metric for analyzing feature importance. Fig.~\ref{fig:gini} shows the Gini score mapped to the time axis of the input signals. It is seen that the GW signal features in the neighbourhood of $\sim\!2.5$ ms are regarded as most relevant by the classifier for splitting nodes across the whole tree. This signifies that in this region the maximum purity in node splitting occurs compared to the rest of the region in the waveforms. In this region, the GW signal is dominated by the ring-down oscillations of the PNS, suggesting that the difference in the initial mass influences the properties of PNS oscillations. However, Gini importance does have sensitivity to correlation of features which influences the quality of node splits along decision tree structures. Therefore higher score is obtained with uncorrelated signals or pure node splits. In our plot, the low Gini score except around $2.5$ ms implies that there must be correlation of features in GW signals, present in these regions. This might have reduced the purity during node splits and thus blurred identification between features.
To verify the robustness of these finding, we have explored sensitivity of the ML algorithm by performing our analysis with multiple different classifiers, including the XGBoost and Neural Network. As we show in Appendix~\ref{appendix-1}, the main finding of our paper remains the same: in all cases, the performance of the classifier is not satisfactory to recover target labels.
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Accuracy.pdf}
\caption{Classifier accuracy score as a function of the signal SNR. It is seen that for low SNR values ($\le30$), accuracy rises sharply with SNR, beyond which accuracy increases marginally with maximum $\sim0.70$ at SNR $=100$.}
\label{fig:accuracy}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Confusion.pdf}
\caption{Confusion matrix for $\mathrm{SNR}=100$ presented for test data of $41$ candidates which is $10\%$ of the full strength catalogue used. The fraction of correct prediction per class can be seen along the diagonal.}
\label{fig:Confusion}
\end{figure}
\iffalse
\begin{table*}
\setlength\extrarowheight{2.09pt}
\begin{center}
\begin{tabular}{ c c c c c| c c c c c c}
\hline
\hline
Train &&&&& Test&&&\\
\hline
\hline
Class & Precision & Recall & F-1 Score & Support & Prescision & Recall & F1-Score & Support \\[.3em]
\hline
$12$ & $1$ &$1$ & $1$&$50$ & $0$&$0$ & $0$&$1$ \\
$15$ & $1$&$1$ &$1$ &$45$ & $0.67$& $0.67$&$0.67$ & $3$ \\
$27$& $1$ &$1$ & $1$&$51$ & $1$& $1$& $1$& $1$ \\
$40$ & $1$ & $1$& $1$&$48$ &$0.80$ & $0.80$&$0.80$ & $5$ \\
$75$ &$1$ &$1$ &$1$ &$49$ & $1$& $1$& $1$& $3$ \\
\hline
Accuracy &&&$1$&$243$&&&$0.77$&$13$\\
Macro Average &$1$&$1$&$1$&$243$&$0.69$&$0.69$&$0.69$&$13$\\
Weighted Average &$1$&$1$&$1$&$243$&$0.77$&$0.77$&$0.77$&$13$\\
\hline
\hline
\end{tabular}
\end{center}
\caption{A summary of the classification diagnostics. The total statistics shown here is $N=256$, which is three less than that in Tab.~\ref{T1}. This is due to the fact, that we found out the interpolation process was not accurately applied on three candidates and biased the classification analysis, hence they were omitted from the analysis.}
\label{T2}
\end{table*}
\fi
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Recall.pdf}
\caption{Recall as a function of the SNR for each progenitor mass. Recall gives insight into the number of positive cases that are misclassified by the model as negatives (i.e. number of false negatives). Higher recall value signifies higher amount of correctly classified positives. In this plot we can see that except for s40 model, recall values are below $0.7$. They only improve up to $\mathrm{SNR}\sim40$, except for s27 which shows a gain in recall at high SNR values. However, its corresponding precision value falls in the same (high) SNR region (Fig.~\ref{fig:precision}) and thus nullifying any additional gain achieved in classification, by the improved recall performance in the high SNR range. }
\label{fig:recall}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Precision.pdf}
\caption{Precision as a function of the SNR for each progenitor mass. Precision gives insight into the quality of positive predictions made by the model. In this case it defines the number of labels the model correctly predicted divided by the total number of labels the model predicted. In this plot we see that beyond SNR$=40$, precision does not change significantly until it reaches maximum $\mathrm{SNR}=100$, where some improvement is observed apart from s27 model, which counteracts this behaviour by showing improvement in its recall values in the same SNR region (Fig.~\ref{fig:recall}). }
\label{fig:precision}
\end{figure}
\begin{figure*}
\centering
\makebox[1.\textwidth]{\includegraphics[width=1.0\textwidth]{Plots/Gini2.jpeg}}
\caption{GW strain for SNR $=15$ (top panel) and Gini importance score (bottom panel) plotted as a function of time. Highest importance score (scaled) is seen around $\sim 2.5$ ms after bounce. The Gini importance score indicates the features where maximum purity in Random Forest node splitting (selection process) occurred.}
\label{fig:gini}
\end{figure*}
\section{Conclusion}
We have presented a detailed analysis of whether bounce and early ring-down GW signals from future observations of rotating CCSNe can be used for identifying iron core masses. We used GW data obtained from numerical simulations and injected white Gaussian noise to mimic the detector noise. We developed a ML classifier with the aim to identify the target classes of these input signals. We considered an idealized optimistic scenario for identifying the mass: we use rapidly rotating models that emit strong GWs, include only progenitor with only four different masses, which simplifies the selection process. In addition, to generate GW signals, we used numerical simulations that use a deleptonization scheme that artificially amplify differences in the collapse and post-bounce dynamics of progenitors with distinct masses.
Despite this idealized favorable scenario, we were unable to identify the mass, purely based on bounce and early ring-down GW signal alone. More realistic treatment of noise modelling like non-Gaussian signals can only reduce the chances of any possible identification of the progenitor mass labels. The fact that even for SNR of 100, the classifier failed to give a high accuracy score, does point to the fact that this is an intrinsic phenomenon and independent of detector quality. Our results therefore show that the information about the iron core mass is not contained in the bounce and early ring-down GW signal. We have also performed similar analysis in Fourier space and classification performance was within $2\%$ of the accuracy score in time domain space. We thus conclude that the weak relation between GW waveforms and the iron core mass is an intrinsic property of the system and the classification via ML is not possible.
Additionally, we have explored sensitivity of the ML algorithm by performing our analysis with multiple different classifiers (e.g., XGBoost and Neural Network, as discussed in Appendix~\ref{appendix-1}). In all cases, our results showed that the performance of the classifier was not satisfactory to recover target labels. We do however note that, as with any ML analysis, numerous other hyperparameter combinations are possible, as do multiple other algorithms. We anticipate to further perform our analysis in future with more complex ML model and representative data. However, we do not expect significant deviation from the results presented in this paper. We expect that to constrain the iron core mass one has to incorporate more information such us longer-term post-bonce GW signal and/or neutrino emission \citep[e.g.,][]{oconnor:13, Yokozawa15, kuroda:17, Nagakura22}.
\section*{Acknowledgements}
We thank Sherwood Richers for carefully reading the manuscript and for helpful comments. This research has been funded by the Science Committee of the Ministry of Education and Science of the Republic of Kazakhstan (Grant No. AP13067834 and AP08856149) and the Nazarbayev University Faculty Development Competitive Research Grant Program No 11022021FD2912 (ssh2022007).
High-performance workstations of ECL/NU have been used to perform all simulations, data analysis, and ML calculations (\href{http://ecl.nu.edu.kz/computational-facilities/}{http://ecl.nu.edu.kz/computational-facilities/}).
\section*{Data Availability}
The data used in this work is available from authors upon request. The gravitational waveforms are publicly available at \href{https://zenodo.org/record/7090935}{https://zenodo.org/record/7090935}.
\bibliographystyle{mnras}
\section{Introduction}
The detection of gravitational waves (GWs) from black hole mergers gave birth to GW astronomy \citep{abbot16PRL}, while the observation of neutron star merger in gravitational and electromagnetic waves is the prime example of the multi-messenger astronomy \citep{abbott17PRL}. One of the most promising multi-messenger sources that are yet to be detected in GWs are core-collapse supernovae (CCSNe).
CCSNe are the powerful explosions of massive stars at the end of their life. The collapse of the stellar core releases $\sim\! 10^{53}$ erg of its gravitational binding energy. While most of this escapes in the form of neutrinos, the rest powers the explosion \citep{Bethe:1990mw}. Supernovae involve powerful aspherical flows that generate GWs with energies up to $\sim\! 10^{47}$ erg \citep[e.g.,][]{ott:09review, kotake:13review, kotake:17}. The explosion front breaks out of the stellar surface hours later \citep{waxman17}, producing a blast of photons across the electromagnetic spectrum \citep[e.g,][]{Nakamura:2016kkl}.
The supernova dynamics, and thus the GW signal, strongly depends on rotation. In non- or slowly rotating models, which represent the majority of CCSNe \cite[e.g.,][]{heger:05}, the explosion is governed by the \emph{neutrino mechanism} \citep[e.g.,][for recent reviews]{kotake:12snreview, janka:12a,mueller:20review, Burrows21review, mezzacappa22}. A fraction of neutrinos emitted by the protoneutron star (PNS) heats the post-shock material, giving rise to neutrino-driven convection \citep{herant:92, bhf:95, janka:95} and standing accretion shock instability (SASI) \citep{blondin:03, foglizzo:06, mueller:12b}. These flows perturb the PNS and excite its oscillations, generating powerful GWs \cite[e.g.,][]{murphy:09, mueller:13,cerda:13, yakunin15, andresen:17, Hayama18, radice:19gw, mezzacappa:20a, Raynaud22}. Asymmetries in neutrino emission also contributes to the GW signal \cite[e.g.,][]{Kotake07, takiwaki:18, Vartanyan20}. The GWs are detectable for sources within our Galaxy with current generation detectors \citep{gossan:16, abbott:20ccsn, Lopez21, Szczepanczyk21, Antelis22}, while the future generation detectors will enable a more detailed observation \citep{Srivastava:2019fcb}.
In rare rapidly rotating stars, the rotational kinetic energy powers the explosion via the \emph{magnetorotational mechanism} \citep{leblanc:70, bisno:76, akiyama:03, burrows:07a, moesta:14a, kuroda:20, obergaulinger:20, Raynaud20}. Due to the centrifugal force, the PNS forms with an oblate perturbation, triggering PNS ring-down oscillations that last for $\sim\! 10 $ ms \citep{ott:12, fuller:15}. In some cases, the PNS may be subject to non-axisymmetric instabilities \citep[e.g.,][]{ott:07cqg, scheidegger:10cqg, shibagaki:20, Takiwaki21}. When this happens, the non-axisymetrically-deformed PNS emits GW for many rotation periods, significantly enhancing the detectability \citep{abbott:20ccsn}. For moderate rotation, both the rotational bounce and convection contribute to the GW signal \citep{andresen19, Pan21, Jardine22}.
Once detected, it is possible to estimate the parameters of source using the GWs emitted by the source \citep[e.g.,][for a recent review]{Christensen22}. Since slowly and rapidly rotating models have significantly different dynamics, their GW signals can be confidently distinguished \citep{Logue12, Powell16, Chan20, Szczepanczyk21, Saiz-Perez22}. Both convection and SASI develop from stochastic perturbations, so the GWs coming from slowly rotating models contain stochastic components. In contrast, the bounce GW signal in rapidly rotating stars can be determined precisely for a given (physical and computational) model parameters \cite[e.g.,][]{zwerger:97, dimmelmeier:08}. Despite the presence of the stochastic contributions, the GW spectra contain the frequencies of the physical processes happening in the central regions \citep[e.g.,][]{Kotake11, mueller:13, Astone18, Roma19, Srivastava:2019fcb, Powell22}. In particular, the PNS oscillations is the dominant component of the signal \citep{murphy:09, cerda:13, morozova:18, radice:19gw, Warren20}, the frequencies of which can be related to the physical parameters of the system, such as the mass and radius of the PNS \cite[e.g.,][]{dimmelmeier:06, mueller:13, Torres-Forne19, Sotani21, Bizouard21}.
For rapidly rotating stars, \citet{abdikamalov:14} explored the possibility to infer the rotation and its distribution in the supernova core. They considered $\sim\!100$ different rotational configurations of a stellar model with 5 different degrees of differential rotation. They found that, for a source at $10$~kpc distance, it is possible to measure rotation with $\sim\! 20\% $ accuracy for rapidly rotating models, in which the rotational kinetic energy exceed $\sim\! 8\%$ of the potential binding energy. In slowly rotating models, due to smaller GW amplitudes and higher stochastic contribution, the error becomes larger. These estimates were further improved by \citet{Engels14, Edwards14}, and \citet{Afle21}. \citet{Hayama16} showed that signs of rapid rotation can be found in the circular polarization of the GW signal. \citet{Yokozawa15} proposed to combine GW and neutrino observations to infer the rotation from the time delays between the bursts of these two signals. \citet{pajkos19, pajkos21} proposed to combine the the core-bounce signal with the dominant frequency mode of the PNS in the pre-explosion to constrain the structure of the progenitor star.
\citet{richers:17} studied the dependence of the GW signal on the parameters of the equation of state (EOS) of high-density matter in protoneutron stars and treatment the electron capture rate during collapse. They find a modest impact of these parameters to the bounce and the early $\sim\! 20$ ms post-bounce GW signal. This emphasizes the importance of accurate modeling \citep[e.g.,][]{lentz:12a, Kotake18, Pan18, Pan19, mezzacappa20, Andresen21}. Using deep convolutional neural networks, \citet{Edwards21} classified these EOSs $72\%$ correctly, while their most probable five EOSs were found with $97\%$ accuracy \cite[see also][]{Chao22}.
In this work, we study if it is possible to probe the iron core mass of CCSN progenitor from the bounce and early ring-down GW signal alone. \citet{ott:12} showed that progenitors with different masses will produce similar GWs if they have a similar angular momentum distributions at a given mass coordinate in the stellar core. This suggest that the progenitor mass at most has a subtle effect on the bounce dynamics. We extend this work further by using machine learning (ML) for signal classification and studying a wider ranges of progenitors and rotational configurations. We look at idealized optimistic scenario for measuring mass from the bounce and early ring-down GW signal alone. First, we consider rapidly rotating models, which produce strong GWs. Second, we use the deleptonization method that is known to artificially amplify the differences between progenitors. Third, we limit ourselves to model progenitors with four different iron core masses only, which simplifies the selection process. Despite these ideally favorable conditions, we show that the iron core mass cannot be accurately measured from the GW bounce signal alone. This suggests that, to measure the mass, one has to incorporate additional information such as longer post-bounce signal or neutrino observations, or both.
\section{Methodology}
\subsection{Gravitational waveforms}
\label{sec:gw_data}
We consider four progenitor models with zero-age main sequence masses ranging from $12$ to $40$ $M_\odot$. At the pre-collapse stage, these models develop iron cores with masses ranging from $1.3M_\odot$ to $1.8M_\odot$, as shown in Table ~\ref{tab:progenitor_param}, which also provides the central densities and entropies. The $12M_\odot$ and $40M_\odot$ models are produced by \cite{woosley:07}, while the $15M_\odot$ model is evolved by \cite{heger:05} with magnetic field prescription by \cite{spruit:02}. The $27M_\odot$ model is produced by \cite{whw:02}. All models have solar metallicity.
To obtain the GW signals, we perform simulations using the {\tt CoCoNuT} code \citep{Dimmelmeier02a, Dimmelmeier02b, dimmelmeier:05MdM}. The simulations are continued until 25 ms after bounce. This time interval contains full bounce and ring-down signal from a~CCSN event \citep[][]{abdikamalov:22review}. We use the same computational setup as in \citet{richers:17} with the SFHo nuclear equation of state \citep{steiner:13b}. During collapse, our code uses the $Y_\mathrm{e}(\rho)$ deleptonization method of \cite{liebendoerfer:05}. Following \citet{richers:17}, we use $Y_\mathrm{e}(\rho)$ profile obtained from spherically symmetric radiation hydrodynamics simulations using the {\tt GR1D} code \citep{oconnor:15a}. This method is known to amplify the contrast between progenitors with different masses \citep{mueller:09phd, pajkos21}: it produces variations in the inner core mass at bounce of $\simeq 10\%$ between progenitors with different masses, whereas full neutrino-transport simulations yields inner core masses that are practically independent of the progenitor \citep{mueller:09phd}. However, this weakness of the method represents the strength of our work: as we show below, even with the artificial amplification, the difference in the GW signals is too small to be distinguishable. In more realistic models, the distinction should be even smaller.
Following \cite{abdikamalov:14}, for each progenitor we consider about $100$ different rotational profiles ranging from slow to rapid rotation, where rotation has little to large effect on the core dynamics, respectively. At distance $\varpi$ from the rotation axis, the angular velocity of pre-collapse models is given by
\begin{equation}
\label{eq:rot_law}
\Omega(\varpi) = \Omega_0 \left[ 1 + \left(\frac{\varpi}{A}\right)^2\right]^{-1},
\end{equation}
where $A$ is a measure of degree of differential rotation, $\Omega_0$ is the central angular velocity \citep{komatsu:89a}. As in \cite{abdikamalov:14}, we consider five different values of $A$, ranging from $300$ to $10,\!000$ km, which correspond to the limits of, respectively, extreme differential and uniform rotations in the stellar core. For a given $\Omega_0$ and $A$, we impose the rotation law (\ref{eq:rot_law}) to the $12M_\odot$ model, as described in \cite{abdikamalov:14}. We then map the specific angular momentum distribution homologously to the other progenitors with different masses. This ensures that the specific angular momentum at a given enclosed mass coordinate is similar in all four progenitors. As summarized in Table~\ref{tab:T1}, the total number of rotational configurations we consider is 97, 99, 102, and 104 for s12, s15, s27, and s40 models, respectively. The difference in this number is caused by the fact that at the extreme rapid rotation limit, due to a fine balance between gravity and centrifugal force in the pre-collapse stage, some of the models do not collapse for some progenitors and collapse for other.
We extract the GW waveforms from the simulations using the Newtonian quadrupole formula in the first moment of momentum density formulation \citep{Dimmelmeier02a, Dimmelmeier02b, dimmelmeier:05MdM}, which yields accurate waveforms for stellar core collapse simulations \citep{reisswig:11ccwave}.
\begin{table}
\setlength\extrarowheight{2.09pt}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
ZAMS mass & $M_\mathrm{IC}$ & $s_\mathrm{c}$ & $\rho_\mathrm{c}$ \\ [.01em]
[$M_\odot$] & [$M_\odot$] & [$k_\mathrm{B}/\mathrm{baryon}$] & [$10^9\,\mathrm{g/cm}^3$] \\ [.3em]
\hline
12 & 1.3 & $0.60$ & $12.0$ \\
15 & 1.4 & $0.62$ & $8.74$ \\
27 & 1.5 & $0.77$ & $4.96$ \\
40 & 1.8 & $1.06$ & $2.17$ \\
\hline
\end{tabular}
\end{center}
\caption{Progenitor model parameters used in our work. The ZAMS mass is the initial (the so-called zero-age) mass of the star. Parameter $M_\mathrm{IC}$ is the inner core mass, while $s_\mathrm{c}$ and $\rho_\mathrm{c}$ are the central entropy and density before collapse.}
\label{tab:progenitor_param}
\end{table}
\begin{table}
\setlength\extrarowheight{2.09pt}
\begin{center}
\begin{tabular}{c c}
\hline
Progenitor Mass & No. of Events \\ [.3em]
[$M_\odot$] & \\ [.3em]
\hline
12 & 97 \\
15 & 99 \\
27 & 102 \\
40 & 104 \\
\hline
Total & $402$\\
\hline
\end{tabular}
\end{center}
\caption{Number of the generated GW events grouped by their progenitor mass types.}
\label{tab:T1}
\end{table}
\subsection{Machine learning}
\begin{figure*}
\centering
\makebox[1.\textwidth]{\includegraphics[width=1.0\textwidth]{Plots/flowchart.pdf}}
\caption{Flowchart of the analysis pipeline of our study.}
\label{fig:flowchart}
\end{figure*}
Our analysis pipeline can be divided in the following steps:
\begin{itemize}
\item Preprocessing: Customize the waveforms for ML input.
\item Noise: Add noise to the signals to model the detector noise.
\item ML: Optimize hyperparameters\footnote{Hyperparameters are parameters of the algorithm, which are tuned to optimize the learning. Their values are chosen before initialising learning. They should not be confused with parameters of the data, which are obtained during the learning. The choice of hyperparameter has a direct impact on the performance of the ML model and hence they are optimized to produce the best performance.} and apply ML algorithm.
\item Analysis: Explore confusion matrix and other metrics.
\end{itemize}
A flowchart depiction of the pipeline is illustrated in Fig.~\ref{fig:flowchart}.
To make the waveforms suitable to ML analysis, we condense all candidate waveforms into a single data cube with the progenitor mass of each candidate treated as target class. We map each candidates to the same cadence along the time axis with uniform starting and ending points of $-5$ and $+15$ ms around the bounce time. The time of bounce corresponds to time equal to zero. We generate $10,000$ interpolated points along the time axis per candidate, resulting in a data cube of dimensions $($402$\times10000\times2)$.
\subsubsection{Random Forest}
\label{sec:rf}
\begin{table}
\setlength\extrarowheight{2.09pt}
\begin{center}
\begin{tabular}{|c | c| }
\hline
hyperparameters & Range \\ [.3em]
\hline \hline
No. of estimators & $[100,400]$ \\
Max. features & [auto, sqrt, log2] \\
Max. depth & $[10,250]$\\
Min. samples split & $[2,100]$ \\
min. samples leaf & $[2,100]$\\
Bootstrap & [True, False]\\
OOB score & [True, False]\\
Warm start & [True, False] \\
\hline
\hline
\end{tabular}
\end{center}
\caption{Hyperparameters and their ranges (or states) used during random grid search.}
\label{tab:rf_hyper}
\end{table}
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/feature.pdf}
\caption{Mean test score of grid search results for $\mathrm{SNR}=100$ plotted against individual hyperparameters listed in Tab.~\ref{tab:rf_hyper}.}
\label{fig:hyper}
\end{figure}
Random forest (RF) is a learning method that operates by constructing an ensemble of decision trees \citep{tree}. It can be applied to classification and regression problems. For classification problems, the RF output is the class selected by the largest number trees, while for regression problems, the result is the average prediction of all trees. RF classifies samples using a forking path of decision points. We move from one decision point to another and at each point we apply a rule that decides which branch to follow \citep{Rf-advntg1}. At the end, we arrive at a leaf, which has a corresponding class label. We conclude the selection process by assigning the data to the class \citep{Denisko18}.
Tree-based approaches such as the RF have generated significant interest recently \citep{kennedy,hernandez,bonjean} because they are robust, difficult to overfit, requires minimal computational resources, and has availability of methods to assist in interpreting the results \citep{RF-Hastie2010,RF-breiman2001}, while being quick to implement. It is also efficient in learning highly non-linear relations between the input and the labels, making them suitable for large training data set and large number of input variables.
We use the accuracy score as the performance metric for optimizing our RF classifier network and to evaluate their performance. To arrive at the final classifier settings for running the pipeline, we split our train-test data into $90-10\%$ ratio and tune several hyperparameters using a randomised grid search algorithm \citep{sklearn_api}. Table~\ref{tab:rf_hyper} lists all the hyperparameters used in the pipeline and their values (or ranges) considered during the grid search. We perform a $100$ fold randomized cross validation for tuning the hyperparameters. Since cross validation can produce an unbiased estimator of a model on unseen data \citep{kuhn2013applied}, we do not use a separate validation dataset. For comparison and consistency of our results, we test additional algorithms, as discussed in Appendix \ref{appendix-1}).
\subsubsection{Detector Noise}
\label{sec:noise}
The GW detectors can be impacted by a wide variety of noises, stemming from, e.g., quantum sensing, seismic disturbances, suspension and mirror coating thermal vibrations, and transient perturbations such as anthropogenic interventions, weather, equipment malfunctions as well as occasional transient noise of unknown origin \citep{LIGOplus, noise-1, noise-2, Driggers19, noise-3, Ormiston20}. In this analysis, we model the combination of all these possible sources of noise as additive white Gaussian noise (AWGN) \citep{awgn1,awgn0,awgn2}. While AWGN cannot model instrumental noise transients and loud GW bursts contributing to non-Gaussian and non-stationary features \citep{Abbot20ligonoise}, as we show below, this approximation is adequate to establish the upper limit for the classification accuracy, which is the main goal of this work.
For a discrete time sequence signal, $x(t_i) \equiv x_i$, the continuum signal $G(t)$ can be defined with the help of a linear correlator as
\begin{equation}
G \equiv \int q(t)x(t)dt,
\label{eq:noise}
\end{equation}
where $q(t)$ encapsulates the information from the inverse noise covariance matrix. This signal can be matched to a known waveform from a linear matched filter. For zero signal, the expectation value $E$ of $G(t)$ becomes zero. One can define the signal to noise ratio (SNR) as
\begin{equation}
\frac{S}{N}=\frac{ E\{G\} }{\sqrt{{\left(\sigma\left\{G\right\}\right)}^2}},
\label{eq:noise2}
\end{equation}
where the numerator is the expectation\footnote{It is derived from the power spectral density of the signal noise \citep{Flanagan:98}.} and denominator is the variance of the signal. In other words, it can be defined as the ratio of signal power to variance. The SNR is a measure of the strength of a signal observed by a detector with a given level of noise. It is directly proportional to the amplitude of the signal buried in the noise.
In our analysis, we add AWGN noise $h_\mathrm{noise}$ to GW strain $h$ in the time domain before the application of ML analysis, for every single chosen value of SNR. The mean value of noise is assumed to be 0. The standard deviation $\sigma_\mathrm{noise}$ of $h_\mathrm{noise}$ is governed by the value of SNR. A consistency check on the stationarity and Gaussianity of the waveform distribution can be done by transforming them in Fourier space and studying their real and imaginary parts individually \citep{Abbot20ligonoise}. If noise distributions for both these components yield mean value of $0$, then white Gaussianity can be assumed. In this analysis, we have performed this check on our noise data and found them to be consistent with this requirement.
As the final step before applying ML, we standardize the dataset using the Z-score normalization \citep{zscore,zscore2}. Roughly speaking, this normalization moves the mean of the signal to $0$ and adjusts the average strength to $1$.
\section{Results}
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/TW1_A3_v4.pdf}
\caption{GW strain as a function of time for different progenitors. The top, center, and bottom panels show slowly, moderately rapidly, and rapidly rotating models, respectively. The time value of $0$ ms corresponds to the time of bounce.}
\label{fig:GW_strain}
\end{figure}
The top, center, and bottom panels of Fig.~\ref{fig:GW_strain} show the GW strain as a function of time for three sets of models with slow, moderate and rapid rotations, respectively. The models in each set have similar distributions of angular momentum in their cores. The plot shows that they produce GW signals that are similar to each other. This supports the observation that different progenitors produce similar bounce GW signal for the same distribution of rotation across mass coordinate \citep{ott:12}. Slowly rotating models exhibit contribution from prompt convection $\gtrsim 6 \, \mathrm{ms}$ after bounce. Since convection is a stochastic process, there is no obvious correlation with the progenitor mass within the time scale considered in this work.
The minor difference in the GW signal between different progenitor masses is caused by the differences in the specific entropy of the iron core. The latter tend to increase with the progenitor mass, but the dependence can be non-monotonic \citep[e.g.,][]{whw:02}. The variation of the inner core mass is amplified by our treatment of deleptonization \citep{mueller:09phd}. As discussed above, our deleptonization scheme produces inner core masses bounce that are $\simeq 10\%$ larger in progenitors with large masses than those in with small masses. In full neutrino transport simulations, the inner core mass is practically independent of the progenitor mass \citep{mueller:09phd}. For this reason, we can only put the upper limit to the accuracy of the classification accuracy, which, as we show below, is too low for any reliable identification of the progenitor mass from GW bounce signal along.
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Rolling_window.pdf}
\caption{The best accuracy score returned by the RF classifier pipeline for classifying the progenitor mass for different input time ranges using the rolling window of fixed width of $8$ ms. The time value of $0$ ms corresponds to the time of bounce.}
\label{fig:rolling}
\end{figure}
Figure~\ref{fig:rolling} shows the best accuracy score for classifying the progenitor mass for different ranges of time windows computed at maximum SNR of $100$. This score measures how many labels the model gets right out of the total number of predictions. The highest accuracy is achieved for range $[-2,\ +6]$ ms, which is also the range used in the analysis of \citet{richers:17}. This is expected since most of the bounce and ring-down GW signal is emitted in this time interval \citep[e.g.,][]{abdikamalov:22review}. Moreover, the prompt convection develops $\gtrsim\! 6$ ms after bounce, which, due to its stochastic nature, makes it harder to identify the source parameters from the signal. Consequently, we carried out the rest of the analysis with the wave spectra in the range $[-2,\ +6]$ ms.
We run the above pipeline for six different SNR values $1, 10, 25, 50, 70$, and $100$. We then compute the best accuracy score for each run. This score measures how many labels the model gets right out of the total number of predictions. Fig.~\ref{fig:accuracy} shows the accuracy rate of the classifier as a function of the SNR. The accuracy curve is obtained using cubic interpolation between the discrete SNR values. At the highest SNR of $100$, we see that there is $\sim70\%$ probability of correct classification of the progenitor mass.
This finding is mirrored in the behavior of the confusion matrix used to asses the classifier performance shown in Fig.~\ref{fig:Confusion}. Along each rows, the fraction of correct prediction is shown in percentage. The sum of these fractions is $1$. The correct predictions are along the diagonal. All classes have more than $50\%$ correct predictions. The s40 model is the most successfully predicted class with $10$ out of $11$ cases accurately predicted.
Apart from studying the accuracy metric and the confusion matrix, we also check the classification performance using the recall (also known as completeness) and precision (also known as correctness) of each progenitor mass class. The recall is defined as
\begin{equation}
\mathrm{recall = \frac{TP}{TP + FN}},
\label{eq:recall}
\end{equation}
while precision is defined as
\begin{equation}
\mathrm{precision = \frac{TP}{TP + FP}},
\label{eq:precision}
\end{equation}
where TP, FN and FP stand for the number of true positives, false negatives, and false positives for classification of target labels, respectively. Precision measures the quality of positive predictions made by the machine, i.e., the fraction of results which are relevant. Recall reflects the percentage of total true positive results correctly classified by the algorithm (hence also called as true positive rate or sensitivity). However, there is always a trade-off between recall and precision, and one can not maximize both these metrics at the same time \citep{recallprecision}.
The recall and precision for our data are shown as a function of SNR in Fig.~\ref{fig:recall} and \ref{fig:precision}, respectively. We see from Fig.~\ref{fig:recall} that the s40 model achieves the highest recall. For $\mathrm{SNR}\gtrsim 15$, we see that recall value for s40 peaks off around $1$, implying zero or low false negatives for $\mathrm{SNR} \gtrsim 15$ (see Eq.~\ref{eq:recall}), making it the most successfully classified progenitor. This is also consistent with the values of the confusion matrix shown in Fig.~\ref{fig:Confusion}, which has an accuracy of $\sim 91\%$ for s40. The s15 model is seen to be the worst performing model. s12 and s27 show similar trends in their recall and precision curves. Except in model s27, where recall improves in high SNR range and correspondingly the precision falls, thus negating any additional gain in classification results, achieved by the improved recall at high SNR values. In conclusion, we find that the recall and precision plots in combination, show overall poor classification performance and also beyond SNR $=40$ the improvement in performance levels off and minor gain is achieved.
To understand the classification mechanism and which signal features the classifier labelled as most dominating, we use the Gini importance score or the Gini index \citep{Gini,Gini2}. The Gini index is commonly used as the splitting criterion in classification trees, the corresponding impurity importance is often called Gini importance or mean decrease in impurity. While it is debated that correlation between data features can hinder importance interpretation, it is still a robust metric for analyzing feature importance. Fig.~\ref{fig:gini} shows the Gini score mapped to the time axis of the input signals. It is seen that the GW signal features in the neighbourhood of $\sim\!2.5$ ms are regarded as most relevant by the classifier for splitting nodes across the whole tree. This signifies that in this region the maximum purity in node splitting occurs compared to the rest of the region in the waveforms. In this region, the GW signal is dominated by the ring-down oscillations of the PNS, suggesting that the difference in the initial mass influences the properties of PNS oscillations. However, Gini importance does have sensitivity to correlation of features which influences the quality of node splits along decision tree structures. Therefore higher score is obtained with uncorrelated signals or pure node splits. In our plot, the low Gini score except around $2.5$ ms implies that there must be correlation of features in GW signals, present in these regions. This might have reduced the purity during node splits and thus blurred identification between features.
To verify the robustness of these finding, we have explored sensitivity of the ML algorithm by performing our analysis with multiple different classifiers, including the XGBoost and Neural Network. As we show in Appendix~\ref{appendix-1}, the main finding of our paper remains the same: in all cases, the performance of the classifier is not satisfactory to recover target labels.
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Accuracy.pdf}
\caption{Classifier accuracy score as a function of the signal SNR. It is seen that for low SNR values ($\le30$), accuracy rises sharply with SNR, beyond which accuracy increases marginally with maximum $\sim0.70$ at SNR $=100$.}
\label{fig:accuracy}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Confusion.pdf}
\caption{Confusion matrix for $\mathrm{SNR}=100$ presented for test data of $41$ candidates which is $10\%$ of the full strength catalogue used. The fraction of correct prediction per class can be seen along the diagonal.}
\label{fig:Confusion}
\end{figure}
\iffalse
\begin{table*}
\setlength\extrarowheight{2.09pt}
\begin{center}
\begin{tabular}{ c c c c c| c c c c c c}
\hline
\hline
Train &&&&& Test&&&\\
\hline
\hline
Class & Precision & Recall & F-1 Score & Support & Prescision & Recall & F1-Score & Support \\[.3em]
\hline
$12$ & $1$ &$1$ & $1$&$50$ & $0$&$0$ & $0$&$1$ \\
$15$ & $1$&$1$ &$1$ &$45$ & $0.67$& $0.67$&$0.67$ & $3$ \\
$27$& $1$ &$1$ & $1$&$51$ & $1$& $1$& $1$& $1$ \\
$40$ & $1$ & $1$& $1$&$48$ &$0.80$ & $0.80$&$0.80$ & $5$ \\
$75$ &$1$ &$1$ &$1$ &$49$ & $1$& $1$& $1$& $3$ \\
\hline
Accuracy &&&$1$&$243$&&&$0.77$&$13$\\
Macro Average &$1$&$1$&$1$&$243$&$0.69$&$0.69$&$0.69$&$13$\\
Weighted Average &$1$&$1$&$1$&$243$&$0.77$&$0.77$&$0.77$&$13$\\
\hline
\hline
\end{tabular}
\end{center}
\caption{A summary of the classification diagnostics. The total statistics shown here is $N=256$, which is three less than that in Tab.~\ref{T1}. This is due to the fact, that we found out the interpolation process was not accurately applied on three candidates and biased the classification analysis, hence they were omitted from the analysis.}
\label{T2}
\end{table*}
\fi
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Recall.pdf}
\caption{Recall as a function of the SNR for each progenitor mass. Recall gives insight into the number of positive cases that are misclassified by the model as negatives (i.e. number of false negatives). Higher recall value signifies higher amount of correctly classified positives. In this plot we can see that except for s40 model, recall values are below $0.7$. They only improve up to $\mathrm{SNR}\sim40$, except for s27 which shows a gain in recall at high SNR values. However, its corresponding precision value falls in the same (high) SNR region (Fig.~\ref{fig:precision}) and thus nullifying any additional gain achieved in classification, by the improved recall performance in the high SNR range. }
\label{fig:recall}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{Plots/Precision.pdf}
\caption{Precision as a function of the SNR for each progenitor mass. Precision gives insight into the quality of positive predictions made by the model. In this case it defines the number of labels the model correctly predicted divided by the total number of labels the model predicted. In this plot we see that beyond SNR$=40$, precision does not change significantly until it reaches maximum $\mathrm{SNR}=100$, where some improvement is observed apart from s27 model, which counteracts this behaviour by showing improvement in its recall values in the same SNR region (Fig.~\ref{fig:recall}). }
\label{fig:precision}
\end{figure}
\begin{figure*}
\centering
\makebox[1.\textwidth]{\includegraphics[width=1.0\textwidth]{Plots/Gini2.jpeg}}
\caption{GW strain for SNR $=15$ (top panel) and Gini importance score (bottom panel) plotted as a function of time. Highest importance score (scaled) is seen around $\sim 2.5$ ms after bounce. The Gini importance score indicates the features where maximum purity in Random Forest node splitting (selection process) occurred.}
\label{fig:gini}
\end{figure*}
\section{Conclusion}
We have presented a detailed analysis of whether bounce and early ring-down GW signals from future observations of rotating CCSNe can be used for identifying iron core masses. We used GW data obtained from numerical simulations and injected white Gaussian noise to mimic the detector noise. We developed a ML classifier with the aim to identify the target classes of these input signals. We considered an idealized optimistic scenario for identifying the mass: we use rapidly rotating models that emit strong GWs, include only progenitor with only four different masses, which simplifies the selection process. In addition, to generate GW signals, we used numerical simulations that use a deleptonization scheme that artificially amplify differences in the collapse and post-bounce dynamics of progenitors with distinct masses.
Despite this idealized favorable scenario, we were unable to identify the mass, purely based on bounce and early ring-down GW signal alone. More realistic treatment of noise modelling like non-Gaussian signals can only reduce the chances of any possible identification of the progenitor mass labels. The fact that even for SNR of 100, the classifier failed to give a high accuracy score, does point to the fact that this is an intrinsic phenomenon and independent of detector quality. Our results therefore show that the information about the iron core mass is not contained in the bounce and early ring-down GW signal. We have also performed similar analysis in Fourier space and classification performance was within $2\%$ of the accuracy score in time domain space. We thus conclude that the weak relation between GW waveforms and the iron core mass is an intrinsic property of the system and the classification via ML is not possible.
Additionally, we have explored sensitivity of the ML algorithm by performing our analysis with multiple different classifiers (e.g., XGBoost and Neural Network, as discussed in Appendix~\ref{appendix-1}). In all cases, our results showed that the performance of the classifier was not satisfactory to recover target labels. We do however note that, as with any ML analysis, numerous other hyperparameter combinations are possible, as do multiple other algorithms. We anticipate to further perform our analysis in future with more complex ML model and representative data. However, we do not expect significant deviation from the results presented in this paper. We expect that to constrain the iron core mass one has to incorporate more information such us longer-term post-bonce GW signal and/or neutrino emission \citep[e.g.,][]{oconnor:13, Yokozawa15, kuroda:17, Nagakura22}.
\section*{Acknowledgements}
We thank Sherwood Richers for carefully reading the manuscript and for helpful comments. This research has been funded by the Science Committee of the Ministry of Education and Science of the Republic of Kazakhstan (Grant No. AP13067834 and AP08856149) and the Nazarbayev University Faculty Development Competitive Research Grant Program No 11022021FD2912 (ssh2022007).
High-performance workstations of ECL/NU have been used to perform all simulations, data analysis, and ML calculations (\href{http://ecl.nu.edu.kz/computational-facilities/}{http://ecl.nu.edu.kz/computational-facilities/}).
\section*{Data Availability}
The data used in this work is available from authors upon request. The gravitational waveforms are publicly available at \href{https://zenodo.org/record/7090935}{https://zenodo.org/record/7090935}.
\bibliographystyle{mnras}
| caeeaa82e204f907210ca5d1b30528abad74de90 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
We consider the Schr\"odinger equation with a logarithmic nonlinearity,
\begin{equation} \tag{logGP}
i \partial_t u + \Delta u = \lambda u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2}, \quad
u_{\mid t=0} =u_0, \label{logGP}
\end{equation}
where $x \in \mathbb{R}^d$, $d \geq 1$, $\lambda > 0$, and with the
boundary condition at infinity
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{u (t, x)} \rightarrow 1 \quad \text{as} \quad \@ifstar{\oldabs}{\oldabs*}{x} \rightarrow \infty.
\end{equation*}
Such boundary condition is reminiscent of the standard
Gross-Pitaevskii equation
\begin{equation}
\label{eq:GP}
i \partial_t u + \Delta u =\(|u|^2-1\)u,\quad
u_{\mid t=0} =u_0,
\end{equation}
whose Cauchy problem was studied in
\cite{BeSa99,GalloGP,Gerard_Cauchy_GP,PG08,Zhidkov92,Zhidkov01}. The
most complete result regarding this aspect is found in
\cite{Gerard_Cauchy_GP,PG08}, where, for $d\le 3$, \eqref{eq:GP} is proved to be
globally well posed in the energy space
\begin{equation*}
E_\textnormal{GP}\coloneqq \{ u \in H^1_\textnormal{loc} (\mathbb{R}^d) \, | \, \mathcal{E}_\textnormal{GP} (u) < \infty \},
\end{equation*}
where $\mathcal{E}_\textnormal{GP}$ is the Ginzburg-Landau energy
\begin{equation*}
\mathcal{E}_{\textnormal{GP}} (u) \coloneqq \@ifstar{\oldnorm}{\oldnorm*}{\nabla
u}_{L^2(\R^d)}^2 + \frac{1}{2}\int_{\R^d} \(|u|^2-1\)^2\diff x.
\end{equation*}
See also \cite{KiMuVi16} for an analogous energy-critical problem.
The logarithmic nonlinearity was introduced in the context of
Schr\"odinger equations in \cite{BiMy76}, as it is the only
nonlinearity satisfying the following tensorization property: if
$u_1(t,x_1),\dots ,u_d(t,x_d)$ are solutions to the one-dimensional
equation, then $u(t,x):=u_1(t,x_1)\times\dots\times u_d(t,x_s)$ solves
the $d$-dimensional equation. This model has regained interest in
various domains of physics: quantum mechanics \cite{yasue},
quantum optics \cite{BiMy76,hansson,KEB00,buljan}, nuclear
physics \cite{Hef85}, Bohmian mechanics \cite{DMFGL03}, effective
quantum gravity \cite{Zlo10}, theory of
superfluidity and Bose-Einstein condensation \cite{BEC}.
The logarithmic model may generalize the Gross-Pitaevskii equation, used in
the case of two-body interaction, to the case of three-body
interaction; see \cite{Zlo10,Zlo11}.
\subsection{Cauchy problem}
\label{sec:cauchy}
From the mathematical point of view, the Cauchy problem for
\eqref{logGP} is more intricate than it may seem at first
glance. The difficulty does not lie in the behavior of the logarithm
at infinity, of course, but in its singularity at the origin. In the
case of vanishing boundary condition at infinity,
$u_0\in H^1(\R^d)$, the Cauchy problem was solved in
\cite{cazenave-haraux} (case $\lambda<0$, see also
\cite{HayashiM2018}) and \cite{carlesgallagher,GuLoNi10} (for any
$\lambda\in \R$), by constructing solutions of a regularized equation
converging to a solution to the exact equation, which turns out to be
unique. Even when nontrivial boundary conditions at infinity are
considered, like in the present paper, the singularity of the
logarithm at the origin is the main difficulty, as the cancellation of
the wave function is difficult (if not impossible) to rule out for
\emph{all} $(t,x)\in \R\times \R^d$. The nonlinearity fails to be
locally Lipschitz continuous, so the fixed point argument employed in
\cite{Gerard_Cauchy_GP} in the case of \eqref{eq:GP} is hard to
implement. Even if a solution of the form $u=1+H^1$ is considered in
the one dimensional case (where the $H^1$-norm control the $L^\infty$
norm), it is not obvious to make a solution global in time, even if
$\|u_0-1\|_{H^1}\ll 1$. Indeed, the conservation of the energy in the
context of \eqref{logGP} reads
$d\mathcal{E}_{\textnormal{logGP}}/dt=0$, where
\begin{equation*}
\mathcal{E}_{\textnormal{logGP}} (u) \coloneqq \@ifstar{\oldnorm}{\oldnorm*}{\nabla
u}_{L^2(\R^d)}^2 + \lambda \int_{\R^d} \( |u|^2\ln|u|^2-|u|^2+1\)\diff x.
\end{equation*}
Consider $F$ the antiderivative of the logarithm, such that $F(1) = 0$:
$F(y) = y \ln{y} - y + 1$ for $y > 0$, and $F(0) = 1$. The above
potential energy is the integral in space of $F(|u|^2)$.
We note that $F \geq 0$ on $[0, \infty)$, and Taylor's formula yields
\begin{equation*}
0\le F(y)=y^2\int_0^1\frac{1-s}{1+sy}\diff s\le y^2,
\end{equation*}
and so $0\le \mathcal{E}_{\textnormal{logGP}}(u)\le
\mathcal{E}_{\textnormal{GP}}(u)$. But this is not enough to solve
\eqref{logGP} in the energy space, neither locally or globally in
time. As a byproduct of our analysis, we will see that in low space
dimensions, $d\le 4$, the energy spaces for \eqref{logGP},
\begin{equation*}
E_\textnormal{logGP} \coloneqq \{ v \in H^1_\textnormal{loc} (\mathbb{R}^d) \, | \, \mathcal{E}_\textnormal{logGP} (v) < \infty \},
\end{equation*}
and
\eqref{eq:GP} coincide,
$E_{\textnormal{logGP}}= E_{\textnormal{GP}}$. Note that the energy
space $E_{\textnormal{GP}}$ was described very accurately in
\cite{Gerard_Cauchy_GP,PG08}.
We consider the notion of mild solution, based on Duhamel's formula:
\begin{equation*}
u(t) = e^{i t \Delta} u_\textnormal{in} - i \lambda \int_0^t e^{i (t - s) \Delta} \( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \) \diff s.
\end{equation*}
As we construct solutions to \eqref{logGP} as weak solutions, it is
sensible to ask whether or not these solutions are also mild
solutions, in the same fashion as in \cite{Segal63}. Our main result
regarding the Cauchy problem for \eqref{logGP} is the following:
\begin{theorem} \label{th:Cauchy_th_main}
For any $u_0 \in \mathcal{E}_\textnormal{logGP}$, there exists a
unique $u \in L^\infty_\textnormal{loc} (\mathbb{R},
E_\textnormal{logGP})$ solution to \eqref{logGP}. Moreover, the
flow of \eqref{logGP} enjoys the following properties:
\begin{itemize}
\item The energy is conserved, $\mathcal{E}_{\textnormal{logGP}}
(u(t))= \mathcal{E}_{\textnormal{logGP}}(u_0)$ for all $t\in
\R$.
\item $u - u_0\in
\mathcal{C}^0 (\mathbb{R}, L^2)$.
\item $u$ is also a mild solution.
\item If $\Delta u_0 \in L^2$, then
$u - u_0 \in W^{1, \infty}_\textnormal{loc}
(\mathbb{R}, L^2) \cap L^\infty_\textnormal{loc} (\mathbb{R},
H^2)$.
\end{itemize}
\end{theorem}
\subsection{Solitary and traveling waves}
There are many references regarding the existence, description and the
stability of solitary and traveling waves for \eqref{eq:GP}, as can
be seen for instance from the results evoked in \cite{volumeGP}. In
order to consider both solitary and traveling waves, we look for
solutions of the form
\begin{equation} \label{eq:trav_wave}
u (t,x) = e^{i \omega t} \phi (x - ct),
\end{equation}
where $\omega \in \mathbb{R}$, $c \in \mathbb{R}^d$ and $\phi \in
E_\textnormal{logGP}$. In the case $c=0$, we classically call the
solution a solitary waves, and if $c\not =0$, we call it traveling
wave (stationary wave if $\omega=c=0$). From the mathematical point of
view, the study of traveling
waves for \eqref{eq:GP} (with $\omega=0$) goes back to \cite{BeSa99},
and has known many
developments since; see e.g. \cite{BeGrSa08,BeGrSa09,Ch12,ChMa17,Ma13}
and references therein. In the one dimensional case, traveling waves for
\eqref{eq:GP} are either constant, or such that $0\le c<\sqrt 2$, and
given explicitly by
\begin{equation*}
\phi_c(x) =
\sqrt{1-\frac{c^2}{2}}\tanh\(\sqrt{1-\frac{c^2}{2}}\frac{x}{\sqrt
2}\)+i\frac{c}{\sqrt 2},
\end{equation*}
up to space translation and a multiplicative constant of modulus $1$; see
e.g. \cite{BeGrSa08}. This solution $u(t,x)=\phi_c(x-ct)$ is usually
refered to as
\emph{dark soliton}, and \emph{black soliton} (or \emph{kink solution})
in the case $c=0$ (the
only case where $\phi$ has a zero). In the case of \eqref{logGP}, we do
not have
such an explicit formula, but somehow a very similar description. In
the general case $d\ge 1$, we have:
\begin{theorem} \label{th:value_omega}
If $\phi \in E_\textnormal{logGP}$ and $u$ is a traveling wave of the form \eqref{eq:trav_wave}, then $\omega = 0$.
\end{theorem}
We now focus on the one-dimensional case. First, like for
\eqref{eq:GP}, if $|c|$ is too large, then there is no non-constant
traveling wave:
\begin{theorem} \label{th:trav_waves_1d}
Let $d = 1$ and $c^2 \ge 2 \lambda$. Any solution to
\eqref{logGP} of the form \eqref{eq:trav_wave} is constant.
\end{theorem}
We now come to the description of non-constant traveling waves:
\begin{theorem} \label{th:sol_wave}
Assume $d=1$ and $c$ such that $c^2 < 2 \lambda$. There exists a
unique non-constant traveling wave in the following sense: there
exists a traveling wave $\phi_c$ such that any non-constant
traveling wave $u$ of the form \eqref{eq:trav_wave} is such that
\begin{equation*}
\phi = e^{i \theta} \phi_c (\cdot- x_0),
\end{equation*}
%
for some constants $\theta, x_0 \in \mathbb{R}$. If $c\not =0$,
$\phi_c$ never vanishes.
%
In the case $c=0$, $\phi_0$ can be taken real-valued and
increasing, and then
\begin{equation*}
\lim_{x \to \pm \infty} \phi_0 (x) = \pm 1.
\end{equation*}
\end{theorem}
We emphasize that qualitatively, we obtain the same properties as the
black soliton, even though it is probably hopeless to get an explicit
expression in the case of \eqref{logGP}.
\subsection{Content and notations}
In Section~\ref{sec:prelim}, we reduce the study of \eqref{logGP}
in the energy space, by analyzing the potential energy and
characterizing the energy space $E_\textnormal{logGP}$. In
Section~\ref{sec:regularity}, we prove that any solution in the
energy space is unique, and that it is a mild solution. The heart of Theorem~\ref{th:Cauchy_th_main} is established in
Section~\ref{sec:construction}, where a weak solution is constructed
by an approximation procedure. In Section~\ref{sec:higher}, we show
that $H^2$ regularity is propagated by the flow, completing the proof
of Theorem~\ref{th:Cauchy_th_main}. We prove general results regarding
solitary and
traveling waves in Section~\ref{sec:dynamics}, in particular
Theorem~\ref{th:value_omega}. In Section~\ref{sec:solitary}, we
analyze stationary waves in the one-dimensional case
(Theorem~\ref{th:sol_wave} in the case $c=0$). One-dimensional traveling waves are
studied in Section~\ref{sec:traveling}, where we complete the proof of
Theorems~\ref{th:sol_wave} and \ref{th:trav_waves_1d}.
We
conclude this paper by open questions in Section~\ref{sec:open}.
\smallbreak
We define the $L^2$-bracket by
\begin{equation*}
\langle f, g \rangle = \int_{\R^d} f \overline{g} \diff x.
\end{equation*}
For $k$ an integer, we denote by $\mathcal{C}^k$ the class of $k$ times
continuously differentiable functions, and if $0<s<1$, by
$\mathcal{C}^{k,s}$ the class of $k$ times
continuously differentiable functions whose derivative of order $k$ is
H\"older continuous with exponent $s$. If $s=1$, $\mathcal{C}^{k,1}$ is
the class of $k$ times
continuously differentiable functions whose derivative of order $k$ is
Lipschitzian.
\section{Preliminary results}
\label{sec:prelim}
\subsection{Generalities}
First, we recall some properties about the logarithm.
The first property was discovered in \cite{cazenave-haraux}, and is
crucial for uniqueness issues related to \eqref{logGP}.
\begin{lem}[{\cite[Lemma~1.1.1]{cazenave-haraux}}] \label{lem:log_inequality}
There holds
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{\Im \left( (z_2 \ln \@ifstar{\oldabs}{\oldabs*}{z_2}^2 - z_1 \ln \@ifstar{\oldabs}{\oldabs*}{z_1}^2) (\overline{z_2} - \overline{z_1}) \right)} \leq 2 \@ifstar{\oldabs}{\oldabs*}{z_2 - z_1}^2, \qquad \forall z_1, z_2 \in \mathbb{C}.
\end{equation*}
\end{lem}
The next lemma measures the continuity of the nonlinearity in \eqref{logGP}:
\begin{lem} \label{lem:prop_lip_log}
For all $\varepsilon \in (0, 1)$ and $x, y \in \mathbb{C}$,
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{x \ln{\@ifstar{\oldabs}{\oldabs*}{x}^2} - y \ln{\@ifstar{\oldabs}{\oldabs*}{y}^2}} \leq C \Bigl( \@ifstar{\oldabs}{\oldabs*}{x}^\varepsilon \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{x}}} + \@ifstar{\oldabs}{\oldabs*}{y}^\varepsilon \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{y}}} \Bigr) \@ifstar{\oldabs}{\oldabs*}{x - y}^{1 - \varepsilon} + 2 \@ifstar{\oldabs}{\oldabs*}{x - y}.
\end{equation*}
%
\end{lem}
\begin{proof}
Let $x, y \in \mathbb{C}$ and assume without loss of generality that $\@ifstar{\oldabs}{\oldabs*}{x} \leq \@ifstar{\oldabs}{\oldabs*}{y}$. We also assume $x \neq 0$, otherwise the estimate is obvious.
%
Then, we have
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{x \ln{\@ifstar{\oldabs}{\oldabs*}{x}^2} - x \ln{\@ifstar{\oldabs}{\oldabs*}{y}^2}} = 2 \@ifstar{\oldabs}{\oldabs*}{x} \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{x}} - \ln{\@ifstar{\oldabs}{\oldabs*}{y}}}
\leq 2 \@ifstar{\oldabs}{\oldabs*}{x} \frac{\@ifstar{\oldabs}{\oldabs*}{\@ifstar{\oldabs}{\oldabs*}{x} - \@ifstar{\oldabs}{\oldabs*}{y}}}{\@ifstar{\oldabs}{\oldabs*}{x}} \leq 2 \@ifstar{\oldabs}{\oldabs*}{x - y}.
\end{equation*}
%
On the other hand,
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{x \ln{\@ifstar{\oldabs}{\oldabs*}{y}^2} - y \ln{\@ifstar{\oldabs}{\oldabs*}{y}^2}} \leq 2 \@ifstar{\oldabs}{\oldabs*}{x - y} \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{y}}} \leq 2^{1+\varepsilon} \@ifstar{\oldabs}{\oldabs*}{x - y}^{1 - \varepsilon} \@ifstar{\oldabs}{\oldabs*}{y}^\varepsilon \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{y}}},
\end{equation*}
%
hence the conclusion.
\end{proof}
Then, we also recall a result about norms on finite dimensional subspace of $L^2$.
\begin{lem} \label{lem:equiv_norm_finite_dim}
If $\{ f_k \}_{k \leq n}$ is a finite family of $L^2$ linearly
independent functions, then there exists $C > 0$ such that, for all $\lambda_k \in \mathbb{C}$,
\begin{equation*}
\max \@ifstar{\oldabs}{\oldabs*}{\lambda_k} \leq C \@ifstar{\oldnorm}{\oldnorm*}{\sum_{k = 0}^n \lambda_k f_k}_{L^2}.
\end{equation*}
\end{lem}
\begin{proof}
Since $\{ f_k \}$ is a basis of $\operatorname{Vect} \{ f_k \}$, $(\lambda_k) \mapsto \sum_{k = 0}^m \lambda_k f_k$ is an isomorphism from $\mathbb{C}^n$ into $\operatorname{Vect} \{ f_k \}$. Since we are in finite dimension, the conclusion easily follows.
\end{proof}
\subsection{Potential energy}
We start by a result on the potential energy $\mathcal{E}_{\textnormal{\textnormal{pot}}}$, given by
\begin{equation*}
\mathcal{E}_{\textnormal{\textnormal{pot}}} (v) \coloneqq \int \Bigl( \@ifstar{\oldabs}{\oldabs*}{v}^2 \ln{\@ifstar{\oldabs}{\oldabs*}{v}^2} - \@ifstar{\oldabs}{\oldabs*}{v}^2 + 1 \Bigr) \diff x.
\end{equation*}
Let
\begin{equation*}
E_\textnormal{\textnormal{pot}} (v) \coloneqq \int ( \@ifstar{\oldabs}{\oldabs*}{v} - 1 )^2 \ln{(2 + \@ifstar{\oldabs}{\oldabs*}{v})} \diff x.
\end{equation*}
We know that both previous quantities are non-negative. The following result shows that they are actually equivalent.
\begin{lem} \label{lem:energy_equiv}
There exists $K_0 > 0$ such that for all $v \in E_\textnormal{logGP}$,
\begin{equation*}
\frac{1}{K_0} E_\textnormal{pot} (v) \leq \mathcal{E}_\textnormal{pot} (v) \leq K_0 E_\textnormal{pot} (v).
\end{equation*}
%
\end{lem}
\begin{proof}
Taylor formula yields, for $y\ge 0$,
\begin{equation*}
y^2\ln y^2-y^2+1 = 4 \(y-1\)^2\int_0^1
\(1+\ln\(1-s+sy\)\)(1-s)ds.
\end{equation*}
Distinguishing $y<1$ and $y>1$, we get:
\begin{equation*}
\ln \(2+y\)\lesssim \int_0^1\(1+\ln\(1-s+sy\)\)(1-s)ds\lesssim
\ln \(2+y\),
\end{equation*}
and the result follows, by considering $y=|v|$.
\end{proof}
This second functional $E_\textnormal{pot}$ appears to be more
convenient than $\mathcal{E}_\textnormal{pot}$. For instance, since
$\ln{2} \leq \ln{(2 + \@ifstar{\oldabs}{\oldabs*}{x})} \leq \ln{2} + C_\delta \@ifstar{\oldabs}{\oldabs*}{\@ifstar{\oldabs}{\oldabs*}{x} -
1}^\delta$ for all $\delta > 0$, we can relate $E_\textnormal{pot}$
to the $L^2$-norm of $\@ifstar{\oldabs}{\oldabs*}{v} - 1$, but also to any $L^p$-norm, for
$p > 2$ as follows:
\begin{lem} \label{lem:E_pot_Lp}
Let $p>2$. There exists $C_p$ such that for all $v \in E_\textnormal{logGP}$,
\begin{equation*}
\ln{(2)} \, \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{v} - 1}_{L^2}^2 \leq E_\textnormal{pot} (v) \leq \ln{(2)} \, \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{v} - 1}_{L^2}^2 + C_p \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{v} - 1}_{L^p}^p.
\end{equation*}
\end{lem}
Last, we state a differential result about the potential
energy, which follows from standard arguments, that may be found in
\cite{Cazenave_semlin_lognls} ([Proposition~1.5.2, proof of
Theorem~3.3.5. Recall that the function $F$ is given on $(0,\infty)$
by $F(y)=y\ln y-y+1$.
\begin{lem} \label{lem:deriv_pot_en_regu}
Let $I$ be an open interval. If $u \in u_0 + \mathcal{C}^0 (I,
L^2) \cap L^\infty_\textnormal{loc} (I, H^1) \cap \mathcal{C}^1 (I,
H^{-1})$ and $\varepsilon > 0$, then $\int (1 - \varepsilon
\Delta)^{-2} F(\@ifstar{\oldabs}{\oldabs*}{u (t)}^2) \diff x$ is well defined for all $t
\in I$ and
\begin{equation*}
\frac{\diff}{\diff t} \int (1 - \varepsilon \Delta)^{-2} F(\@ifstar{\oldabs}{\oldabs*}{u (t)}^2) \diff x = 2 \< (1 - \varepsilon \Delta)^{-1} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (t)}^2} \Bigr), (1 - \varepsilon \Delta)^{-1} \partial_t u (t) \>.
\end{equation*}
%
Moreover, for all $t \in I$,
\begin{equation*}
\int (1 - \varepsilon \Delta)^{-2} F(\@ifstar{\oldabs}{\oldabs*}{u (t)}^2) \diff x \underset{\varepsilon \rightarrow 0}{\longrightarrow} \mathcal{E}_\textnormal{pot} (u(t)).
\end{equation*}
\end{lem}
\subsection{Energy space}
As for the energy space, Lemma~\ref{lem:energy_equiv} leads to
\begin{equation*}
E_\textnormal{logGP} = \{ v \in H^1_\textnormal{loc} (\mathbb{R}^d) \, | \, \nabla v \in L^2(\R^d) \, \textnormal{and} \, E_\textnormal{pot} (v) < \infty \}.
\end{equation*}
We prove an even more explicit description of $E_\textnormal{logGP}$:
\begin{lem} \label{lem:E_logGP_descr}
The energy space is characterized by:
\begin{align*}
E_\textnormal{logGP} &= \{ v \in H^1_\textnormal{loc} (\R^d)\, | \, \nabla
v \in L^2(\R^d) \, \textnormal{and} \, \@ifstar{\oldabs}{\oldabs*}{v} - 1 \in L^2 (\R^d)\} \\
&= \{ v \in
H^1_\textnormal{loc} (\R^d)\, | \, \nabla v \in L^2(\R^d) \, \textnormal{and} \,
\@ifstar{\oldabs}{\oldabs*}{v} - 1 \in H^1(\R^d) \} .
\end{align*}
Moreover, there exists $C>0$ such that for all $v \in E_\textnormal{logGP}$, $\@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{v} - 1}_{H^1}^2 \leq C \mathcal{E}_\textnormal{logGP} (v)$.
\end{lem}
\begin{proof}
First, we show that $E_\textnormal{logGP} \subset \{ v \in
H^1_\textnormal{loc} \, | \, \nabla v \in L^2 \, \textnormal{and}
\, \@ifstar{\oldabs}{\oldabs*}{v} - 1 \in L^2 \}$.
Let $v \in E_\textnormal{logGP}$: $\@ifstar{\oldnorm}{\oldnorm*}{\nabla v}_{L^2} < \infty$, and $\mathcal{E}_\textnormal{pot} (v) < \infty$, which is equivalent to $E_\textnormal{pot} (v) < \infty$. Lemma~\ref{lem:E_pot_Lp} yields $\@ifstar{\oldabs}{\oldabs*}{v} - 1 \in L^2$.
%
\medskip
Conversely, let $v \in {H}^1_\textnormal{loc}$ such that $\@ifstar{\oldabs}{\oldabs*}{v} -
1 ,\nabla v\in L^2$. Then $\nabla (\@ifstar{\oldabs}{\oldabs*}{v} - 1) = \nabla v \cdot \frac{v}{\@ifstar{\oldabs}{\oldabs*}{v}}$ and $\@ifstar{\oldabs}{\oldabs*}{\nabla (\@ifstar{\oldabs}{\oldabs*}{v} - 1)} \leq \@ifstar{\oldabs}{\oldabs*}{\nabla v}$, which shows that $f = \@ifstar{\oldabs}{\oldabs*}{v} - 1 \in H^1$. Since $f \geq -1$, we readily have
\begin{equation*}
E_\textnormal{pot} (v) = \int \@ifstar{\oldabs}{\oldabs*}{f}^2 \ln{(3 + f)} \diff x
\leq \int \@ifstar{\oldabs}{\oldabs*}{f}^2 \ln{(3 + \@ifstar{\oldabs}{\oldabs*}{f})} \diff x \lesssim \int \@ifstar{\oldabs}{\oldabs*}{f}^2 \(1+ \@ifstar{\oldabs}{\oldabs*}{f}^\varepsilon\) \diff x < \infty,
\end{equation*}
where $\varepsilon>0$ is arbitrarily small, and where we have used Sobolev
embedding to conclude.
%
\medskip
From the previous arguments, if $v \in E_\textnormal{logGP}$,
the $L^2$-norm of $|v|-1$ can be estimated thanks to Lemma~\ref{lem:E_pot_Lp} and Lemma~\ref{lem:energy_equiv}, whereas $\@ifstar{\oldabs}{\oldabs*}{\nabla (\@ifstar{\oldabs}{\oldabs*}{v} - 1)} \leq \@ifstar{\oldabs}{\oldabs*}{\nabla v}$, hence the conclusion.
\end{proof}
\begin{cor} \label{cor:v_ln_v_Lp}
For all $\delta\ge 1$, $\varepsilon>0$, and $p$ such that
\begin{equation} \label{eq:p_cases}
p,p+\varepsilon \in
\begin{cases}
(2, \infty] \quad \text{if } d = 1, \\
(2, \infty) \quad \text{if } d = 2, \\
(2, 2^*] \quad \text{if } d \geq 3,\quad 2^* \coloneqq \frac{2d}{d-2},
\end{cases}
\end{equation}
there exists $C_{p,\varepsilon,\delta}$ such that
if $v \in E_\textnormal{logGP}$, then $v \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{v}^2}}^\delta
\in L^p(\R^d)$ and
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{v \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{v}^2}}^\delta}_{L^p} \leq C_{p, \varepsilon, \delta} \, (\mathcal{E}_\textnormal{logGP} (v)^{1/2} + \mathcal{E}_\textnormal{logGP} (v)^{(1 + \varepsilon)/2}).
\end{equation*}
\end{cor}
\begin{proof}
We know that $\@ifstar{\oldabs}{\oldabs*}{v} - 1 \in H^1$ and, since $\delta \geq 1$,
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{ v} \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{v}}^2}^\delta \le C_\delta \@ifstar{\oldabs}{\oldabs*}{\@ifstar{\oldabs}{\oldabs*}{v} -
1} (\ln{(2 + \@ifstar{\oldabs}{\oldabs*}{\@ifstar{\oldabs}{\oldabs*}{v} - 1})})^\delta \leq
C_{\varepsilon, \delta} (\@ifstar{\oldabs}{\oldabs*}{\@ifstar{\oldabs}{\oldabs*}{v} - 1} + \@ifstar{\oldabs}{\oldabs*}{\@ifstar{\oldabs}{\oldabs*}{v} -
1}^{1+\frac{\varepsilon}{p}}).
\end{equation*}
%
This inequality is easily established, for instance by considering
separately the
regions
$\{|v|<1/2\}$, $\{|v|>2\}$, and $\{1/2\le |v|\le 2\}$.
The result then follows from Sobolev embedding,
Lemma~\ref{lem:E_logGP_descr} and Lemma~\ref{lem:energy_equiv}.
\end{proof}
\begin{rem}
This corollary reveals an important difference with the case of
vanishing boundary condition at infinity: in
\cite{GuLoNi10,carlesgallagher}, momenta in $L^2$
($\||x|^\alpha u\|_{L^2}$ for some $\alpha\in (0,1]$) are considered
in order to control the nonlinearity in the region $\{|u|<1\}$. The
situation is obviously different in the case of nontrivial boundary
condition at infinity, and the above corollary is crucial for the
rest of this paper.
\end{rem}
Lemma~\ref{lem:E_logGP_descr} shows
that for $d\le 4$, $E_\textnormal{logGP} = E_\textnormal{GP}$, in view of the
identity
\begin{equation*}
|u|^2-1 = \(|u|-1\)\(|u|-1+2\),
\end{equation*}
and the Sobolev embedding $H^1(\R^d)\subset L^4(\R^d)$ for $d\le 4$.
In particular,
G\'erard \cite{Gerard_Cauchy_GP} proved that $E_\textnormal{GP} + H^1 \subset
E_\textnormal{GP}$ for $d \ge 1$, and therefore the same holds for
$E_\textnormal{logGP}$ when $d\le 4$. This is also true in all dimensions :
\begin{lem} \label{lem:Energy_plus_H1}
Let $d \geq 1$. For all $v \in E_\textnormal{logGP}$ and for all
$f \in H^1$, we have $v + f \in E_\textnormal{logGP}$. In
addition, if $p > 2$ satisfies \eqref{eq:p_cases}, there exists
$C_p$ such that for every such $v$ and $f$,
\begin{equation*}
\mathcal{E}_\textnormal{logGP} (v + f) \leq C_p
\mathcal{E}_\textnormal{logGP} (v) + C_p
\mathcal{E}_\textnormal{logGP} (v)^p + C _p\@ifstar{\oldnorm}{\oldnorm*}{f}_{H^1}^2 +
C_p \@ifstar{\oldnorm}{\oldnorm*}{f}_{H^1}^p.
\end{equation*}
%
\end{lem}
\begin{proof}
Let $v \in E_\textnormal{logGP}$ and $f \in H^1$. We know that
$\nabla v ,\nabla f\in L^2$, so $\nabla(v + f) \in L^2$. From Lemma \ref{lem:E_logGP_descr}, there only remains to show that $\@ifstar{\oldabs}{\oldabs*}{v + f} - 1 \in L^2$. For this, we have for all $x, y \in \mathbb{C}$
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{x + y} = \@ifstar{\oldabs}{\oldabs*}{x} + \int_0^1 \frac{x + r y}{\@ifstar{\oldabs}{\oldabs*}{x + r y}} \cdot y \diff r.
\end{equation*}
%
Applying this equality to $u$ and $f$, we get
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{u + f} - 1 = \@ifstar{\oldabs}{\oldabs*}{u} - 1 + \int_0^1 \frac{u + r f}{\@ifstar{\oldabs}{\oldabs*}{u + r f}} \cdot f \diff r,
\end{equation*}
%
hence
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{\@ifstar{\oldabs}{\oldabs*}{u + f} - 1} \leq \@ifstar{\oldabs}{\oldabs*}{\@ifstar{\oldabs}{\oldabs*}{u} - 1} + \@ifstar{\oldabs}{\oldabs*}{f},
\end{equation*}
%
and, for any $p \in [1, \infty]$,
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{u + f} - 1}_{L^p} \leq \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{u} - 1}_{L^p} + \@ifstar{\oldnorm}{\oldnorm*}{f}_{L^p}
\end{equation*}
%
Moreover, Lemmas \ref{lem:energy_equiv} and \ref{lem:E_pot_Lp} yield, for any $p > 2$ satisfying \eqref{eq:p_cases},
\begin{align*}
\mathcal{E}_\textnormal{logGP} (v + f) &\leq \@ifstar{\oldnorm}{\oldnorm*}{\nabla v + \nabla f}^2 + C \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{v + f} - 1}_{L^2}^2 + C_p \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{v + f} - 1}_{L^p}^p \\
&\lesssim \@ifstar{\oldnorm}{\oldnorm*}{\nabla v}_{L^2}^2 + \@ifstar{\oldnorm}{\oldnorm*}{\nabla f}_{L^2}^2 + \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{u} - 1}_{L^2}^2 + \@ifstar{\oldnorm}{\oldnorm*}{f}_{L^2}^2 + \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{u} - 1}_{L^p}^p + \@ifstar{\oldnorm}{\oldnorm*}{f}_{L^p}^p \\
&\lesssim \mathcal{E}_\textnormal{logGP} (v) + \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{u} - 1}_{H^1}^p + \@ifstar{\oldnorm}{\oldnorm*}{f}_{H^1}^2 + \@ifstar{\oldnorm}{\oldnorm*}{f}_{H^1}^p \\
&\lesssim \mathcal{E}_\textnormal{logGP} (v) + \mathcal{E}_\textnormal{logGP} (v)^p + \@ifstar{\oldnorm}{\oldnorm*}{f}_{H^1}^2 + \@ifstar{\oldnorm}{\oldnorm*}{f}_{H^1}^p. \qedhere
\end{align*}
\end{proof}
We conclude by adapting \cite[Lemma~1]{Gerard_Cauchy_GP}.
Introduce the Zhidkov space
\begin{equation*}
X^1(\R^d) = \{u\in L^\infty(\R^d),\ \nabla u\in L^2(\R^d)\}.
\end{equation*}
\begin{lem}\label{lem:zhidkov}
Let $d\ge 1$. We have $E_\textnormal{logGP}\subset X^1(\R^d)+H^1(\R^d)$.
\end{lem}
\begin{proof}
Proceding like in the proof of \cite[Lemma~1]{Gerard_Cauchy_GP},
consider $\chi\in \mathcal C_0^\infty(\C)$ a cutoff function such
that $0\le \chi\le 1$, $\chi(z)=1$ for $|z|\le 2$ and $\chi(z)=0$
for $|z|\ge 3$. Decompose $u\in E_\textnormal{logGP}$ as
\begin{equation*}
u=u_1+u_2,\quad u_1=\chi(u)u,\quad u_2=\(1-\chi(u)\)u.
\end{equation*}
Then $\|u_1\|_{L^\infty}\le 3$, and since $|u|\ge 2$ on the support of
$u_2$, for any $\delta>1$, we may find $C_\delta$ such that
\begin{equation*}
|u_2|\le C_\delta \left||u|-1\right|^\delta.
\end{equation*}
Since $|u|-1\in H^1(\R^d)$ from Lemma~\ref{lem:E_logGP_descr},
$|u|-1\in L^{2\delta}(\R^d)$ by Sobolev embedding (provided that
$\delta\le\frac{d}{d-2}$ if $d\ge 3$), hence $u_2\in L^2$. The
properties $\nabla u_1,\nabla u_2\in L^2$ are straightforward, and we
refer to \cite{Gerard_Cauchy_GP} for details.
\end{proof}
\section{Regularity and uniqueness}
\label{sec:regularity}
We begin by proving the following point in Theorem \ref{th:Cauchy_th_main}:
\begin{lem} \label{lem:u_in_plusH1}
Let $I$ an interval. If $u \in L^\infty_\textnormal{loc} (I,
E_\textnormal{logGP})$ (that is, $t\mapsto \mathcal{E}_\textnormal{logGP}
(u(t)) \in L^\infty_\textnormal{loc} (I)$) satisfies \eqref{logGP}
in $\mathcal{D}' (I \times \mathbb{R}^d)$, then $u - u_0 \in
\mathcal C^0 (I, L^2)$.
\end{lem}
\begin{proof}
Since $u \in L^\infty_\textnormal{loc} (I, E_\textnormal{logGP})$,
we know from Corollary~\ref{cor:v_ln_v_Lp} that $u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2}
\in L^\infty_\textnormal{loc} (I, L^2)$. Moreover, $\Delta u \in
L^\infty_\textnormal{loc} (I, H^{-1})$. Therefore, \eqref{logGP}
yields $\partial_t u =
\partial_t (u - u_0) \in L^\infty_\textnormal{loc} (I, H^{-1})$,
with $u(0) - u_0 = 0$. This proves that $u - u_0 \in \mathcal{C}^0
(I, H^{-1})$. Since $\nabla(u - u_0) \in L^\infty_\textnormal{loc}
(I, L^2)$, we obtain the result by interpolation.
\end{proof}
This result allows us to infer the uniqueness of the solution of \eqref{logGP}.
\begin{theorem} \label{th:regu_sol}
Let $u_0 \in E_\textnormal{logGP}$ and $I$ be a bounded interval
containing $0$. There exists at most one solution $u \in L^\infty (I, E_\textnormal{logGP})$ to \eqref{logGP} in $\mathcal{D}' (I \times \mathbb{R}^d)$.
\end{theorem}
\begin{proof}
Let $u,v \in L^\infty (I, E_\textnormal{logGP})$ solve
\eqref{logGP}. Then $u (t) - u_0$ and $v (t) - u_0$ are continuous
from $I$ to $L^2$ from Lemma \ref{lem:u_in_plusH1}. Therefore, $w
\coloneqq u - v \in \mathcal C^0(I;L^2)$. This error satisfies (in $\mathcal{D}' (I \times \mathbb{R}^d)$)
\begin{equation*}
i \partial_t w + \Delta w = \lambda (u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2} - v \ln{\@ifstar{\oldabs}{\oldabs*}{v}^2}).
\end{equation*}
%
Since on the one hand $\Delta w \in L^\infty (I, H^{-1})$, and on
the other hand $u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2},v \ln{\@ifstar{\oldabs}{\oldabs*}{v}^2}\in L^\infty (I,
L^2)$, the previous equality is also satisfied in $L^\infty (I,
H^{-1})$. As $w \in L^\infty (I, H^1)$, we can take the $H^{-1}
\times H^1$ bracket against $w$ of the previous equality, which
yields
\begin{equation} \label{eq:pde_w}
i \langle \partial_t w, w \rangle_{H^{-1}, H^1} - \@ifstar{\oldnorm}{\oldnorm*}{\nabla w}_{L^2}^2 = \lambda \int (u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2} - v \ln{\@ifstar{\oldabs}{\oldabs*}{v}^2}) (\overline{u - v}) \diff x.
\end{equation}
%
Thanks to Lemma \ref{lem:log_inequality}, we have
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{\Im \int (u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2} - v \ln{\@ifstar{\oldabs}{\oldabs*}{v}^2}) (\overline{u - v}) \diff x} \leq 2 \@ifstar{\oldnorm}{\oldnorm*}{u - v}_{L^2}^2 = 2 \@ifstar{\oldnorm}{\oldnorm*}{w}_{L^2}^2.
\end{equation*}
%
Therefore, taking the imaginary part of \eqref{eq:pde_w} leads to
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{\Re \langle \partial_t w, w \rangle_{H^{-1}, H^1}} = \frac{1}{2} \@ifstar{\oldabs}{\oldabs*}{\frac{\diff}{\diff t} \@ifstar{\oldnorm}{\oldnorm*}{w}_{L^2}^2} \leq 2 \lambda \@ifstar{\oldnorm}{\oldnorm*}{w}_{L^2}^2.
\end{equation*}
%
Since $w(0) = 0$, Gronwall lemma concludes the proof.
\end{proof}
We end this section with a link between regularity and mild solution.
\begin{lem} \label{lem:mild_sol}
Let $I$ be an open interval of $\mathbb{R}$, $u_0 \in
E_\textnormal{logGP}$ and $u \in L^\infty_\textnormal{loc} (I,
E_\textnormal{logGP})$ satisfying \eqref{logGP} in $\mathcal{D}'
(I \times \mathbb{R}^d)$. Then $u$ is a mild solution, and
\begin{equation*}
\int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr)
\diff s \in H^1(\R^d)\quad \text{for all }t\in I.
\end{equation*}
\end{lem}
\begin{proof}
From Corollary~\ref{cor:v_ln_v_Lp}, we have $u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2} \in L^\infty_\textnormal{loc} (I, L^2)$, thus we can define
\begin{equation*}
v_\textnormal{NL} (t) = - i \lambda \int_0^t e^{-i s \Delta} u (s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \diff s \in \mathcal{C}^0 (I, L^2) \cap W^{1, \infty}_\textnormal{loc} (I, H^{-2}).
\end{equation*}
%
On the other hand, since $\nabla u_0 \in L^2$ and $u_0 \in
L^\infty + L^2$ (from Lemma~\ref{lem:zhidkov}), we can define
$e^{i t \Delta} u_0$ and we have $e^{i t \Delta} u_0 - u_0 \in
\mathcal{C}^0 (I, L^2) \cap \mathcal{C}^1 (I, H^{-1})$ (see
\cite{Gerard_Cauchy_GP}).
%
Therefore,
\begin{equation*}
v \coloneqq e^{i t \Delta} u_0 +e^{i t \Delta}
v_\textnormal{NL} (t) \in u_0 + \mathcal{C}^0 (I, L^2) \cap
W^{1, \infty}_\textnormal{loc} (I, H^{-2}),
\end{equation*}
%
and we can compute
\begin{align*}
\partial_t v &= i \Delta e^{i t \Delta} \Bigl[ u_0 - i \lambda
\int_0^t e^{- i s \Delta} u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2}
\diff s \Bigr] \\
&\quad+ e^{i t \Delta} \partial_t \Bigl[ u_0 - i \lambda \int_0^t e^{- i s \Delta} u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \diff s \Bigr] \\
&= i \Delta v + e^{i t \Delta} \Bigl[ - i \lambda e^{- i t \Delta} u(t) \ln{\@ifstar{\oldabs}{\oldabs*}{u (t)}^2} \Bigr] \\
&= i \Delta v - i \lambda u(t) \ln{\@ifstar{\oldabs}{\oldabs*}{u (t)}^2},
\end{align*}
%
where the equality is to be understood in
$L^\infty_\textnormal{loc} (I, H^{-2})$. Let $w = u - v$. Then, $w
\in \mathcal{C}^0 (I, L^2)$ from Lemma \ref{lem:u_in_plusH1} along
with the previous arguments, and with the previous equality, we
have, in $L^\infty_\textnormal{loc} (I, H^{-2})$,
\begin{equation*}
\partial_t w - i \Delta w = 0.
\end{equation*}
%
Since $w (0) = 0$ by construction, we get $w = 0$, which proves the mild formulation.
%
Last, we know that $\nabla (u(t) - u_0) \in L^2$ and $\nabla (e^{i t \Delta} u_0 - u_0) \in L^2$ for all $t \in I$, therefore
\begin{equation*}
\nabla v_\textnormal{NL} (t) \in L^2, \qquad \forall t \in I.
\end{equation*}
%
Since we already know that $v_\textnormal{NL} (t) \in L^2$, we get the conclusion.
\end{proof}
\section{Construction of a solution}
\label{sec:construction}
In this section, we construct a global solution $u \in L^\infty
(\mathbb{R}, \mathcal{E}_\textnormal{logNLS})$ to \eqref{logGP}. We
adapt the method used by Ginibre and Velo
\cite{Ginibre_Velo__Cauchy_NLS_revisited} to construct global weak
solutions to non-linear Schrödinger equations (NLS) by
compactness. Here, the framework is different since the solution is
not in $L^2$. However, we know that we should have $u (t) - u_0 \in
H^1$. Therefore, we will approximate the solution on $u_0 + X_m$ with
$X_m$ a sequence of finite dimensional linear subspaces approximating $H^1$.
\begin{rem}
In \cite{carlesgallagher}, for vanishing boundary condition at
infinity, another approximation is considered,
consisting in removing the singularity of the logarithm by
saturating the nonlinearity, for $\varepsilon>0$,
\begin{equation}\label{eq:logNLSreg}
i \partial_t u^\varepsilon + \Delta u^\varepsilon = \lambda u^\varepsilon \ln\({\@ifstar{\oldabs}{\oldabs*}{u^\varepsilon}^2+\varepsilon}\), \quad
u^\varepsilon_{\mid t=0} =u_0,
\end{equation}
and letting $\varepsilon\to 0$. This approximation has the advantage of working
whichever the sign of $\lambda$ is, while the approach introduced
initially in \cite{cazenave-haraux} (see also \cite{HayashiM2018})
seems to be bound to the
nondispersive case $\lambda<0$. We do not consider this approach here.
\end{rem}
\subsection{Finite dimensional approximation}
Let $\{ w_j \}_{j \in \mathbb{N}}$ an Hilbertian basis of $L^2$ with
all $w_j \in H^1$ and $m \in \mathbb{N}$. One may think for instance
of Hermite functions, all the more since in
Lemma~\ref{lem:dt_um_unif_bound} below, in order to prove the
propagation of $\dot H^2$ regularity, we further require $w_j\in
H^2$. We take $X_m \coloneqq
\operatorname{Vect} (w_j)_{j \leq m}$ and look for an approximation
$u_m$ of the form
\begin{equation} \label{eq:form_u_m}
u_m (t,x) = u_0(x) + \sum_{k=0}^m g_{m, k} (t) w_k(x) \eqqcolon u_0(x) +
\varphi_m (t,x),
\end{equation}
satisfying
\begin{equation} \label{eq:eq_u_m}
\< w_j, i \partial_t u_m + \Delta u_m - \lambda u_m
\ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2} \>_{H^1, H^{-1}} = 0,\quad 0\le j\le m,
\end{equation}
with initial condition $u_m (t) = u_0$, which is equivalent to $g_{m, k} (0) = 0$ for all $0 \leq k \leq m$.
By substitution, \eqref{eq:eq_u_m} is equivalent to
\begin{equation} \label{eq:ode_g_mk}
\begin{aligned}
i \dot{g}_{m, j} - \< \nabla w_j, \nabla u_0 \> - &\sum_{k = 0}^m
g_{m, k} (t) \< \nabla w_j, \nabla w_k \> \\
&= \lambda \< w_j, f \(
u_0 + \sum_{k=0}^m g_{m, k} (t) w_k \) \>,
\end{aligned}
\end{equation}
where $f(x) = x \ln{\@ifstar{\oldabs}{\oldabs*}{x}^2}$.
First, we study the last term.
\begin{lem} \label{lem:cont_log_term_um}
The map $\varphi \mapsto \< w_j, f (\varphi) \>$ is well defined
and continuous from $u_0 + H^1$ into $\mathbb{C}$ for any $j \in
\mathbb{N}$. More precisely, it is $\mathcal{C}^{0, \varepsilon} (u_0
+ H^1, \mathbb{C})$ for all $\varepsilon \in (0, 1)$.
\end{lem}
\begin{proof}
By Corollary~\ref{cor:v_ln_v_Lp} along with
Lemma~\ref{lem:Energy_plus_H1}, we know that the right-hand side
is well
defined. Moreover, at fixed $j$, taking $\varphi, \psi \in u_0 +
H^1$, we get
\begin{align*}
\@ifstar{\oldabs}{\oldabs*}{\< w_j, f ( \varphi ) \> - \< w_j, f ( \psi ) \>} &= \@ifstar{\oldabs}{\oldabs*}{\< w_j, f ( \varphi ) - f ( \psi ) \>} \leq \int \@ifstar{\oldabs}{\oldabs*}{w_j} \@ifstar{\oldabs}{\oldabs*}{f ( \varphi ) - f ( \psi )}
\diff x \\
&\leq C \int \@ifstar{\oldabs}{\oldabs*}{w_j} \Bigl( \@ifstar{\oldabs}{\oldabs*}{\varphi}^\varepsilon
\@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{\varphi}}} + \@ifstar{\oldabs}{\oldabs*}{\psi}^\varepsilon
\@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{\psi}}} \Bigr) \@ifstar{\oldabs}{\oldabs*}{\varphi - \psi}^{1 -
\varepsilon} \diff x \\
&\quad+ 2 \int \@ifstar{\oldabs}{\oldabs*}{w_j} \@ifstar{\oldabs}{\oldabs*}{\varphi -
\psi} \diff x,
\end{align*}
%
for any $\varepsilon \in (0, 1)$, where we have used Lemma \ref{lem:prop_lip_log}.
The second term of the last inequality is obviously Lipschitzian with
respect to the $L^2$-norm of the difference. As for the first
term, by H\"older inequality with exponents $2$,
$\frac{2}{\varepsilon}$ and $\frac{2}{1 - \varepsilon}$, we
control it by
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{w_j}_{L^2} \@ifstar{\oldnorm}{\oldnorm*}{\varphi - \psi}_{L^2}^{1 - \varepsilon} \Bigl( \@ifstar{\oldnorm}{\oldnorm*}{\varphi \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{\varphi}}}^\frac{1}{\varepsilon}}_{L^2}^\varepsilon + \@ifstar{\oldnorm}{\oldnorm*}{\psi \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{\psi}}}^\frac{1}{\varepsilon}}_{L^2}^\varepsilon \Bigr).
\end{equation*}
%
Now, $\@ifstar{\oldnorm}{\oldnorm*}{\varphi \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{\varphi}}}^\frac{1}{\varepsilon}}_{L^2}$ and $\@ifstar{\oldnorm}{\oldnorm*}{\psi \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{\psi}}}^\frac{1}{\varepsilon}}_{L^2}$ are locally bounded thanks to Corollary \ref{cor:v_ln_v_Lp} and Lemma \ref{lem:Energy_plus_H1} once again. These estimates yield the conclusion.
\end{proof}
\begin{lem} \label{lem:exist_sol_app}
There exists a solution $\{ g_{m, k} \}_{k \leq m}$ to
\eqref{eq:ode_g_mk} on a maximal time interval $(- T_m, T^m)$,
meaning that $T^m < \infty$ if and only if
\begin{equation*}
\limsup_{t \rightarrow
T^m} \sup_{j\le m}\@ifstar{\oldabs}{\oldabs*}{g_{m, j} (t)} = \infty.
\end{equation*}
(And similarly for $T_m$.) Moreover, the $g_{m, k}$'s are
$\mathcal{C}^{1, 1 - \varepsilon} (- T_m, T^m)$ for all
$\varepsilon \in (0, 1)$.
\end{lem}
\begin{proof}
The last term of \eqref{eq:ode_g_mk} is continuous (and even
$\mathcal{C}^{0, \varepsilon}$ for all $\varepsilon \in (0, 1)$)
with respect to the $g_{m, k}$'s thanks to Lemma~\ref{lem:cont_log_term_um}, and so are obviously the other
terms. The conclusion comes from Peano Theorem and the fact that
\eqref{eq:ode_g_mk} is an autonomous ODE.
\end{proof}
We will then show that such a solution is global. First, we prove an intermediate result.
\begin{lem} \label{lem:nabla_lin_indep}
For all $m \in \mathbb{N}$, $(\nabla w_k)_{k \leq m}$ are linearly independent.
\end{lem}
\begin{proof}
Let $\lambda_k \in \mathbb{C}$ such that $\sum_{k=0}^m \lambda_k \nabla w_k = 0$. Then $\nabla \psi = 0$, with $\psi \coloneqq \sum_{k=0}^m \lambda_k w_k \in H^1$. Therefore, $\psi = 0$, and we conclude by the fact that $\{ w_k \}$ is an Hilbertian basis of $L^2$.
\end{proof}
\begin{lem} \label{lem:appr_sol_global}
The solution given by Lemma \ref{lem:exist_sol_app} is
global. Moreover, it satisfies
\begin{equation*}
\mathcal{E}_\textnormal{logGP} (u_m (t)) =
\mathcal{E}_\textnormal{logGP} (u_0), \quad\text{and}
\quad\@ifstar{\oldnorm}{\oldnorm*}{\nabla \varphi_m (t)}_{L^2} \leq 2
\sqrt{\mathcal{E}_\textnormal{logGP} (u_0)}\quad \text{for all
}t \in \R.
\end{equation*}
\end{lem}
\begin{proof}
For this, we only have to prove that all the $g_{m, k}$s are (locally) bounded.
Come back to \eqref{eq:eq_u_m}: by multiplying by $\dot{g}_{m, j}$
and summing over $j \leq m$, we get, in view of \eqref{eq:form_u_m},
\begin{equation*}
\< \partial_t u_m, i \partial_t u_m +\Delta u_m - \lambda u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2} \> = 0.
\end{equation*}
%
By taking the real part of this equation, we obtain
\begin{equation*}
\frac{\diff}{\diff t} \mathcal{E}_\textnormal{logGP} (u_m (t)) = 0.
\end{equation*}
%
Therefore, $\mathcal{E}_\textnormal{logGP} (u_m (t)) =
\mathcal{E}_\textnormal{logGP} (u_0)$ for all $t \in (- T_m,
T^m)$. In particular, $\@ifstar{\oldnorm}{\oldnorm*}{\nabla u_m (t)}_{L^2}$ is uniformly
bounded, and thus so is $\@ifstar{\oldnorm}{\oldnorm*}{\nabla \varphi_m (t)}_{L^2}$: for $
t \in (- T_m, T^m)$,
\begin{align*}
\@ifstar{\oldnorm}{\oldnorm*}{\nabla \varphi_m (t)}_{L^2} & \le
\@ifstar{\oldnorm}{\oldnorm*}{\nabla u_0 }_{L^2} + \@ifstar{\oldnorm}{\oldnorm*}{\nabla u_m(t) }_{L^2}
\le \@ifstar{\oldnorm}{\oldnorm*}{\nabla u_0 }_{L^2} + \sqrt{\mathcal{E}_\textnormal{logGP}
(u_m(t))} \\
&\le 2 \sqrt{\mathcal{E}_\textnormal{logGP} (u_0)}.
\end{align*}
%
We know that $\nabla \varphi_m (t) = \sum_{k=0}^m g_{m, k} (t) \nabla w_k$, thus Lemma~\ref{lem:nabla_lin_indep} shows that we can apply Lemma \ref{lem:equiv_norm_finite_dim}, which gives
\begin{equation*}
\max_{0\le k\le m} \@ifstar{\oldabs}{\oldabs*}{g_{m,k} (t)} \leq C_m \@ifstar{\oldnorm}{\oldnorm*}{\nabla
\varphi_m (t)}_{L^2} \le 2C_m \sqrt{\mathcal{E}_\textnormal{logGP} (u_0)} < \infty, \qquad \forall t \in (- T_m, T^m).
\end{equation*}
%
This proves that the $g_{m,k}$'s are actually globally bounded and therefore the solution is global.
\end{proof}
\subsection{Uniform estimates}
We have already shown the conservation of the energy and an estimate of the $L^2$-norm of $\nabla\varphi_m$, uniform both in $t$ and $m$. Now, we proceed with its $L^2$-norm.
\begin{lem} \label{lem:appr_L2_norm}
Let $u_m = u_0 + \varphi_m$ the solution to \eqref{eq:eq_u_m}
given by Lemma \ref{lem:exist_sol_app}. Then $\varphi_m$ is
bounded in $\mathcal{C}^{0, \frac{1}{2}} (I, L^2)$ uniformly in
$m$, for every bounded interval $I$.
\end{lem}
\begin{proof}
Since all $g_{m, k}$'s are continuous, we already know that
$\varphi_m $ is continuous in time with values in $L^2$.
By multiplying \eqref{eq:eq_u_m} by $g_{m, j} (t)$ and summing over $j$, we obtain
\begin{equation*}
\< \varphi_m, i \partial_t \varphi_m + \Delta u_0 + \Delta
\varphi_m - \lambda u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2} \>_{H^1, H^{-1}} = 0,
\end{equation*}
%
and the imaginary part gives
\begin{equation*}
\frac{1}{2} \frac{\diff}{\diff t} \@ifstar{\oldnorm}{\oldnorm*}{\varphi_m}_{L^2}^2 =
\Im \< \nabla \varphi_m, \nabla u_0 \> + \lambda \Im\<
\varphi_m, u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2} \>,
\end{equation*}
%
since $\Im \< \varphi_m, \Delta \varphi_m \>_{H^1, H^{-1}} =
0$. Then, we can estimate the right-hand side :
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{\< \nabla \varphi_m, \nabla u_0 \>} \leq \@ifstar{\oldnorm}{\oldnorm*}{\nabla
\varphi_m}_{L^2} \@ifstar{\oldnorm}{\oldnorm*}{\nabla u_0}_{L^2} \leq 2
\mathcal{E}_\textnormal{logGP}(u_0),
\end{equation*}
%
thanks to Lemma \ref{lem:appr_sol_global}, and
\begin{align*}
\@ifstar{\oldabs}{\oldabs*}{\< \varphi_m, u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2} \>}& \leq
\@ifstar{\oldnorm}{\oldnorm*}{\varphi_m}_{L^2}\@ifstar{\oldnorm}{\oldnorm*}{u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2}}_{L^2}
\\
&\lesssim \@ifstar{\oldnorm}{\oldnorm*}{\varphi_m}_{L^2} \(\mathcal{E}_\textnormal{logGP}
(u_0)^{1/2} + \mathcal{E}_\textnormal{logGP} (u_0)\),
\end{align*}
%
by Corollary \ref{cor:v_ln_v_Lp}. Since $\varphi_m (0) = 0$, Gronwall lemma gives the uniform boundedness of the $L^2$ norm on every bounded interval $I$. Define
\begin{equation*}
\xi_m (t) \coloneqq - \Delta \varphi_m - \Delta u_0 + \lambda u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2}.
\end{equation*}
%
Then $\xi_m$ is bounded in $L^\infty (I, H^{-1})$ uniformly in $m$ for every bounded interval $I$. Moreover, for every $t, s \in I$,
\begin{equation} \label{eq:int_diff_L2_norm}
\@ifstar{\oldnorm}{\oldnorm*}{\varphi_m (t) - \varphi_m (s)}_{L^2}^2 = 2 \int_s^t \< \varphi_m (\tau) - \varphi_m (s), \partial_t \varphi_m (\tau) \>_{H^1, H^{-1}} \diff \tau.
\end{equation}
%
By multiplying again \eqref{eq:eq_u_m} at time $\tau$ by $g_{m, j} (\tau) - g_{m, j} (s)$ and summing over $j$, there holds
\begin{equation*}
\< \varphi_m (\tau) - \varphi_m (s), \partial_t \varphi_m (\tau) \>_{H^1, H^{-1}} = \< \varphi_m (\tau) - \varphi_m (s), \xi_m (\tau) \>_{H^1, H^{-1}}.
\end{equation*}
%
Since we know that the $H^1$-norms of $\varphi_m (\tau)$ and
$\varphi_m (s)$ are bounded uniformly in $m$ for $\tau, s \in I$,
we get that $\@ifstar{\oldabs}{\oldabs*}{\< \varphi_m (\tau) - \varphi_m (s), \xi_m
(\tau) \>_{H^1, H^{-1}}}$ is bounded uniformly in $m \in
\mathbb{N}$, for $s, \tau \in I$. Then, we get by \eqref{eq:int_diff_L2_norm}
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{\varphi_m (t) - \varphi_m (s)}_{L^2}^2 \leq C_I \@ifstar{\oldabs}{\oldabs*}{t - s},
\end{equation*}
%
for all $t, s \in I$, with $C_I$ depending on $I$ but not on $m$, hence the conclusion.
\end{proof}
\subsection{Convergence}
\begin{lem} \label{lem:convergence}
Let $u_m = u_0 + \varphi_m$ the solution to \eqref{eq:eq_u_m}
given by Lemma~\ref{lem:exist_sol_app}. There exists a subsequence
of $(\varphi_m)_m$ (still denoted by $\varphi_m$) and
\begin{equation*}
\varphi \in
\mathcal{C}^{0, 1} (\mathbb{R}, H^{-1}) \cap \mathcal{C}^{0,
\frac{1}{2}} (\mathbb{R}, L^2) \cap L^\infty_\textnormal{loc}
(\mathbb{R}, H^1)
\end{equation*}
such that $\varphi_m$ converges to $\varphi$ as
$m \to \infty$ in the following sense:
\begin{itemize}
\item $\varphi_m \overset{\ast}{\rightharpoonup} \varphi$ in $L^\infty (I, H^1)$ for its weak-$*$ topology, for every bounded interval $I$,
\item $\varphi_m (t) \rightharpoonup \varphi (t)$ in $L^2$ for its weak topology, for every $t \in \mathbb{R}$.
\end{itemize}
%
Moreover, $\lambda u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2}$ converges to $
i\partial_t \varphi + \Delta \varphi + \Delta u_0 $ for the weak-$*$ topology of $L^\infty (\mathbb{R}, L^2)$.
\end{lem}
\begin{proof}
We know by Lemmas \ref{lem:appr_sol_global} and
\ref{lem:appr_L2_norm} that $\varphi_m$ is uniformly bounded in
$L^\infty (I, H^1)$, which is the dual of $L^1 (I, H^{-1})$, for
every bounded interval $I$. Then, this sequence is relatively
compact for the weak-$*$ topology of $L^\infty (I, H^1)$. By diagonal
extraction, there is a subsequence (still denoted by $\varphi_m$) which converges to some $\varphi \in L^\infty_\textnormal{loc} (\mathbb{R}, H^1)$ for the weak-$*$ topology of $L^\infty (I, H^1)$, for every bounded interval $I$.
%
Besides, $u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2}$ is also uniformly bounded in $L^\infty (\mathbb{R}, L^2) = \Bigl( L^1 (\mathbb{R}, L^2) \Bigr)'$ thanks to Corollary \ref{cor:v_ln_v_Lp} and the conservation of the energy. Therefore, it is also relatively compact for the weak-$*$ topology of $L^\infty (\mathbb{R}, L^2)$.
%
Let $\theta \in \mathcal{C}_c^\infty (\mathbb{R})$ and $j
\le m$:
\begin{align*}
\lambda \int \theta (\tau) \< w_j, u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2} \>
\diff \tau
&= \int \dot{\theta} (\tau) \< w_j, i \varphi_m (\tau) \>
\diff \tau - \int \theta (t) \< \nabla w_j, \nabla
\varphi_m (\tau) \> \diff \tau \\
&\quad - \int \theta (t) \<
\nabla w_j, \nabla u_0 \> \diff \tau \\
&\underset{m \rightarrow \infty}{\longrightarrow} \int
\dot{\theta} (\tau) \< w_j, i \varphi (\tau) \> \diff
\tau - \int \theta (t) \< \nabla w_j, \nabla \varphi
(\tau) \> \diff \tau \\
&\qquad -\int \theta (t) \< \nabla w_j,
\nabla u_0 \> \diff \tau.
\end{align*}
%
Since $(w_j)_j$ is an Hilbertian basis of $L^2$, this proves that
$\lambda u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2}$ also converges to $i\partial_t
\varphi + \Delta \varphi + \Delta u_0$ in $\mathcal{D}'
(\mathbb{R}, L^2)$. Therefore, the convergence is also in the
weak-$*$ topology of $L^\infty (\mathbb{R}, L^2)$, thus we get
$i\partial_t \varphi + \Delta \varphi + \Delta u_0\in L^\infty
(\mathbb{R}, L^2)$.
%
From this and the fact that $\Delta \varphi \in
L^\infty_\textnormal{loc} (\mathbb{R}, H^{-1})$, we get
$\partial_t \varphi \in L^\infty_\textnormal{loc} (\mathbb{R},
H^{-1})$. Therefore, $\varphi \in W^{1, \infty}_\textnormal{loc}
(\mathbb{R}, H^{-1})$ and thus $\varphi \in \mathcal C^{0,
1}_\textnormal{loc} (\mathbb{R}, H^{-1}) \cap
L^\infty_\textnormal{loc} (\mathbb{R}, H^1)$. By interpolation,
this leads to $\varphi \in \mathcal C^{0, \frac{1}{2}}_\textnormal{loc}
(\mathbb{R}, L^2)$.
%
As for the convergence in the second item, let $t \in
\mathbb{R}$. Since $\varphi_m (t)$ is bounded in $L^2$ uniformly
in $m$, the sequence is relatively compact for the weak topology
of $L^2$. Let $\chi$ be the limit of a subsequence (still denoted
by $\varphi_m$). It is then enough to prove that $\chi = \varphi
(t)$. Let $\gamma \in (0, 1)$ Then,
\begin{align*}
\@ifstar{\oldnorm}{\oldnorm*}{\varphi (t) - \chi}_{L^2}^2 &= \< \varphi (t) - \chi, \varphi_m (t) - \chi \> + \frac{1}{2 \gamma} \int_{t - \gamma}^{t + \gamma} \< \varphi (t) - \chi, \varphi (t) - \varphi_m (t) \> \diff \tau \\
&= \< \varphi (t) - \chi, \varphi_m (t) - \chi \> \\
&+ \frac{1}{2 \gamma} \int_{t - \gamma}^{t + \gamma} \< \varphi (t) - \chi, \varphi (t) - \varphi (\tau) + \varphi (\tau) - \varphi_m (\tau) + \varphi_m (\tau) - \varphi_m (t) \> \diff \tau.
\end{align*}
%
Since $\varphi, \varphi_m \in \mathcal{C}^{0, \frac{1}{2}} ([t-1,
t+1], L^2)$ uniformly in $m$, we have for all $\tau \in [t-1,
t+1]$
\begin{align*}
\@ifstar{\oldabs}{\oldabs*}{\< \varphi (t) - \chi, \varphi (t) - \varphi (\tau) \>}
&\leq C_t \@ifstar{\oldabs}{\oldabs*}{t - \tau}^\frac{1}{2} \@ifstar{\oldnorm}{\oldnorm*}{\varphi (t) -
\chi}_{L^2}, \\
\@ifstar{\oldabs}{\oldabs*}{\< \varphi (t) - \chi, \varphi_m (t) - \varphi_m (\tau)
\>} &\leq C_t \@ifstar{\oldabs}{\oldabs*}{t - \tau}^\frac{1}{2} \@ifstar{\oldnorm}{\oldnorm*}{\varphi (t) -
\chi}_{L^2}, \\
\end{align*}
%
hence
\begin{equation}\label{eq:est_conv_phi_t}
\begin{aligned}
\@ifstar{\oldnorm}{\oldnorm*}{\varphi (t) - \chi}_{L^2}^2 &\leq \< \varphi (t) - \chi,
\varphi_m (t) - \chi \>+ \frac{C_t}{2 \gamma} \@ifstar{\oldnorm}{\oldnorm*}{\varphi (t) - \chi}_{L^2} \int_{t -
\gamma}^{t+\gamma} \@ifstar{\oldabs}{\oldabs*}{t - \tau}^\frac{1}{2} \diff
\tau \\
& \quad+\frac{1}{2 \gamma} \int_{t
- \gamma}^{t + \gamma} \< \varphi (t) - \chi,
\varphi (\tau) - \varphi_m (\tau) \> \diff \tau
\\
&\leq \< \varphi (t) - \chi, \varphi_m (t) - \chi \> + C_t \gamma^\frac{1}{2} \@ifstar{\oldnorm}{\oldnorm*}{\varphi (t) -
\chi}_{L^2}\\
&\quad+ \frac{1}{2 \gamma} \int_{t - \gamma}^{t + \gamma} \<
\varphi (t) - \chi, \varphi (\tau) - \varphi_m (\tau) \>
\diff \tau .
\end{aligned}
\end{equation}
%
Moreover, we know that
\begin{itemize}
\item $\< \varphi (t) - \chi, \varphi_m (t) - \chi \> \underset{m \rightarrow \infty}{\longrightarrow} 0$ by weak convergence in $L^2$ of $\varphi_m (t)$ to $\chi$,
%
\item $\int_{t - \gamma}^{t + \gamma} \< \varphi (t) - \chi, \varphi (\tau) - \varphi_m (\tau) \> \diff \tau \underset{m \rightarrow \infty}{\longrightarrow} 0$ by the weak-$*$ convergence in $L^\infty ((t - 1, t + 1), L^2)$ of $\varphi_m$ to $\varphi$.
\end{itemize}
%
Therefore, taking the limit $m \rightarrow \infty$ in \eqref{eq:est_conv_phi_t}, we get
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{\varphi (t) - \chi}_{L^2}^2 \leq C_t \gamma^\frac{1}{2} \@ifstar{\oldnorm}{\oldnorm*}{\varphi (t) - \chi}_{L^2}.
\end{equation*}
%
Since this is true for all $\gamma \in (0, 1)$, we get
$\@ifstar{\oldnorm}{\oldnorm*}{\varphi (t) - \chi}_{L^2}=0$, hence the conclusion.
\end{proof}
\subsection{Equation and initial data for the limit}
We are now able to prove that $u = u_0 + \varphi$ is indeed a solution to \eqref{logGP}.
\begin{lem}
Let $\varphi \in \mathcal{C}^{0, 1} (\mathbb{R}, H^{-1}) \cap
\mathcal{C}^{0, \frac{1}{2}} (\mathbb{R}, L^2) \cap
L^\infty_\textnormal{loc} (\mathbb{R}, H^1)$ defined in
Lemma~\ref{lem:convergence}. Then $u \coloneqq u_0 + \varphi$
satisfies \eqref{logGP}.
\end{lem}
\begin{proof}
First, since $\varphi_m (0) \rightharpoonup \varphi (0)$ and
$\varphi_m (0) = 0$ by construction, we get $\varphi (0) = 0$ and
thus $u_{\mid t=0} = u_0$.
%
We also know from Lemma~\ref{lem:convergence} that $\lambda u_m
\ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2}$ converges to $i \partial_t \varphi +
\Delta \varphi + \Delta u_0 $ for the weak-$*$
topology of $L^\infty (\mathbb{R}, L^2)$, and thus also for the
weak topology of $L^2 (I \times \Omega)$ for every bounded
interval $I$ and every bounded open subset of $\mathbb{R}^d$.
%
We also know that $\varphi_m$ is $\frac{1}{2}$-H\"older continuous
in time with values in $L^2 (\mathbb{R}^d)$ on $I$, uniformly in
$m$ from Lemma \ref{lem:appr_L2_norm}, and thus also on $L^2
(\Omega)$. Last, $\varphi_m (t) \in H^1$ for all $t \in I$ along
with a uniform bound in $m$ by Lemmas \ref{lem:appr_sol_global}
and \ref{lem:appr_L2_norm}. Thus, Arzela--Ascoli Theorem for
${\varphi_m}_{|\Omega}$ yields its relative compactness in
$\mathcal{C} (I, L^2 (\Omega))$, which shows that
${\varphi_m}_{|\Omega}$ converges strongly to
${\varphi}_{|\Omega}$ in $\mathcal{C} (I, L^2 (\Omega))$, and thus
also in $L^2 (I \times \Omega)$. Up to a further subsequence,
$\varphi_m$ converges a.e. to $\varphi$ in $I \times \Omega$, and
so does $u_m$ to $u$. By continuity of $z\mapsto z \ln{\@ifstar{\oldabs}{\oldabs*}{z}^2}$ on $\mathbb{C}$, we obtain the convergence of $u_m (t, x) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t,x)}^2}$ to $u (t, x) \ln{\@ifstar{\oldabs}{\oldabs*}{u (t,x)}^2}$ a.e. in $I \times \Omega$.
%
Besides, we also know that $u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2}$ is bounded in
$L^\infty (I, L^2)$ uniformly in $m$, and thus also in $L^2 (I
\times \Omega)$. Therefore, along this further subsequence, $u_m
\ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2}$ converges weakly to $u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2}$ in $L^2 (I \times \Omega)$.
%
Therefore, by uniqueness of the limit, we get $\lambda u
\ln{\@ifstar{\oldabs}{\oldabs*}{u}^2} = i \partial_t \varphi +\Delta \varphi + \Delta u_0 $, which gives the conclusion.
\end{proof}
\subsection{Conservation law}
We now prove that the energy of the solution that we have just
constructed is independent of time. Note that this is necessarily \emph{the}
solution to \eqref{logGP}, in view of Theorem~\ref{th:regu_sol}.
\begin{lem}
Let $\varphi \in \mathcal{C}^{0, 1} (\mathbb{R}, H^{-1}) \cap \mathcal{C}^{0, \frac{1}{2}} (\mathbb{R}, L^2) \cap L^\infty_\textnormal{loc} (\mathbb{R}, H^1)$ defined in Lemma~\ref{lem:convergence} and $u \coloneqq u_0 + \varphi$. Then $u$ satisfies $\mathcal{E}_\textnormal{log} (u (t)) = \mathcal{E}_\textnormal{log} (u_0)$ for all $t \in \mathbb{R}$.
\end{lem}
\begin{proof}
This proof is an adaptation of the techniques developed by Ozawa
\cite{Ozawa__Conservation_laws} to prove conservation laws from
the mild formulation, and resumes the regularization procedure
presented in e.g. \cite{Cazenave_semlin_lognls}.
We know that $\varphi \in L^\infty_\textnormal{loc} (\mathbb{R}, H^1)$, thus Lemma \ref{lem:Energy_plus_H1} shows that $u \in L^\infty_\textnormal{loc} (\mathbb{R}, E_\textnormal{logGP})$. From Lemma \ref{lem:mild_sol}, $u$ is a mild solution. Moreover, we also know that $\nabla u (t) \in L^2$ for \textit{all} $t \in \mathbb{R}$ (and not only a.e.). Therefore, for all $\varepsilon > 0$, $(1 - \varepsilon \Delta)^{-1} \nabla u (t)$ is well defined in $L^2$ and
\begin{equation*}
(1 - \varepsilon \Delta)^{-1} \nabla u (t) \underset{\varepsilon \rightarrow 0}{\longrightarrow} \nabla u (t) \qquad \text{in } L^2.
\end{equation*}
%
On the other hand, we also have
\begin{align*}
\@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla u (t)}_{L^2}^2 &= \@ifstar{\oldnorm}{\oldnorm*}{e^{- i t \Delta} (1 - \varepsilon \Delta)^{-1} \nabla u (t)}_{L^2}^2 \\
&= \@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} e^{- i t \Delta} \nabla u (t)}_{L^2}^2 \\
&= \@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} e^{- i t \Delta} \nabla u_0 + (1 - \varepsilon \Delta)^{-1} e^{- i t \Delta} \nabla (u (t) - u_0)}_{L^2}^2.
\end{align*}
%
The last line is well defined since $\nabla u_0 \in L^2$. Moreover, since $u (t) - u_0 \in H^1$, we have $e^{- i t \Delta} \nabla (u (t) - u_0) = \nabla e^{- i t \Delta} (u (t) - u_0)$, i.e.
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla u (t)}_{L^2}^2 = \@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} e^{- i t \Delta} \nabla u_0 + (1 - \varepsilon \Delta)^{-1} \nabla e^{- i t \Delta} (u (t) - u_0)}_{L^2}^2
\end{equation*}
Since $(1 - \varepsilon \Delta)^{-1} \nabla$ is bounded on $L^2$
(e.g. as a Fourier multiplier), we can substitute $u (t)$ by the mild formulation :
\begin{align*}
(1 - \varepsilon \Delta)^{-1} &\nabla e^{- i t \Delta} (u (t)
- u_0) \\
&= (1 - \varepsilon \Delta)^{-1} \nabla \biggl( u_0 - i \lambda \int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s - e^{- i t \Delta} u_0 \biggr).
\end{align*}
%
Moreover, since the Schrödinger semigroup $e^{i t \Delta}$ maps $u_0 + H^1$ into itself, we have
\begin{equation*}
e^{- i t \Delta} \nabla u_0 = \nabla e^{- i t \Delta} u_0,
\end{equation*}
which leads to
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla u (t)}_{L^2}^2 = \@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla \biggl( u_0 - i \lambda \int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s \biggr)}_{L^2}^2.
\end{equation*}
%
Now, we know thanks to Lemma \ref{lem:mild_sol} that $\int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s \in H^1$. Therefore,
\begin{align*}
&\@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla u (t)}_{L^2}^2\\
&\quad= \@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla u_0 - i \lambda (1 - \varepsilon \Delta)^{-1} \nabla \int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s}_{L^2}^2 \\
&\quad= \@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla u_0}_{L^2}^2 \\
&\qquad- 2 \lambda \Re \< (1 - \varepsilon \Delta)^{-1} \nabla u_0, i (1 - \varepsilon \Delta)^{-1} \nabla \int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s \> \\
&\qquad+ \lambda^2 \@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla \int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s}_{L^2}^2.
\end{align*}
%
Then, we compute the last two terms of the right-hand side: for the last term,
\begin{multline*}
\@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla \int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s}_{L^2}^2 \\
\begin{aligned}
&= \@ifstar{\oldnorm}{\oldnorm*}{\int_0^t e^{- i s \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s}_{L^2}^2 \\
&= \Re \< \int_0^t e^{- i s \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s, \int_0^t e^{- i s' \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s') \ln{\@ifstar{\oldabs}{\oldabs*}{u (s')}^2} \Bigr) \diff s' \> \\
&= \Re \int_0^t \int_0^t \< e^{- i s \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), e^{- i s' \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s') \ln{\@ifstar{\oldabs}{\oldabs*}{u (s')}^2} \Bigr) \> \diff s \diff s' \\
&= 2 \Re \iint_{0 \leq s' \leq s \leq t} \< e^{- i s \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), e^{- i s' \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s') \ln{\@ifstar{\oldabs}{\oldabs*}{u (s')}^2} \Bigr) \> \diff s \diff s',
\end{aligned}
\end{multline*}
%
by symmetry $s \leftrightarrow s'$. Therefore,
\begin{multline*}
\@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla \int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s}_{L^2}^2 \\
\begin{aligned}
&= 2 \Re \iint_{0 \leq s' \leq s \leq t} \< (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), e^{i (s - s') \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s') \ln{\@ifstar{\oldabs}{\oldabs*}{u (s')}^2} \Bigr) \> \diff s \diff s' \\
&= 2 \Re \int_0^t \< (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), \int_0^{s} e^{i (s - s') \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s') \ln{\@ifstar{\oldabs}{\oldabs*}{u (s')}^2} \Bigr) \diff s' \> \diff s \\
&= \frac{2}{\lambda} \Im \int_0^t \< (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), - i \lambda (1 - \varepsilon \Delta)^{-1} \nabla \int_0^{s} e^{i (s - s') \Delta} \Bigl( u(s') \ln{\@ifstar{\oldabs}{\oldabs*}{u (s')}^2} \Bigr) \diff s' \> \diff s \\
&= \frac{2}{\lambda} \Im \int_0^t \< (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), (1 - \varepsilon \Delta)^{-1} \nabla \Bigl[ u(s) - e^{i s \Delta} u_0 \Bigr] \> \diff s.
\end{aligned}
\end{multline*}
%
For the other term,
\begin{multline*}
\Re \< (1 - \varepsilon \Delta)^{-1} \nabla u_0, i (1 - \varepsilon \Delta)^{-1} \nabla \int_0^t e^{- i s \Delta} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \diff s \> \\
\begin{aligned}
&= - \Im \int_0^t \< (1 - \varepsilon \Delta)^{-1} \nabla u_0, e^{- i s \Delta} (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \> \diff s \\
&= - \Im \int_0^t \< e^{i s \Delta} (1 - \varepsilon \Delta)^{-1} \nabla u_0, (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr) \> \diff s \\
&= \int_0^t \Im \< (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), (1 - \varepsilon \Delta)^{-1} \nabla e^{i s \Delta} u_0 \> \diff s.
\end{aligned}
\end{multline*}
%
Therefore, there holds
\begin{multline*}
\@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla u (t)}_{L^2}^2 - \@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla u_0}_{L^2}^2 \\
\begin{aligned}
&= - 2 \lambda \Im \int_0^t \< (1 - \varepsilon \Delta)^{-1} \nabla \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), (1 - \varepsilon \Delta)^{-1} \nabla u (s) \> \diff s \\
&= 2 \lambda \Im \int_0^t \< (1 - \varepsilon \Delta)^{-1} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), (1 - \varepsilon \Delta)^{-1} \Delta u (s) \> \diff s \\
&= 2 \lambda \Im \int_0^t \< (1 - \varepsilon \Delta)^{-1} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), (1 - \varepsilon \Delta)^{-1} \Bigl[ \Delta u (s) - \lambda u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr] \> \diff s \\
&= - 2 \lambda \Im \int_0^t \< (1 - \varepsilon \Delta)^{-1} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), (1 - \varepsilon \Delta)^{-1} i \partial_t u (s) \> \diff s \\
&= - 2 \lambda \Re \int_0^t \< (1 - \varepsilon \Delta)^{-1} \Bigl( u(s) \ln{\@ifstar{\oldabs}{\oldabs*}{u (s)}^2} \Bigr), (1 - \varepsilon \Delta)^{-1} \partial_t u (s) \> \diff s.
\end{aligned}
\end{multline*}
%
Thanks to Lemma \ref{lem:deriv_pot_en_regu}, we get
\begin{align*}
\@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1} \nabla u (t)}_{L^2}^2 &-
\@ifstar{\oldnorm}{\oldnorm*}{(1 - \varepsilon \Delta)^{-1}\nabla u_0}_{L^2}^2
\\
&= - \lambda \int_0^t \frac{\diff}{\diff s} \Bigl[ \int (1 - \varepsilon \Delta)^{-2} F(\@ifstar{\oldabs}{\oldabs*}{u (s)}^2) \diff x \Bigr] \diff s \\
&= \lambda \int (1 - \varepsilon \Delta)^{-2}
F(\@ifstar{\oldabs}{\oldabs*}{u_0}^2) \diff x
- \lambda \int (1 - \varepsilon \Delta)^{-2} F(\@ifstar{\oldabs}{\oldabs*}{u (s)}^2) \diff x,
\end{align*}
%
which gives $\mathcal{E}_\textnormal{logGP} (u(t)) = \mathcal{E}_\textnormal{logGP} (u_0)$ after taking the limit $\varepsilon \rightarrow 0$.
\end{proof}
\section{Higher regularity}
\label{sec:higher}
In this section, we prove the propagation of the $\dot{H}^2$
regularity by the flow of \eqref{logGP}. We would like to proceed like
in \cite{carlesgallagher} (see also \cite{cazenave-haraux,
Cazenave_semlin_lognls}) directly on the solution $u$ constructed,
i.e. differentiating \eqref{logGP} with respect to time and get an
$L^2$-energy estimate for ${\partial}_t \varphi$. A direct formal computation
would lead to the local boundedness of this $L^2$-norm with the
additional argument that $\partial_t u (0) \in L^2$ as soon as $\Delta
u_0 \in L^2$. However, this computation cannot directly be made
rigorous since $u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2}$ cannot be differentiated due to the
lack of regularity of the logarithm.
To overcome this difficulty, we shall work on the approximate
solutions $u_m$. In \cite{carlesgallagher}, the authors worked on the
approximate solutions to prove the propagation of the $H^2$
regularity, since the above mentioned energy estimate is licit in the
case of \eqref{eq:logNLSreg} (and provides bounds which are uniform in
$\varepsilon$). However, we cannot reproduce the same proof here: our
approximate solutions do not satisfy an equation with a regularized
nonlinearity. Indeed, the equations \eqref{eq:eq_u_m} can be put under
the form
\begin{equation*}
i\partial_t \varphi_m + \mathbb{P}_m \Bigl( \Delta u_m - \lambda u_m \ln{\@ifstar{\oldabs}{\oldabs*}{u_m}^2} \Bigr) = 0,
\end{equation*}
where $\mathbb{P}_m$ is the orthogonal projector from $L^2$ onto $X_m$ (we can assume $w_j \in H^2$ to make this rigorous).
Thus, the nonlinearity is still not smooth enough to make the
computation rigorous. However, by assuming $w_j \in H^2$ (for all
$j$), we automatically have $\varphi_m (t) \in H^2$ for all $t \in
\mathbb{R}$ and $m \in \mathbb{N}$. Even more, we have $\varphi_m \in
\mathcal{C}^1 (\mathbb{R}, H^2)$ since the $g_{m, k}$'s belong to
$\mathcal{C}^1 (\mathbb{R})$ (see Lemma
\ref{lem:exist_sol_app}). Those crude bounds are \textit{a priori} not
uniform in $m$, but we show that we can improve them.
\begin{lem} \label{lem:dt_um_unif_bound}
Let $u_m$ given by Lemma~\ref{lem:exist_sol_app} and assume $w_j \in H^2$ for all $j \in \mathbb{N}$. Then, for all bounded interval $I$, $\partial_t u_m$ is bounded in $\mathcal{C}^0 (I, L^2)$ uniformly in $m$.
\end{lem}
\begin{proof}
From Lemma \ref{lem:exist_sol_app}, we know that the $g_{m, j}$'s
are $\mathcal{C}^{1, \varepsilon}$ for all $\varepsilon \in (0,
1)$, therefore $\partial_t u_m \in\mathcal C^{0, \varepsilon} (\mathbb{R},
H^2)$. Let $\tau > 0$ and
\begin{equation*}
\psi_{m, \tau} (t) \coloneqq \frac{\@ifstar{\oldnorm}{\oldnorm*}{\partial_t \varphi_m (t + \tau)}_{L^2}^2 - \@ifstar{\oldnorm}{\oldnorm*}{\partial_t \varphi_m (t)}_{L^2}^2}{\tau},
\end{equation*}
%
which is well defined for all $t \in \mathbb{R}$ and any $\tau >
0$. Our goal is to prove a bound on $\psi_{m, \tau} (t)$
independent of $\tau \leq 1$. For this, we first rewrite
\begin{equation*}
\psi_{m, \tau} (t) = \frac{1}{\tau} \Re \langle \partial_t \varphi_m (t + \tau) + \partial_t \varphi_m (t), {\partial}_t\varphi_m (t + \tau) - \partial_t \varphi_m (t) \rangle.
\end{equation*}
%
Since $\partial_t \varphi_m (t+\tau)$ and $\partial_t \varphi_m (t)$ are in $\operatorname{Vect} (w_j)_{j \leq m}$, we can use \eqref{eq:eq_u_m} and get
\begin{multline*}
\tau \, \psi_{m, \tau} (t) = - \Im \langle \partial_t \varphi_m (t + \tau) + \partial_t \varphi_m (t), \Delta \varphi_m (t + \tau) - \Delta \varphi_m (t) \rangle \\ + \lambda \Im \langle \partial_t \varphi_m (t + \tau) + \partial_t \varphi_m (t), u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t)}^2} \rangle.
\end{multline*}
%
For the first term, we get
\begin{align*}
\langle \partial_t \varphi_m (t), \Delta \varphi_m (t + \tau) - \Delta \varphi_m (t) \rangle &= \langle \partial_t \varphi_m (t), \Delta \varphi_m (t + \tau) - \Delta \varphi_m (t) - \tau \partial_t \Delta \varphi_m (t) \rangle \\&\quad+ \tau \langle \partial_t \varphi_m (t), \partial_t \Delta \varphi_m (t) \rangle.
\end{align*}
%
Since $\langle \partial_t \varphi_m (t), \partial_t \Delta \varphi_m (t) \rangle = - \@ifstar{\oldnorm}{\oldnorm*}{\partial_t \nabla \varphi_m}_{L^2}^2$, the last term vanishes when we take the imaginary part. On the other hand, we have shown that $\Delta \varphi_m \in \mathcal{C}^{1, \varepsilon} (\mathbb{R}, L^2)$. Hence, for all $\tau \leq 1$,
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{ \Delta \varphi_m (t + \tau) - \Delta \varphi_m (t) - \tau \partial_t \Delta \varphi_m (t) }_{L^2} \leq C_{m, \varepsilon, t} \tau^{1 + \varepsilon}.
\end{equation*}
%
Therefore, for all $t \in \mathbb{R}$, $\tau, \varepsilon \in (0, 1)$,
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{\Im \langle \partial_t \varphi_m (t), \Delta \varphi_m (t + \tau) - \Delta \varphi_m (t) \rangle} \leq C_{m, \varepsilon, t} \tau^{1 + \varepsilon},
\end{equation*}
%
and similarly for $\Im \langle \partial_t \varphi_m (t + \tau), \Delta \varphi_m (t + \tau) - \Delta \varphi_m (t) \rangle$.
%
Now, we also have
\begin{multline*}
\langle \partial_t \varphi_m (t), u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t)}^2} \rangle \\
\begin{aligned}
&= \frac{1}{\tau} \langle \varphi_m (t + \tau) - \varphi_m (t), u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} \rangle \\ &- \frac{1}{\tau} \langle \varphi_m (t + \tau) - \varphi_m (t) - \tau \partial_t \varphi_m (t), u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} \rangle.
\end{aligned}
\end{multline*}
%
For the first term, we can use Lemma \ref{lem:log_inequality} since $\varphi_m (t + \tau) - \varphi_m (t) = u_m (t+\tau) - u_m (t)$, so that
\begin{multline*}
\@ifstar{\oldabs}{\oldabs*}{\Im \langle \varphi_m (t + \tau) - \varphi_m (t), u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} \rangle} \\ \leq 2 \@ifstar{\oldnorm}{\oldnorm*}{u_m (t+\tau) - u_m (t)}_{L^2}^2 = 2 \@ifstar{\oldnorm}{\oldnorm*}{\varphi_m (t+\tau) - \varphi_m (t)}_{L^2}^2.
\end{multline*}
%
For the second term, we can use Lemma \ref{lem:prop_lip_log} so that
\begin{multline*}
\@ifstar{\oldabs}{\oldabs*}{\langle \varphi_m (t + \tau) - \varphi_m (t) - \tau \partial_t \varphi_m (t), u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} \rangle} \\
\leq \@ifstar{\oldnorm}{\oldnorm*}{\varphi_m (t+\tau) - \varphi_m (t) - \tau \partial_t \varphi_m (t)}_{L^2} \@ifstar{\oldnorm}{\oldnorm*}{u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2}}_{L^2}
\end{multline*}
%
The first factor can be estimated like previously with the fact that $\varphi_m \in \mathcal{C}^{1, \varepsilon} (\mathbb{R}; L^2)$:
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{\varphi_m (t+\tau) - \varphi_m (t) - \tau \partial_t \partial_t \varphi_m (t)}_{L^2} \leq C_{m, t, \varepsilon} \tau^{1 + \varepsilon}.
\end{equation*}
%
As for the other factor, we use Lemma \ref{lem:prop_lip_log} and then Corollary \ref{cor:v_ln_v_Lp} to obtain
\begin{align*}
& \@ifstar{\oldnorm}{\oldnorm*}{u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2}}_{L^2} \\
\lesssim &
\@ifstar{\oldnorm}{\oldnorm*}{\Bigl( \@ifstar{\oldabs}{\oldabs*}{u_m (t +
\tau)}^\frac{\varepsilon}{2} \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{u_m
(t + \tau)}}} + \@ifstar{\oldabs}{\oldabs*}{u_m
(t)}^\frac{\varepsilon}{2} \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{u_m
(t)}}} \Bigr) \@ifstar{\oldabs}{\oldabs*}{u_m (t + \tau) - u_m (t)}^{1
- \frac{\varepsilon}{2}}}_{L^2} \\
&+ \@ifstar{\oldnorm}{\oldnorm*}{u_m (t + \tau) - u_m (t)}_{L^2}\\
\lesssim & \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{u_m (t +
\tau)}^\frac{\varepsilon}{2} \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{u_m
(t + \tau)}}} + \@ifstar{\oldabs}{\oldabs*}{u_m
(t)}^\frac{\varepsilon}{2} \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{u_m
(t)}}}}_{L^\frac{4}{\varepsilon}}
\@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{u_m (t + \tau) - u_m (t)}^{1 -
\varepsilon}}_{L^\frac{2}{1 -
\frac{\varepsilon}{2}}} \\
&+ \@ifstar{\oldnorm}{\oldnorm*}{u_m (t + \tau) - u_m (t)}_{L^2}\\
\lesssim &
\Bigl( \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{u_m (t + \tau)}
\@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t +
\tau)}}}^\frac{2}{\varepsilon}}_{L^2}^\varepsilon
+ \@ifstar{\oldnorm}{\oldnorm*}{\@ifstar{\oldabs}{\oldabs*}{u_m (t)} \@ifstar{\oldabs}{\oldabs*}{\ln{\@ifstar{\oldabs}{\oldabs*}{u_m
(t)}}}^\frac{2}{\varepsilon}}_{L^2}^\varepsilon \Bigr)
\@ifstar{\oldnorm}{\oldnorm*}{u_m (t + \tau) - u_m (t)}_{L^2}^{1 -
\frac{\varepsilon}{2}} \\
&+ \@ifstar{\oldnorm}{\oldnorm*}{u_m (t + \tau) - u_m (t)}_{L^2}\\
\lesssim &\Bigl( \mathcal{E}_\textnormal{logGP} (u_0)^{\varepsilon/2} + \mathcal{E}_\textnormal{logGP} (u_0)^{(\varepsilon + \varepsilon^2)/2} \Bigr) \@ifstar{\oldnorm}{\oldnorm*}{u_m (t + \tau) - u_m (t)}_{L^2}^{1 - \frac{\varepsilon}{2}} + \@ifstar{\oldnorm}{\oldnorm*}{u_m (t + \tau) - u_m (t)}_{L^2}.
\end{align*}
%
Therefore, we obtain
\begin{multline*}
\@ifstar{\oldabs}{\oldabs*}{\langle \varphi_m (t + \tau) - \varphi_m (t) - \tau \partial_t \varphi_m (t), u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} \rangle} \\
\leq C_{m, t, \varepsilon} \Bigl[ \tau^{2 +
\frac{\varepsilon}{2}} \Bigl( \mathcal{E}_\textnormal{logGP}
(u_0)^{\varepsilon/2} +
\mathcal{E}_\textnormal{logGP} (u_0)^{\varepsilon} \Bigr)
\@ifstar{\oldnorm}{\oldnorm*}{\delta_\tau \varphi_m}_{L^2}^{1 - \frac{\varepsilon}{2}}
+ \tau^{2+\varepsilon} \@ifstar{\oldnorm}{\oldnorm*}{\delta_\tau \varphi_m}_{L^2}
\Bigr],
\end{multline*}
%
where $\delta_\tau \varphi_m \coloneqq \frac{1}{\tau} (\varphi_m (t+\tau) - \varphi_m(t))$. Thus, we get
\begin{align*}
& \@ifstar{\oldabs}{\oldabs*}{\Im \langle \partial_t \varphi_m (t), u_m (t+\tau)
\ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t)}^2}
\rangle} \\
& \phantom{\Im \langle \partial_t \varphi_m (t)} \le (2 \tau + C_{m, t, \varepsilon} \tau^{1 +
\varepsilon}) \@ifstar{\oldnorm}{\oldnorm*}{\delta_\tau \varphi_m}_{L^2}^2 \\
& \phantom{\Im \langle \partial_t \varphi_m (t)} \quad+
C_{m, t, \varepsilon} \tau^{1 + \frac{\varepsilon}{2}} \(
\mathcal{E}_\textnormal{logGP} (u_0)^{\varepsilon/2} +
\mathcal{E}_\textnormal{logGP} (u_0)^{\varepsilon} \)&
\@ifstar{\oldnorm}{\oldnorm*}{\delta_\tau \varphi_m}_{L^2}^{1 - \frac{\varepsilon}{2}} .
\end{align*}
%
Similar computations can be done for $\Im \langle \partial_t
\varphi_m (t + \tau), u_m (t+\tau) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t+\tau)}^2} - u_m
(t) \ln{\@ifstar{\oldabs}{\oldabs*}{u_m (t)}^2} \rangle$, so we get
\begin{equation} \label{eq:est_psi_m_tau}
\begin{aligned}
\@ifstar{\oldabs}{\oldabs*}{\psi_{m, \tau} (t)} &\leq C_{m, t, \varepsilon}
\tau^\varepsilon + (4 \lambda + C_{m, t, \varepsilon}
\tau^{\varepsilon}) \@ifstar{\oldnorm}{\oldnorm*}{\delta_\tau
\varphi_m}_{L^2}^2 \\
&\quad + C_{m, t, \varepsilon} \tau^{1 +
\frac{\varepsilon}{2}} \Bigl(
\mathcal{E}_\textnormal{logGP} (u_0)^{\varepsilon/2} +
\mathcal{E}_\textnormal{logGP} (u_0)^{\varepsilon}
\Bigr) \@ifstar{\oldnorm}{\oldnorm*}{\delta_\tau \varphi_m}_{L^2}^{1 -
\frac{\varepsilon}{2}}.
\end{aligned}
\end{equation}
%
At $m$ fixed, since $\varphi_m \in \mathcal{C}^{1, \varepsilon}(\R;L^2)$,
we know that as $\tau\to 0$, $\delta_\tau \varphi_m (t)$ converges strongly to $\partial_t \varphi_m (t)$ in $L^2$ for any $t \in \mathbb{R}$, but also that, for $t \in \mathbb{R}$ fixed, $\@ifstar{\oldnorm}{\oldnorm*}{\delta_\tau \varphi_m}_{L^2}$ is uniformly bounded in $\tau \leq 1$. The previous estimate shows that $\psi_{m, \tau} (t)$ is uniformly bounded in $\tau \leq 1$ at $m \in \mathbb{N}$ and $t \in \mathbb{R}$ fixed, which means that
\begin{equation*}
t\mapsto \@ifstar{\oldnorm}{\oldnorm*}{\partial_t \varphi_m (t)}_{L^2}^2 \in W^{1, \infty}_\textnormal{loc} (\mathbb{R}, \mathbb{R}).
\end{equation*}
%
The limit $\tau \rightarrow 0$ in \eqref{eq:est_psi_m_tau} then yields
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{\frac{\diff}{\diff t} \@ifstar{\oldnorm}{\oldnorm*}{\partial_t \varphi_m (t)}_{L^2}^2} \leq 4 \lambda \@ifstar{\oldnorm}{\oldnorm*}{\partial_t \varphi_m (t)}_{L^2}^2.
\end{equation*}
%
On the other hand, we also have $i\partial_t \varphi_m (0) = \mathbb{P}_m \Bigl[ \lambda u_0 \ln{\@ifstar{\oldabs}{\oldabs*}{u_0}^2} - \Delta u_0 \Bigr]$, and since $\mathbb{P}_m$ is an orthogonal projector, we get
\begin{align*}
\@ifstar{\oldnorm}{\oldnorm*}{\partial_t \varphi_m (0)}_{L^2} &\le \lambda \@ifstar{\oldnorm}{\oldnorm*}{u_0
\ln{\@ifstar{\oldabs}{\oldabs*}{u_0}^2}}_{L^2}
+ \@ifstar{\oldnorm}{\oldnorm*}{\Delta
u_0}_{L^2} \\
& \le C_\varepsilon \lambda (\mathcal{E}_\textnormal{logGP} (u_0)^{\varepsilon/2} + \mathcal{E}_\textnormal{logGP} (u_0)^{\varepsilon}) + \@ifstar{\oldnorm}{\oldnorm*}{\Delta u_0}_{L^2}.
\end{align*}
%
Thanks to Gronwall Lemma, we get a bound which is independent of $m$: for all $m \in \mathbb{N}$ and $t \in \mathbb{R}$,
\begin{equation*}
\@ifstar{\oldnorm}{\oldnorm*}{\partial_t \varphi_m (t)}_{L^2}^2 \leq \Bigl[
C_\varepsilon \lambda (\mathcal{E}_\textnormal{logGP}
(u_0)^{\varepsilon/2} +
\mathcal{E}_\textnormal{logGP} (u_0)^{\varepsilon})
+ \@ifstar{\oldnorm}{\oldnorm*}{\Delta u_0}_{L^2} \Bigr]^2 e^{4
\lambda t}.\qedhere
\end{equation*}
\end{proof}
\begin{cor}
Let $u$ given by Lemma \ref{lem:convergence}. Then, for all bounded interval $I$, $\partial_t u$ is bounded in $L^\infty (I; L^2)$, and so is $\Delta u$.
\end{cor}
\begin{proof}
We want to take the limit $m \rightarrow \infty$ in Lemma
\ref{lem:dt_um_unif_bound}. For this, we use the fact that
$\varphi_m \overset{\ast}{\rightharpoonup} \varphi$ in $L^\infty (I; H^1)$ for its
weak-$*$ topology for any bounded interval $I$ by Lemma
\ref{lem:convergence}. Therefore, $\partial_t \varphi_m
\rightharpoonup \partial_t \varphi$ in $\mathcal{D}' (I \times
\mathbb{R}^d)$. As the $L^\infty (I; L^2)$ norm of $\partial_t
\varphi_m$ is uniformly bounded in $m$ and since $\mathcal{C}_c (I
\times \mathbb{R}^d)$ is dense in $L^\infty (I; L^2) = \Bigl( L^1
(I;L^2) \Bigr)'$, this limit shows that $\partial_t \varphi =
\partial_t u$ belongs to $L^\infty (I; L^2)$.
%
To conclude, we show that $u \ln{\@ifstar{\oldabs}{\oldabs*}{u}^2} \in L^\infty (I;
L^2)$: in view of Corollary~\ref{cor:v_ln_v_Lp}, for $t\in I$,
\begin{align*}
\@ifstar{\oldnorm}{\oldnorm*}{u (t) \ln{\@ifstar{\oldabs}{\oldabs*}{u (t)}^2} }_{L^2} \lesssim
\mathcal{E}_\textnormal{logGP} (u_0)^{1/2} +
\mathcal{E}_\textnormal{logGP} (u_0) .
\end{align*}
%
Hence, using \eqref{logGP}, $\Delta u
\in L^\infty (I; L^2)$.
\end{proof}
\section{On stationary and traveling waves}
\label{sec:dynamics}
In this section, we prove some rather general results regarding
solitary and traveling waves.
First, plugging \eqref{eq:trav_wave} into \eqref{logGP}, we get the equation for $\phi$:
\begin{equation} \label{eq:trav_phi}
- \omega \phi - i c \cdot \nabla \phi + \Delta \phi = \lambda \phi \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2}.
\end{equation}
\subsection{The only possible value for $\omega$}
\begin{proof}[Proof of Theorem \ref{th:value_omega}]
%
$u$ satisfies \eqref{logGP} with initial data $\phi$ and $u \in
L^\infty (\mathbb{R}; E_\textnormal{logGP})$. From Lemma
\ref{lem:u_in_plusH1}, we thus know that $u - \phi \in
\mathcal{C}^0 (\mathbb{R}; L^2)$.
%
On the other hand, we know that, for any $t \in \mathbb{R}$, $x \in \mathbb{R}^d$ and any $f$ smooth enough,
\begin{equation*}
f (x - tc) - f(x) = - t \int_0^1 \nabla f (x - \tau c) \cdot c \diff \tau,
\end{equation*}
%
which leads to
\begin{align*}
\@ifstar{\oldnorm}{\oldnorm*}{f (x - tc) - f(x)}_{L^2} &\leq \@ifstar{\oldabs}{\oldabs*}{t} \int_0^1 \@ifstar{\oldnorm}{\oldnorm*}{\nabla f (x - \tau c) \cdot c}_{L^2} \diff \tau\leq \@ifstar{\oldabs}{\oldabs*}{t} \int_0^1 \@ifstar{\oldnorm}{\oldnorm*}{\nabla f (x - \tau c)}_{L^2} \@ifstar{\oldabs}{\oldabs*}{c} \diff \tau \\
&\leq \@ifstar{\oldabs}{\oldabs*}{t} \@ifstar{\oldabs}{\oldabs*}{c} \int_0^1 \@ifstar{\oldnorm}{\oldnorm*}{\nabla f}_{L^2} \diff
\tau \leq \@ifstar{\oldabs}{\oldabs*}{t} \@ifstar{\oldabs}{\oldabs*}{c} \@ifstar{\oldnorm}{\oldnorm*}{\nabla f}_{L^2}.
\end{align*}
%
Since $\nabla \phi \in L^2$, we can also apply this inequality to $\phi$, which proves that for all $t \in \mathbb{R}$, $\phi (. - ct) - \phi \in L^2$.
%
The above two claims imply
that $u - \phi(\cdot -ct) \in L^2$ for all $t \in
\mathbb{R}$. In view of \eqref{eq:trav_wave}, this means that $(e^{i \omega t} - 1) \phi \in L^2$
for all $t \in \mathbb{R}$. Suppose that $\omega \neq 0$:
considering $t = \frac{\pi}{\omega}$, we get $- 2 \phi \in L^2$,
which is in contradiction with the fact that $\@ifstar{\oldabs}{\oldabs*}{\phi} - 1 \in
L^2$ from Lemma~\ref{lem:E_logGP_descr}. Therefore, $\omega=0$.
\end{proof}
\subsection{Regularity and limits of a traveling wave in dimension $d=1$}
In this section, we prove a result about the regularity and limits at
infinity of traveling waves of the form \eqref{eq:trav_wave} solution
to \eqref{logGP} in dimension $d=1$. In view of
Theorem~\ref{th:value_omega}, such traveling waves satisfy
\begin{equation} \label{eq:ode_phi_trav}
-ic \phi' + \phi'' = \lambda \phi \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2}.
\end{equation}
\begin{lem} \label{lem:solit_reg_lim}
A traveling wave $\phi$ solution to \eqref{eq:ode_phi_trav} satisfies
\begin{itemize}
\item $\phi \in \mathcal{C}^2_b$ and $\lim_{\@ifstar{\oldabs}{\oldabs*}{x} \rightarrow \infty} \@ifstar{\oldabs}{\oldabs*}{\phi (x)} = 1$,
%
\item $\phi' \in H^1$ and $\lim_{\@ifstar{\oldabs}{\oldabs*}{x} \rightarrow \infty} \phi' (x) = 0$.
\end{itemize}
\end{lem}
\begin{proof}
Since $\phi \in E_\textnormal{logGP}$, we have $\phi \in H^1_\textnormal{loc}$ and in particular $\phi \in \mathcal{C}^{0, \frac{1}{2}} (\mathbb{R})$ by Sobolev embedding in dimension $d=1$.
Actually, $\phi$ satisfies \eqref{eq:ode_phi_trav} with $\omega = 0$ and $\phi \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2} \in L^2 \cap \mathcal{C}^0$ from Corollary \ref{cor:v_ln_v_Lp}. Moreover, $\phi' \in L^2$ since $\phi \in E_\textnormal{logGP}$. Therefore, we get $\phi'' \in L^2$, which leads to $\phi' \in H^1$. By Sobolev embedding, this leads in particular to $\phi' \in \mathcal{C}^0_b$ and, coming back to \eqref{eq:ode_phi_trav} once again, we get $\phi'' \in \mathcal{C}^0$. This leads to $\phi \in H^2_\textnormal{loc} \cap \mathcal{C}^2$. $\phi' \in H^1$ also gives
$\lim_{\@ifstar{\oldabs}{\oldabs*}{x} \rightarrow \infty} \phi' (x) = 0$. Furthermore, with Lemma \ref{lem:E_logGP_descr}, $\@ifstar{\oldabs}{\oldabs*}{\phi} - 1 \in H^1$, which means that $\lim_{\@ifstar{\oldabs}{\oldabs*}{x} \rightarrow \infty} \@ifstar{\oldabs}{\oldabs*}{\phi (x)} = 1$ and that $\phi$ is bounded, which shows that $\phi''$ is also bounded by \eqref{eq:ode_phi_trav}.
\end{proof}
\section{Solitary waves in the one-dimensional case}
\label{sec:solitary}
In this section, we prove Theorem \ref{th:sol_wave} in the case $c=0$.
By Theorem~\ref{th:value_omega}, we know that $\omega = 0$. Thus, the
equation we address in this section is:
\begin{equation} \label{eq:solit_phi_ode}
\phi'' = \lambda \phi \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2}.
\end{equation}
\subsection{Properties of a solitary wave in dimension $d=1$}
We first assume that there exists a non-constant traveling wave $u$ of
the form \eqref{eq:trav_wave} with $c = 0$, solution to
\eqref{logGP} in the energy space, and we gather some of its properties.
First, we show that $\phi$ satisfies an interesting energy equality.
\begin{lem} \label{lem:energy_ode}
If $\phi\in E_\textnormal{logGP}$, then
\begin{equation} \label{eq:energy_ode}
|\phi'|^2 = \lambda \( |\phi|^2 \ln{|\phi|^2} - |\phi|^2 + 1 \) .
\end{equation}
%
\end{lem}
\begin{proof}
Multiplying \eqref{eq:solit_phi_ode} by $2 \bar\phi' (x)$, and
taking the real value, we get
\begin{equation*}
\frac{\diff}{\diff x} |\phi'|^2 = \lambda \frac{\diff}{\diff
x} \( |\phi|^2 \ln{|\phi|^2} - |\phi|^2 + 1 \).
\end{equation*}
%
By the facts that $\phi \in E_\textnormal{logGP}$ and $|\phi|^2
\ln{|\phi|^2} - |\phi|^2 + 1 \ge 0$, we get $|\phi'|^2 \in L^1$
and $|\phi|^2 \ln{|\phi|^2} - |\phi|^2 + 1 \in L^1$, so we can
integrate the above equation, and no additional constant of
integration appears.
%
\end{proof}
\begin{cor} \label{cor:d_phi_sign}
If $\phi\in E_\textnormal{logGP}$ is real-valued on some unbounded interval $I$, then either $\phi'$ never vanishes and does not change sign on $I$, or $\phi$ is constant on $I$ (equals to $\pm 1$).
\end{cor}
\begin{proof}
If there is some $y \in I$ such that $\phi' (y) = 0$, then Lemma \ref{lem:energy_ode} shows that $\phi (y) = \pm 1$ as $x \mapsto x \ln{x} - x + 1$ vanishes only for $x = 1$ on $\mathbb{R}_+$, and thus the uniqueness part of the Cauchy theorem applied to \eqref{eq:solit_phi_ode} gives (as long as it does not vanish) $\phi \equiv \pm 1$ on $I$. On the other hand, if there is no such $y$, $\phi'$ cannot change sign on $I$ since it is continuous and never vanishes.
\end{proof}
Then, we show that $\phi$ can be taken real-valued up to a gauge on
some neighborhood of $+ \infty$.
\begin{lem} \label{lem:infinite_int}
There exists $x^- \in \mathbb{R} \cup \{ - \infty \}$ and $\theta
\in [0, 2 \pi)$ such that $\phi (x) e^{-i \theta}$ is real-valued
and positive for all $x \geq x^-$, and, either $x^- = - \infty$, or
$\lim_{x\to x^-} \phi(x) = 0$.
\end{lem}
\begin{proof}
We can write \eqref{eq:solit_phi_ode} as a first order differential system,
\begin{equation} \label{eq:ode_phi_psi}
\begin{pmatrix}
\phi \\
\psi
\end{pmatrix}'
=
\begin{pmatrix}
\psi \\
\lambda \phi \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2}
\end{pmatrix}
\eqqcolon F \(
\begin{pmatrix}
\phi \\
\psi
\end{pmatrix}
\).
\end{equation}
%
$F$ is a continuous function on $\mathbb{C}^2$ and is of class
$\mathcal{C}^1$ on $(\mathbb{C} \setminus \{ 0 \}) \times
\mathbb{C}$.
%
From Lemma~\ref{lem:solit_reg_lim}, there exists $x_0 \in
\mathbb{R}$ such that $\frac{3}{2}\ge \@ifstar{\oldabs}{\oldabs*}{\phi (x)} \ge \frac{1}{2}$ and
$\@ifstar{\oldabs}{\oldabs*}{\phi' (x)} \leq \frac{1}{2}$ for all $x \geq
x_0$. Therefore, by Cauchy-Lipschitz Theorem, there exists a maximal
interval $I_+ = (x^-, x^+)$ such that $x_0 \in I_+$ and $(\phi,
\phi')$ is the unique solution of \eqref{eq:ode_phi_psi} on $I_+$
with initial data $\phi (x_0)$ and $\psi (x_0) = \phi' (x_0)$,
with values in $\mathbb{C}^2 \setminus (\{ 0 \} \times
\mathbb{C})$. Moreover, if $\@ifstar{\oldabs}{\oldabs*}{x^+} < \infty$, then either
$\lim_{x \rightarrow x^+} \phi (x) = 0$ or $\lim_{x \rightarrow
x^+} (\@ifstar{\oldabs}{\oldabs*}{\phi (x)} + \@ifstar{\oldabs}{\oldabs*}{\psi (x)}) = \infty$, and similarly
for the limit $x \rightarrow x^-$.
%
However, since we already know that $\frac{3}{2}\ge \@ifstar{\oldabs}{\oldabs*}{\phi (x)}
\ge \frac{1}{2}$ and $\@ifstar{\oldabs}{\oldabs*}{\phi' (x)} \leq \frac{1}{2}$ for all
$x \geq x_0$, we can already conclude that $x^+ = \infty$.
%
On the other hand, since $\phi \in \mathcal{C}^2 (\mathbb{R})$ by Lemma~\ref{lem:solit_reg_lim}, we know that $\lim_{x \rightarrow x^-} (\@ifstar{\oldabs}{\oldabs*}{\phi (x)} + \@ifstar{\oldabs}{\oldabs*}{\psi (x)}) = \infty$ cannot be true if $x^- > - \infty$.
On $I_+$, $\phi$ does not vanish. Hence, we can use a polar decomposition: $\phi = \rho e^{i \theta}$, with $\rho$ and $\theta$ real-valued and defined as
\begin{equation} \label{eq:def_rho_theta}
\rho = \@ifstar{\oldabs}{\oldabs*}{\phi}, \qquad
\theta' = \frac{\Im ( \phi' \overline{\phi} )}{\rho^2}, \quad
\theta (x_0) \in [0, 2 \pi) \text{ such that } e^{i \theta (x_0)} = \frac{\phi (x_0)}{\rho (x_0)}.
\end{equation}
%
The regularity of $\phi$ and the fact that it does not vanish on $I_+$ ensure that $\rho$ and $\theta$ are well defined and $\mathcal{C}^2$ on $I_+$.
Thus, we can compute $\phi''$ in terms of $\rho$, $\theta$, and
their derivatives:
\begin{equation*}
\phi'' = \(\rho'' - \rho (\theta')^2 + 2 i \rho' \theta'\) e^{i \theta}.
\end{equation*}
%
Substituting into \eqref{eq:solit_phi_ode}, we get (after simplification by $e^{i \theta}$)
\begin{equation} \label{eq:polar_form_ode}
\rho'' - \rho (\theta')^2 + 2 i \rho' \theta' + i \rho \theta'' = \lambda \rho \ln{\rho^2}.
\end{equation}
%
By taking the imaginary part, we get $\rho \theta'' + 2 \rho' \theta' = 0$. Since $\rho$ does not vanish on $I_+$, we get for all $x \in I_+$:
\begin{equation*}
\theta' (x) = \frac{c_0}{\rho (x)^2},
\end{equation*}
%
where $c_0$ is a constant.
%
Then, we can compute for all $x \in I_+$
\begin{equation*}
\phi' (x) = \(\theta' (x) + i \rho (x) \theta' (x)\) e^{i \theta (x)} = \(\theta' (x) + i \frac{c_0}{\rho (x)}\) e^{i \theta (x)},
\end{equation*}
%
which leads to $\frac{c_0}{\rho} \in L^2 (I_+)$ and at the same time $\lim_{+ \infty} \frac{c_0}{\rho} = c_0$ since we know that $\lim_{x \rightarrow + \infty} \rho (x) = 1$. Combining these two arguments yields $c_0 = 0$.
%
Hence, $\theta$ is constant on $I_+$, which gives the conclusion.
\end{proof}
From now on, without loss of generality, we can assume that $\theta =
0$ in Lemma~\ref{lem:infinite_int}, up to changing $\phi$ to $e^{- i
\theta}\phi$.
\begin{lem} \label{lem:x_minus_prop}
Let $x^- \in \mathbb{R} \cup \{ - \infty \}$ from
Lemma~\ref{lem:infinite_int}. Then $x^- > - \infty$. Moreover,
$\phi (x) \in (0, 1)$ and $\phi' (x) > 0$ for all $x > x^-$. Last,
$\phi (x^-) = 0$ and $\phi' (x^-) = 1$.
\end{lem}
\begin{proof}
From Corollary \ref{cor:d_phi_sign}, we know that $\phi'$ never vanishes and does not change sign on $(x^-, \infty)$.
%
By contradiction, if $\phi (x_2) > 1$ for some $x_2 > x^-$, then
\begin{itemize}
\item Either $\phi' > 0$ on $(x^-, \infty)$, and thus $\phi (x) \geq \phi (x_2)$ for all $x \geq x_2$, which is in contradiction with the fact that $\lim_\infty \@ifstar{\oldabs}{\oldabs*}{\phi} = 1$.
%
\item Or $\phi' < 0$ on $(x^-, \infty)$, which means that $\phi (x) \geq \phi (x_2) > 1$ for all $x^- < x \leq x_2$. Therefore, $x^- = - \infty$, but then we have the same contradiction as in the previous case, for $- \infty$.
\end{itemize}
%
Therefore, $\phi (x) \in (0, 1)$ for all $x > x^-$, and this leads to $\phi' > 0$ on $(x^-, \infty)$.
%
Now, remark that $f(x) = x^2 \ln{x^2} - x^2 + 1$ is decreasing on $[0, 1]$. Thus, $\lambda f(\phi) = (\phi')^2$ is decreasing, which shows that, taking $x_0 \in (x^-, \infty)$, we have $\phi' (x)^2 \geq \phi' (x_0)^2$ for all $x^- < x \leq x_0$, and thus $\phi' (x) \geq \phi' (x_0)$. By integration, we get for all $x \in (x^-, x_0]$,
\begin{equation*}
\phi (x) \leq \phi (x_0) + (x - x_0) \phi' (x_0).
\end{equation*}
%
Since the right hand side goes to $- \infty$ as $x \rightarrow -
\infty$, and since $\phi > 0$ on $(x^-, \infty)$, this yields $x^- >
- \infty$. Thus, we are in the second case of Lemma
\ref{lem:infinite_int}, and since $\phi$ is continuous, we get
$\phi (x^-) = 0$. Last, \eqref{eq:energy_ode} holds for $x = x^-$,
and we know by continuity that $\phi' (x^-) \geq 0$, which gives
$\phi' (x^-) = 1$.
\end{proof}
\begin{lem} \label{lem:inifnite_int_2}
Let $x^- \in \mathbb{R}$ from Lemma \ref{lem:infinite_int}. Then, for all $x < x^-$, $\phi (x)$ is real-valued, negative and increasing.
\end{lem}
\begin{proof}
From Lemma \ref{lem:x_minus_prop}, we know that $\phi (x^-) = 0$
and $\phi' (x^-) = 1$. Therefore, there exists $\delta > 0$ such
that $\phi \neq 0$ on $[x^- - \delta, x^-)$. Define
\begin{equation*}
x^-_0 \coloneqq \inf \{ y < x^- \, | \, \forall x \in (y, x^-), \phi (x) \neq 0 \} < x^- - \delta.
\end{equation*}
%
On $I_2 \coloneqq (x^-_0, x^-)$, we can use a polar factorization in
the same way as in \eqref{eq:def_rho_theta}: $\phi = \rho e^{i
\theta}$, where $\rho$ and $\theta$ also satisfy
\eqref{eq:polar_form_ode} on $I_2$. Taking again the imaginary
part and integrating the ODE, we have, like before,
\begin{equation*}
\theta' (x) = \frac{c_0}{\rho (x)^2}, \qquad \forall x \in I_2.
\end{equation*}
%
We still have then
\begin{equation*}
\phi' (x) = \(\theta' (x) + i \frac{c_0}{\rho (x)}\) e^{i \theta (x)},
\end{equation*}
%
and once again we must have $\frac{c_0}{\rho (x)} \in L^2 (I_2)$
since $\phi' \in L^2$. However, we know that $\phi (x^-) = 0$ and
$\phi' (x^-) = 1$, which means that $\phi (x) \sim x - x^-$ as $x\to x^-$.
%
In terms of $\rho$, this implies $\rho (x) \sim\@ifstar{\oldabs}{\oldabs*}{x - x^-}$
as $x\to x^-$, and $\frac{c_0}{\rho (x)} \in
L^2 (I_2)$ if and only if $c_0 = 0$. Once again, $\theta$ is
constant on $I_2$, and therefore so is $\frac{\phi}{\rho}$. With
the previous asymptotics, we know that it tends to $-1$ at
$x^-$. Thus, we get
\begin{equation*}
\phi = - \rho.
\end{equation*}
%
This shows that $\phi$ is real-valued and negative on $I_2$. Then, we can apply Lemma~\ref{lem:energy_ode} on $(x^-_0, \infty)$, which shows that $\phi'$ does not vanish on this interval, and thus $\phi'$ does not change sign once again, which proves that $\phi' > 0$ on $I_2$. Therefore, $\phi (x) \leq \phi (x^- - \delta) < 0$ for all $x \in (x^-_0, x^- - \delta)$, which proves that $x^-_0 = - \infty$.
\end{proof}
It only remains to analyze the limit of $\phi$ at $- \infty$.
\begin{lem}
We have
\begin{equation*}
\lim_{x\to - \infty} \phi (x)= - 1
\end{equation*}
\end{lem}
\begin{proof}
We have seen that $\phi$ is increasing, thus it has a limit at $-
\infty$. From Lemma~\ref{lem:inifnite_int_2}, we know it is
negative. Moreover, $\@ifstar{\oldabs}{\oldabs*}{\phi} - 1 \in L^2$, which means that $\phi
+ 1 \in L^2 ((- \infty, 0))$. The conclusion easily follows.
\end{proof}
The conclusion of all the previous lemmas is the following:
\begin{cor} \label{cor:necessary_cond}
If $u$ is a non-constant traveling wave solution to \eqref{logGP}
of the form \eqref{eq:trav_wave} with $c=0$, then there exists $\theta \in \mathbb{R}$ such that $e^{- i \theta} \phi$ is a real-valued, increasing, $\mathcal{C}^2$ function with values in $(-1, 1)$ which vanishes at a unique point $x_0$.
%
Moreover, $\phi_0 (x) = e^{- i \theta} \phi (x + x_0)$ also
satisfies \eqref{logGP},
\begin{equation*}
\lim_{x \rightarrow \pm \infty} \phi_0 (x) = \pm 1.
\end{equation*}
and
\begin{equation} \label{eq:energy_ode_compl}
(\phi_0')^2 = \lambda \Bigl( \phi_0^2 \ln{\phi_0^2} - \phi_0^2 + 1 \Bigr) \qquad \text{ on } \mathbb{R}.
\end{equation}
\end{cor}
\subsection{Analysis of the new ODE and proof of Theorem \ref{th:sol_wave}}
From Corollary~\ref{cor:necessary_cond}, $\phi_0$ vanishes at $x=0$,
satisfies \eqref{eq:energy_ode_compl} and is strictly increasing,
i.e. $\phi_0' > 0$. Thus, $\phi_0$ satisfies the ODE
\begin{equation} \label{eq:new_ode}
\phi_0' = \sqrt{\lambda} \sqrt{\phi_0^2 \ln{\phi_0^2} - \phi_0^2 + 1} \qquad \text{ on } \mathbb{R}.
\end{equation}
The uniqueness of this function is the topic of the following lemma.
\begin{lem} \label{lem:exist_unique_new_ode}
There exists a unique function $\phi_0$ satisfying \eqref{eq:new_ode} with the initial data $\phi_0 (0) = 0$. It is defined on $\mathbb{R}$ and satisfies
\begin{equation*}
\lim_{x \rightarrow \pm \infty} \phi_0 (x) = \pm 1.
\end{equation*}
%
Moreover, $\phi_0 \in E_\textnormal{logGP}$.
\end{lem}
\begin{proof}
We already know that $f(x) = x \ln{x} - x + 1 \geq 0$ for all $x
\geq 0$, and $f$ vanishes only at $x=1$. Moreover, by simple
computations, we check that $g(x) = f(x^2)$ is $\mathcal{C}^1$ on
$\mathbb{R}$, and $\mathcal C^2$ on $\R\setminus\{0\}$. We
also compute $f' (1) = 0$ and $f'' (1) = 1$, hence
\begin{equation*}
g(x)\Eq x {\pm 1} \frac{1}{2}\(x^2-1\)^2.
\end{equation*}
%
From these facts, we deduce that $h = \sqrt{g}$ is a
$\mathcal{C}^0 (\mathbb{R}) \cap \mathcal{C}^1 (\mathbb{R}
\setminus \{ -1, 1 \})$ function such that $h(x) \sim_{x
\rightarrow \pm 1} \@ifstar{\oldabs}{\oldabs*}{x^2 - 1}/\sqrt 2$. Therefore, it is
$\mathcal{C}^{0, 1}$ locally on $\mathbb{R}$, and we may invoke
Cauchy-Lipschitz Theorem: there exists a unique $\phi_0$ satisfying
$\phi_0' = \sqrt{\lambda} h(\phi_0)$ with the initial condition
$\phi_0 (0) = 0$, on a maximal interval $I\ni 0$ of existence.
%
Then, $\phi_0$ cannot reach $1$ or $-1$, because the constant
function $1$ and $-1$ are both solutions to this ODE, and it would
contradict the uniqueness in the Cauchy-Lipschitz Theorem. Since $\phi_0$ is
continuous, we infer that $\phi_0 (x) \in (-1, 1)$ for all $x
\in \mathbb{R}$. Moreover, we can easily prove that there exists
$c_1> 0$ such that
\begin{align*}
h(y) &\ge c_1 (y + 1), \quad\forall y \in [-1, 0], \\
h(y) &\ge c_1 (1 - y),\quad \forall y \in [0, 1].
\end{align*}
%
Since we know that $\phi_0 (x) \in (-1, 0]$ for all $x \leq 0$, we can estimate
\begin{equation*}
\phi_0' (x) \geq c_1 \sqrt{\lambda} (\phi_0 (x) + 1), \qquad \forall x \leq 0.
\end{equation*}
%
By integrating backward, we get for all $x \leq 0$,
\begin{align*}
-1 < \phi_0 (x) &\leq -1 + e^{\sqrt{\lambda} c_1 x}, \\
0 < \phi_0' (x) &\leq c_1 \sqrt{\lambda} e^{\sqrt{\lambda} c_1 x}.
\end{align*}
%
Similarly, we also have for all $x \geq 0$
\begin{align*}
1 - e^{- \sqrt{\lambda} c_1 x} &\leq \phi_0 (x) < -1, \\
0 < \phi_0' (x) &\leq c_1 \sqrt{\lambda} e^{\sqrt{\lambda} c_1 x}.
\end{align*}
%
Those estimates prove the expected limits at $\pm \infty$, and also
show that $\phi_0' \in L^2$ together with $\@ifstar{\oldabs}{\oldabs*}{\phi_0} - 1 \in L^2$. We
conclude that $\phi_0 \in E_\textnormal{logGP}$ by
Lemma~\ref{lem:E_logGP_descr}.
\end{proof}
The previous result obviously proves that we have a set of solitary
waves for \eqref{logGP}.
\begin{cor} \label{cor:set_soli_waves}
For any $\theta \in \mathbb{R}$ and any $x_0 \in \mathbb{R}$,
$e^{i \theta} \phi_0 (. - x_0) \in E_\textnormal{logGP}$ is a
solitary (stationary) wave of \eqref{logGP}.
\end{cor}
It remains to show that they are the only possible solitary waves in
order to prove Theorem~\ref{th:sol_wave} in the case $c=0$.
\begin{proof}[Proof of Theorem \ref{th:sol_wave}: case $c=0$.]
Let $\phi \in E_\textnormal{logGP}$ and $u$ be a traveling wave of
the form \eqref{eq:trav_wave} with $c=0$. From Corollary
\ref{cor:necessary_cond}, we can find $\theta \in \mathbb{R}$ and
$x_0 \in \mathbb{R}$ such that $e^{- i \theta} \phi (. + x_0)$ is
a real-valued, increasing, $\mathcal{C}^2$ function
with values in $(-1, 1)$, which vanishes at $0$, and satisfies
\eqref{eq:energy_ode_compl}, i.e. \eqref{eq:new_ode} by
positivity of its derivative.
%
It is therefore $\phi_0$, defined in
Lemma~\ref{lem:exist_unique_new_ode} by uniqueness given by the
same lemma: $e^{- i \theta} \phi (. + x_0) = \phi_0$, i.e. $\phi =
e^{i \theta} \phi_0 (. - x_0)$.
\end{proof}
\section{Traveling waves in the one-dimensional case}
\label{sec:traveling}
\subsection{Admissible velocities}
\label{sec:velocities}
The goal of this section is to prove Theorem~\ref{th:trav_waves_1d},
and characterize velocities of nontrivial (that is, nonconstant)
traveling waves. We assume that there exists a traveling
wave of the form \eqref{eq:trav_wave} solution to \eqref{logGP},
i.e. $\phi \in E_\textnormal{logGP}$ solution to
\eqref{eq:ode_phi_trav}.
\begin{lem} \label{lem:another_energy_eq}
The function $\eta = 1 - \@ifstar{\oldabs}{\oldabs*}{\phi}^2$ satisfies, on $\mathbb{R}$,
\begin{equation*}
\frac{1}{2} (\eta')^2 = h_c (\eta),
\end{equation*}
where
\begin{equation*}
h_c (y) \coloneqq \lambda \Bigl( (1-y)^2 \ln(1-y)^2 - (1-y)^2 + 1 \Bigr) - \frac{2 \lambda +c^2}{2} y^2.
\end{equation*}
\end{lem}
\begin{proof}
We follow the same lines as in \cite{BeGrSa08}. Writing
$\phi=\phi_1+i \phi_2$, \eqref{eq:ode_phi_trav} becomes
\begin{System} \label{eq:vR}
\phi_1''+ c \phi_2' = \lambda \phi_1\ln\(\phi_1^2+\phi_2^2\),\\[5pt]
\phi_2''- c \phi_1' = \lambda \phi_2\ln\(\phi_1^2 + \phi_2^2\).
\end{System}
We multiply the first equation by $\phi_2$ and the second one by $\phi_1$, and we subtract in order to get:
\begin{equation*}
\( \phi_1' \phi_2 - \phi_1 \phi_2'\)' = -c (\phi_2 \phi_2' + \phi_1 \phi_1') = \frac{c}{2} \eta', \quad
\eta \coloneqq 1-\@ifstar{\oldabs}{\oldabs*}{\phi}^2.
\end{equation*}
Moreover, we know that $\phi$ is bounded by
Lemma~\ref{lem:solit_reg_lim}, thus $\phi_1' \phi_2, \phi_1
\phi_2' \in L^2$, and $\eta \in L^2$ by using
Lemma~\ref{lem:E_logGP_descr}. Thus, integrating the above
identity yields
\begin{equation}\label{eq:wronskien}
\phi_1' \phi_2 - \phi_1 \phi_2' = \frac{c}{2} \eta.
\end{equation}
%
Now, we multiply the first equation of \eqref{eq:vR} by $\phi_1'$, the second by $\phi_2'$ and we sum in order to get:
\begin{equation*}
\phi_1' \phi_1'' + \phi_2' \phi_2'' = \lambda \(\phi_1 \phi_1' + \phi_2 \phi_2'\)\ln \(\phi_1^2 + \phi_2^2\),
\end{equation*}
which can be written as
\begin{equation*}
\(\@ifstar{\oldabs}{\oldabs*}{\phi'}^2\)' = \lambda \(\@ifstar{\oldabs}{\oldabs*}{\phi}^2 \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2} -
\@ifstar{\oldabs}{\oldabs*}{\phi}^2 + 1\)'.
\end{equation*}
%
We know that both $\@ifstar{\oldabs}{\oldabs*}{\phi'}^2$ and $\@ifstar{\oldabs}{\oldabs*}{\phi}^2 \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2} - \@ifstar{\oldabs}{\oldabs*}{\phi}^2 + 1$ (which is positive and whose integral over $\mathbb{R}$ is bounded by $\mathcal{E}_\textnormal{logGP} (\phi)$) are in $L^1$, hence:
\begin{equation*}
\@ifstar{\oldabs}{\oldabs*}{\phi'}^2 = \lambda \bigl( \@ifstar{\oldabs}{\oldabs*}{\phi}^2 \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2} - \@ifstar{\oldabs}{\oldabs*}{\phi}^2 + 1 \Bigr) \qquad \text{on } \mathbb{R}.
\end{equation*}
We deduce then
\begin{align*}
\eta'' &= -2|\phi'|^2 - 2 (\phi_1 \phi_1'' + \phi_2 \phi_2'')\\
&= -2|\phi'|^2 -2 \phi_1 \(-c \phi_2' + \lambda \phi_1 \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2}\) - 2 \phi_2 \( c \phi_1' + \lambda \phi_2 \ln{\@ifstar{\oldabs}{\oldabs*}{\phi}^2}\)\\
&=-4 \lambda (1-\eta)\ln(1-\eta) - 2 \lambda \eta - c^2\eta = h_c' (\eta).
\end{align*}
%
Multiplying by $\eta'$, we obtain
\begin{equation*}
\frac{1}{2} \( (\eta')^2 \)' = ( h_c (\eta) )'.
\end{equation*}
%
Since $\lim_{\pm \infty} \eta' = \lim_{\pm \infty} \eta = 0$ from
Lemma~\ref{lem:solit_reg_lim}, we conclude by integrating.
\end{proof}
\begin{rem}
This lemma may be extended to the case of any non-linearity $F(u)$
with $F$ a continuous function on $\mathbb{C}$ of the form $F (u)
= u \, f(\@ifstar{\oldabs}{\oldabs*}{u}^2)$ with $f(1) = 0$.
\end{rem}
\begin{cor} \label{cor:trav_wave_never_vanish}
If $c \neq 0$, then $\phi$ never vanishes.
\end{cor}
\begin{proof}
It comes from the fact that $h_c (1) = - \frac{c^2}{2} < 0$ if $c
\neq 0$, while $h_c (\eta) = \frac{1}{2} (\eta')^2 \geq
0$. Therefore, $\eta (x) \neq 1$, that is $\phi (x) \neq 0$, for
all $x \in \mathbb{R}$.
\end{proof}
\begin{cor} \label{cor:eta_zero_large_c}
If $c^2 \ge 2 \lambda$, then $\eta \equiv 0$.
\end{cor}
\begin{proof}
The limiting case $c^2=2\lambda$ is treated like in
\cite{BeGrSa08}: if $\eta$ is not identically zero, then in view
of Lemma~\ref{lem:solit_reg_lim}, by
translation invariance, we may assume
\begin{equation*}
|\eta(0)|=\max\{|\eta(x)|,\ x\in \R\}>0,\quad \eta'(0)=0.
\end{equation*}
In view of Lemma~\ref{lem:another_energy_eq}, $h_c(\eta(0))=0$. On the
other hand, Corollary~\ref{cor:trav_wave_never_vanish} implies that
$1-\eta(x)>0$ for all $x\in \R$, and direct computation yields
\begin{equation*}
h''_c(y) =4\lambda\ln(1-y),\quad 0\le y<1.
\end{equation*}
Therefore, $h''_c(y)$ is positive for $y<0$, negative for $0<y<1$. As
$h'_c(0)=0$, $h'_c$ is negative on $(-\infty,0)\cup (0,1)$: $h_c$ is
decreasing on $(-\infty,1)$.
Since $h_c(0)=0$, the only zero of $h_c$ on $[0,1)$ is the origin, and so
$\eta(0)=0$, hence a contradiction: $\eta$ is identically zero.
\smallbreak
Suppose now $c^2>2\lambda$. Since $h_c(0)=h_c'(0)=0$, and
$h_c''(0)= 2\lambda-c^2$,
there exists $\varepsilon > 0$ such that $h_c (y) < 0$ for all $y
\in [- \varepsilon, \varepsilon] \setminus \{ 0 \}$. On the other
hand, $h_c (\eta)$ is continuous, nonnegative (from
Lemma~\ref{lem:another_energy_eq}) and tends to $0$ at $\pm
\infty$. The conclusion follows easily.
\end{proof}
\begin{lem}
If $c^2 \ge 2 \lambda$, then $\phi$ is constant.
\end{lem}
\begin{proof}
By Corollary \ref{cor:eta_zero_large_c}, we know that $\@ifstar{\oldabs}{\oldabs*}{\phi
(x)} = 1$ for all $x \in \mathbb{R}$. Since $\phi$ is a
$\mathcal{C}^2_b (\mathbb{R})$ function, there exists a
real-valued function $\theta \in \mathcal{C}^2$ such that $\phi =
e^{i \theta}$ (defined like in \eqref{eq:def_rho_theta} for
instance).
%
By substitution in \eqref{eq:ode_phi_trav}, $\theta$ satisfies then
\begin{equation*}
c \theta' + i \theta'' - (\theta')^2 = 0.
\end{equation*}
%
By taking the imaginary part, we get $\theta'' = 0$. Since we must
have $\phi' = i \theta' e^{i \theta} \in L^2$, we get $\theta' \in
L^2$ and thus $\theta' = 0$.
\end{proof}
\subsection{Nontrivial traveling waves}
We now consider the case $0 < c^2 < 2 \lambda$.
Define, for all $y > 0$,
\begin{align*}
f_c (y) &\coloneqq \frac{c^2}{4} \Bigl( \frac{1}{y^3} - y \Bigr) + \lambda y \ln{y^2}, \\
g_c (y) &\coloneqq - \frac{c^2}{4} \frac{(1 - y^2)^2}{y^2} +
\lambda (y^2 \ln{y^2} - y^2 + 1).
\end{align*}
The following lemma is established by direct calculations:
\begin{lem} \label{lem:analysis_fcts}
Let $c$ such that $0 < c^2 < 2 \lambda$. There exts $0<y_0<y_1<1$
such that the following holds:
\begin{itemize}
\item $f_c$ has exactly two zeroes on $(0,+\infty)$: $y_1$ and $1$.
\item $f_c$ is positive on $(0,y_1)\cup (1,+\infty)$, negative on
$(y_1,1)$.
\item $g_c$ has exactly two zeroes on $(0,+\infty)$: $y_0$ and $1$.
\item $g_c$ is negative on $(0,y_0)$, positive on
$(y_0,+\infty)\setminus \{0\}$.
\item There exsits $C_c>0$ such that $\frac{1}{C_c} (1 - y)^2 \leq g_c
(y) \leq C_c (1 - y)^2$ for all $y \in (y_1, 1)$.
\end{itemize}
\end{lem}
\begin{proof}
We compute
\begin{equation*}
f_c'(y) = \frac{c^2}{4}\(-\frac{3}{y^4}-1\)+2\lambda \ln y
+2\lambda,\quad f_c''(y) = \frac{3c^2}{y^5}+\frac{2\lambda}{y}>0.
\end{equation*}
As
\begin{equation*}
\lim_{y\to 0} f_c'(y) = -\infty,\quad \lim_{y\to +\infty} f_c'(y) = +\infty,
\end{equation*}
the derivative $f_c'$ has a unique zero on $(0,+\infty)$, which we
denote by $y_2>0$. Since $f_c'(1) = 2\lambda-c^2>0$, we know that
$0<y_2<1$: $f_c(1)=0$, hence $f_c(y_2)<0$, and there exists a unique
$0<y_1<y_2$ such that $f_c(y_1)=0$.
\smallbreak
We note that $f_c=\frac{1}{2}g_c'$, and in view of the above pieces
of information, we can draw:
\smallbreak
\begin{tikzpicture}
\tkzTabInit[lgt=1.2,espcl=2]{$y$ / 1 , $f_c(y)$ / 1 , $g_c(y)$ / 1}{0,
$y_1$, $y_2$, $1$, $+\infty$}
\tkzTabVar{D+ / $+\infty$ , R/ , - / $-$, R/ , + / $+\infty$ }
\tkzTabIma{1}{3}{2}{$0$}
\tkzTabIma{3}{5}{4}{$0$}
\tkzTabVar{D- / $-\infty$ , + / $+$ , R/ , - / $0$ , + / $+\infty$ }
\end{tikzpicture}
In particular, $g_c$ is increasing on $(0,y_1)$, from $-\infty$ to a
positive value, and there exists a unique $y_0\in (0,y_1)$ such that
$g_c(y_0)=0$, and the lemma follows easily.
\end{proof}
\begin{lem} \label{lem:carac_trav_waves}
If $c$ is such that $0 < c^2 < 2 \lambda$, then $\rho = \@ifstar{\oldabs}{\oldabs*}{\phi}$ satisfies
\begin{equation} \label{eq:rho_ode}
\rho'' = f_c (\rho),
\end{equation}
\begin{equation} \label{eq:energy_ode_c_compl}
(\rho')^2 = g_c (\rho).
\end{equation}
%
Moreover, there exists $x_0 \in \mathbb{R}$ such that $\rho (x_0)
= y_0$, where $y_0$ is defined in Lemma~\ref{lem:analysis_fcts},
and $\theta \in \mathcal{C}^2$, defined so that $\phi = \rho e^{i
\theta}$, satisfies
\begin{equation} \label{eq:theta_ode}
\theta' = \frac{c}{2} \Bigl( 1 - \frac{1}{\rho^2} \Bigr).
\end{equation}
\end{lem}
\begin{proof}
By Corollary \ref{cor:trav_wave_never_vanish}, $\phi$ never vanishes. It is also a $\mathcal{C}^2_b$ function. Therefore, we can define $\rho$ and $\theta$ as in \eqref{eq:def_rho_theta} so that $\phi = \rho e^{i \theta}$. They satisfy
\begin{equation} \label{eq:trav_ode_theta_rho}
- i c \rho' + c \rho \theta' + \rho'' - \rho (\theta')^2 + 2 i \rho' \theta' + i \rho \theta'' = \lambda \rho \ln{\rho^2}.
\end{equation}
%
By taking the imaginary part, we get
\begin{equation} \label{eq:ode_theta}
- c \rho' + \rho \theta'' + 2 \rho' \theta' = 0.
\end{equation}
%
As $\rho$ never vanishes, the solution of this ODE on $\theta'$ takes the form
\begin{equation*}
\theta' (x) = \frac{c}{2} + \frac{c_0}{\rho (x)^2},
\end{equation*}
%
where $c_0$ is a real constant. Then, there holds
\begin{equation*}
\phi' = \Bigl( \rho' + i \Bigl( \frac{c}{2} \rho + \frac{c_0}{\rho} \Bigr) \Bigr) e^{i \theta} \in L^2.
\end{equation*}
%
On the other hand, we already know that $\lim_{\pm \infty} \rho = 1$, thus we must have $c_0 = - \frac{c}{2}$, so that
\begin{equation*}
\theta' (x) = \frac{c}{2} \Bigl( 1 - \frac{1}{\rho (x)^2} \Bigr).
\end{equation*}
%
Now, \eqref{eq:trav_ode_theta_rho} reads
\begin{align*}
\rho'' &= - \frac{c^2}{2} \Bigl( \rho - \frac{1}{\rho} \Bigr) + \frac{c^2}{4} \rho \Bigl( 1 - \frac{1}{\rho^2} \Bigr)^2 + \lambda \rho \ln{\rho^2} \\
&= - \frac{c^2}{2} \Bigl( \rho - \frac{1}{\rho} \Bigr) + \frac{c^2}{4} \Bigl( \rho - \frac{2}{\rho} + \frac{1}{\rho^3} \Bigr) + \lambda \rho \ln{\rho^2} \\
&= \frac{c^2}{4} \Bigl( \frac{1}{\rho^3} - \rho \Bigr) + \lambda \rho \ln{\rho^2}.
\end{align*}
%
After multiplication by $\rho'$ and integration, we find:
\begin{equation*}
(\rho')^2 = - \frac{c^2}{4} \frac{(1 - \rho^2)^2}{\rho^2} + \lambda (\rho^2 \ln{\rho^2} - \rho^2 + 1).
\end{equation*}
%
We also know that if $\rho \equiv 1$, then $\theta' \equiv 0$ and
thus $\phi$ is constant. On the contrary, if $\phi$ is not
constant, then $\rho$ is not either, and thus has an extremum since
$\lim_{\pm \infty} \rho = 1$. Let $x_0$ be extremal. Then $\rho'
(x_0) = 0$, and therefore $0 = g_c
(\rho (x_0))$, with $\rho (x_0) \neq 1$.
Lemma~\ref{lem:analysis_fcts} implies $\rho (x_0) = y_0$.
\end{proof}
\begin{lem} \label{lem:cauchy_rho_0}
There exists a unique $\rho_c \in \mathcal{C}^2$ satisfying
\eqref{eq:rho_ode}, $\rho_c (0) = y_0$ and $\rho_c' (0) =
0$. Moreover, $\rho_c \geq y_0 > 0$, $\lim_{\pm \infty} \rho_c =
1$ and $\rho_c - 1 \in H^1$.
\end{lem}
\begin{proof}
Since $f_c$ is $\mathcal{C}^\infty$ on $(0, \infty)$, we can apply
Cauchy-Lipschitz Theorem to get a local solution. By multiplying
\eqref{eq:rho_ode} by $\rho_c'$ and integrating, this solution
satisfies also \eqref{eq:energy_ode_c_compl}.
Lemma~\ref{lem:analysis_fcts} then yields $\rho_c\ge y_0 $,
which proves that this solution is global.
Then, by \eqref{eq:rho_ode} and Lemma \ref{lem:analysis_fcts}
again, we get $\rho_c'' (0) > 0$. Hence there exists $\varepsilon > 0$
such that $\rho_c' < 0$ on $(- \varepsilon, 0)$ and $\rho_c' > 0$
on $(0, \varepsilon)$, and $\rho_c > y_0$ on $(- \varepsilon,
\varepsilon) \setminus \{ 0 \}$.
Let
\begin{align*}
x^+ &\coloneqq \sup \{ x > 0 \, | \, \rho_c' (y) > 0 \text{ for all } y \in (0, x) \}, \\
x^- &\coloneqq \inf \{ x < 0 \, | \, \rho_c' (y) < 0 \text{ for all } y \in (x, 0) \}.
\end{align*}
From \eqref{eq:energy_ode_c_compl} and
Lemma~\ref{lem:analysis_fcts} along with the fact that $\rho_c$ and
$\rho_c'$ are continuous, we know that either $x^+ = + \infty$ or
$\rho_c (x^+) = 1$, and similarly either $x^- = - \infty$ or $\rho_c
(x^-) = 1$. However, if $\rho (x_1) = 1$ at some point $x_1 \in
\mathbb{R}$, then \eqref{eq:energy_ode_c_compl} gives $\rho_c' (x_1)
= 0$ and the uniqueness in Cauchy-Lipschitz Theorem for \eqref{eq:rho_ode}
leads to $\rho_c \equiv 1$, hence a contradiction. Therefore, $x^+ =
+ \infty$ and $x^- = - \infty$, which implies that $\rho_c$ is
decreasing on $(- \infty, 0)$, increasing on $(0, \infty)$ and
$\rho_c (x) \in [y_0, 1)$ for all $x \in \mathbb{R}$. Thus, it has
limits $\ell_\pm > y_0$ at $\pm \infty$ and must also satisfy
$\lim_{\pm \infty} \rho_c' = 0$ in view of
\eqref{eq:energy_ode_c_compl}. Therefore, the limit must satisfy
$g_c (\ell_\pm) = 0$, which leads to $\ell_\pm = 1$.
%
Last, since $\rho_c' > 0$ on $(0, \infty)$,
\eqref{eq:energy_ode_c_compl} reads on this interval
\begin{equation*}
\rho_c' = \sqrt{g_c (\rho_c)}.
\end{equation*}
%
By Lemma \ref{lem:analysis_fcts} and since $\rho_c (x) \geq y_1$ for $x$ large enough, we get
\begin{equation*}
\rho_c' \geq \sqrt{C_c} (1 - \rho_c).
\end{equation*}
%
Thus $\rho_c$ converges exponentially fast to $1$ at $+ \infty$ by
Gronwall lemma, and therefore $\rho_c'$ converges exponentially
fast to $0$. The same
holds as $x \rightarrow - \infty$, hence
the conclusion $\rho_c - 1 \in H^1$.
\end{proof}
\begin{lem} \label{lem:exist_trav_waves_c}
For $c$ such that $0 < c^2 < 2 \lambda$ and $\theta_0 \in
\mathbb{R}$, define $\Theta_c$ by
\begin{equation*}
\Theta_c' = \frac{c}{2} \Bigl( 1 - \frac{1}{\rho_c^2} \Bigr), \qquad
\Theta_c (0) = \theta_0,
\end{equation*}
where $\rho_c$ is given by Lemma~\ref{lem:cauchy_rho_0}.
Then $\phi_c \coloneqq \rho_c e^{i \Theta_c} \in
E_\textnormal{logGP}$ and $u$ defined by \eqref{eq:trav_wave} is a
traveling wave for \eqref{logGP}.
\end{lem}
\begin{proof}
Direct computations show that $\Theta_c$ satisfies
\eqref{eq:ode_theta}, and thus $(\rho_c,\Theta_c)$ satisfies
\eqref{eq:trav_ode_theta_rho}, which is equivalent to $\phi_c$
satisfying \eqref{eq:ode_phi_trav} or $u$ satisfying
\eqref{logGP}. Therefore, we only have to prove that $\phi_c \in
E_\textnormal{logGP}$.
%
By Lemma~\ref{lem:cauchy_rho_0}, we already know that $\@ifstar{\oldabs}{\oldabs*}{\phi_c} - 1 = \rho_c - 1 \in H^1$. Moreover,
\begin{equation*}
\phi_c' = \( \rho_c' + i \rho_c \Theta_c\) e^{i \Theta_c} = \( \rho_c' + i \frac{c}{2} \frac{\rho_c^2 - 1}{\rho_c} \) e^{i \Theta_c}.
\end{equation*}
%
By Lemma \ref{lem:cauchy_rho_0} again, $\rho_c' \in L^2$ and
$\rho_c$ is bounded, and bounded away from $0$, so that
$\frac{\rho_c^2 - 1}{\rho_c} \in L^2$. We conclude $\phi_c \in
E_\textnormal{logGP}$ thanks to Lemma~\ref{lem:E_logGP_descr}.
\end{proof}
\begin{proof}[End of the proof of Theorem~\ref{th:sol_wave}]
Let $\phi \in E_\textnormal{logGP}$ and $u$ a traveling wave of
the form \eqref{eq:trav_wave}. By applying Lemma
\ref{lem:carac_trav_waves}, we get $x_0 \in \mathbb{R}$ such that
$\rho (x_0) = y_0$ where $\rho = \@ifstar{\oldabs}{\oldabs*}{\phi}$, which yields $\rho'
(x_0) = 0$ through \eqref{eq:energy_ode_c_compl}. Since $\rho (. +
x_0)$ also satisfies \eqref{eq:rho_ode}, this function satisfies
the assumptions of Lemma \ref{lem:cauchy_rho_0}. Thus $\rho (. +
x_0) = \rho_c$, and since $\theta$ satisfies \eqref{eq:theta_ode},
$\phi$ is of the form $\phi_c$ defined in Lemma
\ref{lem:exist_trav_waves_c} for some $\theta_0$.
\end{proof}
\section{Some open questions}
\label{sec:open}
\subsection*{Multidimensional solitary and traveling waves}
We have classified solitary and traveling waves in the one-dimensional
setting, using ODE techniques. The picture is of course rather
different for $d\ge 2$, see typically the case of traveling waves for
the Gross-Pitaevskii equation \eqref{eq:GP}, in
\cite{BeSa99,BeGrSa09}, where the role of hydrodynamical formulation
(via Madelung transform) is decisive. See also \cite{BeGrSa08} and references
therein.
\smallbreak
We note that contrary to the case with vanishing boundary condition at
infinity, one
cannot easily construct multidimensional waves from one-dimensional
ones by using the tensorization property evoked in the introduction. Recall that if
$u_1(t,x_1),\dots ,u_d(t,x_d)$ are solutions to the one-dimensional
logarithmic Schr\"odinger equation, then
$u(t,x):=u_1(t,x_1)\times\dots\times u_d(t,x_d)$ solves
the $d$-dimensional logarithmic Schr\"odinger equation. In the case of
vanishing boundary condition at
infinity, $u_j(t,\cdot)\in L^2(\R)$, hence $u(t,\cdot)\in
L^2(\R^d)$; see e.g. \cite{CaFe21} for some applications. On the other
hand, if $u_j(t,\cdot)\in
E_\textnormal{logGP}=E_\textnormal{GP}$ (the sets are
the same in the one-dimensional case), then even for $d=2$,
\begin{equation*}
(x_1,x_2)\mapsto u_1(t,x_1)u_2(t,x_2)
\end{equation*}
does not belong to the energy space
$E_\textnormal{logGP}=E_\textnormal{GP}$ (case of $\R^2$), as the
spatial derivatives are not in $L^2(\R^2)$.
\subsection*{Orbital stability}
In the case of the one-dimensional Gross-Pitaevskii equation, the
orbital stability of traveling waves was proven in \cite{GeZh09} by
using the fact that this equation is integrable, in the sense that it
admits a Lax pair, and solutions are analyzed thanks to
Zakharov-Shabat's inverse scattering method. It seems unlikely that
even for $d=1$, \eqref{logGP} is completely integrable, so the
stability of traveling waves will require a different
approach.
\smallbreak
In \cite{Lin02,BGSS08}, the orbital stability of the black soliton is
studied without using the integrable structure, but variational
arguments.
In \cite{BGS15} (dark soliton) and \cite{GrSm15} (black soliton), the
orbital stability is improved to asymptotic
stability (the modulation factors, which consist of a translation
parameter and a phase shift, are analyzed). These approaches may be
more suited to
\eqref{logGP}, but new
difficulties due to the special form of the nonlinearity arise,
starting with the control of the presence of vacuum, $\{u=0\}$.
\subsection*{Scattering}
Another way to study stability properties consists in establishing a
scattering theory, considering the plane wave solution, $u(t,x) =
e^{ik\cdot x-i|k|^2t}$, $k\in \R^d$, as a reference solution. By
Galilean invariance, the analysis is reduced to the case $k=0$, and
the question is the behavior of the solution as $t\to +\infty$ when
the initial datum $u_0\equiv 1$ is perturbed. In such a framework, a
crucial object is the linearized Schr\"odinger operator about the
constant $1$. Using in addition normal form techniques, a scattering
theory for \eqref{eq:GP} was developed in
\cite{GuNaTs06,GuNaTs07,GuNaTs09} for $d\ge 2$, and resumed in
\cite{KiMuVi16,KiMuVi18} in the case of the three-dimensional
cubic-quintic Schr\"odinger equation (the quintic term introduces the
new difficulty of an energy-critical factor).
\smallbreak
In the case of \eqref{logGP}, the new difficulty is due to the
fact that the logarithmic nonlinearity is not multilinear in $(u,\bar
u)$, reminding of the difficulty of giving sense to linearization in
such a context (see e.g. \cite{Fe21,Pel17}).
\subsection*{Convergence toward other models}
It is well-established that in the long wave r\'egime, the
Gross-Pitaevskii equation converges to the KdV equation in the
one-dimensional case, and to the KP-I equation in the multidimensional
case; see e.g. \cite{BGSS09,BGSS10,ChRo10}. More precisely, we may
resume the general presentation from \cite{ChRo10}: considering the
nonlinear Schr\"odinger equation
\begin{equation*} i{\partial}_t u + \Delta u = f\(|u|^2\)u,
\end{equation*} where the nonlinearity $f$ is such that $f(1)=0$ and
$f'(1)>0$, the KdV and KP-I equations appear in the limit $\varepsilon\to
0$, after the rescaling
\begin{equation*} t= c\varepsilon^3\tau,\quad X_1=\varepsilon(x_1-ct),\quad
X_j=\varepsilon^2 x_j, \ j\in \{2,\dots,d\}.
\end{equation*} Writing
\begin{equation*} u(t,X) = \(1+\varepsilon^2
A^\varepsilon(t,X)\)e^{i\varepsilon\varphi^\varepsilon(t,X)},
\end{equation*} and plugging this expression into the equation solved
by $u$, the formal limit $\varepsilon\to 0$ in the obtained system provides a
limit $A$ solution to the KdV equation if $d=1$, and to the KP-I
equation if $d\ge 2$. We emphasize that in the usual Gross-Pitaevskii
equation \eqref{eq:GP}, $f(y)=y-1$, in the logarithmic case
\eqref{logGP} with $\lambda=1$, $f(y)=\ln y$, so we always have
$f(1)=0$ and $f'(1)>0$. The general case considered in \cite{ChRo10}
is analyzed under the assumption $f\in \mathcal C^\infty (\R,\R)$,
which is not satisfied in the logarithmic case. This is probably not
the sharp assumption in order for the arguments presented in
\cite{ChRo10} to remain valid, but it is most likely that the proof
uses a $\mathcal C^0$ regularity on $[0,+\infty)$, like in
\cite{Ch12,ChMa17,Ma13} (these papers are devoted to the existence of
traveling waves in the same framework; $\mathcal C^1$
regularity of $f$ is required only near $1$). Again, the main difficulty in
the case of the logarithmic nonlinearity lies in the control of
vacuum, as $f$ fails to be continuous at the origin.
\bibliographystyle{abbrv}
| 5c5b54293a6116a1242429447b4101b7fce30b39 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Leptoquarks (LQs) are hypothetical particles that can convert quarks into leptons and vice versa with great interest in elementary particle physics. From the theoretical aspect, it has been predicted naturally by the Pati-Salam unification of quarks and leptons \cite{Pati:1973uk,Pati:1974yy} where leptoquark is first raised as well as many other grand unified theories \cite{Senjanovic:1982ex,Buchmuller:1986iq,Frampton:1991ay,Gershtein:1999gp,Fuentes-Martin:2020pww,King:2021jeo,FernandezNavarro:2022gst}. From the experimental side, the existence of leptoquarks is strongly indicated by lepton flavour universality violation (LFUV) in semi-leptonic $B$ decay \cite{LHCb:2019hip,Belle:2019rba,LHCb:2021trn,LHCb:2021lvy,Belle:2019rba,Angelescu:2021lln,Becirevic:2022tsj}. Besides LFUV, leptoquarks can also be related to a wide variety of phenomena beyond the standard model, including the muon $g-2$ \cite{Cheung:2001ip,ColuccioLeskow:2016dox,Crivellin:2019dwb,Athron:2021iuf,Du:2021zkq,Chen:2022hle}, the neutrino mass \cite{Mahanta:1999xd,Deppisch:2016qqd,Popov:2016fzr,Cai:2017wry,BhupalDev:2020zcy} and the $W$ boson mass \cite{Crivellin:2020ukd,DAlise:2022ypp,Athron:2022qpo,Cheung:2022zsb,Bhaskar:2022vgk,He:2022zjz}.
Despite the theoretical and experimental attraction from leptoquarks, they have not been found by any collider experiment so far. One of the possibilities to find leptoquark is through its connection with Higgs phenomenology \cite{Kolb:1997rb,Dorsner:2016wpm}. Generically, the scalar leptoquarks can couple to Higgs boson in the scalar potential. After electroweak symmetry breaking, the leptoquark-Higgs operator induces the couplings to the physical Higgs boson which can further affect loop-induced Higgs production and decay processes. Such effects can be probed with the Higgs signal strength measurements at colliders and thus are potential smoking guns for leptoquarks.
At the meantime, the Higgs portal allows leptoquarks to modify the EWPT in the early universe. It has been shown that first order EWPT can be induced by an additional singlet scalar field without any vacuum expectation value (VEV) \cite{Curtin:2014jma}. And the stochastic gravitational wave background produced during the cosmological phase transition can be potentially tested by detectors \cite{Caprini:2015zlo,Weir:2017wfa}. This provide us with a new possibility of testing scalar leptoquarks, using a similar approach to the singlet, from cosmic signals.
In this paper, we extend the study of first order EWPT induced by an extra singlet scalar to the case of scalar leptoquarks in $SU(2)$ singlet, doublet and triplet representations, and show how such leptoquarks can affect the EWPT through their coupling to the standard model Higgs boson. By computing the effective scalar potential, we find the range of Higgs portal where eligible first order EWPT can happen for different types of scalar leptoquark with a mass around TeV scale. Then we calculate the gravitational wave background produced during the first order EWPT induced by leptoquark for some benchmark cases and compare it with the detector sensitivities. We found that in some range of the parameter space, the first order EWPT induced by a leptoquark is able to produce gravitational wave signals which are strong enough to be detected.
The paper is organised as follows. In Sec.\ref{sec:1stPT}, we discuss the first order EWPT induced by leptoquark through the Higgs portal. We also show the constraint from Higgs physics to the parameter space. In Sec.\ref{sec:GW}, we show the gravitation wave signal produced during leptoquark-induced first order EWPT for benchmark points. Finally, we summarise and conclude in Sec.\ref{sec:con}.
\section{First order EWPT induced by scalar leptoquarks \label{sec:1stPT}}
In this section, we discuss how first order EWPT can be induced by leptoquarks. The scalar potential of the SM scalar doublet $H$ and an extra complex scalar leptoquark $S$ with a $SU(2)$ index $a$, corresponding to a singlet, doublet or triplet representation, can be written as
\begin{eqnarray}
V_0 = -\mu^2 |H|^2 + \lambda_H |H|^4 +\mu_S^2|S_a|^2 + \lambda_S |S_a|^4 + 2\lambda_{HS} |H|^2|S_a|^2
\end{eqnarray}
For simplicity, we only consider the minimal quartic interaction between Higgs and scalar leptoquark in the form of $|H|^2|S|^2$. Other forms of quartic interactions such as $|H^\dagger S|^2$ for $SU(2)$ doublet leptoquark and $H^\dagger (\sigma^i S_i)(\sigma^j S_j)^\dagger H$ for $SU(2)$ triplet leptoquark can lead to mass shifts between the $SU(2)$ components of leptoquarks after spontaneous symmetry breaking (SSB) as well as extra contributions to the thermal mass of the SM Higgs field. Focussing on the field $h$ in $H=(G^+,(h+iG^0)/\sqrt2)$ which becomes the SM Higgs boson after spontaneous symmetry breaking, the scalar potential becomes
\begin{eqnarray}
V_0 = -\frac{\mu^2}{2} h^2 + \frac{\lambda_H}{4} h^4 +\frac{\mu_S^2}{2}\left(s_{a,1}^2+s_{a,2}^2\right) + \frac{\lambda_S}{4}\left(s_{a,1}^2+s_{a,2}^2\right)^2 + \frac{\lambda_{HS}}{2} h^2\left(s_{a,1}^2+s_{a,2}^2\right)
\end{eqnarray}
where $S_a=(s_{a,1}+i\,s_{a,2})/\sqrt2$. As the leptoquark is typically heavier than the electroweak scale, we assume $\mu_S^2 > 0$ in this research. Then the leptoquark mass after SSB is $m_S^2 = \mu_S^2 + \lambda_{HS} v_0^2$ with $v_0$ the standard model Higgs VEV. At tree level, the phase transition is second order as the participation of $S$ does not vary the minimum of the scalar potential. However, by considering the finite temperature effective potential, the existence of a leptoquark modifies the minimum through the Higgs portal at loop order. In this study, we consider the scalar effective potential at one-loop level for simplicity, neglecting higher order effects \cite{Niemi:2021qvp} which may vary the transition strength by 20\%. We also neglect renormalisation group corrections which have a smaller effect \cite{Schicho:2022wty}.
At one-loop level, the effective scalar potential receives contribution from zero-temperature correction $\Delta V_0^{\rm 1-loop}$ (Coleman-Weinberg potential) and finite-temperature correction $\Delta V_T^{\rm 1-loop}$ \cite{Quiros:1999jp}
\begin{eqnarray}
V_{\rm eff} (h,T)= V_0 + \Delta V_0^{\rm 1-loop}(h) + \Delta V_T^{\rm 1-loop}(h,T)\,.\label{eq:Veff}
\end{eqnarray}
The one-loop zero-temperature correction reads
\begin{eqnarray}
\Delta V_0^{\rm 1-loop}(h) &=& \sum_{i\in b,f} \frac{n_i}{64\pi^2} \left[ m_i^4(h) \left(\ln\frac{m_i^2(h)}{m_i^2(v_0)} -\frac32\right) + 2m_i^2(h)m_i^2(v_0)\right]\,,
\end{eqnarray}
where $m_i^2=m_{0i}^2 +a_i h^2$ are the shifted masses with
\begin{eqnarray}
&&m_{0\{t, W, Z, h, G, S\}}^2 = \{0, 0, 0, -\mu^2, -\mu^2, \mu_S^2\}\,,\\
&&a_{\{t, W, Z, h, G, S\}} = \{\frac{y_t^2}{2}, \frac{g^2}{4},\frac{g^2+{g'}^2}{4}, 3\lambda_H, \lambda_H, \lambda_{HS}\}\,,\\
&&n_{\{t, W, Z, h, G, S\}} = \{-12, 6, 3, 1, 3, n_S\}\,.
\end{eqnarray}
The quantity $v_0$ is the SM Higgs VEV at zero temperature. The degree of freedom $n_S$ in the complex $SU(3)$ triplet $S$, depending on the SU(2) nature of the leptoquark, can be 6 for $SU(2)$ singlet, 12 for $SU(2)$ doublet or 18 for $SU(2)$ triplet.
The one-loop finite-temperature correction in Eq.\ref{eq:Veff} is
\begin{eqnarray}
\Delta V_T^{\rm 1-loop}(h,T) &=& \sum_{i\in b} \frac{n_i T^4}{2\pi^2} J_b \left(\frac{m_i^2}{T^2}\right) + \sum_{i\in f} \frac{n_i T^4}{2\pi^2} J_f \left(\frac{m_i^2}{T^2}\right)\,
\end{eqnarray}
where $b$ and $f$ stand for bosons and fermions and
\begin{eqnarray}
J_{b/f} \left(\frac{m_i^2}{T^2}\right) &=& \int_0^\infty dx x^2\ln\left[1 \mp e^{-\sqrt{x^2+m_i^2(h)/T^2}}\right]\,,
\end{eqnarray}
At high temperature $T \gtrsim m_i $, the functions $J_b$ and $J_f$ can be expressed approximately as
\begin{eqnarray}
J_b \left(\frac{m_i^2}{T^2}\right) &\simeq& -\frac{\pi^4}{45} + \frac{\pi^2}{12}\frac{m_i^2}{T^2} - \frac{\pi}{6}\frac{m_i^3}{T^3} - \frac{1}{32}\frac{m_i^4}{T^4}\left(\ln\frac{m_i^2}{T^2} - c_b\right) +... \label{eq:Jb_app}\\
J_f \left(\frac{m_i^2}{T^2}\right) &\simeq&\frac{7\pi^4}{360} - \frac{\pi^2}{24}\frac{m_i^2}{T^2} - \frac{1}{32}\frac{m_i^4}{T^4}\left(\ln\frac{m_i^2}{T^2} - c_f\right) +...\label{eq:Jf_app}
\end{eqnarray}
with $c_b \simeq 5.4$ and $c_f \simeq 2.6$. At low temperature $T < m_i $, $J_b$ is exponentially suppressed as its argument increases.
To maintain the perturbativity of gauge couplings at high temperature \cite{Linde:1980ts,Croon:2020cgk}, the so-called ring (daisy) diagrams should be resummed. There are two different methods wildly used for resummation. In the Parwani method \cite{Parwani:1991gq}, the shifted masses of bosons in the effective potential are replaced by the Debye masses $M_i^2(h,T)=m_i^2(h)+\Pi_i(T)$, where the self-energies $\Pi_i(T)$ are given by $\Pi_i (T)= b_i T^2$ with \cite{Curtin:2016urg}
\begin{eqnarray}
&&b_h = b_G = \frac{3g^2+{g'}^2}{16} + \frac{\lambda_H}{2} + \frac{y_t^2}{4}+ \frac{n_S \lambda_{HS}}{12} , \quad
b_{W} = b_{Z}(T) = \frac{11}{6} g^2 \,, \quad
b_{\gamma} = \frac{11}{6} {g'}^2 \,, \\
&&b_S =
\begin{dcases}
\frac{\lambda_{HS}}{3} + \frac{(n_S+2)\lambda_S}{12} + \frac34 g_3^2 + \frac14 Y^2 + \frac{{g'}^2}{16} & SU(2) \text{singlet} \,,\\
\frac{\lambda_{HS}}{3} + \frac{(n_S+2)\lambda_S}{12} + \frac34 g_3^2 + \frac14 Y^2 + \frac{3g^2+{g'}^2}{16} & SU(2) \text{ doublet and triplet.}
\end{dcases}
\nonumber\\
\end{eqnarray}
In the Arnold-Espinosa method \cite{Arnold:1992rz}, the replacement only happens in the mass cubic terms. In this paper, we adapt the Parwani method. While the leptoquark coupling $Y$ is typically smaller than unitarity \cite{Hiller:2017bzc}, the $SU(3)$ coupling can have significant contribution to the Debye mass of the leptoquarks. However, the contributions, not only from the $SU(3)$ coupling but also from other gauge couplings, play the same role as the self-interaction coupling $\lambda_S$ in phase transition and thus can be absorbed effectively by $\lambda_S$, turning it into $\tilde{\lambda}_S$. As $\lambda_S$ is unconstrained, relevant discussion is commonly avoided by fixing it to certain value \cite{Curtin:2014jma,Beniwal:2017eik}. Here, we consider the contributions to the thermal mass from the gauge couplings and leptoquark-fermion couplings as an effective contribution to the quartic coupling $\lambda_S$ and fix the resulting effective $\tilde{\lambda}_S$ to be 2.
When the phase transition happens at a low temperature, the effective potential can develop an imaginary part as the thermal masses of Goldstone bosons become negative. It has been pointed out in \cite{Weinberg:1987vp} that such an imaginary part remarks the decay rate of the quantum state minimising the Hamiltonian.
In the simplest case, the sufficient conditions for a eligible first order EWPT to occur are
\begin{enumerate}
\item The electroweak minimum is the true minimum at zero temperature $T=0$ and $h=0$ is a local maximum ($V''(0,0)<0$).
\item At the temperature $T_2$ that $h=0$ transfer from a local maximum to a local minimum, there is another non-zero local minimum.
\end{enumerate}
The first condition ensures that the phase transition is completed today. If $h=0$ is a local minimum at zero temperature, the phase transition can only happen through tunnelling and the probability is too low for the vacuum to transfer to the electroweak vacuum until today. The second condition ensures that there are two minima existing simultaneously during the phase transition. In a scenario satisfying both of the conditions, the two minima of the scalar potential are degenerate at a critical temperature $T_c$. The allowed parameter spaces for first-order phase transition to happen are shown as the coloured regions in Fig.\ref{fig:PT}. The strength of the transition can be estimated by the ratio of the non-zero VEV and the critical temperature, $v_c/T_c$, which is shown are the colour in Fig.\ref{fig:PT}. Above the coloured regions, the first order EWPT is not eligible as condition 1 is not satisfied; below the coloured regions, first order EWPT cannot happen because condition 2 is not satisfied.
In Fig.\ref{fig:PT_S} to Fig.\ref{fig:PT_T}, the required coupling for first order EWPT increases as the leptoquark becomes heavier in each $SU(2)$ representation of leptoquark. By comparing different panels and also by comparing the lines with different colours in Fig.\ref{fig:PT_all}, it can be figured out that the Higgs portal coupling required for first order EWPT becomes smaller as the dimension of the leptoquark $SU(2)$ representation increases. Empirical expressions of the interesting parameter spaces can be obtained when the leptoquark is heavy. For leptoquark mass above 1 TeV, the allowed Higgs portal for eligible first order EWPT to happen is roughly between $\{3.59,\,4.99\}\times (m_{S_1}/1\,\text{TeV})^{0.685}$ for singlet leptoquark, between $\{2.87,\,4.00\}\times (m_{S_2}/1\,\text{TeV})^{0.679}$ for doublet leptoquark and between $\{2.52,\,3.50\}\times (m_{S_3}/1\,\text{TeV})^{0.676}$ for triplet leptoquark.
\begin{figure}[t!]
\begin{center}
\subfigure[\,$SU(2)$ singlet scalar leptoquark\label{fig:PT_S}]{\includegraphics[width=0.49\textwidth]{Figures/PT_S.PNG}}
\subfigure[\,$SU(2)$ doublet scalar leptoquark\label{fig:PT_D}]{\includegraphics[width=0.49\textwidth]{Figures/PT_D.PNG}}
\subfigure[\,$SU(2)$ triplet scalar leptoquark\label{fig:PT_T}]{\includegraphics[width=0.49\textwidth]{Figures/PT_T.PNG}}
\subfigure[\,$SU(2)$ singlet, doublet and triplet scalar leptoquarks\label{fig:PT_all}]{\includegraphics[width=0.49\textwidth]{Figures/PT_all.PNG}}
\includegraphics[width=0.49\textwidth]{Figures/Colorbar.PNG}
\caption{\label{fig:PT} Allowed parameter space for first-order phase transition induced by different types of scalar leptoquark. }
\end{center}
\end{figure}
A more complicated case can occur when the scalar potential develops two non-zero minima simultaneously after the temperature drops below $T_2$. In such a case, the scalar configuration transfer to the nearest non-zero minimum continuously through second order phase transition and tunnel to the larger non-zero minimum through second order phase transition. The regions where such cases happen are marked as green in Fig.\ref{fig:PT_D} and Fig.\ref{fig:PT_T}. However, as the leptoquark is typically above 1 TeV, such regions are not of interest in this study.
\subsection{Constraints on the Higgs portal coupling}
The new interaction between a scalar leptoquark and the Higgs doublet can affect the Higgs boson production and decay processes. The discrepancy between SM prediction and experimental measurement is commonly characterised by the $\varkappa$-factor, defined as $\varkappa_i=\sqrt{\Gamma_i^{\rm exp}/\Gamma_i^{\rm SM}}$ \cite{TheATLASandCMSCollaborations:2015bln,CMS:2015kwa}. The loop-induced contribution from leptoquark to the Higgs boson decay process into photons and the gluon-gluon production of Higgs boson are given by \cite{Dorsner:2016wpm}
\begin{eqnarray}
\varkappa_g &=& 1 + 0.24 \, \frac{\lambda_{HS}\,v^2}{m_S^2} N_S\\
\varkappa_\gamma &=& 1 - 0.052 \, \frac{\lambda_{HS}\,v^2}{m_S^2} N_c \sum_i Q_i^2
\end{eqnarray}
where the sum is taken over all $SU(2)$ components of the leptoquark and $Q_i$ is the electric charge of the $i$th component. $N_S$ is the number of the leptoquark $SU(2)$ components. The experimental measurements by the ATLAS collaboration are $\varkappa_g = 1.01^{+0.11}_{-0.09}$ and $\varkappa_\gamma = 1.02^{+0.08}_{-0.07}$ \cite{ATLAS:2022tnm}. Similar contribution appears in the decay channel of Higgs into a $Z$ boson and a photon as well, in the form of \cite{Dorsner:2016wpm}
\begin{eqnarray}
\varkappa_{Z\gamma} &=& 1 + 0.036\, \frac{\lambda_{HS}\,v^2}{m_S^2} N_c \sum_i Q_i \left(I_i^W - 0.23 Q_i\right)
\end{eqnarray}
where $I_i^W$ is the value of the weak isospin of the leptoquark. The value of $\varkappa_{Z\gamma}$ measured by CMS collaboration is $1.65^{+0.34}_{-0.37}$ \cite{CMS:2022dwd}. Despite abundant collider phenomena caused by the Higgs portal to leptoquarks, none of the observables can constrain the portal coupling restrictedly. When multiple leptoquarks appear in a model, the contributions from different types of leptoquarks can have opposite contributions to the $\varkappa$ parameters above. In order to visualise the effects of these observables, we consider the collider constraints under the assumption of a single leptoquark multiplet and show the maximal values of the Higgs portal allowed by $h\rightarrow\gamma\gamma$ and $gg\rightarrow h$ as the dashed and dotted lines in Fig.\ref{fig:PT_S} to Fig.\ref{fig:PT_T}. While the $gg\rightarrow h$ cross section is affected by the $SU(2)$ representation of the leptoquark, the $h\rightarrow\gamma\gamma$ cross section depends on the electric charge. For scalar leptoquark, assuming direct interaction to SM fermions, there are two different possible assignments of hypercharge for $SU(2)$ singlet and doublet and one assignment for $SU(2)$ triplet \cite{ParticleDataGroup:2020ssz}: $4/3$ or $1/3$ for singlet, $7/6$ or $1/6$ for doublet and $1/3$ for triplet. Although those constraints are currently weak, they are expected to be improved foreseeably by future experiments like HL-LHC \cite{CMS:2018qgz}, FCC \cite{dEnterria:2017dac,FCC:2018evy}, ILC \cite{Bambade:2019fyw} and CEPC \cite{An:2018dwb,Ruan:2021gap}. Moreover, the Higgs portal coupling also affects flavour violating processes like the $h\rightarrow\mu\tau$ or $\tau\rightarrow\mu\gamma$ decay which can be tested by precious measurements at colliders \cite{Dorsner:2016wpm,Crivellin:2020mjs}.
\section{Gravitational wave signals \label{sec:GW}}
During a first-order phase transition, the scalar field configuration transfer from zero vacuum to non-zero vacuum locally in the form of bubbles through tunnelling. The scalar bubbles can then move, collide and expand. Sound waves and magnetohydrodynamic turbulence can be produced after the collision of bubbles. The gravitational wave can be produced through three different mechanisms \cite{Caprini:2015zlo,Weir:2017wfa}: {\bf collision} of the scalar bubbles, overlap of the {\bf sound wave} in the plasma and the fluid {\bf turbulence}. The total gravitational wave spectrum is the sum of the three contributions
\begin{eqnarray}
\Omega_{\rm tot}(f)h^2 = \Omega_{\rm coll}(f)h^2 + \Omega_{\rm sw}(f)h^2 + \Omega_{\rm turb}(f)h^2\,.
\end{eqnarray}
All three contributions depend on the phase transition dynamics which is described by four key parameters: the wall velocity $v_w$, the inverse phase transition duration $\beta/\mathcal{H}_*$, the phase transition strength $\alpha_{T_*}$ and the transition temperature $T_*$. After these parameters are determined, the gravitational wave spectrum can be computed using results from numerical simulations.
The crucial step in computing these key parameters is to compute the Euclidean action. To find the Euclidean action which is defined as the spacial integration of the effective Lagrangian, a solution of the Euclidean equation of motion is required, which is generally not solvable analytically. For further details see Appendix \ref{app:GWinPT}. A common treatment for particles of electroweak scale or below is to make an approximation using Eq.\eqref{eq:Jb_app} and Eq.\eqref{eq:Jf_app} after which the effective potential can be simplified into a quartic function of the scalar field and a semi-analytical bounce solution can be derived \cite{Dine:1992wr,DiBari:2021dri}. However, as the leptoquark is typically above TeV scale \cite{CMS:2020gru,CMS:2020wzx,ATLAS:2021oiz,ATLAS:2022fho}, the one-loop finite-temperature correction from leptoquark is exponentially suppressed and thus negligible. On the other hand, no eligible expansion can be made to the one-loop zero-temperature correction from leptoquark in the parameter space of interest. Therefore we choose to solve the Euclidean equation of motion numerically in this work.
\begin{figure}[t!]
\begin{center}
\subfigure[\,$SU(2)$ singlet leptoquark $m_S=1$ TeV, $v_*/T_*=3.73$]{\includegraphics[width=0.43\textwidth]{Figures/GW_S_1.pdf}}
\subfigure[\,$SU(2)$ singlet leptoquark $m_S=5$ TeV, $v_*/T_*=1.74$]{\includegraphics[width=0.43\textwidth]{Figures/GW_S_2.pdf}}
\subfigure[\,$SU(2)$ doublet leptoquark $m_S=1$ TeV, $v_*/T_*=3.82$]{\includegraphics[width=0.43\textwidth]{Figures/GW_D_1.pdf}}
\subfigure[\,$SU(2)$ doublet leptoquark $m_S=5$ TeV, $v_*/T_*=3.91$]{\includegraphics[width=0.43\textwidth]{Figures/GW_D_2.pdf}}
\subfigure[\,$SU(2)$ triplet leptoquark $m_S=1$ TeV, $v_*/T_*=3.42$]{\includegraphics[width=0.43\textwidth]{Figures/GW_T_1.pdf}}
\subfigure[\,$SU(2)$ triplet leptoquark $m_S=5$ TeV, $v_*/T_*=3.64$]{\includegraphics[width=0.43\textwidth]{Figures/GW_T_2.pdf}}
\caption{\label{fig:fix_rep} Gravitational wave signals for difference benchmark cases. The left panels show the strongest gravitational wave signals from first order EWPT induced by 1 TeV leptoquarks for $SU(2)$ singlet, doublet and triplet from top to bottom. The right panels show similar results for 5 TeV leptoquarks. }
\end{center}
\end{figure}
In Fig.\ref{fig:fix_rep}, we show the gravitational wave produced from first order EWPT for six benchmark cases. From top to bottom, the benchmark cases in each row are chosen for $SU(2)$ singlet, doublet and triplet leptoquark. For each $SU(2)$ representation, the strongest gravitational wave signals that leptoquark-induced first order EWPT can produce when the leptoquark mass is 1 TeV and 5 TeV are presented on the left and right panels respectively. In all the cases, the gravitational wave signals can be detected by BBO \cite{Corbin:2005ny}, DECIGO \cite{Kawamura:2019jqt,Kawamura:2020pcg}, while LISA \cite{Caprini:2015zlo} and $\mu$Ares \cite{Sesana:2019vho} can potentially find the signal in Benchmark Point 1 for singlet leptoquark. We also show the gravitational waves produced by different sources during the phase transition independently. In most of the frequency range, the gravitational wave is dominantly produced by the magnetohydrodynamic (MHD) turbulence.
\begin{figure}[t!]
\begin{center}
\subfigure[]{\includegraphics[width=0.45\textwidth]{Figures/GWMax_vT.pdf}}
\subfigure[]{\includegraphics[width=0.45\textwidth]{Figures/Coupling_GW.pdf}}
\caption{\label{fig:peak_strength} Left panel: Maximal strength of gravitational wave produced as a function of transition strength $v_*/T_*$. Right panel: Maximal strength of gravitational wave produced by first order EWPT induced by different type of leptoquarks of 1 TeV as a function of the Higgs portal coupling. }
\end{center}
\end{figure}
By comparing panels, it can be observed that only the shape of the gravitational wave spectrum for 5 TeV $SU(2)$ singlet leptoquarks shows a significant difference from the others. In fact, the result follows from the fact that the gravitational wave produced from first order EWPT relies on the strength of the transition. To illustrate the relation more explicitly, we show the dependence of gravitational wave signal peak values on the strength of the phase transition in Fig.\ref{fig:peak_strength}. Here, instead of $v_c/T_c$ in the previous section, the phase transition strength is evaluated by the ration of the non-zero minimum of the scalar potential and temperature when the phase transition happens, i.e. when the probability of bubble nucleation is significant. The temperature $T_*$ is defined by the temperature when one bubble is nucleated per unit volume per unit time and the non-zero VEV at $T_*$ is denoted as $v_*$. We find that the gravitational wave is testable when the phase transition strength is roughly larger than $1.34$, corresponding to the Higgs portal roughly larger than $3.95\times (m_{S_1}/1\,\text{TeV})^{0.685}$ in the singlet case, $3.17\times (m_{S_2}/1\,\text{TeV})^{0.679}$ in the doublet case and $2.79\times (m_{S_3}/1\,\text{TeV})^{0.676}$ in the triplet case. In the case with a 5 TeV $SU(2)$ singlet leptoquark, the Higgs portal is constrained by its perturbativity limit and as a consequence, the strongest gravitational wave signal that eligible first order EWPT can produce is less than the other cases.
\begin{figure}[t!]
\begin{center}
\subfigure[\,$SU(2)$ singlet leptoquark $m_S=1$ TeV, $\lambda_{HS}=3.85$]{\includegraphics[width=0.43\textwidth]{Figures/GW_S_3.pdf}}
\subfigure[\,$SU(2)$ doublet leptoquark $m_S=1$ TeV, $\lambda_{HS}=3.85$]{\includegraphics[width=0.43\textwidth]{Figures/GW_D_1.pdf}}
\subfigure[\,$SU(2)$ doublet leptoquark $m_S=1$ TeV, $\lambda_{HS}=3.35$]{\includegraphics[width=0.43\textwidth]{Figures/GW_D_3.pdf}}
\subfigure[\,$SU(2)$ triplet leptoquark $m_S=1$ TeV, $\lambda_{HS}=3.35$]{\includegraphics[width=0.43\textwidth]{Figures/GW_T_1.pdf}}
\caption{\label{fig:fix_BP} Gravitational wave signals for the same benchmark cases in different $SU(2)$ representations. The upper panels show the gravitational wave signals for a benchmark case when the leptoquark is $SU(2)$ singlet and doublet. The lower panels show the gravitational wave signals for another benchmark case when the leptoquark is $SU(2)$ doublet and triplet.}
\end{center}
\end{figure}
For the same benchmark point, the gravitational wave produced during first order EWPT induced by leptoquark with a smaller dimension is stronger. In Fig.\ref{fig:fix_BP}, we choose the benchmark points 3 and 5 in Fig.\ref{fig:PT} and show the gravitational wave produced during singlet- and doublet-induced first order EWPT for the former case and the gravitational wave produced during doublet- and triplet-induced first order EWPT for the later one. It is clear that for the same coupling, the first order EWPT induced by the $SU(2)$ multiplet with a higher dimension produces stronger gravitational waves. Supposing the Higgs portal is measured to be in the region where first-order phase transition appears by future collider experiments, the gravitational waves detection provides an alternative method to further test the Higgs portal as well as determine the $SU(2)$ representation of leptoquarks.
\section{Conclusion\label{sec:con}}
In this paper, we have explored the possibility that first order EWPT induced by the coupling between a scalar leptoquark and the SM Higgs boson produces detectable gravitational wave signals. We have considered different $SU(2)$ representations of the scalar leptoquark, including singlet, doublet and triplet. Despite the lack of VEV for leptoquark itself, a first order EWPT can be induced due to the 1-loop order effects. In general, with first order EWPTs, gravitational waves can be produced by multiple processes in the dynamical evolution of the scalar bubbles. The resulting gravitational waves form a stochastic background that can be probed by gravitational wave detectors.
We have calculated the effective potential of the SM Higgs field in the presence of a scalar leptoquark,
including tree level and 1-loop level contributions as well as the resummation over the ring/daisy diagrams. By applying the conditions for first order EWPT, we have found that the leptoquark can induce a first order EWPT in the parameter space allowed by collider constraints and can be tested by future Higgs precision experiments. Enhanced by the internal degree of freedom of the particular leptoquark, we found that the leptoquark in the $SU(2)$ representation with a higher dimension requires smaller coupling in order to trigger a first order EWPT.
We have followed the standard procedure to compute the gravitational wave spectrum during first order EWPTs. It turns out that the gravitational wave spectrum is mainly determined by the strength of the phase transition characterised by the ratio of the non-zero VEV and the temperature at the time that the transition happens. However, due to the difference in internal degrees of freedom, the strengths of first order EWPTs induced by leptoquarks with the same masses and Higgs portal couplings but different $SU(2)$ nature are different. Since the gravitational wave signals differ, this provides a possibility to determine the $SU(2)$ representation of the leptoquarks through the observations of gravitational wave in particular regions of parameter space.
\section*{Acknowledgments}
BF acknowledges the Chinese Scholarship Council (CSC) Grant No.\ 201809210011 under agreements [2018]3101 and [2019]536. SFK acknowledges the STFC Consolidated Grant ST/L000296/1 and the European Union's Horizon 2020 Research and Innovation programme under Marie Sklodowska-Curie grant agreement HIDDeN European ITN project (H2020-MSCA-ITN-2019//860881-HIDDeN).
| 750554fa58162c02d8beb392cbf3d19d63f08570 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:LJKO2intro}
In this paper we are concerned with the construction of second-order in time discretizations for the following system of PDEs,
describing the time evolution of a density $\rhoe: [0,T]\times \Omega \rightarrow \mathbb{R}_+ $ on a convex bounded domain $\Omega$ and over the time interval $[0,T]$:
\begin{equation}\label{eq:model}
\partial_t \rhoe - \mathrm{div} \left(\rhoe \nabla \frac{\delta \mathcal E}{\delta \rho} (\rhoe) \right) = 0 \quad \text{on } (0,T)\times \Omega \,,
\end{equation}
with initial and boundary conditions:
\begin{equation}\label{eq:bcs}
\rhoe(0,\cdot) = \rho_0\,, \quad \rhoe \nabla \frac{\delta \mathcal E}{\delta \rho} (\rhoe) \cdot n_{\partial \Omega} = 0 \quad \text{on } (0,T) \times \partial \Omega\,,
\end{equation}
for a given initial density $\rho_0$, and where $n_{\partial \Omega}$ denotes the outward pointing normal to $\partial \Omega$. In equation \eqref{eq:model}, $\mathcal{E}:L^1(\Omega; \mathbb{R}_+) \rightarrow \mathbb{R}$ is a functional of the density and describes the energy of the system. Different choices for $\mathcal{E}$ yield different equations modelling a wide range of phenomena. Typical examples are the Fokker-Planck equation \cite{jordan1998fokkerplanck}, the porous medium equation \cite{otto2001geometry} or the Keller-Segel equation \cite{blanchet2013KellerSegel}, but also more complex cases such as multiphase flows \cite{cances2017multiphase,laurenccot2013gradient,cances2019two} or crowd motion models \cite{santambrogio2018crowd} can be considered.
Since the density satisfies the continuity equation with zero boundary flux, its total mass is conserved. Moreover, the energy decreases along the evolution:
\[
\frac{\mathrm{d}}{\mathrm{d} t} \mathcal{E}(\rhoe(t,\cdot) ) \leq 0\,.
\]
This behaviour is a consequence of the fact that system \eqref{eq:model}, under suitable assumptions on the energy, can be interpreted as a gradient flow in the space of probability measures $\mc{P}(\Omega)$ equipped with the Wasserstein distance $W_2$. This interpretation is well-known since the pioneering work of Jordan, Kinderlehrer, and Otto \cite{jordan1998fokkerplanck}, who showed that one recovers the Fokker-Planck equation when following the steepest descent curve of an entropy functional with respect to the Wasserstein metric. Such result is best explained in the time-discrete setting: given a uniform decomposition $0 = t_0 <t_1 < \ldots < t_N= T$ of the interval $[0,T]$ with time step $\tau \coloneqq t_{n+1} -t_n$, consider the sequence $(\rho_n)_n$ defined for $1\leq n\leq N$ by
\begin{equation}\label{eq:jko}
\rho_{n} = \argmin_{\rho \in \mathcal{P}(\Omega)} \frac{W^2_2(\rho,\rho_{n-1})}{2\tau} + \mc{E}(\rho)\,,
\end{equation}
where the energy is given by
\begin{equation}\label{eq:energyfp}
\mathcal{E}(\rho) = \int_\Omega V\rho + \rho \log \rho \,,
\end{equation}
with $V:\Omega\rightarrow \mathbb{R}$ being a Lipschitz function, if $\rho$ is absolutely continuous with respect to the Lebesgue measure and $+\infty$ otherwise. Then, one can show that the discrete curve $t\mapsto \tilde{\rhoe}(t)$, defined by $\tilde{\rhoe}(t,\cdot) = \rho_{n-1}$ for $t\in(t_{n-1}, t_{n}]$ and $1 \leq n \leq N$, converges uniformly in the $W_2$ distance to the unique solution of the Fokker-Planck equation
\begin{equation}\label{eq:fp}
\partial_t \rhoe - \mathrm{div} (\rhoe \nabla V) - \Delta \rhoe = 0 \quad \text{on } (0,T)\times \Omega \,,
\end{equation}
satisfying \eqref{eq:bcs}.
The numerical scheme defined in equation \eqref{eq:jko} is known as JKO scheme and it allows one to interpret many different models as Wasserstein gradient flows. It also provides a convenient framework both for the analysis of such models (e.g., to prove {existence of solutions or} exponential convergence towards steady states) \cite{ambrosio2008gradient,santambrogio2017euclidean}, and for the design of numerical discretizations \cite{benamou2016augmented,carrillo2019primal,cances2020LJKO,leclerc2020lagrangian,carlier2017convergence}. In fact, reproducing the JKO scheme at the discrete level generally implies energy stability even in very degenerate settings. Moreover in the case of convex energies one can use robust convex optimization tools that, e.g., can easily take into account the positivity constraint on the density or even other type of strong constraints (as in the case of incompressible immiscible multiphase flows in porous media, see Section \ref{ssec:multiphase}).
Since the JKO scheme is a variational version of the implicit Euler scheme, it is an order one method. Recently, several higher-order alternatives to the JKO scheme have been proposed, but it is not trivial to translate them into a fully-discrete setting (see \cite{Matthes2019bdf2,Legendre2017VIM}, and Section \ref{ssec:vimbdf2} below for a detailed description of such approaches). In fact, to the best of our knowledge, there exist no viable fully-discrete approach able to compute with second order accuracy general Wasserstein gradient flows while preserving (to some extent) the underlying variational structure.
In this paper we will accomplish this by reformulating the classical multistep scheme based on the Backward Differentiation Formula of order two (BDF2) as the composition of two inner steps: a geodesic extrapolation step, and a standard JKO step. We refer to the resulting scheme as Extrapolated Variational BDF2 (\scheme) scheme. We prove that, as for the JKO scheme, the discrete solutions generated by the EVBDF2 scheme with the energy \eqref{eq:energyfp} converge to weak solutions of the Fokker-Planck equation. Importantly, we also propose a simple and efficient (space-time discrete) implementation which realizes numerically the desired second-order accuracy.
\subsection{Description of the BDF2 approach and main results} In the Euclidean setting, the gradient flow associated to a smooth real-valued convex function $F:\mathbb{R}^d\rightarrow{\mathbb R}$ and a starting point $x_0\in\mathbb{R}^d$, is the unique solution to the Cauchy problem
\begin{equation}\label{eq:gradF}
\left \{
\begin{array}{ll}
x'(t) = - \nabla F(x(t))\,, & \quad \forall\, t>0\,,\\
x(0) = x_0\,. \\
\end{array}
\right.
\end{equation}
The BDF2 scheme applied to such a system, with time step $\tau>0$, can be written as follows: given $x_0,x_1 \in \mathbb{R}^d$, for $n\geq 2$ find $x_n\in \mathbb{R}^d$ satisfying
\begin{equation}\label{eq:BDF2_eu_OC}
\frac{3}{2\tau} \Big( x_n- \frac{4}{3} x_{n-1}+\frac{1}{3} x_{n-2} \Big) = -\nabla F(x_n)\,.
\end{equation}
This can be interpreted as an implicit Euler step, with starting point
\[
x^\alpha_{n-1} \coloneqq x_{n-2} + \alpha(x_{n-1}- x_{n-2}) = x_{n-1} + \beta(x_{n-1}- x_{n-2}) \,,
\] where $\alpha =4/3$ and $\beta = \alpha -1=1/3$, and with time step $(1-\beta) \tau = 2\tau/3$. In turn, $x^\alpha_{n-1}$ coincides with the Euclidean extrapolation at time $\alpha$, from $x_{n-2}$ (at time $0$) to $x_{n-1}$ (at time $1$), with respect to a fictitious time variable (see Figure \ref{fig:time} for a graphical representation of the time intervals involved in the scheme).
\begin{figure}
\begin{overpic}[scale=.8]{time.pdf}
\put(214,42){$(1-\beta)\tau$}
\put(182,42){$\beta\tau$}
\put(125,42){$\tau$}
\put(70,11){$t_{n-2}$}
\put(163,11){$t_{n-1}$}
\put(259,11){$t_n$}
\put(140,-6){$\alpha \tau$}
\end{overpic}
\caption{A graphical representation of the time intervals involved in the definition of the EVBDF2 scheme.} \label{fig:time}
\end{figure}
In order to define a counterpart to the BDF2 scheme \eqref{eq:BDF2_eu_OC} for Wasserstein gradient flows, one needs to replace the Euclidean extrapolation at time $\alpha>1$ by an analogous operation in the space of probability measures equipped with the $W_2$ metric. In this paper, we will represent such an operation by a map $\mathsf{E}_\alpha:\mc{P}_2(\mathbb{R}^d)\times \mc{P}_2(\mathbb{R}^d) \rightarrow \mc{P}_2(\mathbb{R}^d)$ (where $\mc{P}_2(\mathbb{R}^d)$ is the set of probability measures on $\mathbb{R}^d$ with finite second moments), which we will refer to as an $\alpha$-extrapolation operator. Given such a map, we define the \scheme\ scheme as follows: given $\rho_0,\rho_1 \in \mathcal{P}(\Omega)$, for $n\geq 2$ find $\rho_n \in \mc{P}(\Omega)$ satisfying
\begin{equation}\label{eq:bdf2metric}
\displaystyle \rho_{n} \in \argmin_{\rho \in \mathcal{P}(\Omega)} \frac{W^2_2(\rho,\rho^\alpha_{n-1})}{2(1-\beta)\tau} +\mathcal{E}(\rho)\,,
\quad \rho_{n-1}^\alpha = \mathsf{E}_\alpha (\rho_{n-2},\rho_{n-1})\,,
\end{equation}
where here $\mc{E}:\mc{P}(\Omega) \rightarrow \mathbb{R}$ is defined on the whole space $\mc{P}(\Omega)$.
The extrapolation operator $\mathsf{E}_\alpha$ plays a crucial role in the scheme, but it is not trivial to propose an appropriate definition for it due to the structure of $W_2$ geodesics on $\mc{P}_2(\mathbb{R}^d)$.
To clarify this, recall that
a (globally length-minimizing) geodesic with respect to the $W_2$ metric is a curve $\omega: [t_0,t_1] \rightarrow \mc{P}_2(\mathbb{R}^d)$ such that
\begin{equation}\label{eq:geodesic}
W_2(\omega(s_0),\omega(s_1)) = \frac{|s_1-s_0|}{|t_1-t_0|} W_2(\omega(t_0),\omega(t_1))\,,
\end{equation}
for all $s_0,s_1 \in (t_0,t_1)$. Given two measures $\mu_0,\mu_1 \in\mc{P}_2(\mathbb{R}^d)$ there always exists a geodesic connecting the two.
Furthermore, due to Brenier's theorem, supposing that $\mu_0$ is absolutely continuous with respect to the Lebesgue measure, there exists a unique geodesic $\omega:[0,1]\rightarrow \mathcal{P}_2(\mathbb{R}^d)$ such that $\omega(0) = \mu_0$ and $\omega(1)=\mu_1$, and this has a very
simple expression:
\begin{equation}\label{eq:brenierth}
\omega(t) = ((1-t) \mathrm{Id} + t \nabla u)_\# \mu_0\,,
\end{equation}
where $\mathrm{Id}$ is the identity map on $\Omega$ and $u:\Omega \rightarrow \mathbb{R}$ is a convex function. This means that particles travel on straight lines along the interpolation, without colliding into each other. However, for a given $\alpha>1$, there may exist no geodesic defined on $[0,\alpha]$ that coincide on $[0,1]$ with $\omega$. This is because following their straight trajectories particles may collide immediately after time $t=1$, even if both $\mu_0$ and $\mu_1$ have smooth and strictly positive densities. This means that one cannot use such geodesic extensions to define the extrapolation operator $\mathsf{E}_\alpha$ in a unique way. Therefore, instead of focusing on a particular definition, we only require a uniform stability bound on the extrapolation which we will need to prove the convergence of the scheme. In particular, we will focus on extrapolation operators that are dissipative in the following sense:
\begin{definition}[Dissipative extrapolations] \label{def:dissipation} An extrapolation operator $\mathsf{E}_\alpha$ is $\theta$-dissipative if it satisfies
\begin{equation}\label{eq:dissipation}
W_2(\mu_1, \mathsf{E}_\alpha(\mu_0,\mu_1)) \leq \theta W_2(\mu_0,\mu_1)\,,
\end{equation}
for any $\mu_0,\mu_1 \in \mc{P}_2(\mathbb{R}^d)$ and for a constant $\theta\geq 0$.
\end{definition}
Note that by equation \eqref{eq:geodesic}, if the extrapolation is consistent with the geodesic extension when this exists, then we have $\beta =\alpha-1 \leq \theta$.
Upon adding a further consistency assumption on the extrapolation (see equation \eqref{eq:consistency} below) we can establish the following convergence result:
\begin{theorem}\label{th:convergencefp} Let $\rho_0\in\mc{P}(\Omega)$ and $\mc{E}$ given by \eqref{eq:energyfp}. For any given $N\geq 1$, let $(\rho_n)_{n=0}^N$ be the discrete solution defined by the scheme \eqref{eq:bdf2metric} for given $\rho_1 \in \mc{P}(\Omega)$ (dependent on $N$), with time step $\tau = T/N$, and with $\mathsf{E}_\alpha$ being a $\theta$-dissipative extrapolation operator with $0\leq \beta=\alpha-1<1$ and $\theta<1/2$, and such that for all $\mu_0,\mu_1 \in \mc{P}(\Omega)$ and $\varphi \in C^\infty_c(\mathbb{R}^d)$,
\begin{equation}\label{eq:consistency}
\left| \int \varphi\, ( \mathsf{E}_\alpha(\mu_0,\mu_1) - \alpha \mu_1 + \beta \mu_0) \right|\leq C_\varphi W^2_2(\mu_0,\mu_1)\,,
\end{equation}
where $C_\varphi>0$ only depends on $\alpha$, $\varphi$ and $\Omega$. Suppose that $W^2_2(\rho_0,\rho_1) \leq C \tau$, for a constant $C>0$ independent of $\tau$, and that $\mc{E}(\rho_1)\leq \mc{E}(\rho_0)$.
Then, the curve $t\mapsto \tilde{\rho}_\tau(t)$ defined by $\tilde{\rho}_\tau(t) \coloneqq \rho_{n-1}$ for all $t\in(t_{n-1}, t_{n}]$ and $1 \leq n \leq N$, converges as $N\rightarrow \infty$, uniformly in the $W_2$ distance, to a distributional solution to the Fokker-Panck equation on $[0,T]\times \Omega$ and initial conditions given by $\rho_0$.
\end{theorem}
Of course, in order to achieve second order accuracy, we must set $\alpha =4/3$ and require in addition that, if there exists a geodesic $\omega:[0,\alpha]\rightarrow\mc{P}(\Omega)$ such that $\omega|_{[0,1]}$ is a geodesic from $\mu_0$ to $\mu_1$, then $\mathsf{E}_\alpha(\mu_0,\mu_1)$ must coincide with $\omega(\alpha)$. Importantly, we will show that there exist several different ways to define such an operator, providing therefore different convergent approaches.
One of such approaches, which enjoys a particularly rich structure, consists in reproducing the variational charaterization of the linear extrapolation in the metric setting. Given two points $x_0, x_1 \in \mathbb{R}^d$, the Euclidean extrapolation at time $\alpha$ from $x_0$ to $x_1$ is the point $x_\alpha = \alpha x_1 - \beta x_0$ with $\beta =\alpha-1$. This can be obtained as the unique solution to
\begin{equation}\label{eq:euclideanvar}
x_\alpha = \underset{x\in \mathbb{R}^d}{\mathrm{argmin}} ~ \alpha |x-x_1|^2 - \beta |x-x_0|^2\,.
\end{equation}
Similarly, we define the metric extrapolation in the Wasserstein space as follows:
\begin{equation}\label{eq:metricextra}
\mathsf{E}_\alpha(\mu_0,\mu_1) \coloneqq \underset{\rho \in \mathcal{P}_2(\mathbb{R}^d)}{\mathrm{argmin}} ~ \alpha W_2^2(\rho,\mu_1) - \beta W_2^2(\rho,\mu_0)\,.
\end{equation}
Problem \eqref{eq:metricextra} is not a convex optimization problem in the classical sense. To see this, consider the following simple counterexample. In dimension $d=1$, take \[
\mu_0 = (\delta_{-1} + \delta_{1})/2, \quad \mu_1 = \delta_0, \quad \nu_0=\delta_{-1}, \quad \nu_1=\delta_{1}. \]Along the interpolation $\nu(t) = (1-t)\nu_0+t\nu_1$, the first term of the functional in \eqref{eq:metricextra} is constant whereas the second one is concave.
Nonetheless, we will show that problem \eqref{eq:metricextra} always admits a unique solution (see Proposition \ref{prop:existence}) and it also satisfies the assumptions in Theorem \ref{th:convergencefp}. Furthermore, exploiting the variational formulation of the metric extrapolation \eqref{eq:metricextra}, we can prove a more general convergence result using the Evolution Variational Inequality (EVI) characterization of gradient flows in metric spaces. More precisely, we prove the following result:
\begin{theorem}\label{th:convergenceevi}
Let $\rho_0\in\mc{P}(\Omega)$ and $\mc{E}:\mc{P}(\Omega) \rightarrow \mathbb{R}$ being a $\lambda$-convex energy in the generalized geodesic sense, for $\lambda\in{\mathbb R}_+$. For any given $N\geq 1$, let $(\rho_n)_{n=0}^N$ be the discrete solution defined by the scheme \eqref{eq:bdf2metric} for given $\rho_1 \in \mc{P}(\Omega)$ (dependent on $N$), with time step $\tau = T/N$, and with $\mathsf{E}_\alpha$ being the metric extrapolation \eqref{eq:metricextra} with $\beta = \alpha -1$. Suppose that $W^2_2(\rho_0,\rho_1) \leq C \tau$, for a constant $C>0$ independent of $\tau$, and that $\mc{E}(\rho_1)\leq \mc{E}(\rho_0)$. Then, the curve $t\mapsto \tilde{\rho}_\tau(t)$ defined by $\tilde{\rho}_\tau(t) \coloneqq \rho_{n-1}$ for $t\in(t_{n-1}, t_{n}]$ and $1 \leq n \leq N$, converges as $N \rightarrow \infty$, uniformly in the $W_2$ distance, to the unique absolutely continuous curve $\rhoe:[0,T]\rightarrow\mc{P}(\Omega)$ satisfying $\rhoe(0) = \rho_0$ and such that
for any $\nu\in\Pc(\Om)$ it holds
\begin{equation*}
\frac{\text{d}}{\text{d} t} \frac{1}{2} {W}_2^2(\varrho(t),\nu) \le \mathcal{E}(\nu)-\mathcal{E}(\varrho(t)) - \frac{\lambda}{2} {W}_2^2(\varrho(t),\nu), \quad \forall t\in(0,T) \,.
\end{equation*}
\end{theorem}
\noindent
Remarkably, problem \eqref{eq:metricextra} admits a convex dual formulation, see Remark \ref{rem:metricextra_dual}.
\subsection{Relation with previous works and numerical implementation issues}\label{ssec:vimbdf2}
Going back to the discretization of system \eqref{eq:gradF}, each step of the BDF2 scheme \eqref{eq:BDF2_eu_OC} can also be obtained as the optimality conditions of the following problem:
\begin{equation}\label{eq:BDF2_eu}
x_n = \argmin_{x\in\mathbb{R}^d} \alpha \frac{\left|x-x_{n-1}\right|^2}{2(1-\beta)\tau} - \beta \frac{\left|x-x_{n-2}\right|^2}{2(1-\beta)\tau} + F(x) \,.
\end{equation}
This suggests defining a similar formulation in Wasserstein space as follows
\begin{equation}\label{eq:BDF2_mat}
\rho_n \in \argmin_{\rho\in\Pc(\Om)} \alpha \frac{{W}_2^2(\rho,\rho_{n-1})}{2(1-\beta)\tau} - \beta\frac{{W}_2^2(\rho,\rho_{n-2})}{2(1-\beta)\tau} + \mathcal{E}(\rho) \,.
\end{equation}
This approach has been proposed by Matthes and Plazotta \cite{Matthes2019bdf2,plazotta2018fokkerplanck}, who proved analogues of Theorems \ref{th:convergencefp} and \ref{th:convergenceevi}. It is easy to see that this leads to a different discrete solution than the scheme \eqref{eq:bdf2metric} (see, e.g., the example in Figure \ref{fig:equivalence}). However, just as for the metric extrapolation problem \eqref{eq:metricextra}, \eqref{eq:BDF2_mat} is not a convex optimization problem in the classical sense. For this reason, it is not easy to provide a numerical implementation of \eqref{eq:BDF2_mat} when $d\geq 2$. The same is true for the \scheme\ scheme \eqref{eq:bdf2metric} when using the metric extrapolation. Nonetheless, the advantage of using the \scheme\ scheme is that one has some freedom in choosing the extrapolation operator, which makes it more amenable to computations.
\begin{figure}
\begin{overpic}[scale=.9]{equivalence.pdf}
\put(205,74){$\rho_{n-2}$}
\put(205,52){$\rho_{n-1}$}
\put(142,44){$x$}
\put(82,107){$y$}
\put(205,30){$\rho_{n-1}^\alpha$}
\end{overpic}
\caption{An example for which the schemes \eqref{eq:bdf2metric} and \eqref{eq:BDF2_mat} provide different results, e.g., for the energy given by the convex indicator function of the set $\{\mu : \mu(\mathbb{R}^d\setminus\{x=0\}) =0\}$. In the figure $\rho_{n-2}$, $\rho_{n-1}$ and $\rho_{n-1}^\alpha$ are uniformly distributed on the segments $(t,-t)$, $(t,(1-\beta)t/\alpha)$ and $(t,t)$ for $t\in[-1,1]$, respectively (in this case the geodesic from $\rho_{n-2}$ to $\rho_{n-1}$ on the time interval $[0,1]$ can be extendend up to time $\alpha$, yielding $\rho_{n-1}^\alpha$). For the scheme \eqref{eq:bdf2metric} the measure $\rho_n$ is uniformly distributed on the segment $(0,t)$ for $t\in[-1,1]$, whereas for the scheme \eqref{eq:BDF2_mat} the measure $\rho_n$ can be obtained as the extrapolation of the projections of $\rho_{n-2}$ and $\rho_{n-1}$ on the axis $y$, and can be shown to have a strictly smaller support.
}\label{fig:equivalence}
\end{figure}
Another second-order variation of the JKO scheme was proposed by Legendre and Turinici \cite{Legendre2017VIM}, and it is based on the implicit midpoint rule, which applied to system \eqref{eq:gradF} leads to the scheme: for $n\geq1$ find $x_n \in \mathbb{R}^d$ satisfying
\[
\frac{1}{\tau}(x_n-x_{n-1}) = -\nabla F\Big(\frac{x_n+x_{n-1}}{2}\Big) \,,
\]
which can be obtained as the optimality conditions of the problem
\begin{equation}\label{eq:VIM_eu}
x_n = \argmin_{x\in\mathbb{R}^d} \frac{ \left|x-x_{n-1}\right|^2}{2\tau} + 2 F\Big(\frac{x+x_{n-1}}{2}\Big) \,.
\end{equation}
Translating such a scheme to the Wasserstein setting yields the Variational Implicit Midpoint (VIM) scheme proposed in \cite{Legendre2017VIM}: for $n\geq1$ find $\rho_n \in \Pc(\Om)$ satisfying
\begin{equation}\label{eq:VIM}
\rho_n \in \argmin_{\rho\in\Pc(\Om)} \frac{{W}_2^2(\rho,\rho_{n-1})}{2\tau} + 2 \mathcal{E}(\rho_{n-1/2}) \,,
\end{equation}
where ${\rho}_{n-1/2}$ is the midpoint of the (not necessarily unique) geodesic between $\rho$ and $\rho_{n-1}$.
Also in this case, it is not evident how to implement such a scheme, as it requires an explicit formula for the midpoint given the initial and final measures. This may also lead to convexity issues. Notice however that in the same spirit of our formulation of the BDF2 scheme, the implicit midpoint scheme can be formulated in the following alternative way: for $n\geq1$ find $\rho_n \in \Pc(\Om)$ satisfying
\begin{equation}\label{eq:VIM2}
\rho_n = {\sf E}_2 (\rho_{n-1},\rho_{n-1/2})\,, \quad \rho_{n-1/2} \in \argmin_{\rho\in\Pc(\Om)} \frac{{W}_2^2(\rho,\rho_{n-1}) }{\tau} + \mathcal{E}(\rho) \,,
\end{equation}
where ${\sf E}_2 (\rho_{n-1},\rho_{n-1/2})$ denotes the extrapolation at time $\alpha =2$ of a geodesic from $\rho_{n-1}$ (at time 0) to $\rho_{n-1/2}$ (at time 1). In general, this leads to a different discrete solution than the one obtained with \eqref{eq:VIM}, although the two schemes coincide if there exists a unique geodesic extension from $\rho_{n-1}$ to $\rho_{n-1/2}$ which stays globally length-minimizing up to time 2 for all $n$.
Nevertheless, the behavior of scheme \eqref{eq:VIM2} is radically different from that of the \scheme\ \eqref{eq:bdf2metric}, due to the different way JKO steps and extrapolations are performed. Namely, the order of the operations as well as the length of the steps play a crucial role. We will investigate this phenomenon numerically by considering a fully-discrete version of the VIM scheme and show that in general this approach may lead to persistent oscillations in the solution (Section \ref{ssec:comparison}).
Providing a fully discrete version of problem \eqref{eq:model}, via the \scheme\ scheme \eqref{eq:bdf2metric}, comes with an additional challenge since the chosen space discretization should also be second-order accurate in space, in order to exploit the increased accuracy of the time discretization.
We propose a discretization in the Eulerian framework of finite volumes. Specifically, we implement Two Point Flux Approximation (TPFA) finite volumes, which have been extensively analyzed lately for the discretization of optimal transport and Wasserstein gradient flows \cite{gladbach2018scaling,erbar2020computation,natale2021computation,cances2020LJKO,natale2020FVCA}.
Following these last two works in particular, we propose a scheme in which the Wasserstein distance is locally linearized, at each step of the scheme, in order to decrease the computational complexity of the approach, without dropping the second-order accuracy in time. In addition, we propose one possible discrete version of the extrapolation in this setting, which can be implemented in a robust way, and we verify numerically the second-order accuracy of the resulting approach.
\section{Preliminaries and notation}\label{sec:prem}
Let $\mc{P}_2(\mathbb{R}^d)$ be the space of probability measures with finite second moments. Given $\mu_0, \mu_1 \in \mc{P}_2(\mathbb{R}^d)$, we denote by $W_2(\mu_0,\mu_1)$ the $L^2$-Wasserstein distance between $\mu_0$ and $\mu_1$ (see, e.g.,
Chapter 5 in \cite{santambrogio2015optimal}). This can be defined via the following minimization problem:
\begin{equation}\label{eq:w2}
W_2^2(\mu_0,\mu_1)\coloneqq \min_{\substack{\gamma \in \Pi(\mu_0,\mu_1)}} \int {|x-y|^2}\,\mathrm{d} \gamma(x,y) \,,
\end{equation}
where $\Pi(\mu_0,\mu_1)$ is the set of probability measures on $\mathbb{R}^d\times \mathbb{R}^d$ with marginals $\mu_0$ and $\mu_1$. This problem always admits a solution $\gamma^*$, although it is not necessarily unique, which we refer to as an optimal transport plan from $\mu_0$ to $\mu_1$. By linearity of the constraint and of the function minimized in \eqref{eq:w2}, one can easily check that the function $W_2^2$ is jointly convex with respect to its arguments (with respect to the linear structure of $\mc{P}_2(\mathbb{R}^d)$). We will refer to the space of probability measures $\mc{P}_2(\mathbb{R}^d)$
equipped with the metric $W_2$ as the Wasserstein space.
Problem \eqref{eq:w2} admits an alternative dynamical formulation, which was introduced by Benamou and Brenier in \cite{benamou2000computational}, and which reads as follows
\begin{equation}\label{eq:w2dynamic}
W_2^2(\mu_0,\mu_1) = |t_1 - t_0| \min_{(\omega, v) \in \mathcal{C}} \int_{t_0}^{t_1} \mathrm{d} t \int_\Omega \omega(t) |v(t,\cdot)|^2
\end{equation}
where $\mc{C}$ is the set of curves $(\omega,v)$, with $\omega:[t_0,t_1] \rightarrow \mc{P}_2(\mathbb{R}^d)$ and $ v: [t_0,t_1] \rightarrow L^2(\omega(t);\mathbb{R}^d)$, satisfying weakly the continuity equation
\begin{equation}\label{eq:continuity}
\partial_t \omega + \mathrm{div}(\omega v ) = 0
\end{equation}
with zero flux boundary conditions (i.e.\ $\omega \,v \cdot n_{\partial \Omega} = 0 $), and initial and final conditions $
\omega(t_0) = \mu_0 $, $ \omega(t_1) = \mu_1$. The minimum in \eqref{eq:w2dynamic} is always achieved although there might be multiple minimizers. In particular, one can use formula \eqref{eq:w2dynamic} to deduce that the Wasserstein space is a geodesic space and the minimizers $\omega$ are geodesics.
By the optimality conditions of problem \eqref{eq:w2dynamic}, a curve $\omega$ is a geodesic if and only if there exists a time-dependent vector field $t\in[t_0,t_1] \mapsto \nabla \phi(t,\cdot) \in L^2(\omega(t);\mathbb{R}^d)$ such that $(\omega,\nabla \phi) \in \mc{C}$, and where the potential $\phi:[t_0,t_1]\times \mathbb{R}^d \rightarrow \mathbb{R}$ verifies:
\begin{enumerate}
\item $\phi(t_0, \cdot)$ is a continuous $(-|t_1-t_0|^{-1})$-convex function, i.e.\ such that the so-called Brenier potential
\begin{equation}\label{eq:u}
x\mapsto u(x) \coloneqq |t_1-t_0| \phi(t_0,x) + \frac{|x|^2}{2} \quad \text{is convex}\,;
\end{equation}
\item the potential $\phi$ is the unique viscosity solution of the Hamilton-Jacobi equation
\begin{equation}\label{eq:hj}
\displaystyle \partial_t \phi + \frac{|\nabla \phi |^2}{2} = 0,
\end{equation}
or equivalently, it verifies the Hopf-Lax representation formula,
\begin{equation}\label{eq:hamiltonjacobiintro}
\phi(t,x) = \inf_{y\in\mathbb{R}^d} \frac{|x-y|^2}{2(t-t_0)} + \phi(t_0,y)
\end{equation}
\end{enumerate}
We say that a function $\phi$ verifying these condition is an optimal potential from $\mu_0$ to $\mu_1$ on the time interval $[t_0,t_1]$.
Furthermore, for any optimal potential $\phi$, it holds:
\[
\frac{W_2^2(\mu_0,\mu_1)}{2} = \int_{\Omega} \phi(t_1,\cdot) \mu_1-\int_{\Omega} \phi(t_0,\cdot) \mu_0\,.
\]
Because of the semi-convexity of $\phi(t_0,\cdot)$, the maps $X(t,\cdot)$, defined a.e.\ by
\begin{equation}\label{eq:xflow}
X(t,\cdot) \coloneqq \mathrm{Id} + (t-t_0) \nabla \phi(t_0,\cdot)
\end{equation}
are injective for all $t\in[t_0,t_1)$ (as the gradient of a strongly convex function), and the resulting curve of maps $X:[t_0,t_1]\times \mathbb{R}^d \rightarrow \mathbb{R}^d$ is the Lagrangian flow of the time-dependent vector field $\nabla \phi(t,\cdot)$, i.e., for a.e.\ $x\in\mathbb{R}^d$, $X(\cdot,x)$ solves the flow equation
\[
\frac{\mathrm{d}}{\mathrm{d} t} X(t,x) = \nabla \phi(t,X(t,x)),\quad X(t_0,x) = x \,.
\]
If $\mu_0$ is absolutely continuous, given an optimal potential $\phi$ and the associated Lagrangian flow $X$ defined by \eqref{eq:xflow}, one can easily verify that the curve
\begin{equation}\label{eq:rhopush}
\omega(t) = X(t,\cdot)_\# \mu_0 \,,
\end{equation}
solves the continuity equation with velocity $\nabla \phi$ and boundary conditions $\omega(0)=\mu_0$ and $\omega(1)= \mu_1$ (in distributional sense), and therefore it is a geodesic. Moreover, using the absolute continuity of $\mu_0$, one can also show that the initial potential $\phi(t_0,\cdot)$ is uniquely defined $\mu_0$-a.e., and no other geodesic curve exists connecting $\mu_0$ and $\mu_1$.
Note also that from \eqref{eq:rhopush}, one can recover Brenier's result \eqref{eq:brenierth} with the Brenier potential $u$ as in \eqref{eq:u}, and also verify the equivalence with formulation \eqref{eq:w2}. As a matter of fact, in this case the optimal transport plan is also unique and is given by $\gamma^* = (\mathrm{Id},\nabla u)_\# \mu_0$, where the map $\nabla u$ is the so-called optimal transport map from $\mu_0$ to $\mu_1$. On the other hand, for any convex function $u$, setting $\phi(0,\cdot)$ via \eqref{eq:u}, the curve $\omega$ defined in \eqref{eq:rhopush} is a geodesic between $\mu_0$ and $(\nabla u)_\# \mu_0$ (and the unique one, if $\mu_0$ is absolutely continuous).
\section{Analysis of the \scheme\ scheme}\label{sec:LJKO2analysis}
In this section we collect the main properties of the \scheme\ discretization \eqref{eq:bdf2metric}, and in particular we prove Theorem \ref{th:convergencefp}, which establishes the convergence of the discrete flow generated by the scheme to the linear Fokker-Planck equation. Throughout the section, $(\rho_n)_n$ denotes a sequence of measures generated by the \scheme\ scheme \eqref{eq:bdf2metric}, where $\sf{E}_\alpha$ is a $\theta$-dissipative extrapolation, with $\theta<1/2$.
\subsection{Well-posedness and classical estimate}\label{ssec:well-posedness} We start by stating some a priori bounds, which are valid for a general class of energies.
In particular, in this paragraph, we only assume that $\mc{E}$ is lower-semicontinuous and bounded from below, so that problem \eqref{eq:bdf2metric} admits a minimizer at each step $n$.
\begin{lemma}\label{lem:almostED}
At each step $n$, the solution $\rho_{n}$ satisfies the following inequality
\begin{equation}\label{eq:almostED}
(1-\theta)\frac{{W}_2^2(\rho_{n},\rho_{n-1})}{2(1-\beta)\tau} + \mathcal{E}(\rho_{n}) \leq \theta \frac{{W}_2^2(\rho_{n-1},\rho_{n-2})}{2(1-\beta)\tau} + \mathcal{E}(\rho_{n-1})\,.
\end{equation}
\end{lemma}
\begin{proof}
Due to the optimality of $\rho_{n}$ and using \eqref{eq:dissipation}, we can write
\[
\begin{aligned}
\frac{{W}_2^2(\rho_{n},\rho^\alpha_{n-1})}{2(1-\beta)\tau} + \mathcal{E}(\rho_{n}) & \le \frac{{W}_2^2(\rho_{n-1},\rho^\alpha_{n-1}) }{2(1-\beta)\tau} + \mathcal{E}(\rho_{n-1}) \\
&\le \frac{\theta^2}{2(1-\beta)\tau} {W}_2^2(\rho_{n-1},\rho_{n-2}) + \mathcal{E}(\rho_{n-1})\,.
\end{aligned}
\]
If $\theta=0$ this coincides with \eqref{eq:almostED}. If $\theta>0$, observe that
by the triangular and Young's inequalities, for any $c>0$,
\[
{W}_2^2(\rho_{n},\rho_{n-1}) \le \Big(1+\frac{1}{c}\Big){W}_2^2(\rho_{n},\rho^\alpha_{n-1})+(1+c){W}_2^2(\rho_{n-1},\rho^\alpha_{n-1})\,.
\]
Setting $c=\theta^{-1}-1$ in this last inequality and using again \eqref{eq:dissipation}, we can estimate the left-hand side from below using
\[
\begin{aligned}
\frac{{W}_2^2(\rho_{n},\rho^\alpha_{n-1})}{2(1-\beta)\tau} &\ge \frac{1}{2(1-\beta)\tau} \left(\frac{c}{c+1} {W}_2^2(\rho_{n},\rho_{n-1}) -c{W}_2^2(\rho_{n-1},\rho^\alpha_{n-1})\right) \\
&\ge \frac{1-\theta}{2(1-\beta)\tau} {W}_2^2(\rho_{n},\rho_{n-1}) - \frac{(1-\theta)\theta}{2(1-\beta)\tau}{W}_2^2(\rho_{n-1},\rho_{n-2}) \,.
\end{aligned}
\]
Rearranging, we obtain \eqref{eq:almostED}.
\end{proof}
Note that if we take $\beta =0$, i.e.\ we remove the extrapolation step, we can take $\theta=0$ in \eqref{eq:almostED} and recover the standard dissipation estimate for the JKO scheme.
\begin{lemma}\label{lem:bound_W}
Let $C_1>0$ be a constant such that ${W}_2^2(\rho_1,\rho_0)\le C_1\tau$ and $\mathcal{E}(\rho_1)\leq \mathcal{E}(\rho_0)$. Then, it holds:
\begin{equation}\label{eq:bound_W}
\frac{1}{\tau} \sum_{n=0}^{N_{\tau}} {W}_2^2(\rho_{n},\rho_{n-1}) \le C
\end{equation}
for a constant $C>0$ depending only on $C_1$, $\beta$, $\theta$, $\mc{E}$ and $\rho_0$.
\end{lemma}
\begin{proof}
Summing over $n$ the inequality \eqref{eq:almostED} we obtain
\begin{equation}\label{eq:EDlong}
\frac{1-2\theta}{2(1-\beta)\tau} \sum_{n=0}^{N} {W}_2^2(\rho_{n},\rho_{n-1}) \le \mathcal{E}(\rho_1)-\mathcal{E}(\rho_{n}) + \frac{\theta}{(1-\beta)\tau} {W}_2^2(\rho_1,\rho_0)\,,
\end{equation}
Then, since $\theta<1/2$ and thanks to the lower bound on the energy and the assumption $\mc{E}(\rho_1)\leq \mc{E}(\rho_0)$, we have
\[
\frac{1}{\tau} \sum_{n=0}^{N} {W}_2^2(\rho_{n},\rho_{n-1}) \le \frac{2(1-\beta)}{1-2\theta} \Big( \mathcal{E}(\rho_0)- \inf \mathcal{E} \Big) +\frac{2\theta}{1-2\theta}C_1 \,.
\]
\end{proof}
\begin{remark} For a given $\rho_0$, one can always choose $\rho_1$ so that the constant $C_1$ above is independent of $\tau$ and $\mc{E}(\rho_1)\leq \mc{E}(\rho_0)$, which are also the assumptions in the statements of Theorems \ref{th:convergencefp} and \ref{th:convergenceevi}. For example, it is sufficient to take $\rho_1$ as the solution obtained after a finite number $N_0\in \mathbb{N}$ of JKO steps with time step $\tau/N_0$ and initial condition given by $\rho_0$, with $\mathcal{E}(\rho_0)<\infty$. In fact, in this case, by the same proof as for Lemma \ref{lem:bound_W} (with $\beta=\theta=0$), one can take $C_1 = 2 (\mathcal{E}(\rho_0)- \inf \mathcal{E} )$.
\end{remark}
\subsection{Convergence towards the Fokker-Planck equation}\label{ssec:LJKO2convergence}
Given a Lipschitz continuous exterior potential $V\in W^{1,\infty}(\Omega)$, the Fokker-Planck equation is given by
\begin{equation}\label{eq:FokkerPlanckLJKO2}
\partial_t \rhoe = \Delta \rhoe + \nabla \cdot (\rhoe \nabla V) \quad \text{in } (0,T)\times\Omega \,,
\end{equation}
complemented with no-flux boundary conditions $(\nabla \rhoe +\rhoe \nabla V) \cdot n_{\partial \Omega} = 0$ on $\p \Omega$ and an initial condition $\rhoe(0,\cdot)=\rho_0\in\Pc(\Om)$. Equation \eqref{eq:FokkerPlanckLJKO2} can be interpreted as a Wasserstein gradient flow with respect to the energy functional $\mc{E}:\mc{P}(\Omega) \rightarrow \mathbb{R}$ given by
\begin{equation}\label{eq:FokkerPlanckEnergy}
\mathcal{E}(\rho) = \mathcal{U}(\rho) + \int_{\Om} \rho V \,,
\end{equation}
where the internal energy $\mathcal{U}:\mc{P}(\Omega) \rightarrow \mathbb{R}$ (the entropy) is defined by
\begin{equation}\label{eq:entropy}
\mathcal{U}(\rho) \coloneqq
\begin{cases}
\int_{\Om} \rho \log\rho \quad &\text{if $\rho \ll \mathrm{d} x\mres \Omega$} \,, \\
+\infty \quad &\text{otherwise} \,,
\end{cases}
\end{equation}
where $\mathrm{d} x\mres \Omega$ denotes the restriction of the Lebesgue measure to the domain $\Omega$.
Since the function $x\mapsto x\log x$ is strictly convex and superlinear, the energy $\mc{E}$ is also strictly convex (with respect to the linear structure of $\mc{P}(\Omega)$) and lower semi-continuous (with respect to the weak-* topology: see, e.g., Proposition 7.7 in \cite{santambrogio2015optimal}). Since $W_2^2$ is continuous and convex in its arguments, there exists a unique solution $\rho_{n}$ to problem \eqref{eq:bdf2metric} at each step $n$, and this is furthermore absolutely continuous with respect to $\mathrm{d} x \mres \Omega$.
As in the previous paragraph, we assume that $\sf{E}_\alpha$ is a $\theta$-dissipative extrapolation with $\theta<1/2$, and $(\rho_n)_n$ denotes a sequence of measures generated by the associated \scheme\ scheme \eqref{eq:bdf2metric}.
Although the discrete flow does not move by strictly minimizing the energy at each step (see Lemma \ref{lem:almostED}), we will show that it converges to the maximal slope curve of $\mathcal{E}$. For this, we will rely on the same arguments as in the original work of Jordan, Kinderlehrer, and Otto \cite{jordan1998fokkerplanck} for the JKO scheme.
Relying on the estimate \eqref{eq:bound_W}, the compactness arguments for obtaining a limit curve are rather standard.
We introduce two density curves on the interval $[0,T]$, given by
\begin{equation}\label{eq:interpolations}
\begin{aligned}
&\rhoe_{\tau}(t) = \sum_{n=1}^{N} \rho_{n-1} \mathds{1}_{(t_{n-1},t_{n}]} \,, \quad \rho_{\tau}(0) = \rho_0 \,, \\
&\tilde{\rhoe}_{\tau}(t) = \sum_{n=1}^{N} \tilde{\rhoe}_{n}(t) \mathds{1}_{(t_{n-1},t_{n}]} \,, \quad \tilde{\rho}_{\tau}(0) = \rho_0 \,,
\end{aligned}
\end{equation}
with $t \mapsto \tilde{\rhoe}_{n}(t)$ being the geodesic curve between $\rho_{n-1}$ and $\rho_{n}$ on the time interval $[t_{n-1},t_{n}]$ (i.e.\ the minimizer of problem \eqref{eq:w2dynamic} on this interval). Let $\tilde{v}_n$ be the associated optimal vector field as in problem \eqref{eq:w2dynamic} for all $1\leq n \leq N$. By definition of $\tilde{\rhoe}_\tau$, we have that
\[
\p_t \tilde{\rhoe}_{\tau} +\nabla \cdot (\tilde{\rhoe}_{\tau}\tilde{v}_{\tau}) = 0
\]
in the distributional sense on $(0,T) \times\Omega$,
where $\tilde{v}_{\tau}$ is the vector field defined by $\tilde{v}_{\tau}|_{(t_{n-1},t_n]} = \tilde{v}_n$ for all $1\leq n \leq N$. Moreover, on each interval $[t_{n-1},t_{n}]$ it holds:
\[
{W}_2^2(\rho_{n},\rho_{n-1}) = \tau \int_{t_{n-1}}^{t_{n}}\int_{\Om} \tilde{\rhoe}_{\tau} |\tilde{v}_{\tau}|^2 \,.
\]
The curve $\rhoe_{\tau}$ is a piecewise constant measure-valued curve whereas $\tilde{\rhoe}_{\tau}$ is a (absolutely) continuous one, interpolating the discrete densities.
\begin{proposition}\label{prop:conv_rho}
For a given $\rho_0$ and any given $N\geq 1$, let $\rho_\tau$ be the curve defined as in equation \eqref{eq:interpolations}, with $\rho_1$ being such that $W^2_2(\rho_0,\rho_1) \leq C \tau$, for a constant $C>0$ independent of $\tau$, and $\mc{E}(\rho_1)\leq \mc{E}(\rho_0)$. Then, the sequence $(\rhoe_{\tau})_{\tau}$ converges uniformly in the ${W}_2$ distance to an absolutely continuous curve $\rhoe:[0,T]\rightarrow \mc{P}(\Omega)$.
\end{proposition}
\begin{proof}
The sequence of curves $(\tilde{\rhoe}_\tau)_{\tau\in{\mathbb R}_+}$, defined from $[0,T]$ to the (compact) space $\mc{P}(\Omega)$ equipped with the Wasserstein distance, is uniformly H\"older continuous.
Indeed, for any $r,s\in (0,T], s>r$, denote $N_r,N_s$ the two integers such that $r\in(t_{N_r},t_{N_r+1}], s\in(t_{N_s},t_{N_s+1}]$. By the dynamical formulation of the Wasserstein distance \eqref{eq:w2dynamic}, it holds
\begin{equation}\label{eq:holder_estim}
\begin{aligned}
{W}_2(\tilde{\rhoe}_{\tau}(s),\tilde{\rhoe}_{\tau}(r)) &\le |s-r|^{\frac{1}{2}} \left(\int_{r}^{s}\int_{\Om} \tilde{\rhoe}_{\tau} |\tilde{v}_{\tau}|^2\right)^{\frac{1}{2}} \le |s-r|^{\frac{1}{2}} \left( \sum_{n=N_r}^{N_s} \int_{t_{n}}^{t_{n+1}}\int_{\Om} \tilde{\rhoe}_{\tau} |\tilde{v}_{\tau}|^2\right)^{\frac{1}{2}} \\
&= |s-r|^{\frac{1}{2}} \left( \sum_{n=N_r}^{N_s} \frac{1}{\tau}{W}_2^2(\rho_{n},\rho_{n+1})\right)^{\frac{1}{2}} \le C |s-r|^{\frac{1}{2}}
\end{aligned}
\end{equation}
where in the last inequality we used the estimate \eqref{eq:bound_W}. By the generalized Ascoli-Arzelà theorem, the sequence converges uniformly in ${W}_2$, up to a subsequence, to a limit curve $\rhoe$. As the inequality \eqref{eq:holder_estim} passes to the limit, $\rhoe$ is also an absolutely continuous curve with respect to the Wasserstein metric.
Finally, for any $r\in[0,T]$,
\[
{W}_2(\rhoe_{\tau}(r),\tilde{\rhoe}_{\tau}(r)) = {W}_2(\tilde{\rhoe}_{\tau}(t_{N_r}),\tilde{\rhoe}_{\tau}(r)) \le \left( \int_{t_{N_r}}^{t_{N_r+1}}\int_{\Om} \tilde{\rhoe}_{\tau} |\tilde{v}_{\tau}|^2 \right)^{1/2}\le C \sqrt{\tau} \,,
\]
by the same computations.
Therefore, the piecewise continuous curve $\rhoe_\tau$ converges uniformly with order $\sqrt{\tau}$ to the same limit curve $\rhoe$.
\end{proof}
To characterize the limit curve $\rhoe$ we will rely on the optimality conditions of the minimization problem in \eqref{eq:bdf2metric}, which is equivalent to a single JKO step.
Consider an absolutely continuous measure $\rho$ and a smooth vector field ${\xi}$ tangent to the boundary of $\Omega$. We define $\omega$ as the absolutely continuous curve solution to
\begin{equation}\label{eq:curve_var}
\partial_s \omega + \nabla \cdot (\omega {\xi}) = 0\,, \quad \text{in } (-\delta,\delta)\times\Omega\,, \quad \omega(0) = \rho\,,
\end{equation}
for $\delta>0$.
The variations of the energy and the Wasserstein distance along curves defined in this way can be computed explicitly as follows.
\begin{lemma}\label{lem:variations_ac}
Consider two measures $\rho\in \mc{P}(\Omega)$ and $\nu\in\mc{P}_2(\mathbb{R}^d)$, with $\rho$ absolutely continuous, and denote by $\gamma$ the optimal transport plan from $\rho$ to $\nu$.
For any ${\xi}\in\C^\infty(\Omega;\mathbb{R}^d)$ such that ${\xi}\cdot n_{\partial \Omega} = 0$ on $\partial \Omega$, let $\omega$ be the curve of measures defined by \eqref{eq:curve_var} with $\omega(0)=\rho$. It holds:
\begin{equation}\label{eq:var_W2}
\frac{\text{d} {W}_2^2(\omega(s),\nu)}{\text{d} s} \Big|_{s=0} = 2 \int_{\Omega\times\Omega} \langle x-y,{\xi}({x}) \rangle \, \text{d} \gamma(x,y) \,,
\end{equation}
and
\begin{equation}\label{eq:var_energy}
\frac{\text{d} \mathcal{E}(\omega(s))}{\text{d} s} \Big|_{s=0} = -\int_{\Om} \nabla \cdot {\xi} ({x}) \mathrm{d} \rho({x}) + \int_{\Om} \langle \nabla V({x}) , {\xi}({x}) \rangle \mathrm{d} \rho({x}) \,.
\end{equation}
\end{lemma}
\begin{proof}
See \cite[Theorem 5.1]{jordan1998fokkerplanck}.
\end{proof}
We are now ready to prove Theorem \ref{th:convergencefp} which states the convergence of the sequence of curves $(\rhoe_\tau)_{\tau}$ towards a distributional solution of equation \eqref{eq:FokkerPlanckLJKO2}. Specifically, we need to prove that, for all $\varphi\in\C^{\infty}_c([0,T)\times\Omega)$, the limit curve $\rhoe$ satisfies:
\begin{equation}\label{eq:FokkerPlanck_dist}
-\int_0^T\int_{\Om} \p_t \varphi \rhoe-\int_{\Om} \varphi(0)\rhoe(0) -\int_0^T \int_{\Om} \Delta \varphi \rhoe + \int_0^T\int_{\Om} \langle \nabla V , \nabla \varphi \rangle \rhoe = 0 \,.
\end{equation}
\begin{proof}[Proof of Theorem \ref{th:convergencefp}]
Let us define for all $\rho \in \mc{P}(\Omega)$,
\begin{equation}
\mathcal{G}(\rho_{n-1},\rho_{n-2};\rho) \coloneqq \frac{ {W}_2^2(\rho,\rho^\alpha_{n-1}) }{2(1-\beta)\tau}+ \mathcal{E}(\rho) \,,
\end{equation}
which is minimized by $\rho_n$, by the definition of the scheme \eqref{eq:bdf2metric}. Consider a smooth function $\varphi\in\C^{\infty}_c([0,T]\times\Omega)$ such that $\nabla\varphi\cdot n_{\partial \Omega}=0$ on $\p\Omega$. We define the sequence $(\varphi_{n})_n\subset\C_c^{\infty}(\Omega)$ as $\varphi_{n}=\varphi(t_{n},\cdot)$.
Consider then a curve $\omega$ defined as in \eqref{eq:curve_var} with $\omega(0)=\rho_{n}$ and ${\xi}= \nabla \varphi_{n-2}$.
Denoting by ${\gamma}_{n}$ the optimal transport plan from $\rho_{n}$ to $\rho^\alpha_{n-1}$, and using \eqref{eq:var_W2}-\eqref{eq:var_energy} as well as the optimimality of $\rho_n$, we obtain
\begin{equation}\label{eq:BDF2_mod_OC}
\begin{multlined}
\frac{\text{d} \mathcal{G}(\rho_{n-1},\rho_{n-2}; \omega(s))}{\text{d} s}\Big|_{s=0} = \frac{1}{(1-\beta)\tau} \int_{\Omega\times\Omega} \langle x-x_\alpha,\nabla \varphi_{n-2}({x}) \rangle \text{d} {\gamma}_{n} (x,x_\alpha) \\ -\int_{\Om} ( \Delta \varphi_{n-2}({x}) ) \mathrm{d} \rho_{n}({x}) + \int_{\Om} \langle \nabla V({x}) , \nabla \varphi_{n-2} ({x}) \rangle \mathrm{d} \rho_{n}({x}) = 0 \,.
\end{multlined}
\end{equation}
Thanks to Proposition \ref{prop:conv_rho} and the regularity of $\varphi$, we immediately have
\begin{multline*}
\Bigg|
\sum_{n=2}^{N} \tau \left(-\int_{\Om} ( \Delta \varphi_{n-2}) \rho_{n} + \int_{\Om} \langle \nabla V , \nabla \varphi_{n-2} \rangle \rho_{n} \right) \\
-\left( -\int_0^T \int_{\Om} \Delta \varphi \rhoe + \int_0^T\int_{\Om} \langle \nabla V , \nabla \varphi \rangle \rhoe \right)
\Bigg| \longrightarrow 0 \,,
\end{multline*}
for $\tau\rightarrow0$. In order to prove that the measure $\rhoe$ is a distributional solution of equation \eqref{eq:FokkerPlanckLJKO2} we need to show that
\begin{multline*}
I_1 \coloneqq \Bigg|
\sum_{n=2}^{N} \frac{1}{1-\beta} \int_{\Omega\times\Omega} \langle x-x_\alpha,\nabla \varphi_{n-2}({x}) \rangle \text{d} \gamma_{n} (x,x_\alpha) \, \\
-\left(-\int_0^T\int_{\Om} \p_t \varphi \rhoe-\int_{\Om} \varphi(0)\rhoe(0) \right)
\Bigg| \longrightarrow 0 \,,
\end{multline*}
as well.
We can bound the latter quantity as $I_1\le I_2+I_3$, where $I_2 = \big| \sum_{n=2}^{N} I_2^n \big|$ with
\[
I_2^n \coloneqq
\frac{1}{1-\beta} \int_{\Omega\times\Omega} \langle x-x_\alpha,\nabla\varphi_{n-2}({x}) \rangle \text{d} \gamma_{n}(x,x_\alpha)
-\frac{1}{1-\beta} \int_{\Omega} (\rho_{n}-\alpha\rho_{n-1}+\beta\rho_{n-2}) \varphi_{n-2} \,,
\]
and
\begin{equation*}
I_3 \coloneqq \left| \sum_{n=2}^{N} \frac{1}{1-\beta} \int_{\Omega} (\rho_{n}-\alpha\rho_{n-1}+\beta\rho_{n-2}) \varphi_{n-2} -
\left(-\int_0^T\int_{\Om} \p_t \varphi \rhoe-\int_{\Om} \varphi(0)\rhoe(0) \right)
\right| \,.
\end{equation*}
Integrating by part the discrete derivative in this last term,
\begin{multline*}
\sum_{n=2}^{N} \frac{1}{1-\beta}\int_{\Omega} (\rho_{n}-\alpha\rho_{n-1}+\beta\rho_{n-2}) \varphi_{n-2} = \\
\begin{aligned}
&=\sum_{n=2}^{N} \frac{1}{1-\beta} \int_{\Omega} (\varphi_{n-2}-(\alpha\varphi_{n-1}-\beta\varphi_{n})) \rho_{n} + \frac{1}{1-\beta} \int_{\Omega} \beta \varphi_0\rho_0 +(\beta\varphi_1-\alpha\varphi_0)\rho_1\,.
\end{aligned}
\end{multline*}
Then, since $\alpha = 1+\beta$, and thanks to the smoothness of the function $\varphi$ and Proposition \ref{prop:conv_rho}, we obtain $I_3\le C\tau$ for some constant $C$ independent of $\tau$.
Let us focus then on the term $I_2$. Adding and subtracting $(1-\beta)^{-1}\int_{\Omega} (\rho_{n}-\rho^\alpha_{n-1})\varphi_{n-2}$ at each step $n$, we obtain
\begin{align*}
I_2^n
&\le
\begin{aligned}[t]
&\frac{1}{1-\beta} \left| \int_{\Omega\times\Omega} \langle x-x_\alpha,\nabla\varphi_{n-2}({x}) \rangle \text{d} \gamma_{n}(x,x_\alpha) - \int_{\Omega} (\rho_{n}-\rho^\alpha_{n-1})\varphi_{n-2} \right| \\
&+\frac{1}{1-\beta} \left| \int_{\Omega} (\alpha \rho_{n-1}-\beta \rho_{n-2}-\rho^\alpha_{n-1}) \varphi_{n-2} \right| \eqqcolon\frac{1}{1-\beta}( I_4^n+ I_5^n) \,.
\end{aligned}
\end{align*}
Rewriting
\[
\int_{\Omega} (\rho_{n}-\rho^\alpha_{n-1})\varphi_{n-2} = \int_{\Om\times\Om} (\varphi_{n-2}({x})-\varphi_{n-2}(x_\alpha)) \text{d} {\gamma}_{n}(x,x_\alpha) \,,
\]
we can bound $I_4^n$ as
\[
\begin{aligned}
I_4^n &=\left| \int_{\Om\times\Om} \varphi_{n-2}({x}) - \varphi_{n-2}(x_\alpha) - \langle x-x_\alpha,\nabla\varphi_{n-2}({x}) \rangle \text{d} {\gamma}_{n}(x,x_\alpha) \right| \\
&\le \frac{1}{2} ||\text{Hess}(\varphi_{n-2})||_{\infty} \left( \int_{\Om\times\Om} |x-x_\alpha|^2 \text{d} {\gamma}_{n}(x,x_\alpha) \right) \\
&= \frac{1}{2} ||\text{Hess}(\varphi_{n-2})||_{\infty} {W}_2^2(\rho_{n},\rho^\alpha_{n-1}) \\
&\le ||\text{Hess}(\varphi_{n-2})||_{\infty} \left({W}_2^2(\rho_{n},\rho_{n-1})+{W}_2^2(\rho_{n-1},\rho^\alpha_{n-1})\right) \\
&\le ||\text{Hess}(\varphi_{n-2})||_{\infty} \Big({W}_2^2(\rho_{n},\rho_{n-1})+\theta^2 {W}_2^2(\rho_{n-1},\rho_{n-2})\Big) \,,
\end{aligned}
\]
where we used the dissipation estimate \eqref{eq:dissipation}. Similarly by the consistency assumption \eqref{eq:consistency} on the extrapolation, there exists a constant $C_\varphi$ only depending on $\varphi$ and $\Omega$ such that
\[
I_5^n \leq C_\varphi W^2_2(\rho_{n-1},\rho_{n-2})\,.
\]
Using the bound \eqref{eq:bound_W}, the estimates above imply that there exists a constant $C>0$ such that $I_2 \leq C \tau$.
The whole term $I_1$ is therefore converging to zero and $\rhoe$ satisfies equation \eqref{eq:FokkerPlanck_dist}.
\end{proof}
\section{Extrapolation in Wasserstein space}\label{sec:extrapolation}
In this section we consider the issue of defining geodesic extrapolations in the Wasserstein space. In particular, we propose several notions of extrapolation operators $\sf{E}_\alpha$, which in some cases verify the assumptions of Theorem \ref{th:convergencefp}, and discuss their relationship. To simplify the exposition, in the whole section we consider the extrapolation problem on the whole space $\mc{P}_2(\mathbb{R}^d)$, even though some of the proposed definitions may be adapted so that the extrapolation of two measures in $\mc{P}(\Omega)$ stays in $\mc{P}(\Omega)$ (see Remark \ref{rem:omega}). This last property is not required in our definition of the EVBDF2 scheme \eqref{eq:bdf2metric}, but it is useful to produce a fully-discrete scheme (see Section \ref{sec:discreteextra}).
As recalled in the introduction, a globally-minimizing geodesic with respect to the $W_2$ metric is a curve $\omega: [t_0,t_1] \rightarrow \mc{P}_2(\mathbb{R}^d)$ such that
\begin{equation}\label{eq:geodesicglob}
W_2(\omega(s_0),\omega(s_1)) = \frac{|s_1-s_0|}{|t_1-t_0|} W_2(\omega(t_0),\omega(t_1))\,,
\end{equation}
for all $s_0,s_1 \in (t_0,t_1)$. We say that $\omega: [t_0,t_1] \rightarrow \mc{P}_2(\mathbb{R}^d)$ is a locally-minimizing geodesic if for all $t\in(t_0,t_1)$ there exists an open interval $J \ni t$ such that \eqref{eq:geodesicglob} holds for all $s_0,s_1 \in J \cap (t_0,t_1)$.
From the discussion in Section \ref{sec:prem}, given two measures $\mu_0, \mu_1 \in \mc{P}_2(\mathbb{R}^d)$, if $\mu_0$ is absolutely continuous there exists a unique globally length-minimizing geodesic connecting the two, which is given by
\begin{equation}\label{eq:brenier2}
\omega(t) = ((1-t) \mathrm{Id} + t \nabla u)_\# \mu_0
\end{equation}
for $t\in[0,1]$, where $u$ is a uniquely defined convex function $\mu_0$-a.e.\ (up to an additive constant). As a matter of fact, we have for all $s_0,s_1 \in (0,1)$,
\begin{equation}\label{eq:geoineq}
\begin{aligned}
W^2_2(\omega(s_0),\omega(s_1))& \leq \int_{\mathbb{R}^d} |(1-s_0) x + s_0 \nabla u(x) - (1-s_1) x - s_1 \nabla u(x)|^2 \mathrm{d} \mu_0(x)\\ &=|s_1 - s_0|^2 W_2^2(\mu_0,\mu_1),
\end{aligned}
\end{equation}
where for the first inequality we used as competitor the plan $((1-s_0) \mathrm{Id} + s_0 \nabla u, (1-s_1) \mathrm{Id} + s_1 \nabla u)_\# \mu_0$, and for the second equality the optimality of the plan $(\mathrm{Id},\nabla u)_\#\mu_0$ for the transport problem from $\mu_0$ to $\mu_1$. On the other hand, for $s_1>s_0$, by the triangular inequality and \eqref{eq:geoineq}
\[
\begin{aligned}
W_2(\mu_0,\mu_1)& \leq W_2(\mu_0,\omega(s_0)) + W_2(\omega(s_0),\omega(s_1)) +W_2(\omega(s_1),\mu_1) \\
& \leq (s_0 + 1 - s_1) W_2(\mu_0,\mu_1) + W_2(\omega(s_0),\omega(s_1))\,,
\end{aligned}
\]
and therefore the inequality in \eqref{eq:geoineq} is an equality.
Moreover, by similar calculations one can verify that for any $\alpha\geq 1$ the curve $t\in[0,\alpha] \mapsto \omega(t)$, still defined as in \eqref{eq:brenier2}, is a globally length-minimizing geodesic if and only if $u$ is $\beta/\alpha$-convex, i.e.\ the function
\begin{equation}\label{eq:convexityextra}
x \mapsto \alpha u(x) - \beta \frac{|x|^2}{2} \quad \text{is convex},
\end{equation}
with $\beta =\alpha -1$.
However, in general, there is no guarantee that $u$ is strongly-convex even if $\mu_0$ and $\mu_1$ have smooth and strictly positive densities and for arbitrarily small $\beta$, as shown by the following example.
\begin{example}[Contraction flow]
Take $u = \frac{\beta}{2\alpha}|\cdot|^2$, for $\alpha >1$ and $\beta = \alpha-1$. Then, for any absolutely continuous $\mu_0 \in \mc{P}_2(\mathbb{R}^2)$ and $\mu_1 = (\nabla u)_\# \mu_0$, there exists a unique globally length-minimizing geodesic on $(-\infty,\alpha]$ such that $\omega(0) = \mu_0$ and $\omega(1)=\mu_1$, which is given by \eqref{eq:brenier2}. On the other hand, since all trajectories cross at time $\alpha$ (i.e.\ $(1-\alpha) \mathrm{Id} + \alpha \nabla u = 0$), there exists no geodesic on $(-\infty,\alpha']$ (either local or global) with $\alpha'>\alpha$ satisfying the same property.
\end{example}
In the case of compactly supported measures, globally length-minimizing geodesic extensions may not exist even if particle trajectories do not cross. In this case, however, locally length-minimizing extensions may still exist as shown in the next example.
\begin{example}[Shear flow]\label{ex:shear} For $d=2$, let
\[
\mu_0 = \frac{1}{2}(\delta_{z} + \delta_{-z}) \,, \quad \mu_1 = \frac{1}{2}(\delta_{z -v} + \delta_{-z+v})
\]
where $z = (1,1)$ and $v=(1,0)$. In this case, there exists a unique geodesic $\omega:\mathbb{R}\rightarrow \mathcal{P}_2(\mathbb{R}^2)$ which is locally length-minimizing, and such that $\omega(0) = \mu_0$ and $\omega(1)= \mu_1$, which is given by
\begin{equation}\label{eq:shear}
\omega(t) = \frac{1}{2}(\delta_{z - tv} + \delta_{-z+tv}).
\end{equation}
However, $\omega$ is globally length-minimizing only when restricted on $(-\infty,2]$.
\end{example}
In order to define our scheme, we need an extrapolation operator which is well-defined even when the geodesic extension (either globally or locally length-minimizing) does not exist. In the following we will introduce different possible definitions and describe their properties.
\subsection{Free-flow extrapolations} One possible strategy for defining an extrapolation consists in disregarding the convexity condition on the Brenier potential in \eqref{eq:convexityextra}, and allowing particles to cross each other while keeping their straight trajectories at constant speed. If $\mu_0\in \mc{P}_2(\mathbb{R}^d)$ is absolutely continuous, this amounts to defining, for any $\mu_1 \in \mc{P}_2(\mathbb{R}^d)$ and $\alpha>1$,
\begin{equation}\label{eq:lagextra}
{\sf E}_\alpha(\mu_0,\mu_1) = ((1-\alpha) \mathrm{Id} + \alpha \nabla u)_\# \mu_0 \,,
\end{equation}
where $u$ is a Brenier potential from $\mu_0$ to $\mu_1$ (uniquely defined $\mu_0$-a.e.). If $\mu_0$ is not absolutely continuous, there may exist multiple geodesics and optimal transport plans from $\mu_0$ to $\mu_1$. In general, we say that an extrapolation operator $\sf{E}_\alpha$ yields a \emph{free-flow extrapolation} if, denoting by $\Gamma(\mu_0,\mu_1)$ the set of optimal plans from $\mu_0$ to $\mu_1$, one has:
\begin{equation}\label{eq:freeflow}
\forall \mu_0,\mu_1 \in \mc{P}_2(\mathbb{R}^d)\,, ~ \exists\, \gamma^* \in \Gamma(\mu_0,\mu_1)~ :~ {\sf E}_\alpha(\mu_0,\mu_1) =(\pi_{\alpha})_\# \gamma^* \,,
\end{equation}
where $\pi_\alpha: \mathbb{R}^d\times \mathbb{R}^d \rightarrow \mathbb{R}^d$ is the map defined by $\pi_\alpha(x,y) = x+ \alpha (y-x)$.
By construction, when the geodesic induced by $\gamma^*$ in \eqref{eq:freeflow} admits a locally (or globally) length-minimizing geodesic extension, the resulting free-flow extrapolation is always consistent with it (for example, free-flow extrapolations yield the curve \eqref{eq:shear} in the case of Example \ref{ex:shear}). Furthermore, such extrapolation operators are admissible for our scheme in the sense of Theorem \ref{th:convergencefp}, as shown by the following proposition.
\begin{proposition} \label{prop:lagextra} Any free-flow extrapolation operator ${\sf E}_\alpha: \mc{P}_2(\mathbb{R}^d)\times \mc{P}_2(\mathbb{R}^d) \rightarrow \mc{P}_2(\mathbb{R}^d)$, i.e.\ any map satisfying \eqref{eq:freeflow}, is $\beta$-dissipative with $\beta = \alpha -1$, and in addition it verifies the consistency assumption \eqref{eq:consistency}.
\end{proposition}
\begin{proof} For simplicity, we only consider the case where $\mu_0$ is absolutely continuous. Let $\nabla u$ the optimal transport map from $\mu_0$ to $\mu_1$.
To prove the dissipativity, let $\bar{\gamma} = (\nabla u, (1-\alpha) \mathrm{Id} + \alpha \nabla u)_\# \mu_0$. Then $\bar{\gamma} \in \Pi(\mu_1, {\sf E}_\alpha(\mu_0,\mu_1))$ and by equation \eqref{eq:w2},
\[
W^2_2(\mu_1, {\sf E}_\alpha(\mu_0,\mu_1)) \leq \int |x-y|^2 \mathrm{d} \bar{\gamma}(x,y) = (1-\alpha)^2 \int_{\Omega} | \mathrm{Id} - \nabla u|^2 \mu_1 = \beta^2W^2_2(\mu_0,\mu_1)\,.
\]
For the consistency, let $\varphi \in C^\infty_c(\mathbb{R}^d)$ and observe that, by the definition of pushforward,
\[
\int \varphi\, ( \mathsf{E}_\alpha(\mu_0,\mu_1) - \alpha \mu_1 + \beta \mu_0) = \int \left[ \varphi\big( (1-\alpha)x + \alpha \nabla u(x)\big) -\alpha \varphi\big(\nabla u(x)\big) +\beta \varphi(x) \right]\mathrm{d} \mu_0(x) \,.
\]
Using the Taylor expansion of $\varphi$ around the point $x$ in the integral on the right-hand side, we find
\[
\left|\int \varphi\, ( \mathsf{E}_\alpha(\mu_0,\mu_1) - \alpha \mu_1 + \beta \mu_0)\right| \leq \frac{\alpha\beta}{2} \|\mathrm{Hess}(\varphi)\|_\infty W^2_2(\mu_0,\mu_1)\,.
\]
In the general case where $\mu_0$ is not absolutely continuous, the proof is analogous replacing transport maps by optimal plans.
\end{proof}
\subsection{Extrapolation with collisions}\label{sec:collisions} Free-flow extrapolations are the simplest way to extend geodesics after their maximal time of existence, but they are purely Lagrangian and they cannot be easily implemented in an Eulerian setting. Here we describe an alternative route to construct an extrapolation operator which prevents particles to cross, and which is based on viscosity solutions of the Hamilton-Jacobi equation. The resulting operator can be implemented in a robust way, but unfortunately it falls outside the hypotheses of the convergence results presented in this work. In Section \ref{sec:numerics}, we will describe its implementation in detail and verify numerically that it leads to a second-order scheme.
Given $\mu_0, \mu_1 \in\mc{P}_2(\mathbb{R}^d)$, let us suppose that the optimal potential $\phi$ for the transport from $\mu_0$ to a given measure $\mu_1$ on the time interval $[0,1]$, is such that
\begin{equation}\label{eq:phi0}
\phi(0,\cdot) \text{ is globally Lipschitz.}
\end{equation}
Then, the curve $\omega:[0,\infty) \rightarrow \mc{P}_2(\mathbb{R}^d)$ satisfying
\begin{equation}\label{eq:omegahj}
\omega(t) = \left[\nabla \mathrm{co} \left ((1-t) \frac{|\cdot|^2}{2} + t u\right)\right]_\# \mu_0 \,,
\end{equation}
where $u = |\cdot|^2/2 + \phi(0,\cdot)$ is a Brenier potential from $\mu_0$ to $\mu_1$, and where $\mr{co}$ denotes the convex hull, is well-defined. We remark that \eqref{eq:omegahj} coincides at time $t=\alpha$ with the free-flow extrapolation \eqref{eq:lagextra} as long as the convexity condition \eqref{eq:convexityextra} holds. On the other hand, if such condition is not verified, taking the convex envelope in \eqref{eq:omegahj} guarantees that the flow stays monotone and particles cannot cross.
If \eqref{eq:phi0} holds, one also has that the Hamilton-Jacobi equation \eqref{eq:hj} with intial condition $\phi(0,\cdot)$ has a unique viscosity solution, which is given by the Hopf-Lax formula
\begin{equation}\label{eq:hopf}
\phi(t,\cdot)=\mathcal{H}_t(\phi(0,\cdot)), \quad \mathcal{H}_t(\phi(0,\cdot))(x) \coloneqq \inf_{y\in\mathbb{R}^d} \frac{|x-y|^2}{2t} + \phi(0,y)\,.
\end{equation}
Note that the evolution of the density transported by the velocity field $\nabla \phi(t,\cdot)$ (via the continuity equation) is also well-defined since so is its Lagragian flow \cite{khanin2010particle,bogaevsky2004matter}. In the following lemma we show that equations \eqref{eq:hopf} and \eqref{eq:omegahj} are closely related.
\begin{lemma}
Let $\phi:[0,\infty) \times \mathbb{R}^d\rightarrow \mathbb{R}$ be the unique viscosity solution to the Hamilton-Jacobi equation, or equivalently verifying \eqref{eq:hopf} for $t>0$, with $\phi(0,\cdot)$ being a Lipschitz function, and denote $u \coloneqq \phi(0,\cdot) + \frac{|\cdot|^2}{2}$. Let $\mu_0\in\mathcal{P}_2(\mathbb{R}^d)$ be an absolutely continuous measure and $\omega :[0,\infty) \rightarrow \mathcal{P}_2(\mathbb{R}^d)$ be the curve defined by \eqref{eq:omegahj}
for all $t \geq 0$. Then,
\begin{enumerate}
\item for all $t \geq 0$,
$\omega(t)$ solves
\begin{equation}\label{eq:varextra}
\min_{\mu\in\mathcal{P}_2(\mathbb{R}^d)} \frac{W^2_2(\mu_0, \mu)}{2t} - \int_{\mathbb{R}^d} \phi(t, \cdot) \mu\,;
\end{equation}
\item if $d=1$, $\omega$ is a weak solution to the continuity equation with velocity $\nabla \phi(t,\cdot)$.
\end{enumerate}
\end{lemma}
\begin{proof}
Concerning the first point, note that for any function $\psi$ it holds:
\begin{equation}\label{eq:hopf-leg}
\begin{aligned}
\frac{|y|^2}{2}-t\mathcal{H}_t(\psi)(y)
&=\frac{|y|^2}{2}-\inf_x \frac{|x-y|^2}{2} +t \psi(x) \\
&=\sup_x y \cdot x -\left(\frac{|x|^2}{2}+t\psi(x)\right)=\left(\frac{|\cdot|^2}{2}+t\psi(\cdot)\right)^*(y), \quad \forall y\,.
\end{aligned}
\end{equation}
In particular, this implies that $-t\phi(t,\cdot) + |\cdot|^2/2$ is convex. Hence, by the optimality conditions of problem \eqref{eq:varextra} \cite[Example 7.21]{santambrogio2015optimal}, an optimal potential $\tilde{\phi}:[0,t] \times \mathbb{R}^d \rightarrow \mathbb{R}$ from $\mu_0$ to $\mu$ can be selected by setting $\tilde{\phi}(t,\cdot) = \phi(t,\cdot)$. Since by definition $\tilde{\phi}(t,\cdot) = \mathcal{H}_t(\tilde{\phi}(0,\cdot))$ one can deduce that $\tilde{\phi}(0,\cdot) = - \mathcal{H}_t(-\phi(t,\cdot))$.
Therefore, the optimal transport map from $\mu_0$ to the optimal measure $\mu$ is the gradient of $\frac{|\cdot|^2}{2}-t\mathcal{H}_t(-\phi(t,\cdot))=\frac{|\cdot|^2}{2}-t\mathcal{H}_t(-\mathcal{H}_t(\phi(0,\cdot)))$. Applying two times \eqref{eq:hopf-leg} we conclude.
For the second part, we refer to Proposition 4.1 in \cite{ben2003system}, where an explicit expression for the measure transported by the flow is provided.
\end{proof}
\begin{remark}\label{rem:viscosity+continuity}
For $d>1$, the curve \eqref{eq:omegahj} does not coincide in general with the solution of the continuity equation with velocity $\nabla \phi(t,\cdot)$. This is because \eqref{eq:omegahj} completely disregards the dynamics of mass within the shocks, which may be non-trivial \cite{ben2003system,bogaevsky2004matter}.
\end{remark}
There are two main problems with using \eqref{eq:omegahj} to define an extrapolation operator, i.e.\ setting $\mathsf{E}_\alpha(\mu_0,\mu_1)=\omega(\alpha)$. First, the initial potential $\phi(0,\cdot)$ is uniquely defined only $\mu_0$-a.e., however the value of the potential outside the support of $\mu_0$ does affect the final measure $\omega(\alpha)$ for $\alpha>1$. Second, because of the same reason one can easily construct solutions that are not dissipative in the sense of Definition \ref{def:dissipation}: for example, one can take $\mu_0 =\mu_1$ with compact support and select an initial potential outside the support in such a way that $\omega(\alpha)$ (defined as in the previous lemma) is different from $\mu_1$.
\begin{remark}[Extrapolation via pressureless fluids]\label{rem:pressureless}
With the same notation as above, one could construct geodesic continuations also by looking for solutions $\omega:[0,\infty)\rightarrow \mc{P}_2(\mathbb{R}^d)$, $v:[0,\infty)\rightarrow L^2(\omega(t);\mathbb{R}^d)$, of the following system of PDEs:
\begin{equation}\label{eq:pressureless}\left\{
\begin{array}{l}
\partial_t \omega + \mathrm{div}(\omega v) = 0\,,\\
\displaystyle \partial_t (\omega v) + \mathrm{div}\left( \omega {v\otimes v}\right) = 0\,,\\
\end{array}\right.,
\end{equation}
with zero boundary flux and initial conditions given by
\[
\omega (0) = \mu_0\,, \quad v(0,\cdot) = \nabla \phi(0,\cdot)\,.
\]
System \eqref{eq:pressureless} describes the evolution of a pressureless fluid with given initial density and velocity. In fact, any sufficiently regular solution $(\omega,v)$ of problem \eqref{eq:w2dynamic} on the time interval $[0,1]$ also solves \eqref{eq:pressureless}, since the absence of shocks implies that the Hamilton-Jacobi equation is equivalent to the conservation of momentum, i.e.\ the second equation in \eqref{eq:pressureless}. Moreover, dissipative solutions to such system, i.e.\ for which the kinetic energy $\mc{E}:[0,\infty) \rightarrow \mathbb{R}_+$ given by
\[
\mc{E}(t)\coloneqq \int_{\mathbb{R}^d} \omega(t)|v(t)|^2
\]
is nonincreasing,
provide a dissipative notion of extrapolation, since by equation \eqref{eq:w2dynamic}, for any $\alpha =1+\beta>1$
\[
W^2_2(\mu_1,\omega(\alpha)) \leq \beta \int_1^\alpha \mathrm{d} t \int_{\mathbb{R}^d} \omega(t) |u(t)|^2 \leq \beta^2 \int_0^1 \mathrm{d} t \int_{\mathbb{R}^d} \omega(t) |u(t)|^2 = \beta^2 W^2_2(\mu_0,\mu_1)\,.
\]
Such solutions can be constructed by requiring a \emph{sticky collision} condition, which enforces particles to share the same position after their collision.
In dimension higher than one, few results exist on the well-posedness of system \eqref{eq:pressureless}, so we will not consider this case in detail. On the other hand, in dimension one, sticky solutions to system \eqref{eq:pressureless} have been widely studied in the literature. In particular, Brenier and Grenier \cite{brenier1998one} showed that one can construct solutions to \eqref{eq:pressureless} using the unique entropy solution of a scalar conservation law, and in particular a solution to \eqref{eq:pressureless} is given by the curve
\[
\omega(t) = \tilde{X}(t,\cdot)_\# \mu_0,
\]
with
\begin{equation}\label{eq:pushenvelope}
\tilde{X}(t,x) \coloneqq (\partial_x \mathrm{co}\, \psi(t,\cdot)) \circ F_0(x)\,, \quad \psi(t,s) \coloneqq \int_{0}^s X(t, F_0^{[-1]}(s')) \, \mathrm{d} s'\,,
\end{equation}
where $X$ is defined as in \eqref{eq:rhopush}, and $F_0^{[-1]}: [0,1] \rightarrow \overline{\mathbb{R}}$ is the quantile function of $\mu_0$, i.e.\ the pseudo-inverse of its comulative distribution function $F_0:x \rightarrow \int_{-\infty}^x \mathrm{d} \mu_0(x)$. Note that as long as the geodesic can be extended $\psi(t,\cdot)$ stays convex (as it is the integral of a monotone function) and therefore the definitions for $X(t,\cdot)$ and $\tilde{X}(t,\cdot)$, respectively in \eqref{eq:rhopush} and \eqref{eq:pushenvelope}, coincide. We will show that in this case the resulting notion of extrapolation coincide with that provided by the metric extrapolation, which is discussed in detail in the next section.
\end{remark}
\subsection{Metric extrapolation}\label{ssec:metricextra}
In analogy with the Euclidean case (see equation \eqref{eq:euclideanvar}), one can adopt a variational definition for the extrapolation, which we refer to as \emph{metric extrapolation}, and which is defined for all $\alpha>1$ and for all $\mu_0,\mu_1 \in\mc{P}_2(\mathbb{R}^d)$ by
\begin{equation}\label{eq:metricextraF}
\mathsf{E}_\alpha(\mu_0,\mu_1) \coloneqq \underset{\rho \in \mathcal{P}_2(\mathbb{R}^d)}{\mathrm{argmin}} ~ \mc{F}(\mu_0,\mu_1; \rho) \,, \quad \mc{F}(\mu_0,\mu_1; \rho) \coloneqq \alpha W_2^2(\rho,\mu_1) - \beta W_2^2(\rho,\mu_0)\,,
\end{equation}
where $\beta =\alpha -1$. In Proposition \ref{prop:existence} we will show that problem \eqref{eq:metricextraF} admits indeed a unique solution, which justifies the definition of the metric extrapolation.
\begin{remark}\label{rem:omega} Alternatively, one can define the metric extrapolation as in equation \eqref{eq:metricextra}, via a minimization on probability measures in $\mc{P}(\Omega)$ over a given domain $\Omega$. In this case, differently from the free-flow case \eqref{eq:lagextra}, the support of the extrapolated measures is always contained in $\bar{\Omega}$. The results of this section hold also in this case without major changes.
\end{remark}
First of all, we observe that by the triangular and Young's inequalities, for any $\rho, \mu_0, \mu_1 \in \mc{P}_2(\mathbb{R}^d)$
\[
W^2_2(\rho,\mu_0) \leq \left(1+\frac{1}{\beta}\right) W^2_2(\rho,\mu_1) + (1+\beta) W^2_2(\mu_0,\mu_1)
\]
and therefore
\begin{equation}\label{eq:boundbelow}
\mc{F}(\mu_0,\mu_1; \rho) \geq - \alpha \beta W^2_2(\mu_0,\mu_1).
\end{equation}
Then, if there exists a unique geodesic \eqref{eq:brenier2} from $\mu_0$ to $\mu_1$ and this can be continued up to time $\alpha$, i.e.\ if the associated Brenier potential $u$ is $\beta/\alpha$-convex, then the lower bound is attained only by $\rho=\omega(\alpha)$ with
\[
\omega(\alpha) = ((1-\alpha) \mathrm{Id} + \alpha \nabla u)_\# \mu_0,
\]
since by equation \eqref{eq:geodesic}
\[
W^2_2(\mu_0,\omega(\alpha)) = \alpha^2 W^2_2(\mu_0,\mu_1) \,, \quad
W^2_2(\mu_1,\omega(\alpha)) = \beta^2 W^2_2(\mu_0,\mu_1)\,.
\]
\begin{remark} Note that if the geodesic extension is only locally (but not globally) minimising, then it may not be recovered as a solution of problem \eqref{eq:metricextraF}: for instance, this is the case for the shear flow example \ref{ex:shear}, in which case one can compute the explicit solution to the metric extrapolation problem, which is represented in Figure \ref{fig:shear}.
\end{remark}
\begin{figure}
\begin{overpic}[scale=.9]{shear.pdf}
\put(205,74){$\mu_0$}
\put(205,52){$\mu_1$}
\put(142,44){$x$}
\put(82,107){$y$}
\put(205,30){$\mu_\alpha$}
\end{overpic}
\caption{Metric extrapolation in the setting of Example \ref{ex:shear}. The black solid line connecting the support of the three measures represents the trajectory followed by the extrapolated measure for different values of the parameter $\alpha$.}\label{fig:shear}
\end{figure}
Existence and uniqueness for minimizers of problem \eqref{eq:metricextraF} actually hold in general due to the fact that the functional $\mathcal{F}$ is strongly convex along particular curves known as generalized geodesics. To describe such curves we will need the following lemma:
\begin{lemma}[Lemma 5.3.2 in \cite{ambrosio2008gradient}] \label{lem:glue} Let $\gamma_{0,1}, \gamma_{0,2} \in \mc{P}_2(\mathbb{R}^d\times \mathbb{R}^d)$ be two plans with the same first marginal, i.e.\ such that for all $\varphi \in C^0_b(\mathbb{R}^d)$
\[
\int \varphi(x_0) \mathrm{d} \gamma_{0,1}(x_0,x_1) = \int \varphi(x_0) \mathrm{d} \gamma_{0,2}(x_0,x_2) \,.
\]
Then, there exists a plan $\gamma \in \mc{P}_2(\mathbb{R}^d\times \mathbb{R}^d\times \mathbb{R}^d)$ such that for all $\psi \in C^0_b(\mathbb{R}^d\times \mathbb{R}^d)$
\begin{equation}\label{eq:gammagen}
\begin{aligned}
\int \psi (x_0,x_1) \mathrm{d} \gamma(x_0,x_1,x_2) &= \int \psi(x_0,x_1) \mathrm{d} \gamma_{0,1} (x_0,x_1) \,, \\
\int \psi(x_0,x_2) \mathrm{d} \gamma(x_0,x_1,x_2) &= \int \psi(x_0,x_2) \mathrm{d} \gamma_{0,2} (x_0,x_2) \,.
\end{aligned}
\end{equation}
\end{lemma}
Given three measures $\nu_0, \nu_1, \nu_2 \in \mc{P}_2(\mathbb{R}^d)$, let $\gamma_{0,1}\in \mc{P}_2(\mathbb{R}^d\times \mathbb{R}^d)$ and $\gamma_{0,2}\in \mc{P}_2(\mathbb{R}^d\times \mathbb{R}^d)$ optimal transport plans from $\nu_0$ to $\nu_1$ and from $\nu_0$ to {$\nu_2$}, respectively. A generalized geodesic from $\nu_1$ to $\nu_2$ with base $\nu_0$ is a curve $\omega:[0,1]\rightarrow\mathcal{P}_2(\mathbb{R}^d)$ satisfying, for all $\varphi\in C^0_b(\mathbb{R}^d)$,
\[
\int \varphi \omega(t) = \int \varphi(x_1(1-t) +x_2 t) \mathrm{d} \gamma (x_0,x_1,x_2)
\]
where $\gamma \in \mathcal{P}_2(\mathbb{R}^d\times \mathbb{R}^d \times \mathbb{R}^d)$ is a plan verifying \eqref{eq:gammagen}.
The existence of such a plan is a consequece of Lemma \ref{lem:glue}. In the case where $\nu_0$ is absolutely continuous, denoting by $T_{0,1}$ and $T_{0,2}$ the optimal transport plans from $\nu_0$ to $\nu_1$ and from $\nu_0$ to $\nu_2$ respectively, there exists a unique generalized geodesic from $\nu_1$ to $\nu_2$ with base $\nu_0$ which is given by
\begin{equation}\label{eq:gengeo}
\omega(t) = ((1-t) T_{0,1} +t T_{0,2})_\#\nu_0\,.
\end{equation}
A functional $\mathcal{J}:\mathcal{P}_2(\mathbb{R}^d)\rightarrow \mathbb{R}$ is $\lambda$-convex along generalized geodesics based in $\nu_0$, if for all $\nu_1$ to $\nu_2$ and for all generalized geodesics $\omega :[0,1]\rightarrow\mathcal{P}_2(\mathbb{R}^d)$ from $\nu_1$ to $\nu_2$ with base $\nu_0$,
\begin{equation}\label{eq:lambgenconv}
\mathcal{J}(\omega(t)) \leq (1-t) \mathcal{J}(\nu_1) + t \mathcal{J}(\nu_2) - \lambda \frac{t(1-t)}{2} \int |x_1-x_2|^2 \mathrm{d} \gamma(x_0,x_1,x_2)
\end{equation}
with $\gamma$ satisfying equation \eqref{eq:gammagen}. {We say that the functional $\mathcal{J}$ is $\lambda$-convex along generalized geodesics if the previous definition holds true for any $\nu_0\in\mathcal{P}_2(\mathbb{R}^d)$.}
The following result was proven in \cite{Matthes2019bdf2} and provides the strong convexity of the functional $\mc{F}$ along generalized geodesics. For completeness, we include the proof in our setting.
\begin{lemma}[Theorem 3.4 in \cite{Matthes2019bdf2}]\label{lem:abconvex}
For any $\mu_0,\mu_1 \in \mc{P}_2(\mathbb{R}^d)$, the functional $\mathcal{F}(\mu_0,\mu_1; \cdot):\mathcal{P}_2(\mathbb{R}^d) \rightarrow\mathbb{R}$ defined in \eqref{eq:metricextraF} is 2-convex along generalized geodesics based in $\mu_1$. In particular, for any $\mu_2, \mu_3 \in \mc{P}_2(\mathbb{R}^d)$ there exists a curve $\omega:[0,1]\rightarrow \mathcal{P}_2(\mathbb{R}^d), \omega(0)=\mu_2, \omega(1)=\mu_3$, such that for all $t\in[0,1]$, it holds:
\begin{equation}\label{eq:abconvex}
\mathcal{F}(\mu_0,\mu_1; \omega(t)) \le (1-t) \mathcal{F}(\mu_0,\mu_1; \mu_2) + t \mathcal{F}(\mu_0,\mu_1; \mu_3)- t(1-t)W^2_2(\mu_2,\mu_3).
\end{equation}
\end{lemma}
\begin{proof} We take as $\omega$ any generalized geodesic based in $\mu_1$ and from $\mu_2$ to $\mu_3$. In particular, for all $\psi \in C^0_b(\mathbb{R}^d)$,
\[
\int \varphi \,\omega(t) = \int \varphi(x_2(1-t) +x_3 t) \mathrm{d} \gamma (x_1,x_2,x_3)
\]
where $\gamma \in \Pi(\mu_1,\mu_2,\mu_3)$ verifies for all $\psi \in C^0_b(\mathbb{R}^d\times\mathbb{R}^d)$,
\[
\begin{aligned}
\int \psi (x_1,x_2) \mathrm{d} \gamma(x_1,x_2,x_3) &= \int \psi(x_1,x_2) \mathrm{d} \gamma_{1,2} (x_1,x_2) \,, \\
\int \psi(x_1,x_3) \mathrm{d} \gamma(x_1,x_2,x_3) &= \int \psi(x_1,x_3) \mathrm{d} \gamma_{1,3} (x_1,x_3) \,,
\end{aligned}
\]
with $\gamma_{1,2}$ and $\gamma_{1,3}$ being optimal plans from $\mu_1$ to $\mu_2$ and from $\mu_1$ to $\mu_3$, respectively.
For a fixed $t\in(0,1)$, we denote by $\pi_{1,t}:\mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^d \rightarrow \mathbb{R}^d \times \mathbb{R}^d$ the map defined by $\pi_{1,t} (x_1,x_2,x_3) = (x_1, (1-t)x_2+ tx_3)$. Then $\gamma_{1,t}\coloneqq (\pi_{1,t})_\# \gamma \in \Pi(\mu_1,\omega(t))$ and therefore:
\begin{equation}\label{eq:est1}
\begin{aligned}
W^2_2(\mu_1,\omega(t)) & \leq \int |x_t - x_1|^2 \mathrm{d} \gamma_{1,t} (x_1,x_t) \\ &= \int |(1-t) x_2 +t x_3 - x_1|^2 \mathrm{d} \gamma(x_1,x_2,x_3) \\ &= (1-t) W^2(\mu_1,\mu_2) + t W^2(\mu_1,\mu_3) - t(1-t) \int |x_2 - x_3 |^2 \mathrm{d} \gamma(x_1,x_2,x_3)\,.
\end{aligned}
\end{equation}
On the other hand, again for a given $t\in(0,1)$, let us define the plan ${\gamma}_{t,3} \in \Pi(\omega(t),\mu_3)$, satisfying for all $\psi\in C^0_b(\mathbb{R}^d\times \mathbb{R}^d)$:
\begin{equation}\label{eq:gammat2}
\int \psi(x_t,x_3) \mathrm{d}{\gamma}_{t,3}(x_t,x_3) = \int \psi((1-t)x_2 + t x_3, x_3) \mathrm{d} \gamma(x_1,x_2,x_3).
\end{equation}
Denoting by $\gamma_{t,0}$ an optimal transport plan from $\omega(t)$ to $\mu_0$, by Lemma \ref{lem:glue}, there always exist a plan $\eta\in \Pi(\omega(t),\mu_3,\mu_0)$ such that for all $\psi\in C^0_b(\mathbb{R}^d\times \mathbb{R}^d)$:
\begin{equation}\label{eq:eta}
\begin{aligned}
\int \psi (x_t,x_3) \mathrm{d} \eta(x_t,x_3,x_0) &= \int \psi(x_t,x_3) \mathrm{d} \gamma_{t,3} (x_t,x_3) \,, \\
\int \psi(x_t,x_0) \mathrm{d} \eta(x_t,x_3,x_0) &= \int \psi(x_t,x_0) \mathrm{d} \gamma_{t,0} (x_t,x_0) \,.
\end{aligned}
\end{equation}
Let $\Lambda:\mathbb{R}^{3d} \rightarrow \mathbb{R}^{3d}$ be the diffeomorphism defined by $\Lambda(x_2,x_3,x_0) = ((1-t)x_2 + t x_3,x_3,x_0)$ and denote $\tilde{\eta} = \Lambda^{-1}_\# \eta$. Note that by equations \eqref{eq:gammat2} and \eqref{eq:eta}, $\tilde{\eta} \in \Pi(\mu_2,\mu_3,\mu_0)$. Hence, we have that
\begin{equation}\label{eq:est2}
\begin{aligned}
W^2_2(\mu_0,\omega(t)) & = \int |x_t - x_0|^2 \mathrm{d} \eta(x_t,x_3,x_0) \\ &= \int | (1-t)x_2 + t x_3 - x_0|^2 \mathrm{d} \tilde{\eta}(x_2,x_3,x_0)\\
& = \int \left[ (1-t) | x_2 - x_0|^2 + t | x_3 - x_0|^2 - t(1-t) |x_2-x_3|^2 \right] \mathrm{d} \tilde{\eta}(x_2,x_3,x_0)\\
& \geq (1-t) W^2_2(\mu_0,\mu_2) + t W^2_2(\mu_0,\mu_3) - t(1-t) \int |x_2 - x_3|^2 \mathrm{d} \gamma(x_1,x_2,x_3).
\end{aligned}
\end{equation}
We obtain the result combining the estimates in \eqref{eq:est1} and \eqref{eq:est2}, and using the relation $\alpha -\beta = 1$. Inequality \eqref{eq:abconvex} is finally obtained bounding the transport cost from $\mu_2$ to $\mu_3$ by $W^2_2(\mu_2,\mu_3)$.
\end{proof}
\begin{proposition}\label{prop:existence}
The metric extrapolation problem \eqref{eq:metricextraF} admits a unique solution $\mu_\alpha$. Moreover, the metric extrapolation is $\beta$-dissipative, i.e.
\begin{equation}\label{eq:betadisalp}
W_2(\mu_1, \mu_\alpha) \leq \beta W_2(\mu_0,\mu_1)\,,
\end{equation}
and for all $\mu\in\mathcal{P}_2(\mathbb{R}^d)$,
\begin{equation}
\label{eq:rhoab_ineq}
{W}_2^2(\mu,\mu_\alpha) + \mathcal{F}(\mu_0,\mu_1; \mu_\alpha) \le \mathcal{F}(\mu_0,\mu_1; \mu )\,.
\end{equation}
\end{proposition}
\begin{proof}
The functional $\mathcal{F}$ is bounded from below by \eqref{eq:boundbelow} and it is stongly convex along generalized geodesics by Lemma \ref{lem:abconvex}. This latter fact implies uniqueness of the solution.
Regarding existence, notice first that we need to resort to Prokhorov's theorem for compactness in the weak-* topology, since the base space is not compact (see \cite[Chapter 1]{santambrogio2015optimal}).
Second, the Wasserstein distance on $\mc{P}_2(\mathbb{R}^d)$ is not continuous with respect the weak-* topology, but it is continuous along sequences of probability measures whose second-order momenta are converging as well \cite[Theorem 5.11]{santambrogio2015optimal}. Convergence of momenta can be established for a minimizing sequence of $\mathcal{F}$ thanks again to strong convexity. Let $(\mu^n)_n$ be a minimizing sequence. For $N\in\mathbb{N}$ and $n>N$, consider a generalized geodesic $\omega$ as in Lemma \ref{lem:abconvex}, based in $\mu_1$ and with $\omega(0) =\mu^n$ and $\omega(1)=\mu^N$. From \eqref{eq:abconvex} and the bound \eqref{eq:boundbelow}, we obtain for a fixed $t\in(0,1)$:
\[
\begin{aligned}
t(1-t){W}_2^2(\mu^n,\mu^N) & \le (1-t) \mathcal{F}(\mu_0,\mu_1; \mu^n)+t\mathcal{F}(\mu_0, \mu_1; \mu^N) + \alpha\beta W^2_2(\mu_0,\mu_1) \\
& \le \mathcal{F}(\mu_0,\mu_1; \mu^N) + \alpha\beta W^2_2(\mu_0,\mu_1) \,.
\end{aligned}
\]
As this inequality holds for all $n>N$, this implies that the sequence of momenta of $\mu^n$ is uniformly bounded.
In turn, this further implies that the sequence is tight. Hence, by Prokhorov's theorem we can extract a subsequence converging to $\mu_\alpha$. Up to extraction of another subsequence, the momenta converge as well and therefore $\mu_\alpha$ is the minimizer of $\mathcal{F}$.
Inequality \eqref{eq:rhoab_ineq} derives again from Lemma \ref{lem:abconvex}. For a given $\mu\in\mathcal{P}_2(\mathbb{R}^d)$, consider again a generalized geodesic $\omega$ as in Lemma \ref{lem:abconvex}, with $\omega(0) =\mu_\alpha$ and $\omega(1)=\mu$. By optimality of $\mu_\alpha$, it holds
\[
\begin{aligned}
0 &\le \mathcal{F}(\mu_0,\mu_1; \omega(t))-\mathcal{F}(\mu_0, \mu_1; \mu_\alpha) \\
&\le t\big(\mathcal{F}(\mu_0,\mu_1; \mu)-\mathcal{F}(\mu_0, \mu_1; \mu_\alpha )\big)- t(1-t){W}_2^2(\mu,\mu_\alpha)\,,
\end{aligned}
\]
which, dividing by $t$ and taking the limit $t\rightarrow0$, gives \eqref{eq:rhoab_ineq}. Using \eqref{eq:boundbelow} on the left-hand side of \eqref{eq:rhoab_ineq} and then taking $\mu=\mu_1$, we obtain the estimate \eqref{eq:betadisalp}.
\end{proof}
In order to prove the consistency assumption we will use the following optimality conditions for problem \eqref{eq:metricextraF}.
\begin{lemma}\label{lem:optgamma} Let $\mu_\alpha$ be the unique solution to problem \eqref{eq:metricextraF}.
There exist two optimal transport plans $\gamma_{0,\alpha}$ and $\gamma_{1,\alpha}$ from $\mu_0$ to $\mu_\alpha$ and from $\mu_1$ to $\mu_{\alpha}$, respectively, such that
\begin{equation}\label{eq:rhoab_OC}
\alpha \int \langle x_\alpha - x_1, {\xi}(x_\alpha) \rangle \text{d} \gamma_{1,\alpha} (x_1,x_\alpha)-\beta \int \langle x_\alpha - x_0, {\xi}(x_\alpha) \rangle \text{d} \gamma_{0,\alpha} (x_0,x_\alpha) = 0 \,,
\end{equation}
for any ${\xi}\in\C^\infty_c(\mathbb{R}^d;\mathbb{R}^d)$.
\end{lemma}
\begin{proof}
In order to prove the result we construct a sequence of approximated smooth variational problems and pass to the limit in the optimality conditions.
Let us define for $\varepsilon>0$,
\begin{equation}\label{eq:Feps}
\mathcal{F}_{\varepsilon}(\mu_0,\mu_1; \mu) \coloneqq \mathcal{F}(\mu_0,\mu_1; \mu) + \varepsilon \, \mathcal{U}(\mu) \,,
\end{equation}
where $\mc{U}$ denotes the entropy defined in \eqref{eq:entropy} (replacing $\Omega$ with $\mathbb{R}^d$),
and the regularized problem
\begin{equation}\label{eq:defrhoabEps}
\inf_{\rho\in\mathcal{P}_2(\mathbb{R}^d)} \mathcal{F}_{\varepsilon}(\mu_0,\mu_1; \rho) \,.
\end{equation}
Problem \eqref{eq:defrhoabEps} admits a unique solution $\mu_\alpha^\varepsilon$ which is now absolutely continuous. The proof is the same as in Proposition \ref{prop:existence} since $\mathcal{U}$ is convex along generalized geodesics, bounded from below and lower semi-continuous for the weak-* topology. By applying Lemma \ref{lem:variations_ac} (adapted to the case where $\Omega =\mathbb{R}^d$), we can write down the necessary optimality conditions of problem \eqref{eq:defrhoabEps}:
\begin{multline}\label{eq:defrhoabEps_oc}
\left. \frac{\text{d} \mathcal{F}_{\varepsilon}(\mu_0,\mu_1; \omega(s))}{\text{d} s}\right|_{s=0} = 2\alpha \int \langle x_\alpha - x_1, {\xi}(x_\alpha) \rangle \text{d} \gamma^\varepsilon_{1,\alpha} (x_1,x_\alpha)\\-2\beta \int \langle x_\alpha - x_0, {\xi}(x_\alpha) \rangle \text{d} \gamma^\varepsilon_{0,\alpha} (x_0,x_\alpha) -\varepsilon \int \nabla \cdot {\xi} (x) \mathrm{d} \mu_\alpha^{\varepsilon}(x) = 0 \,,
\end{multline}
for any ${\xi}\in\C^\infty_c(\mathbb{R}^d;\mathbb{R}^d)$, where $\omega:(-\delta,\delta)\rightarrow \mc{P}_2(\mathbb{R}^d)$ is the curve of measures defined by \eqref{eq:curve_var} with $\omega(0)=\mu_\alpha$, and where we now denote by $\gamma^\varepsilon_{0,\alpha}$ and $\gamma^\varepsilon_{1,\alpha}$ the optimal transport plans from $\mu_0$ to $\mu_\alpha^\varepsilon$ and from $\mu_1$ to $\mu_\alpha^\varepsilon$, respectively.
We want to show that problem \eqref{eq:defrhoabEps} $\Gamma$-converges towards problem \eqref{eq:metricextraF} in order to pass to the limit in the optimality conditions. By the lower semi-continuity of $\mathcal{F}$ and the fact that $\mathcal{U}$ is bounded from below, the $\Gamma$-$\liminf$ is obvious,
\[
\mathcal{F}(\mu_0,\mu_1; \rho) \le \liminf_{\varepsilon} \mathcal{F}(\mu_0,\mu_1; \rho_{\varepsilon}) \le \liminf_{\varepsilon}\mathcal{F}_{\varepsilon}(\mu_0,\mu_1; \rho_\varepsilon) \,,
\]
for any $\rho_{\varepsilon}\rightarrow\rho$ in the Wasserstein sense. Concerning the $\Gamma$-$\limsup$, if $\mathcal{U}(\mu_\alpha)<+\infty$ we can take $\rho_{\varepsilon}=\mu_\alpha$ as recovering sequence. Otherwise, since the set of absolutely continuous measures is dense in $\mathcal{P}_2(\mathbb{R}^d)$,
we can take a sequence of absolutely continuous measures $\rho_{\varepsilon}$ converging to $\mu_\alpha$ with respect to the Wasserstein metric.
Since $\mathcal{U}(\mu_\alpha)=\infty$, up to a reparametrization we can assume that the entropy is increasing and that
\[
\mathcal{U}(\rho_{\varepsilon})\le \frac{C}{\sqrt{\varepsilon}} \,,
\]
for a constant $C$ independent of $\varepsilon$.
Then it holds:
\[
\limsup_{\varepsilon} \mathcal{F}_{\varepsilon}(\mu_0,\mu_1; \rho_{\varepsilon}) = \lim_{\varepsilon} \mathcal{F}_{\varepsilon}(\mu_0,\mu_1; \rho_{\varepsilon}) = \mathcal{F}(\mu_0,\mu_1; \mu_\alpha) \,.
\]
Therefore problem \eqref{eq:defrhoabEps} $\Gamma$-converges to problem \eqref{eq:metricextraF}, which implies that $\mu_\alpha^\varepsilon \rightarrow\mu_\alpha$ in the Wasserstein sense. By the stability of optimal transport plans {\cite[Theorem 5.20]{villani2009optimal}}, there exist optimal plans $\gamma_{0,\alpha}$ and $\gamma_{1,\alpha}$ from $\mu_0$ to $\mu_\alpha$ and from $\mu_1$ to $\mu_{\alpha}$, respectively, such that (up to the extraction of a subsequence)
\[
\gamma_{0,\alpha}^\varepsilon \longrightarrow \gamma_{0,\alpha}\,, \quad \gamma_{1,\alpha}^\varepsilon \longrightarrow \gamma_{1,\alpha} \,,
\]
in the Wasserstein sense. As the vector field ${\xi}$ is smooth, passing to the limit in \eqref{eq:defrhoabEps_oc} we obtain \eqref{eq:rhoab_OC}.
\end{proof}
\begin{proposition}
The metric extrapolation defined via \eqref{eq:metricextraF} verifies the consistency assumption \eqref{eq:consistency}.
\end{proposition}
\begin{proof}
Using the same notation as in the statement of Lemma \ref{lem:optgamma}, we have that for all $\varphi \in C^\infty_c(\mathbb{R}^d)$
\[
\int \varphi\, ( \mu_\alpha - \alpha \mu_1 + \beta \mu_0) = \alpha \int ( \varphi(x_1) - \varphi(x_\alpha)) \mathrm{d} \gamma_{1,\alpha}(x_1,x_\alpha) - \beta \int ( \varphi(x_0) - \varphi(x_\alpha)) \mathrm{d} \gamma_{0,\alpha}(x_0,x_\alpha)\,.
\]
Using the Taylor expansion of $\varphi$ at $x_\alpha$ in both integrals on the right-hand side, Lemma \ref{lem:optgamma}, and the dissipation property \eqref{eq:betadisalp}, we obtain
\[
\begin{aligned}
\left| \int \varphi\, ( \mu_\alpha - \alpha \mu_1 + \beta \mu_0) \right| &\leq \|\mathrm{Hess}\varphi\|_\infty ( \alpha W^2_2(\mu_1,\mu_{\alpha}) + \beta W^2_2(\mu_0,\mu_{\alpha}) ) \\& \leq \alpha \beta \|\mathrm{Hess}\varphi\|_\infty W^2_2(\mu_0,\mu_1).
\end{aligned}
\]
\end{proof}
\begin{remark}[Relation with pressureless fluids]\label{rem:metricextra_L2}
In dimension one, the Wasserstein distance $W_2$ coincides with the $L^2$ distance between the quantile functions. In particular, the metric extrapolation $\mu_\alpha$ is given by
\[
\mu_\alpha = (G_\alpha)_\# \mathrm{d} x |_{[0,1]} \,, \quad G_\alpha \coloneqq \underset{\substack{G\in L^2([0,1],\mathbb{R})\\ \text{monotone}}}{\mathrm{argmin}} ~ \alpha \| G - F_1^{[-1]}\|^2_{L^2} - \beta \| G - F_0^{[-1]}\|^2_{L^2}\,,
\]
where $F_0^{[-1]}$ and $F_1^{[-1]}$ are the quantiles of $\mu_0$ and $\mu_1$, respectively. The solution to this problem coincides with the sticky particle model described in Remark \ref{rem:pressureless}, i.e.\ $G_\alpha = \tilde{X}(\alpha,\cdot)$ with $\tilde{X}$ as in \eqref{eq:pushenvelope}.
\end{remark}
\begin{remark}[Dual formulation of the metric extrapolation]\label{rem:metricextra_dual}
Let us recall that the optimal transport problem \eqref{eq:w2} admits the following dual formulation \cite[Theorem 5.10]{villani2009optimal}:
\begin{equation}\label{eq:w2dual}
\frac{W^2(\mu_0,\mu_1)}{2} = \sup_{\phi_0} \left\{ \int \mc{H}_1(\phi_0) \mu_1 - \int \phi_0 \mu_0 ~:~ \frac{|\cdot|^2}{2} + \phi_0(\cdot) ~ \text{ is convex}\right\} \,,
\end{equation}
and if $\mu_0$ is absolutely continuous, this admits a unique maximiser $\phi_0$, and $u(\cdot) \coloneqq \frac{|\cdot|^2}{2} + \phi_0(\cdot)$ is the Brenier potential from $\mu_0$ to $\mu_1$.
However, the associated geodesic from $\mu_0$ to $\mu_1$ can be extended up to time $\alpha>1$ only if \eqref{eq:convexityextra} holds, or equivalently if
\begin{equation}\label{eq:alphaconvex}
x\mapsto \frac{|x|^2}{2} + \alpha \phi_0(x) ~~\text{is convex}.
\end{equation} Therefore, in order to construct an extrapolation, one can instead consider the problem
\begin{equation}\label{eq:metricextra_dual}
\sup_{\phi_0} \left\{
\int \mc{H}_1(\phi_0) \mu_1 - \int \phi_0 \mu_0 ~:~ \frac{|\cdot|^2}{2} + \alpha \phi_0(\cdot) ~ \text{ is convex}
\right\},
\end{equation}
and, if $\mu_0$ is absolutely continuous, set
\[
{\sf E}_\alpha(\mu_0, \mu_1) = (\nabla u_\alpha)_\# \mu_0,
\]
where $u_\alpha(\cdot)\coloneqq\frac{|\cdot|^2}{2} + \alpha \phi_0(\cdot)$ and $\phi_0$ solves \eqref{eq:metricextra_dual}.
This extrapolation is well defined and it turns out to be a dual formulation for the metric extrapolation in the spirit of \cite{carlier2008Toland}. However, even if very natural, this dual point of view was not needed for the results presented here, and therefore it will be developed in a future work.
\end{remark}
\section{Convergence in the EVI sense}\label{ssec:EVI}
Let us now make the further assumption that the energy functional $\mathcal{E}$ is $\lambda$-convex in the generalized geodesic sense on $\mc{P}(\Omega)$, for $\lambda\in{\mathbb R}_+$ (see equation \eqref{eq:lambgenconv}, and recall that $\Omega$ is supposed to be convex, so generalized geodesics with endpoints in $\mc{P}(\Omega)$ are well-defined as curves on $\mc{P}(\Omega)$).
We recall that a curve $\varrho:[0,T]\rightarrow\Pc(\Om)$, $\varrho(0)=\rho_0$, is a Wasserstein gradient flow in the EVI sense if for any $\nu\in\Pc(\Om)$ it holds
\begin{equation}\label{eq:evi}
\frac{\text{d}}{\text{d} t} \frac{1}{2} {W}_2^2(\varrho(t),\nu) \le \mathcal{E}(\nu)-\mathcal{E}(\varrho(t)) - \frac{\lambda}{2} {W}_2^2(\varrho(t),\nu), \quad \forall t\in(0,T) \,,
\end{equation}
or, equivalently, if for all $r,s\in(0,T)$ with $r\le s$ it holds
\begin{equation}\label{eq:evi_integral}
\frac{1}{2} {W}_2^2(\varrho(s),\nu) -\frac{1}{2} {W}_2^2(\varrho(r),\nu)\le \mathcal{E}(\nu)(s-r)-\int_{r}^{s} (\mathcal{E}(\varrho(t)) + \frac{\lambda}{2} {W}_2^2(\varrho(t),\nu)) \text{d} t \,.
\end{equation}
In this section, we show that the limit curve extracted from the time discretization \eqref{eq:bdf2metric} using the metric extrapolation \eqref{eq:metricextraF} satisfies the inequality \eqref{eq:evi_integral}.
We first show that for scheme \eqref{eq:bdf2metric}-\eqref{eq:metricextraF} a discrete version of the inequality \eqref{eq:evi_integral} holds.
As the Wasserstein distance ${W}_2^2(\cdot,\rho^\alpha_{n-1})$ is $2$-convex along any generalized geodesic based in $\rho^\alpha_{n-1}$ (see, e.g., the proof of Lemma \ref{lem:abconvex}), the overall functional
\begin{equation}
\mathcal{G}(\rho_{n-1},\rho_{n-2};\rho) = \frac{{W}_2^2(\rho,\rho^\alpha_{n-1})}{2(1-\beta) \tau} + \mathcal{E}(\rho) \,,
\end{equation}
is $\frac{1}{(1-\beta) \tau}+\lambda>0$ convex along any generalized geodesic on $\mc{P}(\Omega)$ based in $\rho^\alpha_{n-1}$.
Note that in order to consider the case $\lambda<0$ one should explicitly add a restriction on the time step $\tau$ so that $\frac{1}{(1-\beta) \tau}+\lambda>0$.
\begin{lemma}
At each step $n$, for all $ \nu \in \mathcal{P}(\Omega)$, the following inequality holds:
\begin{multline}\label{eq:evi_discrete}
\Big(\frac{1}{2(1-\beta) \tau}+\frac{\lambda}{2}\Big){W}^2_2(\rho_{n},\nu) - \alpha \frac{{W}_2^2(\nu,\rho_{n-1})}{2(1-\beta)\tau} +\beta \frac{{W}_2^2(\nu,\rho_{n-2})}{2(1-\beta)\tau} \\ \le \mathcal{E}(\nu)-\mathcal{E}(\rho_{n}) +\alpha\beta \frac{W^2_2(\rho_{n-1},\rho_{n_2})}{2(1-\beta)\tau} -\frac{{W}_2^2(\rho_{n},\rho^\alpha_{n-1})}{2(1-\beta)\tau} \,.
\end{multline}
\end{lemma}
\begin{proof}
By the discussion above, considering the generalized geodesic $\omega$ between $\nu$ and $\rho_{n}$ with base $\rho^\alpha_{n-1}$, and using the optimality of $\rho_{n}$, we obtain
\[
\begin{aligned}
0&\le \mathcal{G}(\rho_{n-1},\rho_{n-2};\omega(t)) - \mathcal{G}(\rho_{n-1},\rho_{n-2};\rho_{n}) \\
&\le t (\mathcal{G}(\rho_{n-1},\rho_{n-2};\nu) - \mathcal{G}(\rho_{n-1},\rho_{n-2};\rho_{n})) -\frac{1}{2}\Big(\frac{1}{(1-\beta) \tau}+\lambda\Big)t(1-t){W}^2_2(\rho_{n},\nu).
\end{aligned}
\]
Dividing by $t$ and taking the limit $t\rightarrow0$, this yields
\[
\Big(\frac{1}{2(1-\beta) \tau}+\frac{\lambda}{2}\Big){W}^2_2(\rho_{n},\nu) -\frac{{W}_2^2(\nu,\rho^\alpha_{n-1})}{2(1-\beta)\tau} \le \mathcal{E}(\nu)-\mathcal{E}(\rho_{n})-\frac{{W}_2^2(\rho_{n},\rho^\alpha_{n-1})}{2(1-\beta)\tau} \,.
\]
Adding on both side the term $-\frac{1}{2(1-\beta)\tau}\mathcal{F}(\rho_{n-1},\rho_{n-2};\rho^\alpha_{n-1})$, using \eqref{eq:rhoab_ineq} on the left-hand side, we obtain
\begin{multline*}
\Big(\frac{1}{2(1-\beta) \tau}+\frac{\lambda}{2}\Big){W}^2_2(\rho_{n},\nu) - \alpha \frac{{W}_2^2(\nu,\rho_{n-1})}{2(1-\beta)\tau} +\beta \frac{{W}_2^2(\nu,\rho_{n-2})}{2(1-\beta)\tau} \\ \le \mathcal{E}(\nu)-\mathcal{E}(\rho_{n}) -\frac{1}{2(1-\beta)\tau}\mathcal{F}(\rho_{n-1},\rho_{n-2};\rho^\alpha_{n-1}) -\frac{{W}_2^2(\rho_{n},\rho^\alpha_{n-1})}{2(1-\beta)\tau} \,.
\end{multline*}
Finally, using \eqref{eq:boundbelow} on the right-hand side we conclude.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:convergenceevi}]
We recall that thanks to the classical estimate \eqref{eq:bound_W} (Lemma \ref{lem:bound_W}), the piecewise constant curve
\[
\rho_{\tau}(t) = \sum_{n=1}^{N} \rho_{n-1} \mathds{1}_{(t_{n-1},t_{n}]} \,, \quad \rho_{\tau}(0) = \rho_0 \,,
\]
converges uniformly in the ${W}_2$ distance to an absolutely continuous limit curve $\rhoe:[0,T]\rightarrow \mc{P}(\Omega)$ (see Proposition \ref{prop:conv_rho}).
In order to prove convergence of the scheme in the EVI sense, we show that this curve satisfies inequality \eqref{eq:evi_integral}. Thanks to the uniform convergence in time, the procedure is the same as in \cite[Theorem 5.1]{Matthes2019bdf2}.
For simplicity, assume that given $r,s\in(0,T), r\le s$, there exist $N_{\tau},M_{\tau}\in\mb{N}, N_{\tau}\le M_{\tau}$, such that $r=N_{\tau} \tau, s=M_{\tau} \tau$, $\forall \tau$.
We multiply by $\tau$ inequality \eqref{eq:evi_discrete} and sum over $n$ from $N_{\tau}$ to $M_{\tau}$ to obtain the discrete integral form of the EVI:
\begin{multline}\label{eq:EVIt}
\frac{1}{2(1-\beta)} \sum_{n=N_{\tau}}^{M_{\tau}}\left( {W}_2^2(\rho_{n},\nu) -\alpha {W}_2^2(\nu,\rho_{n-1}) +\beta {W}_2^2(\nu,\rho_{n-2}) \right) \\
\le \mathcal{E}(\nu)(t-s)-\sum_{n=N_{\tau}}^{M_{\tau}} \tau \Big(\mathcal{E}(\rho_{n})+\frac{\lambda}{2}{W}_2^2(\rho_{n},\nu)\Big) \\
+\frac{1}{2(1-\beta)} \sum_{n=N_{\tau}}^{M_{\tau}} \left( \alpha \beta {W}_2^2(\rho_{n-1},\rho_{n-2})-{W}_2^2(\rho_{n},\rho^\alpha_{n-1})\right) .
\end{multline}
By canceling out terms, the left-hand side is equal to
\begin{multline}\label{eq:EVIt_left}
\frac{1}{2(1-\beta)} \left(-\alpha {W}_2^2(\nu,\rho_{N_{\tau}-1}) + \beta {W}_2^2(\nu,\rho_{N_{\tau}-2})
+ \beta {W}_2^2(\nu,\rho_{N_{\tau}-1}) \right. \\ \left.+{W}_2^2(\rho_{M_{\tau}-1},\nu) +{W}_2^2(\rho_{M_{\tau}},\nu) -\alpha {W}_2^2(\nu,\rho_{{M_{\tau}-1}})\right) \,,
\end{multline}
and thanks to the uniform convergence in the Wasserstein distance, \eqref{eq:EVIt_left} converges to
\[
\frac{1}{2} {W}_2^2(\varrho(s),\nu) -\frac{1}{2} {W}_2^2(\varrho(r),\nu) \,,
\]
for $\tau\rightarrow0$, where we recall $\alpha-\beta =1$.
Concerning the right-hand side, thanks again to the uniform convergence in the Wasserstein distance, the lower semi-continuity of $\mathcal{E}$ and Fatou's lemma, we have
\[
\limsup_{n\rightarrow\infty} -\sum_{n=N_{\tau}}^{M_{\tau}} \tau \Big(\mathcal{E}(\rho_{n})+\frac{\lambda}{2}{W}_2^2(\rho_{n},\nu)\Big) \le
-\int_{r}^{s} \Big(\mathcal{E}(\rhoe(t)) + \frac{\lambda}{2} {W}_2^2(\rhoe(t),\nu)\Big) \text{d}\text{t} \,.
\]
Finally, owing to bound \eqref{eq:bound_W}, we estimate the last contribution of \eqref{eq:EVIt} as
\[
\sum_{n} \alpha \beta {W}_2^2(\rho_{n-1},\rho_{n-2})-{W}_2^2(\rho_{n},\rho^\alpha_{n-1}) \le \sum_{n} \alpha \beta {W}_2^2(\rho_{n-1},\rho_{n-2}) \le C \tau,
\]
which converges to zero. As a consequence, we recover the continuous inequality \eqref{eq:evi_integral}.
\end{proof}
\section{Finite volume discretization}\label{sec:LJKO2scheme}
In this section we describe a space-time discretization of the proposed approach which yields numerically second-order accuracy both in space and time.
We consider a discretization in the Eulerian framework of finite volumes. In this setting, neither the free-flow extrapolation nor the metric one have a straightforward implementation. For this reason, we will construct a discrete extrapolation operator based on formula \eqref{eq:varextra}: in this way the extrapolation step is cast in a variational way allowing for a robust implementation. Although not satisfying the hypotheses of theorem \eqref{th:convergencefp}, this choice leads to a convergent and second order accurate scheme, as we will show numerically.
As explained in Remark \ref{rem:viscosity+continuity}, the variational step \eqref{eq:varextra} differs from the direct forward integration of the continuity equation. This latter is a viable alternative to define a discrete extrapolation and leads to second order accuracy as well (see \cite{todeschi2021finite}), but it is not clear how to discretize this in a robust way.
The fundamental tool is the solution of JKO steps, which requires the expensive problem of computing the Wasserstein distance. Following \cite{cances2020LJKO,natale2020FVCA}, we linearize the Wasserstein distance obtaining LJKO steps, a more affordable problem to solve. Remarkably, this approach preserves the second order accuracy in time of our time discretization. The discretization in space is based instead on Two-Point Flux Approximation (TPFA) finite volumes with a centered choice for the mobility, which leads to simple and flexible schemes which are second order accurate in space.
\subsection{Discrete setting}
TPFA finite volumes require a sufficiently regular partitioning of the domain $\Omega$, according to \cite[Definition 9.1]{EGH00}.
For simplicity, we describe the methodology in two dimensions only, although generalizations to arbitrary dimensions are possible, and for $\Omega\subset \mathbb{R}^2$ being a polygonal domain.
The discretization of $\Omega$ consists of three sets: the set of cells $K\in\mathcal{T}$; the set of edges $\sigma\in\overline{\Sigma}$, which is composed of the two subsets of internal edges $\Sigma$ and external edges $\overline{\Sigma}\setminus\Sigma$; the set of cell centers ${(\mathbf{x}_K)}_{K\in\mathcal{T}}$.
We will denote the finite volume mesh as $\left(\mathcal{T},\overline{\Sigma}, {(\mathbf{x}_K)}_{K\in\mathcal{T}}\right)$.
The fundamental regularity hypothesis we need to construct TPFA schemes is the orthogonality between each internal edge $\sigma=K|L\in\Sigma$ and the segment $\mathbf{x}_L-\mathbf{x}_K$.
Typical example of meshes that can be used to this end are Cartesian grids, Voronoi tessellations and Delaunay triangulations, by taking the circumcenters of the polygonal cells as cell centers.
For each cell $K\in\mathcal{T}$, we denote $\Es_K$ and $\EsIn_K$ the subsets of edges and internal edges belonging to $K$, and by $m_K$ the measure of the cell. The mesh size $h$ is the largest among all cells' diameters, $h\coloneqq \max_{K\in\mathcal{T}} \text{diam}(K)$, and characterizes the refinement of the mesh. For every internal edge, the diamond cell $\Delta_\sigma$ is the quadrilateral with vertices given by the cell centers, $\mathbf{x}_K$ and $\mathbf{x}_L$, and the vertices of the edge. Denoting by $d_{\sigma}\coloneqq |\mathbf{x}_L-\mathbf{x}_K|$ and $m_\sigma$ the measure of the edge, the measure of the diamond cell is equal to $m_{\Delta_\sigma}=\frac{m_\sigma d_\sigma}{d}$, where $d$ stands for the space dimension. Finally, we denote by $d_{K,\sigma}$ the Euclidean distance between the cell center $\mathbf{x}_K$ and the midpoint of the edge $\sigma\in\Es_K$, and by $\bs{n}_{K,\sigma}$ the outward unit normal of the cell $K$ on the edge $\sigma$.
The finite volume methodology introduces two levels of discretization, on cells and edges.
The first one is used to discretize scalar quantities whereas the second one for vectorial ones. To this end, we introduce three discrete inner product spaces $({\mathbb R}^{\mathcal{T}},\langle \cdot, \cdot \rangle_{\mathcal{T}}), \, ({\mathbb R}^{\Sigma},\langle \cdot, \cdot \rangle_{\Sigma})$ and $(\mathbb{F}_\Cs,\langle \cdot, \cdot \rangle_{\mb{F}_{\mathcal{T}}})$. The scalar products $\langle \cdot, \cdot \rangle_{\mathcal{T}}$ and $\langle \cdot, \cdot \rangle_{\Sigma}$ are defined as
\[
\begin{aligned}
&\langle \cdot, \cdot \rangle_{\mathcal{T}}: (\boldsymbol{a},\boldsymbol{b})\in [{\mathbb R}^{\mathcal{T}}]^2 \mapsto \sum_{K\in\mathcal{T}} a_K b_K m_K \,, \\
&\langle \cdot, \cdot \rangle_{\Sigma}: (\boldsymbol{u},\boldsymbol{v})\in [{\mathbb R}^{\Sigma}]^2 \mapsto \sum_{\sigma\in\Sigma} u_{\sigma} v_{\sigma} m_{\sigma} d_{\sigma} \,.
\end{aligned}
\]
The space $\mathbb{F}_\Cs$ is the space of consevative fluxes, it is defined by
\begin{equation*}\label{eq:LJKO2space_fluxes}
\mathbb{F}_\Cs=\{\boldsymbol{F}=(F_{K,\sigma},F_{L,\sigma})_{\sigma\in\Sigma}\in\mb{R}^{2\Sigma}: F_{K,\sigma}+F_{L,\sigma}=0\} \,,
\end{equation*}
and its scalar product is
\[
\langle \cdot, \cdot \rangle_{\mb{F}_{\mathcal{T}}}: (\bs{F},\bs{G})\in [\mb{F}_{\mathcal{T}}]^2 \mapsto \sum_{\sigma\in\Sigma} (F_{K,\sigma} G_{K,\sigma}+F_{L,\sigma} G_{L,\sigma}) \frac{m_{\sigma}d_{\sigma}}{2} \,.
\]
Note that the space $\mathbb{F}_\Cs$ is defined on internal edges only. This is sufficient, since we are dealing with no flux boundary value problems, and therefore we can neglect the flux variables on the boundary.
We denote $F_{\sigma} = |F_{K,\sigma}| = |F_{L,\sigma}|$ the modulus of the flux on each internal edge $\sigma=K|L\in\Sigma$ and, by convention, $|\boldsymbol{F}| = (F_{\sigma})_{\sigma\in\Sigma} \in \mb{R}^{\Sigma}$ and $|\boldsymbol{F}|^2 = (F_{\sigma}^2)_{\sigma\in\Sigma} \in \mb{R}^{\Sigma}$, for $\boldsymbol{F}\in\mb{F}_{\mathcal{T}}$.
According to finite volumes, the discrete divergence operator $\mathrm{div}_{\mathcal{T}}: \mb{F}_\mathcal{T} \rightarrow {\mathbb R}^\mathcal{T}$ is defined in an integral sense as
\[
(\mathrm{div}_{\mathcal{T}} \bs{F})_K \coloneqq \mathrm{div}_{K} \bs{F} \coloneqq \frac{1}{m_K} \sum_{\sigma\in\Sigma_K} F_{K,\sigma} m_{\sigma}\,,
\]
that is, for each cell, the discrete divergence is computed as the sum of the fluxes across its boundary.
The discrete gradient $\nabla_\Sigma: \mb{R}^{\mathcal{T}} \rightarrow \mb{F}_\mathcal{T}$ is defined by duality, requiring that $\langle\nabla_\Sigma \bs{a}, \bs{F} \rangle_{\mb{F}_\mathcal{T}}= - \langle \bs{a} , \mathrm{div}_\mathcal{T} \bs{F}\rangle_{\mathcal{T}}$, for all $\bs{a}\in{\mathbb R}^{\mathcal{T}}$ and $\bs{F}\in\mathbb{F}_\Cs$. Then, it holds
\[
(\nabla_\Sigma \bs{a})_{K,\sigma} \coloneqq \mathrm{\nabla}_{K,\sigma} \bs{a} \coloneqq \frac{a_L -a_K}{d_\sigma} \,.
\]
Both the discrete divergence and gradient operators automatically inherit the zero flux boundary condition from the definition of $\mathbb{F}_\Cs$.
The space $({\mathbb R}^{\Sigma},\langle \cdot, \cdot \rangle_{\Sigma})$ is introduced in order to match the two different discretizations on cells and edges. In order to reconstruct variables defined on cells to the edges, and viceversa, we need two reconstruction operators.
We use a centered reconstruction for the mobility in order to attain the second order accuracy in space.
To this end, we use the weighted arithmetic average operator $\mathcal{L}_{\Sigma}:{\mathbb R}^{\mathcal{T}}\rightarrow{\mathbb R}^{\Sigma}$ and its adjoint $\mathcal{L}^*_{\Sigma}:{\mathbb R}^{\Sigma}\rightarrow{\mathbb R}^{\mathcal{T}}$ (with respect to the two scalar products):
\begin{equation}\label{eq:reconstruction}
(\mathcal{L}_\Sigma \bs{a})_{\sigma}\coloneqq \lambda_{K,\sigma} a_K +\lambda_{L,\sigma} a_L \,, \quad ( \mathcal{L}_\Sigma^* \bs{u})_K \coloneqq \sum_{\sigma\in\Sigma_K} \lambda_{K,\sigma} u_{\sigma} \frac{m_{\sigma}d_{\sigma}}{m_K}\,,
\end{equation}
for $\bs{a}\in{\mathbb R}^{\mathcal{T}}$ and $\bs{u}\in{\mathbb R}^{\Sigma}$, with $\lambda_{K,\sigma}+\lambda_{L,\sigma}=1, \forall \sigma=K|L\in\Sigma$. Two possible choices for the weights are $(\lambda_{K,\sigma},\lambda_{L,\sigma}) = (\frac{d_{K,\sigma}}{d_{\sigma}},\frac{d_{L,\sigma}}{d_{\sigma}})$ or $(\frac{1}{2},\frac{1}{2})$, both leading to second order accurate schemes in space \cite{natale2020FVCA}. The former choice is possible only if $\mathbf{x}_K\in K$, which may not be always the case for arbitrary admissible meshes.
\begin{remark}
The definition of the reconstruction operators and the choice of weights may be delicate in general for the discretization of dynamical optimal transport, depending on the discretization chosen for $\Omega$. See \cite{gladbach2018scaling,natale2021computation} for details. Notice in particular that the choice $(\lambda_{K,\sigma},\lambda_{L,\sigma})=(\frac12,\frac12)$ may lead to convergence failure in very simple settings \cite[Section 5]{gladbach2018scaling}. Nevertheless, in the context of the discretization of Wasserstein gradient flows the definition of the reconstruction is more flexible, see \cite{cances2020LJKO,forkert2020evolutionary}.
\end{remark}
\subsection{Discrete ${\stackrel{.}{H}}{}^{-1}$ norm}
As suggested in \cite{lavenant2018dynamical,erbar2020computation,natale2021computation}, a convenient choice for the time discretization of the Wasserstein distance \eqref{eq:w2dynamic} is to use a staggered time discretization for the velocity and the density on subintervals of the time interval $[0,1]$, and reconstruct the density on intermediate steps via arithmetic average. It has been shown numerically in \cite{cances2020LJKO,natale2020FVCA} that a single step discretization on the whole interval is sufficient in order to preserve the first-order accuracy of the JKO scheme \eqref{eq:jko}.
Following the same ideas, here we approximate the Wasserstein distance between two measures $\mu,\nu\in\Pc(\Om)$ as
\begin{equation}\label{eq:Hm1norm}
\frac{1}{2} W_2^2(\mu,\nu)
\approx \sup_{\phi} \int_{\Om} \phi(\mu-\nu)-\frac{1}{2}\int_{\Om} \Big(\frac{\mu+\nu}{2}\Big) |\nabla \phi|^2 \,.
\end{equation}
Formula \eqref{eq:Hm1norm} is obtained by discretizing in one step problem \eqref{eq:w2dynamic} and by applying a duality result thanks to the change of variables $(\omega,v)\mapsto(\omega,\omega v)$. For more details on this construction see \cite{cances2020LJKO,natale2020FVCA}. This approximation consists in replacing the Wasserstein distance with the weighted dual norm $\frac{1}{2}||\mu-\nu||_{\dot{H}_{\frac{\mu+\nu}{2}}^{-1}}$. The choice of the arithmetic average of the two measures as weight is fundamental in order to achieve second order accuracy in time for the scheme we will propose in the following.
Using the finite volume discretization introduced above we can provide a discrete analogous of the weighted norm.
Given the discrete measures $\boldsymbol{\mu},\boldsymbol{\nu}\in{\mathbb R}^{\mathcal{T}}_+$ and for any ${\boldsymbol h}\in{\mathbb R}^{\mathcal{T}}$, the discrete counterpart of the weighted $\dot{H}^{-1}$ norm squared is
\begin{equation}\label{eq:Adiss_dual}
\mathcal{A}_{\mathcal{T}}\Big(\frac{\boldsymbol{\mu}+\boldsymbol{\nu}}{2};{\boldsymbol h}\Big) \coloneqq \sup_{\boldsymbol{\phi}\in{\mathbb R}^{\mathcal{T}}} \langle {\boldsymbol h}, \boldsymbol{\phi} \rangle_{\mathcal{T}}- \frac{1}{2} \Big\langle \mathcal{L}_\Sigma \Big(\frac{\boldsymbol{\mu}+\boldsymbol{\nu}}{2}\Big), |\nabla_\Sigma \boldsymbol{\phi}|^2 \Big\rangle_{\Sigma} \,.
\end{equation}
A few remarks are in order about such a discretization.
\begin{itemize}
\item For any $\boldsymbol{\rho}\in{\mathbb R}^{\mathcal{T}}_+$, the function $\mathcal{A}_{\mathcal{T}}(\boldsymbol{\rho};\cdot)$ is proper, convex and lower semi-continuous as supremum of convex and lower semi-continuous functions.
\item The supremum is unbouded if the condition $\langle {\boldsymbol h} , \boldsymbol{1} \rangle_{\mathcal{T}}=0$ is not satisfied. On other hand, if $\langle {\boldsymbol h} , \boldsymbol{1} \rangle_{\mathcal{T}}=0$, there exists a maximizer $\boldsymbol{\phi}$, which is however not uniquely defined, since the function maximised in \eqref{eq:Hm1norm} is invariant with respect to addition of a global constant or perturbations sufficiently far from the support of ${\boldsymbol h}$, $\boldsymbol{\mu}$, and $\boldsymbol{\nu}$.
\item Setting ${\boldsymbol h} = \boldsymbol{\nu} - \boldsymbol{\mu}$ in \eqref{eq:Adiss_dual}, with $\boldsymbol{\mu}$ and $\boldsymbol{\nu}$ being a discrete approximation of two measures $\mu$ and $\nu$, we obtain a discrete version of $W_2^2(\mu,\nu)/2$. In this case the optimal potential $\boldsymbol{\phi}$
can be interpreted as a discrete counterpart of a continuous optimal potential $\phi$, satisfying the Hamilton-Jacobi equation on the time interval $[0,1]$, evaluated at time $1/2$.
\item The total kinetic energy is discretized on the diamond cells.
Notice that due to the definition of the scalar product $\langle\cdot,\cdot\rangle_{\Sigma}$, the measure of each diamond cell is taken $m_{\sigma} d_{\sigma} = d m_{\Delta_{\sigma}}$, i.e. $d$ times the actual measure. This is done in order to compensate for the unidirectional discretization, since each term $|\nabla_{K,\sigma} \boldsymbol{\phi}|$ is meant as an approximation of the quantity $|\nabla \phi \cdot\bs{n}_{K,\sigma}|$, and have a consistent discretization. See \cite{natale2021computation} for more details on this construction.
\end{itemize}
\subsection{Discrete extrapolation}\label{sec:discreteextra}
We now construct a discrete version of the extrapolation operator $\mathsf{E}_\alpha$ at time $\alpha$, by discretizing the procedure described in Section \ref{sec:collisions}, and in particular of equation \eqref{eq:varextra}. The proposed strategy requires three subsequent steps: i) compute the interpolation between the two measures; ii) integrate forward in time the optimal potential; and finally iii) solve a JKO step.
Let us consider two discrete densities $\boldsymbol{\mu},\boldsymbol{\nu}\in{\mathbb R}^{\mathcal{T}}_+$ with the same total discrete mass $\langle \boldsymbol{\mu}, \boldsymbol{1} \rangle_{\mathcal{T}} = \langle \boldsymbol{\nu}, \boldsymbol{1} \rangle_{\mathcal{T}}$.
The first step requires to solve problem \eqref{eq:Adiss_dual} for ${\boldsymbol h}=\boldsymbol{\nu}-\boldsymbol{\mu}$ in order to find an optimal potential ${\boldsymbol{\phi}}$, which approximates the continuous one, solution to the Hamilton-Jacobi equation \eqref{eq:hj}, at the midpoint of the time interval $[0,1]$.
In the second step, we evolve the optimal potential according to the Hamilton-Jacobi equation until the final time $\alpha$, that is considering a temporal step of length $\frac{1}{2}+\beta=\frac{\alpha+\beta}{2}$. This can be done with an explicit Euler step as follows:
\begin{equation}\label{eq:HJevolve2}
\bphi^\alpha = {\boldsymbol{\phi}} - \frac{2}{\alpha+\beta}\frac{1}{2} \mathcal{L}^*_{\Sigma} |\nabla_\Sigma {\boldsymbol{\phi}}|^2 \,.
\end{equation}
Note that we use the operator $\mathcal{L}_{\Sigma}^*$ to reconstruct the square of the gradient of the potential. However, as this step is not variational, it is not mandatory to use the adjoint of the reconstruction $\mathcal{L}_{\Sigma}$ and any other (second order) strategy can be adopted.
Finally, for the third step, we approximate problem \eqref{eq:varextra} using again the discrete weighted $\dot{H}^{-1}$ norm. Specifically, we define a discrete extrapolation operator as a map $\mathsf{E}_\alpha^{\mathcal{T}}: {\mathbb R}^{\mathcal{T}}_+\times {\mathbb R}^{\mathcal{T}}_+\rightarrow {\mathbb R}^{\mathcal{T}}_+$ verifying
\begin{equation}\label{eq:rhoabLJKO}
\mathsf{E}_\alpha^{\mathcal{T}}(\boldsymbol{\mu},\boldsymbol{\nu}) \in \argmin_{\boldsymbol{\rho}\in{\mathbb R}^{\mathcal{T}}_+} \frac{1}{\alpha} \mathcal{A}_{\mathcal{T}} \Big(\frac{\boldsymbol{\rho}+\boldsymbol{\mu}}{2};\boldsymbol{\mu}-\boldsymbol{\rho}\Big) - \langle \bphi^\alpha, \boldsymbol{\rho} \rangle_{\mathcal{T}} \,,
\end{equation}
for all $\boldsymbol{\mu},\boldsymbol{\nu}\in \mathbb{R}^\mathcal{T}_+$ and where $\bphi^\alpha$ is given by equation \eqref{eq:HJevolve2}.
Due to the definition of $\mathcal{A}_{\mathcal{T}}$, any solution $\boldsymbol{\rho}$ satisfies $\langle \boldsymbol{\rho}, \boldsymbol{1} \rangle_{\mathcal{T}} = \langle \boldsymbol{\nu}, \boldsymbol{1} \rangle_{\mathcal{T}}$. However, since ${\boldsymbol{\phi}}$ is in general not unique, in order to specify a discrete extrapolation operator one needs to select a specific optimal potential for any $\boldsymbol{\mu},\boldsymbol{\nu}\in \mathbb{R}^\mathcal{T}_+$.
\subsection{A space-time discrete \scheme\ scheme}
We can finally formulate our second order finite volume scheme. Consider a convex discrete energy function $\mathcal{E}_{\mathcal{T}}:{\mathbb R}^{\mathcal{T}}\rightarrow{\mathbb R}$ and the two initial densities $\boldsymbol{\rho}_0,\boldsymbol{\rho}_1\in{\mathbb R}^{\mathcal{T}}_+$, with the same total discrete mass.
We define the subspace of discrete probability measures $\mathbb{P}_{\Cs}\subset{\mathbb R}^{\mathcal{T}}$ as
\[
\mathbb{P}_{\Cs} = \{ \boldsymbol{\rho}\in{\mathbb R}^{\mathcal{T}}_+: \langle \boldsymbol{\rho}, \boldsymbol{1} \rangle_{\mathcal{T}} = \langle \boldsymbol{\rho}_0, \boldsymbol{1} \rangle_{\mathcal{T}} \} \,.
\]
For the time step $\tau>0$, we compute the sequence of densities $(\boldsymbol{\rho}_{n})_{n\ge2}\subset\mathbb{P}_{\Cs}$ defined by the following recursive scheme:
\vspace{0.5em}
\begin{equation}\label{eq:LJKO2th}
\left\{
\begin{aligned}
&\boldsymbol{\rho}^\alpha_{n-1}=\mathsf{E}_\alpha^{\mathcal{T}}(\boldsymbol{\rho}_{n-2},\boldsymbol{\rho}_{n-1})\,, \\[0.5em]
&\boldsymbol{\rho}_n\in\argmin_{\boldsymbol{\rho}\in{\mathbb R}^{\mathcal{T}}_+} \frac{1}{\tau(1-\beta)} \mathcal{A}_{\mathcal{T}} \Big(\frac{\boldsymbol{\rho}+\brho^\alpha_{n-1}}{2};\brho^\alpha_{n-1}-\boldsymbol{\rho}\Big) + \mathcal{E}_{\mathcal{T}}(\boldsymbol{\rho})\,.
\end{aligned}
\right.
\end{equation}
The LJKO step in \eqref{eq:LJKO2th} is a well posed convex optimization problem. Uniqueness of the solution at each step is guaranteed if $\mathcal{E}_{\mathcal{T}}$ is strictly convex. Moreover, due to the definition of $\mathcal{A}_{\mathcal{T}}$, any solution $\boldsymbol{\rho}$ belongs to $\mb{P}_{\mathcal{T}}$.
\begin{remark}[Efficient implementation via the interior method]
Problem \eqref{eq:rhoabLJKO} and the LJKO step in \eqref{eq:LJKO2th} can be solved efficiently thanks to an interior point algorithm, as suggested in \cite{natale2020FVCA} (see also \cite{natale2021computation}).
This implies that the density will be always strictly greater than zero, up to the tolerance set for the solver.
Hence, one can compute the solution ${\boldsymbol{\phi}}$, required to define $\mathsf{E}_\alpha^\mathcal{T}$, solving directly the linear system given by the optimality condition of problem \eqref{eq:Adiss_dual}:
\begin{equation}\label{eq:Adiss_dual_OC}
\boldsymbol{\nu}-\boldsymbol{\mu} + \mathrm{div}_{\mathcal{T}} (\mathcal{L} \Big(\frac{\boldsymbol{\mu}+\boldsymbol{\nu}}{2}\Big) \odot \nabla {\boldsymbol{\phi}}) = 0 \,,
\end{equation}
where $\odot$ denotes the component-wise product, which has then a unique solution defined up to a global additive costant.
\end{remark}
\subsection{Other implementations}\label{ssec:others_order2}
We now propose a discrete version of the extrapolation-based version of the VIM scheme \eqref{eq:VIM2} and the BDF2 scheme \eqref{eq:BDF2_mat} within the same TPFA finite volume setting introduced above.
We will study these numerically in Section \ref{sssec:1dconvergence} by comparing their solutions to the solutions provided by scheme \eqref{eq:LJKO2th} on one-dimensional test cases.
Our formulation of the VIM scheme \eqref{eq:VIM2} requires solving a JKO step with time step $\frac{\tau}{2}$ and then computing the $2$-extrapolation. Using the tools introduced above, in the discrete setting this can be formulated as follows. Given the initial density $\boldsymbol{\rho}_0\in\mathbb{P}_{\Cs}$ and a time step $\tau>0$, construct the sequence of densities $(\boldsymbol{\rho}_{n})_{n\ge1}\subset\mathbb{P}_{\Cs}$ by solving at each step $n$
\begin{equation}\label{eq:VIMdiscrete}
\left\{
\begin{aligned}
&\boldsymbol{\rho}_{n-\frac{1}{2}} \in\argmin_{\boldsymbol{\rho}\in{\mathbb R}^{\mathcal{T}}_+} \frac{2}{\tau} \mathcal{A}_{\mathcal{T}} \Big(\frac{\boldsymbol{\rho}+\boldsymbol{\rho}_{n-1}}{2};\boldsymbol{\rho}_{n-1}-\boldsymbol{\rho}\Big) + \mathcal{E}_{\mathcal{T}}(\boldsymbol{\rho}) \,,\\
&\boldsymbol{\rho}_n={\sf E}^{\mathcal{T}}_2(\boldsymbol{\rho}_{n-1},\boldsymbol{\rho}_{n-\frac{1}{2}})\,.
\end{aligned}
\right.
\end{equation}
As before, the discrete LJKO steps can be computed thanks to an interior point algorithm.
From a computational point of view, this scheme is cheaper to compute than \eqref{eq:LJKO2th}, as in this case the value of the optimal potential in the discrete weighted $\dot{H}^{-1}$ norm from $\boldsymbol{\rho}_{n-1}$ to $\boldsymbol{\rho}_{n-\frac{1}{2}}$ is already known from the LJKO step and does not need to be computed. However, in the next section, we will show numerically that the solutions produced by the VIM scheme \eqref{eq:VIMdiscrete} are much more oscillatory than those obtained with the \scheme\ scheme.
We can also propose a naive discretization of the BDF2 scheme \eqref{eq:BDF2_mat} by replacing the Wasserstein distances with discrete weighted $\dot{H}^{-1}$ norms. Consider two initial conditions $\boldsymbol{\rho}_0,\boldsymbol{\rho}_1\in\mathbb{P}_{\Cs}$ and the time parameter $\tau>0$. At each step $n$, compute $\boldsymbol{\rho}_{n}$ as solution to
\begin{equation}\label{eq:bdf2discrete}
\inf_{\boldsymbol{\rho}\in{\mathbb R}^{\mathcal{T}}_+} \frac{\alpha}{(1-\beta)\tau} \mathcal{A}_{\mathcal{T}} \Big(\frac{\boldsymbol{\rho}+\boldsymbol{\rho}_{n-1}}{2};\boldsymbol{\rho}_{n-1}-\boldsymbol{\rho}\Big) -\frac{\beta}{(1-\beta)\tau} \mathcal{A}_{\mathcal{T}} \Big(\frac{\boldsymbol{\rho}+\boldsymbol{\rho}_{n-2}}{2};\boldsymbol{\rho}_{n-2}-\boldsymbol{\rho}\Big) + \mathcal{E}_{\mathcal{T}}(\boldsymbol{\rho}) \,.
\end{equation}
Problem \eqref{eq:bdf2discrete} is not a convex optimization problem. Notice that it is not even bounded from below in general. Indeed, the function $\mathcal{A}_{\mathcal{T}} (\frac{\boldsymbol{\rho}+\boldsymbol{\rho}_{n-2}}{2};\boldsymbol{\rho}_{n-2}-\boldsymbol{\rho})$ is not bounded from above if the density $\boldsymbol{\rho}_{n-2}$ is not supported everywhere.
We can nevertheless try to compute stationary points of the objective function in \eqref{eq:bdf2discrete} using again an interior point algorithm. Despite not being a robust and completely meaningful strategy, in some cases it is possible to solve the problem, which enables us to compare it to our implementation.
\begin{remark}
In one dimension, as pointed out in Remark \ref{rem:metricextra_L2}, both the metric extrapolation \eqref{eq:metricextra} and the BDF2 scheme \eqref{eq:BDF2_mat} can be recast as convex optimization problems. In this case it is possible then to design effective discretizations for these (as originally done in \cite{Matthes2019bdf2}). Nevertheless, this approach requires, at least in the Eulerian framework, to be able to switch between discrete densities and discrete quantile functions, and it does not appear obvious how to achieve this while preserving the second order accuracy of the space discretization.
\end{remark}
\section{Numerical validation of the \scheme\ scheme}\label{sec:numerics}
The objective of this section is to validate our numerical scheme \eqref{eq:LJKO2th}.
We will first show qualitatively its behavior with simple one-dimensional examples and compare it to the schemes \eqref{eq:VIMdiscrete} and \eqref{eq:bdf2discrete}. We then show that all these three approaches lead to a second order accurate discretization in both time and space.
We consider for these purposes two specific problems that exhibit a gradient flow structure in the Wasserstein space: the Fokker-Planck equation we presented in Section \ref{ssec:LJKO2convergence} and the porous medium equation. This latter writes
\begin{equation}\label{eq:porouseqLJKO2}
\partial_t \rhoe = \Delta \rhoe^{\delta} + \nabla \cdot (\rhoe \nabla V) \,,
\end{equation}
and it is a Wasserstein gradient flow with respect to the energy
\begin{equation}\label{eq:porousenLJKO2}
\mathcal{E}(\rho) = \int_{\Omega} \frac{1}{\delta-1} \rho^{\delta} + \rho V \,,
\end{equation}
for a given $\delta>1$ and with $V\in W^{1,\infty}(\Omega)$ a Lipschitz continuous exterior potential \cite{otto2001geometry}.
The energy functionals \eqref{eq:FokkerPlanckEnergy} and \eqref{eq:porousenLJKO2} are both of the form $\mathcal{E}(\rho)=\int_{\Om} E(\rho) \text{d} x$ for a strictly convex function $E:\mathbb{R}_+ \rightarrow \mathbb{R}$. They can be straightforwardly discretized as $\mathcal{E}_{\mathcal{T}} = \sum_{K\in\mathcal{T}} E(\rho_K) m_K$.
Finally, we will test scheme \eqref{eq:LJKO2th} on a more challenging application in order to show its flexibility and robustness, that is an incompressible immiscible multiphase flow in a porous medium.
We remark that when two initial conditions $\boldsymbol{\rho}_0,\boldsymbol{\rho}_1$ are needed, we compute first $\boldsymbol{\rho}_1$ from $\boldsymbol{\rho}_0$ via an LJKO step:
\[
\boldsymbol{\rho}_1=\argmin_{\boldsymbol{\rho}\in{\mathbb R}^{\mathcal{T}}_+} \frac{1}{\tau} \mathcal{A}_{\mathcal{T}} \Big(\frac{\boldsymbol{\rho}+\boldsymbol{\rho}_0}{2};\boldsymbol{\rho}_0-\boldsymbol{\rho}\Big) + \mathcal{E}_{\mathcal{T}}(\boldsymbol{\rho}) \,.
\]
In the ODE setting, computing the second initial condition via a first step of implicit Euler scheme ensures the overall second order accuracy \cite{deuflhard2002scientific}. This strategy reveals to be numerically effective also in this setting.
\subsection{Comparison of the three approaches}\label{ssec:comparison}
We compare the three different approaches on simple one dimensional tests for the diffusion equation and the porous medium equation.
For both system we set $\Omega = [0,1]$, discretized in subintervals of equal length $m_K=0.02$.
We first consider the diffusion equation, which is problem \eqref{eq:FokkerPlanckLJKO2} with zero external potential $V$. We take as initial condition
\[
\rho_0 = \operatorname{exp}\Big(-50\Big(x-\frac{1}{2}\Big)^2\Big) \,,
\]
which we discretize as $\boldsymbol{\rho}_0= (\rho_0(\mathbf{x}_K))_{K\in\mathcal{T}}$, and the time step $\tau=0.01$. In Figure \ref{fig:1dtest_FP}, we show the density obtained with the three schemes at three different times.
Using the VIM scheme \eqref{eq:VIMdiscrete}, spurious oscillations appear in the solution and these persist along the integration in time. Such oscillations can be explained as the result of the interaction of the extrapolation step, causing the mass to exit the domain, and the boundary conditions, forcing the mass to stay within $\Omega$. Neither the \scheme\ scheme \eqref{eq:LJKO2th} nor the BDF2 scheme \eqref{eq:bdf2discrete} suffer from this probem. However, notice that in both cases the dynamics slightly differ from pure diffusion due to the presence of bumps in the solution.
\begin{figure}[t!]
\centering
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{BDF2_FP_002.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{BDF2_FP_004.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{BDF2_FP_006.pdf} \\
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{VIM_FP_002.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{VIM_FP_004.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{VIM_FP_006.pdf} \\
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{BDF2mod_FP_002.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{BDF2mod_FP_004.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{BDF2mod_FP_006.pdf}
\caption{Comparison between the three schemes for the diffusion equation. From top to bottom, the BDF2 scheme \eqref{eq:bdf2discrete}, the VIM scheme \eqref{eq:VIMdiscrete} and the \scheme\ scheme \eqref{eq:LJKO2th}. From left to right, three different time steps: $t=0.02, 0.04, 0.06$.}
\label{fig:1dtest_FP}
\end{figure}
Consider now the porous medium equation \eqref{eq:porouseqLJKO2} with $\delta=2$ and the external potential $V(x)=-x$, which causes the mass to drift towards the positive direction. We take as initial condition
\[
\rho_0(x) = \mathds{1}_{x\le\frac{3}{10}} \,,
\]
discretized again as $\boldsymbol{\rho}_0= (\rho_0(\mathbf{x}_K))_{K\in\mathcal{T}}$, and the time step $\tau=0.002$. In this case, the naive implementation we proposed for the BDF2 scheme does not converge, which is not surprising since the objective function in \eqref{eq:bdf2discrete} is unbounded from below. The results for the VIM scheme \eqref{eq:VIMdiscrete} and the \scheme\ scheme \eqref{eq:LJKO2th} are shown in Figure \ref{fig:1dtest_PM}. Again, the VIM sheme is unstable whereas the \scheme\ scheme controls and smooths the oscillations generated by the extrapolation step. Note that in this case the oscillations are due to the compact support of the density and the explicit integration in time of the Hamilton-Jacobi equation: in the extrapolation step the mass cannot flow outside the support, which acts then like a boundary.
Finally, we observe that, as in the continuous setting, we cannot expect any regularity on the measure obtained after the extrapolation, and the JKO step is the only source of regularity for both the \scheme\ and the VIM scheme.
One may argue that the two schemes perform the same operations up to a temporal shift, which should contradict the different behavior shown in Figure \ref{fig:1dtest_FP}. However, notice that scheme \eqref{eq:LJKO2th} performs a smaller extrapolation and a bigger JKO step with respect to scheme \eqref{eq:VIMdiscrete}. Furthermore, in \eqref{eq:LJKO2th} one needs to compute an extrapolation between two minimizers of the JKO step, whereas in \eqref{eq:VIMdiscrete} the extrapolation is between an extrapolated measure and a JKO minimizer.
\begin{figure}[t!]
\centering
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{VIM_PM_0004.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{VIM_PM_0008.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{VIM_PM_0020.pdf} \\
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{BDF2mod_PM_0004.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{BDF2mod_PM_0008.pdf}
\includegraphics[trim={3.6cm 8.5cm 3.6cm 8cm},clip,width=0.3\textwidth]{BDF2mod_PM_0020.pdf}
\caption{Comparison between the VIM scheme \eqref{eq:VIMdiscrete} (top row) and the \scheme\ scheme \eqref{eq:LJKO2th} (bottom row) for the porous medium equation. The BDF2 scheme \eqref{eq:bdf2discrete} does not converge in this case. From left to right, three different time steps: $t=0.004, 0.008, 0.020$.}
\label{fig:1dtest_PM}
\end{figure}
\subsection{Convergence tests}
We now compare the three schemes in terms of order of convergence with respect to an exact one-dimensional solution of the Fokker-Planck equation \eqref{eq:FokkerPlanckLJKO2}. For the \scheme\ scheme \eqref{eq:LJKO2th}, we will also perform two dimensional tests using the porous medium equation \eqref{eq:porouseqLJKO2}. For all tests, we consider a sequence of meshes $\left(\mathcal{T}_m,\overline{\Sigma}_m, {(\mathbf{x}_K)}_{K\in\mathcal{T}_m}\right)$ with decreasing meshsize $h_m$ and a sequence of decreasing time steps $\tau_m$ such that $\frac{h_{m+1}}{h_m}=\frac{\tau_{m+1}}{\tau_m}$.
We solve the discrete problem for each couple $(h_m,\tau_m)$ and evaluate the convergence with respect to the discrete $L^1((0,T);L^1(\Omega))$ error:
\[
\epsilon_m = \sum_{n} \tau \sum_{K \in \mathcal{T}_m} |\rho_{K,n}-\rhoe(\mathbf{x}_K,n \tau)| m_K \,.
\]
We compute the rate of convergence as: \[\frac{\log(\epsilon_{m-1})-\log(\epsilon_m)}{\log(\tau_{m-1})-\log(\tau_m )} \,.
\]
\subsubsection{One-dimensional tests}\label{sssec:1dconvergence}
On the domain $\Omega=[0,1]$ and for the external potential $V(x)=-gx$, we consider the following exact solution to the Fokker-Planck equation \eqref{eq:FokkerPlanckLJKO2}:
\begin{equation}\label{eq:FPsol}
\rhoe(t,x) = \operatorname{exp}\left(-\Big(\pi^2+\frac{g^2}{4}\Big) t+\frac{g}{2}x\right)\left(\pi \cos(\pi x)+\frac{g}{2}\sin(\pi x)\right)+\pi \operatorname{exp}\Big(g\Big(x-\frac{1}{2}\Big)\Big) .
\end{equation}
We consider the value $g=1$. For each mesh $\left(\mathcal{T}_m,\overline{\Sigma}_m, {(\mathbf{x}_K)}_{K\in\mathcal{T}_m}\right)$ and time step $\tau_m$, we compute then the discrete solution using the three schemes, starting from the initial condition $\boldsymbol{\rho}_0= (\rhoe(0,\mathbf{x}_K))_{K\in\mathcal{T}}$.
The results are presented in Table \ref{tab:errors1d_order2}. Both the BDF2 and the \scheme\ schemes are second order accurate, whereas the order of convergence is less than one for the VIM scheme. This is due to the presence of oscillations in the solutions obtained with the VIM scheme, which are however only present at the beginning of the time interval $[0,0.25]$. Repeating the test on the interval $[0.05,0.25]$, the convergence significantly improves and attains second order accuracy as well.
\begin{table}
\centering
\setlength{\tabcolsep}{5pt}
\caption{Errors and convergence rates for the three schemes for the Fokker-Planck equation in one dimension. Integration time $[0,0.25]$ for the first three cases, $[0.05,0.25]$ for the last one.}
\label{tab:errors1d_order2}
\begin{tabular}{cccccccccc}
\toprule
& & \multicolumn{2}{c}{BDF2 \eqref{eq:bdf2discrete}} & \multicolumn{2}{c}{\scheme\ \eqref{eq:LJKO2th}} & \multicolumn{2}{c}{VIM \eqref{eq:VIMdiscrete}} & \multicolumn{2}{c}{VIM \eqref{eq:VIMdiscrete}} \\
\toprule
$h_m$ & $\tau_m$ & $\epsilon_m$ & rate & $\epsilon_m$ & rate & $\epsilon_m$ & rate & $\epsilon_m$ & rate \\
\midrule
0.100 & 0.050 & 2.091e-02 & / & 2.217e-02 & / & 5.895e-02 & / & 4.667e-03 & / \\
0.050 & 0.025 & 6.376e-03 & 1.713 & 7.016e-03 & 1.660 & 3.615e-02 & 0.706 & 1.024e-03 & 2.188 \\
0.025 & 0.013 & 1.791e-03 & 1.832 & 2.044e-03 & 1.779 & 2.294e-02 & 0.656 & 2.517e-04 & 2.025 \\
0.013 & 0.006 & 4.849e-04 & 1.885 & 5.653e-04 & 1.854 & 1.468e-02 & 0.644 & 6.264e-05 & 2.007 \\
0.006 & 0.003 & 1.280e-04 & 1.922 & 1.508e-04 & 1.906 & 1.234e-02 & 0.251 & 1.562e-05 & 2.003 \\
0.003 & 0.002 & 3.324e-05 & 1.945 & 3.933e-05 & 1.939 & 9.983e-03 & 0.306 & 3.901e-06 & 2.002 \\
\bottomrule
\end{tabular}
\end{table}
\subsubsection{Two-dimensional tests} We now estimate the order of convergence of the \scheme\ scheme on two-dimensional test cases. Here, we set $\Omega= [0,1]^2$ and use the same sequence of grids that have been used in \cite{cances2020LJKO,natale2020FVCA}, which allows for a direct comparison of the results therein.
We repeat first the test on the Fokker-Planck equation in two dimensions using the same solution \eqref{eq:FPsol} on the domain $\Omega=[0,1]^2$. The results are shown in Table \ref{tab:errors2d_FP_LJKO2} and confirm the second order accuracy of the scheme.
\begin{table}[t]
\centering
\caption{Errors and convergence rate for the \scheme\ scheme \eqref{eq:LJKO2th} for the Fokker-Planck equation in two dimensions.}
\label{tab:errors2d_FP_LJKO2}
\begin{tabular}{cccc}
\toprule
$h_m$ & $\tau_m$ & $\epsilon_m$ & rate \\
\midrule
0.2986 & 0.0500 & 2.111e-02 & / \\
0.1493 & 0.0250 & 6.800e-03 & 1.634 \\
0.0747 & 0.0125 & 2.017e-03 & 1.754 \\
0.0373 & 0.0063 & 5.669e-04 & 1.831 \\
0.0187 & 0.0031 & 1.535e-04 & 1.884 \\
\bottomrule
\end{tabular}
\end{table}
We also perform a convergence test with respect to an explicit solution of the porous medium equation \eqref{eq:porouseqLJKO2} with zero exterior potential $V$. This equation admits a solution called Barenblatt profile \cite{otto2001geometry}:
\begin{equation}\label{eq:porous_sol}
\rhoe(t,{x}) = \frac{1}{t^{d \lambda}}\Big(\frac{\delta-1}{\delta}\Big)^{\frac{1}{\delta-1}} \max\Big(M-\frac{\lambda}{2}\Big|\frac{x-x_0}{t^\lambda}\Big|^2,0\Big)^{\frac{1}{\delta-1}} \,,
\end{equation}
where $\lambda=\frac{1}{d(\delta-1)+2}$, $d$ standing for the space dimension, and $x_0$ is the point where the mass is centered. The parameter $M$ can be chosen to fix the total mass. The value
\[
M=\Big(\frac{\delta}{\delta-1}\Big)^{-\frac{1}{\delta}} \Big(\frac{\lambda \delta}{2\pi(\delta-1)}\Big)^\frac{\delta-1}{\delta}
\]
sets it equal to one.
The function \eqref{eq:porous_sol} solves \eqref{eq:porouseqLJKO2} on the domain $\Omega=[0,1]^d$, with $\mathbf{x}_0$ in the interior of $\Omega$, starting from $t_0>0$ and for a sufficiently small time horizon $T$, such that the mass does not reach the boundary of the domain.
We consider the two-dimensional case and $x_0=(0.5,0.5)$.
We solve the problem for $\delta=2,3,4$, with initial condition $\boldsymbol{\rho}_0= (\rhoe(t_0,\mathbf{x}_K))_{K\in\mathcal{T}}$, starting respectively from $t_0=10^{-4},10^{-5},10^{-6}$ and up to time $T=t_0+10^{-3}$.
The results are presented in Table \ref{tab:errors2d_PM}. The convergence profile is not clean, probably due to the low precision of the discretization in space. We can nevertheless notice that in the case $\delta=2$ the rate of convergence is approaching order two with refinement. In the cases $\delta=3,4$, where the solution is less regular, the order tends to $1.5$.
\begin{table}[t]
\centering
\caption{Errors and convergence rates for the \scheme\ scheme \eqref{eq:LJKO2th} for the porous medium equation.}
\label{tab:errors2d_PM}
\begin{tabular}{cccccccc}
\toprule
& & \multicolumn{2}{c}{$\delta=2$} & \multicolumn{2}{c}{$\delta=3$} & \multicolumn{2}{c}{$\delta=4$} \\
\toprule
$h_m$ & $\tau_m$ & $\epsilon_m$ & rate & $\epsilon_m$ & rate & $\epsilon_m$ & rate \\
\midrule
0.2986 & 2.000e-04 & 5.139e-04 & / & 7.515e-04 & / & 9.537e-04 & / \\
0.1493 & 1.000e-04 & 1.999e-04 & 1.363 & 2.780e-04 & 1.435 & 3.085e-04 & 1.628 \\
0.0747 & 5.000e-05 & 6.429e-05 & 1.636 & 4.630e-05 & 2.586 & 1.103e-04 & 1.485 \\
0.0373 & 2.500e-05 & 1.471e-05 & 2.127 & 2.903e-05 & 0.674 & 3.847e-05 & 1.519 \\
0.0187 & 1.250e-05 & 4.129e-06 & 1.833 & 7.521e-06 & 1.949 & 1.340e-05 & 1.522 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Incompressible immiscible multiphase flows in porous media}\label{ssec:multiphase}
Incompressible immiscible multiphase flows in porous media can be described as Wasserstein gradient flows, as shown in \cite{cances2017multiphase}. We recall quickly the model problem in a simplified way.
In the porous medium $\Omega$, $N+1$ phases are flowing and we denote by $\boldsymbol{s}=(s_0,...,s_N)$ the saturations of each phase, i.e. the portion of volume occupied by each phase in each point. The evolution of each saturation obeys the following equations:
\begin{equation}\label{eq:multiphase}
\left\{
\begin{aligned}
&\frac{\partial s_i}{\partial t} + \nabla \cdot (s_i v_i) = 0 \,,\\
&v_i= - \frac{1}{\mu_i} (\nabla p_i - \rho_i g) \,, \\
&p_i-p_0 = \pi_i(\boldsymbol{s},x) \, ,
\end{aligned}
\right.
\end{equation}
$i\in \{0,...,N\}$ for the first two equations, $i\in \{1,...,N\}$ for the third one, plus the total saturation condition $\sum_{i=0}^{N} s_i(t,x) = 1$ and the no-flux boundary conditions. The densities $\rho_i$ and the viscosities $\mu_i$, both constant in the whole domain, are characteristic of each phase. In \eqref{eq:multiphase} the porosity of the medium is considered constant and neglected. The term $\rho_i g$ reflects the influence of the potential energy on the motion ($g$ is the gravitational acceleration), but other types of potential energy could be considered. The model is completed specifying the $N$ capillary pressure relations, decribed by the functions $\pi_i$.
We introduce the probability spaces
\[
\mathcal{P}_i = \Big\{ s_i \in \mathcal{P}(\Omega): s_i(\Omega) = c_i \Big\}, \quad i\in \{0,...,N\},
\]
with the constant $c_i$ denoting the total mass of each phase. Each space $\mathcal{P}_i$ is endowed with the following quadratic Wasserstein distance,
\[
W_{2,i}^2(s_i^1,s_i^2) = \min_{\gamma\in \Pi(s_i^1,s_i^2)} \int \mu_i |x-y|^2 \text{d} \gamma(x,y) \,,
\]
for $s_i^1,s_i^2\in\mathcal{P}_i$ and we can define the global quadratic Wasserstein distance $\boldsymbol{W}_2$ on $\boldsymbol{\mathcal{P}}:= \mathcal{P}_0 \times ... \times \mathcal{P}_N$ by setting
\[
\boldsymbol{W}_2^2(\boldsymbol{s}^1,\boldsymbol{s}^2) = \sum_{i=0}^{N} W_{2,i}^2(s_i^1,s_i^2), \quad \forall \boldsymbol{s}^1, \boldsymbol{s}^2 \in \boldsymbol{\mathcal{P}}.
\]
Problem \eqref{eq:multiphase} can then be represented as the gradient flow in the space $\boldsymbol{\mathcal{P}}$ with respect to the (strictly convex) energy functional
\begin{equation}\label{eq:Fmultiphase}
\mathcal{E}(\boldsymbol{s}) = \int_{\Omega} \boldsymbol{\Psi} \cdot \boldsymbol{s} + \int_{\Omega} \Pi(\boldsymbol{s},x) + i_{\boldsymbol{\mathcal{S}}}(\boldsymbol{s}) \,,
\end{equation}
where $\boldsymbol{\Psi}=(\Psi_0,\ldots,\Psi_N)$ is the exterior gravitational potential given by
\[
\Psi_i(x) = -\rho_i g \cdot x, \quad \forall x \in \Omega \,,
\]
$\Pi(s,x)$ is a strictly convex potential such that
\[
\pi_i(\boldsymbol{s},x) = \frac{\partial \Pi(\boldsymbol{s},x)}{\partial s_i}, \quad i \in \{1,...,N\},
\]
and $i_{\boldsymbol{\mathcal{S}}}$ is the indicator function of the set
\[
\boldsymbol{\mathcal{S}}=\left\{ \boldsymbol{s} \in \boldsymbol{\mathcal{P}}: \sum_{i=0}^{N} s_i(x)=1, \text{for a.e.} \,\, x\in \Omega \right\}.
\]
When applying the \scheme\ scheme to such gradient flow, the extrapolation may be taken in each space $\mathcal{P}_i$ independently, i.e. we define the extrapolation in the space $\boldsymbol{\mathcal{P}}$ as
\[
\mathsf{E}_\alpha(\boldsymbol{s}^1,\boldsymbol{s}^2)\coloneqq (\mathsf{E}_\alpha(s_i^1,s_i^2))_{i=0}^N \,,
\]
for all $\boldsymbol{s}^1,\boldsymbol{s}^2\in\boldsymbol{\mathcal{P}}$. This does not guarantee at all that at each step $n$ of the scheme the extrapolation is a feasible point for $\mathcal{E}(\boldsymbol{s})$, that is $\mathsf{E}_\alpha(\boldsymbol{s}^1,\boldsymbol{s}^2)\notin\boldsymbol{\mathcal{S}}$ in general even though $\boldsymbol{s}^1,\boldsymbol{s}^2\in\boldsymbol{\mathcal{S}}$. Nevertheless, the resulting scheme is well defined as well as the numerical approach \eqref{eq:LJKO2th}. In our implementation, we linearize each Wasserstein distances independently. The energy functional can be discretized straightforwardly.
As a specific instance of problem \eqref{eq:multiphase}, we consider a two-phase flow, where water ($s_0$) and oil ($s_1$) are competing in the porous medium. We choose the classical Brooks-Corey capillary pressure model,
\[
p_1-p_0 = \pi_1(s_1) = \lambda (1-s_1)^{-\frac{1}{2}}\,,
\]
and take $g$ acting along the negative direction of the $y$ axis, $|g|=9.81$. We set the model parameter $\lambda=0.05$. The densities and the viscosities of the two fluids are, respectively, $\rho_0=1$ and $\rho_1 = 0.87$, $\mu_0=1$ and $\mu_1=100$. We consider a non convex domain $\Omega$ shaped as an hourglass and set an initial condition where the water is distributed uniformly in a layer in the upper part, whereas the oil takes the complementary space (see Figure \ref{fig:twophaseflow_0}). The evolution of the oil saturation $s_1$ is presented in Figure \ref{fig:twophaseflow}.
\begin{figure}[t!]
\centering
\begin{minipage}{0.89\textwidth}
\centering
\subfloat[][$t=0$\label{fig:twophaseflow_0}]{\includegraphics[trim={4.7cm 7.8cm 4cm 7.25cm},clip,width=0.29\textwidth]{twophaseflow_s1-0.pdf}} \,
\subfloat[][$t=8$]{\includegraphics[trim={4.7cm 7.8cm 4cm 7.25cm},clip,width=0.29\textwidth]{twophaseflow_s1-4.pdf}} \,
\subfloat[][$t=16$]{\includegraphics[trim={4.7cm 7.8cm 4cm 7.25cm},clip,width=0.29\textwidth]{twophaseflow_s1-8.pdf}} \\
\subfloat[][$t=24$]{\includegraphics[trim={4.7cm 7.8cm 4cm 7.25cm},clip,width=0.29\textwidth]{twophaseflow_s1-12.pdf}} \,
\subfloat[][$t=48$]{\includegraphics[trim={4.7cm 7.8cm 4cm 7.25cm},clip,width=0.29\textwidth]{twophaseflow_s1-24.pdf}} \,
\subfloat[][$t=120$]{\includegraphics[trim={4.7cm 7.8cm 4cm 7.25cm},clip,width=0.29\textwidth]{twophaseflow_s1-60.pdf}} \\
\end{minipage}
\hspace{-2em}
\begin{minipage}{0.09\textwidth}
\centering
\vspace{0.4em}
\includegraphics[trim={17.2cm 8.6cm 1.7cm 8.1cm},clip,height=0.2\textheight,]{colorbar_east2.pdf}
\end{minipage}
\caption{Evolution of the saturation of the oil phase in the hourglass. The evolution of the water is complementary. As expected, the water, the denser phase, flows down the hourglass under the effect of gravity up until reaching the bottom.}
\label{fig:twophaseflow}
\end{figure}
\section{Conclusion}
In this work we proposed and analyzed different notions of extrapolation in the Wasserstein space. We showed how these can be used to construct a second-order time discretization of Wasserstein gradient flows, based on a two-step reformulation of the classical BDF2 scheme. According to the specific notion considered, we could prove different types of convergence guarantees for the scheme. We also proposed a fully-discrete version of the method, and demonstrated numerically its second-order accuracy in space and time. The possibility to provide an implementable scheme is in fact the main advantage of our approach compared to previous works also based on the BDF2 scheme \cite{Matthes2019bdf2}, or on the midpoint rule \cite{Legendre2017VIM}. The different type of extrapolations and their properties are summarized in Table \ref{tab:extra}.
\begin{table}
\centering
\caption{Summary of the different types of extrapolation proposed in the present work.}
\label{tab:extra}
\renewcommand{\arraystretch}{1.3}
\begin{tabular}{C{3.5cm} C{3cm} C{3cm} C{3cm}}
\toprule
& Free-flow extrapolation \eqref{eq:freeflow} & Viscosity extrapolation \eqref{eq:omegahj} & Metric extrapolation \eqref{eq:metricextraF} \\
\toprule
Fokker-Planck conv. & \cmark & \textbf{?} & \cmark \\ \hline
EVI conv.& \textbf{?} & \textbf{?} & \cmark \\ \hline
Implementation & \textbf{?} & \cmark & \textbf{?}\\ \hline
Second order & \textbf{?} & \cmark & \textbf{?} \\
\toprule
\end{tabular}
\end{table}
In order to provide our fully discrete scheme, we worked in the framework of Eulerian discretizations and considered an extrapolation based on viscosity solutions of the Hamilton-Jacobi equation. The resulting scheme is robust and allows to achieve second order of accuracy both in space and time, but it does not verify the hypotheses of our convergence results. The free-flow extrapolation could be implemented straightforwardly in the framework of Lagrangian discretizations (see, e.g., \cite{matthes2014convergence,calvez2014particle} for Lagrangian discretizations of Wasserstein gradient flows), although in this setting it would be challenging to achieve second order accuracy in space.
The metric extrapolation enjoys the nicest mathematical structure, and in priniciple one could exploit its dual formulation \eqref{eq:metricextra_dual}, which is a convex optimization problem, to implement it numerically. However, dealing with the strong-convexity constraint on the Brenier potential requires the development of dedicated tools. We will investigate this direction in a future work.
\section*{Acknowledgements}
TOG acknowledges the support of the french Agence Nationale de la Recherche through the project MAGA (ANR-16-CE40-0014).
GT acknowledges that this project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 754362.
The authors would like to thank Clément Cancès and Guillaume Carlier for fruitful discussions and suggestions on the topic.
\begin{center}
\vspace{0.5em}
\includegraphics[width=0.15\textwidth]{eu_flag.jpg}
\end{center}
\bibliographystyle{plain}
| 0461cfe14f405abce8374e70e562e883a72fcb76 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In his paper~\cite{For68} from 1968, Forsythe considered the ``asymptotic directions of the $s$-dimensional optimum gradient method'' or, in modern terms, the asymptotic behavior of the conjugate gradient (CG) method~\cite{HesSti52} when restarted every~$s$ steps. Based on what he could prove in the simplest case of restart length $s=1$, and what he observed numerically for $s=2$, Forsythe formulated a conjecture about the behavior of the method for general~$s$. In short, the Forsythe conjecture is that the normalized error vectors (or normalized residual vectors) of the restarted method eventually cycle back and forth between two limiting directions. This unwanted oscillatory behavior leads to slow (at most linear) convergence, which Stiefel had illustrated with his cage (``K\"afig'') already in 1952~\cite{Sti52}.
In light of the importance and wide-spread use of the CG method, surprisingly few results about the Forsythe conjecture have been published beyond the case $s=1$, and the conjecture remains largely open after more than 50 years. In addition to Forsythe in~\cite{For68}, proofs of the oscillatory behavior for $s=1$ were published by Akaike~\cite{Aka59} and Gonzaga and Schneider~\cite{GonSch16}, as well as Afanasjew, Eiermann, Ernst, and G\"uttel~\cite{AfaEieErnGue08} (for a somewhat different algorithm). The nature of the oscillations for $s=1$ were studied by Nocedal, Sartenaer, and Zhu~\cite{NocSarZhu02}. Prozanto, Wynn, and Zhigljavsky~\cite{ProWynZhi09} analyzed the rate of convergence of the $s$-dimensional optimum gradient method, re-derived several results of Forsythe in a different formulation, and numerically studied the limiting behavior particularly in the case $s=2$. This case was also analyzed by Zhuk and Bondarenko~\cite{ZhuBon83}, and we will comment on their work after the proof of our Theorem~\ref{thm:limit-degree} below.
Forsythe already suspected, and even conjectured, that the results he proved apply to a wider class of (restarted) gradient-type methods when applied to sufficiently smooth functions; see~\cite[p.~58]{For68}. And indeed, a cyclic behavior has sometimes been observed also in other restarted Krylov subspace methods, particularly when applied to symmetric or normal matrices; see, e.g.,~\cite{BakJesMan05,VecLan10}.
In this paper we generalize the Forsythe conjecture (formulated only for symmetric positive definite matrices) to symmetric and nonsymmetric matrices. Instead of the projection process underlying the CG method, we use the projection process of the Arnoldi (or for symmetric matrices the Lanczos) algorithm. The resulting \emph{Arnoldi cross iteration} alternatingly uses the given matrix $A$ and its transpose~$A^T$. While such a cross iteration usually does not solve a linear algebraic system or eigenvalue problem for~$A$, it can be of interest for analyzing the worst-case behavior of the underlying projection process. Motivated by work of Zavorin~\cite{Zav01} as well as Zavorin, O'Leary, and Elman~\cite{ZavOLeElm03} (who studied complete stagnation of the GMRES method~\cite{SaaSch86} for~$A$ and~$A^T$), we have previously introduced a \emph{GMRES cross iteration} in the context of worst-case GMRES~\cite{FabLieTic13}. In this paper we do not focus on the worst-case behavior, but we will nevertheless show that the Arnoldi cross iteration with $s=1$ solves the \emph{ideal Arnoldi problem} for real orthogonal matrices; see~\eqref{eqn:idealArnoldi} and the end of Section~\ref{sec:ACI-orth}.
Because it uses $A$ and $A^T$, the Arnoldi cross iteration for $s=1$ is related to Ostrowski's two-sided iteration~\cite{Ost59} and Parlett's Alternating Rayleigh Quotient Iteration (ARQI) for computing eigenpairs of nonsymmetric matrices~\cite{Par74}. The precise nature of these relations are yet to be explored.
For symmetric matrices the iterative steps with $A$ and~$A^T$ in the Arnoldi cross iteration coincide, and we recover essentially the same algorithm as originally studied by Forsythe, only with a different (namely the Arnoldi or the Lanczos) projection process. We conjecture that the Arnoldi cross iteration in general has the same oscillatory limiting behavior as the one conjectured by Forsythe for the restarted CG method. We point out that usually such a behavior does not exist in restarted Krylov subspace method for nonsymmetric matrices which use only $A$ (and not $A^T$); see~\cite{VecLan11}.
We generalize or extend several results from the original Forsythe formulation and from our paper~\cite{FabLieTic13} to the Arnoldi cross iteration. We also prove a few new results about the limiting behavior of this iteration. In the simplest case $s=1$ and $A=A^T$, the iteration reduces to the one studied in~\cite{AfaEieErnGue08}, and we give an alternative (and in particular simpler) proof of the limiting behavior in this case. Another new result in this paper is the proof of the limiting behavior of the Arnoldi cross iteration for $s=1$ and orthogonal matrices with eigenvalues having only positive (or only negative) real parts.
Besides being more general than the original formulation of Forsythe, we hope that our translation of the theory in the context of the conjecture into more modern notation based on projection processes is more transparent, and therefore will motivate further research beyond this paper.
The paper is organized as follows. As mentioned above, in Section~\ref{sec:forsythe} we present the original Forsythe conjecture in modern terms and discuss previous attempts to prove it. In Section~\ref{sec:arnoldi} we introduce the Arnoldi cross iteration, prove several results about this iteration and its limiting behavior, and generalize the conjecture. In Section~\ref{sec:special} we consider the Arnoldi cross iteration for symmetric and orthogonal matrices. In Section~\ref{sec:Conclusions} we give concluding remarks.
\smallskip{}
\emph{Notation.} Throughout the paper we consider only real matrices for notational simplicity. Many results can be easily extended to the complex case. The degree of the minimal polynomial of a matrix $A\in {\mathbb R}^{n\times n}$ is denoted by $d(A)$, and the grade of a vector $v\in{\mathbb R}^{n}$ with respect to $A$ by $d(A,v)$; cf.~\cite[Definition~4.2.1]{LieStr13}. For each $k\geq 1$, the $k$th Krylov subspace of $A\in\mathbb{R}^{n\times n}$ and $v\in\mathbb{R}^n$ is denoted by $\mathcal{K}_{k}(A,v):={\rm span} \{v,Av,\dots,A^{k-1}v\}$. The Euclidean norm on $\mathbb{R}^n$ is denoted by $\|\cdot\|$, and for a symmetric positive definite matrix $A\in \mathbb{R}^{n \times n}$ the $A$-norm on $\mathbb{R}^n$ is denoted by $\|\cdot\|_A$, where $\|v\|_A:=(v^TAv)^{1/2}$.
\section{The original Forsythe conjecture}\label{sec:forsythe}
Forsythe~\cite{For68} considered minimizing functions of the form
\begin{equation}\label{eqn:f}
f(x)=\frac12 x^TAx-x^Tb,
\end{equation}
where $A\in {\mathbb R}^{n\times n}$ is symmetric positive definite, by an iterative method he called the \emph{optimum $s$-gradient method}. We remark that Forsythe actually considered $b=0$ in~\eqref{eqn:f} for simplicity, so that the unique minimizer of~$f$ is given by $x=0$. In order to avoid any possible confusion of the reader that may be caused by studying a method for computing a solution that is already known trivially, we here consider a possibly nonzero vector~$b\in\mathbb{R}^n$ in~\eqref{eqn:f}. In any case, the unique minimizer of~$f$ is equal to the uniquely determined solution of the linear algebraic system $Ax=b$.
We now briefly describe Forsythe's optimum $s$-gradient method using modern notation. Let $x_0\in {\mathbb R}^n$ be a given initial vector. Then the gradient of $f$ evaluated at $x_0$ is given by
$$\nabla f(x_0)=Ax_0-b=:-r_0,$$
and the optimum 1-gradient method amounts to finding the minimum of $f$ with respect to the $A$-norm on the line $x_0-\alpha\nabla f(x_0)=x_0+\alpha r_0$, which is the \emph{line of steepest descent} of $f$ at $x_0$. Thus, using the notation of Krylov subspaces, the new iterate of the optimum 1-gradient method satisfies
\begin{equation}\label{eqn:CG1}
x_1\in x_0+{\rm span}\{r_0\}=x_0+\mathcal{K}_1(A,r_0),
\end{equation}
and it is found so that
$$f(x_1) = \min_{z\in x_0+\mathcal{K}_1(A,r_0)} f(z).$$
Since $\|x-x_1\|_A^2=2 f(x_1)+x^TAx$, solving the minimization problem for~$f$ is equivalent with finding~$x_1$ so that
$$\|x-x_1\|_A=\min_{z\in x_0+\mathcal{K}_1(A,r_0)}\|x-z\|_A.$$
It is well known (see, e.g.,~\cite[Theorem~2.3.1]{LieStr13}) that~$x_1$ is uniquely determined by the orthogonality condition
\begin{equation}\label{eqn:CG2}
x-x_1\perp_A \mathcal{K}_1(A,r_0).
\end{equation}
The equations \eqref{eqn:CG1}--\eqref{eqn:CG2} give a complete mathematical characterization of the optimum 1-gradient method.
For each $z=x_0+\alpha r_0\in x_0+\mathcal{K}_1(A,r_0)$ we have
$$\nabla f(z)= A(x_0+\alpha r_0)-b=-r_0+\alpha Ar_0.$$
Forsythe called the set
$$\{x_0+\alpha_1 r_0+\alpha_2 Ar_0\,:\,\alpha_1,\alpha_2\in {\mathbb R}\}
= x_0+\mathcal{K}_2(A,r_0)$$
the \emph{$2$-plane of steepest descent} of $f$ at $x_0$. Analogously, the \emph{$s$-dimensional plane of steepest descent} of $f$ at $x_0$ is given by
$$\Big\{x_0+\sum_{j=0}^{s-1} \alpha_j A^jr_0\,:\,\alpha_j\in {\mathbb R},\,j=0,1,\dots,s-1\Big\}
= x_0+\mathcal{K}_s(A,r_0).$$
In analogy with \eqref{eqn:CG1}--\eqref{eqn:CG2}, the iterate $x_1$ of the optimum $s$-gradient method is then defined by the relations
\begin{equation}\label{eqn:forsythe}
x_1\in x_0+\mathcal{K}_s(A,r_0)\quad\mbox{such that}\quad x-x_1\perp_A \mathcal{K}_s(A,r_0).
\end{equation}
This is nothing but the mathematical characterization of the $s$th iterate of the CG method applied to $Ax=b$ with the initial vector $x_0$; see, e.g.,~\cite[Theorem~2.3.1]{LieStr13}. Forsythe was of course aware that the method he considered is mathematically equivalent to the CG method. He also pointed out that the implementation of Hestenes and Stiefel ``in practice ... may usually be superior to the optimum $s$-gradient methods''~\cite[p.~58]{For68}.
Forsythe was interested in the behavior of the iterates in the optimum $s$-gradient method when it is applied multiple times or, in modern terms, \emph{restarted}. It is well known that the restarted method converges to the uniquely determined minimizer of~$f$, given by $x=A^{-1}b$. The interesting question in this context is \emph{from which directions} the iterates approach their limit.
In order to study this behavior, one considers an integer~$s$ with $1\leq s< d(A)$, an initial vector $x_0\in {\mathbb R}^{n}$ with $d(A,x_0)\geq s+1$, and a sequence of vectors constructed using \eqref{eqn:forsythe} with an additional normalization:
\begin{align}
& \mbox{For \ensuremath{k=0,1,2,\dots}}\nonumber\\
& \hspace{1cm}\mbox{$y_k=r_k/\|r_k\|$,}\label{eqn:yk}\\
& \hspace{1cm}\mbox{$x_{k+1}\in x_k+\mathcal{K}_s(A,r_k)$
such that $x-x_{k+1}\perp_A \mathcal{K}_s(A,r_k)$.}\label{eqn:forsythe1}
\end{align}
For the case $s=1$, and hence the steepest descent method, Forsythe and Motzkin had conjectured already in 1951 that the two sequences of normalized vectors with even and odd indices, i.e., $\{y_{2k}\}$ and $\{y_{2k+1}\}$, alternate asymptotically between two limit vectors that are determined by the eigenvectors of $A$ corresponding to its smallest and its largest eigenvalue. In the words of Forsythe, ``the iteration behaves asymptotically, as $k\rightarrow\infty$, as though it were entirely in the two-space $\pi_{1,n}$''~\cite[p.~64]{For68}. Since asymptotically the vectors with the even indices become arbitrarily close to being collinear, and the same happens for the vectors with the odd indices, the error norms of the non-normalized restarted iteration can converge to zero (only) linearly; see~\cite[pp.~63--64]{For68} for Forsythe's original discussion of this important observation.
The conjecture of Forsythe and Motzkin (for $s=1$) was proven by Akaike in 1959 using methods from probability theory~\cite{Aka59}, and Forsythe gave another proof in~\cite{For68} using orthogonal polynomials. Based on numerical evidence he suspected that the behavior of the optimum $s$-gradient method is similar for all $s$, and we can therefore state his conjecture from~\cite[p.~66]{For68} as follows:
\medskip
{\sc Forsythe conjecture.} \emph{For $2\leq s<d(A)$, each of the two subsequences $\{y_{2k}\}$ and $\{y_{2k+1}\}$ in \eqref{eqn:yk}--\eqref{eqn:forsythe1} has a single limit vector.}
\section{The Arnoldi cross iteration}\label{sec:arnoldi}
Since the (oblique) projection process~\eqref{eqn:forsythe1} on which the Forsythe conjecture is based uses the $A$-inner product, the matrix $A$ must be symmetric positive definite. Our goal in this section is to obtain a generalization of the Forsythe conjecture to symmetric and nonsymmetric matrices that is as straightforward as possible. We therefore consider a closely related projection process, which is also well known in the area of Krylov subspace methods, and which is well defined for general~$A$.
Given $A\in\mathbb{R}^{n\times n}$, an integer $s\geq1$, and a vector $v\in\mathbb{R}^{n}$, we define the vector $w\in\mathbb{R}^{n}$ such that
\begin{equation}\label{eqn:arnoldi}
w\in A^{s}v+\mathcal{K}_{s}(A,v)\quad\mbox{and}\quad w\perp\mathcal{K}_{s}(A,v).
\end{equation}
The construction of the vector $w$ is the $s$th step in the orthogonalization of the Krylov sequence $v,Av,A^{2}v,\dots$, where the vector $A^{s}v$ is orthogonalized with respect to the previous vectors $v,\dots,A^{s-1}v$. Since the standard method for computing orthogonal Krylov subspace bases is the Arnoldi algorithm~\cite{Arn51}, we call $w$ the \emph{Arnoldi projection} of~$v$ with respect to $A$ and $s$.
By construction, $w=p(A)v$ for some polynomial $p\in\mathcal{M}_{s}$, which is the set of (real) monic polynomials of degree~$s$. Further basic properties of the Arnoldi projection are shown in the next result.
\medskip{}
\begin{lemma}\label{lem:basic}
If $A\in\mathbb{R}^{n\times n}$, $1\leq s<d(A)$, $v\in\mathbb{R}^{n}$, and $w$ is given by \eqref{eqn:arnoldi}, then the following
hold:
\begin{itemize}
\item[(1)] The vector $w$ satisfies
\[
\|w\|=\min_{z\in\mathcal{K}_{s}(A,v)}\|A^{s}v+z\|=\min_{q\in\mathcal{M}_{s}}\|q(A)v\|,
\]
and hence $w=0$ if and only if $d(A,v)\leq s$.
\item[(2)] If $d(A,v)\geq s$, then $w=\monic{A}{v}{v}$ for a uniquely determined polynomial
$\mpol{z}{v} \in\mathcal{M}_{s}$.
\end{itemize}
\end{lemma}
\begin{proof}
(1) By construction, $w=A^{s}v+u$ for some $u\in\mathcal{K}_{s}(A,v)$,
and hence for any $z\in\mathcal{K}_{s}(A,v)$ we obtain
\begin{align*}
\|A^{s}v+z\|^{2}=\|w-(u-z)\|^{2}=\|w\|^{2}+\|u-z\|^{2}\geq\|w\|^{2},
\end{align*}
with equality if and only if $z=u$. This proves the minimization
property of $w$. From $\|w\|=\min_{q\in\mathcal{M}_{s}}\|q(A)v\|$
we see that $w=0$ holds if and only if $d(A,v)\leq s$.
(2) Suppose that $d(A,v)\geq s$. If $w=p(A)v=q(A)v$ with $p,q\in\mathcal{M}_{s}$,
then $(p-q)(A)v=0$. Since the polynomials $p$ and $q$ are both
monic, the polynomial $p-q$ has degree at most $s-1$. But then $d(A,v)\geq s$
implies $p-q=0$.
\end{proof}
\medskip{}
In order to generalize the Forsythe conjecture to nonsymmetric matrices we will focus on the limiting behavior of a sequence of vectors obtained from repeatedly computing~\eqref{eqn:arnoldi} alternatingly with $A$ and $A^T$, and with additional normalizations.
The algorithm we consider here is similar to~\cite[Algorithm~1]{FabLieTic13}, but based on the Arnoldi projection~\eqref{eqn:arnoldi} instead of the projection process that is used in the GMRES method.
Given $A\in\mathbb{R}^{n\times n}$, an integer $s$ with $1\leq s<d(A)$, and a vector $v_{0}\in\mathbb{R}^{n}$ with $\|v_0\|=1$ and $d(A,v_0)\geq s+1$, we consider the following algorithm:
\begin{align}
& \mbox{For \ensuremath{k=0,1,2,\dots}}\nonumber \\
& \hspace{1cm}\mbox{\ensuremath{\widetilde{w}_{k}=\monic{A}{v_k}{v_k}},}\label{eqn:twk}\\
& \hspace{1cm}\mbox{\ensuremath{w_{k}=\widetilde{w}_{k}/\|\widetilde{w}_{k}\|},}\label{eqn:wk}\\
& \hspace{1cm}\mbox{\ensuremath{\widetilde{v}_{k+1}=\monic{A^T}{w_k}{w_k}},}\label{eqn:tvk}\\
& \hspace{1cm}\mbox{\ensuremath{v_{k+1}=\widetilde{v}_{k+1}/\|\widetilde{v}_{k+1}\|}.}\label{eqn:vkp1}
\end{align}
We call this algorithm the \emph{Arnoldi cross iteration with restart length~$s$}, or shortly \emph{ACI($s$)}.
Because of the mathematical similarities between the ACI($s$) and the GMRES cross iteration in~\cite{FabLieTic13}, we expect that their limiting behavior is similar.
We will now transfer some of the results about the GMRES cross iteration in~\cite[Section~2]{FabLieTic13} to the ACI($s$). That paper is about worst-case GMRES, and here we need an analogous definition for the worst case in the projection process \eqref{eqn:arnoldi}.
\medskip{}
\begin{definition}
Let $A\in\mathbb{R}^{n\times n}$ and an integer $s$ with $1\leq s<d(A)$ be given.
Denote
\begin{equation}\label{eq:wcArnoldi}
\Phi_s(A):=\max_{\substack{v\in\mathbb{R}^n\\ \|v\|=1}}\min_{p\in\mathcal{M}_s}\|p(A)v\|
= \max_{\substack{v\in\mathbb{R}^n\\ \|v\|=1}}
\|\monic{A}{v}{v}\| \,.
\end{equation}
A unit norm vector $v\in\mathbb{R}^n$ and a corresponding monic polynomial $P_s(z;v)$
for which the value $\Phi_s(A)$ is attained are called a worst-case Arnoldi vector
and a worst-case Arnoldi polynomial for $A$ and $s$, respectively.
\end{definition}
\medskip{}
Since a worst-case vector or the corresponding worst-case polynomial for the GMRES method need not be unique in general (see \cite[Theorem~4.1]{FabLieTic13}), we expect the same to be true for the worst-case Arnoldi problem \eqref{eq:wcArnoldi}. Note that in general we have
\begin{equation}\label{eqn:idealArnoldi}
\Phi_s(A)=\max_{\substack{v\in\mathbb{R}^n\\ \|v\|=1}}
\min_{p\in\mathcal{M}_s}\|p(A)v\| \leq \min_{p\in\mathcal{M}_s} \max_{\substack{v\in\mathbb{R}^n\\ \|v\|=1}}\|p(A)v\|=
\min_{p\in\mathcal{M}_s} \|p(A)\|.
\end{equation}
The expression on the right hand side is called the \emph{ideal Arnoldi problem} for~$A$ and~$s$; see \cite{GreTre94}. If equality holds in \eqref{eqn:idealArnoldi}, then the worst-case Arnoldi polynomial in~\eqref{eq:wcArnoldi} is unique; see~\cite[Lemma~2.4]{TicLieFab07} and~\cite{LieTic09}.
The next result transfers~\cite[Theorem~2.2]{FabLieTic13} and the first part of~\cite[Theorem~2.5]{FabLieTic13} to the ACI($s$).
\medskip{}
\begin{theorem}\label{thm:monotone}
Let $A\in\mathbb{R}^{n\times n}$, let $s$ be an integer with $1\leq s<d(A)$, and let $v_{0}\in\mathbb{R}^{n}$ be such that $\|v_0\|=1$ and $d(A,v_{0})\geq s+1$. Then the vectors in \eqref{eqn:twk}--\eqref{eqn:vkp1} are all well defined, and
\begin{equation}\label{eqn:interlace}
\|\widetilde{w}_k\|\leq\|\widetilde{v}_{k+1}\|\leq\|\widetilde{w}_{k+1}\|\leq\|\widetilde{v}_{k+2}\|\leq \Phi_s(A^T)=\Phi_s(A),\quad k=0,1,2,\dots.
\end{equation}
Equality holds in the first inequality if and only if $v_{k}=\alpha v_{k+1}$ for some $\alpha\neq0$, and in the second if and only if $w_{k}=\beta w_{k+1}$ for some $\beta\neq0$.
\end{theorem}
\begin{proof}
We start by showing inductively that the vectors in \eqref{eqn:twk}--\eqref{eqn:vkp1} are all well defined and satisfy \eqref{eqn:interlace}. Suppose that for some $k\geq 0$ we have $\widetilde{w}_{k}=\monic{A}{v_k}{v_k}\neq 0$ for some uniquely determined polynomial $\mpol{z}{v_k}\in\mathcal{M}_s$, so that $w_{k}=\widetilde{w}_{k}/\|\widetilde{w}_{k}\|$ is well defined. By construction we have $\widetilde{w}_{k}\perp\mathcal{K}_{s}(A,v_{k})$, and therefore
\begin{align*}
\|\widetilde{w}_{k}\|^{2} =\langle \monic{A}{v_k}{v_k},\widetilde{w}_{k}\rangle=\langle A^{s}v_{k},\widetilde{w}_{k}\rangle=
\langle q(A)v_{k},\widetilde{w}_{k}\rangle
&=\langle v_{k},q(A^T)\widetilde{w}_{k}\rangle,
\end{align*}
where $q\in\mathcal{M}_{s}$ is arbitrary. Using $w_{k}=\widetilde{w}_{k}/\|\widetilde{w}_{k}\|$ we obtain
\begin{align}\label{eqn:orth1}
\|\widetilde{w}_{k}\| =
\langle v_{k},q(A^T)w_{k}\rangle.
\end{align}
Since $0<\|\widetilde{w}_{k}\|$ and $q\in\mathcal{M}_{s}$ is arbitrary, we have $d(A^T,w_k)\geq s+1$. Using Lemma~\ref{lem:basic},
there is a unique polynomial $\mpol{z}{w_k} \in \mathcal{M}_s$ such that
$\widetilde{v}_{k+1}=\monic{A^T}{w_k}{w_k} \neq 0$, and, using \eqref{eqn:orth1}
and choosing $q=\mpol{z}{w_k}$ yields
\begin{align}\label{eqn:ineq1}
\|\widetilde{w}_{k}\| = \langle v_{k},\widetilde{v}_{k+1}\rangle \leq\|\widetilde{v}_{k+1}\|,
\end{align}
where we have used the Cauchy-Schwarz inequality and $\|v_{k}\|=1$.
Therefore, $v_{k+1} = \widetilde{v}_{k+1}/\|\widetilde{v}_{k+1}\|$ is well defined.
Similarly, we can prove that $\|\widetilde{v}_{k+1}\|=\langle w_{k},q(A) v_{k+1}\rangle$
where $q\in\mathcal{M}_{s}$ is arbitrary.
Since $0<\|\widetilde{v}_{k+1}\|$, we have $d(A,v_{k+1})\geq s+1$, and it holds that
\begin{align}\label{eqn:ineq1a}
\|\widetilde{v}_{k+1}\|
=\langle w_{k},\widetilde{w}_{k+1}\rangle \leq \| \widetilde{w}_{k+1} \|.
\end{align}
By induction, all vectors in \eqref{eqn:twk}--\eqref{eqn:vkp1} are well defined
and satisfy \eqref{eqn:interlace}.
Moreover, since we have used the Cauchy-Schwarz inequality to obtain \eqref{eqn:ineq1}, we have $\|\widetilde{w}_{k}\|\leq\|\widetilde{v}_{k+1}\|$ with equality if and only if $v_{k}$ and $\widetilde{v}_{k+1}$ are linearly dependent, or equivalently $v_{k}=\alpha v_{k+1}$ for some $\alpha\neq0$. Similarly, the equality in \eqref{eqn:ineq1a} holds
if and only if $w_k$ and $\widetilde{w}_{k+1}$ are linearly dependent, or equivalently $w_k=\beta w_{k+1}$ for some $\beta\neq 0$.
It is obvious that $\| \widetilde{w}_{k} \| \leq \Phi_s(A)$
and $\| \widetilde{v}_{k} \| \leq \Phi_s(A^T)$ for each $k$. We only need to prove that
$\Phi_s(A) = \Phi_s(A^T)$.
Let $v$ be a unit norm worst-case Arnoldi vector for $A$ and $s$, and denote $\widetilde{w}=\monic{A}{v}{v}$ and $w\equiv \widetilde{w}/\|\widetilde{w}\|$,
so that $\|\widetilde{w}\| = \Phi_s(A)$. Then
\begin{align*}
\Phi_s(A)
= \left\langle \monic{A}{v}{v}, w \right\rangle
= \left\langle v, q(A^T) w \right\rangle
\leq
\min_{p\in \mathcal{M}_s} \| p(A^T) w \|
\leq
\Phi_s(A^T),
\end{align*}
where $q\in\mathcal{M}_{s}$ is arbitrary. Exchanging the roles of $A$ and $A^T$ shows the reverse inequality and completes the proof.
\end{proof}
\medskip{}
As shown in \eqref{eqn:interlace}, the two nondecreasing sequences $\{\|\widetilde{w}_k\|\}$ and $\{\|\widetilde{v}_k\|\}$ interlace each other and are both bounded from above by $\Phi_s(A)=\Phi_s(A^T)<\infty$. Thus, the sequences converge to the same limit,
\begin{align}\label{eqn:same-limit}
\|\widetilde{w}_k\|\rightarrow\tau\quad\mbox{and}\quad \|\widetilde{v}_k\|\rightarrow\tau\quad \mbox{for $k\rightarrow\infty$,}
\end{align}
where $\tau\leq \Phi_s(A)$.
This observation is helpful for proving the next result, which is the ACI($s$) version of the second part of~\cite[Theorem~2.5]{FabLieTic13}.
\medskip{}
\begin{theorem}\label{thm:evenodd}
Let $A\in\mathbb{R}^{n\times n}$, let $s$ be an integer with $1\leq s<d(A)$, and let $v_{0}\in\mathbb{R}^{n}$ be such that $\|v_0\|=1$ and $d(A,v_{0})\geq s+1$. Then the sequences of the normalized vectors in \eqref{eqn:twk}--\eqref{eqn:vkp1} satisfy
\[
\lim_{k\rightarrow\infty}\|w_{k+1}-w_{k}\|=0\quad\mbox{and}\quad\lim_{k\rightarrow\infty}\|v_{k+1}-v_{k}\|=0.
\]
\end{theorem}
\begin{proof}
Using similar ideas as in the proof of Theorem~\ref{thm:monotone}, for any $k\geq0$ we have
\begin{align*}
\frac{1}{2}\|v_{k+1}-v_{k}\|^{2} & =1-\langle v_{k+1},v_{k}\rangle\\
& =1-\frac{1}{\|\widetilde{v}_{k+1}\|}\langle \monic{A^T}{w_k}{w_k},v_k\rangle\\
& =1-\frac{1}{\|\widetilde{v}_{k+1}\|}\langle w_k,q(A) v_k\rangle\qquad\mbox{(for any \ensuremath{q\in\mathcal{M}_s})}\\
& =1-\frac{1}{\|\widetilde{v}_{k+1}\|}\langle w_k,\monic{A}{v_k}{v_k}\rangle\\
& =1-\frac{1}{\|\widetilde{v}_{k+1}\|}\langle w_k,\widetilde{w}_k\rangle\\
& =1-\frac{\|\widetilde{w}_k\|}{\|\widetilde{v}_{k+1}\|},
\end{align*}
where we have used $w_k\perp \mathcal{K}_s(A,v_k)$.
Since the sequences $\{\|\widetilde{v}_k\|\}$ and $\{\|\widetilde{w}_k\|\}$ converge to the same limit for $k\rightarrow\infty$, we have $\frac{\|\widetilde{w}_k\|}{\|\widetilde{v}_{k+1}\|}\rightarrow 1$ and hence
$\|v_{k+1}-v_{k}\|\rightarrow 0$ for $k\rightarrow\infty$. The proof for the sequence $\{w_k\}$ is analogous.
\end{proof}
\medskip{}
In the following we will focus on the properties of sequence $\{v_k\}$, with the understanding that analogous results can be formulated also for the sequence $\{w_k\}$. The unit norm vectors $v_{k}$ form a bounded sequence in ${\mathbb{R}}^{n}$. By the Bolzano-Weierstrass Theorem this sequence has a convergent subsequence, and thus it has limit vectors. The existence of a single limit vector of the sequence $\{y_{2k}\}$ in the iteration \eqref{eqn:yk}--\eqref{eqn:forsythe1} is the content of the original Forsythe conjecture (see Section~\ref{sec:forsythe}), and here we formulate the same conjecture for the ACI($s$):
\medskip{}
\textsc{ACI($s$) conjecture.} \emph{Let $A\in\mathbb{R}^{n\times n}$, let $s$ be an integer with $1\leq s<d(A)$, and let $v_{0}\in\mathbb{R}^{n}$ be such that $\|v_0\|=1$ and $d(A,v_{0})\geq s+1$. Then the sequence $\{v_{k}\}$ in \eqref{eqn:twk}--\eqref{eqn:vkp1} has a single limit vector.} \medskip{}
As shown in Theorem~\ref{thm:evenodd}, the (Euclidean) distance between consecutive vectors of the sequence $\{v_k\}$ shrinks to zero for $k\rightarrow\infty$. Because of this property it may be difficult to find a counterexample for the conjecture numerically. On the other hand, this property is not sufficient for the existence of a single limit vector. For example, the complex points $\mu_k=e^{\mathbf{i} s_k\pi}$, where $s_k=\sum_{j=1}^k 1/j$ for $k\geq 1$, satisfy
$$|\mu_k-\mu_{k+1}|=|1-e^{\mathbf{i} \pi/(k+1)}|\rightarrow 0\quad\mbox{for}\quad k\rightarrow\infty,$$
but the sequence $\{\mu_k\}$ does not converge, since $\{s_k\}$ diverges.
Using a similar notation as Forsythe in~\cite{For68}, we define for the given (possibly nonsymmetric) matrix $A\in\mathbb{R}^{n\times n}$ and integer $s$, where $1\leq s<d(A)$, the set
\[
\Sigma^{A}:=\{v\in\mathbb{R}^{n}\,:\,\|v\|=1\;\mbox{and}\;d(A,v)\geq s+1\},
\]
and the transformation
\begin{align}
T_{A}:\Sigma^{A}\rightarrow\Sigma^{A},\quad v\mapsto T_{A}(v):=\frac{\monic{A}{v}{v}}{\|\monic{A}{v}{v}\|}.
\end{align}
In the notation $\Sigma^A$ and $T_A$ we have suppressed the dependence on~$s$ for simplicity. However, for the analysis of the iteration \eqref{eqn:twk}--\eqref{eqn:vkp1}, it is important to explicitly consider the dependence on the matrix, since we operate with both $A$ and $A^T$.
In terms of \eqref{eqn:twk}--\eqref{eqn:vkp1}, if $v_0\in\Sigma^A$, then
$$v_{k+1}=T_{A^T}(T_{A}(v_{k}))\quad\mbox{for all $k\geq0$,}$$
where Theorem~\ref{thm:monotone} shows that the transformation $T_{A^T}\circ T_A:\Sigma^A\rightarrow\Sigma^A$ is indeed well defined. Moreover, both $T_{A^T}$ and $T_A$ are continuous, so that $T_{A^T}\circ T_{A}$ is also continuous.
The next result adapts~\cite[Theorems~3.8]{For68} to our context.
\medskip{}
\begin{theorem}\label{thm:limit-set}
Let $A\in\mathbb{R}^{n\times n}$, let $s$ be an integer with $1\leq s<d(A)$, and let $v_{0}\in\mathbb{R}^{n}$ be such that $\|v_0\|=1$ and $d(A,v_{0})\geq s+1$. Then the set
$\Sigma_{*}^{A}$ of limit vectors of the sequence $\{v_{k}\}$ in \eqref{eqn:twk}--\eqref{eqn:vkp1} satisfies:
\begin{itemize}
\item[(1)] $\Sigma_{*}^{A}$ is a closed and connected set in $\mathbb{R}^{n}$.
\item[(2)] $\Sigma_{*}^{A}\subseteq\Sigma^{A}$, and each $v_{*}\in\Sigma_{*}^{A}$
satisfies $v_{*}=T_{A^T}(T_{A}(v_{*}))$.
\end{itemize}
\end{theorem}
\begin{proof}
(1) Suppose that the sequence $\{v_{k}\}$ has two isolated limit vectors, $v_{*,1}$ and $v_{*,2}$. Then there exist neighborhoods $U_{1}$ and $U_{2}$ of $v_{*,1}$ and $v_{*,2}$, respectively, which do not contain other limit vectors of the sequence. Let $\{v_{k_{i}}\}$
be set of vectors in the sequence that is contained in $U_{1}$. Then the sequence $\{v_{k_{i}}\}$ converges to $v_{*,1}$, since this sequence can have only one limit vector. We know that the distance $\|v_{k+1}-v_{k}\|$ converges to zero for $k\rightarrow\infty$; see Theorem~\ref{thm:evenodd}. Therefore almost all (i.e., all except finitely many) successors of vectors in the set $\{v_{k_{i}}\}$ are contained in $U_{1}$. But this implies that only a finite number of the vectors of the sequence $\{v_{k}\}$ can be outside $U_{1}$, which contradicts that $v_{*,2}$ is a limit vector. Consequently, the sequence $\{v_{k}\}$ cannot have two isolated limit vectors.
(2) If $v_{*}\in\Sigma_{*}^{A}$, then there exists a subsequence $\{v_{k_{i}}\}$ of the sequence $\{v_{k}\}$ that converges to $v_{*}$. For each vector $v_{k_{i}}$ in the subsequence we have $T_{A^T}(T_{A}(v_{k_{i}}))=v_{k_{i}+1}$, and therefore
$$\|T_{A^T}(T_{A}(v_{k_{i}}))-v_{k_{i}}\|=\|v_{k_{i}+1}-v_{k_{i}}\|\rightarrow 0\quad \mbox{for $k\rightarrow\infty$};$$
see Theorem~\ref{thm:evenodd}. Since $T_{A^T}\circ T_{A}$ is continuous and $v_{k_{i}}\rightarrow v_{*}$,
we have $T_{A^T}(T_{A}(v_{k_{i}}))\rightarrow T_{A^T}(T_{A}(v_{*}))$, and hence $v_{*}=T_{A^T}(T_{A}(v_{*}))$.
It is clear that every $v_{*}\in\Sigma_{*}^{A}$ satisfies $\|v_{*}\|=1$. If $d(A,v_{*})\leq s$, then Lemma~\ref{lem:basic} implies $T_{A^T}(T_{A}(v_{*}))=0$, in contradiction to $T_{A^T}(T_{A}(v_{*}))=v_{*}$. Thus, $d(A,v_{*})\geq s+1$, which shows that $\Sigma_{*}^{A}\subseteq\Sigma^{A}$.
\end{proof}
\section{Results for special cases}\label{sec:special}
In this section we will first derive some general results about the ACI($s$) for symmetric matrices. Then we prove the ACI($1$) conjecture for symmetric, and for real orthogonal matrices with $d(A)=n$
and eigenvalues having only positive (or only negative) real parts.
\subsection{The ACI($s$) for symmetric matrices}\label{sec:4.1}
If $A=A^T\in\mathbb{R}^{n\times n}$, the steps \eqref{eqn:twk}--\eqref{eqn:wk} and \eqref{eqn:tvk}--\eqref{eqn:vkp1} in the ACI($s$) are identical, and hence we can write the algorithm in following simpler form:
\begin{align}
& \mbox{For \ensuremath{k=0,1,2,\dots}}\nonumber \\
& \hspace{1cm}\mbox{\ensuremath{\widetilde{v}_{k+1}=\monic{A}{v_k}{v_k}},}\label{eqn:tvk-simple}\\
& \hspace{1cm}\mbox{\ensuremath{v_{k+1}=\widetilde{v}_{k+1}/\|\widetilde{v}_{k+1}\|}.}\label{eqn:vk-simple}
\end{align}
Our conjecture now is that for each integer $s$ with $1\leq s<d(A)$, and unit norm initial vector $v_{0}\in\mathbb{R}^{n}$ with $d(A,v_{0})\geq s+1$, the sequence $\{v_{2k}\}$ in \eqref{eqn:tvk-simple}--\eqref{eqn:vk-simple} has a single limit vector.
Theorem~\ref{thm:monotone} for $A=A^T$ and the simplified algorithm \eqref{eqn:tvk-simple}--\eqref{eqn:vk-simple} says that
$$\|\widetilde{v}_k\|\leq \|\widetilde{v}_{k+1}\|,\quad k=0,1,2,\dots,$$
with equality if and only if $v_k=\alpha v_{k+2}$ for some $\alpha\neq 0$.
Knowing that $v_k$ and $v_{k+2}$ have unit norm, and that $\langle v_k,v_{k+2}\rangle > 0$ (see \eqref{eqn:ineq1}), we must have $\alpha =1$, i.e., $v_k=v_{k+2}$.
This can actually happen, as shown in the next result, which adapts~\cite[Theorem~4.8]{For68} to our context.
\medskip{}
\begin{theorem}
\label{thm:collinear} Let $A=A^{T}\in\mathbb{R}^{n\times n}$, let $s$ be an integer with $1\leq s<d(A)$, and let $v_{0}\in\mathbb{R}^{n}$ be such that $\|v_0\|=1$ and $d(A,v_{0})=s+1$. Then the vectors $v_{0},v_{2},v_{4},\dots$ in \eqref{eqn:tvk-simple}--\eqref{eqn:vk-simple} are all equal.
\end{theorem}
\begin{proof}
First note that if $d(A,v_0)=s+1$, then the subspace $\mathcal{K}_{s+1}(A,v_{0})$ is $A$-invariant. Therefore, all vectors $v_k$ are contained in the same $s+1$ dimensional space, and by Theorem \ref{thm:monotone} it holds that $d(A,v_{k})=s+1$ for all $k\geq 0$. As a consequence,
$$
\dim \mathcal{K}_{s}(A,v_{k}) = s,\quad k = 0,1,\dots .
$$
It suffices to show that $v_{0}= v_{2}$, then the equality of all vectors $v_{0},v_{2},v_{4},\dots$ follows inductively.
By construction, $v_1\perp\mathcal{K}_{s}(A,v_{0})$ and $ v_2\perp\mathcal{K}_{s}(A,v_{1})$, where the second orthogonality condition is equivalent to
$v_1\perp\mathcal{K}_{s}(A,v_{2})$. Therefore,
$$
v_1\;\perp\;\mathcal{K}_{s}(A,v_{0})\quad\mbox{and}\quad v_1\;\perp\;\mathcal{K}_{s}(A,v_{2}).
$$
Since $\dim \mathcal{K}_{s}(A,v_{0}) = \dim \mathcal{K}_{s}(A,v_{2}) = s$
and since
all vectors are contained in the same $s+1$ dimensional space,
we have $\mathcal{K}_{s}(A,v_{0})=\mathcal{K}_{s}(A,v_{2})$. Hence,
\[
v_{2}=\sum_{j=0}^{s-1}\gamma_{j}A^{j}v_{0},
\]
for some coefficients $\gamma_{0},\dots,\gamma_{s-1}\in\mathbb{R}$.
We will show that $\gamma_{0}=1$ and $\gamma_{j}=0$ for $j>0$.
Using $v_2 \perp A^i v_{1}$ for $i=0,\dots,s-1$ we obtain
\[
0=\langle A^i v_{1},v_{2}\rangle=
\langle v_{1},A^i v_{2}\rangle =
\sum_{j=0}^{s-1}\gamma_{j}\langle v_{1},A^{j+i}v_{0}\rangle,\quad i=0,\dots,s-1.
\]
Now realize that
$\langle v_{1},A^{m}v_{0}\rangle= 0$ for $m=0,\dots,s-1$
and $\langle v_{1},A^{m}v_{0}\rangle \neq 0$ for $m\geq s$
because $A^m v_0 \in {\mathcal K}_{s+1}(A,v_0)\,\backslash\,{\mathcal K}_{s}(A,v_0)$.
In particular, for $i=1$ we obtain
\[
0=\langle A v_{1},v_{2}\rangle=\gamma_{s-1}\langle v_{1},A^{s}v_{0}\rangle.
\]
Since $\langle v_{1},A^{s}v_{0}\rangle\neq 0$, we get $\gamma_{s-1}=0$. Proceeding by induction we obtain $\gamma_{s-2}=\cdots=\gamma_{1}=0$. Therefore, $v_{2}=\gamma_{0}v_{0}$. Since $v_0$ and $v_{2}$ are unit norm vectors and $\langle v_0,v_{2}\rangle > 0$ (see \eqref{eqn:ineq1}) we get $\gamma_0=1$.
\end{proof}
\medskip{}
We next prove a result about the limit vectors of the sequence \eqref{eqn:tvk-simple}--\eqref{eqn:vk-simple}. The part about their grades adapts~\cite[Theorem~4.7]{For68} to our context.
\medskip{}
\begin{theorem}\label{thm:limit-degree}
Let $A=A^T\in\mathbb{R}^{n\times n}$, let $s$ be an integer with $1\leq s<d(A)$, and let $v_{0}\in\mathbb{R}^{n}$ be such that $\|v_0\|=1$ and $d(A,v_{0})\geq s+1$.
Then each limit vector $v_*$ of \eqref{eqn:tvk-simple}--\eqref{eqn:vk-simple} satisfies
\begin{equation}\label{eqn:interpolation}
(\Qpol{A}{v_*}-\tau^2 I)v_{*}=0,
\end{equation}
where $\tau$ is the limit value of the sequence $\{\|\widetilde{v}_k\|\}$ (as in~\eqref{eqn:same-limit}),
$$
\Qpol{z}{v_*} := \mpol{z}{w_*}\mpol{z}{v_*}, \quad\mbox{and}\quad w_* := \frac{\monic{A}{v_*}{v_*}}{\tau},
$$
so that, in particular, $s<d(A,v_{*})\leq 2s$.
\end{theorem}
\smallskip
\begin{proof}
We use the notation of Theorem~\ref{thm:limit-set}. It is clear that each limit vector $v_*\in\Sigma^{A}_*$ satisfies $d(A,v_*)>s$, so we only need to show that $d(A,v_*)\leq 2s$. Moreover, since $A=A^T$, we have $T_A^2(v_{*})=v_{*}$.
By construction,
$$
T_A(v_*)=\frac{\monic{A}{v_*}{v_*}}{\tau}
$$ for a uniquely determined polynomial $\mpol{z}{v_*}\in\mathcal{M}_{s}$, and
$$
T_A^{2}(v_*)=\frac{\monic{A}{w_*}{w_*}}{\tau}, \quad w_*=T_A(v_*),
$$
for a uniquely determined polynomial $\mpol{z}{w_*}\in\mathcal{M}_{s}$; cf. Lemma~\ref{lem:basic}. Note that
$\tau$ is independent of the limit vector $v_*$. Thus,
\[
T_A^{2}(v_*)=\frac{\mpol{A}{w_*}\monic{A}{v_*}{v_*}}{{\tau^2}} =:
\frac{\monicQ{A}{v_*}{v_*}}{\tau^2}
\]
for a uniquely determined polynomial $\Qpol{z}{v_*}\in\mathcal{M}_{2s}$. Finally, using $T_A^{2}(v_{*})=v_{*}$ we obtain $(\Qpol{A}{v_*}-{\tau^2} I)v_{*}=0$, and hence $d(A,v_{*})\leq2s$.
\end{proof}
\medskip{}
Consider an orthogonal diagonalization of $A$, i.e., $A=U\Lambda U^T$, where $\Lambda={\rm diag}(\lambda_1,\dots,\lambda_n)$ and $U$ is orthogonal.
Let $v_* = U \nu$ where $\nu = [\nu_1,\dots,\nu_n]^T \in \mathbb{R}^n$, is the coordinate vector of $v_*$ in the eigenvector basis. Then the condition \eqref{eqn:interpolation} can also be written in the form
$$
\Qpol{\lambda_i}{v_*}\nu_i = {\tau^2} \nu_i,\quad i=1,\dots,n.
$$
If $d(A,v_*)= d$, then there are exactly~$d$ nonzero coordinates $\nu_{k_i}$, $i=1,\dots,d$,
and the polynomial $\Qpol{z}{v_*}$ satisfies the~$d$ interpolation conditions
$$
\Qpol{\lambda_{k_i}}{v_*}= \tau^2,\quad i=1,\dots,d.
$$
In particular, if $d=2s$, then $\Qpol{z}{v_*}$ is uniquely determined by the corresponding interpolation conditions, so that
$$
\Qpol{z}{v_*} = \prod_{i=1}^{2s} (z-\lambda_{k_i}) + \tau^2.
$$
Since there exist only finitely many combinations of $2s$ distinct eigenvalues of~$A$, there are only finitely many polynomials $\Qpol{z}{v_*}$ that correspond to limit vectors~$v_*$ having degree $2s$.
The set $\Sigma_{*}^{A}$ of limit vectors of the sequence~$\{v_{2k}\}$, is a closed and connected set on the unit sphere; see Theorem~\ref{thm:limit-set}. Therefore, in order to show that $\Sigma_{*}^{A}$ consist of a single vector, it is sufficient to show that $\Sigma_{*}^{A}$ contains only finitely many vectors. The existence of only finitely many limit polynomials $\Qpol{z}{v_*}$ can potentially be used to prove that there can be only finitely many limit vectors $v_*$, thereby obtaining a proof of the ACI($s$) conjecture for symmetric matrices and general~$s$.
We point out that the convergence of the coefficients of the iteration polynomials and its consequences for the convergence of the sequence $\{v_{2k}\}$ is an essential ingredient in the work of Zhuk and Bondarenko~\cite{ZhuBon83} on the original Forsythe conjecture and the case $s=2$. In particular, they show that for $s=2$ the coefficients of the monic polynomials $\Qpol{z}{v_k}$ converge to their limit values. Quoting a paper by Zabolotskaya, they use as a proven fact that the convergence of the polynomial coefficients \emph{implies} the existence of a single limit vector of the even iterates; see~\cite[property~4, p.~429]{ZhuBon83}. However, the English translation of Zabolotskaya's paper~\cite[p.~238]{Zab79} states that the convergence of the polynomial coefficients would only \emph{indicate} the existence of a single limit vector. This may well be an imprecise translation of Zabolotskaya's Russian original, but it would nevertheless be very useful to have a more transparent proof of this essential property. Until then we consider the Forsythe conjecture for the case $s=2$ to be still open.
\subsection{Proof of the ACI($1$) conjecture for symmetric matrices}
\label{sec:s1}
If $s=1$, then Theorem~\ref{thm:limit-degree} shows that every limit vector $v_*$ of the sequence $\{v_{2k}\}$ in \eqref{eqn:tvk-simple}--\eqref{eqn:vk-simple} satisfies $d(A,v_*)=2$, i.e., $v_*$ is a linear combination of exactly two linearly independent eigenvectors of $A$. This observation is essential for proving the existence of a single limit vector of the sequence $\{v_{2k}\}$ for $s=1$.
Consider the ACI($1$) for $A=A^T\in\mathbb{R}^{n \times n}$, $s=1$. In this case we have (cf. \eqref{eqn:arnoldi})
\[
w=\monic{A}{v}{v}=Av-\alpha v\;\perp\;{\rm span}\{v\}.
\]
Thus, $0=v^{T}w=v^{T}Av-\alpha v^{T}v$, which yields
\[
\alpha=\frac{v^{T}Av}{v^{T}v},
\]
i.e., $\alpha$ is the \emph{Rayleigh quotient} of $A$ and $v$. Therefore, the algorithm \eqref{eqn:tvk-simple}--\eqref{eqn:vk-simple} can be written as follows:
\begin{align}
& \mbox{For \ensuremath{k=0,1,2,\dots}}\nonumber \\
& \hspace{1cm}\widetilde{v}_{k+1}=(A-\rho_k I)v_{k},\quad \rho_{k}=v_{k}^{T}Av_{k},\label{eqn:s1-1}\\
& \hspace{1cm}v_{k+1}=\widetilde{v}_{k+1}/\|\widetilde{v}_{k+1}\|.\label{eqn:s1-2}
\end{align}
Let $A=U\Lambda U^{T}$ be an orthogonal diagonalization of $A$ with $\Lambda={\rm diag}(\lambda_{1},\dots,\lambda_{n})$. Then $\rho_k=(U^Tv_k)\Lambda (U^T v_k)$, and in \eqref{eqn:s1-1} we can write $U^T\widetilde{v}_{k+1}=(\Lambda-\rho_{k}I)U^{T}v_{k}$. This shows that without loss of generality we can consider the behavior of the \eqref{eqn:s1-1}--\eqref{eqn:s1-2} for a diagonal matrix~$A$. Moreover, without loss of generality we will assume that $A$ has $n$ distinct eigenvalues.
\smallskip
\begin{theorem}\label{thm:s1}
Let $A={\rm diag}(\lambda_{1},\dots,\lambda_{n})\in\mathbb{R}^{n\times n}$
with $\lambda_{1}<\lambda_{2}<\cdots<\lambda_{n}$, and let $v_{0}$ be a unit norm initial vector with $d(A,v_0)\geq 2$. Then the sequence $\{v_{2k}\}$ in \eqref{eqn:s1-1}--\eqref{eqn:s1-2} converges to a single limit vector.
\end{theorem}
\smallskip
\begin{proof}
It is
sufficient to show that $\Sigma_{*}^{A}$ contains only finitely many vectors.
Let $v_*\in \Sigma_{*}^{A}$. We know that there exists a fixed $\tau>0$ (independent of~$v_*$) with $\|\widetilde{v}_{k}\|=\|Av_k-(v_k^TAv_k)v_k\|\rightarrow\tau$ for $k\rightarrow\infty$; see \eqref{eqn:same-limit}. Thus,
\begin{equation}\label{eqn:tau-eq1}
\tau=\left\Vert Av_{*}-\left(v_{*}^{T}Av_{*}\right)v_{*}\right\Vert.
\end{equation}
Moreover, Theorem~\ref{thm:limit-degree} implies that $d(A,v_*)=2$, and hence $v_*=\alpha e_i+\beta e_j$ for some canonical basis vectors $e_{i}$ and $e_{j}$, $1\leq i< j\leq n$, and nonzero $\alpha,\beta\in\mathbb{R}$ that satisfy $\alpha^2+\beta^2=1$.
Now suppose that some vector $v=\alpha e_i+\beta e_j$ with $\alpha^2+\beta^2=1$ satisfies the equation \eqref{eqn:tau-eq1}. Then
\begin{align*}
\tau^{2} &=\left\Vert Av-\left(v^{T}Av\right)v\right\Vert ^{2}=
v^{T}A^{2}v-\left(v^{T}Av\right)^{2} \\ &=\alpha^{2}\lambda_{i}^{2}+\beta^{2}\lambda_{j}^{2}-\left(\alpha^{2}\lambda_{i}+\beta^{2}\lambda_{j}\right)^{2}\\
& = \alpha^{2}\lambda_{i}^{2}(1-\alpha^{2})+\beta^{2}\lambda_{j}^{2}(1-\beta^{2})-2\alpha^{2}\beta^{2}\lambda_{i}\lambda_{j}\\
& = \alpha^{2}\beta^{2}\left(\lambda_{i}^{2}-2\lambda_{i}\lambda_{j}+\lambda_{j}^{2}\right)\\
& = \alpha^{2}\beta^{2}\left(\lambda_{i}-\lambda_{j}\right)^{2}\\
& = \alpha^{2}\left(1-\alpha^{2}\right)\left(\lambda_{i}-\lambda_{j}\right)^{2}.
\end{align*}
There are only finitely many combinations of distinct $i,j\in\{1,2,\dots,n\}$, and for each such combination there are only finitely many values of $\alpha\in\mathbb{R}$ that satisfy
$$\alpha^{2}\left(1-\alpha^{2}\right)=\frac{\tau^2}{\left(\lambda_{i}-\lambda_{j}\right)^{2}}.$$
Consequently, there are only finitely many vectors of the form $v=\alpha e_i+\beta e_j$ with $\alpha^2+\beta^2=1$ that satisfy the equation \eqref{eqn:tau-eq1}. Therefore there can be only finitely many $v_*\in \Sigma_{*}^{A}$, which shows that the sequence $\{v_{2k}\}$ converges to a single limit vector.
\end{proof}
\medskip
Afanasjew, Eiermann, Ernst, and G\"uttel~\cite[Section~3]{AfaEieErnGue08} also have shown that the sequence $\{v_{2k}\}$ in \eqref{eqn:s1-1}--\eqref{eqn:s1-2} converges to a single limit vector. Similar to the original proof of Forsythe~\cite{For68}, their proof is based on first showing that each limit vector is a linear combination of $e_1$ and $e_n$ (or eigenvectors corresponding to the smallest and largest eigenvalue of~$A$), and then showing that there can be only finitely many such combinations. Variations of this approach have appeared also in other proofs for symmetric positive definite matrices and $s=1$; see, e.g.,~\cite{Aka59,GonSch16}. The approach is longer and more technical than our proof of Theorem~\ref{thm:s1}, but the results give more information about the structure of the limiting vectors.
\subsection{On the ACI($1$) conjecture for orthogonal matrices}\label{sec:ACI-orth}
In this section we first prove the ACI($1$) conjecture for orthogonal matrices $A\in\mathbb{R}^{n\times n}$ with $0\notin F(A)$, the field of values of~$A$, and we then comment on the behavior when $0\in F(A)$.
Let $A\in\mathbb{R}^{n\times n}$ be orthogonal. Starting from a unit norm vector $v_0\in\mathbb{R}^n$ with $d(A,v_0)\geq s+1$, the ACI($1$) is as follows:
\begin{align}
& \mbox{For $k=0,1,2,\dots$}\nonumber\\
& \hspace{1cm}\widetilde{w}_k = (A-\alpha_k I)v_k,\quad \alpha_k=v_{k}^TAv_k,\label{eqn:ACI1a}\\
& \hspace{1cm}w_k = \widetilde{w}_k / \|\widetilde{w}_k\|,\\
& \hspace{1cm}\widetilde{v}_{k+1} = (A^T-\beta_k I)w_k,\quad \beta_k=w_{k}^TA^Tw_k=w_{k}^TAw_k,\\
& \hspace{1cm}v_{k+1} =\widetilde{v}_k/\|\widetilde{v}_k\|.\label{eqn:ACI1b}
\end{align}
The Rayleigh quotients $\alpha_k$ and $\beta_k$ in the ACI($1$) are real elements in $F(A)$. Moreover,
\begin{align}
\|\widetilde{w}_k\|^2 &= v_{k}^T (A^T-\alpha_k I)(A-\alpha_k I)v_k=1-\alpha_k^2,
\label{eqn:norw}
\\
\|\widetilde{v}_{k+1}\|^2 &= w_{k}^T (A-\beta_k I)(A^T-\beta_k I)w_k=1-\beta_k^2,
\label{eqn:norv}
\end{align}
where we have used that $A$ is orthogonal. By construction, $\|\widetilde{w}_k\|>0$ and $\|\widetilde{v}_{k+1}\|>0$, so that $\alpha_k,\beta_k\in (-1,1)$.
Since $\|\widetilde{w}_k\|\leq \|\widetilde{v}_{k+1}\|$, we always have
\begin{equation}\label{eqn:betalpha}
|\beta_k| \leq |\alpha_k|\,.
\end{equation}
We know from \eqref{eqn:same-limit} that the sequences $\{\|\widetilde{w}_k\|\}$ and $\{\|\widetilde{v}_k\|\}$ converge to the same limit. Therefore, $\alpha_k^2-\beta_k^2\rightarrow 0$, so that there exist $\alpha,\beta\in (-1,1)$ with
\begin{equation}\label{eqn:alpha-lim}
|\alpha|=|\beta|,\quad \alpha_k\rightarrow\alpha,\quad\mbox{and}\quad \beta_k\rightarrow\beta.
\end{equation}
Here $\alpha$ and $\beta$ are independent of the limit vectors of the sequences $\{v_k\}$ and $\{w_k\}$.
It is well known that an orthogonal matrix~$A\in \mathbb{R}^{n\times n}$ can be orthogonally block-diagonalized as $A=UGU^T$ with
\begin{equation*}
U=[U_1,\dots,U_m,u_1,\dots,u_k],\quad G={\rm diag}(G_1,\dots,G_m,[\pm 1],\dots,[\pm 1]),
\end{equation*}
where $U^TU=I$, $U_1,\dots,U_m\in\mathbb{R}^{n\times 2}$, $u_1,\dots u_k\in\mathbb{R}^n$, and
\begin{equation}\label{eqn:Dj}
G_j=\begin{bmatrix} c_j & s_j\\ -s_j & c_j\end{bmatrix}\in\mathbb{R}^{2\times 2},
\quad c_j^2+s_j^2=1,\quad s_j\neq 0,\quad j=1,\dots,m.
\end{equation}
The blocks $G_j$ in \eqref{eqn:Dj} correspond to the non-real eigenvalues of~$A$, which occur in complex conjugate pairs $c_j\pm \mathbf{i}\, s_j$ with real parts $c_j\in (-1,1)$. By transforming the iterates of the ACI($1$) with $U^T$ (similarly to the transformation for symmetric matrices in Section~\ref{sec:s1}) we can assume without loss of generality that~$A$ is in the block-diagonal form, i.e., that $A={\rm diag}(G_1,\dots,G_m,[\pm 1],\dots,[\pm 1])$.
We will now assume that~$A\in\mathbb{R}^{n\times n}$ is orthogonal with $d(A)=n$ and $0\notin F(A)$. Then in the orthogonal block-diagonalization of $A$ there can be at most one block of size $1\times 1$, either $[1]$ or $[-1]$, and $c_1,\dots,c_m$ are pairwise distinct and either all positive or all negative. For simplicity, we will state and prove the next results only for the positive case.
\smallskip
\begin{lemma}\label{lem:orthogonal1}
Let $A={\rm diag}(G_1,\dots,G_m,[1])\in\mathbb{R}^{n\times n}$, where possibly the block $[1]$ does not occur, have blocks $G_j$ as in \eqref{eqn:Dj} with $0<c_1<\cdots<c_m$. If $v_0\in\mathbb{R}^n$ is any unit norm initial vector with $d(A,v_0)\geq 2$, then any limit vector $v_*\in\Sigma_*^A$ of the sequence $\{v_k\}$ in \eqref{eqn:ACI1a}--\eqref{eqn:ACI1b} satisfies $d(A,v_*)=2$, and $\alpha=\beta$ is equal to the real part of an eigenvalue of~$A$.
\end{lemma}
\smallskip
\begin{proof}
Let $v_*$ be any limit vector of the sequence $\{v_k\}$. It is clear that $d(A,v_*)\geq 2$, so it suffices to show that $d(A,v_*)\leq 2$.
According to Theorem~\ref{thm:limit-set} we have $v_{*}=T_{A^T}(T_{A}(v_{*}))$, i.e.,
\begin{align*}
& \widetilde{w}_* = (A-\alpha I)v_*,\quad \alpha=v_{*}^TAv_*,\\
& w_* = \widetilde{w}_* / \|\widetilde{w}_*\|,\\
& \widetilde{v}_{*} = (A^T-\beta I)w_*,\quad \beta=w_{*}^TA^Tw_*=w_{*}^TAw_*,\\
& v_{*} =\widetilde{v}_*/\|\widetilde{v}_*\|.
\end{align*}
Moreover, \eqref{eqn:norw}--\eqref{eqn:norv} and \eqref{eqn:alpha-lim} yield $\|\widetilde{v}_*\|^2=\|\widetilde{w}_*\|^2=1-\alpha^2$, so that
\begin{align*}
v_* &=T_{A^T}(T_{A}(v_{*}))=
\frac{(A^T-\beta I)(A-\alpha I)v_*}{\|\widetilde{w}_*\|\, \|\widetilde{v}_*\|}=\frac{((1+\alpha\beta)I-\alpha A^T-\beta A)v_*}{1-\alpha^2}.
\end{align*}
This yields the equation
\begin{equation}\label{eqn:v-eq}
(\alpha A^T+\beta A)v_*=(\alpha^2+\alpha\beta)v_*.
\end{equation}
Since $\alpha$ and $\beta$ are real elements in~$F(A)$ with $|\alpha|=|\beta|$, we must have $\alpha=\beta\in(0,1)$. Then \eqref{eqn:v-eq} implies that the limit vector $v_*$ satisfies
\begin{equation}\label{eqn:v-eq-1}
\frac12 (A+A^T)v_*=\alpha v_*.
\end{equation}
Note that
$$\frac12 (A+A^T)={\diag}(c_1 I_2,\dots, c_m I_2, [1]),$$
where possibly the block $[1]$ does not occur. Since $c_1,\dots,c_m\in (0,1)$ are pairwise distinct, we must have $\alpha=c_j$ for exactly one index $j\in\{1,\dots,m\}$. Consequently, $v_*=[0,\dots,0,y^T,0,\dots,0]^T$ for some unit norm vector $y\in\mathbb{R}^2$ corresponding to the $j$th block, which shows that $d(A,v_*)\leq 2$.
\end{proof}
\medskip
For the given orthogonal matrix $A={\rm diag}(G_1,\dots,G_m,[1])\in\mathbb{R}^{n\times n}$, where possibly the block $[1]$ does not occur, we will consider the corresponding block-partitioning of the vectors $v_k$, i.e.,
\begin{equation}\label{eqn:partitioning}
v_k = \begin{bmatrix}
v_k^{(1)}\\
\vdots\\
v_k^{(m)}\\
v_k^{(m+1)}
\end{bmatrix},\quad v_k^{(j)}\in \mathbb{R}^2, \quad j=1,\dots,m,\quad v_k^{(m+1)} \in \mathbb{R},
\end{equation}
where possibly the block $v_k^{(m+1)}$ does not occur.
\medskip
\begin{lemma}\label{lem:c1}
Let $A={\rm diag}(G_1,\dots,G_m,[1])\in\mathbb{R}^{n\times n}$, where possibly the block $[1]$ does not occur, have blocks $G_j$ as in \eqref{eqn:Dj} with $0<c_1<\cdots<c_m$. If
$v_0\in\mathbb{R}^n$ is any unit norm initial vector with $d(A,v_0)\geq 2$ and $v_0^{(1)} \neq 0$, then any limit vector $v_*\in\Sigma_*^A$ of the sequence $\{v_k\}$ in \eqref{eqn:ACI1a}--\eqref{eqn:ACI1b} has zero entries except for the block $v_*^{(1)}$, $\| v_*^{(1)}\|=1$, and $\alpha=\beta=c_1$.
\end{lemma}
\smallskip
\begin{proof}
The ACI($1$) yields
\begin{align*}
v_{k+1}=\frac{(A^{T}-\beta_{k}I)(A-\alpha_{k}I)}{\|\widetilde{w}_{k}\|\,\|\widetilde{v}_{k+1}\|}\,v_{k},
\quad k=0,1,2,\dots.
\end{align*}
If the block $[1]$ occurs, then
$$v_{k+1}^{(m+1)}=\frac{(1-\beta_k)(1-\alpha_k)}{\sqrt{1-\beta_{k}^{2}}\sqrt{1-\alpha_{k}^{2}}}v_k^{(m+1)}
=\left(\frac{(1-\beta_k)(1-\alpha_k)}{(1+\beta_k)(1+\alpha_k)}\right)^{1/2}v_k^{(m+1)},\ \
k=0,1,2,\dots,$$
where we have used \eqref{eqn:norw}--\eqref{eqn:norv}. Since the factor that multiplies $v_k^{(m+1)}$ is less than~$1$, we have $v_k^{(m+1)}\rightarrow 0$ and hence $v_*^{(m+1)}=0$.
The assertion is trivial if $m=1$, so we can assume that $m>1$. For the other blocks we have the equation
\begin{equation}\label{eqn:blocks}
v_{k+1}^{(j)}=
\frac{(G_{j}^{T}-\beta_{k}I)(G_{j}-\alpha_{k}I)}{\sqrt{1-\beta_{k}^{2}}\sqrt{1-\alpha_{k}^{2}}}\,v_{k}^{(j)},
\quad j=1,\dots,m.
\end{equation}
Taking the squared norm in \eqref{eqn:blocks} and using the fact that
\[
(G_{j}^{T}-\delta I)(G_{j}-\delta I)=(1+\delta^{2}-2c_{j}\delta)I
\]
holds for any real $\delta$, we obtain
\begin{align*}
\|v_{k+1}^{(j)}\|^{2} & = \left(\frac{1+\beta_{k}^{2}-2\beta_{k}c_{j}}{1-\beta_{k}^{2}}\right)\left(\frac{1+\alpha_{k}^{2}-2\alpha_{k}c_{j}}{1-\alpha_{k}^{2}}\right)\|v_{k}^{(j)}\|^{2}\\
& = \left(1+2\beta_{k}\frac{\beta_{k}-c_{j}}{1-\beta_{k}^{2}}\right)\left(1+2\alpha_{k}\frac{\alpha_{k}-c_{j}}{1-\alpha_{k}^{2}}\right)\|v_{k}^{(j)}\|^{2}.
\end{align*}
We will now prove by contradiction that
$\beta_{k}\rightarrow c_{1}$ and $\alpha_{k}\rightarrow c_{1}$.
Suppose that $\alpha_{k}\rightarrow c_{\ell}$ and $\beta_{k}\rightarrow c_{\ell}$ for some $\ell>1$. Using \eqref{eqn:betalpha} we know that
$$
\alpha_{k}\geq\beta_{k} \geq \alpha_{k+1} \geq \beta_{k+1} \geq \dots \geq c_{\ell},
$$
and hence, for $j=1$,
\[
\|v_{k+1}^{(1)}\|\geq
\left( 1+2c_{\ell}\frac{c_{\ell}-c_{1}}{1-c_{\ell}^{2}} \right)
\|v_{k}^{(1)}\|\geq
\left( 1+2c_{\ell}\frac{c_{\ell}-c_{1}}{1-c_{\ell}^{2}} \right)^{k+1}
\|v_{0}^{(1)}\|.
\]
But since $1+2c_{\ell}\frac{c_{\ell}-c_{1}}{1-c_{\ell}^{2}}>1$ and $v_0^{(1)}\neq 0$, this implies $\|v_{k}^{(1)}\|\rightarrow\infty$, in contradiction to the normalization of the vectors $v_k$. Therefore, $\beta_{k}\rightarrow c_{1}$ and $\alpha_{k}\rightarrow c_{1}$, and Lemma~\ref{lem:orthogonal1} yields that $v_*$ has the required form.
\end{proof}
\medskip{}
Our next goal is show that under the assumptions of Lemma~\ref{lem:c1} there is only one uniquely determined vector in $\Sigma_*^A$.
In the following lemma we show that norms of the blocks $v_{k}^{(j)}$ for $j>1$ converge to zero at least linearly.
\smallskip
\begin{lemma}\label{lem:unit0}
If $A$ and $v_0$ are as in Lemma~\ref{lem:c1}, then exist an index $k_{0}$
and $0<\varrho<1$, such that for all $k\geq k_{0}$,
\[
\|v_{k+1}^{(j)}\|\leq \varrho \,\|v_{k}^{(j)}\|,\quad j=2,\dots,m+1.
\]
\end{lemma}
\begin{proof}
We know from the proof of Lemma~\ref{lem:c1} that
\begin{equation*
\|v_{k+1}^{(j)}\|^{2} \,= \, \zeta_{k}^{(j)} \|v_{k}^{(j)}\|^{2},
\quad j=1,\dots,m+1,
\end{equation*}
where
\begin{equation}\label{eqn:zetaj}
\zeta_{k}^{(j)}
=
\left(1+2\beta_{k}\frac{\beta_{k}-c_{j}}{1-\beta_{k}^{2}}\right)\left(1+2\alpha_{k}\frac{\alpha_{k}-c_{j}}{1-\alpha_{k}^{2}}\right),\quad j=1,\dots,m,
\end{equation}
and
\begin{equation}\label{eqn:zetam}
\zeta_{k}^{(m+1)} = \frac{(1-\beta_k)(1-\alpha_k)}{(1+\beta_k)(1+\alpha_k)}.
\end{equation}
Using that $\alpha_{k}\geq\beta_{k}\geq\alpha_{k+1}\geq\beta_{k+1}\geq\dots\geq c_{1}$ we obtain
$$
\zeta_k^{(j)} \rightarrow \left(1+2c_{1}\frac{c_{1}-c_{j}}{1-c_{1}^{2}}\right)^2
\;\; j=1,\dots,m,
\quad\mbox{and}\quad
\zeta_k^{(m+1)} \rightarrow \left(\frac{1-c_1}{1+c_1}\right)^2,
$$
where the limit values are both strictly less than 1.
Consequently there exist $k_{0}\geq 0$ and $0<\varrho<1$ such that
\[
\|v_{k+1}^{(j)}\|^{2}\leq\varrho^{2}\|v_{k}^{(j)}\|^{2},\quad j=2,\dots,m+1,
\]
for all $k\geq k_{0}$.
\end{proof}\smallskip
Since we are interested only in convergence of the sequence of vectors, we can assume for simplicity and without loss of generality that $k_{0}=0$ in Lemma~\ref{lem:unit0}. For all $k\geq 0$ we then have
\begin{eqnarray*}
1 & = & \|v_{k+1}\|^{2}=\|v_{k+1}^{(1)}\|^{2}+\sum_{j=2}^{m+1}\|v_{k+1}^{(j)}\|^{2}
\leq \|v_{k+1}^{(1)}\|^{2}+\varrho^{2}\sum_{j=2}^{m+1}\|v_{k}^{(j)}\|^{2}\\
& \leq & \|v_{k+1}^{(1)}\|^{2}+\varrho^{2(k+1)}\sum_{j=2}^{m+1}\|v_{0}^{(j)}\|^{2}.
\end{eqnarray*}
Therefore,
\begin{equation}\label{eqn:key}
1-\|v_{k+1}^{(1)}\|^{2}\leq\varrho^{2(k+1)}.
\end{equation}
The next result shows that the sequence $\{v_{k}^{(1)}\}\subset {\mathbb R}^2$ converges.
\smallskip
\begin{lemma} \label{lem:unit1}
Let $A$ and $v_0$ be as in Lemma~\ref{lem:c1}, and assume without loss of generality that $k_0=0$ in Lemma~\ref{lem:unit0}. Then, for all $k\geq 0$,
\[
\|v_{k+1}^{(1)}-v_{k}^{(1)}\|\leq \varrho^{k}.
\]
\end{lemma}
\begin{proof}
We know that
\[
\|v_{k+1}^{(1)}-v_{k}^{(1)}\|^{2}=\|v_{k+1}^{(1)}\|^{2}+\|v_{k}^{(1)}\|^{2}
-2\left(v_{k+1}^{(1)}\right)^{T}\left(v_{k}^{(1)}\right),
\]
where
\begin{eqnarray*}
\left(v_{k+1}^{(1)}\right)^{T}\left(v_{k}^{(1)}\right) & = & \left(v_{k}^{(1)}\right)^{T}\frac{(G_{1}^{T}-\beta_{k}I)(G_{1}-\alpha_{k}I)}{\sqrt{1-\alpha_{k}^{2}}\sqrt{1-\beta_{k}^{2}}}\,v_{k}^{(1)}\\
& = & \frac{1-c_{1}\left(\alpha_{k}+\beta_{k}\right)+\alpha_{k}\beta_{k}}{\sqrt{1-\alpha_{k}^{2}}\sqrt{1-\beta_{k}^{2}}}\left\Vert v_{k}^{(1)}\right\Vert ^{2}\\
& \geq &
\frac{1-\beta_k\left(\alpha_{k}+\beta_{k}\right)+\alpha_{k}\beta_{k}}{\sqrt{1-\alpha_{k}^{2}}\sqrt{1-\beta_{k}^{2}}}\left\Vert v_{k}^{(1)}\right\Vert ^{2} \geq
\left\Vert v_{k}^{(1)}\right\Vert ^{2} .
\end{eqnarray*}
Note that the next to last inequality follows from $\beta_k \geq c_1 > 0$. Therefore
\begin{eqnarray*}
\|v_{k+1}^{(1)}-v_{k}^{(1)}\|^{2}&\leq&\|v_{k+1}^{(1)}\|^{2}-\|v_{k}^{(1)}\|^{2}
\leq 1-\|v_{k}^{(1)}\|^{2} \leq \varrho^{2k},
\end{eqnarray*}
where we have used \eqref{eqn:key}.
\end{proof}
\smallskip
Now we are ready to prove the convergence theorem.
\smallskip
\begin{theorem} \label{thm:convergence}
Let $A={\rm diag}(G_{1},\dots,G_{m},[1])\in\mathbb{R}^{n\times n}$, $m\geq1$, be an orthogonal matrix with blocks $G_{j}$ as in \eqref{eqn:Dj}, and $0<c_{1}<c_{2}<\cdots<c_{m}<1$, where possibly the block $[1]$ does not occur. Let $v_{0}\in\mathbb{R}^{n}$ be any unit norm initial vector with $d(A,v_{0})\geq2$ such that $v_{0}^{(1)}\neq0$. Then the sequence $\{v_{k}\}$ in \eqref{eqn:ACI1a}--\eqref{eqn:ACI1b} converges to a single limit vector.
\end{theorem}
\begin{proof}
Using Lemmas~\ref{lem:unit0} and \ref{lem:unit1}, and assuming without loss of generality that $k_0=0$ in Lemma~\ref{lem:unit0}, we obtain for all $k\geq 0$,
\begin{eqnarray*}
\|v_{k+1}-v_{k}\|^{2} &=& \|v_{k+1}^{(1)}-v_{k}^{(1)}\|^{2}+\sum_{j=2}^{m+1}\|v_{k+1}^{(j)}-v_{k}^{(j)}\|^{2}\\
&\leq& \varrho^{2k}+2\varrho^{2k}\sum_{j=2}^{m+1}\|v_{0}^{(j)}\|^{2}
\,\leq\, 3\varrho^{2k}
\end{eqnarray*}
which implies $\sum_{k=0}^{\infty}\|v_{k+1}-v_{k}\|<\infty$, and hence
finishes the proof.
\end{proof}
\smallskip
We now consider an orthogonal matrix $A\in\mathbb{R}^{n\times n}$ with $d(A)=n$ and $0\in F(A)$. Analogously to Lemma~\ref{lem:c1} we will show that (with appropriate assumptions on the initial vector~$v_0$) the sequences of the Rayleigh quotients $\{\alpha_k\}$ and $\{\beta_k\}$ in \eqref{eqn:ACI1a}--\eqref{eqn:ACI1b} converge to $\min_{z\in F(A)} \mathrm{Re}(z)$. Under the assumption $0\in F(A)$ this means that $\alpha=\beta=0$ in \eqref{eqn:alpha-lim}.
We can again assume without loss of generality that $A$ is block diagonal,
\[
A={\rm diag}([-1],G_{1},\dots,G_{m},[1])\in\mathbb{R}^{n\times n},\quad m\geq 1,
\]
with blocks $G_{j}$ as in \eqref{eqn:Dj}, and $-1<c_{1}<c_{2}<\cdots<c_{m}<1$, where the blocks $[-1]$ or $[1]$ possibly do not occur. Let us also formally define $c_0 = -1$ and $c_{m+1}=1$. We will consider
a block-partitioning of the vectors $v_{k}$ as in \eqref{eqn:partitioning}, where we add the block $v_{k}^{(0)}$ if the block $[-1]$ occurs.
\smallskip
\begin{lemma} \label{lem:zeroFA}
Let $A={\rm diag}([-1],G_{1},\dots,G_{m},[1])\in\mathbb{R}^{n\times n}$,
$m\geq1$, be an orthogonal matrix with blocks $G_{j}$ as in \eqref{eqn:Dj},
and $-1=c_0<c_{1}<c_{2}<\cdots<c_{m}<1=c_{m+1}$, where possibly the blocks $[-1]$
or $[1]$ do not occur. Let $v_{0}\in\mathbb{R}^{n}$ be a unit norm initial vector with $d(A,v_{0})\geq2$. Suppose that $v_{0}^{(\ell)}\neq0$ for some $\ell$, $1\leq \ell\leq m$, and that $v_{0}^{(j)}\neq0$ for some $j$, $0\leq j\leq m+1$, such that $c_{\ell}c_j \leq 0$. Then $\alpha=\beta=0$ in \eqref{eqn:alpha-lim}.
\end{lemma}
\smallskip
\begin{proof}
From \eqref{eqn:betalpha} and \eqref{eqn:alpha-lim} we know that $|\alpha_{k}|\geq|\beta_{k}|$ and that $|\alpha_{k}|\rightarrow\gamma$ and $|\beta_{k}|\rightarrow\gamma$. Since $d(A,v_{0})\geq2$ and
$v_{0}^{(\ell)}\neq0$ for some $1\leq \ell\leq m$, we have $|\alpha_{0}|<1$, and therefore $0\leq\gamma<1$. We will prove by contradiction that $\gamma=0$. Let $\gamma>0$, i.e., $\alpha_{k}\rightarrow\alpha$ and $\beta_{k}\rightarrow\beta$, where $1>|\alpha|=|\beta|=\gamma>0$.
Suppose first that $\alpha=-\beta\neq0$, and consider a block
for which $v_{0}^{(\ell)}\neq0$, $1\leq \ell\leq m$. Then (see \eqref{eqn:zetaj})
\begin{eqnarray*}
\zeta_k^{(\ell)} \rightarrow\left(1+2\alpha\frac{\alpha+c_{j}}{1-\alpha{}^{2}}\right)\left(1+2\alpha\frac{\alpha-c_{j}}{1-\alpha^{2}}\right) & = & 1+\frac{4\alpha^{2}}{1-\alpha{}^{2}}\left(\frac{1-c_{j}^{2}}{1-\alpha{}^{2}}\right)>1,
\end{eqnarray*}
Therefore $\|v_{k}^{(\ell)}\|\rightarrow\infty$, which is a contradiction.
Suppose now that $1>\alpha=\beta>0$, and consider a block
with $c_j\leq 0$ such that $v_{0}^{(j)}\neq0$, $0\leq j \leq m$.
If $j=0$, then
\[
\zeta_k^{(0)} \rightarrow \frac{(1+\alpha)^{2}(1+\alpha)^{2}}{\left(1-\alpha^{2}\right)\left(1-\alpha^{2}\right)}>1,
\]
and if $j>0$, then
\[
\zeta_k^{(j)} \rightarrow \left(1+2\frac{\alpha^{2}-\alpha c_{j}}{1-\alpha^{2}}\right)^{2}>1
\]
giving a contradiction. Note that if $-1<\beta=\alpha<0$, then we derive a contradiction by considering
a block with $c_j\geq 0$ such that $v_{0}^{(j)}\neq0$.
\end{proof}
\smallskip
In order to prove that the sequence $\{v_{k}\}$ converges to a single limit vector also
when $0\in F(A)$, it would be sufficient to show that there exists $0<\varrho<1$ such that
\[
\ensuremath{\|v_{k+1}-v_{k}\|=\mathcal{O}\left(\varrho^{k}\right)}
\]
for $k$ sufficiently large; cf. the proof of Theorem~\ref{thm:convergence}. With some effort we can show that there is a constant $C>0$ such that
\[
\|v_{k+1}-v_{k}\|\leq C |\alpha_{k}|
\]
for $k$ sufficiently large. It remains to prove that the coefficients $\alpha_{k}$ converge to zero linearly, but this steps needs further investigation.
We will now show by an example that for an orthogonal matrix $A$ with $0\in F(A)$ a limit vector $v_*$ of the sequence $\{v_{k}\}$ in \eqref{eqn:ACI1a}--\eqref{eqn:ACI1b} may satisfy $d(A,v_*)>2$. This is a significant difference to the case $0\notin F(A)$, where $d(A,v_*)=2$ holds for any limit vector $v_*$; see Lemma~\ref{lem:orthogonal1}.
\smallskip
\begin{example}
Let $A={\rm diag}(G_1,G_2)\in\mathbb{R}^{4\times 4}$ with blocks $G_{j}$ as in \eqref{eqn:Dj}, where $c_1\in (-1,0)$ and $c_2=-c_1$, so that in particular $0\in F(A)$. Let $v_0\in \mathbb{R}^4$ be any unit norm vector with $v_0^{(1)}=v_0^{(2)}$. Then $d(A)=d(A,v_0)=4$, and in \eqref{eqn:ACI1a}--\eqref{eqn:ACI1b} we get $\alpha_0=v_0^TAv_0=0$, and hence
\begin{align*}
\widetilde{w}_0 &=Av_0,\quad \|\widetilde{w}_0\|=1,\quad w_0=\widetilde{w}_0=Av_0,\quad \beta_0 =w_0^TAw_0=v_0^TAv_0=0,\\
\widetilde{v}_1&=A^Tw_0=v_0,\quad v_1=v_0.
\end{align*}
Consequently, $v_k=v_0$ and $w_k=Av_0$ for all $k\geq 0$. The unique limit point of the sequence $\{v_k\}$ is given by $v_*=v_0$, and $d(A,v_*)=4$.
\end{example}
\medskip
The observation in this example that a limit vector $v_*$ satisfies $d(A,v_*)=n$ is not surprising if we look at the proof of Lemma~\ref{lem:zeroFA}: Since $\alpha=0$, we have
\[
\zeta_{k}^{(j)}\rightarrow 1,\quad j=0,1,\dots,m+1,
\]
which would indicate that $v_*^{(j)}\neq 0$ if $v_{0}^{(j)}\neq 0$, and even $d(A,v_*)=d(A,v_0)$. However, we did not prove that this holds in general.
Finally, recall that the ideal Arnoldi problem for a matrix $A\in\mathbb{R}^{n\times n}$ and $s=1$ is given by
$$\min_{\alpha\in\mathbb{R}}\|A-\alpha I\|;$$
see \eqref{eqn:idealArnoldi}. A straightforward computation shows that
$$\min_{\alpha\in\mathbb{R}}\|A-\alpha I\|=\max_{\substack{v\in\mathbb{R}^n\\ \|v\|=1}} (\|Av\|^2-\langle v,Av\rangle^2)^{1/2}.$$
If $A\in\mathbb{R}^{n\times n}$ is normal and has the eigenvalues $\lambda_1,\dots,\lambda_n\in {\mathbb C}$, then
$$\min_{\alpha\in\mathbb{R}}\|A-\alpha I\|=\min_{\alpha\in\mathbb{R}}\max_{1\leq j\leq n} |\lambda_j-\alpha|,$$
and the unique solution of this problem is given by the center of the closed disk of smallest radius in the complex plane that contains $\lambda_1,\dots,\lambda_n$; see~\cite[Section~2.4]{FabLieTic10}.
If $A$ and $v_0$ are as in Lemma~\ref{lem:c1}, then $c_1$ is the center of that disk, and for the unique limit vector $v_*$ we have
$$\|Av_*\|^2-\langle v_*,Av_*\rangle^2=\|G_1v_*^{(1)}\|^2-\langle v_*^{(1)},G_1v_*^{(1)}\rangle^2=1-c_1^2,$$
so that
$$\min_{\alpha\in\mathbb{R}}\|A-\alpha I\|=\|A-c_1 I\|=\|Av_*-c_1v_*\|=(1-c_1^2)^{1/2}.$$
On the other hand, if $A$ and $v_0$ are as in Lemma~\ref{lem:zeroFA}, then $0$ is the center of that disk. Any $v_*\in\Sigma_*^A$ then satisfies $v_*^TAv_*=0$, which gives $\|Av_*\|^2-\langle v_*,Av_*\rangle^2=1$, and $\min_{\alpha\in\mathbb{R}}\|A-\alpha I\|=\|A\|=\|Av_*\|=1$.
In short, any $v_*\in\Sigma_*^A$ solves the ideal Arnoldi problem for an orthogonal matrix~$A\in\mathbb{R}^{n\times n}$ and $s=1$ (when $v_0$ satisfies the appropriate assumptions.) Note that this observation gives no insight into the uniqueness of the limit vectors.
\section{Concluding remarks}\label{sec:Conclusions}
In this paper we have introduced and analyzed the ACI($s$) in order to generalize the Forsythe conjecture from symmetric positive definite matrices to symmetric and nonsymmetric matrices. We were able to prove the existence of a single limit vector of the sequence $\{v_{2k}\}$ in \eqref{eqn:s1-1}--\eqref{eqn:s1-2} (the case $s=1$ and $A=A^T$), and the sequence $\{v_k\}$ in \eqref{eqn:ACI1a}--\eqref{eqn:ACI1b} (the case $s=1$ and orthogonal~$A$) when $0\notin F(A)$. Our uniqueness proof for $s=1$ and $A=A^T$ is much simpler than other previously published proofs.
In the case $s=1$, the property of monotonically increasing norms in Theorem~\ref{thm:evenodd}, and its proof based on orthogonality and the Cauchy-Schwarz inequality, appear to the be closely related to the property of monotonically decreasing residual norms in the Rayleigh quotient iteration (for symmetric matrices) as well as Ostrowski's two-sided iteration and Parlett's two-sided iteration (for general matrices). The monotonic residual norms ultimately yield the global convergence of these iterations; see, e.g.,~\cite{Par74}. Working out the exact relations between the iterations studied in the context of the Forsythe conjecture and the Rayleigh quotient based iterations remains a subject of further research.
Also, the Forsythe conjecture and its generalization to the ACI($s$) still remain largely open.
\section*{Acknowledgements} We thank G\'erard Meurant for helpful comments.
\bibliographystyle{siam}
| 315339ef16766a313c3525424e3a77663124f10c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Many model-based computational imaging methods assume that data acquisition can be represented as a linear system of equations, such that
\begin{equation}
\mathbf{b} = \mathbf{A}\mathbf{f} + \mathbf{n},
\end{equation}
where $\mathbf{b}$ is the length-$M$ vector of measured data samples, $\mathbf{f}$ is the length-$N$ vector of unknown image voxel values, $\mathbf{A}$ is an $M\times N$ matrix modeling the data acquisition operator, and $\mathbf{n}$ is a length-$M$ vector of additive noise perturbations. Based on this forward model, it is relatively popular to formulate image reconstruction as a regularized linear least-squares optimization problem:
\begin{equation}
\hat{\mathbf{f}} = \arg\min_\mathbf{f} \frac{1}{2}\|\mathbf{A}\mathbf{f} - \mathbf{b}\|_2^2 + R(\mathbf{f}),\label{eq:complex}
\end{equation}
where $\|\cdot\|_2$ is the standard Euclidean ($\ell_2$) norm, and $R(\cdot)$ is a regularization functional that is designed to reduce ill-posedness by encouraging the reconstructed image to have desirable characteristics. For example, it is common to encourage the reconstructed image to be spatially-smooth by choosing $R(\cdot)$ as a penalty on the norm of the image gradient or as a penalty on the norm of the high-frequency subbands in a wavelet representation of the image.
Although the image $\mathbf{f}$ may be real-valued in many scenarios (i.e., $\mathbf{f} \in \mathbb{R}^N$), there are several imaging modalities (including certain kinds of optical imaging, ultrasound imaging, magnetic resonance imaging (MRI), synthetic aperture radar, etc.) for which the image of interest is most naturally modeled as complex-valued (i.e., $\mathbf{f} \in \mathbb{C}^N$). In such cases, the image magnitude and image phase often represent manifestations of different physical phenomena, and have distinct spatial characteristics from one another that are not easily captured by a single unified regularization penalty. This has led several authors, working on image reconstruction across a variety of different modalities and application domains, to consider optimization problems in which the magnitude and phase of the image are regularized separately \cite{cetin2001,fessler2004,tuysuzoglu2012,JustinDecovSLIM2011,FesslerSepMagPhs2012,guven2016, ZibettiSepMagPhs2017,LustigPhaseCycling2018,moradikia2018,haldar2019}.
While these approaches were developed independently in many cases, the optimization problems can largely be unified by the following general formulation:
\begin{equation} \label{eq:GeneralReconModel}
\{\hat{\mathbf{m}},\hat{\mathbf{p}}\} = \arg\min_{\substack{\mathbf{m} \in \mathbb{R}^N \\\mathbf{p} \in \mathbb{R}^N}} \tilde{J}(\mathbf{m,p}) + R_1(\mathbf{m}) + \tilde{R}_2(\mathbf{p}),
\end{equation}
where
\begin{equation}\label{eq:F}
\tilde{J}(\mathbf{m,p}) \triangleq \frac{1}{2}\|\mathbf{A}\left(\mathbf{m}\odot e^{i\mathbf{p}}\right) - \mathbf{b}\|_2^2.
\end{equation}
In this expression, $\mathbf{m}$ and $\mathbf{p}$ respectively represent the magnitude\footnote{Note that this formulation allows the ``magnitude'' vector $\mathbf{m}$ to have negative entries, which can avoid unnecessary phase discontinuities in $\mathbf{p}$ \cite{ FesslerSepMagPhs2012}.} and phase of the complex image $\mathbf{f}$ such that $\mathbf{f} = \mathbf{m}\odot e^{i\mathbf{p}}$, and $R_1(\cdot)$ and $\tilde{R}_2(\cdot)$ respectively represent separate regularization penalties for the magnitude and phase. We have used the notation $\odot$ to represent the Hadamard product (elementwise multiplication) of two vectors, have used the notation $e^\mathbf{z}$ to denote the elementwise exponentiation of the vector $\mathbf{z}$, and have used $i$ to denote the unit imaginary number ($i=\sqrt{-1}$).
On one hand, Eq.~\eqref{eq:GeneralReconModel} is desirable because it allows independent customization of the regularization penalties applied to the magnitude and phase components of the image, which has previously been shown to be beneficial \cite{cetin2001,fessler2004,tuysuzoglu2012,JustinDecovSLIM2011,FesslerSepMagPhs2012, guven2016,ZibettiSepMagPhs2017,LustigPhaseCycling2018,moradikia2018,haldar2019}. On the other hand, this formulation can be substantially more difficult to solve than Eq.~\eqref{eq:complex}. For example, while Eq.~\eqref{eq:complex} is convex if $R(\cdot)$ is chosen to be convex (and therefore can be globally optimized from arbitrary initializations using standard optimization methods!), the formulation in Eq.~\eqref{eq:GeneralReconModel} is generally non-convex and is thus more challenging to optimize. As a result, many authors have decided to make use of a simple alternating minimization (AM) strategy that alternates between optimizing $\mathbf{m}$ for a fixed value of $\mathbf{p}$ and optimizing $\mathbf{p}$ for a fixed value of $\mathbf{m}$ \cite{fessler2004,JustinDecovSLIM2011,FesslerSepMagPhs2012,ZibettiSepMagPhs2017, LustigPhaseCycling2018, haldar2019}. However, while AM can successfully decrease the objective function, it is not necessarily computationally efficient.
In this work, we propose and evaluate a new algorithm named PALMNUT to solve Eq.~\eqref{eq:GeneralReconModel}. This algorithm is based on combining the proximal alternating linearized minimization (PALM) algorithm \cite{BoltePALM2014} with Nesterov's momentum \cite{NesterovBook2004} and a novel approach that uses uncoupled coordinatewise step sizes derived from coordinatewise Lipschitz-like bounds. PALMNUT is evaluated in the context of several different MRI-related inverse problems, where it is demonstrated to outperform popular existing methods. A preliminary account of portions of this work was previously presented in a recent conference \cite{liuISMRM2019}.
\section{Background}\label{sec:back}
The following subsections review the AM approach (one of the dominant existing algorithms for solving Eq.~\eqref{eq:GeneralReconModel}) as well as the related PALM algorithm (which serves as the foundation for our proposed PALMNUT approach).
Since both AM and PALM are generic algorithms that are broadly applicable beyond just the magnitude and phase problem from Eq.~\eqref{eq:GeneralReconModel}, we will start by introducing both of these algorithms in a more general setting before specializing to our problem of interest. Specifically, we will consider the generic optimization problem of the form
\begin{equation}\label{eq:AMcost}
\{\hat{\mathbf{x}},\hat{\mathbf{y}}\} = \arg\min_{\substack{\mathbf{x} \in \mathbb{R}^{N_{1}} \\ \mathbf{y} \in \mathbb{R}^{N_{2}}}} \Psi(\mathbf{x,y}),
\end{equation}
where $\mathbf{x} \in \mathbb{R}^{N_{1}}$ and $\mathbf{y} \in \mathbb{R}^{N_{2}}$ are optimization variables, and the objective function $\Psi(\cdot,\cdot)$ can be decomposed as
\begin{equation}\label{eq:PALMcost}
\Psi(\mathbf{x,y}) = H(\mathbf{x},\mathbf{y}) + F(\mathbf{x}) + G(\mathbf{y})
\end{equation}
for some real-valued or extended-real-valued scalar functions $H(\cdot,\cdot)$, $F(\cdot)$, and $G(\cdot)$.
\subsection{Alternating Minimization}
Given the problem setting described above, the AM algorithm proceeds according to Alg.~\ref{alg:AM}.
\begin{algorithm}
\caption{Alternating Minimization for Eq.~\eqref{eq:AMcost}}
\label{alg:AM}
\begin{algorithmic}
\REQUIRE Set $k=1$ and initialize $\hat{\mathbf{x}}_0$ and $\hat{\mathbf{y}}_0$.
\WHILE {not converge}
\vspace{0.05in}
\STATE $\displaystyle \hat{\mathbf{x}}_{k} = \arg\hspace{-0.5em}\min_{\mathbf{x} \in \mathbb{R}^{N_1}} \Psi(\mathbf{x},\hat{\mathbf{y}}_{k-1})$\\
\hspace{2.4em}$\displaystyle = \arg\hspace{-0.5em}\min_{\mathbf{x} \in \mathbb{R}^{N_1}} H(\mathbf{x},\hat{\mathbf{y}}_{k-1})+F(\mathbf{x})$ \\
\STATE $\displaystyle \hat{\mathbf{y}}_{k} = \arg\hspace{-0.5em}\min_{\mathbf{y} \in \mathbb{R}^{N_2}} \Psi(\hat{\mathbf{x}}_{k},\mathbf{y})$\\
\hspace{2.4em}$\displaystyle = \arg\hspace{-0.5em}\min_{\mathbf{y} \in \mathbb{R}^{N_2}} H(\hat{\mathbf{x}}_{k},\mathbf{y})+G(\mathbf{y})$ \\
\STATE $k \gets k + 1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
As can be seen, AM alternates between optimizing the estimate of $\mathbf{x}$ for a fixed estimated value of $\mathbf{y}$ and optimizing the estimate of $\mathbf{y}$ for a fixed estimated value of $\mathbf{x}$. This can be viewed as a block coordinate descent algorithm, and thus has well-studied theoretical characteristics \cite{wright2015}. The algorithm leads to monotonic decrease of the objective function value, and therefore (by the monotone convergence theorem) convergence of the objective function value if $\Psi(\mathbf{x},\mathbf{y})$ is bounded from below. In practice, it is not necessary to solve the optimization subproblems for $\hat{\mathbf{x}}_{k}$ and $\hat{\mathbf{y}}_{k}$ exactly in each step, as the overall objective function will still decrease monotonically as long as the subproblem objective function values always decrease in each iteration.
In the case of optimizing Eq.~\eqref{eq:GeneralReconModel} with AM \cite{fessler2004,JustinDecovSLIM2011,FesslerSepMagPhs2012,ZibettiSepMagPhs2017, LustigPhaseCycling2018, haldar2019}, one should associate $\mathbf{x}$ with $\mathbf{m}$, $\mathbf{y}$ with $\mathbf{p}$, $H(\cdot,\cdot)$ with $\tilde{J}(\cdot,\cdot)$, $F(\cdot)$ with $R_1(\cdot)$, and $G(\cdot)$ with $\tilde{R}_2(\cdot)$.
When defining $\tilde{J}(\mathbf{m},\mathbf{p})$ as in Eq.~\eqref{eq:F}, the magnitude subproblem has the form of a classical regularized linear least-squares problem. As such, there exist many different algorithms to solve this kind of problem efficiently. However, the phase subproblem is more complicated, as it is generally nonlinear and nonconvex. Existing methods have frequently relied on either the nonlinear conjugate gradient (NCG) algorithm \cite{fessler2004,JustinDecovSLIM2011,FesslerSepMagPhs2012, ZibettiSepMagPhs2017,haldar2019} or a phase cycling heuristic \cite{LustigPhaseCycling2018} to address this subproblem.
\subsection{Proximal Alternating Linearized Minimization (PALM)}
The PALM algorithm considers the same setup described previously, but with the additional assumptions that $H(\cdot,\cdot)$ is a smooth real-valued function and $F(\cdot)$ and $G(\cdot)$ are proper and lower semicontinuous (but potentially nonsmooth) extended-real-valued functions. None of these functions are required to be convex.
The structure of PALM is strongly motivated by the AM algorithm, and similar to AM, the PALM algorithm for Eq.~\eqref{eq:AMcost} proceeds by updating the $\mathbf{x}$ and $\mathbf{y}$ variables in alternation. However, rather than directly using the AM update formulas, PALM first applies proximal linearization of $H$ to each subproblem prior to computing each update step. Specifically, PALM uses
\begin{equation}
\begin{split}\label{eq:prox_phs}
\hat{\mathbf{x}}_{k} = \arg\hspace{-0.5em}\min_{\mathbf{x} \in \mathbb{R}^N_{1}} & H(\hat{\mathbf{x}}_{k-1},\hat{\mathbf{y}}_{k-1}) \\ &+ \inner{\mathbf{x} - \hat{\mathbf{x}}_{k-1}, \nabla_\mathbf{x} H(\hat{\mathbf{x}}_{k-1},\hat{\mathbf{y}}_{k-1})} \\
&+ \frac{c_k}{2} \norm{\mathbf{x} - \hat{\mathbf{x}}_{k-1}}_2^2 + F(\mathbf{x}) + G(\hat{\mathbf{y}}_{k-1})
\end{split}
\end{equation}
and
\begin{equation}\begin{split}\label{eq:prox_mag}
\hat{\mathbf{y}}_{k} = \arg\hspace{-0.5em}\min_{\mathbf{y} \in \mathbb{R}^{N_{2}}} &H(\hat{\mathbf{x}}_{k},\hat{\mathbf{y}}_{k-1}) \\ &+ \inner{\mathbf{y} - \hat{\mathbf{y}}_{k-1}, \nabla_\mathbf{y} H(\hat{\mathbf{x}}_{k},\hat{\mathbf{y}}_{k-1})} \\
&+ \frac{d_k}{2} \|\mathbf{y} - \hat{\mathbf{y}}_{k-1}\|_2^2 + F(\hat{\mathbf{x}}_{k}) + G(\mathbf{y}),
\end{split}
\end{equation}
where $c_k$ and $d_k$ are real-valued positive scalars, $\nabla_\mathbf{x}$ represents the gradient with respect to $\mathbf{x}$, and $\inner{\cdot,\cdot}$ denotes the standard dot product.
One way to justify Eqs.~\eqref{eq:prox_phs} and \eqref{eq:prox_mag} is to invoke the majorize-minimize algorithmic framework \cite{hunter2004}. Specifically, for any real-valued objective function $Q(\mathbf{x})$ with $\mathbf{x}\in\mathbb{R}^N$, we say that the function $S_{k}(\mathbf{x})$ is a ``majorant'' of $Q(\mathbf{x})$ at the point $\hat{\mathbf{x}}_{k-1}$ if it satisfies two conditions: (i) $S_{k}(\hat{\mathbf{x}}_{k-1}) = Q(\hat{\mathbf{x}}_{k-1})$; and (ii) $Q(\mathbf{x}) \leq S_{k}(\mathbf{x})$ for $\forall \mathbf{x} \in \mathbb{R}^N$. An important feature of majorant functions is that if $S_{k}(\hat{\mathbf{x}}_{k}) <S_{k}(\hat{\mathbf{x}}_{k-1})$, then it is guaranteed that $Q(\hat{\mathbf{x}}_{k}) < Q(\hat{\mathbf{x}}_{k-1})$. This fact allows the majorant $S_{k}(\mathbf{x})$ to be used as a surrogate for the original objective function $Q(\mathbf{x})$, since descending on $S_{k}(\mathbf{x})$ guarantees descent on $Q(\mathbf{x})$. The use of a surrogate can be beneficial whenever the surrogate is easier to optimize than the original function \cite{hunter2004}.
With these concepts defined, we return to Eqs.~\eqref{eq:prox_phs} and \eqref{eq:prox_mag}. Assume that for a fixed value of $\mathbf{y}$, $\nabla_\mathbf{x} H(\mathbf{x},{\mathbf{y}})$ is globally Lipschitz in $\mathbf{x}$ such that
\begin{equation} \label{eq:PALM_Lipx}
\norm{ \nabla_\mathbf{x} H(\mathbf{x}_1,{\mathbf{y}}) - \nabla_\mathbf{x} H(\mathbf{x}_2,{\mathbf{y}}) }_2 \leq L_1(\mathbf{y}) \norm{ \mathbf{x}_1 - \mathbf{x}_2 }_2
\end{equation}
for $\forall \mathbf{x}_1,\mathbf{x}_2 \in \mathbb{R}^{N_1}$, where $L_1(\mathbf{y})$ is the corresponding Lipschitz constant. Similarly, assume that for a fixed value of $\mathbf{x}$, $\nabla_\mathbf{y} H(\mathbf{x},{\mathbf{y}})$ is globally Lipschitz in $\mathbf{y}$ such that
\begin{equation} \label{eq:PALM_Lipy}
\norm{ \nabla_\mathbf{y} H(\mathbf{x},\mathbf{y}_1) - \nabla_\mathbf{y} H(\mathbf{x},\mathbf{y}_2) }_2 \leq L_2(\mathbf{x}) \norm{ \mathbf{y}_1 - \mathbf{y}_2 }_2
\end{equation}
for $\forall \mathbf{y}_1,\mathbf{y}_2 \in \mathbb{R}^{N_2}$, where $L_2(\mathbf{x})$ is the corresponding Lipschitz constant. Under these assumptions, it can be shown that the objective function in Eq.~\eqref{eq:prox_phs} will be a majorant of $\Psi(\mathbf{x},\hat{\mathbf{y}}_{k-1})$ at the point $\mathbf{x} = \hat{\mathbf{x}}_{k-1}$ whenever $c_k \geq L_1(\hat{\mathbf{y}}_{k-1})$, and the objective function in Eq.~\eqref{eq:prox_mag} will be a majorant of $\Psi(\hat{\mathbf{x}}_{k},\mathbf{y})$ at the point $\mathbf{y} = \hat{\mathbf{y}}_{k-1}$ whenever $d_k \geq L_2(\hat{\mathbf{x}}_{k})$ \cite{BoltePALM2014}. This means that if $c_k$ and $d_k$ are chosen large enough, then Eqs.~\eqref{eq:prox_phs} and \eqref{eq:prox_mag} can be interpreted as majorize-minimize steps. This will guarantee monotonic decrease of the objective function value, and therefore convergence of the objective function value will be guaranteed for PALM if $\Psi(\mathbf{x},\mathbf{y})$ is bounded from below \cite{hunter2004}. Under some additional mild conditions, the iterates of PALM are guaranteed to converge to a critical point of $\Psi(\mathbf{x},\mathbf{y})$ \cite{BoltePALM2014}.
The solutions to the optimization problems in Eqs.~\eqref{eq:prox_phs} and \eqref{eq:prox_mag} can be written in a more concise ``proximal operator'' form \cite{BoltePALM2014}. Specifically, Eq.~\eqref{eq:prox_phs} is equivalent to the ``proximal operator''
\begin{equation} \label{eq:PALM_xUpdate}
\begin{split}
\hat{\mathbf{x}}_{k} &= \arg\hspace{-0.5em}\min_{\mathbf{x} \in \mathbb{R}^{N_1}} F(\mathbf{x}) + \frac{c_k}{2} \norm{ \mathbf{x} - \mathbf{w}_k }_2^2\\
&\triangleq \prox{F}{\mathbf{w}_k,c_k}
\end{split}
\end{equation}
with
\begin{equation}\label{eq:w}
\mathbf{w}_k = \hat{\mathbf{x}}_{k-1} - \left( 1/c_k \right) \nabla_\mathbf{x} H(\hat{\mathbf{x}}_{k-1},\hat{\mathbf{y}}_{k-1}),
\end{equation}
and Eq.~\eqref{eq:prox_mag} is equivalent to the ``proximal operator''
\begin{equation} \label{eq:PALM_yUpdate}
\begin{split}
\hat{\mathbf{y}}_{k} &= \arg\hspace{-0.5em}\min_{\mathbf{y} \in \mathbb{R}^{N_2}} G(\mathbf{y}) + \frac{d_k}{2} \norm{ \mathbf{y} - \mathbf{z}_k }_2^2,\\
&\triangleq \prox{G}{\mathbf{z}_k,d_k}
\end{split}
\end{equation}
with
\begin{equation}
\mathbf{z}_k = \hat{\mathbf{y}}_{k-1} - \left( 1/d_k \right) \nabla_\mathbf{y} H(\hat{\mathbf{x}}_{k},\hat{\mathbf{y}}_{k-1}).
\end{equation}
These representations are useful, because for many common regularization penalties, the corresponding proximal operators often have simple closed-form solutions \cite{parikh2013, beck2017}.
In summary, the PALM algorithm proceeds according to Alg.~\ref{alg:PALM}.
\begin{algorithm}
\caption{PALM for Eq.~\eqref{eq:AMcost}}
\label{alg:PALM}
\begin{algorithmic}
\REQUIRE Set $k=1$ and initialize $\hat{\mathbf{x}}_0$ and $\hat{\mathbf{y}}_0$.
\WHILE {not converge}
\vspace{0.05in}
\STATE Choose $c_k \geq L_1(\hat{\mathbf{y}}_{k-1})$\\
$\mathbf{w}_k = \hat{\mathbf{x}}_{k-1} - \left( 1/c_k \right) \nabla_\mathbf{x} H(\hat{\mathbf{x}}_{k-1},\hat{\mathbf{y}}_{k-1})$\\
$\hat{\mathbf{x}}_{k} = \prox{F}{\mathbf{w}_k,c_k}$ \\[5pt]
\STATE Choose $d_k \geq L_2(\hat{\mathbf{x}}_{k})$ \\
$\mathbf{z}_k = \hat{\mathbf{y}}_{k-1} - \left( 1/d_k \right) \nabla_\mathbf{y} H(\hat{\mathbf{x}}_{k},\hat{\mathbf{y}}_{k-1})$ \\
$\hat{\mathbf{y}}_{k} = \prox{G}{\mathbf{z}_k,d_k}$ \\[5pt]
\STATE $k \gets k + 1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
Applying PALM to Eq.~\eqref{eq:GeneralReconModel} is nontrivial and, to the best of our knowledge, has never been done before. We will describe our implementation (as well as the enhancements needed for PALMNUT) in the next section.
\section{Methods}
This section is organized as follows. In Subsection~\ref{sec:palm}, we first demonstrate how to apply PALM to solve Eq.~\eqref{eq:GeneralReconModel}. Next, Subsection~\ref{sec:unc} will introduce the principles of our novel approach that relies on uncoupled coordinatewise step sizes. Afterwards, Subsection~\ref{sec:palmnut} will describe the incorporation of Nesterov's momentum strategy and describe the full PALMNUT algorithm.
For the sake of concreteness, the remainder of this paper will make the further assumption that the phase regularization term $\tilde{R}_2(\mathbf{p})$ can be written as
\begin{equation}\label{eq:PhsRegAssum}
\tilde{R}_2(\mathbf{p}) = R_2(e^{i\mathbf{p}})
\end{equation}
for some appropriate penalty function $R_2(\cdot)$, such that we are regularizing the exponentiated phase. This choice is beneficial because it means that $\mathbf{p}$ always appears in exponentiated form in every place it appears in the objective function, which will enable simplifications later on. This choice has also been used in previous work \cite{JustinDecovSLIM2011,FesslerSepMagPhs2012,haldar2019} because it offers several other benefits. First, the exponentiated phase $e^{i\mathbf{p}}$ is unique, even though the phase vector $\mathbf{p}$ is never unique due to the 2$\pi$-periodicity of phase. Second, the exponentiated phase $e^{i\mathbf{p}}$ can be spatially-smooth even if the phase $\mathbf{p}$ is nonsmooth due to issues associated with phase wrapping. And finally, the spatial derivatives of the exponentiated phase image can be shown to have the same magnitude as the spatial derivatives of the optimally-unwrapped non-exponentiated phase image under common regularity conditions \cite{ZhiPeiPhaseUnwrap1996}, which is important because phase derivatives are often used by regularization strategies that are designed to promote smooth phase \cite{fessler2004, JustinDecovSLIM2011,FesslerSepMagPhs2012,haldar2019,LustigPhaseCycling2018}.
\subsection{Applying PALM to Eq.~\eqref{eq:GeneralReconModel}}\label{sec:palm}
In order to apply PALM to Eq.~\eqref{eq:GeneralReconModel}, it may be tempting to associate $H(\cdot,\cdot)$ with $\tilde{J}(\cdot,\cdot)$, $F(\cdot)$ with $R_1(\cdot)$, and $G(\cdot)$ with $R_2(\cdot)$ as we had also done for AM. However, it turns out that $\tilde{J}(\cdot,\cdot)$ does not have favorable Lipschitz bounds for the phase subproblem, and a different approach may be preferred.
To see this, note for the magnitude that
\begin{equation}
\begin{split}
&\nabla_{\mathbf{m}} \tilde{J}(\mathbf{m},\mathbf{p}) = \\
&\hspace{5em}\Re \left\{ e^{-i\mathbf{p}} \odot \left[ \mathbf{A}^{H} \mathbf{A} \left( \mathbf{m} \odot e^{i\mathbf{p}} \right) - \mathbf{A}^{H} \mathbf{b} \right] \right\}
\end{split}
\end{equation}
and that
\begin{equation}\label{eq:mlip}
\begin{split}
&\|\nabla_{\mathbf{m}} \tilde{J}(\mathbf{m}_1,\mathbf{p}) - \nabla_{\mathbf{m}} \tilde{J}(\mathbf{m}_2,\mathbf{p})\|_2 \leq \\
&\hspace{2em}\left\|\Re \left\{ \mathrm{diag}(e^{-i\mathbf{p}})\mathbf{A}^H\mathbf{A} \mathrm{diag}(e^{i\mathbf{p}})\right\} \right\| \|\mathbf{m}_1 - \mathbf{m}_2\|_2,
\end{split}
\end{equation}
where $\Re\{\cdot\}$ denotes the operator that extracts the real part of its input, $\| \cdot \|$ denotes the spectral norm, and $\mathrm{diag}(e^{i\mathbf{p}})$ is the square diagonal matrix whose diagonal entries are equal to the elements of $ e^{i\mathbf{p}}$. Due to the characteristics of spectral norms, Eq.~\eqref{eq:mlip} provides a tight Lipschitz bound.
However, for the phase, note that
\begin{equation}\label{eq:gradp}
\begin{split}
&\nabla_{\mathbf{p}} \tilde{J}\left( \mathbf{m,p} \right) = \\
&\hspace{2em}\Im \left\{ e^{-i\mathbf{p}} \odot \mathbf{m}_{k}\odot \left( \mathbf{A}^{H} \mathbf{A} \left( \mathbf{m}_{k} \odot e^{i\mathbf{p}} \right) - \mathbf{A}^{H} \mathbf{b} \right) \right\},
\end{split}
\end{equation}
where $\Im\{\cdot\}$ denotes the operator that extracts the imaginary part of its input. Because of the form of this gradient expression (i.e., $\mathbf{p}$ appears nonlinearly in Eq.~\eqref{eq:gradp}), deriving a good Lipschitz bound for the phase subproblem is nontrivial. We have been partially successful in deriving valid Lipschitz upper bounds for this case (results not shown), but none of the bounds we've derived have been anywhere close to tight. This is problematic for the implementation of PALM, because these loose Lipschitz bounds could cause us to choose $d_k$ values that are much larger than necessary, which will result in smaller-than-necessary step sizes in the phase update problem, which will ultimately lead to slow convergence speed.
To avoid this issue, instead of working directly with the original variable $\mathbf{p}$ and its exponentiated version $e^{i\mathbf{p}}$, we will instead consider an equivalent formulation using the change of variables $\mathbf{q} \triangleq e^{i\mathbf{p}}$ under the constraint that all entries of $\mathbf{q}$ have magnitude one. This allows us to equivalently express Eq.~\eqref{eq:GeneralReconModel} as
\begin{equation} \label{eq:GeneralReconModel2}
\{\hat{\mathbf{m}},\hat{\mathbf{q}}\} = \arg\min_{\substack{\mathbf{m} \in \mathbb{R}^N \\ \mathbf{q} \in \mathcal{V}}} {J}(\mathbf{m,q}) + R_1(\mathbf{m}) + R_2(\mathbf{q}),
\end{equation}
where
\begin{equation}\label{eq:F2}
{J}(\mathbf{m,q}) \triangleq \frac{1}{2}\|\mathbf{A}\left(\mathbf{m}\odot \mathbf{q} \right) - \mathbf{b}\|_2^2
\end{equation}
and
\begin{equation}
\mathcal{V} \triangleq \left\{ \mathbf{q} \in \mathbb{C}^{N}: \abs{q_{n}} = 1,\ n=1,2,...,N \right\}.
\end{equation}
Once $\hat{\mathbf{m}}$ and $\hat{\mathbf{q}}$ are obtained, the corresponding value of $\hat{\mathbf{p}}$ can be obtained, if so desired, by computing the angle of each of the entries of $\mathbf{q}$.
This reformulation is beneficial, because the gradients simplify substantially. For the magnitude, we now have that
\begin{equation} \label{eq:DataFid_Mag_Grad}
\nabla_\mathbf{m} J(\mathbf{m},\mathbf{q}) = \Re\left\{ \bar{\mathbf{q}} \odot \left[\mathbf{A}^H\mathbf{A} (\mathbf{m} \odot \mathbf{q}) - \mathbf{A}^H\mathbf{b} \right] \right\}
\end{equation}
where $\bar{\mathbf{q}}$ denotes the elementwise complex conjugation of $\mathbf{q}$, which gives the tight Lipschitz bound
\begin{equation} \label{eq:DataFid_Mag_Lipschitz}
\begin{split}
&\|\nabla_{\mathbf{m}} {J}(\mathbf{m}_1,\mathbf{q}) - \nabla_{\mathbf{m}} {J}(\mathbf{m}_2,\mathbf{q})\|_2 \leq \\
&\hspace{4em}\norm{\Re \left\{ \mathrm{diag}(\bar{\mathbf{q}})\mathbf{A}^H\mathbf{A} \mathrm{diag}(\mathbf{q})\right\}} \|\mathbf{m}_1 - \mathbf{m}_2\|_2.
\end{split}
\end{equation}
For the phase, we now have that
\begin{equation} \label{eq:DataFid_Phs_Grad}
\nabla_\mathbf{q} J(\mathbf{m},\mathbf{q}) = \mathbf{m} \odot \left[\mathbf{A}^H\mathbf{A} (\mathbf{m} \odot \mathbf{q}) - \mathbf{A}^H\mathbf{b} \right]
\end{equation}
which gives the tight Lipschitz bound
\begin{equation} \label{eq:DataFid_Phs_Lipschitz}
\begin{split}
&\|\nabla_{\mathbf{q}} {J}(\mathbf{m},\mathbf{q}_1) - \nabla_{\mathbf{q}} {J}(\mathbf{m},\mathbf{q}_2)\|_2 \leq \\
&\hspace{8em}\left\| \mathbf{A} \mathrm{diag}(\mathbf{m}) \right\|^2 \|\mathbf{q}_1 - \mathbf{q}_2\|_2.
\end{split}
\end{equation}
Although this reformulation simplifies the calculation of Lipschitz constants, the introduction of the constraint set $\mathcal{V}$ makes the optimization problem more complicated. To alleviate this concern, we will further rewrite the constrained problem from Eq.~\eqref{eq:GeneralReconModel2} in an equivalent unconstrained form using indicator functions \cite{AfonsoADMMsparseRecon2010} as
\begin{equation}\label{eq:GeneralReconModel_uncons}
\begin{split}
\{\hat{\mathbf{m}},\hat{\mathbf{q}}\} = \arg\min_{\substack{\mathbf{m} \in \mathbb{R}^N \\ \mathbf{q} \in \mathbb{C}^N}} {J}(\mathbf{m,q}) &+ R_1(\mathbf{m}) \\
&+ R_2(\mathbf{q}) + \mathcal{I}_{\mathcal{V}}(\mathbf{q}),
\end{split}
\end{equation}
where
\begin{equation}
\mathcal{I_{V}}(\mathbf{q}) \triangleq \begin{cases} 0, & \mathbf{q} \in \mathcal{V} \\ +\infty, & \mathbf{q} \notin \mathcal{V}. \end{cases}
\end{equation}
The new objective function from Eq.~\eqref{eq:GeneralReconModel_uncons} now has four terms in it (i.e., $J(\cdot,\cdot)$, $R_1(\cdot)$, $R_2(\cdot)$, and $\mathcal{I}_\mathcal{V}(\cdot)$), and to apply PALM, it is necessary to associate these with the PALM terms $H(\cdot,\cdot)$, $F(\cdot)$, and $G(\cdot)$. The function $J(\cdot,\cdot)$ is always smooth and involves both $\mathbf{m}$ and $\mathbf{q}$, so we will associate it with $H(\cdot,\cdot)$. The function $\mathcal{I}_\mathcal{V}(\cdot)$ is always nonsmooth and only involves $\mathbf{q}$, so we will associate it with $G(\cdot)$. While these associations are straightforward (we have no other options), we potentially have options for $R_1(\cdot)$ and $R_2(\cdot)$ depending on their smoothness characteristics. If $R_1(\cdot)$ is nonsmooth, then it has to be associated with $F(\cdot)$, but if it is smooth (Lipschitz) then we could either choose to associate it with $F(\cdot)$ or incorporate it into $H(\cdot,\cdot)$. Similarly, if $R_2(\cdot)$ is nonsmooth, then it has to be associated with $G(\cdot)$, but if it is smooth (Lipschitz) then we could either choose to associate it with $G(\cdot)$ or incorporate it into $H(\cdot,\cdot)$.
Although we have different options, the remainder of this paper will assume (for simplicity and without loss of generality) that smooth regularization functions will always be incorporated into $H(\cdot,\cdot)$. This choice leads to function associations that are summarized in Table~\ref{tab:summ}.
\bgroup
\def1.5{1.5}
\begin{table}[h]
\resizebox{\columnwidth}{!}{
\begin{tabular}{>{\centering}m{3.5em} >{\centering}m{3.5em} | >{\centering}m{7.5em} | >{\centering}m{3em} | >{\centering\arraybackslash}m{6.5em} |}
$R_1(\cdot)$ is Smooth? & $R_2(\cdot)$ is Smooth? & $H(\cdot,\cdot)$ & $F(\cdot)$ & $G(\cdot)$ \\ \hline
Y & Y & $J(\mathbf{m},\mathbf{q}) + R_1(\mathbf{m}) + R_2(\mathbf{q}) $ & 0 & $\mathcal{I}_\mathcal{V}(\mathbf{q})$ \\ \hline
N & Y & $J(\mathbf{m},\mathbf{q}) + R_2(\mathbf{q})$ & $R_1(\mathbf{m})$ & $\mathcal{I}_\mathcal{V}(\mathbf{q})$ \\ \hline
Y & N & $J(\mathbf{m},\mathbf{q}) + R_1(\mathbf{q})$ & 0 & $R_2(\mathbf{q}) + \mathcal{I}_\mathcal{V}(\mathbf{q})$\\ \hline
N & N & $J(\mathbf{m},\mathbf{q})$ & $R_1(\mathbf{m})$ & $R_2(\mathbf{q}) + \mathcal{I}_\mathcal{V}(\mathbf{q})$ \\ \hline
\end{tabular}}
\caption{Associations between PALM and Eq.~\eqref{eq:GeneralReconModel_uncons} depending on whether $R_1(\cdot)$ and $R_2(\cdot)$ are smooth (Lipschitz).}
\label{tab:summ}
\end{table}
\egroup
With these assignments, the PALM algorithm from Alg.~\ref{alg:PALM} can be directly applied, although it is still necessary to specify the computation of $c_k$, $d_k$, $\mathrm{prox}_F(\cdot,\cdot)$, and $\mathrm{prox}_G(\cdot,\cdot)$. Although these computations will necessarily depend on the characteristics of $R_1(\cdot)$ and $R_2(\cdot)$, we will provide concrete illustrations of these calculations for two typical choices of regularization penalties (we will use these same choices of regularization penalties in the validation study presented later in the paper). Of course, these two illustrations do not encompass every possibility, and interested readers are referred to Refs.~\cite{CombettesProx2011,parikh2013,beck2017} for further discussion and examples of computating proximal operators. However, it should be noted that the two illustrations below focus on the case where $R_2(\cdot)$ is smooth (the first two rows of Table~\ref{tab:summ}), as we have found that it is frequently nontrivial to derive the $\mathrm{prox}_G(\cdot,\cdot)$ operator when $G(\cdot)$ incorporates both $R_2(\cdot)$ and $\mathcal{I}_\mathcal{V}(\cdot)$ (the last two rows of Table~\ref{tab:summ}).
\subsubsection{Huber-function regularization of $\mathbf{m}$ with Tikhonov regularization of $\mathbf{q}$}
For our first illustration, we will consider the case where magnitude regularization takes the form of either
\begin{equation} \label{eq:Huber_Mag_Tikhonov_Phs}
R_1(\mathbf{m}) = \lambda_1 \sum_{\ell=1}^L h_{\xi}\left(\left[ \mathbf{B} \mathbf{m}\right]_\ell \right)
\end{equation}
or
\begin{equation} \label{eq:Huber_Mag_Tikhonov_Phs2}
R_1(\mathbf{m}) = \lambda_1 \sum_{\ell=1}^L h_{\xi}\left( \sqrt{\sum_{t=1}^T \left|\left[\mathbf{B}_{t} \mathbf{m}\right]_\ell\right|^2} \right),
\end{equation}
and phase regularization takes the form of
\begin{equation}\label{eq:r222}
R_2(\mathbf{q}) = \frac{\lambda_2}{2} \norm{\mathbf{Cq}} _2^2.
\end{equation}
In these expressions, $\lambda_1$ and $\lambda_2$ are positive scalar regularization parameters that can be respectively adjusted to control the influence of the magnitude and phase regularization terms, $[\mathbf{z}]_\ell$ is used to denote the $\ell$th entry of the vector $\mathbf{z}$, and $h_\xi(\cdot)$ is the Huber function defined as
\begin{equation} \label{eq:Huber}
h_{\xi}(t) = \begin{cases}
\frac{1}{2\xi} \abs{t}^2 , & \abs{t} \leq \xi \\
\abs{t} - \frac{1}{2}\xi, & \abs{t} > \xi.
\end{cases}
\end{equation}
The Huber function is a smooth, convex function that is commonly used for both robust statistics (to mitigate the effects of outliers) \cite{huber1981} and for edge/discontinuity/sparsity-preserving image regularization \cite{nikolova2005,black1995}. As can be seen from Eq.~\eqref{eq:Huber}, the Huber function is similar to the $\ell_1$-norm for large values of its argument, but unlike the $\ell_1$-norm, is also smooth at the origin because it behaves like a squared $\ell_2$-norm for small values of its argument. The Huber function with a small value of $\xi$ is frequently chosen as a differentiable surrogate for sparsity-promoting $\ell_1$-norm regularization, while choosing larger values of $\xi$ can make the Huber function more tolerant to smoothly-varying image regions, more resilient to noise, and easier to characterize theoretically \cite{nikolova2005,haldar2013,haldar2011}.
The regularization in Eq.~\eqref{eq:Huber_Mag_Tikhonov_Phs} is based on applying the Huber function to a single transformation $\mathbf{B}\in\mathbb{C}^{L \times N}$ (e.g., a wavelet transform, a finite-difference transform, etc.) of the magnitude vector $\mathbf{m}$. The more general regularization in Eq.~\eqref{eq:Huber_Mag_Tikhonov_Phs2} applies the Huber function to a combination of $T$ different transforms $\mathbf{B}_{t}\in\mathbb{C}^{L \times N}$ of $\mathbf{m}$, which can be useful for imposing additional transform-domain structure. For example, combining a horizontal finite-difference transform with a vertical finite-difference transform within Eq.~\eqref{eq:Huber_Mag_Tikhonov_Phs2} is a common way to achieve isotropic regularization \cite{rudin1991}. In addition, Eq.~\eqref{eq:Huber_Mag_Tikhonov_Phs2} is related to concepts of simultaneous sparsity \cite{tropp2006}, and our previous work has used regularization penalties of this form to impose the constraint that multi-contrast images of the same anatomy will frequently have correlated edge characteristics \cite{haldar2008,haldar2011,haldar2013,haldar2019}. Since Eq.~\eqref{eq:Huber_Mag_Tikhonov_Phs} is a special case of Eq.~\eqref{eq:Huber_Mag_Tikhonov_Phs2} with $T=1$, our description below will assume use of the more general form of Eq.~\eqref{eq:Huber_Mag_Tikhonov_Phs2}.
The regularization in Eq.~\eqref{eq:r222} corresponds to standard quadratic (Tikhonov) regularization. If $\mathbf{C}$ is chosen as a spatial finite-difference operator, this type of regularization can be good at imposing the constraint that the image phase should be spatially smooth without major discontinuities \cite{fessler2004,JustinDecovSLIM2011,haldar2019,FesslerSepMagPhs2012,ZibettiSepMagPhs2017}. While not every MRI image will have smooth phase characteristics, most do, and smooth phase is a common constraint within the image reconstruction literature \cite{liang1992,haldar2020}.
For this case, both $R_1(\cdot)$ and $R_2(\cdot)$ are smooth, corresponding to the situation in the first row of Table \ref{tab:summ}. As such, to implement PALM, we use the assignments:
\begin{equation}
\begin{split}
H(\cdot,\cdot) &\leftarrow \frac{1}{2}\|\mathbf{A}\left(\mathbf{m}\odot \mathbf{q} \right) - \mathbf{b}\|_2^2 \\
& \hspace{0.2in}+ \lambda_1 \sum_{\ell=1}^L h_{\xi}\left( \sqrt{\sum_{t=1}^T \left|\left[\mathbf{B}_{t} \mathbf{m}\right]_\ell\right|^2} \right) \\
&\hspace{0.2in}+ \lambda_2 \norm{\mathbf{Cq}} _2^2 \\
F(\cdot) &\leftarrow 0 \\
G(\cdot) &\leftarrow \mathcal{I}_\mathcal{V}(\mathbf{q}).
\end{split}
\end{equation}
The gradients of $H(\cdot,\cdot)$ needed for PALM are
\begin{equation}
\begin{split}
\nabla_{\mathbf{m}}H(\mathbf{m},\mathbf{q}) = \Re&\left\{ \bar{\mathbf{q}} \odot \left[\mathbf{A}^H\mathbf{A} (\mathbf{m} \odot \mathbf{q}) - \mathbf{A}^H\mathbf{b} \right] \right. \\
&+ \left. \lambda_1 \sum_{t=1}^T \mathbf{B}_t^H \mathbf{W}(\mathbf{m}) \mathbf{B}_t \mathbf{m} \right\}
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
\nabla_{\mathbf{q}}H(\mathbf{m},\mathbf{q}) =\mathbf{m} \odot& \left[\mathbf{A}^H\mathbf{A} (\mathbf{m} \odot \mathbf{q}) - \mathbf{A}^H\mathbf{b} \right] \\
&+\lambda_2 \mathbf{C}^H\mathbf{C}\mathbf{q},
\end{split}
\end{equation}
where $\mathbf{W}(\mathbf{m})$ is an $L\times L$ diagonal matrix depending on $\mathbf{m}$ with $\ell$th diagonal entry given by
\begin{equation}
[\mathbf{W}(\mathbf{m})]_{\ell \ell} = 1/\max\left\{\xi,\sqrt{\sum_{t=1}^T \left|\left[\mathbf{B}_{t} \mathbf{m}\right]_\ell\right|^2} \right\}.
\end{equation}
These gradient expressions give rise to Lipschitz-type upper bounds in the form of Eqs.~\eqref{eq:PALM_Lipx} and \eqref{eq:PALM_Lipy}, such that the majorization and descent conditions will be satisfied whenever
\begin{equation}
c_k \geq \norm{\mathbf{A}}^2+\frac{\lambda_1}{\xi} \norm{\sum_{t=1}^T \mathbf{B}_t^H\mathbf{B}_t}
\end{equation}
and
\begin{equation}\label{eq:d1}
d_k \geq \norm{\mathbf{A}}^2 \norm{\hat{\mathbf{m}}_{k}}_\infty^2+\lambda_2 \norm{\mathbf{C}}^2,
\end{equation}
where $\|\cdot\|_\infty$ denotes the $\ell_\infty$-norm. The spectral norms in these expressions are not iteration-dependent, and can be precomputed and reused throughout the iterative process. If it is difficult to analytically calculate the spectral norm values, they can also be evaluated using standard computationally-efficient numerical methods like Lanczos iteration \cite{golub2013}.
Finally, the proximal operators needed for PALM are given by
\begin{equation}
\prox{F}{\mathbf{w}_k,c_k} = \mathbf{w}_k
\end{equation}
and
\begin{equation}\label{eq:proxg}
\prox{G}{\mathbf{z}_k,d_k} = \frac{\mathbf{z}_k}{|\mathbf{z}_k|},
\end{equation}
where division is performed elementwise and we choose the convention that $\frac{0}{0}=1$.
\subsubsection{$\ell_1$ regularization of $\mathbf{m}$ with Huber-function regularization of $\mathbf{q}$}
For our second illustration, we will consider the case where
\begin{equation} \label{eq:L1_Mag_Huber_Phs}
R_1(\mathbf{m}) = \lambda_1 \norm{\mathbf{Tm}}_1
\end{equation}
and
\begin{equation}\label{eq:L1_Mag_Huber_Phs2}
R_2(\mathbf{q}) = \lambda_2 \sum_{\ell=1}^L h_{\xi}\left( \sqrt{\sum_{t=1}^T \left|\left[\mathbf{B}_{t} \mathbf{q}\right]_\ell\right|^2} \right).
\end{equation}
The $\ell_1$-norm penalty with sparsifying transform matrix $\mathbf{T}$ from Eq.~\eqref{eq:L1_Mag_Huber_Phs} is standard for promoting transform-domain sparsity, and has been previously used to regularize the magnitude vector $\mathbf{m}$ in several applications \cite{cetin2001,tuysuzoglu2012,FesslerSepMagPhs2012,guven2016,ZibettiSepMagPhs2017,LustigPhaseCycling2018,moradikia2018}. The characteristics of the Huber function from Eq.~\eqref{eq:L1_Mag_Huber_Phs2} have been discussed previously. By taking a small value of the parameter $\xi$, the Huber function can be used as a smooth approximation of the $\ell_1$-norm in order to enable sparsity-promoting and/or edge-preserving regularization of the phase image, which can useful for some applications with more complicated phase characteristics \cite{LustigPhaseCycling2018, FesslerSepMagPhs2012}.
In this case, $R_1(\cdot)$ is non-smooth while $R_2(\cdot)$ is smooth, corresponding to the second row of Table \ref{tab:summ}. As such, to implement PALM, we use the assignments:
\begin{equation}
\begin{split}
H(\cdot,\cdot) &\leftarrow \frac{1}{2}\|\mathbf{A}\left(\mathbf{m}\odot \mathbf{q} \right) - \mathbf{b}\|_2^2 \\
& \hspace{0.2in}+ \lambda_2 \sum_{\ell=1}^L h_{\xi}\left( \sqrt{\sum_{t=1}^T \left|\left[\mathbf{B}_{t} \mathbf{q}\right]_\ell\right|^2} \right) \\
F(\cdot) &\leftarrow \lambda_1 \norm{\mathbf{Tm}}_1 \\
G(\cdot) &\leftarrow \mathcal{I}_\mathcal{V}(\mathbf{q}).
\end{split}
\end{equation}
The gradients of $H(\cdot,\cdot)$ needed for PALM are
\begin{equation}
\begin{split}
\nabla_{\mathbf{m}}H(\mathbf{m},\mathbf{q}) = \Re&\left\{ \bar{\mathbf{q}} \odot \left[\mathbf{A}^H\mathbf{A} (\mathbf{m} \odot \mathbf{q}) - \mathbf{A}^H\mathbf{b} \right] \right\}
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
\nabla_{\mathbf{q}}H(\mathbf{m},\mathbf{q}) =\mathbf{m} \odot& \left[\mathbf{A}^H\mathbf{A} (\mathbf{m} \odot \mathbf{q}) - \mathbf{A}^H\mathbf{b} \right] \\
&+ \lambda_2 \sum_{t=1}^T \mathbf{B}_t^H \mathbf{W}(\mathbf{q}) \mathbf{B}_t \mathbf{q}.
\end{split}
\end{equation}
These gradient expressions give rise to Lipschitz-type upper bounds in the form of Eqs.~\eqref{eq:PALM_Lipx} and \eqref{eq:PALM_Lipy}, such that the majorization and descent conditions will be satisfied whenever
\begin{equation}
c_k \geq \norm{\mathbf{A}}^2
\end{equation}
and
\begin{equation}\label{eq:d2}
d_k \geq \norm{\mathbf{A}}^2 \norm{\hat{\mathbf{m}}_{k}}_\infty^2 + \frac{\lambda_2}{\xi} \norm{\sum_{t=1}^T \mathbf{B}_t^H\mathbf{B}_t}.
\end{equation}
As before, the spectral norms in these expressions are not iteration-dependent, and can be precomputed and reused throughout the iterative process.
Finally, assuming that $\mathbf{T}$ is a unitary transform such that $\mathbf{T}^H = \mathbf{T}^{-1}$, the proximal operator for $F(\cdot)$ is given by \cite{beck2017}
\begin{equation} \label{eq:SoftThr}
\begin{split}
\prox{F}{\mathbf{w}_k,c_k} &= \\
&\hspace{-2em}\mathbf{T}^H\mathrm{diag}\left(\frac{\max\left\{|\mathbf{T}\mathbf{w}_k| - \lambda_1/c_k,0\right\}}{|\mathbf{T}\mathbf{w}_k|}\right)\mathbf{T}\mathbf{w}_k,
\end{split}
\end{equation}
where maximization, absolute value, and division operations are performed elementwise. Note that $G(\cdot)$ is the same as in the previous illustration, and therefore has the same proximal operator (Eq.~\eqref{eq:proxg}).
\subsection{Uncoupled Step Sizes}\label{sec:unc}
Although the PALM algorithm described in the previous section provides a novel and effective approach for solving Eq.~\eqref{eq:GeneralReconModel}, we observe in this section that it may be very conservative and computationally inefficient for PALM to use the same value of $d_k$ (and therefore, the same step size $1/d_k$) for all elements of the phase vector $\mathbf{p}$. This inefficiency stems from the fact that $d_k$ is set based on the global Lipschitz constant (effectively, the worst-case rate of change of the gradient along any possible direction), while we have observed that the rate of change of the gradient can be much smaller than the worst-case along specific directions. Concretely, using the global Lipschitz constant means that the step size will depend on the maximum value of $\mathbf{m}$, while we observe that it can be much better for the step size for each coordinate to instead depend on the coordinatewise values of $\mathbf{m}$. This observation motivates us to investigate and utilize coordinatewise bounds on the rate of change of the gradient, enabling uncoupled coordinatewise step sizes.
For the sake of generality, we will first describe this approach for the general setting of Section~\ref{sec:back}, where we are given a generic smooth real-valued function $H(\mathbf{x},\mathbf{y})$. The PALM approach utilized majorants of $H(\mathbf{x},\mathbf{y})$ that were derived based on the global (scalar-valued) Lipschitz constants $L_1(\mathbf{y})$ and $L_2(\mathbf{x})$ of Eqs.~\eqref{eq:PALM_Lipx} and \eqref{eq:PALM_Lipy}. In this section, we instead make the assumption that a vector $\mathbf{L}_1(\mathbf{y}) \in \mathbb{R}^{N_1}$ can be found such that, for a fixed value of $\mathbf{y}$, we have
\begin{equation} \label{eq:Lipd1}
\begin{split}
\inner{ \nabla_\mathbf{x} H(\mathbf{x}_1,{\mathbf{y}}) - \nabla_\mathbf{x} H(\mathbf{x}_2,{\mathbf{y}}), \mathbf{x}_1-\mathbf{x}_2 } & \\
& \hspace{-6em} \leq \norm{\sqrt{\mathbf{L}_1(\mathbf{y})} \odot (\mathbf{x}_1 - \mathbf{x}_2) }_2^2
\end{split}
\end{equation}
for $\forall \mathbf{x}_1,\mathbf{x}_2 \in \mathbb{R}^N_{1}$, where the square-root operation is applied elementwise. Similarly, we assume that a vector $\mathbf{L}_2(\mathbf{x}) \in \mathbb{R}^{N_2}$ can be found such that, for a fixed value of $\mathbf{x}$, we have
\begin{equation}\label{eq:Lipd2}
\begin{split}
\inner{ \nabla_\mathbf{y} H(\mathbf{x},{\mathbf{y}_1}) - \nabla_\mathbf{y} H(\mathbf{x},{\mathbf{y}_2}), \mathbf{y}_1-\mathbf{y}_2 } & \\
& \hspace{-6em} \leq \norm{\sqrt{\mathbf{L}_2(\mathbf{x})} \odot (\mathbf{y}_1 - \mathbf{y}_2) }_2^2
\end{split}
\end{equation}
for $\forall \mathbf{y}_1,\mathbf{y}_2 \in \mathbb{R}^{N_2}$.
It should be noted that if the global Lipschitz continuity conditions of Eqs.~\eqref{eq:PALM_Lipx} and \eqref{eq:PALM_Lipy} are known to be satisfied, then a vector $\mathbf{L}_1(\mathbf{y})$ satisfying Eq.~\eqref{eq:Lipd1} can be trivially obtained by setting all of its entries equal to $L_1(\mathbf{y})$, with an analogous argument holding true for $\mathbf{L}_2(\mathbf{x})$. In particular, the Cauchy-Schwarz inequality combined with Eq.~\eqref{eq:PALM_Lipx} implies that
\begin{equation}
\begin{split}
\inner{ \nabla_\mathbf{x} H(\mathbf{x}_1,{\mathbf{y}}) - \nabla_\mathbf{x} H(\mathbf{x}_2,{\mathbf{y}}), \mathbf{x}_1-\mathbf{x}_2 } &\\
& \hspace{-14em} \leq \norm{\nabla_\mathbf{x} H(\mathbf{x}_1,{\mathbf{y}}) - \nabla_\mathbf{x} H(\mathbf{x}_2,{\mathbf{y}})}_2 \norm{\mathbf{x}_1-\mathbf{x}_2}_2\\
& \hspace{-14em} \leq L_1(\mathbf{y})\norm{\mathbf{x}_1-\mathbf{x}_2}_2^2\\
& \hspace{-14em}= \norm{\sqrt{L_1(\mathbf{y})}(\mathbf{x}_1-\mathbf{x}_2)}_2^2.
\end{split}
\end{equation}
However, Eqs.~\eqref{eq:Lipd1} and \eqref{eq:Lipd2} are more flexible than Eqs.~\eqref{eq:PALM_Lipx} and \eqref{eq:PALM_Lipy} because a different Lipschitz-like constant can be used for every coordinate, and many of these entries can be potentially much smaller than the global Lipschitz constant (because the function gradient may change much more slowly along these directions).
From an optimization perspective, Eqs.~\eqref{eq:Lipd1} and \eqref{eq:Lipd2} are important because they enable the use of potentially better majorants than were used by PALM, as described by the following theorem.
\begin{theorem} \label{thm:UT_Conv}
Consider the setting described in Section~\ref{sec:back}, and assume that the smooth real-valued function $H(\mathbf{x},\mathbf{y})$ satisfies the conditions of Eq.~\eqref{eq:Lipd1}. Then given a vector $\mathbf{c}_k \in \mathbb{R}^{N_1}$ and assuming $\mathbf{y}$ is held fixed at $\mathbf{y} = \hat{\mathbf{y}}_{k-1}$, the function
\begin{equation}
\begin{split}
H(&\hat{\mathbf{x}}_{k-1},\hat{\mathbf{y}}_{k-1}) + \inner{\mathbf{x}-\hat{\mathbf{x}}_{k-1}, \nabla_\mathbf{x} H(\hat{\mathbf{x}}_{k-1},\hat{\mathbf{y}}_{k-1})} \\ &+ \frac{1}{2} \norm{\sqrt{\mathbf{c}_k}\odot(\mathbf{x} - \hat{\mathbf{x}}_{k-1})}_2^2 + F(\mathbf{x}) + G(\hat{\mathbf{y}}_{k-1})
\end{split}
\end{equation}
is a majorant of $\Psi(\mathbf{x},\hat{\mathbf{y}}_{k-1})$ at the point $\mathbf{x}=\hat{\mathbf{x}}_{k-1}$ whenever $\mathbf{c}_k \geq \mathbf{L}_1(\hat{\mathbf{y}}_{k-1})$ (elementwise). Similarly, given a vector $\mathbf{d}_k \in \mathbb{R}^{N_2}$, assuming $H(\mathbf{x},\mathbf{y})$ satisfies the conditions of Eq.~\eqref{eq:Lipd2}, and assuming $\mathbf{x}$ is held fixed at $\mathbf{x} = \hat{\mathbf{x}}_{k}$, the function
\begin{equation}
\begin{split}
H(&\hat{\mathbf{x}}_{k},\hat{\mathbf{y}}_{k-1}) + \inner{\mathbf{y}-\hat{\mathbf{y}}_{k-1}, \nabla_\mathbf{y} H(\hat{\mathbf{x}}_{k},\hat{\mathbf{y}}_{k-1})} \\ &+ \frac{1}{2} \norm{\sqrt{\mathbf{d}_k}\odot(\mathbf{y} - \hat{\mathbf{y}}_{k-1})}_2^2 + F(\hat{\mathbf{x}}_{k}) + G(\mathbf{y})
\end{split}
\end{equation}
is a majorant of $\Psi(\hat{\mathbf{x}}_{k},\mathbf{y})$ at the point $\mathbf{y} = \hat{\mathbf{y}}_{k-1}$ whenever $\mathbf{d}_k \geq \mathbf{L}_2(\hat{\mathbf{x}}_{k})$ (elementwise).
\end{theorem}
The proof of this theorem is given in Appendix~\ref{app:proof}. Although this theorem is stated for real-valued vectors (for consistency with previous descriptions), the same also holds true for complex-valued vectors $\mathbf{x}$ and $\mathbf{y}$.
Following the same approach as PALM but replacing the original PALM majorants (Eqs.~\eqref{eq:prox_phs} and \eqref{eq:prox_mag}) with the new majorants from Thm.~\ref{thm:UT_Conv} results in the new PALM algorithm with uncoupled step sizes given in Alg.~\ref{alg:PALMUT}. This algorithm uses proximal operators with vector-valued $\mathbf{c}_k$ and $\mathbf{d}_k$, which we define as
\begin{equation}
\begin{split}
\prox{F}{\mathbf{w}_k,\mathbf{c}_k} \triangleq \arg\hspace{-0.5em}\min_{\mathbf{x} \in \mathbb{R}^{N_1}} &F(\mathbf{x}) \\&+ \frac{1}{2} \norm{\sqrt{\mathbf{c}_k} \odot( \mathbf{x} - \mathbf{w}_k )}_2^2
\end{split}
\end{equation}
and
\begin{equation}
\begin{split}
\prox{G}{\mathbf{z}_k,\mathbf{d}_k} \triangleq \arg\hspace{-0.5em}\min_{\mathbf{y} \in \mathbb{R}^{N_2}} &G(\mathbf{y}) \\&+ \frac{1}{2} \norm{ \sqrt{\mathbf{d}_k} \odot (\mathbf{y} - \mathbf{z}_k) }_2^2.
\end{split}
\end{equation}
\begin{algorithm}
\caption{PALM for Eq.~\eqref{eq:AMcost} with Uncoupled Step Sizes}
\label{alg:PALMUT}
\begin{algorithmic}
\REQUIRE Set $k=1$ and initialize $\hat{\mathbf{x}}_0$ and $\hat{\mathbf{y}}_0$.
\WHILE {not converge}
\vspace{0.05in}
\STATE Choose $\mathbf{c}_k \geq \mathbf{L}_{1}(\hat{\mathbf{y}}_{k-1})$ (elementwise)\\
$\mathbf{w}_k = \hat{\mathbf{x}}_{k-1} - \text{diag}(\mathbf{c}_k)^{-1} \nabla_\mathbf{x} H(\hat{\mathbf{x}}_{k-1},\hat{\mathbf{y}}_{k-1})$\\
$\displaystyle \hat{\mathbf{x}}_{k} = \prox{F}{\mathbf{w}_k,\mathbf{c}_k}$ \\[5pt]
\STATE Choose $\mathbf{d}_k \geq \mathbf{L}_{2}(\hat{\mathbf{x}}_{k})$ (elementwise) \\
$\mathbf{z}_k = \hat{\mathbf{y}}_{k-1} - \text{diag}(\mathbf{d}_k)^{-1} \nabla_\mathbf{y} H(\hat{\mathbf{x}}_{k},\hat{\mathbf{y}}_{k-1})$ \\
$\displaystyle \hat{\mathbf{y}}_{k} = \prox{G}{\mathbf{z}_k,\mathbf{d}_k}$ \\[5pt]
\STATE $k \gets k + 1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
Just like for the original PALM algorithm, the use of valid majorants in our new algorithm guarantees that it will monotonically decrease the objective function value, and that the objective function value converge if $\Psi(\mathbf{x},\mathbf{y})$ is bounded from below. Further, this new algorithm reduces to the original PALM algorithm if the $\mathbf{c}_k$ and $\mathbf{d}_k$ are treated as scalars (i.e., the values in different entries are always the same) instead of choosing different values in each of the entries.
Now that the general approach has been described, let's consider the application to the specific magnitude and phase optimization problem of interest from Eq.~\eqref{eq:GeneralReconModel}. Without further assumptions about the structure of $\mathbf{A}$, we do not observe any special coordinatewise structure for the magnitude subproblem. As such, we can simply utilize the global Lipschitz constant in this case, setting $\mathbf{c}_k = c_k \mathbf{1}$, where $\mathbf{1}$ is a vector whose entries are all equal to 1 and $c_k$ is the value obtained for the PALM algorithm as described in the previous section.
However, for the phase subproblem with a fixed value of $\mathbf{m}$, we observe that
\begin{equation}
\begin{split}
&\inner{\nabla_{\mathbf{q}} {J}(\mathbf{m},\mathbf{q}_1) - \nabla_{\mathbf{q}} {J}(\mathbf{m},\mathbf{q}_2), \mathbf{q}_1 - \mathbf{q}_2} \leq \\
&\hspace{8em}\|\norm{\mathbf{A}} |\mathbf{m}| \odot (\mathbf{q}_1 - \mathbf{q}_2)\|_2^2,
\end{split}
\end{equation}
where the magnitude operation is applied elementwise to the vector $\mathbf{m}$. Thus, for the first illustration from the previous subsection (Huber-function regularization of $\mathbf{m}$ with Tikhonov regularization of $\mathbf{p}$), our new algorithm can adopt
\begin{equation}
\mathbf{d}_k \geq \norm{\mathbf{A}}^2 |\hat{\mathbf{m}}_k|^2 + \lambda_2 \norm{\mathbf{C}}^2
\end{equation}
instead of the previous expression from Eq.~\eqref{eq:d1}. Similarly, for the second illustration from the previous subsection ($\ell_1$-regularization of $\mathbf{m}$ with Huber-function regularization of $\mathbf{p}$), our new algorithm can adopt
\begin{equation}
\mathbf{d}_k \geq \norm{\mathbf{A}}^2 |\hat{\mathbf{m}}_k|^2 +\frac{\lambda_2}{\xi} \norm{\sum_{t=1}^T \mathbf{B}_t^H\mathbf{B}_t}
\end{equation}
instead of the previous expression from Eq.~\eqref{eq:d2}. Note also that for both of the previous illustrations, the $\prox{G}{\cdot,\cdot}$ expressions had no dependence on the actual value of $d_k$, which allows us to simply reuse the same proximal operators for the new algorithm without modification. As such, for these illustrations, the main difference between the PALM algorithm and our new algorithm is that PALM takes a uniform step size for the phase update that depends on the maximium value of $\mathbf{m}$, while the new algorithm can take larger step sizes for coordinates where the corresponding value of $\mathbf{m}$ is small.
\subsection{Nesterov's Momentum Acceleration and PALMNUT}\label{sec:palmnut}
Our proposed PALMNUT (PALM with Nesterov's momentum and Uncoupled sTep sizes) algorithm is obtained by combining Alg.~\ref{alg:PALMUT} with Nesterov's momentum technique. The basic idea of Nesterov's technique is that, instead of computing the next iterate $\hat{\mathbf{x}}_{k}$ based on values derived from $\hat{\mathbf{x}}_{k-1}$, it can be better to instead find the next iterate using values derived from a combination of the previous iterates \cite{NesterovBook2004,BeckFISTA2009}, which can be interpreted as using ``momentum'' from previous iterations. For convex optimization problems, this approach can even result in convergence rates that have optimal order \cite{NesterovBook2004,BeckFISTA2009}. Of course, the problem of interest in this work is not convex, although it has been shown empirically (oftentimes without rigorous theoretical justification) that Nesterov's momentum technique can often substantially accelerate the iterative solution of nonconvex optimization problems.
The idea of applying momentum to accelerate the convergence of PALM has been studied in Ref.~\cite{PockiPALM2016}, where the resulting algorithm was called inertial PALM (iPALM). Theoretical convergence results for iPALM were proven with restrictive parameter choices (different from Nesterov's original parameter choices), although it was also shown empirically that using Nesterov's original parameter choices generally led to much faster convergence, despite the lack of theoretical guarantees. Our empirical experience is also consistent with these previous observations, so our proposed PALMNUT algorithm similarly utilizes Nesterov's original parameter choices (following the concise form described by Ref.~\cite{SuODE2014}). The final PALMNUT algorithm, incorporating both uncoupled coordinatewise step sizes and Nesterov's momentum, is given in Alg.~\ref{alg:PALMNUT}.
\begin{algorithm}
\caption{PALMNUT}
\label{alg:PALMNUT}
\begin{algorithmic}
\REQUIRE Set $k=1$ and initialize $\hat{\mathbf{x}}_0$ and $\hat{\mathbf{y}}_0$. \\
Set $\mathbf{u}_{0}=\hat{\mathbf{x}}_{0}$ and $\mathbf{v}_{0}=\hat{\mathbf{y}}_{0}$.
\WHILE {not converge}
\vspace{0.05in}
\STATE Choose $\mathbf{c}_k \geq \mathbf{L}_1(\mathbf{v}_{k-1})$ (elementwise) \\
$\mathbf{w}_k = \hat{\mathbf{x}}_{k-1} - \text{diag}(\mathbf{c}_k)^{-1} \nabla_\mathbf{x} H({\mathbf{u}}_{k-1},\mathbf{v}_{k-1})$\\
$\displaystyle \hat{\mathbf{x}}_{k} = \prox{F}{\mathbf{w}_k,\mathbf{c}_k}$ \\[5pt]
$\mathbf{u}_{k} = \hat{\mathbf{x}}_{k} + \frac{k-1}{k+2}(\hat{\mathbf{x}}_{k} -\hat{\mathbf{x}}_{k-1})$\\[5pt]
\STATE Choose $\mathbf{d}_k \geq \mathbf{L}_{2}({\mathbf{u}}_{k})$ (elementwise) \\
$\mathbf{z}_k = \hat{\mathbf{y}}_{k-1} - \text{diag}(\mathbf{d}_k)^{-1} \nabla_\mathbf{y} H(\mathbf{u}_{k},{\mathbf{v}}_{k-1})$ \\
$\displaystyle \hat{\mathbf{y}}_{k} = \prox{G}{\mathbf{z}_k,\mathbf{d}_k}$ \\[5pt]
$\mathbf{v}_{k} = \hat{\mathbf{y}}_{k} + \frac{k-1}{k+2}(\hat{\mathbf{y}}_{k} -\hat{\mathbf{y}}_{k-1})$\\[5pt]
\STATE $k \gets k + 1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\section{Numerical Experiments}
In the following subsections, we evaluate PALMNUT in three different simulations that are representative of a diverse set of real problems in MRI: sparsity-promoting reconstruction of undersampled k-space data \cite{FesslerSepMagPhs2012,ZibettiSepMagPhs2017,LustigPhaseCycling2018}, regularization-based denoising of a complex image \cite{haldar2019,haldar2013, haldar2008}, and using phase correction to enable the combination of multiple images acquired in the presence of experimental phase instabilities \cite{haldar2019, bernstein1989,mckinnon2000,liu2004,ChenMUSE2013,eichner2015}.
In each of these cases, we compare PALMNUT against AM combined with NCG \cite{fessler2004,JustinDecovSLIM2011,FesslerSepMagPhs2012, ZibettiSepMagPhs2017,haldar2019} and AM combined with phase cycling \cite{LustigPhaseCycling2018}. PALMNUT and AM combined with NCG are directly comparable, because they can both be used to regularize the exponentiated phase $e^{i\mathbf{p}}$, and therefore can both be applied to the exact same optimization problem. As a result, for both of these algorithms, the phase was regularized with $R_2(e^{i\mathbf{p}})$ as described previously, and we used the cost function value and the computation time to judge algorithm performance. However, the phase cycling heuristic \cite{LustigPhaseCycling2018} is intended to be used for the direct regularization of $\mathbf{p}$. For this algorithm, we therefore had to instead use phase regularization of the form $R_2(\mathbf{p})$. In addition, by its nature, the phase cycling heuristic employs a different phase-cycled cost function at each iteration, which makes it difficult to plot a meaningful cost function value. To allow for comparisons with this different approach, we therefore also computed a normalized root-mean-squared error (NRMSE) metric for each algorithm. If $\mathbf{f}$ represents the ground truth vector and $\hat{\mathbf{f}}$ represents an estimate, the NRMSE of the estimate is given by
\begin{equation}
\mathrm{NRMSE} \triangleq \|\hat{\mathbf{f}}-\mathbf{f}\|_2/\|\mathbf{f}\|_2.
\end{equation}
In order to focus on consequential errors, the NRMSE values were always computed after masking out the empty (background) parts of the field-of-view. This was particularly beneficial for AM with phase cycling, which frequently showed higher error levels in the image background compared to the other two approaches.
Regularization parameters for each cost function in each scenario were empirically optimized to achieve the smallest possible final NRMSE values for the two AM algorithms. Our implementation of AM with NCG used the Polak-Ribiere version of NCG \cite{press1992}. Our implementation of AM with phase cycling was based on code provided by the authors of Ref.~\cite{LustigPhaseCycling2018} (available from \url{https://mrirecon.github.io/bart/}).
\begin{figure}[t]
\centering
\subfloat[Magnitude]{\includegraphics[height=1.12in]{./figs/CSpMRI_GroundTruthMag.png}}
\hfil
\subfloat[Phase]{\includegraphics[height=1.12in]{./figs/CSpMRI_GroundTruthPhs.png}}
\hfil
\subfloat[Sampling Mask]{\includegraphics[height=1.12in]{./figs/mask.png}}
\caption{The ground truth (a) magnitude and (b) phase images used for the undersampled MRI reconstruction scenario, along with (c) the $8\times$-accelerated k-space sampling mask used to simulate the undersampled acquisition.}\label{fig:goldstandard}
\end{figure}
\subsection{Undersampled MRI Reconstruction}\label{sec:under}
In the first set of evaluations, we considered the reconstruction of an MR image from $8\times$-undersampled k-space data. The gold-standard magnitude and phase images, which were obtained from a real fully-sampled in vivo T1-weighted MRI acquisition with $256 \times 256$ in-plane matrix size, are shown in Fig.~\ref{fig:goldstandard}. This figure also shows the k-space sampling mask (corresponding to $8\times$ undersampling) that we used to simulate an accelerated acquisition.
\begin{figure}
\centering
\subfloat[]{\includegraphics[width=1.7in]{./figs/CSpMRI_cost.eps}}
\hfil
\subfloat[]{\includegraphics[width=1.7in]{./figs/CSpMRI_NRMSE_complx.eps}}\\
\subfloat[Zero-filled]{\includegraphics[width=.8in]{./figs/ZeroFillMag.png}\includegraphics[width=.8in]{./figs/ZeroFillPhs.png}}
\hfil
\subfloat[PALMNUT]{\includegraphics[width=.8in]{./figs/CSpMRI_NAPALMMag.png}\includegraphics[width=.8in]{./figs/CSpMRI_NAPALMPhs.png}}
\caption{Convergence plots and reconstructed images for the undersampled MRI reconstruction scenario. The convergence plots show (a) the cost function value as a function of computation time and (b) the NRMSE value as a function of computation time. Also shown are the magnitude and phase images corresponding to (c) zero-filled reconstruction and (d) PALMNUT.}
\label{fig:CSpMRIConvCurv}
\end{figure}
For this simulation, the original fully-sampled k-space data (originally measured with 32 channels) was coil-compressed down to 8 virtual channels to reduce computational complexity, and was then retrospectively undersampled using the aforementioned k-space sampling mask. For reconstruction, the $\mathbf{A}$ matrix was chosen according to the standard SENSE model \cite{pruessmann2001}, with sensitivity maps estimated using ESPIRiT \cite{LustigESPIRiT}. The magnitude regularization took the form of an $\ell_1$ penalty as given by Eq.~\eqref{eq:L1_Mag_Huber_Phs}, where, following Ref.~\cite{LustigPhaseCycling2018}, the sparsifying transform $\mathbf{T}$ was chosen to be the unitary Daubechies-4 wavelet transform. The phase regularization took the form of a Huber-function penalty as given by Eq.~\eqref{eq:L1_Mag_Huber_Phs2}, where the Huber parameter $\xi$ was chosen to be a small number (i.e., $\xi = 0.001$) in order to approximate the $\ell_1$-norm. Following Ref.~\cite{LustigPhaseCycling2018}, the transform we used for phase regularization was also a unitary Daubechies-4 wavelet transform. All three algorithms were initialized by applying SENSE-based coil-combination to the multi-channel images obtained by zero-filling the unmeasured data.
Convergence results for all three algorithms are shown in Fig.~\ref{fig:CSpMRIConvCurv}, along with representative image reconstruction results. As can be seen, the cost function and NRMSE values converge to similar levels for both PALMNUT and AM with NCG, although PALMNUT converged much faster. Although AM with phase cycling converged to a result with a reasonably-good NRMSE, it was substantially worse than PALMNUT in both convergence speed and the final achieved NRMSE value.
\begin{figure}[t]
\centering
\subfloat[Ground Truth]{\includegraphics[width=.85in]{./figs/GroundTruthMag.png}\includegraphics[width=.85in]{./figs/GroundTruthPhs.png}}
\hfil
\subfloat[Noisy]{\includegraphics[width=.85in]{./figs/NoisyMag.png}\includegraphics[width=.85in]{./figs/NoisyPhs.png}}
\caption{The (a) ground truth and (b) noisy magnitude and phase images used for the regularization-based MRI denoising scenario.}
\label{fig:denoise_gold}
\end{figure}
\subsection{Regularization-based MRI Denoising}\label{sec:denoise}
In the second set of evaluations, we considered regularization-based denoising of a 230$\times$180 single-channel T1-weighted MR image, obtained by applying complex coil-combination to an 8-channel dataset and subsequently adding simulated complex Gaussian noise. The ground truth and noisy images are shown in Fig.~\ref{fig:denoise_gold}.
For reconstruction, the $\mathbf{A}$ matrix was an identity matrix. Following Refs.~\cite{haldar2013,haldar2019,haldar2011,haldar2008}, the magnitude was regularized using a Huber-function penalty as given by Eq.~\eqref{eq:Huber_Mag_Tikhonov_Phs}, where a finite difference transformation was used to enforce spatial smoothness of the image. Following Ref.~\cite{haldar2019,fessler2004,JustinDecovSLIM2011,FesslerSepMagPhs2012,ZibettiSepMagPhs2017}, the phase was regularized using a Tikhonov penalty as given by Eq.~\eqref{eq:Huber_Mag_Tikhonov_Phs2}, also using a finite difference transformation to enforce spatial smoothness. All algorithms were initialized with the noisy image.
Convergence results are shown in Fig.~\ref{fig:DenoiseConvCurv}, along with representative reconstruction results. As can be seen, the results in this case are consistent with the previous case: PALMNUT and AM with NCG converged to similar NRMSE values, although PALMNUT was substantially faster, while AM with phase cycling was reasonably successful yet still substantially worse than the others in both speed and NRMSE.
\begin{figure}[t]
\centering
\subfloat[]{\includegraphics[width=1.7in]{./figs/Denoising_cost.eps}}
\hfil
\subfloat[]{\includegraphics[width=1.7in]{./figs/Denoising_NRMSE_complx.eps}} \\
\subfloat[PALMNUT]{\includegraphics[width=0.85in]{./figs/Denoise_NAPALMMag.png}\includegraphics[width=0.85in]{./figs/Denoise_NAPALMPhs.png}}
\caption{Convergence plots and reconstructed images for the regularization-based MRI denoising scenario. The convergence plots show (a) the cost function value as a function of computation time and (b) the NRMSE value as a function of computation time. Also shown are the (c) magnitude and phase images corresponding to PALMNUT.}
\label{fig:DenoiseConvCurv}
\end{figure}
\subsection{Phase-Corrected MR Image Combination}
In the third experiment, we simulated a scenario that is common in diffusion MRI, in which multiple measurements are made of the same image to enable averaging to improve SNR, but the phase varies randomly with each measurement due to experimental instabilities. This case was simulated based on actual diffusion MRI magnitude and phase data from Ref.~\cite{haldar2019}. We simulated a case with four repetitions, based on one ground truth magnitude image and four ground truth phase images, all with matrix size $180 \times 332$. The magnitude image was combined with each of the different phase images to yield four different complex images, and then complex Gaussian noise was added to each result. The ground truth images and a representative noisy image are shown in Fig.~\ref{fig:combo_truth}.
\begin{figure}[t]
\centering
\subfloat[Magnitude]{\includegraphics[width=1.6in]{./figs/ShareMag_GroundTruthMag.png}}
\hfil
\subfloat[Noisy]{\includegraphics[width=1.6in]{./figs/ShareMag_NoisyMag.png}}\\
\subfloat[Phases]{\includegraphics[width=0.82in]{./figs/ShareMag_GroundTruthPhs1.png}\hfil\includegraphics[width=0.82in]{./figs/ShareMag_GroundTruthPhs2.png}\hfil\includegraphics[width=0.82in]{./figs/ShareMag_GroundTruthPhs3.png}\hfil\includegraphics[width=0.82in]{./figs/ShareMag_GroundTruthPhs4.png}}
\caption{(a) The ground truth magnitude and (b) a representative noisy magnitude image, along with (c) the four different ground truth phase images used for the scenario with phase-corrected combination of multiple images. }\label{fig:combo_truth}
\end{figure}
For reconstruction, we estimated a single shared magnitude image $\mathbf{m}$ and four different phase images $\mathbf{p}_j$ corresponding to the four different noisy measured images $\mathbf{b}_j$, $j=1,2,3,4$, by minimizing the cost function
\begin{equation}
R_1(\mathbf{m}) + \sum_{j=1}^4 \frac{1}{2} \|\mathbf{m}\odot e^{i \mathbf{p}_j} - \mathbf{b}_j \|_2^2 + R_2(e^{i \mathbf{p}_j}),
\end{equation}
where the magnitude and phase regularization penalties $R_1(\cdot)$ and $R_2(\cdot)$ were chosen in exactly the same way as for the denoising scenario from the previous subsection. All optimization algorithms were initialized with the noisy phase images and by taking the average of the noisy magnitude images.
The convergence plots shown in Fig.~\ref{fig:ShareMagConvCurv} show similar characteristics to those observed in the previous scenarios, with PALMNUT having a distinct advantage over the two alternative algorithms.
\begin{figure}
\centering
\subfloat[]{\includegraphics[width=1.7in]{./figs/ShareMag_cost.eps}}
\hfil
\subfloat[]{\includegraphics[width=1.7in]{./figs/ShareMag_NRMSE_complx.eps}}\\
\subfloat[PALMNUT]{\includegraphics[width=1.7in]{./figs/ShareMag_NAPALMMag.png}}
\caption{Convergence plots and reconstructed images for the phase-corrected MRI image combination scenario. The convergence plots show (a) the cost function value as a function of computation time and (b) the NRMSE value as a function of computation time. Also shown is (c) the combined magnitude image obtained from PALMNUT.}
\label{fig:ShareMagConvCurv}
\end{figure}
\section{Discussion}
The results of the previous section demonstrated that PALMNUT can have major advantages relative to AM methods. However, PALMNUT represents a combination of three different ideas (PALM, Nesterov's momentum, and uncoupled stepsizes), and the previous experiments did not investigate which parts of PALMNUT contribute most to its performance. In order to gain more insight, we did another set of experiments in which we compared PALMNUT against the original PALM algorithm (Alg.~\ref{alg:PALM}), PALM with Nesterov's momentum but without uncoupled stepsizes (iPALM), and PALM with uncoupled stepsizes but without Nesterov's momentum (Alg.~\ref{alg:PALMUT}). Results are shown in Fig.~\ref{fig:ablation} for both the undersampled MRI reconstruction (Section~\ref{sec:under}) and the regularization-based MRI denoising (Section~\ref{sec:denoise}) scenarios described previously. From these plots, we observe that both uncoupled step sizes and Nesterov's momentum independently improve the convergence speed relative to the original PALM method, with Nesterov's momentum contributing a little more than the use of uncoupled step sizes. However, PALMNUT's combination of all of these elements leads to the best overall results.
\begin{figure}[t]
\centering
\subfloat[Undersampled MRI Reconstruction]{\includegraphics[width=1.7in]{./figs/CSpMRI_PALM_self_cost.eps}\includegraphics[width=1.7in]{./figs/CSpMRI_PALM_self_nrmse.eps}} \\
\subfloat[Regularization-based MRI Denoising]{\includegraphics[width=1.7in]{./figs/Denoise_PALM_self_cost.eps}\includegraphics[width=1.7in]{./figs/Denoise_PALM_self_nrmse.eps}}
\caption{The convergence plots show the cost function value and the NRMSE value as a function of computation time for (a) the undersampled MRI reconstruction scenario and (b) the regularization-based MRI denoising scenario.}
\label{fig:ablation}
\end{figure}
An interesting phenomenon we observed with PALMNUT is that the magnitude estimate frequently converged faster than the phase estimate (results not shown). As a result, it could be potentially beneficial to update the phase estimate more frequently than the magnitude estimate, although we believe that such an exploration is beyond the scope of this paper.
One of the key ingredients of PALMNUT is the use of coordinatewise step sizes based on coordinatewise Lipschitz-like bounds in the form of Eqs.~\eqref{eq:Lipd1} and \eqref{eq:Lipd2}. Although these bounds were formulated in the context of PALM, we believe that Lemma 1 (found in the Appendix and used in the proof of Theorem 1) represents a novel coordinatewise majorization relationship, which could also potentially be useful to construct majorants in more general optimization scenarios.
Finally, we should mention that while PALM combined with uncoupled step sizes has guaranteed monotonic convergence properties, the convergence of the PALMNUT approach (which additionally includes Nesterov's momentum) has not yet been theoretically proven. This represents another potentially interesting topic for future research.
\section{Conclusion}
We proposed and evaluated a new algorithm called PALMNUT, which combines the PALM algorithm with Nesterov's momentum with uncoupled coordinatewise step sizes derived from coordinatewise Lipschitz-like bounds. Although our approach is general and can be applied to other computational imaging scenarios, our evaluation studies focused on MRI scenarios involving separate regularization of the image magnitude and phase. Our empirical results demonstrated that PALMNUT consistently had substantial advantages over previous approaches based on alternating minimization across several different MRI scenarios. As a result, we expect that PALMNUT will be useful for these kinds of MRI scenarios, and may also prove useful for more general computational imaging problems with similar optimization structure.
\onecolumn
\appendices
\section{Proof of Thm.~\ref{thm:UT_Conv}}\label{app:proof}
Theorem~\ref{thm:UT_Conv} is an immediate consequence of the following Lemma.
\begin{lemma} \label{Lema:DesctIneq}
Let $Q(\mathbf{x}): \mathbb{R}^N \rightarrow \mathbb{R}$ be smooth, and assume that a non-negative vector $\mathbf{L} \in \mathbb{R}^{N}$ exists such that
\begin{equation}
\begin{split}
\inner{\nabla Q(\mathbf{x}_1) - \nabla Q(\mathbf{x}_2), \mathbf{x}_1 - \mathbf{x}_2} \leq \norm{\sqrt{\mathbf{L}} \odot (\mathbf{x}_1 - \mathbf{x}_2)}_2^2
\end{split}
\end{equation}
for $\forall \mathbf{x}_1,\mathbf{x}_2 \in \mathbb{R}^N$. Then for $\forall \mathbf{x},\hat{\mathbf{x}}_k \in \mathbb{R}^N$ and $\forall \mathbf{c} \geq \mathbf{L}$ (elementwise), we have
\begin{equation} \label{eq:CompDesct}
\begin{split}
Q(\mathbf{x}) &\leq Q(\hat{\mathbf{x}}_k) + \inner{\mathbf{x}-\hat{\mathbf{x}}_k,\nabla Q(\hat{\mathbf{x}}_k)} + \frac{1}{2} \norm{\sqrt{\mathbf{c}}\odot\left( \mathbf{x}-\hat{\mathbf{x}}_k\right)}_{2}^{2}.
\end{split}
\end{equation}
Further, the right hand side of Eq.~\eqref{eq:CompDesct} can be written compactly as
\begin{equation}\label{eq:CompDesct2}
\begin{split}\\
Q(\hat{\mathbf{x}}_k) + \inner{\mathbf{x}-\hat{\mathbf{x}}_k,\nabla Q(\hat{\mathbf{x}}_k)} + \frac{1}{2} \norm{\sqrt{\mathbf{c}}\odot\left( \mathbf{x}-\hat{\mathbf{x}}_k\right)}_{2}^{2}
&= \tau_k + \frac{1}{2}\norm{\sqrt{\mathbf{c}} \odot (\mathbf{x} - \mathbf{w}_k)}_2^2,
\end{split}
\end{equation}
where
\begin{equation}
\mathbf{w}_k = \hat{\mathbf{x}}_k - \mathrm{diag}(\mathbf{c})^{-1} \nabla Q(\hat{\mathbf{x}}_k)
\end{equation}
and $\tau_k$ is a constant that does not depend on the variable $\mathbf{x}$, given by
\begin{equation}
\tau_k = Q(\hat{\mathbf{x}}_k) - \inner{\hat{\mathbf{x}}_k,\nabla Q(\hat{\mathbf{x}}_k)}+\frac{1}{2} \norm{\sqrt{\mathbf{c}}\odot \hat{\mathbf{x}}_k}_2^2 - \frac{1}{2}\norm{\sqrt{\mathbf{c}} \odot \mathbf{w}_k}_2^2.
\end{equation}
\end{lemma}
\begin{proof}
Inspired by the proof of Proposition A.24 from \cite{BertsekasNonlinearProgram}, define $\alpha(t) \triangleq Q(t\mathbf{x} + (1-t)\hat{\mathbf{x}}_k)$. Then $\alpha(0) = Q(\hat{\mathbf{x}}_k)$, $\alpha(1) = Q(\mathbf{x})$, and
\begin{equation}
\frac{d}{dt}\alpha(t) = \inner{\nabla Q(t\mathbf{x} + (1-t) \hat{\mathbf{x}}_k), \mathbf{x} - \hat{\mathbf{x}}_k}.
\end{equation}
We have that
\begin{equation}
\begin{split}
Q(\mathbf{x}) - Q(\hat{\mathbf{x}}_k) &= \int_{0}^1 \frac{d}{dt}\alpha(t) dt \\
& = \int_0^1 \inner{\nabla Q(t\mathbf{x} + (1-t) \hat{\mathbf{x}}_k), \mathbf{x} - \hat{\mathbf{x}}_k} dt \\
& = \int_0^1 \inner{\nabla Q(t\mathbf{x} + (1-t) \hat{\mathbf{x}}_k) - \nabla Q(\hat{\mathbf{x}}_k), \mathbf{x} - \hat{\mathbf{x}}_k} dt + \inner{\nabla Q(\hat{\mathbf{x}}_k),\mathbf{x}-\hat{\mathbf{x}}_k} \\
& = \int_0^1 \frac{1}{t} \inner{\nabla Q(t\mathbf{x} + (1-t) \hat{\mathbf{x}}_k) - \nabla Q(\hat{\mathbf{x}}_k), t \mathbf{x} - t\hat{\mathbf{x}}_k} dt + \inner{\mathbf{x}-\hat{\mathbf{x}}_k,\nabla Q(\hat{\mathbf{x}}_k)}\\
& = \int_0^1 \frac{1}{t} \inner{\nabla Q(t\mathbf{x} + (1-t) \hat{\mathbf{x}}_k) - \nabla Q(\hat{\mathbf{x}}_k), t \mathbf{x} + (1- t)\hat{\mathbf{x}}_k - \hat{\mathbf{x}}_k} dt +\inner{\mathbf{x}-\hat{\mathbf{x}}_k,\nabla Q(\hat{\mathbf{x}}_k)} \\
& \leq \int_{0}^1 \frac{1}{t} \norm{\sqrt{\mathbf{L}} \odot (t \mathbf{x} + (1- t)\hat{\mathbf{x}}_k - \hat{\mathbf{x}}_k)}_2^2 dt+ \inner{\mathbf{x}-\hat{\mathbf{x}}_k,\nabla Q(\hat{\mathbf{x}}_k)} \\
& = \int_{0}^1 t \norm{\sqrt{\mathbf{L}} \odot (\mathbf{x} - \hat{\mathbf{x}}_k)}_2^2 dt+ \inner{\mathbf{x}-\hat{\mathbf{x}}_k,\nabla Q(\hat{\mathbf{x}}_k)} \\
& = \frac{1}{2}\norm{\sqrt{\mathbf{L}} \odot (\mathbf{x} - \hat{\mathbf{x}}_k)}_2^2 + \inner{\mathbf{x}-\hat{\mathbf{x}}_k,\nabla Q(\hat{\mathbf{x}}_k)} \\
& \leq \frac{1}{2}\norm{\sqrt{\mathbf{c}} \odot (\mathbf{x} - \hat{\mathbf{x}}_k)}_2^2 + \inner{\mathbf{x}-\hat{\mathbf{x}}_k,\nabla Q(\hat{\mathbf{x}}_k)}.
\end{split}
\end{equation}
This derivation proves Eq.~\eqref{eq:CompDesct}, while the simplifications leading to Eq.~\eqref{eq:CompDesct2} come simply from completing the square.
\end{proof}
\twocolumn
| 95929f42bed8249b8f44a1634036ebb8c1f487e6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{intro}
Dense and hot plasma structures are a common sight in the solar corona. These structures are mostly linear (in the sense that their lengths are much longer than their transverse scales) and are believed to trace local magnetic field lines. Some of them are closed, meaning both ends of the structure anchored on the Sun (footpoints) are clearly visible while others that are extended to very large distances concealing their connection from one foot point to other, are considered open. Ray-like structures visible in the polar regions of the Sun, called polar plumes, are open structures. The structures that form part of active regions are typically closed and are called coronal loops. It is also common to have ``open loops'' or ``extended loops'', referring to the loops that are part of active regions but their connectivity to the other foot point is not clear. Extended loops originating in a quiet-Sun network region (or a plage region) are called network plumes or on-disk plume-like structures. Because of the constant churning experienced by all these structures due to convective motions on the surface, it is natural to expect creation of different magnetohydrodynamic (MHD) wave modes which are believed to play a significant role in maintaining the million degree temperatures in the corona. We refer the reader to a review by \citet{2020SSRv..216..140V} in this issue, to find our current understanding on the contribution of MHD waves to coronal heating.
The launch of high-resolution telescopes such as SoHO and TRACE in the late 90s has led to the discovery of different MHD wave modes in the solar corona. Both longitudinal slow magnetoacoustic waves and transverse kink/Alfv\'enic waves were found to exist. Modern telescopes such as Hinode, SDO, and IRIS alongside several ground-based telescopes (for e.g., SST, DST, GREGOR, GST, DKIST) offer us even better resolution in addition to providing increased spatial, temporal, and atmospheric coverage. These extraordinary capabilities have not only expanded the detection of different MHD waves but also improved our knowledge on their characteristic properties. Indeed, as many researchers have shown, current observations are advanced enough to compare the observed wave properties with theoretical models and extract several key parameters through coronal seismology - a newly emergent field in solar physics, thanks to these advances \citep[e.g.,][]{1999Sci...285..862N, 2001A&A...372L..53N, 2012RSPTA.370.3193D, 2020ARA&A..58..441N}.
Slow magnetoacoustic waves (or simply slow waves) have been detected in different coronal structures including polar plumes \citep[e.g.,][]{1997ApJ...491L.111O, 1998ApJ...501L.217D}, active region fan loops \citep[e.g.,][]{2002A&A...387L..13D, 2006ApJ...643..540M} and other network or plage related plume-like structures \citep[e.g.,][]{1999SoPh..186..207B, 2000A&A...355L..23D}. Standing slow waves were also found in hot flare loops \citep[e.g,][]{2011SSRv..158..397W}. A dedicated review on standing waves has been presented by \citet{2021SSRv..217...34W} in this issue. Recent studies have shown the origin of propagating slow waves in the lower solar atmosphere \citep[e.g.,][]{2009A&A...505..791S, 2011ApJ...728...84B, 2012ApJ...757..160J, 2015ApJ...812L..15K, 2016ApJ...830L..17Z}, revealed their damping to be frequency-dependent albeit with some inconsistencies with theory \citep[e.g.,][]{2014ApJ...789..118K}, and also highlighted that thermal conduction might not be the dominant cause for their damping as previously thought \citep[e.g.,][]{2015ApJ...811L..13W}. The high-resolution observations also uncovered a few problems, one of which is the possible ambiguity in distinguishing the slow waves from high-speed quasi-periodic flows. Especially, the spectroscopic observations revealed a more complex picture, with perturbations not only in the line intensities but also in other parameters such as the Doppler velocities, line widths and red-blue asymmetry leading to the debate on interpretation in terms of waves \citep[e.g.,][]{2009A&A...503L..25W, 2009A&A...499L..29B, 2010ApJ...721..744K, 2010ApJ...713..573M, 2011A&A...528L...4K, 2011ApJ...734...81M, 2014ApJ...793..117S, 2015RAA....15.1027D, 2016RAA....16...93J, 2016ApJS..224...30Y, 2018ApJ...853..134M, 2021ApJ...909..202C} or flows \citep[e.g.,][]{2007Sci...318.1585S, 2007ApJ...667L.109D, 2008ApJ...678L..67H, 2010ApJ...722.1013D, 2010A&A...516A..14H, 2010RAA....10.1307G, 2010ApJ...715.1012B, 2011ApJ...730...37U, 2011ApJ...736..130T, 2012ApJ...759..144T, 2012ApJ...760...82S, 2016ApJ...829L..18B, 2021SoPh..296...47T}. Despite this, a number of seismological applications of slow waves have been proposed including the determination of plasma temperature, loop geometry, magnetic field, thermal conduction and polytropic index \citep{2009ApJ...697.1674M, 2009A&A...503L..25W, 2011ApJ...727L..32V, 2014A&A...561A..19Y, 2015ApJ...811L..13W, 2016NatPh..12..179J, 2017ApJ...834..103K, 2018SoPh..293....2D, 2018ApJ...868..149K, 2018ApJ...860..107W, 2019ApJ...886....2W}. More recent models discuss the inclusion of unknown coronal heating function in the MHD equations in a parameterised form as a function of the equilibrium parameters such as density and temperature in order to constrain it using the observations of slow waves \citep{2019A&A...628A.133K, 2019PhPl...26h2113Z, 2020A&A...644A..33K, 2021SoPh..296...20P}.
Different MHD wave modes are ubiquitously observed in the solar atmosphere by space and ground-based instruments \citep{2007SoPh..246....3B}.
It is worth noting that early studies since the era of {\it Skylab} have interpreted the observed nonthermal broadening of spectral lines to be the signature of Alfv\'en waves in polar coronal holes \citep[][and references therein]{1990ApJ...348L..77H,1998A&A...339..208B,2009A&A...501L..15B}. However, the observational resolution in that era was not good enough to disentangle different wave modes (torsional Alfv\'en wave or homogenous Alfv\'en wave) and the structure of the solar plasma. After the launch of high resolution instruments, there is an ample evidence that the solar plasma is inhomogeneous. Hence, the term Alfvénic (also called kink) better acknowledges the rich and complex nature of the coupled wave system within a realistic, continuous, inhomogeneous magnetic medium.
Direct signatures of kink waves were first noted in coronal loops \citep{1999Sci...285..862N,1999ApJ...520..880A}. Both decaying and decayless oscillations have been observed in different solar structures \citep{2013A&A...552A..57N,2015A&A...583A.136A}. It has been reported that most of the kink oscillations are excited when coronal loops are displaced from their mean position by the lower coronal eruptions \citep{2015A&A...577A...4Z} including coronal jets \citep{2016SoPh..291.3269S}. After the launch of CoMP and SDO, propagating kink waves were detected in the polar regions of the Sun \citep[][and references therein]{2007Sci...317.1192T,2009ApJ...697.1384T,2014ApJ...790L...2T}.
It has been believed that the sources of these propagating kink waves in polar coronal holes are the supergranular motions. Recently, \citep{2016ApJ...828...89M, 2019NatAs...3..223M} has shown that p-modes could be responsible for exciting observed kink oscillations in the solar atmosphere. While \citet{2016A&A...591L...5N} showed that the observed dependence of the displacement and velocity amplitudes on the oscillation periods does not have any peak at a certain frequency. This casts doubts over the interpretations that p-modes are responsible for observed kink waves in the solar corona. Also, kink waves (standing and propagating) can be used to infer cross density structures and magnetic field using seismology inversion techniques \citep[][and references therein]{2011ApJ...736...10S,2019ApJ...876..106T,2020ApJ...894...79W,2020Sci...369..694Y}. Though, a unique solution of the inversion does not exist \citep[see][and references therein]{2019A&A...622A..44A} except when the density profile of the inhomogeneous layer is kept fixed \citep{2016A&A...589A.136P}.
Transverse waves are believed to carry energy flux enough to heat the quiet Sun\citep{2011Natur.475..477M}. While \citet{2014ApJ...790L...2T} claimed that transverse waves carry insufficient energy to accelerate solar winds in polar coronal holes, \citet{2014Sci...346A.315T} reported that transverse waves observed in the transition regions have enough energies to accelerate the solar wind. Thus, more studies using upcoming space and ground-based instruments are needed to understand the role of transverse waves in accelerating solar winds in the polar coronal holes.
In this review article, we give an overview of some of these recent advancements in the observations of both slow waves and Alf\'{e}nic waves with an emphasis on open coronal structures. We also present relevant theoretical/numerical modelling studies wherever appropriate. We discuss some of the latest advances in the study of slow waves in Section \ref{slow_waves}, and that in Alfv\'{e}nic or kink waves in Section \ref{alfven_waves}, followed by a description of the role of Alfv\'{e}nic waves in the solar wind acceleration in Section \ref{solar_wind} and finally provide a brief summary of all the results along with some future directions in Section \ref{summary}.
\section{Slow magnetoacoustic waves}
\label{slow_waves}
Observations of propagating slow magnetoacoustic waves in the solar atmosphere have generated increased attention over the past decades. The continuous multi-wavelength capture of the full solar disk at high spatial and temporal resolutions by SDO/AIA, has enabled us to perform detailed studies on the slow wave propagation and dissipation characteristics. The perturbations due to slow waves appear as slanted ridges of alternating brightness (or Doppler velocity) in time-distance maps that are commonly constructed to study the evolution of a specific coronal structure. These perturbations are generally referred to as `propagating disturbances (PDs)' or `propagating coronal disturbances (PCDs)'. There has been a debate in the recent years proclaiming the observed PDs could also be generated by high-speed quasi-periodic upflows \citep[e.g.,][and references cited therein]{2010ApJ...722.1013D, 2010A&A...510L...2M, 2011ApJ...727L..37T, 2011ApJ...738...18T, 2012ApJ...749...60M} while their interpretation in terms of slow waves \citep[e.g.,][and references cited therein]{2009A&A...503L..25W, 2010ApJ...724L.194V, 2011A&A...528L...4K, 2012ASPC..456...91W, 2012A&A...546A..93G} is not unique. Importantly, some spectroscopic observations of PDs from \textit{Hinode}/EIS have revealed periodic blueward asymmetry in the line profiles suggesting the presence of periodic upflows \citep{2010ApJ...722.1013D}. However, \citet{2010ApJ...724L.194V} have later demonstrated that such asymmetry in the line profiles is an intrinsic feature of upward propagating slow waves. Although the issue has not been completely resolved yet \citep[e.g.,][and references cited therein]{2012RSPTA.370.3193D, 2015SoPh..290..399D, 2017ApJ...845L..18D}, it has been largely agreed that both periodic waves and aperiodic flows coexist near the foot points but the wave signatures dominate at larger distances along the structure \citep[e.g.,][and references cited therein]{2011ApJ...737L..43N, 2012ApJ...759..144T,2012ApJ...754..111O,2013ApJ...775L..23W, 2015ApJ...807...71P, 2015ApJ...815L..16S}. The discussion in this article presumes this behaviour and considers PDs synonymous with slow waves unless mentioned otherwise explicitly. Interested readers may refer to \cite{2012RSPTA.370.3193D}, \cite{2016GMS...216..419B}, and \cite{2016GMS...216..395W}, for further details on this debate. One must also note that the extended loop structures associated with active regions are considered as open structures in this article in the sense of the wave propagation but not magnetic connectivities that are unclear in observations. This is a valid assumption since the extent of propagation of slow waves is in general much shorter compared to the length (visible extent) of the associated structures. In the following, we discuss the progress made in the last few years in understanding the origin of the slow waves, their damping behaviour in the solar corona, and finally illustrate some of their seismological applications.
\subsection{Source of oscillations}
Slow waves are regularly observed in a variety of coronal structures. Both standing and propagating versions of these waves have been found although the standing waves are thought to be locally generated \citep{2011SSRv..158..397W, 2013ApJ...779L...7K}, whereas the propagating waves originate in the lower solar atmosphere \citep{2009A&A...505..791S, 2011ApJ...728...84B, 2012ApJ...757..160J, 2015ApJ...812L..15K}. Moreover, the propagating waves have been shown to be ubiquitous in the solar corona \citep{2012A&A...546A..50K, 2018ApJ...853..145M}. A wide range of periodicities are observed, but in general, longer periods ($\sim$ tens of minutes) are dominant in the polar regions \citep{2009A&A...499L..29B, 2011A&A...528L...4K} while the active regions are replete with shorter periods ($\sim$ minutes) \citep{2012SoPh..281...67K, 2018ApJ...856L..16W}. However, the longer periods ($>$ 10 min) are difficult to penetrate into the solar corona unless the magnetic field is highly inclined \citep{2015ApJ...809L..17J}. Therefore, even though the general characteristic properties of these waves appear broadly similar across all regions, it is quite possible that they have a different origin in active regions and polar regions. Here, we discuss some recent results on possible source locations across these two regions. It may be noted that longer periods ($\sim$ 25 min) are also found in active regions \citep{2009A&A...503L..25W, 2011A&A...533A.116Y} and similarly, shorter periods ($\sim$ 5 min) are not totally non-existent in the polar regions \citep{2014ApJ...789..118K} but the distinction here is based on their dominant presence.
\subsubsection{Active regions}
Early observations using high-resolution imaging data from \textit{Transition Region and Coronal Explorer} \citep[TRACE;][]{1999SoPh..187..229H} have revealed that the slow wave perturbations in loops that are rooted in a sunspot umbra display 3-minute periods whereas those in loops anchored in plage-related regions display 5-minute periods \citep{2002A&A...387L..13D}. These periods are coincident with that of the oscillations found in the lower atmosphere at the corresponding locations. It was therefore suggested that the driver of the coronal slow waves lies in the lower atmosphere near the footpoints of the supporting structure. \cite{2006ApJ...643..540M} used spectroscopic data corresponding to the transition region of a sunspot along with the co-temporal coronal images of the region, to show the propagation of slow waves from the transition region to corona. The authors further conjectured that these oscillations are driven by the global oscillations of the Sun, i.e., the photospheric $p$-modes. It is possible that the $p$-modes act as a broadband driver and the local magnetic configuration aids the leakage of the natural frequencies in the vicinity of the acoustic cutoff in the form of upward propagating slow waves \citep{2011ApJ...728...84B}.
The amplitude of propagating slow waves is not constant but appears to be modulated over time. Assuming the local medium is relatively stable, the temporal changes in a propagating wave implies changes in the driver properties. Utilising this idea, \cite{2015ApJ...812L..15K} traced the coronal slow waves across different layers of the solar atmosphere to identify their driver. High-resolution multi-wavelength observations of a sunspot, acquired from both ground- and space-based telescopes in 9 different channels, are used to perform this study.
\begin{figure}
\includegraphics[width=0.25\textwidth, trim={1.6cm 0 0 0}, clip]{kpfig1a-eps-converted-to.pdf}
\includegraphics[width=0.75\textwidth]{kpfig1b.jpg}
\caption{Structure of the leading spot of NOAA 11366 across 9 different wavelength channels captured on 10 December 2011 by a suite of telescopes based on ground and space through coordinated observations. The vertical dashed line connects the footpoint location of a fan loop (marked as `B') rooted in the umbra across different channels. The central panels display the corresponding Fourier power spectra of light curves from location `B'. The peak periodicities are listed in the individual panels. The vertical dashed lines denote the locations of 3 and 5 minute periods. The filtered light curves allowing frequencies only within a narrow band (indicated in grey in the central panels) around the peak frequency are shown in the right panels. The names of the individual channels are also listed. Adapted from \cite{2015ApJ...812L..15K}}
\label{kpfig1}
\end{figure}
Fig.~\ref{kpfig1} (left panel) displays the structure of sunspot across 9 wavelength channels. The fan loop structure is evident in the top two panels characterising coronal emission. The location `B' corresponds to the foot point of a selected loop which displays propagating slow waves and the vertical dashed line connects this location across all the channels. The central panels display the corresponding Fourier power spectra of the light curves extracted from this location. The peak periods are listed in each panel. The vertical dashed lines in red denote the locations of the 3 and 5 minute periods. The shift in periodicity from 5 min in the photosphere (bottom two panels) to 3 min in the chromosphere/transition region is long-known with multiple explanations already provided in the literature \citep{1983SoPh...82..369Z, 1991A&A...250..235F}. The grey bands over the spectra delineate the widths of the band-pass filter applied on the original data to obtain the light curves displayed in the right panels. The identification labels for the individual channels are listed at the bottom of the light curves. As may be seen from the light curves, the amplitude of the oscillations is periodically varying with approximately similar behaviour across all the channels. The modulation period is found to be in the range of 20$-$27 min. These correlated amplitude variations across different layers imply that they are clearly non local and therefore should be associated with the driver. \cite{2015ApJ...812L..15K} have further demonstrated that similar modulation periods are found in the global acoustic oscillations at non-magnetic locations outside the sunspot and are fully consistent with those seen in coronal oscillations. The modulation itself is caused by the simultaneous presence of waves with closely spaced frequencies (as may be seen in the Fourier power spectra in Fig.~\ref{kpfig1}) as noted previously by several authors \citep[e.g.,][]{2006ApJ...643..540M} and is a characteristic feature of global oscillations. These results compelled the authors to conclude that the coronal slow waves are externally driven by the global solar oscillations ($p$-modes) in the photosphere. A recent study by \cite{2020A&A...638A...6S} also shows similar amplitude modulation in slow waves observed in a sunspot across different layers of the solar atmosphere possibly confirming the wider applicability of this interpretation. Utilising high-resolution multi-wavelength observations of a sunspot from 16 different channels, \citet{2016ApJ...830L..17Z} studied the propagation of slow waves from photosphere to corona. These authors employed a time-distance helioseismic technique to cross-correlate and track the waves across different layers of the solar atmosphere. They analysed the propagation patterns of different frequencies and compared them with a numerical model to demonstrate that the observed slow waves in the corona originate from a source few megameters beneath the photosphere. It was further suggested that the interaction of $p$-modes with the sub-surface magnetic fields of sunspots can possibly explain the underlying source.
By combining high-spatial and temporal resolution data from the ground-based Dunn Solar Telescope (DST) with the Extreme Ultraviolet (EUV) imaging data from the SDO/AIA, \cite{2012ApJ...757..160J} found $\approx$3 minute oscillations across different layers of the solar atmosphere. It has been further demonstrated that the coronal fan structures, where propagating slow waves are observed, are rooted at specific locations in the umbra identified by small-scale bright features called Umbral Dots (UDs) and the oscillation power at these locations in the 3-minute period band is at least three orders of magnitude higher than the surrounding umbra. The remarkable correspondence between the oscillation properties across different layers led the authors to suggest that the source of the coronal slow waves is located in the umbral photosphere. \cite{2017ApJ...836...18C} studied velocity oscillations in the photosphere of a sunspot umbra and observed similar enhancements in the 3 minute oscillation power in the vicinity of umbral dots and light bridges. Additionally, they found that these oscillations occur independent of the 5 minute oscillations seen in the rest of the umbra and display an upward propagating nature. Based on these results, in addition to the possibility that umbral dots and light bridges are locations of magnetoconvection, the authors speculated that 3 minute oscillations in sunspots are locally driven through turbulent convection as predicted previously by certain models \citep{1967SoPh....2..385S, 1993ApJ...404..372L}. More recently, \cite{2019ApJ...879...67C} identified different oscillation patterns in the umbral photosphere, all centred on umbral dots that exhibit large morphological and dynamical changes. These authors again concluded magnetoconvection as a possible source of sunspot oscillations. \cite{2014AstL...40..576Z} \& \cite{2020ApJ...888...84S} also find the sunspot oscillations are localised over individual cells of few arcseconds, however, to explain the multiple peaks in the oscillation spectra and their spatial distribution, the authors suggest the existence of a sub-photospheric resonator \citep{1984PAZh...10...51Z} that could drive these oscillations.
\subsubsection{Polar regions}
As stated earlier, it is not trivial for long-period slow waves to propagate from photosphere to the outer solar atmosphere unless some special conditions are met. The ambiguity between wave and flow interpretations for PDs further complicates the issue. Using spectroscopic data from \textit{Hinode}, \cite{2011ApJ...737L..43N} found that line profiles near the base of a structure display flow-like signatures while those at larger distances are compatible with slow waves. The authors further concluded that both waves and flows may originate from unresolved explosive events in the lower corona. \cite{2012ApJ...754..111O} and \cite{2013ApJ...775L..23W} performed 3D MHD simulations demonstrating that impulsive flow pulses injected at the base of a coronal loop naturally excite slow waves which propagate along the loop while the flows themselves rapidly decelerate with height. \cite{2012A&A...546A..93G} studied PDs of 14.5 min periodicity observed in a south polar coronal hole using spectroscopic data. They found that the spectral parameters of Ne{\,}\textsc{viii} 770{\,}{\AA} line exhibit correlated periodic fluctuations in intensity and Doppler shift with no corresponding signatures in line width. Besides, the spectral line profiles did not show any visible asymmetry compelling the authors to interpret the observed PDs as slow waves. The co-temporal data from N{\,}\textsc{iv} 765{\,}{\AA} line, representing the transition region, displays periodic enhancements in line width near the base of the PDs with no associated variations in the other two line parameters. Based on these results, the authors suggest that the slow waves observed in the investigated polar region may be triggered by small-scale explosive events in the lower atmosphere. The base of a network plume, where PDs were observed, was examined in detail by \cite{2015ApJ...807...71P} using both imaging and spectroscopic data from SDO/AIA and IRIS. Quasi-periodic brightenings were found near the base with spectral profiles occasionally displaying Doppler shifts with large deviations from the mean coincident with enhanced line widths and intensity consistent with a flow-like behaviour. The PDs observed along the plume were, however, compatible with slow wave properties. The brightenings were interpreted as due to small scale jet-like features \citep[e.g.,][]{2014Sci...346A.315T,2014ApJ...787..118R} and their correlation with the PDs, according to the authors, indicate them as a possible driver for propagating slow waves observed along the plume.
A four-hour-long image sequence, capturing the evolution of a north polar coronal hole at high spatial and temporal resolutions by SDO/AIA, is analysed by \cite{2015ApJ...809L..17J} to investigate the relation between the PDs observed in the corona and the spicules seen in the lower atmosphere. A direct correspondence between these two phenomena is detected in the composite images constructed by overlaying the lower-atmospheric 304{\,}{\AA} channel data on top of the coronal 171{\,}{\AA} channel data.
\begin{figure}
\includegraphics[width=0.5\textwidth]{kpfig2a}
\includegraphics[width=0.47\textwidth]{kpfig2b}
\caption{\textit{Left}: A section of a north polar coronal hole of the Sun captured by SDO/AIA on 6 August, 2010. The image is a composite with the bottom and top parts displaying the emission observed in 304{\,}{\AA} and 171{\,}{\AA} channels, respectively. From \cite{2015ApJ...809L..17J}. \textit{Right}: A composite time-distance map with the bottom part displaying the rise and fall of spicules as seen in IRIS 1400{\,}{\AA} channel and the top part displaying the PDs as seen in AIA 171{\,}{\AA} channel. The slanted black lines tracking the individual features give an estimate of the propagation speed through their slope. The measured speeds (in km{\,}s$^{-1}$) are listed next to these lines. From \cite{2015ApJ...815L..16S}.}
\label{kpfig2}
\end{figure}
Fig.~\ref{kpfig2} shows one such image from their work in the left panel. While noting that a one-to-one correspondence is not always possible because of the overlapping structures along the line of sight especially near the limb, the authors demonstrate that the periodic behaviour of PDs is because of the repeated occurrence of spicules at the same location. Based on the observed speeds, it is further claimed that both type I and type II spicules contribute equally to the generation of PDs. The exact nature of the PDs is not explicitly addressed but the authors appeared to consider them as mass motions. \cite{2015ApJ...815L..16S} performed a similar study by combining the high-resolution slit-jaw images of a south polar region from IRIS with the co-temporal multi-wavelength data from SDO/AIA. Data from IRIS 2796{\,}{\AA}, and 1400{\,}{\AA} channels along with the AIA 304{\,}{\AA}, 171{\,}{\AA}, and 193{\,}{\AA} channels were used in this study. All these channels, when combined, encompass the observations of solar atmosphere from chromosphere, through transition region, to corona. A correspondence between the spicules in the lower atmosphere and the PDs in the corona is again observed. Furthermore, it is found that the beginning of each PD coincides with the rise phase of a spicule. Brightenings are observed in coronal channels, co-temporal to the fall of spicular material, following which another PD is initiated. A composite time-distance map describing the rise and fall of spicules (as seen in IRIS 1400{\,}{\AA} channel) and the simultaneous propagation of PDs (as seen in AIA 171{\,}{\AA} channel) is shown in the right panel of Fig.~\ref{kpfig2}. Additionally, the authors measured propagation speeds of PDs in 171{\,}{\AA} and 193{\,}{\AA} channels but could not determine if there is any temperature-dependent propagation \citep{2012SoPh..279..427K} because of the large uncertainties in their estimation. However, the presence of alternate bright and dark ridges and their continued existence at large altitudes led the authors to favour a slow-wave interpretation for the observed PDs. Finally, \cite{2015ApJ...815L..16S} conjecture that both spicules and slow waves are simultaneously driven by a reconnection-like process in the lower atmosphere, of which the cold spicular material falls back while the waves continue to propagate upward. Later, based on 1D MHD modelling, \citet{2016ApJS..224...30Y} show that finite-lifetime stochastic transients launched at the base of a magnetic structure can generate quasi-periodic oscillations with signatures similar to PDs along the structure. Gaussian pulses with the typical spicular timescales are used to simulate the transients in their study. An empirical atmospheric model is considered and depending on the smoothness of the thermal structure (with or without chromospheric resonant cavity), it was shown that both long and short-period PDs could be reproduced corresponding to those observed in polar plumes and active region loops, respectively. A more recent study by \citet{2019Sci...366..890S} has further illustrated that the connection between spicules and PD-like coronal moving features is very common even in on-disk observations. \\
\subsection{Damping/Dissipation properties}
The slow waves are observed to fall very rapidly in amplitude with their travelling distance along the solar coronal structures. Various physical mechanisms including thermal conduction, compressive viscosity, and optically thin radiation, which could affect the wave amplitude, were identified and studied through theoretical modelling in the past (see a complete review by \citet{2021SSRv..217...34W}, in this issue). The results indicated, for typical coronal conditions, that thermal conduction is a strong contributor to the observed damping of slow waves and the effects of viscosity and radiation are negligible \citep{2002ApJ...580L..85O,2003A&A...408..755D, 2004A&A...415..705D,2018AdSpR..61..645P}. Later studies revealed exceptions to this scenario in some special conditions, e.g., the dissipation by thermal conduction is insufficient to explain the rapid damping of oscillations with longer periods in typical coronal loops \citep{2011ApJ...734...81M}, while compressive viscosity may become important when the loop is super-hot and with very low density \citep[see][]{2006SoPh..236..127P, 2007SoPh..246..187S, 2021SoPh..296...20P}, or when anomalous transport occurs in hot flaring plasma \citep{ 2019ApJ...886....2W}.
By constructing powermaps in different period ranges, \cite{2012A&A...546A..50K} studied the distribution of oscillatory power in three different open structures, namely, the active region fan loops, network plumes and the polar plumes. It is found that the oscillatory power at longer periods is significant up to larger distances in all the structures indicating a frequency-dependent damping. Furthermore, the measured amplitudes and damping lengths are shorter in hotter wavelength channels. These results are shown to be compatible with a propagating slow wave model incorporating thermal conduction as the main damping mechanism. The frequency dependence has been further explored by \cite{2014ApJ...789..118K}, where the authors attempted to find the exact quantitative dependence of damping length on the oscillation period. For each identified period, the Fourier amplitudes were computed as a function of distance, by fitting with an exponential decay function, from which the associated damping length is derived. Exploiting the simultaneous existence of multiple oscillation periods and combining the results from similar structures, the quantitative dependence of damping length on oscillation period is examined. Considering a power-law dependence of $L_d \propto P^{\alpha}$, where $L_d$ is the damping length and $P$ is the oscillation period, the $\alpha$ values (power-law indices) were obtained as 0.7$\pm$0.2 for the active region loops and network plumes and $-$0.3$\pm$0.1 for the polar region structures using AIA 171{\,}{\AA} channel. The corresponding values for the AIA 193{\,}{\AA} channel are 1.7$\pm$0.5, and $-$0.4$\pm$0.1, respectively. Assuming a 1D linear wave model, the authors also derived the expected theoretical dependences for different damping mechanisms, according to which, the $\alpha$ value should lie between 0 and 2. Notably, for the case of weak thermal conduction, the expected $\alpha$ value is 2 implying a steeper dependence and for the case of strong thermal conduction, this value should be 0 indicating no dependence or uniform damping across all periods. In reality, an intermediate value is possible depending on the strength of the thermal conduction but the obtained values are clearly deviant, especially those from polar regions, uncovering a large discrepancy between the observations and the theory of slow waves. It may be noted that the theoretical dependences ($\alpha$ values) are calculated under the assumption that only a single dissipation mechanism works (or perhaps dominant). Indeed, the negative slopes obtained for polar regions would mean stronger damping at longer periods which is not expected from any of the dissipation mechanisms studied by \cite{2014ApJ...789..118K} in the linear regime. Also, these results appear to be in direct contradiction with \cite{2012A&A...546A..50K}, where the authors reported longer periods travelling to larger distances even in polar regions. However, it was later noted that this behaviour was due to the availability of higher power at longer periods near the base of the corona.
\cite{2014A&A...568A..96G} also studied the damping behaviour of slow waves in polar regions. Their results indicate the presence of two different damping regimes, the stronger damping near the limb ($<10$ Mm height) and weaker damping at larger heights. Furthermore, in order to understand the frequency dependence, damping lengths were measured over three predefined period bands i.e., 4$-$6 min, 6$-$15 min, and 16$-$45 min. No preferable dependence was found near the limb but at larger heights the damping appears to be stronger in the shorter period band which the authors explain as evidence for damping due to thermal conduction. However, it needs to be pointed out that sampling the oscillations at a larger number of periods as obtained in \cite{2014ApJ...789..118K} would be more suitable and better for understanding the period dependence.
\cite{2016ApJ...820...13M} performed 3D MHD simulations based on a uniform cylinder model to investigate the damping of slow waves at multiple periods. By perturbing the pressure at one end of the loop, slow waves are driven at 4 preselected periods. The other boundary is kept open allowing the waves to propagate and escape the domain. Thermal conduction is incorporated into the model as the only damping mechanism for the waves. The gravitational stratification and the field divergence are ignored. The obtained numerical results are forward modelled to aid a direct comparison with the observations. It is found that the damping lengths are dependent on the oscillation period with a power-law index near 1 agreeing with the previous observations for active region loops \citep{2014ApJ...789..118K}. The authors also derived full solutions to the dispersion relation for thermal conduction damping, instead with the approximations made in \cite{2014ApJ...789..118K}, which shows a similar dependence. Based on this, the discrepancy in the previous results was explained as due to the differences in the damping of shorter and longer periods which conform to strong and weak thermal conduction limits, respectively. However, the negative dependence found in the polar regions remains unexplained. In order to verify and examine if the frequency dependence obtained in the polar regions is general, \cite{2018ApJ...853..134M} performed a statistical study by employing 62 datasets of polar coronal hole regions observed between 2010 and 2017 by SDO/AIA.
\begin{figure}
\includegraphics[width=0.98\textwidth]{kpfig3}
\caption{The dependence of damping length on oscillation period extracted from the polar plume (left) and interplume (right) regions using data from the AIA 171{\,}{\AA} channel. The symbols in grey correspond to the individual measurements. The most frequent value (modal value) at each period bin is denoted by an open black circle and the standard width of the associated distribution is shown as an error bar. The slanted black line over the data represents the best fit to the most frequent values. The slope obtained from the fit is listed in the individual panels. From \cite{2018ApJ...853..134M}}
\label{kpfig3}
\end{figure}
Fig.~\ref{kpfig3} displays the results of this study for the data from the AIA 171{\,}{\AA} channel. The left and right panels show the dependences obtained for the plume and interplume regions respectively. The measured slopes obtained by fitting the modal values at each period bin are $-$0.3$\pm$0.1 for plume and $-$0.4$\pm$0.1 for interplume regions. Similar values were obtained for the data from the AIA 193{\,}{\AA} channel. These results not only confirm those from \cite{2014ApJ...789..118K}, that the anomalous dependence of damping length on oscillation period observed in polar regions is real, but also suggest that such a damping relation is typical for slow waves in polar structures which therefore, strongly necessitates an explanation.
More recently, \cite{2019FrASS...6...57S} studied the dependence of damping length on the temperature of the loop. The damping lengths of oscillations were measured in 35 loop structures (selected from 30 different active regions) and the corresponding plasma temperatures were obtained from Differential Emission Measure (DEM) analysis \citep{2012A&A...539A.146H}.
\begin{figure}
\includegraphics[width=0.98\textwidth]{kpfig4}
\caption{\textit{Left}: Relative intensity of oscillation as a function of distance along a sample loop. The orange diamond symbols denote the actual data with the corresponding uncertainties shown as error bars. The solid black curve represents the best fit for an exponentially decaying sine wave model. The damping length value obtained from the fit is listed in the figure. \textit{Right}: Dependence of damping length on loop temperature. The orange diamond symbols denote the values from observations with the corresponding uncertainties shown as horizontal and vertical error bars. The filled green circles highlight the expected values based on a linear slow wave theory incorporating damping due to thermal conduction. The open green circles denote the same but for a constant value (5/3) for the polytropic index. From \cite{2019FrASS...6...57S}}
\label{kpfig4}
\end{figure}
Fig.~\ref{kpfig4} displays the obtained results from this study. The left panel demonstrates the measurement of damping length ($L_d$) by following the phase of a sample oscillation with the obtained $L_d$ value listed in the plot. An alternate method to estimate damping length, by directly tracking the oscillation amplitudes along the loops produces similar results. The right panel of Fig.~\ref{kpfig4} depicts the obtained dependence of damping length on temperature. The expected values from linear wave theory considering thermal conduction as the damping mechanism are also plotted (green filled circles). The oscillation periods, densities, temperatures, and polytropic indices for the individual loop structures, extracted previously by \cite{2018ApJ...868..149K}, are incorporated into these calculations. Additionally, the expected values derived by assuming a constant value of 5/3 (adiabatic index expected in the corona) for the polytropic index are plotted (green open circles). As may be seen, the damping length is expected to decrease evidently to lower values for hotter loops but the observations appear to defy this. Based on these results, the authors concluded that thermal conduction is suppressed in hotter loops which is consistent with previous reports by \cite{2015ApJ...811L..13W}. Furthermore, it may be noted that the theoretical damping length values are 2$-$3 orders of magnitude larger than the observed values. While this discrepancy could be partly explained by the fact that the observed damping length values are projected values (meaning they are lower limits), the authors speculate that such a large difference could alternatively imply that thermal conduction is not the dominant damping mechanism in those coronal loops.
Another aspect that has become popular in recent years is the misbalance in the local thermal equilibrium caused by slow waves which in turn could affect the wave dynamics significantly \citep{2017ApJ...849...62N}. To study this effect, the heating function, $H$, is parametrised as a function of the equilibrium parameters density $\rho$ and temperature $T$, and included in the energy equation of MHD in the recent models \citep{2019A&A...628A.133K, 2019PhPl...26h2113Z, 2020A&A...644A..33K, 2021SoPh..296...20P}. In comparison, in earlier models, $H(\rho,T)$ was either ignored or assumed to be a constant balanced by the other dissipative effects (e.g., radiative cooling). During the linearisation process in the modelling of waves, the partial derivatives of the heating function $\partial H/\partial \rho$ and $\partial H/\partial T$ play a crucial role, introducing new effects on the wave dynamics. In particular, \citet{2019A&A...628A.133K} found that these novel effects may result in additional damping of slow waves, even in the complete absence of thermal conduction. This could also be an interesting avenue to pursue the explanation of earlier results with unexpected dependence of slow wave damping on oscillation period or temperature \citep{2014ApJ...789..118K, 2019FrASS...6...57S}. Indeed, it seems quite promising, because the variation of the polytropic index with loop temperature \citep{2018ApJ...868..149K} is also recovered in the model \citep{2019PhPl...26h2113Z}. \\
\subsection{Seismological applications}
One of the important applications of studying waves in the solar atmosphere is that it gives us the ability to infer important physical parameters through seismology. The common occurrence of slow magnetoacoustic waves makes them even more compelling to pursue this. Over the past decade a number of important parameters including the plasma temperature, coronal loop geometry, coronal magnetic field, thermal conduction coefficient, polytropic index, have been extracted from the observations of slow waves in combination with MHD theories demonstrating their vast potential in this area \citep[for e.g.,][]{2009ApJ...697.1674M, 2009A&A...503L..25W, 2011ApJ...727L..32V, 2014A&A...561A..19Y, 2015ApJ...811L..13W, 2016NatPh..12..179J, 2017ApJ...834..103K, 2018SoPh..293....2D, 2018ApJ...868..149K, 2019ApJ...886....2W, 2020A&A...644A..33K}. A subset of these applications, particularly those reported in the recent years, will be discussed here in detail.
Considering a cylindrical flux tube geometry for coronal loops, within the long-wavelength limit, i.e., when the wavelength of the oscillation, $\lambda$, is much greater than the diameter of the loop, $d$, the slow waves propagate at tube speed, $c_t$, which is dependent on both the sound speed, $c_s$, and the Alfv\'{e}n speed, $v_a$, through the relation $c_t^2 = c_s^2 v_a^2/(c_s^2+v_a^2)$. From the observations of slow waves we know that the long-wavelength limit is generally applicable in the solar corona which means unless the plasma $\beta$ is very small, the propagation speed of slow waves is not only dependent on the plasma temperature but also on the magnetic field strength. By applying this relation to the observations of standing slow waves, \citet{2007ApJ...656..598W} have earlier demonstrated that one could derive the coronal magnetic field strength. More recently, \citet{2016NatPh..12..179J} have further exploited this relation by extending it to the propagating slow waves to derive spatially resolved coronal magnetic fields over a sunspot. \citet{2016NatPh..12..179J} have shown that the propagating slow waves observed in an isolated sunspot from the active region NOAA 11366 possess multiple periodicities with longer periods dominant at farther distances from the sunspot centre. The apparent propagation speeds of longer period waves were found to be significantly smaller compared to that of the shorter periods found near the sunspot centre. By combining the speed-period and period-distance dependences, the authors derived a radial dependence of apparent propagation speed on distance. In order to estimate the real propagation speed the inclination of the magnetic field needs to be known. In principle, the local magnetic field geometry can be obtained from stereoscopic observations of propagating slow waves \citep{2009ApJ...697.1674M, 2020ARA&A..58..441N} or magnetic field extrapolation methods. Although, each of these methods have their own intrinsic difficulties and ambiguities, the authors applied both the methods to derive the inclination of the field incorporating which they deduced the actual propagation speed ($c_t$) as a function of distance from the sunspot centre. The corresponding temperatures and densities were estimated employing a differential emission measure technique. Inputting these values in the expression for $c_t$, \citet{2016NatPh..12..179J} have obtained a magnetic field strength of 32$\pm$5 G near the sunspot centre which is found to rapidly decrease to about 1 G over a distance of 7 Mm. It is not explicitly stated whether these values are consistent with those derived from the extrapolations. Also, in order to improve the signal, the authors had to average the observations in the azimuthal direction so they could only provide an average radial dependence of magnetic field strength but not a fully resolved spatial distribution. Furthermore, the azimuthal averaging would result in the loss of some important structure in this dimension \citep{2014A&A...569A..72S, 2017ApJ...842...59J, 2019ApJ...877L...9K} so this is perhaps not an ideal procedure. Nevertheless, this study clearly highlights the scope of slow waves in determining one of the most important physical parameters of the corona, the magnetic field strength.
Under the condition of low-plasma $\beta$, which is generally assumed for solar corona, the tube speed $c_t$ is almost equivalent to the sound speed $c_s$. So it is common to associate the propagation speed of slow waves with the local sound speed which, in turn, is proportional to the square root of the corresponding plasma temperature, $T$. Thus, one could derive the temperature of the plasma by measuring the propagation speed of slow waves. However, the observations are typically constituted of two-dimensional projected images of the oscillations and hence could only provide a lower-limit on the slow-wave speed. Availing a rare opportunity, \citet{2009ApJ...697.1674M} analysed the propagation of slow waves in a coronal loop observed by two spacecrafts from two different vantage points. These stereoscopic observations made by \textit{STEREO}/EUVI enabled the authors to study the three-dimensional propagation of slow waves and provide their true propagation speed. The obtained propagation speed was about 132$\pm$11{\,}km{\,}s$^{-1}$ corresponding to a plasma temperature of 0.84$\pm$0.15 MK derived under the adiabatic assumption. The observed propagation speeds are, in general, constant along the length of the loop. Earlier observations by \citet{2003A&A...404L...1K} report differential propagation of slow waves in different temperature channels indicating a sub-resolution structure within a coronal loop. However, these authors also did not find any spatial variation of propagation speeds in the individual channels.
\begin{figure}
\includegraphics[width=0.31\textwidth]{kpfig5a}
\includegraphics[width=0.67\textwidth]{kpfig5b}
\caption{The propagation speed of slow waves measured along a selected loop track (left), corresponding deprojected values (middle) and the derived plasma temperatures (right). Different symbols/colors represent results for different channels as described in the figure legend. Note the distances in middle and right panels correspond to deprojected values from the bottom of the loop. The diamond and `x' symbols in the right panel denote the plasma temperatures derived under the isothermal approximation. The dashed lines show corresponding values for an adiabatic case. Adapted from \citet{2017ApJ...834..103K}.}
\label{kpfig5}
\end{figure}
Recently, using the high-spatial resolution observations from \textit{SDO}/AIA \citet{2017ApJ...834..103K} studied the propagation of slow waves along an active region fan loop in two different temperature channels. In addition to the differential propagation, the authors found an accelerated propagation of waves along the loop in both the channels. This enabled them to derive the spatial variation of plasma temperature along the wave propagation path. The authors employed nonlinear force-free magnetic field extrapolations to estimate the inclination of the loop and thereby deduced the deprojected propagation speeds. Fig.{\,}\ref{kpfig5} shows the initial observed speeds, deprojected true speeds, and the obtained temperature values as a function of distance in the left, middle, and right panels, respectively. The results for both wavelength channels are shown in the figure. The temperatures were estimated assuming isothermal propagation of slow waves although corresponding values for adiabatic propagation are also shown (dashed lines in the right panel). It may be noted that the temperature values especially near the bottom of the loop ($\approx$0.2{\,}MK) are much lower compared to the previous results. This is mainly because of the lower propagation speeds observed here. It is possible that there are aperiodic downflows in the loop concurrent with the waves. Although the time-distance maps do not clearly indicate this, such flows are fairly common in sunspots \citep{2015A&A...582A.116S, 2016A&A...587A..20C, 2018ApJ...859..158S, 2020A&A...636A..35N, 2020A&A...640A.120N}. If that is the case, the true propagation speed of the wave (and hence the plasma temperature) could be higher. Nevertheless, the evidence for differential propagation is remarkably clear suggesting a multi-thermal structure within the coronal loop. The authors highlight this as an ability of slow waves to resolve the sub-resolution structure of coronal loops.
Utilising the spectroscopic observations of running slow waves in a coronal loop made by \textit{Hinode}/EIS, \citet{2011ApJ...727L..32V} derived the effective adiabatic index or the polytropic index ($\gamma_\mathrm{eff}$) of coronal plasma. The authors employed the expected relation between the density and temperature perturbations due to a slow wave to achieve this. The required density and temperature estimations were made from appropriate spectroscopic line ratios. The authors obtained a value of $\gamma_\mathrm{eff} =$1.10$\pm$0.02, close to the isothermal value 1, implying that thermal conduction is very efficient in the solar corona. Furthermore, using the observed phase lag between the density and temperature perturbations (under the assumption that it is entirely introduced by thermal conduction), they derived the thermal conduction coefficient, $\kappa_0 =$9$\times$10$^{-11}${\,}W{\,}m$^{-1}${\,}K$^{-1}$, which is of the same order of magnitude as the classical \textit{Spitzer} conductivity. On the other hand, a similar study by \citet{2015ApJ...811L..13W} using the observations of standing slow waves in hot coronal loops, resulted in a polytropic index value of 1.64$\pm$0.08. The proximity of this value to the adiabatic index of 5/3 (for an ideal monatomic gas) implies that thermal conduction is heavily suppressed in the investigated loop. By comparing the observed phase lag between density and temperature perturbations to that of a 1D linear MHD model, the authors suggest that thermal conduction is suppressed at least by a factor of three. \citet{2018ApJ...868..149K} investigated the behaviour of propagating slow waves in multiple fan loops from 30 different active regions. By analysing the relation between corresponding density and temperature perturbations these authors also derived associated polytropic indices which were found to vary from 1.04$\pm$0.01 to 1.58$\pm$0.12. Additionally, the authors studied their temperature dependence which indicates a higher polytropic index value for hotter loops. It was further concluded that this dependence perhaps suggests a gradual suppression of thermal conduction with increase in temperature of the loop thus bringing both the previous studies into agreement.
Coming back to the new models with the thermal misbalance \citep[see e.g.][and references therein]{2019A&A...628A.133K, 2019PhPl...26h2113Z}, these offer an excellent opportunity for slow waves to probe the coronal heating function, which is arguably the greatest mystery in solar or stellar physics. In order to study this \citet{2020A&A...644A..33K} has parameterised the heating function $H$ as power laws of density $\rho$ and temperature $T$. Based on the common existence of long-lived plasma structures and the statistical properties of the slow wave damping, the authors then constrained the power-law indices of these key quantities offering an insight into the characteristics of coronal heating function. It has also been shown that the characteristic thermal misbalance time scales are on the same order as the oscillation periods and damping times of slow waves observed in various structures including flaring loops and coronal plumes. Although this investigation has been largely theoretical, it offers a lot of observational potential to make progress in the decades old question of characterising the coronal heating mechanism. For more details, please see the review by \citet{2020SSRv..216..140V} in this volume. \\
\section{Alfv\'{e}nic/Kink waves}
\label{alfven_waves}
Apart from slow magnetoacoustic waves, kink(Alfv\'enic) waves are also observed in the solar atmosphere at the locations of the open magnetic field regions. Alfv\'enic waves were suggested to have mixed properties as they propagate, having both compression and parallel vorticity and with non-zero radial, perpendicular and parallel components of displacement and vorticity \citep{2009A&A...503..213G,2012ApJ...753..111G,2019FrASS...6...20G,2021A&A...646A..86G}. Thus kink waves under some physical conditions, could share properties of surface Alfv\'en waves \citep[Also see][]{2009A&A...503..213G}.
In this review we will use terms Alfv\'enic and kink interchangeably.
\subsection{Observations}
Both spectroscopic and imaging observations have been used to establish the presence of the Alfv\'enic waves in the solar atmosphere. In the following subsections we discuss them separately.
\subsubsection{Spectroscopic Observations}
\label{section:Spectroscopic Observations}
Perhaps one of the earliest signatures of the presence of Alfv\'enic and Alfv\'en waves (or turbulence) in the solar atmosphere comes from the observations of the nonthermal spectral line broadenings in the solar atmosphere. \citep{1973ApJ...181..547H,2008ApJ...676L..73V,1976ApJ...205L.177D, 1976ApJS...31..417D, 1976ApJS...31..445F, 1990ApJ...348L..77H}. The nonthermal line broadening at the location of the coronal holes was reported by \citet{1998A&A...339..208B, 1998SoPh..181...91D, 2005A&A...436L..35O,1997ApJ...476L..51O,1999ApJ...510L..59K,2016SSRv..201...55A} using the Solar Ultraviolet Measurements of Emitted Radiation (SUMER) spectrometer, Coronal Diagnostic Spectrometer (CDS), and Ultraviolet Coronagraph Spectrometer (UVCS) on-board {\it Solar and Heliospheric Observatory} (SoHO). Furthermore, an increase followed by the flattening of the nonthermal line widths with height above the solar limb was noted \citep{1998A&A...339..208B, 1998SoPh..181...91D, 2005A&A...436L..35O,1990ApJ...348L..77H}. A similar nature of variation of the nonthermal line widths was found at the location of the coronal holes using Extreme Ultraviolet Imaging Spectrometer (EIS) on-board {\it Hinode} \citep{2007ApJ...667L.109D,2009A&A...501L..15B,2012ApJ...753...36H}. Using the nonthermal line broadening these studies have estimated the energy flux carried by the Alfv\'enic waves in the solar atmosphere and found that it is enough to balance total coronal energy losses in coronal holes. Furthermore the flattening or leveling-off of the nonthermal line widths was attributed to Alfv\'en wave damping (see Figure~\ref{ntlw}). The departure from the expected WKB theory (undamped propagation of Alfv\'en wave) happens as early as 0.1 $R_{\odot}$ from the solar surface.
\begin{figure}[ht!]
\centering
\includegraphics[scale=0.25]{lw_wkb.png}
\caption{Variation of the nonthermal line widths with height above a polar coronal hole in different spectral lines. Red and green color represents the nonthermal line widths derived from the eclipse observations in Fe XIII and Fe XIV emission lines, respectively \citep{2011SoPh..270..213S}. Square and diamond in blue are derived from the observations taken in Fe XII using SUMER and EIS, respectively \citep[see,][respectively]{2003ApJ...598..657M,2009A&A...501L..15B}. Solid line represents the nonthermal line widths derived from Fe XII emission line \citep{2012ApJ...751..110B}. The dotted, dashed and dash-dotted curves represent the nonthermal widths expected from WKB propagating for different density profiles as reported in \citet{1999JGR...104.9801G,2012ApJ...751..110B,1999A&A...349..956D}, respectively. Adapted from \citet{2012ApJ...751..110B}}.
\label{ntlw}
\end{figure}
However, recently \citet{2020ApJ...899....1P} have shown analytically and using numerical simulations that for a period averaged Alv\'en(ic) waves (such as those observed by SUMER with large exposure time), the nonthermal line widths are larger than the rms wave amplitude by a factor of $\sqrt{2}$. Further, LOS integration of several structures oscillating with different polarisation and phases of oscillation cause nonthermal line widths of the order of rms wave amplitudes. These results are in contrast with those obtained in some of the above mentioned studies \citep[ e.g,][]{1990ApJ...348L..77H,1998A&A...339..208B, 1998SoPh..181...91D, 2005A&A...436L..35O,1997ApJ...476L..51O,1999ApJ...510L..59K}, where authors assume that the nonthermal line widths are smaller than the rms wave amplitudes by a factor of $\sqrt{2}$. Thus the energy flux, which varies as the square of the rms wave amplitudes, is overestimated at least by a factor of 2 in above mentioned studies.\\
Additionally, several analytical and numerical Alfv\'en wave models have also been made to demonstrate the propagation and dissipation of the Alfv\'en waves and their potential contributions in the observed line-width variations in the solar atmosphere, thus invoking the significant role of these waves in generating turbulence and/or heating \citep[e.g,][and references therein]{2002MNRAS.336.1195P,2006SoPh..237..143D,2013MNRAS.428...40C,2015ApJ...811...88Z,2015A&A...581A.131J}. However, such studies do not consider the present day scenario of the mixed wave modes/Alfv\'enic wave modes and their specific physical properties and role in the line-width broadening or narrowing.\\
The Coronal Multi-channel Polarimeter (CoMP, see \citealt{2007Sci...317.1192T}) has established the ubiquity of Alfv\'enic/kink waves in the solar corona, primarily observed as propagating Doppler velocity fluctuations. \citet{2009ApJ...697.1384T} reported these fluctuations have both an outwardly- and inwardly-propagating nature and a disparity between the outward and inward wave power, which could be a signal of isotropic MHD turbulence. Furthermore, \citet{2015NatCo...6E7813M} observed Alfv\'enic waves that partially reflect while propagating up along the open magnetic field regions, such as polar coronal holes. These results demonstrate that counter-propagating Alfv\'enic waves exist in open coronal magnetic fields, thus providing support for Alfv\'en wave turbulence models (see also section \ref{section:Generation of turbulence in open magnetic field regions}). Interestingly, there is also a disparity in outward-to-inward wave power for closed magnetic fields (see \citealt{2010A&A...524A..23T, 2010ApJ...718L.102V, 2019ApJ...876..106T}). Moreover, low frequency Alfv\'enic waves are more likely to reflect than high frequencies (see the right panel of Figure~\ref{ps}). Earlier, theoretical studies have also reported a similar behaviour for Alfv\'en waves \citep{2005ApJS..156..265C}. These authors reported that these waves are strongly reflected at the transition region (95\%). While beyond transition region, waves with frequency lower than the critical frequency (local gradient of Alfv\'en velocity) reflect, which is about 30 hours ($\sim$1 day) at large distances from Sun. Again using CoMP, an enhancement of the power spectra in 3-5 mHz was noted in the open magnetic field regions (see Figure~\ref{ps}), suggesting the possible role of p-modes in generating Alfv\'enic waves, as a spatially-ubiquitous input \citep{2016ApJ...828...89M} that is sustained over the solar cycle \citep{2019NatAs...3..223M}. An enhancement in power at 3-5 mHz is also seen in the intensity variations from SDO/AIA in quiet sun region and sunspots \citep{2016A&A...592A.153K}.
\begin{figure}[ht!]
\centering
\includegraphics[scale=0.35]{41467_2015_Article_BFncomms8813_Fig3_HTML.jpg}
\caption{{\it (a)} Power spectra of the Doppler velocities showing enhancements at 3-5 mHz. {\it (b)}: Frequency dependent ratio of inward and outward power suggesting that low frequency waves are more likely to reflect \citep[adapted from][] {2015NatCo...6E7813M}. }
\label{ps}
\end{figure}
Furthermore, \citet{2012ApJ...761..138M} have noted a wedge shape correlation between rms Doppler velocities and nonthermal line widths in the solar corona similar to Figure~\ref{wedge}. Using Monte Carlo method, these authors suggested that the observed correlation might be due to the Alfve\'enic wave propagation and LOS integration of several oscillating structures. Since the amplitude of Alfv\'enic wave increases with height, LOS integration of several oscillating structures also increases non-thermal line widths. This lead to a wedge shape correlation. However, they artificially added a substantial amount of nonthermal broadening in the Monte Carlo simulations to match the observed nonthermal broadening. The source of this additional nonthermal bradening was not known. Later, \citet{2019ApJ...881...95P} have studied the correlation between rms Doppler velocities and the nonthermal line widths at different heights at the location of the polar coronal holes using observations from CoMP (see bottom panel of Figure~\ref{wedge}) and numerical simulations. Combining numerical simulations and forward modelling, these authors could reproduce the observed wedge-shaped correlation without adding any artificial source of nonthermal line widths (see left panel of Figure~\ref{fig-wedge}). They proposed that at least a part of the wedge-shape correlation between the nonthermal line widths and rms Doppler velocities is due to the propagation of Alfv\'enic waves in the solar atmosphere and a part is due to height dependence of RMS wave amplitude and non-thermal line widths. (see section \ref{section:forward modelling} for more details). Figure~\ref{wedge} denotes a wedge shape correlation over entire solar corona (top panel) and at the location of the coronal holes (bottom panel) at different heights obtained using the data from CoMP.
\begin{figure}[ht!]
\centering
\includegraphics[scale=0.75]{fig0a.pdf}
\includegraphics[scale=0.55]{fig0b.pdf}
\caption{{\it Top} Wedge-shaped correlation displaying the distribution of the rms Doppler velocities and mean nonthermal line widths over entire CoMP FOV. {\it Bottom}: At the location of coronal holes segregated with height. Adapted from \citet{2019ApJ...881...95P}.}
\label{wedge}
\end{figure}
\subsubsection{Imaging Observations}
\label{section:Imaging Observations}
\begin{figure}[t]
\centering
\includegraphics[scale=0.7]{Figure_from_Weberg_et_al_2020.jpg}
\caption{Box and whisker plot of wave parameters for slits analysed (31 in total). Lower and upper bounds of each coloured box correspond to the first and third data quartiles, respectively. Horizontal lines indicate median values, diamonds denote log-normal means, \lq{whiskers}\rq{} show the log-normal standard deviations. Solely as a visual aide, every five slits appear in the same colour. \citep[adapted from][] {2020ApJ...894...79W}. }
\label{Figure: Weberg et al 2020}
\end{figure}
In addition to the nonthermal broadening, direct observation of transverse waves have been reported in polar coronal holes. Using data from SDO/AIA, \cite{2011Natur.475..477M} observed ubiquitous outward-propagating Alfv\'enic motions in the corona with amplitudes of the order of 25 $\pm$ 5 km s$^{-1}$, uniformly-distributed over periods of 150-550 seconds throughout the quiescent solar atmosphere. These estimates were obtained using Monte Carlo simulations, which we refer to as indirect measurements. Using these indirect measurements, the energy flux carried by these waves was estimated to be enough to accelerate the fast solar wind and power the quiet corona. Again using SDO/AIA, \citet{2014ApJ...790L...2T} provided the first {\it{direct}} measurements of transverse wave motions in solar polar plumes but, in contrast to \cite{2011Natur.475..477M}, found lower Alfv\'enic wave amplitudes around 14 $\pm$ 10 km s$^{-1}$, as well as observing a broad distribution of parameters which is skewed in favour of relatively-smaller amplitude waves. When this whole range of parameters is taken into account to calculate the energy flux carried by these waves, the energy budget, determined by direct measurement, falls 4-10 times below the minimum theoretical requirement (models require 100–200 W m$^{-2}$ of wave energy flux near the coronal base to match empirical data, see, e.g. \citealt{1988ApJ...325..442W} and \citealt{1995JGR...10021577H}). Furthermore, these energy budget calculations assume a filling factor of unity in a homogeneous plasma, which corresponds to the most generous case in terms of energy flux. Thus, more realistic estimates about the inhomogeneity of the medium would yield even smaller energy fluxes (see, e.g. \citealt{2013ApJ...768..191G}).
\citet{2014ApJ...790L...2T} conclude that transverse waves in polar coronal holes are insufficiently energetic to be the dominant energy source of the fast solar wind. Furthermore, it is interesting to note that the directly observed transverse velocity amplitudes are broadly compatible (within standard deviations) with the line-of-sight nonthermal velocities reported in plume plasma by \citet{2009A&A...501L..15B} using Hinode/EIS. If one assumes that these nonthermal velocity measurements contain contributions from both transverse (kink) waves and torsional Alfv\'en waves \citep{2019ApJ...870...55G}, this comparability suggests torsional Alfv\'en waves may not contribute strongly to plume nonthermal velocity measurements (otherwise we would expect nonthermal line-of-sight velocities to be larger). However, we must be cautious since this may also be due to a lack of large-amplitude torsional motions in plumes and/or under-resolved line-of-sight motions using our current spectrometers (see also Section \ref{section:Spectroscopic Observations} for further discussion of nonthermal motions).
Recently, \citet{2020ApJ...894...79W} utilised the NUWT automated algorithm (see \citealt{2018ApJ...852...57W} for details, as well as Section 2.2 of \lq\lq{Novel techniques in coronal seismology data analysis}\rq\rq) to analyse the transverse waves above the polar coronal hole, with a specific interest in how those properties vary with altitude. Between altitudes of 15 to 35 Mm, it was found that measured wave periods were approximately constant, and that the displacement and velocity amplitudes increased at rates consistent with undamped waves (Figure~\ref{Figure: Weberg et al 2020}). Between 5 and 15 Mm above the limb, the relative density was inferred to be larger than that expected from 1D hydrostatic models, which may indicate a more extended transition region with a gradual change in density.
\subsection{Theoretical modelling}
\subsubsection{Generation of turbulence in open magnetic field regions}
\label{section:Generation of turbulence in open magnetic field regions}
Although some observations presented in the previous subsection show that the omnipresent propagating Alfv\'enic waves have roughly enough energy to power the quiet sun, the problem of how this energy is converted to heat still persists. Because of the very high Reynolds numbers of flows in the corona
(e.g. $R_m \approx 10^8 - 10^{12}$, see \citealt{2005psci.book.....A}), any potential damping mechanism must include the generation of scales small enough so that wave energy can be efficiently dissipated. The most popular damping mechanisms of Alfv\'en and kink(Alfv\'enic) waves are phase mixing and resonant absorption respectively.
For more on the linear damping mechansism, see, e.g., \citet{2015RSPTA.37340261A}. In this subsection we will focus instead on nonlinear damping mechanisms, such as turbulence. In MHD turbulence, damping occurs through the nonlinear cascade of wave energy down to the dissipation scales. There are several wave turbulence models which deal with coronal heating and solar wind acceleration in open coronal structures \citep[see, e.g.,][]{2009ApJ...707.1659C,2013ApJ...776..124P,2016ApJ...821..106V,2017ApJ...835...10V,2019ApJ...880L...2S,2019JPlPh..85d9009C}. While a few compressible models are able to maintain a multi-million degree corona and accelerate the solar wind through turbulent heating, it is known that turbulent heating without compressible effects appears to be insufficient in open coronal regions \citep{2016ApJ...821..106V,2019SoPh..294...65V}, even when injecting Alfv\'en waves with rms velocities on the order of the full non-thermal line broadening observed in the corona. All of the previous turbulent models rely on the linear process of reflection of outgoing waves along radial Alfv\'en speed gradients in order to generate turbulence. The presence of nonlinearly interacting counter-propagating Alfv\'en waves is thought to be a necessary condition (see subsection \ref{section:tm_solwind}). While this requirement still holds in incompressible and transversely homogeneous plasma, as employed in most of the previous models, it is now known that MHD waves affected by perpendicular structuring, such as surface Alfv\'en, kink, or Alfv\'enic waves can self-cascade nonlinearly, leading to what is referred to as uniturbulence \citep{2017NatSR...14820M,2019ApJ...882...50M}, a term which is used to differentiate this turbulence generation mechanism from the counter-propagating one. The nonlinear energy cascade rate for kink waves propagating in a cylindrical flux tube was calculated recently by \citet{2020ApJ...899..100V}. It was shown that for thin coronal strands and kink wave amplitudes that are comparable to the observed ones the energy cascade time scale can be shorter than the wave period. Uniturbulence lifts the necessity for reflected waves in turbulence generation, however it relies on the presence of structuring. Such structuring may be obtained through other mechanisms, such as the parametric decay instability (see subsection \ref{section:tm_solwind}). As there is ample evidence for coronal structuring perpendicular to the magnetic field \citep[see, e.g., ][]{2014ApJ...788..152R,2018ApJ...862...18D}, uniturbulence might play a major role in turbulence generation. Preliminary results show this is indeed the case \citep{2021ApJ...907...55M}, however more research needs to be conducted in order to estimate the added contribution to coronal heating and solar wind acceleration.
\subsubsection{Forward modelling of Alfv\'enic waves in the open magnetic field regions}
\label{section:forward modelling}
The emission observed by a telescope is projected in the plane of sky and affected by the line-of-sight (LOS) effects because of the optically thin nature of the solar corona. The LOS superposition of the emission makes the interpretation of the observed spectroscopic or imaging features difficult to interpret in the solar atmosphere. To account for the LOS effects, a forward modelling of 3D numerical simulations is needed, by which the generated synthetic spectra and images can be compared with the real observations to test the validity of the MHD models \citep{10.3389/fspas.2016.00004}. \citet{2013ApJ...778..176O, 2017ApJ...845...98O} performed the forward modelling of Alfv\'en waves in the solar atmosphere incorporating both wave propagation and dissipation in open and closed magnetic field regions in optically thin EUV emission lines. A fairly good match is obtained between the total intensities and the nonthermal line widths derived from the synthetic spectra and those derived from the observations of EIS and SUMER \citep{2013ApJ...778..176O, 2017ApJ...845...98O}. Further, it was noted that the rms wave amplitudes obtained using forward modelling were smaller than those expected for the undamped waves (using WKB assumption) for both open and closed magnetic field regions. Wave dissipation due to counter-propagating waves and wave reflection are proposed to be the main mechanism for damping of Alfv\'en waves in AWSoM, causing rms wave amplitudes smaller than the expected WKB propagation \citep{2017ApJ...845...98O}. It should be noted that the perpendicular structure of Alfv\'en waves is prescribed by the driver, in contrast with collective fast magnetoacoustic modes of perpendicular plasma non-uniformities, the results of such modelling could be sensitive to the perpendicular structuring. In addition to the global model such as Alfven-Wave driven sOlar wind Model (AWSoM), models based on the collection of the single and multiple flux tubes are also used for studying the wave-driven turbulence \citep[][see section \ref{section:Generation of turbulence in open magnetic field regions} for more details.]{2017ApJ...835...10V,2017NatSR...14820M,2019ApJ...881...95P}. \citet{2012ApJ...746...31D} studied the effect of LOS integration of propagating Alfv\'enic waves in multistranded coronal loops using numerical simulations and forward modelling. They reported that the energies estimated from the derived Doppler velocities from their model are an underestimation of the actual energy present in the Alfv\'enic waves. Recently, \citet{2019ApJ...881...95P} has performed the 3D numerical simulations of propagating Alfv\'enic waves in open magnetic field regions with both transverse and longitudinal structuring leading to the reflection of waves and generation of the uniturbulence. However, it should be noted that recently \citet{2021ApJ...907...55M} have shown that the reflected wave power is quite less and the turbulence in transversely inhomogeneous plasma is primarily due to self cascading of propagating kink(Alfv\'enic) waves (uniturbulence). These authors performed the forward modelling for Fe XIII emission line centered at 10794\AA~and compared the synthetic spectra with those obtained from the CoMP. Their model was able to reproduce the observed wedge-shaped correlation between the nonthermal line widths and rms wave amplitudes. The left panel of Figure~\ref{fig-wedge} shows the wedge-shaped correlation between nonthermal line widths ans rms Doppler velocities derived from the synthetic spectra. It matches reasonably well with the correlation observed in the solar corona (see Figure~\ref{wedge}). Right panel of Figure~\ref{fig-wedge} shows the variation of nonthermal line widths for different strength of velocity drivers (shown in different colors). The nature of variation matches with those observed in the solar atmosphere using the data from SUMER \citep{1998A&A...339..208B} and EIS \citep{2012ApJ...753...36H}. The curve in black represents the variation of the nonthermal line widths with heights for a non-stratified plasma. Large nonthermal line widths in these simulations are caused by the LOS integration and uniturbulence due to transverse inhomogeneity. Comparison of the gravitationally stratified cases with the non-stratified case shows that the increase in the nonthermal line widths leading to the wedge-shaped correlation is due to the increase in the amplitude of Alfv\'enic waves with height (due to decrease in the density). Also, the Right panel of Figure~\ref{fig-wedge} shows that the nonthermal line widths levels off with height. Though, the leveling-off of the nonthermal line widths is attributed to the wave damping \citep{2012ApJ...753...36H}, \citet{2020ApJ...899....1P} have suggested that a part of leveling-off of the nonthermal line widths could be due to the non-WKB propagation of the low frequency (large wavelength) Alfv\'enic waves in the gravitationally stratified plasma. The reflection of Alfv\'enic waves is also observed in the solar corona as explained in section~\ref{section:Spectroscopic Observations}. Though the forward modelling of the uniturbulence models have shown that the LOS effects are quite important while reproducing the observed spectroscopic properties of the solar corona, the influence of uniturbulence (with minimal LOS integration) on the nonthermal line broadening is yet to be ascertained.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.4,angle=90]{figure-wedge.pdf}
\vspace{-2cm}
\caption{{\it Left:} Wedge-shape correlation derived from the synthetic spectra obtained by performing forward modelling on 3D MHD simulations and LOS integration over 100 segments oscillating in the random directions and in random phases. Different colors represents different rms velocities used for exciting the Alfv\'enic waves at the bottom boundary. {\it Right:} variation of the nonthermal line widths with height in the solar atmosphere. The variation qualitatively matches with those observed in the solar corona. Different colors represent different strength of the velocity drivers. Curve in black shows the variation of nonthermal line widths without gravity \citep[adapted from][]{2019ApJ...881...95P}.}
\label{fig-wedge}
\end{figure}
\subsection{Coronal Seismology}
Apart from heating the solar corona Kink (Alfv\'enic) waves in the solar atmosphere can be used to diagnose the local plasma conditions because properties of waves depend on the properties of the medium in which they travel. Here, we will discuss the seismological application of the kink(Alfv\'enic) waves in the solar atmosphere, particularly in the open magnetic field regions. Recently, there have been a significant developments in diagnosing the plasma conditions using the principle of Bayesian inference \citep{2021ApJS..252...11A,2018AdSpR..61..655A}. \citet{2019A&A...625A..35A,2020A&A...640L..17M} have successfully used the Bayesian inference techniques to infer the local plasma parameters and field conditions for the standing and propagating kink waves. Though there have been a plethora of studies in coronal seismology dedicated to inferring the cross field density structures and magnetic fields in closed magnetic field regions, there are limited studies on the application of the coronal seismology of propagating kink waves in the open magnetic field regions. \citet{2015NatCo...6E7813M} applied the coronal seismology technique to estimate the density and magnetic field variation at the location of open magnetic field regions. The nature of variation of magnetic field with radial distance was consistent with those predicted by the potential field source surface (PFSS) model. Similarly, \citet{2020ApJ...894...79W} has estimated the density gradients studying the height variation of the amplitude of the Alfv\'enic waves in the open magnetic field regions. The nature of variation of density matches well with those derived in \citet{2012ApJ...751..110B}. The observed slow variation of density leads \citet{2020ApJ...894...79W} to claim that there might exists an extended transition region due to the presence of spicules. \citet{2018ApJ...856..144M} have tested the accuracy of seismologic inversion to infer magnetic field strength in open magnetic field regions using MHD simulations and forward modelling of the propagating Alfv\'enic waves. They found that the accuracy of inversions does not depend on the fine structuring and choice of the spectroscopic lines. Recently, \citet{2020Sci...369..694Y} and \citet{2020ScChE..63.2357Y} observed prevailing transverse MHD waves with CoMP (which they identified as kink waves, which have an Alfv\'enic nature) and used these waves, in conjunction with estimates of electron number density, to map the plane-of-sky component of the global coronal magnetic field (found to be around 1 to 4 gauss, between 1.05 to 1.35 solar radii, where these results encompass both open and closed magnetic fields). Overall, these studies indicate that Alfv\'enic waves are a fundamental feature of the Sun's corona.
\section{Connection with solar wind acceleration}
\label{solar_wind}
Apart from heating the solar atmosphere and diagnosing the plasma conditions, Alfv\'enic waves play an important role in accelerating the solar wind. After the advent of space based observations, there has been a significant observational and theoretical advancement in this topic of research.
\subsection{Observational inputs}
Recent observations show that Alfv\'{e}nic waves are abundant at least at the bottom of the open-field coronal hole regions.
The chromosphere of open-field regions is often filled with long spicules \citep{2012ApJ...759...18P}. High-resolution and high-cadence observations from Hinode/SOT have revealed the prevalence of swaying motions of spicules, which have been interpreted as Alfv\'{e}nic waves \citep{2007Sci...318.1574D}. These transverse waves have been found to reveal amplitudes of 10-25 km/s and periods of 100-500 seconds. The estimated energy flux is on the order of 100 W m$^{-2}$, which is comparable to that required to drive the solar wind (for more detail see \citet{2020SSRv..216..140V}). The authors claimed that the discovery of these waves supports the solar wind models based on dissipation of low-frequency Alfv\'{e}n waves. However, relatively higher frequency ($\sim$60 s) waves have also been detected in some spicules, which may support solar wind models based on high-frequency Alfv\'{e}n waves \citep{2009A&A...497..525H, 2012arXiv1207.6417Y, 2011ApJ...736L..24O}.
Signatures of twisting and torsional motions have also been reported in spicules. From the chromospheric spectra taken by SST, \cite{2012ApJ...752L..12D} identified both swaying motions and torsional motions \citep[also see][]{2017NatSR...743147S}. The amplitude of torsional motion is often slightly higher than that of the transverse motion. A subsequent study using IRIS observations has largely confirmed the existence of these motions \citep{2014Sci...346D.315D}. Some of these torsional motions could be related to torsional Alfv\'en waves, which may be associated with significant heating of the chromospheric and coronal plasma.
IRIS has performed direct imaging and simultaneous spectroscopic observations of the solar transition region for the first time. The high-resolution image sequences taken by IRIS have revealed the prevalence of intermittent small-scale transition region jets from network lanes in the quiet Sun and open-field coronal hole regions \citep{2014Sci...346A.315T,2016SoPh..291.1129N}. These network jets are the most prominent dynamic features in the transition region, and some of them are likely the heating signature of the chromospheric spicules \citep{2014ApJ...792L..15P,2015ApJ...799L...3R}, whereas others appear to be the smallest jetlets discovered from SDO/AIA observations \citep{2014ApJ...787..118R,2018ApJ...868L..27P}. \cite{2014Sci...346A.315T} found that these jets exhibit swaying motions similar to that found in chromospheric spicules. In addition, the simultaneously taken spectra of Si IV 1393.77 are found to be obviously broadened at the locations of these jets, demonstrating that the well-known nonthermal broadening of transition region lines \citep{1998ApJ...505..957C} is actually related to the network jets. As stated in their paper, unresolved transverse motions associated with kink waves and torsional Alfv\'en waves propagating along these jets are most likely responsible for the nonthermal broadening of the transition region spectral lines. The wave amplitude was estimated to be $\sim$20 km/s, and the energy flux associated with the waves is 1-2 orders of magnitude higher than that required to drive the solar wind.
\subsection{Theoretical modelling}
\label{section:tm_solwind}
The acceleration and heating of the solar wind are processes that are strongly coupled. \citet{Leer_1980JGR....85.4681L} realised this by noticing that the density of the wind at the critical point increases exponentially with the temperature at the base of the corona. Hence, the only way of generating a fast wind is to deposit the heat beyond the critical point. This fact naturally led to Alfv\'en waves becoming a favourite candidate for accelerating the solar wind (noting that the slow magnetosonic waves do not show sufficient energy flux in the solar corona). This result has been further supported by the detection of Alfv\'enic fluctuations in-situ at distances 0.7-1AU by the Mariner 5 spacecraft by \citet{Belcher_1971JGR....76.3534B}, and Alfv\'enic waves permeating the base of the corona \citep{2007Sci...317.1192T,2015NatCo...6E7813M,2020Sci...369..694Y,2020ScChE..63.2357Y}. However, the latter observations do not detect sufficient energy flux in the Alfv\'enic fluctuations for coronal heating or solar wind acceleration, and the idea whether Alfv\'en waves and turbulence could account for solar wind acceleration and heating is still controversial \citep[e.g.,][]{2010ApJ...711.1044R}.
Indeed, Alfv\'en waves are notoriously difficult to dissipate at the base of the corona and the wave-action flux is expected to be constant with radial distance due to stratification and no dissipation (assuming radial expansion of the magnetic field in polar regions). This means that the non-linear factor $\delta B/B$ is expected to steadily increase with radial distance \citep{Jacques_1977ApJ...215..942J}. Hence, non-linear effects, which are a major candidate for Alfv\'en wave dissipation, become increasingly important at large distances \citep[please see][for a review on solar wind models]{Hansteen_2012SSRv..172...89H}, while super-radial magnetic field divergence may counter this effect \citep[see the review,][] {2016SSRv..201...55A}. In the nearly collisionless heliospheric solar wind plasma the eventual dissipation of turbulence and heating must occur on kinetic scales and is modeled typically with hybrid or Particle-In-Cell (PIC) or Vlasov's codes (for example, see the review \citet{2010LRSP....7....4O} and recently, \citep{2018ApJ...853...26F,2019SoPh..294..153R,10.3389/fspas.2016.00004,2019ApJ...879...53A,2019SoPh..294..114P,2019ApJ...887..208P}).
Both, compressible and incompressible models for Alfv\'en wave dissipation have been successful in generating a corona and accelerating the solar wind. However, these models usually rely on empirical or WKB description of the waves, and the dissipation/heating processes are not modeled explicitly (i.e., the detailed physics is not included). Moreover, the perpendicular structuring of the plasma, intrinsic for the corona, which affects dramatically the MHD wave propagation, is usually neglected in these models. Regardless of the model the aims are the same: secure a turbulent cascade and/or the generation of small-scales down to kinetic dissipation scales in which the heating can take place in the nearly collisionless plasma. In the lower solar corona and in active regions collisions are more frequent, and MHD dissipation through viscosity and resistivity may be appropriate. One way to achieve this, particularly effective at the base of the corona and in closed loops, is the nonlinear mode conversion of Alfv\'en waves into longitudinal waves due to the centrifugal and ponderomotive forces (see \citet{2020SSRv..216..140V} for details). Since the area expansion factor directly controls the amplitude of these forces, numerical models have been successful at generating both the slow and fast solar wind only by changing this factor and was shown first using 2.5D MHD by \citet{1998JGR...10323677O}. Another popular mechanism is the Alfv\'en wave turbulence, which is discussed in the previous section, is based on the amount of wave-to-wave interaction in very low-beta plasma conditions. Alfv\'en wave turbulence can be achieved from linear processes as in the RMHD models \citep[e.g.][in which wave reflection is only achieved from the change in Alfv\'en speed with height]{Velli_1993A&A...270..304V}, or nonlinear processes such as the parametric decay instability (PDI).
An important constraint in models of incompressible Alfv\'en wave turbulence is the ratio of sunward to anti-sunward Alfv\'en wave energies. This ratio is observed to increase with distance \citep{Bavassano_2000JGR...10515959B}, from a value of $\approx0.3$ at 1.5 AU to $\approx 0.5$ at 4 AU, a fact that appears in opposition to Alfv\'enic turbulence due to the expected dynamic alignment effect \citep{Dobrowolny_1980PhRvL..45..144D}. Indeed, a daughter (reflected) wave is expected to decay faster than a parent wave, thereby leading to a dominant anti-sunward wave fraction and a corresponding decrease with radial distance of the sunward to anti-sunward ratio.
The PDI is an instability of the Alfv\'en wave in compressible low beta plasmas \citep{Sagdeev_1969npt..book.....S,Goldstein_1978ApJ...219..700G} and applies to both circularly polarised or linearly polarised waves. Assuming that the original (or parent) wave is propagating rightward, due to interaction with a compressible wave the rightward propagating Alfv\'en wave decays into a rightward propagating acoustic wave and a leftward propagating Alfv\'en wave. This nonlinear reflection is also known as backscattering. \citet{Shoda_2016ApJ...820..123S} show that this kind of reflection due to compressible waves is substantial and allows to explain the observed increase in the ratio of sunward to anti-sunward Alfv\'enic fluctuations despite of the dynamic alignment effect.
Another important dissipation mechanism for Alfv\'en waves in the solar wind is phase mixing (\citet{Heyvaerts_1983AA...117..220H}, see \citet{2020SSRv..216..140V} for details). This mechanism is efficient in the presence of large transverse density inhomogeneity such as that produced by the PDI. The effects of phase mixing on the dissipation of Alfv\'en waves due to background inhomogeneity in coronal holes were studied using 2.5D MHD and theoretical models in the past \citep[see the reviews][]{2005SSRv..120...67O,2016GMS...216..241O}. One of the first fully self consistent models of fast solar wind acceleration by Alfv\'en waves in a transversely inhomogeneous medium in 2.5D MHD was carried out by \citet{2012ApJ...749....8M}.
By conducting a first of its kind 3D MHD numerical simulation including compressible effects, \citet{Shoda_2018ApJ...859L..17S} show that the early nonlinear stage is dominated by phase mixing \citep[as the kind discussed by][]{2017NatSR...714820M}, thanks to the fast growth of the PDI (see Figure~\ref{fig-pdi} (b) and (d)). On the other hand, Alfv\'en wave turbulence dominates the fully nonlinear stage (see Figure~\ref{fig-pdi} (c) and (d)). It is therefore expected that regions with large density fluctuation (that may be attributed to the PDI), such as the coronal base and solar wind acceleration region \citep{2014ApJ...788..152R}, are likely dominated by phase mixing, while the other regions are dominated by Alfv\'en wave turbulence. The observed fast saturation of the non-thermal line broadening \citep{Hahn_2013ApJ...776...78H} may then be a consequence of the PDI-driven large density fluctuations at the coronal base.
Recently, quasi-periodic fast propagating (QFPs) waves associated with solar flares were discovered with SDO/AIA in EUV observations of coronal active regions \citep{2011ApJ...736L..13L}. These waves were identified as fast magnetosonic waves using 3D MHD modeling as well as 2.5D MHD model in a waveguide, and shown to carry significant energy flux for coronal heating \citep{2011ApJ...740L..33O,2014A&A...569A..12N,2018ApJ...860...54O}. On the other hand, using observations in the radio wavelengths \citet{2018ApJ...861...33K} detected signatures of QFP waves in the open coronal structure at about 1.7 R$_{sun}$, carrying the energy flux at least an order of magnitude lower than the local radiative losses. Since the fast magnetosonic waves can travel across the magnetic field, they can potentially transfer energy flux from flaring loops to open magnetic field regions. Since the initial discovery the QFPs were observed in many events. The observations of QFPs show that the waves damp within short distance (several wavelengths) from the flaring source in an active regions, that is likely due to the combination of magnetic field divergence and wave dissipation processes. At present, the statistics and occurrence rate of the QFPs in the corona is in initial stages of study \citep{2016AIPC.1720d0010L}, and it is not yet known whether the QFPs are produced at sufficient rate and amplitude for coronal heating.
The solar wind speed and the location of the heating are furthermore critically dependent on the photospheric perpendicular correlation length $\lambda_0$, which sets the length scale of the dominant driver for the transverse waves. The solar wind speed generally increases with $\lambda_0$ \citep{Verdini_2010ApJ...708L.116V}, while the maximum temperature is almost constant with respect to it. As explained by \citet{Shoda_2018ApJ...853..190S}, this is because shock heating and turbulence heating readjust. Indeed, a reduction in $\lambda_0$ leads to enhanced turbulence heating, which then increases coronal pressure and reduces compressibility, thereby reducing the efficiency of shock heating. Values of $\lambda_0 \lesssim 1$ Mm lead to dominant turbulence heating. However, compressibility still plays an essential role in this heating since it enhances the wave reflection via the PDI \citep{Shoda_2018ApJ...853..190S}.
The value of $\lambda_0$ is determined by the properties of magnetoconvection, but its dominant spatial range is still a matter of debate. If it is determined by the photospheric kinetic energy peak then $\lambda_0\approx1$~Mm is expected \citep{Rempel_2014ApJ...789..132R}, case in which granulation (and thus the observed swaying motion of flux tubes) is the main driver. However, sub-arcsecond magnetic patches and vortex flows may be more energetically important, setting $\lambda_0\approx0.1$~Mm.
\begin{figure}[!h]
\includegraphics[scale=0.4,angle=90]{fig_pdi.pdf}
\caption{Growth of parametric decay instability (PDI) with time. {\it (a)}: Circularly polarized Alfv\'en wave. Top panel shows a transverse slice and bottom panel shows a longitudinal slice. A random density fluctuation is imposed along z to trigger PDI. {\it (b)}: At 12$\tau_{0}$ (saturation phase), PDI has fully developed in simulations leading to large density imhomogeneities and phase mixing. {\it (c)}: At 24$\tau_{0 }$ (fully non-linear phase), Alfv\'en wave turbulence (AWT) is developed. {\it (d)}: Evolution of the growth rate of PDI, phase mixing, and Alfv\'en wave turbulence of forward and backward propagating (denoted by + and -) with time. Initially, PDI dominates the system. It leads to large transverse density inhomogeneities leading phase mixing dominate the system in the saturation phase. Finally, in the fully non-linear phase, AWT dominates the system. Adapted from \citep{Shoda_2018ApJ...853..190S}}
\label{fig-pdi}
\end{figure}
\section{Summary and Future directions}
\label{summary}
The progress made in recent years in our understanding of MHD waves as observed in open coronal structures is presented here. We particularly discuss the results pertaining to slow magnetoacoustic waves and Alfv\'enic waves alongside the role of latter in the acceleration of solar wind. The openness of coronal structures is determined based on the wave propagation lengths (with respect to the scale of the structure) but not on their magnetic connectivity. So, as such, the relevant research on propagating slow waves in active region loops is also discussed in this article.
It has been long known that the slow waves observed in active region fan loops originate in the lower atmosphere. Recent high-resolution, multi-instrument, multi-wavelength observations have provided a concrete evidence to show that they are actually generated via the leakage of photospheric oscillations into the outer layers \citep[e.g.,][]{2012ApJ...757..160J, 2015ApJ...812L..15K, 2016ApJ...830L..17Z, 2017ApJ...836...18C}. However, different schools of thought still exist on how these photospheric oscillations are generated. Some researchers find a connection with global $p$-modes while others demonstrate that they are locally generated through turbulent magnetoconvection. The possible location of the source region beneath the photosphere has also been discussed. Despite having similar properties, the existence of long period oscillations observed in polar regions is difficult to explain. Their origin has been studied in relation to transient events (spicules, reconnection jets, stochastic transients, etc.) occurring in the lower atmosphere \citep[e.g.,][]{2015ApJ...809L..17J, 2015ApJ...815L..16S, 2016ApJS..224...30Y, 2019Sci...366..890S}. While we are yet to find a solid evidence to pinpoint to a single source, the bigger challenge, in fact, is to clearly distinguish the PDs in terms of either slow waves or the recently discovered high-speed quasi-periodic upflows. Indeed, using simple models of slow waves and periodic flows, \cite{2015SoPh..290..399D} did not find any observational characteristics that would allow to distinguish between the two interpretations in isolation (i.e., without relying on a direct comparison between the two models) within the spatial, temporal, and spectral limits of the current instruments. On the other hand, \citet{2010ApJ...724L.194V} demonstrated that the observables which are usually considered as signatures of upflows could be easily explained in terms of the wave model.
Recent work has highlighted that the PDs observed in network plage regions might be more complex than a simple interpretation in terms of waves or flows. \citet{2005ApJ...624L..61D} and \citet{2006RSPTA.364..383D} already suggested a close relationship between PDs in the corona and (type II) spicules, based on the similarities in their properties. This was recently confirmed by \citet{2017ApJ...845L..18D}, based on combining SDO/AIA and IRIS observations with numerical simulations by \citet{2017Sci...356.1269M}. Their modelling results suggest that in plage region loops, PDs are the signatures of a complex series of events where the generation of spicular flows is linked to shock waves propagating into the corona and plasma heating derived from the dissipation of both waves and electric currents. PDs are not always small-amplitude perturbations of a pre-existing coronal structure but can actually represent the formation of new coronal strands. This complexity results in a mixture (and hence wide range) of properties, particularly the propagation speeds, as aspects of real flows, shock waves and current dissipation are all present, providing a natural explanation for the apparent contradictions in the interpretation of PDs in the literature. Although they are generally reported to only contain very modest energy budgets, PDs could play a significant role in the mass and energy flow in the solar corona through their close link with spicule-driven heating.
The rapid damping of slow waves has been extensively studied. It has been found that PDs in polar regions exhibit anomalous frequency-dependent damping (see Fig.{\,}\ref{kpfig3}) which is not compatible with traditionally considered dissipation mechanisms within the linear regime. It may seem that the damping lengths are underestimated especially for longer periods because of the instrument sensitivity or other detection limitations at large distances from the limb. However, the gradual decrease in damping length observed together with the fact that the damping lengths at longer periods are actually smaller than the possibly more reliable values at shorter periods makes these observational results convincing and therefore, requires an explanation. The temperature dependence of damping lengths in active region PDs also shows notable discrepancies with the theory (see Fig.{\,}\ref{kpfig4}). The results appear to suggest that thermal conduction is suppressed in hotter loops which is, although in agreement with previous observations, counter-intuitive. Further investigations are necessary to develop a proper understanding of this behaviour. The damping of slow waves due to the newly introduced thermal misbalance, is another interesting aspect to pursue further in the future.
A number of applications of slow waves that are useful to probe the plasma temperature, magnetic field strength, thermal conduction, polytropic index, and coronal heating function are discussed. While noting that there are still limitations to the individual techniques applied (e.g., projection effects), one may start exploring the scope of implementing the methods over large datasets.
Although indirect evidence of Alfv\'enic waves in the polar regions has been known since early 1970s, only after the advent of high resolution imaging and spectroscopic instruments such as AIA/SDO, EIS/Hinode, and CoMP, these waves could be resolved both spatially and temporally. Past observations and recent simulations have outlined the importance of studying the nonthermal line widths in establishing the prevalence of Alfv\'enic waves in the solar atmosphere and estimating the wave energy flux. The nonthermal line broadening in the corona is found to be correlated with the Doppler velocities, a feature that can be reproduced by the propagating kink/Alfv\'enic waves in the solar corona. However, in the transition region, IRIS observations have demonstrated the cause of nonthermal broadening is the network jets \citep{2014Sci...346A.315T}. Furthermore, the temperature dependence of the nonthermal broadening \citep{1998ApJ...505..957C} is still not well understood. Could it be related to different properties of Alfv\'{e}n waves at different temperatures? Furthermore the nature of variation of the nonthermal broadening with height is shown to be dependent on temperature. For example, while the nonthermal line widths of Fe X (red, 6347\AA) emission line increases with height, the nonthermal line widths of Fe XIV (green, 5303\AA) emission line decreases with height in the solar atmosphere at the locations of both open and closed magnetic field regions \citep{2002PASJ...54..793S, 2013SoPh..282..427K}. Our understanding is quite limited in explaining this nature of variation. Simultaneous multi-temperature and high-resolution spectroscopic observations of the chromosphere, transition region and corona, such as observations that will be made by the SPICE spectrometer on board Solar Orbiter, the proposed EUVST spectrograph, and the proposed VELC spectrograph onboard ADITYA-L1 may be required to answer these questions.
There have been significant improvements in the numerical models and these have been somewhat successful in explaining a few observed properties of the solar corona. Most of these models assume a hot corona already present. Thus more efforts are needed to build a model that can self-consistently maintain the corona and reproduce the observed spectroscopic properties. Furthermore, the role of Alfv\'en(ic) waves in generating fine structures in the coronal plumes should be investigated in more details and compared with the observations.
While MHD models could be appropriate for modeling waves and heating in the lower coronal plasma, where collisions are frequent, in the heliospheric plasma the collision are rare and the plasma and kinetic instabilities become important, as evident from in-situ observations at 1AU by Wind and other spacecraft, by Helios to 0.3AU and recently from Parker Solar Probe (PSP) in the inner heliosphere. Therefore, while MHD models could be applied to model the inertial range of turbulence and long wavelength MHD waves, kinetic models must be used to model important small-scale and kinetic processes, such as resonant dissipation of turbulence cascade, temperature anisotropies and related plasma instabilities (such as ion-cyclotron, and firehose), as well as likely Landau damping of compressional wave modes. Thus, the future direction of wave studies in open coronal structures must extend the modeling into kinetic regime in order to capture the physics of solar wind kinetic plasma processes, evident in spacecraft observations.
\begin{acknowledgements}
The authors thank the anonymous referees for their helpful suggestions. We wish to acknowledge ISSI-BJ for hosting the workshop at Beijing and for their generous Support for all the participants.
SKP is grateful to FWO Vlaanderen for a senior postdoctoral fellowship (No. 12ZF420N).
VP was supported by the Spanish Ministerio de Ciencia, Innovaci\'on y Universidades through project PGC2018-102108-B-I00 and FEDER funds. VP and TVD are supported by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No 724326).
J.A.M. acknowledges UK Science and Technology Facilities Council (STFC) support from grant ST/T000384/1. L.O. acknowledges support by
NASA grants NNX16AF78G, 80NSSC18K1131 and NASA Cooperative Agreement
NNG11PL10A to CUA. P.A. acknowledges funding from his STFC Ernest Rutherford Fellowship (No. ST/R004285/2). IDM acknowledges support from the UK Science and Technology Facilities Council (consolidated grants ST/N000609/1 and ST/S000402/1), the European Union Horizon 2020 research and innovation programme (grant agreement No. 647214) and the Research Council of Norway through its Centres of Excellence scheme, project number 262622. H.T. is supported by NSFC grants 11825301 and 11790304(11790300).
T.J.W. acknowledges support by NASA grants 80NSSC18K1131, 80NSSC18K0668, and NASA Cooperative Agreement NNG11PL10A to CUA.
\end{acknowledgements}
\bibliographystyle{aasjournal}
| c6e57540afce35e624179b41d2e352171ed79b7f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:introduction}}
\IEEEPARstart{I}{ncreasing} availability and sophistication of digital image editing tools cause a major problem of that we even can not believe what we see \cite{liu2018image}. Image forgery is becoming a global epidemic which deeply affects our daily life for that some forgers use elaborately forged images to spread fake news or do other unscrupulous businesses \cite{wu2018busternet}. Copy-move forgery is a kind of image forgery in which one or several regions are pasted elsewhere in the same image in order to hide or duplicate objects of interest. Copy-move forgery detection techniques have always been a hot topic in image forensics \cite{ryu2010detection,christlein2012evaluation,li2015segmentation,cozzolino2015efficient}, and play important roles in cybersecurity and multimedia security \cite{qian2016separable,qiao2019adaptive}.
Conventional copy-move forgery detection methods adopt handcrafted features, and can be broadly divided into two categories, i.e., block-based approaches \cite{bashar2010exploring,ryu2013rotation,li2013image,li2013efficient,cozzolino2015efficient,mahmood2016copy,bi2018fast}, and keypoint-based approaches \cite{pan2010region,amerini2011sift,kakar2012exposing,pun2015image,li2015segmentation,ardizzone2015copy}. Their major difference is that block-based methods aim at exploring local features from abundant overlapping patches, while keypoint-based methods concentrate on patches of keypoints \cite{zandi2016iterative}. Nowadays, deep learning techniques have dominated various image processing tasks including image forensics \cite{cozzolino2017recasting,wu2017deep,liu2018image,cozzolino2018forensictransfer,zhou2018learning,cun2018image,cozzolino2019noiseprint,liu2019adversarial,yu2019attributing,mayer2019forensic,rossler2019faceforensics++}. Deep learning based copy-move forgery detection has also been investigated in \cite{wu2018image,wu2018busternet}. In \cite{wu2018image}, Wu et al. proposed an end-to-end deep neural network for predicting copy-move forgery masks. They construct a convolutional neural network for feature extraction, then compute self-correlation maps of convolutional features, and finally reconstruct forgery masks through a deconvolutional network. In \cite{wu2018busternet}, Wu et al. extended their network to a two-branch architecture: one branch localizes potential manipulation regions via visual inconsistencies; the other branch detects copy-move regions via visual similarities. According to the observations in \cite{korus2017multi,liu2018image,cozzolino2019noiseprint}, it is a very chanllenging task to localize forged regions in realistic forged images which barely have visual inconsistencies. As for the branch for detecting visual similarities, it tries to explore high-level low-resolution convolutional features \cite{simonyan2014very}, limiting the ability to detect accurate boundaries and small forged regions. Hence, we focus on digging deeper into visual similarity clues in our work.
\begin{figure*}[htp]
\centering
\includegraphics[width=15cm]{twostageframework.eps}
\caption{Overview of our two-stage copy-move forgery detection with self deep matching and Proposal SuperGlue. The first stage is a backbone self deep matching network based on atrous convolution, skip matching, and self-correlation with spatial attention. Red, blue, yellow blocks denote three convolutional blocks with the same scale, which are re-constructed by atrous convolution and skip matching. The second stage is named as Proposal SuperGlue. Proposal selection is conducted based on generated proposals and the backbone score map. Pairwise SuperGlue is conducted among candidate highly suspected proposals. Then, integrated score maps are generated by integrating matched keypoint scores and backbone scores. ConvCRF is constructed to refine integrated score maps.}
\label{fig:framework}
\end{figure*}
In this paper, we propose a novel two-stage copy-move forgery detection framework which integrates end-to-end deep matching with proposal based keypoint matching. The pipeline of this two-stage framework is shown in Fig.~\ref{fig:framework}.
\textbf{In the first stage}, a backbone self deep matching network is constructed to generate backbone score maps which indicate suspicious probabilities of pixels. Our backbone network integrates atrous convolution, skip matching, and spatial attention. Atrous convolution can increase the resolution of feature maps, and skip matching can invesitgate hierarchical information. Particularly, we discover the inherent connections between spatial attention and self-correlation, and propose a self-correlation module with spatial attention. Previous copy-move forgery detection methods \cite{wu2018image,wu2018busternet} only adopt VGG16 \cite{simonyan2014very}, we further study the feasibility of constructing deep matching based on deeper networks (ResNet50, ResNet101 \cite{he2016deep}) and light-weight networks (MobileNet \cite{howard2017mobilenets,sandler2018mobilenetv2,howard2019searching}, ShuffileNet \cite{zhang2018shufflenet,ma2018shufflenet}). The backbone network is regarded as a filter to efficiently detect suspected forged regions, while the results may inevitably contain false-alarmed regions or incomplete regions. Thus, we propose the second stage to remove false-alarmed regions and remedy incomplete regions.
\textbf{In the second stage}, a proposal based keypoint matching method is proposed and named as Proposal SuperGlue. Proposal SuperGlue mainly consists of two components: (1) A proposal selection module obtains several highly suspected proposals from a large number of bounding boxes provided by a proposal generation method \cite{pinheiro2015learning}. Proposal selection takes advantage of both backbone score maps and appearance clues, bridging the gap between deep matching and keypoint matching. (2) Proposal matching and label generation are devised to remove false alarms, remedy incomplete regions, and generate pixel labels from score maps. Deep-learning keypoint extraction (SuperPoint \cite{detone2018superpoint}) and matching (SuperGlue \cite{sarlin2020superglue}) are conducted among candidate proposals. An integrated score map generation method is designed to integrate keypoint matching results and backbone score maps. And an integrated score map refinement method is presented based on an improved fully connected CRF, i.e., ConvCRF (Convolutional Conditional Random Field) \cite{teichmann2018convolutional}.
Specifically, our main contributions of
this paper can be summarized as follows:
\begin{itemize}
\item An innovative two-stage copy-move forgery detection framework is proposed based on self deep matching and Proposal SuperGlue. We imaginatively integrate end-to-end deep matching with keypoint matching through highly suspected proposals.
\item A backbone self deep matching network is constructed based on atrous convolution, skip matching and spatial attention. Inherent connections between self-correlation and spatial attention are elaborately investigated.
\item Proposal SuperGlue, which incorporates proposal generation and deep-learning keypoint matching with a series of postprocessing procedures, is proposed to effectively remove false-alarmed regions and remedy incomplete regions.
\end{itemize}
The structure of this paper is as follows: In Section \ref{sec:rw},
we discuss related work. In Section \ref{sec:methodology}, we elaborate the
proposed framework. In Section \ref{sec:experiment}, experiments are conducted. In Section \ref{sec:conclusion}, we draw conclusions.
\section{Related Work}
\label{sec:rw}
In this section, we briefly review the state-of-the-art copy-move forgery detection methods, attention mechanism, proposal generation and local feature matching which are the key techniques researched in our work.
\textbf{Copy-move forgery detection}. Conventional copy-move forgery detection methods mainly consist of three components \cite{cozzolino2015efficient}: (1) feature extraction: extracting suitable features from pixels of interest; (2) matching: computing their best matching based on their associated features; (3) post-processing: processing and filtering vague detections to reduce false alarms. According to the formulations of feature extraction and subsequent matching schemes, these methods can be classified into two categories, i.e., block-based and keypoint-based methods. In block-based methods, a variety of features have been investigated for describing overlapping blocks and dense matching, e.g., DCT (Discrete Cosine Transform) \cite{mahmood2016copy}, DWT (Discrete Wavelet Transform) and KPCA (Kernel Principal Component Analysis) \cite{bashar2010exploring},
Zernike moments \cite{ryu2013rotation}, PCT (Polar Cosine Transform) \cite{yap2010two,li2013image}, PCET (Polar Complex Exponential Transform) \cite{bi2018fast}, LBP (Local Binary Patterns) \cite{li2013efficient}, Circular Harmonic Transforms (CHT) \cite{cozzolino2015efficient}. In keypoint-based methods, the commonly used features are SIFT (Scale Invariant Feature Transform) \cite{pan2010region,amerini2011sift,li2015segmentation,pun2015image} and SURF (Speeded-Up Robust Features) \cite{ardizzone2015copy,silva2015going}. Although great progress has been made in the study of copy-move forgery detection, it is still an unresolved challenging task for that duplicate regions may be small or smooth, and have gone through complicated rotation, resizing, compression and noise addition \cite{bi2018fast}. Besides, all the above conventional copy-move forgery detection methods rely on hand-crafted features and each module is optimized independently \cite{wu2018busternet}. Consequently, two kinds of end-to-end deep learning based copy-move forgery detection methods were proposed by Wu et al. in \cite{wu2018image,wu2018busternet}.
\textbf{Attention mechanism}. In \cite{sutskever2014sequence}, Sutskever et al. constructed a multi-layer long short term memory (LSTM) to map the input sequence to a fixed-length vector, and another deep LSTM to decode the target sequence from the vector. In \cite{bahdanau2014neural}, Bahdanau et al. adopted the attention mechanism to dynamically generate the vectors. Since then, the attention mechanism has been widely applied to solve sequential decision tasks \cite{lin2017structured}, and numerous attention-based models have been proposed \cite{xu2015show,luong2015effective,vaswani2017attention}. The attention mechanism can bias the allocation of available processing resources to the most informative components of input signals \cite{hu2018squeeze}, and has also been applied to solve multimedia problems, e.g., image classification \cite{hu2018squeeze,wang2017residual}, object detection \cite{woo2018cbam}, image super-resolution \cite{zhang2018image}, video classification \cite{wang2018non}. In these tasks, consistent improvements have been gained by adopting attention mechanisms to recalibrate informative convolutional features.
\textbf{Proposal generation}. In our work, we try to generate bounding boxes enclosing suspected regions. Proposal generation is a kind of technique that has been widely researched before the arrival of end-to-end object detection \cite{ren2015faster}. It aims to find out a set of (ranging from hundreds to thousands per image) proposal regions or bounding boxes which may contain objects \cite{liu2017listnet}. Since we can get hundreds of proposals which are near to contours in images, they cover suspected regions with high probability. Proposal generation approaches can be divided into two categories: conventional methods and deep-learning methods. Conventional methods leverage low-level grouping and saliency clues, e.g., objectness scoring \cite{alexe2012measuring,zitnick2014edge}, seed segmentation \cite{humayun2014rigor,krahenbuhl2014geodesic,krahenbuhl2015learning}, superpixel merging \cite{uijlings2013selective,pont2016multiscale}. Deep-learning approaches construct deep-network architectures to obtain proposals. For example, Deepbox \cite{kuo2015deepbox} learns a convolutional network to rerank proposals generated by EdgeBox \cite{zitnick2014edge}. Multibox \cite{erhan2014scalable} constructs a deep network to generate bounding box proposals. DeepMask \cite{pinheiro2015learning} and SharpMask \cite{pinheiro2016learning} can generate and refine segmentation proposals with high efficiency. These deep-learning approaches give us an opportunity to efficiently generate suspected boxes enclosing forged regions from a small set of candidate proposals.
\textbf{Local feature matching}. It mainly consists of five steps, (1) detecting interest points, (2) computing visual descriptors, (3) matching visual descriptors with a nearest neighbor (NN) search, (4) filtering incorrect matches, (5) estimating a geometric transformation \cite{sarlin2020superglue}. In recent years, researchers have been trying to learn better sparse detectors and local descriptors \cite{detone2018superpoint,dusmanu2019d2,DBLP:conf/nips/RevaudSHW19,ono2018lf,yi2016lift} from data using Convolutional Neural Networks (CNNs), and attempting to improve their discriminative ability by using various strategies, e.g., a wider context using regional features, log-polar patches, unsupervised learning. Although tremendous progress has been made in this field, these sets of matches are still estimated by NN search. In \cite{sarlin2020superglue}, a novel approach based on graph neural networks, i.e., SuperGlue, is proposed to establish pointwise correspondences from off-the-shelf local features: it acts as a middle-end between hand-crafted or learned front-end and back-end. SuperGlue outperforms other learned approaches and achieves state-of-the-art results on pose estimation. In keypoint-based copy-move forgery detection approaches \cite{pan2010region,amerini2011sift,kakar2012exposing,pun2015image,li2015segmentation,ardizzone2015copy}, hand-crafted local features and NN search have been widely researched. In our work, we try to integrate learning based detector, discriptor and matching into a unified copy-move forgery detection framework.
\section{Methodology}
\label{sec:methodology}
Our two-stage copy-move forgery detection framework consists of self deep matching and Proposal SuperGlue, as shown in Fig. \ref{fig:framework}. The first stage is a backbone self deep matching network, which generates score maps in an end-to-end manner. Firstly, we introduce the main architecture of our backbone network, including several alternative formulations in section \ref{sssec:FEAC}. Then, we introduce self-correlation with spatial attention in section \ref{sssec:SCSA}. The second stage is called Proposal SuperGlue, which is proposed to remove false-alarmed regions and remedy incomplete regions. In section \ref{sssec:psel}, we introduce our proposal selection strategy based on deep-learning proposal generation. In section \ref{sssec:PMLG}, we introduce proposal-based point matching, integrated score map generation and refinement.
\subsection{Self Deep Matching}
\label{ssec:sdmarch}
\subsubsection{Backbone Network Architecture}
\label{sssec:FEAC}
In our work, we adopt VGG16 as our basic feature extractor, remove pooling operatons in the fourth and fifth convolutional blocks \cite{simonyan2014very}, and adjust the fifth block by adopting atrous convolution to keep their original field-of-views. Atrous convolution can generalize standard convolution, adjust filter’s field-of-view and control the resolution of convolutional features \cite{chen2018deeplab,chen2018encoder,chen2017rethinking}. Let $\mathbf{y}(i_c,j_c)$ denote the output of the atrous convolution of a $2$-D input signal $\mathbf{x}(i_c,j_c)$, and the atrous convolution can be computed as:
\begin {equation}\label{eq:yij}
\mathbf{y}(i_c,j_c)=\sum_{k_1,k_2}\mathbf{w}(k_1,k_2)\times \mathbf{x}(i_c+r_\mathrm{ac} k_1,j_c+r_\mathrm{ac} k_2)
\end {equation}
where $k_1,k_2\in[-fl(\frac{K}{2}),fl(\frac{K}{2})]$ ($fl(\cdot)$ is a floor function), $\mathbf{w}(k_1,k_2)$ denotes a $K\times K$ filter, atrous rate $r_\mathrm{ac}$ determines the stride with which we sample the input
signal. In the fifth block of our basic architecture, atrous rate $r_\mathrm{ac}$ is set to $2$. Consequently, we generate three groups of larger feature maps with the same size, i.e., $\mathbf{F}_3$, $\mathbf{F}_4$ and $\mathbf{F}_5$ in Fig. \ref{Figure:ADMarch}. These hierarchical feature maps are all fed into our self-correlation module with spatial attention to compute the correlation maps. This kind of skip connections between multi-level feature maps and correlation computation is named as skip matching, and can effectively leverage rich hierarchical information provided by the feature extractor. And in the next section, we will introduce our self-correlation with spatial attention in detail.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.86\columnwidth]{SelfDMarchitecture.eps}
\caption{The architecture of backbone network with VGG16.}
\label{Figure:ADMarch}
\end{figure}
Based on the computed correlation maps, we construct an Atrous Spatial
Pyramid Pooling (ASPP) module to capture their multiscale information. As shown in Fig.~\ref{Figure:ADMarch}, we construct $3$ parallel atrous convolutional layers with $3\times 3$ filters and atrous rates of $\{6,12,18\}$. Besides, we construct a convolutional layer with $1\times 1$ filters, and a global average pooling layer followed by a convolutional layer with $1\times 1$ filters to capture local features and image-level features respectively. All these convotional layers output $48$-channel feature maps, and the five groups of feature maps are concatenated and fed into subsequent layers which are constituted of convolutional and upsampling layers.
\textbf{Alternative formulations}. Previous end-to-end copy-move forgery detection approaches \cite{wu2018image,wu2018busternet} adopt VGG16 for feature extraction, and we do not know the performance of deeper or light-weight networks. Thus, we formulate the popular ResNet50 and ResNet101 \cite{he2016deep} as deeper feature extractor, decrease the strides of fourth and fifth convolutional blocks, and set the atrous rates as $2$ and $4$ respectively. Light-weight networks are specifically tailored for mobile and resource constrained environments \cite{sandler2018mobilenetv2}. We reformulate three popular and competitive light-weight networks, i.e., MobileNetV2 \cite{sandler2018mobilenetv2}, MobileNetV3 \cite{howard2019searching} and ShuffleNetV2 \cite{ma2018shufflenet}. We enlarge feature maps of the last two convolutional blocks by decreasing strides and adopting atrous convolution. In these formulations, we still can get $3$ sets of feature maps with the same size. And these features are fed into correlation layers and subsequent score map generation layers.
\subsubsection{Self-Correlation with Spatial Attention}
\label{sssec:SCSA}
In this section, we detailedly introduce the proposed self-correlation with spatial attention, and discuss the conections between self-correlation and spatial attention. Let $\mathbf{F}_{l}$ denote the $l$-th block feature maps, and $\mathbf{F}_{l}(i,j)$ denotes a $c$-dimensional descriptor at $(i,j)$. Note that $\mathbf{F}_{l}\in\mathbb{R}^{h\times w \times c}$, $i\in[1,h]$, $j\in[1,w]$, $h$ and $w$ indicate the height and width of the feature map, and $h=w$ in our work. Before the attention and correlation computation, L2-normalization is conducted, $\bar{\mathbf{F}}_{l}(i,j)=\mathrm{L2\_norm}(\mathbf{F}_{l}(i,j))={\mathbf{F}_{l}(i,j)}/{||\mathbf{F}_{l}(i,j)||_2}$. By adopting L2-normalization, we can restrict the value ranges of descriptors, and obtain normalized feature maps, i.e., $\bar{\mathbf{F}}_{l}$.
Spatial attention is a kind of self-attention module which calculates response at a position as a weighted sum of the features at all positions \cite{vaswani2017attention}. It can capture long-range dependences and allocate attention according to
similarity of color and texture. In our work, spatial attention is constructed to reinforce $\bar{\mathbf{F}}_{l}$. $\bar{\mathbf{F}}_{l}$ is first transformed into two feature spaces $\bm{f}(\bar{\mathbf{F}}_{l})=\bar{\mathbf{F}}_{l}\bm{W}_{l,\bm{f}}+\bm{b}_{l,\bm{f}}$ and $\bm{g}(\bar{\mathbf{F}}_{l})=\bar{\mathbf{F}}_{l}\bm{W}_{l,\bm{g}}+\bm{b}_{l,\bm{g}}$. Then we compute:
\begin {equation}\label{eq:spatialbeta}
\beta_l^{(m,n)}=\frac{\mathrm{exp}(s_l^{(m,n)})}{\sum_n\mathrm{exp}(s_l^{(m,n)})}
\end {equation}
where
\begin {equation}\label{eq:spatialbetasmn}
s_l^{(m,n)}= \bm{f}(\bar{\mathbf{F}}^{(m)}_{l})^T\bm{g}(\bar{\mathbf{F}}^{(n)}_{l})
\end {equation}
$\beta_l^{(m,n)}$ indicates the extent to which the model attends to the $n$-th location when predicting the $m$-th region, $m,n\in[1,h\times w]$. The output of the attention block is computed as:
\begin {equation}\label{eq:spatialatten}
\mathbf{o}_l^{(m)}=\sum_n\beta_{mn}\bm{h}(\bar{\mathbf{F}}^{(n)}_{l})
\end {equation}
where $\bm{h}(\bar{\mathbf{F}}_{l})=\bar{\mathbf{F}}_{l}\bm{W}_{l,\bm{h}}+\bm{b}_{l,\bm{h}}$. Note that $\bm{W}_{l,\bm{f}}\in\mathbb{R}^{c \times \frac{c}{8}}$, $\bm{W}_{l,\bm{g}}\in\mathbb{R}^{c \times \frac{c}{8}}$, $\bm{W}_{l,\bm{h}}\in\mathbb{R}^{c \times c}$, $\bm{b}_{l,\bm{f}}\in \mathbb{R}^{\frac{c}{8}}$, $\bm{b}_{g,\bm{h}}\in \mathbb{R}^{\frac{c}{8}}$, $\bm{b}_{l,\bm{h}}\in \mathbb{R}^{c}$, which are implemented as $1\times 1$ convolutional layers. $\mathbf{O}_l=\{\mathbf{o}_l^{(1)},\mathbf{o}_l^{(2)},\cdots,\mathbf{o}_l^{(h\times w)}\}$ are the attention values for $\bar{\mathbf{F}}_l$. Consequently, the spatial attention reinforced convolutional feature maps can be computed as:
\begin {equation}\label{eq:sattwefea}
\ddot{\mathbf{F}}_l=\mathrm{Atten}_l(\bar{\mathbf{F}}_l)=\lambda_l\mathbf{O}_l+\bar{\mathbf{F}}_l
\end {equation}
where $\lambda_l$ is a scale parameter which is initialized as $0$ and gradually learned to assign a proper value.
Self-correlation aims to compute the similarity between every two locations in the convolutional feature maps. Scalar product is commonly used:
\begin {equation}\label{eq:selfcorr}
c_l^{(m,n)}=(\ddot{\mathbf{F}}^{(m)}_{l})^T \ddot{\mathbf{F}}^{(n)}_{l}
\end {equation}
Thus, we can get a raw correlation map tensor $\mathbf{C}_l=\{c_l^{(m,n)}|m,n\in[1,h\times w]\}\in \mathbb{R}^{h\times w \times (h\times w)}$. In fact, only a small fraction of features has close relations, and the majority of features are dissimilar. This indicates that a subset of $\mathbf{C}_l$ contains sufficient information to decide which feature is matched. Consequently, $\mathbf{C}_l$ is sorted along the $(h\times w)$ channels, and top-$T$ values are selected:
\begin {equation}\label{eq:sortpool}
\tilde{\mathbf{C}}_l(i,j,1:T)=\mathrm{Top\_T}(\mathrm{Sort}(\mathbf{C}_l(i,j,:)))
\end {equation}
A monotonic decreasing curve with an abrupt drop at some point should be observed along the $T$ channels, as long as $\tilde{\mathbf{C}}_l(i,j)$ has matched regions. Thus, the $T$ channels should cover the most drops, and the selection of $T$ is discussed in experiments. In theory, our network can process arbitrary-sized images since the adoption of the top-$T$ selection, unless $h\times w<T$. Moreover, zero-out and normalization operations are conducted on $\tilde{\mathbf{C}}_l$ to limit correlation values to certain ranges and filter redundant values:
\begin {equation}\label{eq:relunorm}
\bar{\mathbf{C}}_l=\mathrm{L2\_norm}(\mathrm{Max}(\tilde{\mathbf{C}}_l,0))
\end {equation}
Since we get three groups of feature maps with the same size from feature extractor, i.e., $l\in\{3,4,5\}$, we can get three groups of correlation maps, i.e., $\bar{\mathbf{C}}_3$, $\bar{\mathbf{C}}_4$ and $\bar{\mathbf{C}}_5$. Note that the parameters of spatial attention are not shared for the computation of these three groups of correlation maps. Then, we concatenate the three groups of correlation maps, and get a correlation map tensor $\widehat{\mathbf{C}}=\mathrm{Concat}(\bar{\mathbf{C}}_3,\bar{\mathbf{C}}_4,\bar{\mathbf{C}}_5)$, where $\widehat{\mathbf{C}}\in\mathbb{R}^{h \times w \times 3T}$. Since $\widehat{\mathbf{C}}$ is computed from three groups of hierarchical feature maps, it contains rich correlation relations from coarse to fine.
\textbf{Inherent connections between self-correlation and spatial attention.} Both self-correlation and spatial attention attempt to explore correlations between every pair of features in a feature map tensor. Spatial attention conducts a scalar product in transformed spaces as Eq. (\ref{eq:spatialbetasmn}), while self-correlation conducts a scalar product directly on feature maps as Eq. (\ref{eq:selfcorr}). Essentially, they have the same target of finding the close related regions. Spatial attention can allocate attention according to similarity of features, driving correlated regions to have closer feature distributions. Inspired by this inherent connection, we construct spatial attention before self-correlation computation. Consequently, it can reinforce the subsequent self-correlation computation.
Additionally, we have also investigated multi-head spatial attention. Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions \cite{vaswani2017attention}. In fact, it constructs several parallel spatial attention blocks. Furthermore, we also attempt to add channel attention (SE blocks \cite{hu2018squeeze}) before or after spatial attention. Channel attention can highlight channel-wise informative
features, and a weighted scalar
product can be conducted in the correlation computation
procedure. However, they can not achieve better performance with additional parameters which will be discussed in experiments.
\subsection{Proposal SuperGlue}
\label{ssec:psg}
Proposal SuperGlue can be broadly divided into two steps. \textit{In the first step}, a proposal selection method is proposed to obtain highly suspected regions from hundreds of bounding-box proposals. These bounding-box proposals are generated by a proposal generation method which exploits image appearance features to find bounding boxes near to contours. \textit{In the second step}, we devise proposal-based keypoint matching with elaborately designed postprocessing procedures. Firstly, pairwise deep-learning keypoint matching is conducted among candidate proposals. Then, we propose an integrated score map generation method to integrate both self deep matching and keypoint matching results, so that some false-alarmed regions can be removed and incomplete regions can be complemented. Finally, in order to get good score distributions and accurate boundaries, ConvCRF is constructed to refine integrated score maps according to integrated scores and associated apperance similarity in the image. \textit{The first step} is introduced in section \ref{sssec:psel}, and \textit{the second step} is discussed in section \ref{sssec:PMLG}.
\subsubsection{Proposal Selection}
\label{sssec:psel}
Processed by our backbone self deep matching network, we can get a score map $\mathbf{S}\in\mathbb{R}^{h_I\times w_I}$, $h_I$ and $w_I$ denote the size of the score map which is the same as the size of input image $\mathbf{I}$. $\mathbf{S}(i_I,j_I)\in[0,1]$, and $i_I\in [1,h_I],j_I\in [1,w_I]$. Since the feature maps of the backbone network have lower resolutions than original images, and the self-correlation is built on primitive scalar product instead of intricate similarity computing, $\mathbf{S}$ may have false-alarmed or incomplete regions. In order to get rid of false-alarmed regions and complement incomplete regions, a proposal selection strategy is proposed to obtain highly suspected boxes for further matching.
Our motivation is that $\mathbf{S}$ always has some isolated meaningless regions in some complicated images according to our observation. Whether we can enclose meaningful regions while ignore meaningless regions may affect the performance. However, how can we obtain several well enclosed bounding boxes based on score map $\mathbf{S}$ and input image $\mathbf{I}$? After all, there are too many bounding boxes can be generated from a single image, e.g., different scales, aspect ratios and positions. In fact, the majority of copy-move forged regions have clear contours, and might be possible to be covered by genereted proposals which are relied on edges or saliency features \cite{liu2017listnet}. Thus, we propose to conduct proposal generation on the input image, and select several high-quality boxes from hundreds of proposals. Our proposal selection strategy relies on score map $\mathbf{S}$, and consists of selecting and merging operations. We assume that there is a proposal generation function $\mathcal{P}(\cdot)$ with image $\mathbf{I}$ as input, we can get $P$ proposals $\mathbf{P}=\{\bm{p}_p|p\in[1,P]\}$. $\bm{p}_p=\{(x^p_1,y^p_1),(x^p_2,x^p_2)\}$ contains the coordinates of top left and bottom right corners. With score map $\mathbf{S}$ at hand, we can get the average score in proposal $\bm{p}_p$, i.e., $s_p = \mathnormal{f}_{\mathrm{avgs}}(\mathbf{S},\bm{p}_p)$. According to $s_p$ and its relations with other proposals, we can obtain the final proposals. Our proposal selection strategy can be summarized as Algorithm~\ref{pssalg}.
\begin{algorithm}[htp]
\caption{Proposal selection strategy.}
\label{pssalg}
\hspace*{0.02in} {\bf Input:} Image
$\mathbf{I}$ and score map $\mathbf{S}$\\
\hspace*{0.02in} {\bf Output:} Selected proposals $\mathbf{P}_{s}$
\begin{algorithmic}[1]
\State $\mathbf{P}=\mathcal{P}(\mathbf{I})$;
\State $\mathbf{P}_{t}=\{\}$;
\For{$p=1$ to $P$}
\State $s_p = \mathnormal{f}_{\mathrm{avgs}}(\mathbf{S},\bm{p}_p)$;
\If{$s_p > s_{t}$}
\State $flag=1$;
\For{$i_{ps}=1$ to $\mathrm{len}(\mathbf{P}_{t})$}
\State $v_{iou}=\mathrm{IoU}(\mathbf{P}_{t}(i_{ps}),\bm{p}_p)$;
\State $v_{inter}=\mathrm{Inter}(\mathbf{P}_{t}(i_{ps}),\bm{p}_p)$;
\If{$v_{iou}>iou_{t}$}
\If {$s_p>\mathnormal{f}_{\mathrm{avgs}}(\mathbf{S},\mathbf{P}_{t}(i_{ps}))$}
\State $\mathbf{P}_{t}(i_{ps})=\bm{p}_p$; $flag=0$; Break;
\EndIf
\EndIf
\If {$v_{inter}/\mathrm{Size}(\bm{p}_p)>inter_t$ or \\\hspace*{0.8in}$v_{inter}/\mathrm{Size}(\mathbf{P}_{t}(i_{ps}))>inter_t$}
\State $\bm{p}_{m}=\mathrm{Merge}(\bm{p}_p,\mathbf{P}_{t}(i_{ps}))$;
\If {$\mathnormal{f}_{\mathrm{avgs}}(\mathbf{S},\bm{p}_{m})>s_{t}$}
\State $\mathbf{P}_{t}(i_{ps})=\bm{p}_{m}$; $flag=0$; Break;
\EndIf
\EndIf
\EndFor
\If{$flag=1$}
\State $\mathbf{P}_{t}=\mathbf{P}_{t}\cup\bm{p}_p$;
\EndIf
\EndIf
\EndFor
\State $\mathbf{P}_{s}=\{\}$;
\While{$\mathrm{len}(\mathbf{P}_{s})\ne \mathrm{len}(\mathbf{P}_{t})$}
\If {$\mathbf{P}_{s}\ne\oslash$}
\State $\mathbf{P}_{t}=\mathbf{P}_{s}$;
\EndIf
\State $\mathbf{P}_{s}=\{\}$;
\For{$i_{ps1}=1$ to $\mathrm{len}(\mathbf{P}_{t})$}
\For{$i_{ps2}=i_{ps1}+1$ to $\mathrm{len}(\mathbf{P}_{t})$}
\State $v_{inter}=\mathrm{Inter}(\mathbf{P}_{t}(i_{ps1}),\mathbf{P}_{t}(i_{ps2}))$;
\If {$v_{inter}/\mathrm{Size}(\mathbf{P}_{t}(i_{ps1}))>inter_t$ or\\\hspace*{0.6in} $v_{inter}/\mathrm{Size}(\mathbf{P}_{t}(i_{ps2}))>inter_t$}
\State $\bm{p}_{m}=\mathrm{Merge}(\mathbf{P}_{t}(i_{ps1}),\mathbf{P}_{t}(i_{ps2}))$;
\If {$\mathnormal{f}_{\mathrm{avgs}}(\mathbf{S},\bm{p}_{m})>s_{t}$}
\State $\mathbf{P}_{s}=\mathbf{P}_{s}\cup\bm{p}_{m}$;
\Else
\State Insert $\mathbf{P}_{t}(i_{ps1})$ or $\mathbf{P}_{t}(i_{ps2})$ \\\hspace*{1.0in}with higher intersection rate;
\EndIf
\EndIf
\EndFor
\EndFor
\EndWhile
\end{algorithmic}
\end{algorithm}
In Algorithm~\ref{pssalg}, there are some basic functions: $\mathrm{len}(\cdot)$ returns the item number of input list, $\mathrm{IoU}(\cdot,\cdot)$ computes the Intersection over Union (IoU) \cite{liu2017listnet} of two boxes, $\mathrm{Inter}(\cdot,\cdot)$ computes their intersection, and $\mathrm{Size}(\cdot)$ indicates the size of the input box. $\mathrm{Merge}(\cdot,\cdot)$ is used to merge two input boxes, in other words, it generates the smallest box which can cover the two input boxes. The proposal generation function $\mathcal{P}(\cdot)$ is implemented based on DeepMask \cite{pinheiro2015learning}. The basic idea of Algorithm~\ref{pssalg} is that we try to reject proposals with small average scores, select proposals with higher scores from proposals which have high IoU with each other, and merge proposals or select larger boxes when they have large intersection rates. And there are some parameters need to set, proposal threshold score $s_{t}=0.4$, threshold IoU $iou_t=0.5$, threshold intersection rate $inter_t=0.8$. Besides, all proposals or merged boxes should meet the basic requirement that they should be smaller than the half of the input image. Last but not least, iterations are conducted to avoid merged boxes with large intersection rates. By using Algorithm~\ref{pssalg}, we can get $\tilde{P}$ (generally less than $10$) high-quality proposals $\mathbf{P}_{s}=\{\bm{p}_{\tilde{p}}|\tilde{p}\in[1,\tilde{P}]\}$ ($\mathbf{P}_{s}(\tilde{p})$ indicates $\bm{p}_{\tilde{p}}$ in $\mathbf{P}_{s}$).
\subsubsection{Keypoint Matching and Label Generation}
\label{sssec:PMLG}
\textbf{Proposal-based keypoint matching}. With high-qualtiy proposals $\mathbf{P}_{s}$ at hand, we can extract interest points from them and conduct keypoint matching. As we discussed in Section \ref{sec:rw}, CNN-based interest point detection and discription show a good prospect in numerous applications. Thus, we extract keypoints with corresponding descriptors from each proposal using SuperPoint \cite{detone2018superpoint}. It is a fully-convolutional model which operates on full-sized
images and jointly computes pixel-level interest point locations and associated descriptors in one forward pass. It can be denoted as $\mathbf{K}_{\tilde{p}}=\mathrm{SuperPoint}(\bm{p}_{\tilde{p}},\mathbf{I})$, where $\mathbf{K}_{\tilde{p}}$ denotes the extracted keypoints set from proposal $\bm{p}_{\tilde{p}}$, and the $k_p$-th elements is $\mathbf{K}_{\tilde{p}}(k_p)=\{(x_{k_p},y_{k_p}),\mathbf{d}_{k_p}\}$, $\mathbf{d}_{k_p}$ denotes the corresponding descriptor. Then for each pair of point sets $\mathbf{K}_{\tilde{p}}(k_{p1})$ and $\mathbf{K}_{\tilde{p}}(k_{p2})$, we conduct SuperGlue \cite{sarlin2020superglue} to get matched points and corresponding matching scores $\mathbf{M}_{\tilde{p}1},\mathbf{M}_{\tilde{p}2}=\mathrm{SuperGlue}(\mathbf{K}_{\tilde{p}1},\mathbf{K}_{\tilde{p}2})$. SuperGlue uses a graph neural network and attention to solve an assignment optimization problem. Instead of learning better task-agnostic local features followed by simple matching heuristics and tricks, SuperGlue learns the matching process from pre-existing local features using a novel neural architecture for the first time. It matches two sets of local features by jointly finding correspondences and rejecting non-matchable points. Finally, we can get $M$ matched points and their matching scores $\mathbf{M}=\{\mathbf{M}_{\tilde{p}}|\tilde{p}\in[1,\tilde{P}]\}=\{(x_m,y_m),s_m|m\in[1,M]\}$.
\textbf{Integrated score map generation}. In order to map the matching scores to each pixel, we adopt a superpixel algorithm, i.e., SEEDS \cite{van2012seeds,van2015seeds}. By conducting superpixel segmentation, we get the superpixel labels $\mathbf{L}_{sp} = \mathrm{SuperPixel}(\mathbf{I})$. Let $\mathbf{L}_{sp}(x_m,y_m)$ denote pixels whose superpixel labels are the same as $(x_m,y_m)$. We set scores of these pixels the same as the score of matched point $(x_m,y_m)$, i.e., $\mathbf{S}_{sp}(\mathbf{L}_{sp}(x_m,y_m))=s_m$. Thus, we get our pixel-level scores $\mathbf{S}_{sp}$ from superpixel and matched points. Besides, we also generate a pixel-level score map $\mathbf{S}_{p}$ from backbone scores $\mathbf{S}$ and candidate proposals which have matched points. Concretely, we set $\mathbf{S}_{p}(x,y)=\mathbf{S}(x,y)$ for $(x,y)$ in the scope of $\bm{p}_{\tilde{p}}$ which contains matched points, otherwise $\mathbf{S}_{p}(x,y)=0$. Thus, our integrated score map $\mathbf{S}_{in}$ is computed based on $\mathbf{S}_{sp}$ and $\mathbf{S}_{p}$ as follows:
\begin {equation}\label{eq:finalscore}
\mathbf{S}_{in}=\frac{1}{1+\mathrm{exp}(-\phi(\alpha\cdot\mathbf{S}_{sp}+\beta\cdot\mathbf{S}_{p}+\gamma))}
\end {equation}
where $\alpha$, $\beta$ and $\gamma$ are three parameters to balence $\mathbf{S}_{sp}$ and $\mathbf{S}_{p}$. We set $\alpha=\beta=1$ to make $\mathbf{S}_{sp}$ and $\mathbf{S}_{p}$ have the same contribution. We set $\gamma=-0.5$ to make sure it has the same distribution when $\mathbf{S}_{sp}(i)=0$. $\phi$ indicates the amplifying factor to control score distribution of $\mathbf{S}_{in}$, and is set to $4$.
\textbf{Integrated score map refinement for label generation}. The directly computed $\mathbf{S}_{in}$ has some small isolated regions or holes inside detected regions, because there are some false-alarmed or missing-detected regions. In order to neglect regions with lower matching probability and refine contours according to image content, we formulate fully connected CRF (Conditional Random Field) \cite{krahenbuhl2011efficient} based on $\mathbf{S}_{in}$ and image $\mathbf{I}$, to get final labels. Our problem is that we have an image $\mathbf{I}$ which has $N$ pixels, and we try to fulfill a segmentation task with two classes. A segmentation of $\mathbf{I}$ is modelled as a random field $\mathbf{X}=\{X_1,\cdots,X_N\}$ where each random variable $X_n$ takes values of $\{0,1\}$. ``$1$" is used to label forged locations and corresponding genuine ones, while ``$0$" is for remaining parts. A conditional random field $(\mathbf{I}, \mathbf{X})$ is characterized by a Gibbs distribution $P(\mathbf{X}|\mathbf{I})=\frac{1}{Z(\mathbf{I})}\mathrm{exp}(-E(\mathbf{X}|\mathbf{I}))$, where the energy function $E(\mathbf{X}|\mathbf{I})$ is given by:
\begin {equation}\label{eq:crfenergy}
E(\mathbf{X}|\mathbf{I})=\sum_{i\le N}{\psi_u(X_i|\mathbf{I})} + \sum_{i\ne j\le N}{\psi_p(X_i,X_j|\mathbf{I})}
\end {equation}
where $\psi_u(X_i|\mathbf{I})$ is called unary potential. In our work, our computed $\mathbf{S}_{in}$ is treated as the unary potential:
\begin {equation}\label{eq:unarypotential}
\psi_u(X_i|\mathbf{I})=\mathbf{S}_{in}(i)
\end {equation}
And $\psi_p(X_i,X_j|\mathbf{I})$ is called pairwise potential. It accounts for the joint distribution of pixels $i$ and $j$. It allows us to explicitly model interactions between pixels, such as pixels with similar colour are
likely the same class. And $\psi_p$ is formulated as weighted sum of Gaussian kernels:
\begin {equation}\label{eq:pairpotential}
\psi_p(X_i,X_j|\mathbf{I})=\mu(X_i,X_j)k(\mathbf{f}_i,\mathbf{f}_j)
\end {equation}
where $\mu(X_i,X_j)$ is a simple label compatibility function, which is given by the Potts model $\mu(X_i,X_j)=[X_i\ne X_j]$. It
penalizes nearby similar pixels that are assigned different labels. $k(\mathbf{f}_i,\mathbf{f}_j)$ denotes Gaussian kernels with feature vectors $\mathbf{f}_i$ and $\mathbf{f}_j$ in an arbitrary feature space. Specifically, contrast-sensitive two-kernel potentials are formulated in our model:
\begin {equation}\label{eq:gaussiankernel}
\begin{aligned}
k(\mathbf{f}_i,\mathbf{f}_j)=&\underbrace{w^{(1)}\mathrm{exp}(-\frac{|\mathrm{p}_i-\mathrm{p}_j|^2}{2\theta_{\alpha}^2}-\frac{|\mathrm{I}_i-\mathrm{I}_j|^2}{2\theta_{\beta}^2})}_{\text{appearance kernel}}\\&+\underbrace{w^{(2)}\mathrm{exp}(-\frac{|\mathrm{p}_i-\mathrm{p}_j|^2}{2\theta_{\gamma}^2})}_{\text{smoothness kernel}}
\end{aligned}
\end {equation}
where $\mathrm{I}_i$ and $\mathrm{I}_j$ are color vectors, $\mathrm{p}_i$ and $\mathrm{p}_j$ are positions. $w^{(1)}$ and $w^{(2)}$ are linear combination weights. $\theta_{\alpha}$, $\theta_{\beta}$ and $\theta_{\gamma}$ are controlling parameters. The appearance kernel drives nearby pixels with similar color to be in the same class. The smoothness kernel removes small isolated regions.
In our implementation, we adopt ConvCRF \cite{teichmann2018convolutional} for inference. ConvCRF adds the assumptions of conditional independence fully-connected CRF, and reformulates the inference in terms of convolutions which are implemented highly efficiently on GPUs.
\section{Experimental Evaluation}
\label{sec:experiment}
\subsection{Implementation Details}
\label{ssec:impledetails}
Real-world copy-move forgery needs forgers to manually manipulate images and pasted regions. Therefore, the available copy-move forgery datasets are not sufficient for training an end-to-end deep matching network. Thus, we automatically generate a synthetic training set and a synthetic testing set from MS COCO 2014 training images and testing images respectively. For each image, we resize it to $512\times 512$, randomly select one annotated region under different transformations, and paste it to a random position of this image. All pasted regions randomly suffer four types of transformations, i.e., rotation changes in $\mathbb{U}(-60,60)$, scale changes in $\mathbb{U}(0.5,4)$, luminance changes in $\mathbb{U}(-32,32)$, and deformation changes in $\mathbb{U}(0.5,2)$ (decrease or increase the width of a tampered region). Following this strategy, we generate $120,000$ training images and $1,000$ testing images.
The self deep matching network is trained with a single spatial cross
entropy loss, and parameters in the basic feature extraction network are initialized using VGG16 \cite{simonyan2014very} which
is trained for image classification. Similarly, alternative formulations are initilized using corresponding classification networks. We conduct $16$-epoch training, and adopt the Adadelta optimizer \cite{zeiler2012adadelta}. The input images are randomly resized in the range of $[256\times 256,512\times 512]$. Limited by our GPU memory, the batch size is set to $6$ (a larger batch size may further improve our performance). As for the Proposal SuperGlue stage, no further training is needed. We directly adopt their trained DeepMask \cite{pinheiro2015learning}, SuperPoint \cite{detone2018superpoint}, SuperGlue \cite{sarlin2020superglue} models.
\subsection{Backbone Network Ablation Study}
\label{ssec:ablation}
\begin{table*}[htp]
\renewcommand{\arraystretch}{1.3}
\caption{Step-by-step analyses on the synthetic testing set.}
\label{table:combstep}
\centering
\footnotesize
\setlength{\tabcolsep}{1.6mm}{
\begin{tabular}{c | c c c c | c c c c | c c }
\hline
\multirow{2}{*}{Variant} & \multicolumn{4}{c|}{Protocol-All} & \multicolumn{4}{c|}{Protocol-Detected} & \multirow{2}{*}{$T$} & \multirow{2}{*}{Trainable params} \\\cline{2-9}
& IoU & Precision & Recall & F1-score & IoU & Precision & Recall & F1-score & & \\
\hline
encoder-decoder & 0.4982 & 0.5930 & 0.7905 & 0.6328 & 0.4992 & 0.5942 & 0.7921 & 0.6341 & 48 & 7,772,209 \\
encoder-decoder-skip & 0.5523 & 0.6523 & 0.7927 & 0.6835 & 0.5529 & 0.6530 & 0.7935 & 0.6841 & 48 & 14,985,265 \\
encoder-decoder-skip-normfeas & 0.6822 & 0.7793 & 0.8332 & 0.7897 & 0.6822 & 0.7793 & 0.8332 & 0.7897 & 48 & 14,985,265\\
SelfDM & 0.6959 & 0.7813 & 0.8506 & 0.7999 & 0.6959 & 0.7813 & 0.8506 & 0.7999 & 48 & 14,985,265 \\
\hline
SelfDM-16 & 0.6818 & 0.7703 & 0.8427 & 0.7891 & 0.6832 & 0.7719 & 0.8444 & 0.7907 & 16 & 14,851,633 \\
SelfDM-32 & 0.6881 & 0.7899 & 0.8255 & 0.7927 & 0.6881 & 0.7899 & 0.8255 & 0.7927 & 32 & 14,918,449 \\
SelfDM-48 & 0.6959 & 0.7813 & 0.8506 & 0.7999 & 0.6959 & 0.7813 & 0.8506 & 0.7999 & 48 & 14,985,265 \\
SelfDM-64 & 0.6942 & 0.8038 & 0.8221 & 0.7970 & 0.6949 & 0.8046 & 0.8229 & 0.7978 & 64 & 15,052,081 \\
\hline
SelfDM-SA & 0.7233 & 0.8458 & 0.8227 & 0.8216 & 0.7240 & 0.8467 & 0.8235 & 0.8225 & 48 & 15,724,148 \\
SelfDM-MSA & 0.7119 & 0.8466 & 0.8064 & 0.8096 & 0.7126 & 0.8475 & 0.8072 & 0.8104 & 48 & 17,940,797 \\
SelfDM-SACA & 0.7129 & 0.8370 & 0.8204 & 0.8136 & 0.7129 & 0.8370 & 0.8204 & 0.8136 & 48 & 15,799,236 \\
SelfDM-CASA & 0.7195 & 0.8472 & 0.8164 & 0.8182 & 0.7195 & 0.8472 & 0.8164 & 0.8182 & 48 & 15,799,236 \\
\hline
\end{tabular}}
\end{table*}
\begin{table*}[htp]
\renewcommand{\arraystretch}{1.3}
\caption{Feature extractor comparisons on the synthetic testing set.}
\label{table:combmodel}
\centering
\footnotesize
\setlength{\tabcolsep}{2.35mm}{
\begin{tabular}{c | c c c c | c c c c | c }
\hline
\multirow{2}{*}{Variant} & \multicolumn{4}{c|}{Protocol-All} & \multicolumn{4}{c|}{Protocol-Detected} & \multirow{2}{*}{Trainable params} \\\cline{2-9}
& IoU & Precision & Recall & F1-score & IoU & Precision & Recall & F1-score & \\
\hline
SelfDM-SA & 0.7233 & 0.8458 & 0.8227 & 0.8216 & 0.7240 & 0.8467 & 0.8235 & 0.8225 & 15,724,148 \\
SelfDM-SA-ResNet50 & 0.7372 & 0.7809 & 0.9191 & 0.8312 & 0.7372 & 0.7809 & 0.9191 & 0.8312 & 30,664,372 \\
SelfDM-SA-ResNet101 & 0.7176 & 0.8246 & 0.8359 & 0.8059 & 0.7183 & 0.8254 & 0.8368 & 0.8067 & 49,656,500 \\
SelfDM-SA-MobileNetV2 & 0.7126 & 0.7957 & 0.8584 & 0.8118 & 0.7126 & 0.7957 & 0.8584 & 0.8118 & 4,557,012 \\
SelfDM-SA-MobileNetV3 & 0.7512 & 0.8575 & 0.8467 & 0.8412 & 0.7512 & 0.8575 & 0.8467 & 0.8412 & 4,092,268 \\
SelfDM-SA-ShuffleNetV2 & 0.6676 & 0.7692 & 0.8137 & 0.7745 & 0.6676 & 0.7692 & 0.8137 & 0.7745 & 2,920,602 \\
\hline
\end{tabular}}
\end{table*}
Our backbone self deep matching network incorporates encoder-decoder architecture with atrous convolution, skip matching, feature normalization, correlation normalization and spatial attention. In TABLE~\ref{table:combstep}, step-by-step analyses are provided on the synthetic testing set. We compute the pixel-level IoU, precision, recall and F1-score for each image, and caculate their average scores. Two protocols are adopted: ``Protocol-All" means we compute the average scores of all evaluated images, and ``Protocol-Detected" only computes average scores of detected images.
It shows that each component of our backbone network plays an important role in improving its localization performance. Specifically, ``encoder-decoder" denotes a pure architecture with a feature extractor and a decoder, there is no skips, normalization and attention; In ``encoder-decoder-skip", we add skip matching; In ``encoder-decoder-skip-normfeas", input feature maps before correlation computation are normalized; In ``SelfDM", correlation maps are followed by ReLU and L2-normalization. Besides, we make a discussion on the selection of $T$ in Eq.~(\ref{eq:sortpool}). We find that it can even achieve comparable performance when $T=16$. When we set $T=48$, it can get higher scores. There is no further improvement with $T=64$. More importantly, we test four types of attention-based self-correlation formulations, i.e., ``SelfDM-SA" with spatial attention, ``SelfDM-MSA" with multi-head spatial attention \cite{vaswani2017attention}, ``SelfDM-SACA" with spatial attention before channel attention and ``SelfDM-CASA" with spatial attention after channel attention \cite{hu2018squeeze}. Although, ``SelfDM-MSA", ``SelfDM-SACA" and ``SelfDM-CASA" have more parameters, their performance is barely satisfactory. We guess the main reasons are that: (1) single spatial attention can already reinforce correlated regions, while multiple spatial attention with redundant information may mislead our model; (2) SE blocks \cite{hu2018squeeze} (channel attention) improve the classification performance by densely adding them into convolutional blocks, while we only add them into self-correlation which is not sufficient enough. After comprehensive comparison, we select the version with spatial attention, i.e., ``SelfDM-SA".
In section \ref{sssec:FEAC}, alternative formulations are discussed. Two deeper networks (ResNet50, ResNet101) and three light-weight networks (MobileNetV2, MobileNetV3, ShuffleNetV2) are consturcted. We find that ``SelfDM-SA-ResNet50" can slightly improve the performance of our backbone network, while the performance of deeper ``SelfDM-SA-ResNet101" is even worse. It shows that high-level features with richer semantic information are not as important as discriminative features with rich spatial information, in a deep matching task. Furthermore, light-weight networks are compared. The performance of ``SelfDM-SA-MobileNetV3" is even better. So we finally select the default ``SelfDM-SA" with VGG, ``SelfDM-SA-ResNet50" and ``SelfDM-SA-MobileNetV3" for further comparison in the next section.
\subsection{Comparison with State-of-the-art Methods}
\label{ssec:csms}
\begin{table*}[hbtp]
\renewcommand{\arraystretch}{1.3}
\caption{Comparisons with the state-of-the-art methods on the synthetic testing set.}
\label{table:comparecomb}
\centering
\footnotesize
\setlength{\tabcolsep}{3.2mm}{
\begin{tabular}{c | c c c c | c c c c }
\hline
\multirow{2}{*}{Methods} & \multicolumn{4}{c|}{Protocol-All} & \multicolumn{4}{c}{Protocol-Detected} \\\cline{2-9}
& IoU & Precision & Recall & F1-score & IoU & Precision & Recall & F1-score \\
\hline
LiJ \cite{li2015segmentation} & 0.3188 & 0.3620 & 0.3723 & 0.3597 & 0.6826 & 0.7752 & 0.7972 & 0.7702 \\
Cozzolino \cite{cozzolino2015efficient} & 0.2377 & 0.3105 & 0.2584 & 0.2728 & 0.6911 & 0.9027 & 0.7513 & 0.7931 \\
BusterNet \cite{wu2018busternet} & 0.3349 & 0.5814 & 0.3764 & 0.4213 & 0.4412 & 0.7660 & 0.4959 & 0.5550 \\
\hline
SelfDM-SA & 0.7233 & 0.8458 & 0.8227 & 0.8216 & 0.7240 & 0.8467 & 0.8235 & 0.8225 \\
SelfDM-SA+PS & 0.7198 & 0.8445 & 0.8232 & 0.8186 & 0.7205 & 0.8453 & 0.8239 & 0.8194 \\
SelfDM-SA+PS+CRF & 0.7403 & 0.8785 & 0.8176 & 0.8308 & 0.7433 & 0.8820 & 0.8209 & 0.8342 \\
\textit{SelfDM-SA+CRF}$\star$ & \textit{0.7317} & \textit{0.8921} & \textit{0.7958} & \textit{0.8252} & \textit{0.7376} & \textit{0.8993} & \textit{0.8024} & \textit{0.8309} \\
\hline
SelfDM-SA-ResNet50 & 0.7372 & 0.7809 & 0.9191 & 0.8312 & 0.7372 & 0.7809 & 0.9191 & 0.8312 \\
SelfDM-SA-ResNet50+PS & 0.7247 & 0.7742 & 0.9112 & 0.8232 & 0.7247 & 0.7742 & 0.9112 & 0.8232 \\
SelfDM-SA-ResNet50+PS+CRF & 0.7438 & 0.8066 & 0.8986 & 0.8358 & 0.7438 & 0.8066 & 0.8986 & 0.8358 \\
\hline
SelfDM-SA-MobileNetV3 & 0.7512 & 0.8575 & 0.8467 & 0.8412 & 0.7512 & 0.8575 & 0.8467 & 0.8412 \\
SelfDM-SA-MobileNetV3+PS & 0.7364 & 0.8488 & 0.8391 & 0.8302 & 0.7364 & 0.8488 & 0.8391 & 0.8302 \\
SelfDM-SA-MobileNetV3+PS+CRF & 0.7531 & 0.8848 & 0.8281 & 0.8394 & 0.7538 & 0.8856 & 0.8289 & 0.8403 \\
\hline
\end{tabular}}
\end{table*}
We adopt four datasets for comprehensive comparisons: our synthetic testing set, CoMoFoD dataset \cite{tralic2013comofod}, CASIA CMFD dataset \cite{wu2018busternet}, and MICC-F220 dataset \cite{amerini2011sift}.
Pasted regions in our synthetic testing set have gone through multiple changes with greater extent. Besides, there is only one pair of similar regions in each image, and there is no obvious ``disturbance" (similar but genuine regions) for the most cases. So it can indicate the capability and robustness of algorithms to detect appearance similar regions under different transformations. We select a representative keypoint-based method (LiJ \cite{li2015segmentation}), an advanced block-based method (Cozzolino \cite{cozzolino2015efficient}), and an end-to-end deep learning method (BusterNet \cite{wu2018busternet}) for comparison. Scores in TABLE~\ref{table:comparecomb} are generated by their codes provided by the authors. It clearly shows that classical methods (``LiJ" and ``Cozzolino") are not robust enough against different transformations, while their detected regions mostly are accurate (comparable scores in ``Protocol-Detected"). Our backbone network has strong ability to detect similar regions. Since there are few ``disturbances", the ability of Proposal SuperGlue to remove false-alarmed regions is not obvious in this dataset. However, it still can be seen that precisions are increased and overall scores are higher. Specifically, ``SelfDM-SA+PS" indicates the two-stage version without ConvCRF optimization, ``SelfDM-SA+PS+CRF" adds ConvCRF, and ``SelfDM-SA+CRF$\star$" directly conducts ConvCRF on the backbone network which is used to demonstrate the effectiveness of Proposal SuperGlue. Furthermore, visual comparions are provided in Fig.~\ref{Figure:cocovis}. In rows 1 and 2, we provide images with obvious scale changes, SelfDM-SA can already achieve satisfied performance. In the column of ``SelfDM-SA+PS", proposals (light blue rectangular regions with lower scores) can enclose suspected regions, and further SuperGlue with SuperPoint can be conducted. With the help of ConvCRF, the score distribution can be optimized. BusterNet only detects their approximate locations without accurate boundaries. And it is difficult for classical methods to detect regions under severe transformations like scale or rotation changes. In rows 3 to 6, we provide four challengable cases. Our backbone network detects some false-alarmed regions or only detects partial regions, while proposal SuperClue can enclose those regions and clearly optimize the detected regions. In the last two rows, we also provide two failure cases. SelfDM-SA can already generate an accurate result while Proposal SuperGlue causes some false-alarmed regions around boundaries in row 7. Or no meaningful proposals are obtained in row 8.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.99999\columnwidth]{cocovis.eps}
\caption{Example copy-move forgery detection results on the synthetic testing set. In the ``GT" column, red regions indicate forged regions and green regions are original regions.}
\label{Figure:cocovis}
\end{figure}
The CoMoFoD dataset consists of $200$ copy-move forged images with resolution $512\times 512$. Besides the version with no postprocessing, these images are processed under $6$ kinds of postprocessing respectively, namely brightness change (BC), contrast adjustments (CA), color reduction (CR), image
blurring (IB), JPEG compression (JC), and noise adding (NA). In TABLE~\ref{table:CoMoFoDnoattack}, two measure protocols are used. ``Correctly Detected Average" only computes the average score of correctly detected images. An image is referred as correctly detected if its pixel-level F1-score is higher
than $0.5$. ``Overall Average" computes their average scores of all images. We find that although our backbone network can achieve excellent performance on synthetic testing images, there is no obvious advantage on CoMoFoD. The main reason is that pasted regions in CoMoFod have gone through only slight changes, compared methods can already achieve good performance. Besides, limited by the training set, the majority of synthetic images only have a pair of similar objects. In another word, there are few of disturbances with similar appearance. Our backbone network has strong ability to detect similar objects under severe transformations, so that there are inevitably some false-alarmed regions (rows 1, 3, 4, 5, 6, 8 in Fig.~\ref{Figure:comovis}), which affect the scores on CoMoFoD. With the help of Proposal SuperGlue, it can be clearly seen that we can remove some false-alarmed regions and complement miss-detected regions (Fig.~\ref{Figure:comovis}). Thus, the performance can be obviously improved. Besides, the robustness against different postprocessing is evaluated and shown in Fig.~\ref{Figure:f1scorescomofodattacks}. Our two-stage version, i.e., SelfDM-SA+PS+CRF, can achieve consistently higher scores under different attacks, which also demonstrate the high robustness of our method. As for the alternative networks, i.e., SelfDM-SA-ResNet50 and SelfDM-SA-MobileNetV3, our Proposal SuperGlue can also notably improve their performance. Especially, the precision scores are improved.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.99999\columnwidth]{comovis.eps}
\caption{Example copy-move forgery detection results on CoMoFoD.}
\label{Figure:comovis}
\end{figure}
CASIA CMFD dataset is selected from CASIA TIDEv2.0 dataset by Wu et al. \cite{wu2018busternet}. There are $1313$ CMFD samples and their authentic counterparts. They provide $256\times 256$ images and masks as a HDF dataset. In our experiments, both pixel-level and image-level scores are computed. Pixel-level scores are the overall average scores of all CMFD samples (the same as ``Overall Average" in TABLE~\ref{table:CoMoFoDnoattack}). Our backbone network based on VGG, i.e., SelfDM-SA, can already achieve higher scores, and Proposal SuperGlue with ConvCRF can further improve its performance. Dramatically, the MobileNetV3 version can achieve the best performance. It further demonstrates that the discriminative capability of features are more important in the deep matching task, and it is different from image understanding tasks (e.g. image classification, object detection, semantic segmentation) in which high-level features with more semantic informantion play a more important role.
\begin{table*}[htp]
\renewcommand{\arraystretch}{1.3}
\caption{Comparisons on CoMoFoD dataset with no attack.}
\label{table:CoMoFoDnoattack}
\centering
\footnotesize
\setlength{\tabcolsep}{3.55mm}{
\begin{tabular}{c | c c c c | c c c }
\hline
\multirow{2}{*}{Method} & \multicolumn{4}{c|}{Correctly Detected Average} & \multicolumn{3}{c}{Overall Average} \\\cline{2-8}
& Detected Rate & Precision & Recall & F1-score & Precision & Recall & F1-score \\
\hline
Ryu2010\cite{ryu2010detection} & 0.450 & 0.9627 & 0.6984 & 0.7993 & 0.4578 & 0.3435 & 0.3737 \\
LiJ \cite{li2015segmentation} & 0.510 & 0.8042 & 0.9586 & 0.8616 & 0.4247 & 0.6633 & 0.4644 \\
Cozzolino \cite{cozzolino2015efficient} & 0.505 & 0.8132 & 0.9384 & 0.8591 & 0.4174 & 0.5042 & 0.4440 \\
Wu2018 \cite{wu2018image} & 0.265 & 0.6111 & 0.7148 & 0.6313 & 0.3629 & 0.4041 & 0.3113 \\
BusterNet \cite{wu2018busternet} & 0.585 & 0.8352 & 0.7875 & 0.8009 & 0.5734 & 0.4939 & 0.4926\\
\hline
SelfDM-SA & 0.475 & 0.7086 & 0.8210 & 0.7350 & 0.5722 & 0.5216 & 0.4660 \\
SelfDM-SA+PS & 0.575 & 0.7895 & 0.8087 & 0.7641 & 0.6086 & 0.5624 & 0.5151 \\
SelfDM-SA+PS+CRF & 0.545 & 0.8139 & 0.8282 & 0.7943 & 0.6375 & 0.5444 & 0.5172 \\
\hline
SelfDM-SA-ResNet50 & 0.520 & 0.7518 & 0.7394 & 0.7208 & 0.5340 & 0.5467 & 0.4753 \\
SelfDM-SA-ResNet50+PS & 0.545 & 0.8439 & 0.7668 & 0.7786 & 0.5910 & 0.5631 & 0.5108 \\
SelfDM-SA-ResNet50+PS+CRF & 0.555 & 0.8728 & 0.7432 & 0.7773 & 0.6231 & 0.5342 & 0.5088 \\
\hline
SelfDM-SA-MobileNetV3 & 0.465 & 0.6526 & 0.8468 & 0.7096 & 0.4833 & 0.5198 & 0.4299 \\
SelfDM-SA-MobileNetV3+PS & 0.530 & 0.7488 & 0.8137 & 0.7438 & 0.5170 & 0.5260 & 0.4645 \\
SelfDM-SA-MobileNetV3+PS+CRF & 0.505 & 0.7885 & 0.8202 & 0.7742 & 0.5600 & 0.5049 & 0.4695 \\
\hline
\end{tabular}}
\end{table*}
\begin{figure}[htp]
\begin{minipage}[b]{0.8\linewidth}
\centering
\centerline{\includegraphics[width=5cm]{comolegend.eps}}
\end{minipage}
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=5cm]{comoBC.eps}}
\centerline{\footnotesize{Brightness change (BC)}}
\end{minipage}
\hfill
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=5cm]{comoCA.eps}}
\centerline{\footnotesize{Contrast adjustments (CA)}}
\end{minipage}
\vfill
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=5cm]{comoCR.eps}}
\centerline{\footnotesize{Color reduction (CR)}}
\end{minipage}
\hfill
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=5cm]{comoIB.eps}}
\centerline{\footnotesize{ Image blurring (IB)}}
\end{minipage}
\vfill
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=5cm]{comoJC.eps}}
\centerline{\footnotesize{JPEG compression (JC)}}
\end{minipage}
\hfill
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=5cm]{comoNA.eps}}
\centerline{\footnotesize{Noise adding (NA)}}
\end{minipage}
\caption{Pixel-level F1 scores (y-axis) on CoMoFoD under attacks (x-axis).}
\label{Figure:f1scorescomofodattacks}
\end{figure}
\begin{table*}[htp]
\renewcommand{\arraystretch}{1.3}
\caption{Performance analysis on CASIA CMFD dataset.}
\label{table:casia}
\centering
\footnotesize
\setlength{\tabcolsep}{5.3mm}{
\begin{tabular}{c | c c c | c c c}
\hline
\multirow{2}{*}{Method} & \multicolumn{3}{c|}{Pixel Level} & \multicolumn{3}{c}{Image Level} \\\cline{2-7}
& Precision & Recall & F1-score & Precision & Recall & F1-score \\
\hline
Ryu2010 \cite{ryu2010detection} & 0.2271 & 0.1336 & 0.1640 & 0.9701 & 0.2447 & 0.3908 \\
Christlein \cite{christlein2012evaluation} & 0.3709 & 0.0014 & 0.0023 & 0.6849 & 0.6782 & 0.6815 \\
Cozzolino \cite{cozzolino2015efficient} & 0.2492 & 0.2681 & 0.2543 & 0.9951 & 0.3061 & 0.4682 \\
Wu2018 \cite{wu2018image} & 0.2397 & 0.1379 & 0.1464 & 0.6637 & 0.7359 & 0.6980 \\
BusterNet-simi \cite{wu2018busternet} & 0.4723 & 0.4844 & 0.4372 & 0.7153 & 0.8073 & 0.7585 \\
BusterNet \cite{wu2018busternet} & 0.5571 & 0.4383 & 0.4556 & 0.7822 & 0.7389 & 0.7598 \\
\hline
SelfDM-SA & 0.6551 & 0.4353 & 0.4635 & 0.7707 & 0.7807 & 0.7757 \\
SelfDM-SA+PS & 0.6485 & 0.4531 & 0.4709 & 0.7860 & 0.7609 & 0.7732 \\
SelfDM-SA+PS+CRF & 0.6494 & 0.4520 & 0.4782 & 0.7860 & 0.7609 & 0.7732 \\
\hline
SelfDM-SA-ResNet50 & 0.5358 & 0.4500 & 0.4356 & 0.6038 & 0.9238 & 0.7303 \\
SelfDM-SA-ResNet50+PS & 0.5359 & 0.4676 & 0.4464 & 0.6164 & 0.9018 & 0.7323 \\
SelfDM-SA-ResNet50+PS+CRF & 0.5729 & 0.4679 & 0.4595 & 0.6164 & 0.9018 & 0.7323 \\
\hline
SelfDM-SA-MobileNetV3 & 0.6248 & 0.4778 & 0.4843 & 0.6914 & 0.8294 & 0.7542\\
SelfDM-SA-MobileNetV3+PS & 0.6272 & 0.4856 & 0.4891 & 0.7040 & 0.8096 & 0.7531 \\
SelfDM-SA-MobileNetV3+PS+CRF & 0.6362 & 0.4752 & 0.4918 & 0.7040 & 0.8096 & 0.7531 \\
\hline
\end{tabular}}
\end{table*}
\begin{table}[!t]
\renewcommand{\arraystretch}{1.3}
\caption{Image-level performance on MICC-F220 dataset.}
\label{table:miccf220}
\centering
\footnotesize
\begin{tabular}{c | c c c}
\hline
Method & TPR & FPR & F1-score \\
\hline
Cozzolino \cite{cozzolino2015efficient} & 0.8455 & 0.1727 & 0.8378 \\
LiJ \cite{li2015segmentation} & 0.7091 & 0.1727 & 0.7536 \\
GoDeep \cite{silva2015going} & 0.4545 & 0.4182 & 0.4854 \\
Zandi \cite{zandi2016iterative} & 0.7818 & 0.4818 & 0.6908 \\
BusterNet \cite{wu2018busternet} & 0.4909 & 0.2000 & 0.5806 \\
SelfDM-SA & 0.9273 & 0.2545 & 0.8500 \\
SelfDM-SA+PS+CRF & 0.9182 & 0.2272 & 0.8559 \\
SelfDM-SA-ResNet50 & 0.9727 & 0.6909 & 0.7304 \\
SelfDM-SA-ResNet50+PS+CRF & 0.9545 & 0.6454 & 0.7343 \\
SelfDM-SA-MobileNetV3 & 0.9636 & 0.4182 & 0.8092 \\
SelfDM-SA-MobileNetV3+PS+CRF & 0.9545 & 0.3909 & 0.8140 \\
\hline
\end{tabular}
\end{table}
MICC-F220 is composed by $220$ images: $110$ tampered images and $110$ originals. There is no ground truth, and we evaluate the image-level performance. All the methods are evaluated by True Positive Rate (TPR), False Positive Rate (FPR) and corresponding F1-score, which are computed as: $TPR=TP/(TP+FN)$, $FPR=FP/(TN+FP)$, $F_1=2TP/(2TP+FP+FN)$, where $TP$ denotes true positive, $TN$ denotes true negtive, $FN$ denotes false negative and $FP$ denotes false positive. SelfDM-SA and the corresponding Proposal SuperGlue version can achieve better performance than many other state-of-the-art methods. Especially, we can achieve higher TPRs. However, the alternative formulations have higher FPRs, FPRs of the ResNet50 version are even greater than $0.6$. Considering all the experimental results on four datasets, we do not recommend the use of deeper networks for feature extraction, because they have more parameters, low efficiency and high false-alarmed rates. The VGG version is more stable and robust. The MobileNetV3 version has less parameters to learn and can achieve comparable performance on different datasets. Even so, we find that our two-stage framework can be applied to backbone networks with different feature extractors to achieve better performance.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we propose a two-stage framework for deep learning based copy-move forgery detection. Our two-stage framework integrates self deep matching and keypoint matching by obtaining highly suspected proposals. The first stage is a backbone network which adopts atrous convolution, skip matching, and spatial attention. In the second stage, our Proposal SuperGlue is proposed to remove false alarms and complement incomplete regions. Specifically, we build a proposal selection module to enclose suspected regions, and conduct pairwise matching based on SuperPoint and SuperGlue. Integrated score map generation and refinement methods are proposed to obtain final results. Our two-stage framework can achieve consistently better performance on different public datasets. The two-stage framework relies on the performance of the backbone network. In the future, our two-stage framework can be further improved by designing a more powerful backbone network.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
| d3c791d1ba9000b2e69d92c8bd2bcf3c3e3191b0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $n \geq 2$ and $m \geq 2$ be two integers, and let $G_n(m)$ denote the group of $n\times n$ upper triangular matrices with entries in $\mathbb{Z}/m \mathbb{Z}$ and ones along the diagonal. We number the rows of each matrix in $G_n(m)$ from top to bottom. We consider the following Markov chain $(X_t)_{t\geq0}$ on $G_{n}(m)$: $X_t$ is derived from $X_{t-1}$ by picking a row $i \in \{2,\ldots, n\}$ uniformly at random and with probability $1/4$ adding it to row $i-1$, with probability $1/4$ subtracting it from row $i-1$, and otherwise staying fixed.
Let $P^t_A(B)$ be the probability that $X_t=B$ given that $X_0=A$. The walk is stationary with respect to the uniform distribution $U$ on $G_n(m)$, and our main result studies the mixing time of the walk, defined as
\begin{equation*}
t_{mix}(\varepsilon)= \inf \lbrace t\geq 0 : d_n(t) \leq \varepsilon \rbrace,
\end{equation*}
where
\begin{equation*}
d_n(t) = \max_{A \in G_n(m)} \Vert P^t_{A}-U \Vert_{T.V.} = \frac{1}{2} \max_{A \in G_n(m)} \big \lbrace \sum_{B \in G_n(m)} \vert P^t_{A}(B) - U(B) \vert \big \rbrace
\end{equation*}
is the total variation distance of $P^t_A(B)$ from $U$. Our main theorem determines the mixing time of the random walk $X_t$.
\begin{theorem}\label{main}
For $m$ prime, there are universal, positive constants $A,B$ and $D$, such that for $t_{n,m}= D (m^2 n \log n + n^2 e^{9 \sqrt{ \log m}}) + cnm^2 \log \log n ,$ we have that
$$\ d_n(t_{n,m}) \leq B \left( e^{-Ac} + \frac{1}{\sqrt{t_{n,m}}} \right),$$
for $c>2$.
\end{theorem}
For the case where $m$ is not prime, we are missing one of the main tools, namely Theorem 3 of \cite{Hough}. We are still able to prove a similar upper for the mixing time, which is slightly less tight than the one of Theorem \ref{main}.
\begin{theorem}\label{maing}
There are universal, positive constants $A,B,C$ and $D$, such that for $t_{n,m}= D (m^2 n \log n + n^2 e^{C (\log m)^{2/3}}) + cnm^2 \log \log n ,$ we have that
$$\ d_n(t_{n,m}) \leq B \left( e^{-Ac} + \frac{1}{\sqrt{t_{n,m}}} \right),$$
for $c>2$.
\end{theorem}
This process has a long history. The case $n=3$ was first introduced by Zack~\cite{Zack}, and Diaconis and Saloff-Coste proved that order $m^2$ steps are necessary and sufficient for convergence to uniform~\cite{Laurent, Har, Moderate}. This result was later proven again by Bump, Diaconis, Hicks, Miclo and Widom using Fourier analysis~\cite{Hicks}. For $n$ growing, a first upper bound of order $n^7$ was introduced by Ellenberg~\cite{Ellenberg}, which was later improved by Stong to $n^3 m^2 \log m $ \cite{Stong}. The case where $m=2$ was treated by Peres and the second author~\cite{PeresSly}, who proved the mixing time is $O(n^2)$.
Arias-Castro, Diaconis and Stanley \cite{StanleyD} used super-character theory, introduced by Andre \cite{Andre3}, \cite{Andre}, \cite{Andre2} and Yan \cite{Yan}, \cite{Yan2} to bound the spectrum of the random walk. This results in a bound for the mixing time of order $n^4 m^4 \log n $, for the case where $m$ is prime. The first author~\cite{EN} improved their analysis to $n^4 m^2$, which gives the correct order of the mixing time in $m$, but not in $n$.
More recently, some other features of this walk have been studied. Diaconis and Hough \cite{Hough} studied how many steps an element on the $i$--th diagonal of the matrix needs to mix. We state their result in detail in Section \ref{prime}. The projection onto the final column of the matrix is itself a well known Markov chain known as the East Model. Ganguly, Lubetzky and Martinelli~\cite{GLM:15} proved that the East model exhibits cutoff and later Ganguly and Martinelli~\cite{GM} extending this to the last $k$ columns of the upper triangular matrix walk.
Recently, Hermon and Thomas \cite{HermonThomas} considered a different question concerning $G_n(m)$. They sample $k$ generators uniformly at random and they prove cutoff for the case where $k$ is growing with $\vert G_n(m) \vert$.
Our strategy is to study how fast the first row mixes and proceed by induction on $n$. We do so by analysing it as a random sum of the second row at random times. It is important to understand the values that the second row takes to understand how well mixed this random sum becomes. This is easier to do in the case where $m$ is a prime, thanks to the work of Diaconis and Hough \cite{Hough}.
\section{Preliminaries}\label{prem}
For the remainder of the paper we study instead the continuous version of the random walk. For each $ i \in \{ 2,\ldots, n \}$, we consider a rate $1$ Poisson clock, and when the $i$--th clock rings, we either add or subtract row $i$ to row $i-1$, each event happening with probability $1/2$. Our main result for this walk is:
\begin{theorem}\label{mainc}
For $m$ prime, there are universal, positive constants $A, B$ and $D$, such that for $t_{n,m}= D(m^2 \log n + n e^{9 \sqrt{ \log m}} ) + cm^2 \log \log n,$ we have that
$$ d_n(t_{n,m}) \leq B e^{-Ac},$$
for $c>2$.
\end{theorem}
Similarly, we have the following theorem for the continuous time random walk.
\begin{theorem}\label{maingc}
There are universal, positive constants $A,B,C$ and $D$, such that for $t_{n,m}= D (m^2 \log n + n e^{C (\log m)^{2/3}}) + cm^2 \log \log n ,$ we have that
$$\ d_n(t_{n,m}) \leq B e^{-Ac} ,$$
for $c>2$.
\end{theorem}
Theorems \ref{mainc} and \ref{maingc} imply Theorems \ref{main} and \ref{maing} respectively, because of Theorem $20.3$ of \cite{Peresbook} and equation (3.4) of \cite{R}.
\subsection{The induction lemma}
The goal of this section is to prove an inequality that relates $d_n(t) $ to $d_{n-1}(t)$ thus allowing to prove Theorem \ref{maingc} inductively.
We break $X_t$ in two parts: let $r_t$ be the $n \times n$ matrix that has the same first row as $X_t$ and every other entry zero, and let $Y_t= X_t-r_t$. Let $t_1, t_2 \ldots $ be the times that the clock of the second row rings. Let $N(t)= \max \{ j \geq 0 : t_j \leq t\}$. We have that
\begin{equation}\label{separating}
X_t= Y_t + \sum_{j=1}^{N(t)} a_j E(1,2) Y_{t_j} ,
\end{equation}
where the $a_j \in \{-1,1\}$, each value occurring with probability $1/2$.
Equation~\eqref{separating} will allow us to separate the mixing time of the first row from the mixing time of the rest of the matrix.
The main idea is to prove a bound for the $\ell^2$ distance between $r_t$ and the uniform measure on $(\mathbb{Z}/m \mathbb{Z})^{n-1}$, by studying the spectrum of the transition matrix of $r_t$. These eigenvalues will be indexed by vectors $y \in (\mathbb{Z}/m \mathbb{Z})^{n}$ whose first coordinate is zero. For a nonzero such $y $, let $Z_y^t(i)= X_t(i) y $, where $X_t(i) $ is the $i$--th row of $X_t$. From now on, we will write that $y \in (\mathbb{Z}/m \mathbb{Z})^{n-1}$, though we actually mean that $y$ has $n$ coordinates, the first one of which is zero. The $\ell^2$ distance between $r_t$ and its stationary measure at time $t_{n,m}$ is given in terms of $\{Z^s_y(2) \}$ with $y \in (\mathbb{Z}/m \mathbb{Z})^{n-1} $ and $ s \in \{ t_1, \ldots, N(t_{n,m})\}$. We prove that most of these values are conveniently large. For $a,b \in \mathbb{Z}/m \mathbb{Z}$, we use the notation $|a| >b$ to mean that $a \in \{b+1, \ldots, m-b-1\} $ .
Let $P^t$ be the indicator function that the clock of the second row rings at time $t$.
Let $A^t_{y,x}= \int_0^t 1_{\lbrace P^s=1 \rbrace } 1_{\lbrace \vert Z_y^s(2) \vert > x\rbrace } ds$ count the number of times $s$ that the second row clock rings and $\vert Z_y^s(2) \vert > x$. Let $I \in \{2,\ldots, n \}$ and let $P_2=\langle e_1, e_2 \rangle \setminus \langle e_1 \rangle $, $Q_I= \langle e_1, \ldots e_{I-1} \rangle \setminus \langle e_1, e_2 \rangle $ and $W_I=\in (\mathbb{Z}/m \mathbb{Z})^{n-1} \setminus \langle e_1, \ldots e_{I-1} \rangle$. Let $A$ be an appropriately chosen constant, which is universal on $n$ and $m$, and let $E_{t,I,x,w}$ be the event that $A_{y,x}^t \geq A^{-1}t$ for every $y \in W_I$, $A_{y,w}^t \geq A^{-1}t$ for every $y \in Q_I$ and $A_{y,m/8}^t \geq A^{-1}t$ for every $y \in P_2$.
The following lemma will help us prove Theorem \ref{maing} inductively. Let $\mathcal{F}_t$ be the $\sigma-$algebra generated by the all of updates except the random signs used when adding the second row to the first. In particular $\mathcal{F}_t$ contains all the information from rows 2 to $n$ as well as the times at which row 2 is added to row 1.
\begin{lemma}\label{first}
Let $q_{t}$ be the conditional distribution of $r_t$ at time $t$ given $\mathcal{F}_t$. Then
$$d_{n}(t) \leq d_{n-1}(t) + \mathbb{E} [\Vert q_{t}-u \Vert_{T.V.}],$$
where $u$ is the uniform measure on $(\mathbb{Z}/ m \mathbb{Z})^{n-1}$.
\end{lemma}
\begin{proof}
Let $X$ be a uniformly random element of $G_{n}(m)$. We can couple $Y_t$ and the $n-1$ last rows of $X$ except with probability $d_{n-1}(t)$. This coupling moreover can be made $\mathcal{F}_t$ measurable. Conditional on $\mathcal{F}_t$ we can then couple the first row except with probability $\Vert q_{t}-u \Vert_{T.V.}$. The lemma then follows by averaging.
\end{proof}
We use Lemma \ref{first} to prove Theorems \ref{mainc} and \ref{maingc} by induction. In particular, we prove the following lemma.
\begin{lemma}\label{q}
Let $t=t_{n,m}$. There exist positive constants $a, b, f,g$ such that for any $n$ we can find $x,w$ and $I$ such that on the event $E_{t,I,x,w}$
$$ \mathbb{E}[\Vert q_{t}-u \Vert_{T.V.}] \leq a n^{-1}(\log n)^{-c},$$
where $c>2$ is the constant from Theorems \ref{mainc} and \ref{maingc}.
Furthermore,
$$ \pr{E_{t,I,x,w} ^c} \leq b e^{-fc} \left( m^{-gn} + n^{-gm} \right).
$$
\end{lemma}
We now show how to use Lemma \ref{q} to prove Theorems \ref{mainc} and \ref{maingc}.
\begin{proof}[Proof of Theorems \ref{mainc} and \ref{maingc}.]
The proof for the case $n=3$ can be found in \cite{Laurent}. Combining Lemmas \ref{first} and \ref{q} with induction, we get that
\begin{align*}
d_n(t_{n,m}) & \leq B e^{-dc} + \sum_{i=4}^n \left( \frac{1}{i(\log i)^c} + b e^{-fc} \left( m^{-gi} + i^{-gm} \right) \right) ,
\end{align*}
where the term $B e^{-dc}$ comes from the case $n=3$. Therefore,
\begin{align*}
d_n(t_{n,m}) & \leq B e^{-dc} + \tilde{b} e^{-fc}+ \int_{i=3}^{\infty} \frac{1}{x(\log x)^c}dx\\
& \leq B e^{-dc}+ \tilde{b} e^{-fc}+ \frac{(\log 3)^{1-c}}{c-1},
\end{align*}
which completes the proof of Theorems \ref{mainc} and \ref{maingc} for $c>2$.
\end{proof}
\subsection{The $\ell^2$ bound.}
The goal of this section is to establish an inequality, which will be used to bound $\mathbb{E}[\Vert q_{t}-u \Vert_{T.V.}]$. Let $N(t)$ be the number of times that the second clock has rang by time $t$.
For $k \in \mathbb{N}$ and $\underline{w}=(w_1, \ldots, w_k)$, with $w_i \in ( \mathbb{Z}/m \mathbb{Z})^{(n-1)}$ let $G_{k,\underline{w}}$ be the event that $N(t)=k$ and that the second row $X_t(2)$ is equal to $w_j$ at the time of the $j$-th ring for $j=1,\ldots,k$.
Let $q_{\substack{k,\underline{w}}}$ be the distribution of $r_t$, conditioning on $G_{k,\underline{w}}$.
Then we have that
\begin{align}
\mathbb{E}[\Vert q_{t}-u \Vert_{T.V.}] =
\label{l} \sum_{\substack{k, \underline{w}}}\Vert q_{\substack{k,\underline{w}}}-u \Vert_{T.V.} \prcond{G_{k,\underline{w}}}{\mathcal{F}_t}.
\end{align}
Each $q_{\substack{k,\underline{w}}}$ has the same orthonormal eigenbasis as the simple random walk on $ \mathbb{Z}/m \mathbb{Z}$, despite the fact that at each step we are adding/subtracting a different quantity. The corresponding eigenvalues are $e^{-2(k- \sum_{s=1}^k \lambda_{\substack{y, w_s}})},$
where the $ \lambda_{\substack{y, w_s} }= \cos \frac{2 \pi \langle y, w_s \rangle}{m}$ are the eigenvalues of the discrete time Markov chain on $ ( \mathbb{Z}/m \mathbb{Z})^{n-1}$ that adds or subtracts $w_s$ to the current state with probability $1/2$.
Then we use the classical $\ell^2 $ bound,
\begin{align}\label{y}
&4 \Vert q_{\substack{k, \underline{w}}}-u \Vert_{T.V.}^2 \leq \sum_{ y \in (\mathbb{Z}/m \mathbb{Z})^{n-1} \setminus \{ \textbf{0} \} } e^{-2(k- \sum_{s=1}^k \lambda_{\substack{y, w_s}})}.
\end{align}
To continue with bounding \eqref{y}, we will need the following technical lemma.
\begin{lemma}\label{integral}
We have that
$$\sum_{j =1}^{m-1}
e^{-2 x (1- \cos \frac{2 \pi j }{m} )} \leq m e^{-2x} + \frac{\sqrt{3} m}{ 2\sqrt{2 \pi x}} ,$$
where $x>0$.
\end{lemma}
\begin{proof}
\begin{align}
\sum_{j =1}^{m-1}
e^{-2 x (1- \cos \frac{2 \pi j }{m} )}
& \leq 2 \sum_{j =1}^{m/2}
e^{-2 x (1- \cos \frac{2 \pi j }{m} )}\cr
& \label{spl} \leq m e^{-2x} + 2 \sum_{j =1}^{m/4} e^{-2 x (1- \cos \frac{2 \pi j }{m} )} ,
\end{align}
where for the first term in \eqref{spl} we bound the negative cosines. Using the inequality $\cos x \leq 1 - \frac{x^2}{2} + \frac{x^4}{24}$, we get that
\begin{align}
\eqref{spl}& \label{thr} \leq m e^{-2x} + 2 \sum_{j=1}^{m/4} e^{- \frac{8j^2 \pi^2}{3 m^2} x }\\
&\label{i} \leq m e^{-2x} + 2 \int_{0}^{\infty } e^{- \frac{8w^2 \pi^2x }{3m^2} }dw.
\end{align}
Using the substitution $v=\frac{4 \pi \sqrt{x} }{\sqrt{3}m}w,$ we get that
\begin{align}
\eqref{i}&\leq m e^{-2x} + \frac{ \sqrt{3}m}{ 2\pi \sqrt{ x}} \int_{0}^{\infty } e^{-v^2/2}dv \cr
& \leq m e^{-2x} + \frac{\sqrt{3} m}{ 2\sqrt{2 \pi x}} .
\end{align}
\end{proof}
\subsection{Coupling with Exponentials}\label{good}
Recall that $Z^t_y(i):= X_t (i)y,$ for $i=1,\ldots, n$. In this section, we study the time intervals during which $Z^t_y(i) \neq 0$. We want to understand for how long $Z^t_y(i) $ remains equal to zero and for how long it doesn't.
Let $y \in (\mathbb{Z}/m \mathbb{Z})^{n-1} \setminus \langle e_1 \rangle,$ where $ \langle e_1 \rangle$ denotes the subspace of $ (\mathbb{Z}/m \mathbb{Z})^{n-1}$ generated by the vector $e_1=(1,0,\ldots ,0)$.
We start by proving that there is a good chance that $Z^t_y(i)$ will be non-zero after order $ n $ steps.
\begin{lemma}\label{nzero}
Let $T_i$ denote the first time that $Z^t_y(i)$ is non-zero. We have that
$$\pr{T_i> n -1 +c } \leq e^{-c}.$$
\end{lemma}
\begin{proof}
We consider the random walk on $\mathbb{Z}$, starting at zero. When at $x$, we move to $x+1$ according to a rate $1$ Poisson clock or to $x-1$ according to a rate $1/2$ Poisson clock. A Chernoff bound gives that
\begin{equation}\label{t}
\pr{T_i> t } \leq \pr{S_t <n-1}= \pr{e^{- \lambda S_t} > e^{- \lambda (n-1)}} \leq e^{\lambda (n-1)} \expect{e^{-\lambda S_t}}
\end{equation}
We have that $S_t= M_t-N_t,$ where $M_t$ is a Poisson($t$) random variable and $N_t$ is a Poisson($t/2$) random variable. Therefore, we have that
\begin{equation*}
\eqref{t} \leq e^{\lambda (n-1)} e^{-\frac{3}{2}t+t (e^{- \lambda + \frac{1}{2} e^{\lambda}})}.
\end{equation*}
Setting $\lambda= \log 2$, we get that
$$\pr{T_i> t } \leq 2^{n-1} e^{-t (3 - \sqrt{2})}.$$
Setting $t= n-1 +c$ we get the desired result.
\end{proof}
We first want to study for how long $Z^t_y(2)$ can stay equal to zero.
\begin{definition}
Let $\ell_1$ be a time such that $Z^{\ell_1}_y(2)=0$ and $Z^{\ell_1^-}_y(2)\neq 0$. Let $\ell_2= \inf \{t>\ell_1: Z^{t}_y(2)\neq 0 \}$. We will call $[\ell_1, \ell_2] $ a $y$--zero interval.
\end{definition}
\begin{lemma}\label{nstar}
Let $m>2$ be an odd integer. Let $[\ell_1, \ell_2]$ be a $y$--zero interval. Then,
$$\prcond{\ell_2-\ell_1>13k }{\mathcal{F}_t} \leq e^{-k},$$
where $k>0$.
\end{lemma}
\begin{proof}
Consider the column dynamics, $Z^t_y(i), $ where $i =2,\ldots, n$. Let $Z_t$ be the first entry of the column $Z^t_y(i)$ that is not divisible by $m$, when read from top to bottom. If this starts at the second coordinate, then we want to study the first time $\xi$ that $Z$ returns at $2$. We are going to couple $Z_t$ with a biased random walk on $\mathbb{Z}$.
Let $S_x$ is the random walk on $\mathbb{Z}$ at time $x$, which starts at $0$, and moves by $+1$ according to a Poisson($1$) clock and by $-1$ according to a Poisson$(1/2)$ clock. Then, a Chernoff bound gives
$$\pr{\xi>x} \leq \pr{S_x <0} \leq \pr{ e^{-\lambda S_x} \geq 1 } \leq \expect{e^{-\lambda S_x}} \leq e^{-\frac{3x}{2} } e^{x (e^{- \lambda }+\frac{e^{ \lambda}}{2})},$$
since $S_x= M_x-N_x,$ where $M_x$ is a Poisson($x$) random variable and $N_x$ is a Poisson($x/2$) random variable.
Optimizing over $\lambda$, we have that
$$\pr{\xi>x} \leq e^{(-\frac{3}{2} + \sqrt{2})x } \leq e^{- \frac{x}{13}} .$$
Therefore, if $x= 13k $,
$$\prcond{\ell_2-\ell_1>13k }{\mathcal{F}_t} \leq e^{-k}.$$
\end{proof}
We can also study the length of the intervals during which $Z^t_y(2)\neq 0$.
\begin{definition}
Let $\ell_3$ be a time such that $Z^{\ell_1}_y(3)\neq0$ and $Z^{\ell_3^-}_y(2)= 0$. Let $\ell_4= \inf \{t>\ell_3: Z^{t}_y(2)=0 \}$. We will call $[\ell_3, \ell_4] $ a $y$--non-zero interval.
\end{definition}
\begin{lemma}\label{star}
Let $m>2$ be an integer. Let $[\ell_3, \ell_4] $ be a $y$--non-zero interval. Then,
$$\prcond{\ell_4-\ell_3 \leq k }{\mathcal{F}_t} \leq e^{-k},$$
where $k>0$.
\end{lemma}
\begin{proof}
We couple $\ell_4-\ell_3$ with the time it takes for the clock of the third row to ring and the statement follows.
\end{proof}
We are now going to put all this information together to prove that during any interval, $ Z^t_y(2)$ is non-zero for a constant fraction of the time. Let $t_0= n$. We break up the interval $[t_0,t_{n,m}] $ in intervals $[t_j, t_{j+1}]$ of length $L $, so that $t_j=n+jL$. Let $j \in \{ 1,\ldots,n\}$ and let $g=15$.
\begin{definition}\label{g}
An interval $[t_j, t_{j+1}] $ is called $y$--good if $ Z^t_y(i) \neq 0 \mod m $ for at least $1/g$ of $[t_j, t_{j+1}]$.
Let $D_{y}^i$ be the set of all $y$--good intervals by time $t_{n,m}$ and let $M^t_{y}$ be the number of $y$--good intervals that have occurred by time $t$.
\end{definition}
The following lemma is a standard tool that will help us study how likely it is for a given interval to be $y$--good.
\begin{lemma}\label{expon}
Let $B_1, \ldots, B_k$ be independent, exponential random variables with mean one. We have that
\begin{enumerate}
\item[(a)] $\pr{\sum_{i=1}^k B_i > 2k} \leq \left( \frac{2}{e} \right)^{k}.$
\item[(b)] $\pr{\sum_{i=1}^k B_i < \frac{k}{2}} \leq \left(\frac{6}{7}\right)^k$.
\end{enumerate}
\end{lemma}
The following lemma tells that a constant fraction of intervals are $y$--good.
\begin{lemma}\label{prob}
At time $t_{n,m}$ we have that
$$\pr{M_{y}^t \leq \frac{t_{n,m}}{10L}} \leq \frac{2}{m^{2Dn}},$$
for every $j \geq 1$, for a suitable constant $D$.
\end{lemma}
\begin{proof}
Consider the $y$--non-zero intervals $A_b \subset [T_2,t_{n,m}]$ and the $y$--zero intervals $B_k \subset [T_2,t_{n,m}]$. Let $\vert A_b \vert$ be the length of $A_b$. Let
$$W_y= \sum_b \vert A_b \vert$$
be the total time that $Z^s_y(2)$ is not equal to zero.
For the case where $m$ is odd, we have that
\begin{align}
\label{nine1}& \lbrace M_{y}^t \leq \frac{t_{n,m}}{10L}\rbrace \subset \lbrace W_y\leq x t_{n,m}\rbrace ,
\end{align}
where $x= \frac{1}{10}+ \frac{9}{10g}$. Lemmas \ref{nstar} and \ref{star} say that we can couple each $\vert A_b \vert$ with a Exponential random variable with mean $1$ and each $\vert B_k\vert /13$ with an exponential random variable with mean $1$. Let $ r \in \left[\frac{1}{5} (1+9g^{-1})t_{n,m}, \frac{9}{520} (1-g^{-1}) t_{n,m} \right]$.
We have that either $A_1 \cup \ldots \cup A_r \subset [0, t_{n,m}]$ or $(\cup_{i=1}^r B_i) $
contains all $y$--non-zero intervals that $[0, t_{n,m}]$ contains. This is summarized in the following equation
$$W_y \geq \min \big \{\sum_{b=1}^r \vert A_b \vert, t_{n,m} - \sum_{k=1}^r \vert B_k \vert \big \}.$$
Let $\mathcal{B}$ be the event that $\lbrace T_2<2Dn \log m \rbrace$. Therefore,
\begin{equation}
\label{r} \pr{W_y\leq xt_{n,m},\mathcal{B}} \leq \pr{\sum_{b=1}^r \vert A_b \vert \leq x t_{n,m} } + \pr{\sum_{k=1}^r \vert B_k \vert \geq \left( 1-x \right) t_{n,m} }.
\end{equation}
Because of the choice of $r$ we have that
\begin{align}
\eqref{r} & \leq \pr{\sum_{b=1}^r \vert A_b \vert \leq r/2} + \pr{\sum_{k=1}^r \vert B_k \vert \geq 2r}\cr
& \leq \left(\frac{6}{7} \right)^{ r} + \pr{\sum_{k=1}^r \frac{\vert B_k\vert }{13} \geq \frac{1-g^{-1}}{13} t_{n,m
} }\cr
&\label{e} \leq \left(\frac{6}{7} \right)^{r} + \left( \frac{2}{e} \right)^{ r} ,
\end{align}
where \eqref{e} follows from Lemma \ref{expon}.
Putting \eqref{nine1} and \eqref{e} together, we have that
\begin{align}
\label{nine}&\pr{M_{y}^t \leq \frac{t_{n,m}}{10L}} \leq \pr{W_y\leq x t_{n,m},\mathcal{B}} + \frac{1}{m^{2Dn}} \leq \left(\frac{6}{7} \right)^{r} + \left( \frac{2}{e} \right)^{ r}+\frac{1}{m^{2Dn}} .
\end{align}
Using the definition of $r$ and \eqref{nine} we get the desired result.
For the case where $m$ is even, project all values over $\mathbb{Z}/2 \mathbb{Z}$. Equation 2.2 of \cite{PeresSly} says that for every $\varepsilon>0$, we have that
$$\pr{\big \vert W_{t_{n,m}} - \frac{t_{n,m}}{2} \big \vert \geq \varepsilon} \leq 2^{n+1} e^{- \frac{t_{n,m} \varepsilon^2 \lambda}{12}},$$
where $\lambda$ is a positive constant not depending on $n,m$.
Therefore
\begin{align*}
&\pr{W_y\leq x t_{n,m}} \leq \frac{1}{m^{dn}},
\end{align*}
for a suitable constant $d$.
\end{proof}
Finally, we need a lemma that says that if $ Z_y^s(2) $ is sufficiently big with a good probability during a $y$--good interval, then $ Z_y^s(2) $ is sufficiently big for a constant fraction of $[0, t_{n,m}]$.
Let $\mathcal{G}_j$ be the event that $[t_j,t_{j+1}] \in D_y^2$ and let $\mathcal{J} $ be the indices $j$ that satisfy $\pr{\mathcal{G}_j}\neq 0$. Recall that $P^t$ is the indicator function that the clock of the second row rings at time $t$ and $A^t_{y,x}= \int_0^t 1_{\lbrace P^s=1 \rbrace } 1_{\lbrace \vert Z_y^s(2) \vert > x\rbrace } ds$. The following is crucial to proving the second part of Lemma \ref{q}.
\begin{lemma}\label{r15}
For $y \notin \langle e_1 \rangle$, consider $\tilde{B}_{y}= \lbrace A^t_{y,x}
\geq (2A)^{-1} t_{n,m} \rbrace $ and assume that $\prcond{\tilde{B}_{y}}{\mathcal{G}_j} \geq 1/2,$ for every $j \in \mathcal{J}$. We have that there are constants $b$ and $g$ such that
\begin{align*}
\pr{\tilde{B}_{y}} \geq 1- \left( e^{- g \frac{t_{n,m}}{L}} + \frac{b}{m^{gn}} \right) e^{-c},
\end{align*}
where $b,g$ are suitable constants, $c$ is the constant from Theorem \ref{main} and $L$ is the length of the intervals.
\end{lemma}
\begin{proof}
Let $B_{y}$ denote the event that $\vert Z^s_y(2) \vert > x$ for at least $A^{-1}$ of
$ [0, t_{n,m}]$, where $A$ is a suitable constant. Let $\mathcal{F}_{j}$ be the $\sigma$--algebra generated by all the row operations performed before time $t_j$. We have that
\begin{align}
\sum_j I_{B_{j,y}} & = \sum_j \prcond{B_{j,y}}{\mathcal{F}_{j}} + \sum_{j} \left( I_{B_{j,y}}- \prcond{B_{j,y}}{\mathcal{F}_{j}}\right),
\end{align}
where the term $M=\sum_{j} \left( I_{B_{j,y}}- \prcond{B_{j,y}}{\mathcal{F}_{j}}\right)$ is a martingale.
Using Lemma \ref{prob}, we have that
\begin{align}
\pr{B_y} &=\pr{ \sum_j I_{B_{j,y}} \geq A^{-1} t_{n,m}}\cr
& \label{MM12} \geq \pr{ \sum_j \prcond{B_{j,y}}{\mathcal{F}_{j}} \geq 2A^{-1} t_{n,m}, M \geq -A^{-1} t_{n,m} }.
\end{align}
Using the Azuma-Hoeffding inequality we have that
\begin{align}
\label{M12} \pr{ M \geq -A^{-1} t_{n,m} } &\geq 1- e^{-t_{n,m}} \geq 1- m^{-dn}e^{-c},
\end{align}
with $d>2$.
Our assumption that $\prcond{\tilde{B}_{y}}{\mathcal{G}_j} \geq 1/2$, gives that
\begin{equation}
\prcond{B_{j,y}}{\mathcal{F}_{j}} \geq H^y_j I_{G^y_j},
\end{equation}
where $H^y_j$ is a Bernoulli$(1/2)$ random variable, that is independent of $I_{G^y_j}$. Conditioning on the set $D^I_y$ of $y$--good intervals, we have that
\begin{align*}
\pr{ \sum_j \prcond{B_{j,y}}{\mathcal{F}_{j}} \geq 2A^{-1} t_{n,m} } \geq & \pr{M^{t,n,m}_y \geq \frac{t_{n,m}}{10L_1}}\\
& \quad \pr{ \sum_{j\in D^I_y} H^y_j \geq 2A^{-1} \frac{t_{n,m}}{L_1} \bigg \vert M^t_y \geq \frac{t_{n,m}}{10L_1}},
\end{align*}
where $M^{t,n,m}_y $ is the number of $y$--good intervals by time $t_{n,m}$.
Therefore, Lemma \ref{prob} combined with the properties of the binomial distribution and an appropriate choice of $A$ give that
\begin{equation}
\label{y12} \pr{ \sum_j \prcond{B_{j,y}}{\mathcal{F}_{j}} \geq 2A^{-1} t_{n,m} } \geq 1-e^{- D \frac{t_{n,m}}{L}}e^{-c},
\end{equation}
where $D>2$ is an appropriate constant. Combining \eqref{MM12}, \eqref{M12} and \eqref{y12}, we get that
\begin{align*}
\pr{B^c_y} \leq \left( e^{- D \frac{t_{n,m}}{L}}+ \frac{1}{m^{dn}} \right) e^{-c}.
\end{align*}
Given $B_y$, we have that $A^t_{y,x}$ can be coupled with a Poisson random variable with mean $A^{-1}t_{n,m}$, so that
\[\pr{\tilde{B}_{y}^c} \leq \prcond{\tilde{B}_{y}^c}{B_y} +\pr{B_y^c
}\]
Using the tails of a Poisson random variable, we have that
\begin{align*}
\pr{\tilde{B}_{y}^c} & \leq (e/2)^{-(2A)^{-1} t_{n,m}} + \left( e^{- D \frac{t_{n,m}}{L}}+ \frac{1}{m^{dn}} \right) e^{-c} \\
& \leq \left( \frac{1}{m^{\tilde{D}n}} + e^{- D \frac{t_{n,m}}{L}}+ \frac{1}{m^{dn}} \right) e^{-c},
\end{align*}
which finishes the proof.
\end{proof}
\subsection{Inducing the walk to a smaller dimension}
Let $I \in \lbrace 2,\ldots,n \rbrace$ and let $Z_y^t=X_ty$. In this section, we develop the necessary tools to study the walk if we only focus on the top $I$ coordinates of $Z_y^t$.
Let $ s_1 <s_2< \ldots \leq t_{n,m}$ denote the times when the $I$--th clock rings and $Z_y^{s_j}(I) \neq 0$. Let $z_1< z_2<\ldots \leq t_{n,m}$ denote the times when a clock other than the $I$--th one rings, let $W_j$ be the corresponding operation matrix applied and $L(t)= \max \lbrace j \geq 0 :z_{j} \leq t \rbrace$. For $0\leq t \leq t_{n,m}$, we define the the backwards process by $Y_0=I_n$ and
\begin{equation}
Y_t= \prod_{j=0}^{L(t_{n,m} ) - L(t- t_{n,m} ) -1}W_{L(t_{n,m})-j} =W_{L(t_{n,m})}W_{L(t_{n,m})-1}\ldots W_{L(t_{n,m}-t)+1}
\end{equation}
and for $0\leq t'< t \leq t_{n,m}$ we let
\begin{equation}
Y_{t',t}= Y_{t_{n,m}-t}^{-1} Y_{t_{n,m}-t'}=W_{L(t)}\ldots W_{L(t')+1}.
\end{equation}
Notice that the entries of $Y_t$ and $Y_{t,t'}$ that fall on the $[1,I-1] \times [I,n]$ box are equal to zero and that $Y_t$ is a Markov chain on the columns of a matrix on $G_n(m)$.
\begin{lemma}\label{vector}
We have that
\begin{align*}
Z_y^{s_{\ell}} & = Y_{0, s_{\ell} } Z_y^{0} + \sum_{k=1}^{\ell-1} a_k Y_{s_k, s_{\ell} } E(I-1,I) Y_{ 0, s_{k}} Z_y^{0} \quad + a_{\ell} E(I-1,I) Y_{ 0,s_{\ell} } Z_y^{0},
\end{align*}
where $a_k \in \{ \pm 1 \}$ are the random signs corresponding to the $k$--th time the $I$--th clock rings.
\end{lemma}
\begin{proof}
We will prove the statement by induction. For $\ell=0$ both sides are equal to the identity $Z_y^{0}$. By the definition of $s_{\ell+1}$ we have that
\begin{align}
\label{defi}Z^{s_{\ell +1}}_y & = \left( I_n + a_{\ell+1} E(I-1,I) \right) Y_{ s_{\ell}, {s_{\ell +1}}} Z^{s_{\ell }}_y
\end{align}
By the induction hypothesis we have that
\begin{align}
\eqref{defi} & = \left( I_n + a_{\ell +1} E(I-1,I) \right) Y_{ s_{\ell}, {s_{\ell +1}}} \biggl( Y_{0, s_{\ell} } Z_y^{0}+ a_{\ell} E(I-1,I) Y_{0, s_{\ell} } Z_y^{0} \cr
\label{special}&\quad\quad + \sum_{k=1}^{\ell} a_k Y_{ s_k,s_{\ell}} E(I-1,I) Y_{0, s_k} Z_y^{0} \biggr)
\end{align}
Using the facts that $ E(I-1,I) E(I-1,I)= 0$ and $ E(I-1,I)Y E(I-1,I)= 0$ for every $Y \in G$ whose $[I-1] \times [ I,n]$ entries are zero.
\begin{align}
\eqref{special}& = Y_{0, s_{\ell+1} } Z_y^{0}+ a_{\ell +1} E(I-1,I) Y_{0, s_{\ell+1} } Z_y^{0} +\sum_{k=1}^{\ell-1} a_k Y_{s_k, s_{\ell+1}} E(I-1,I) Y_{0, s_{k} } Z_y^{0} \cr
&\label{a} \quad + a_{\ell} Y_{s_{\ell},s_{\ell +1}} E(I-1,I) Y_{0, s_{\ell+1} } Z_y^{0},
\end{align}
which finishes the proof.
\end{proof}
Since we are interested in $Z_y^{t}(2)$, we write a similar version of Lemma \ref{vector}. Using the fact that $Z_y^{t} = Y_{s_{\ell},t } Z_y^{s_{\ell}},$ we get the following.
\begin{corollary}\label{coordinate}
We have that
\begin{align*}
Z_y^{t} &=Y_{0,t } Z_y^{0}+ \sum_{k=1}^{\ell-1} a_k Y_{s_k, t} E(I-1,I) Y_{0, s_{k}} Z_y^{0} \quad + a_{\ell} E(I-1,I) Y_{0,t} Z_y^{0}
\end{align*}
and
$$Z_y^{t}(2) = Y_{0, t } Z_y^{0}(2)+ \sum_{k=1}^{\ell-1} a_k Y_{s_k, t} E(I-1,I) Y_{0, s_{k} } Z_y^{0}(2).$$
\end{corollary}
To study $Y_{0, t } Z_y^{0}(2)$, consider a vector process starting at $ y_{I} e_{I}$ and having the same updates as the original process. Then $Y_{0, t } Z_y^{0}(2)$ is the second coordinate of this process.
Similarly, $Y_{s_k, t } E(I-1,I) Y_{0, s_{k} } Z_y^{0}(2)$ is the same as the second coordinate of the vector process that starts at $y_{I}^{s_k} e_{I-1}$, where $y_{i}^{s_k}$ is the $I$--th coordinate of $Z_y^{s_k}$, and whose updates from are the same as the updates that occur between times $s_k$ and $t$.
Let $\mathcal{G}_{j}^y$ be the event $\{[t_j, t_{j+1}] \in D^I_y \}$ for the $I$ that was chosen. The following lemma introduces a condition under which $\vert Z_y^{t}(2) \vert$ is guaranteed to be big. This will be crucial to proving that the eigenvalues of the walk are sufficiently small for a constant fraction of the time.
\begin{lemma}\label{de20}
Let $s_{1}$ is the first time after $t_{j}$ that the $I$--th clock rings and $Z_y^{s_{1}}(I) \neq 0$ and let $t\in [t_{j+1}, t_{j+2}]$. When $s_1 \leq t_{j+1}$, set $N^t= Y_{s_{1},t } E(I-1,I) Y_{0, s_{1} } Z_y^{0}(2)$. For every $x \in \{0, \ldots m/4\}$, we have that
\begin{equation}\label{DiaHough2}
\prcond{ \vert Z_y^{t}(2) \vert \geq x }{\mathcal{G}_{j}^y} \geq \frac{1}{2}\prcond{ \vert2 N^t\vert \geq 2x }{\mathcal{G}^y_{j}}1_{ \{s_1 \leq t_{j+1} \}}.
\end{equation}
\end{lemma}
\begin{proof}
Lemma \ref{nstar} says that $ s_{1}/13$ can be coupled with an exponential random variable with mean one. Let $\mathcal{Y}_t$ be the event that $ \vert2 N^t\vert \geq 2x$. We have that
\begin{align}
&\prcond{ \vert Z_y^{t}(2) \vert \geq x }{\mathcal{G}^y_{j}} \label{la2} \geq \prcond{ \vert Z_y^{t}(2) \vert \geq x }{\mathcal{Y}_t, \mathcal{G}^y_j} \prcond{ \mathcal{Y}_t }{\mathcal{G}^y_{j}}1_{ \{s_1 \leq t_{j+1}\}}
\end{align}
The condition $\{ \vert2 N^t\vert \geq 2x \} $ combined with the fact that $a_1=1$ or $\break a_1=-1$ with probability $1/2$ will result in $\vert Z_y^{t}(2) \vert \geq x$ with probability $1/2$. Therefore,
\begin{align*}
\eqref{la2}& \geq \frac{1}{2}\prcond{ \vert 2 N^t \vert \geq 2x }{\mathcal{G}^y_{j}} 1_{ \{s_1 \leq t_{j+1} \}}.
\end{align*}
\end{proof}
\section{The case where $m$ is a prime}\label{prime}
In this section $m$ is a prime number. Since the case $m=2$ was covered in \cite{PeresSly}, from now on in this section $m$ will denote an odd prime. We now write the Diaconis--Hough lemma, which we are going to use for the proof of Theorem \ref{mainc}.
\begin{lemma}[Theorem 3, \cite{Hough}]\label{DH}
Let $Z_t$ be the configuration of the rightmost corner of the upper triangular random walk at time $t$. For any set $A \subset \mathbb{Z}/ m \mathbb{Z}$ we have that
$$\Vert \pr{Z_t \in \cdot} - U\Vert_{T.V.} \leq \exp(-r t 2^{-n} m^{-\frac{2}{n-1}}),$$
where $U$ is the uniform measure on $G$ and $r
$ is a universal constant.
\end{lemma}
Diaconis and Hough mainly treat the case where $n$ is fixed. Therefore, the term $2^n$ is not announced in their main result, but can be found in the proof of Proposition 22 of \cite{Hough}.
\subsection{The eigenvalues for $y \in W_I$}
Let $I=2+J$, where $J= \sqrt{\log m}$, so that $2^J \leq m^{\frac{2}{J}}$. Recall that $W_I= \left( \mathbb{Z}/ m \mathbb{Z}\right)^{n-1} \setminus \langle e_1, \ldots, e_{I-2} \rangle$ and let $y \in W_I$. The goal of this section is to prove that $\vert Z_y^{t}(2) \vert \geq \frac{m}{8}$ for a constant fraction of $[0,t_{n,m}]$.
We choose the length of each interval $[t_j, t_{j+1}]$ to be $L= d m^{\frac{4}{J }},$ where $d>0$ is a suitable constant that makes $L$ big enough for Lemma \ref{prob} to work.
Let $T_I$ be the first time that $Z_y^{t}(I) \neq 0$. Let $\mathcal{G}_{j,I}^y$ be the event $\{[t_j, t_{j+1}] \in D^I_y \}$ for the $I$ that was chosen, but for simplicity we will write $\mathcal{G}^y_{j}$. The following lemma is the main tool for proving the Theorem \ref{mainc}.
\begin{lemma}\label{de}
For $y \in W_I$, we have that
\begin{equation}
\prcond{ \vert Z_y^{t}(2) \vert \geq \frac{m}{8} }{\mathcal{G}^y_{j}} \geq \frac{1}{64},
\end{equation}
for every $t \geq t_{j+1}$.
\end{lemma}
\begin{proof}
We just need to check that the assumptions of Lemma \ref{de20} hold.
To bound $\prcond{ \vert2 N^t\vert \geq \frac{m}{4} }{\mathcal{G}^y_{j}}$ from below, we use Lemma \ref{DH}, to get that
$$\prcond{ \vert 2 N^t\vert \geq \frac{m}{4} }{\mathcal{G}^y_{j} } \geq \frac{1}{8}- e^{-r \frac{L}{2^{I}m^{2J^{-1}}}}.$$
By our choice of $L$ and $J$, we have that
\begin{equation}\label{DiaHough}
\prcond{ \vert 2 N^t\vert \geq \frac{m}{4} }{\mathcal{G}^y_{j}} \geq \frac{1}{16}.
\end{equation}
Also, our choice of $L$ gives that
\[\pr{s_1<t_{j+1}} \geq \frac{1}{2}.\]
Lemma \ref{de20} gives the desired result.
\end{proof}
The next lemma says that the event considered in Lemma \ref{de} holds for a constant fraction of the time. Let $S$ be an appropriately chosen constant, which is is uniform on $j,m, n$ and $y$.
\begin{lemma}\label{B1}
Let $B_{j,y}$ denote the event
\[
B_{j,y} :=\Big\{\int_{t_{j+1}}^{t_{j+2}}1_{ \{\vert Z^s_y(2) \vert > m/8 \} } ds \geq \frac{1}{3} L\Big \}.
\]
For $y \in W_I,$ we have that
$$\prcond{B_{j,y}} {\mathcal{G}^y_{j}} \geq \frac{1}{65},
$$
for every $j \geq 1$.
\end{lemma}
\begin{proof}
Let $R_j$ count the number of $s \in [t_{j+1}, t_{j+2}]$, such that $\vert Z^{s}_y(2) \vert \leq m/8$. So $R_j= \int_{t_{j+1}}^{ t_{j+2}} 1_{ \lbrace \vert Z^{s}_y(2) \vert \leq m/8 \vert \mathcal{G}^y_{j} \rbrace}ds $. Lemma \ref{de} gives that
\begin{align*}
&\expect{ R_j \vert \mathcal{G}^y_{j}}\leq \frac{63}{128} L.
\end{align*}
Markov's inequality then gives that
\begin{align*}
&\prcond{ R_j > \frac{4095}{8192} L }{\mathcal{G}^y_{j}} \leq \frac{64}{65}.\end{align*}
The constant $S$ is chosen so that
\begin{align*}
&\prcond{B_{j,y}} {\mathcal{G}_{j}} \geq \prcond{ R_j \leq \frac{4095}{8192} L }{\mathcal{G}^y_{j}} \geq \frac{1}{65}.
\end{align*}
\end{proof}
The rest of the eigenvalues will be studied in Section \ref{general}.
\section{The case where $m$ is not a prime}\label{general}
In this section, we study the quantity $Z_y^t(2)$ for the case where $m$ is not necessarily prime. We start by proving a lemma similar to Lemma \ref{DH} which works for $m \in \mathbb{N}$ that is not necessarily prime.
Let $J= \lfloor (\log m)^{1/3} \rfloor$ and $D=20(J+1)$. Let $A_{J,D,m}= 2m^{2/J}/ \log D$ and let $p$ be a prime such that $\frac{1}{20} A_{J,D,m} \leq 6 r^{-1}2^J p^{2/J} \leq \frac{1}{2} A_{J,D,m}$. Recall that $r$ is the constant from Lemma \ref{DH}. The goal of this section is to prove the following lemma.
\begin{lemma}\label{BH2}
Let $Z_t$ be the the last column of $X_t$. We have that there is a constant $K$ such that
\[\pr{\vert aZ_t(n-J-1)\vert > m e^{-K (\log m)^{2/3}} } \geq 1/2,\]
where $t \in [6r^{-1}2^Jp^{2/J}, A_{J,D,m} ]$ and $a \in \{1,\ldots, m-1\}$.
\end{lemma}
To prove Lemma \ref{BH2}, we will need the following lemma concerning $Z_t$ over~$\mathbb{Z}$.
\begin{lemma}\label{AH}
Let $\mathcal{Z}_t= X_t e_{n-1}$ be the column process over $\mathbb{Z}$ which starts at $(0,\ldots,1)^T$. Set $x=2m^{2/J}/ \log D$. Then, we have that
$$\pr{\max_{\substack{t \leq x,\\ 1\leq i \leq k}} \{ \vert \mathcal{Z}_t(n-i)\vert \} \leq x^{k/2} (\log D)^{k/2}} \geq 1- \frac{2k}{D} ,$$
for $k \leq J+1 \leq n-1$.
\end{lemma}
\begin{proof}
We will prove the result by induction on $k$. For $k=1$, the only possible $i \leq k$ is $i=1$. We have that $\mathcal{Z}_t(n-1)$ is a simple random walk on $\mathbb{Z}$. The reflection principle gives that
\begin{equation}\label{rp}
\pr{\max_{t \leq x} \{ \vert \mathcal{Z}_t(n-1)\vert \} \geq m^{1/J}} \leq \pr{ \vert \mathcal{Z}_{x}(n-1)\vert \geq m^{1/J} }.
\end{equation}
The Azuma-Hoeffding inequality gives that
\[\pr{ \vert \mathcal{Z}_{x}(n-1)\vert \geq m^{1/J} } \leq 2 e^{- \log L}= \frac{2}{D} \]
and therefore, in combination with \eqref{rp}, we get that
$$\pr{\max_{t \leq x} \{ \vert \mathcal{Z}_t(n-1)\vert \} \geq m^{1/J} } \leq 2 e^{- \log L}= \frac{2}{D}.$$
Let $\mathcal{A}_k$ be the event that $\{\max_{\substack{t \leq x,\\ i \leq k}} \{ \vert \mathcal{Z}_t(n-i)\vert \} \leq x^{k/2} (\log D)^{k/2} \}.$
Assume that $\pr{\mathcal{A}_k} \geq \left(1- \frac{2}{D} \right)^k$. Let $b_1\leq \ldots \leq t$ are the times that the $n-k$ clock rings.
Writing $\mathcal{Z}_y^{t}(n-k-1)= \sum a_i \mathcal{Z}_y^{s_i}(n-k)$, the Azuma-Hoeffding inequality gives that $\prcond{\mathcal{A}_{k+1}}{\mathcal{A}_k} \geq 1 - e^{- 2\frac{\log D}{2}}$.
Therefore
$$\pr{\mathcal{A}_{k+1}}\geq (1 - e^{- \log D}) \left(1- \frac{2}{L} \right)^k \geq \left(1- \frac{2}{D} \right)^{k+1}. $$
Using the fact that $\left(1- \frac{2}{L} \right)^{k} \geq 1- \frac{2k}{D}$, we complete the proof.
\end{proof}
Let $T$ be the first time that there is a $j \leq n- J-1$ satisfies $\vert \mathcal{Z}_t(j)\vert > m/6$. Lemma \ref{AH} says that w.h.p. $T >2 m^{2/J}/ \log D$.
Recall that $\frac{1}{20} A_{J,D,m} \leq 6 r^{-1}2^J p^{2/J} \leq \frac{1}{2} A_{J,D,m}$.
Let
\[\theta_k(t):= \max_{\substack{A \subset \mathbb{Z}/ m \mathbb{Z} \\ \vert A \vert \leq k}} \{\pr{Z_t(n-I) \in A}\}.\]
\begin{lemma}\label{important1}
For $t \in [6r^{-1}2^J p^{2/J}, A_{J,D,m} ]$, we have that
$$\theta_{p/3}(t) \leq 1/2.$$
\end{lemma}
\begin{proof}
Let $\Tilde{Z}_t$ be the process over $\mathbb{Z}/p \mathbb{Z}$ and let
$$\Tilde{\theta}_k(t):= \max_{\substack{A \subset \mathbb{Z}/ p \mathbb{Z} \\ \vert A \vert \leq k}} \{\pr{\Tilde{Z}_t(n-J-1) \in A}\}.$$
Since $p <m $, we notice that $\theta_{p/3}(t) \leq \Tilde{\theta}_{p/3}(t) + \pr{T \leq t}$.
If we assume that $\Tilde{\theta}_{p/3}(t) \geq 2/5 $ then there is a set $A \subset \mathbb{Z}/ p \mathbb{Z}$ with $\vert A \vert \leq p/3$ such that $$\pr{\Tilde{Z}_t(n-J-1) \in A}- \pi_p(A) \geq 1/15,$$
where $\pi_p$ is the uniform measure over $\mathbb{Z}/ p \mathbb{Z}$. This implies that
$$d_{T.V.}(\Tilde{Z}_t(n-J-1), \pi_p) \geq 1/15 ,$$
which contradicts the fact that $t \geq 6r^{-1}2^J p^{2/J}$, which means that Lemma \ref{DH} gives that
$$d_{T.V.}(\Tilde{Z}_t(n-J-1), \pi_p) \leq e^{-3} .$$
Therefore, $\theta_{p/3}(t) \leq \Tilde{\theta}_{p/3}(t) + 1/10\leq 1/2$.
\end{proof}
\begin{corollary}\label{important3}
For $t \in [6r^{-1}2^J p^{2/J}
, A_{J,D,m} ]$, we have that there is a universal constant $K$ such that
\[\pr{\vert Z_t(n-J-1)\vert > m e^{-K (\log m)^{2/3}} } \geq 1/2.\]
\end{corollary}
\begin{proof}
Lemma \ref{important1} gives that
\[\pr{\vert Z_t(n-J-1)\vert > p/6} \geq 1/2.\]
The fact that $\frac{1}{20} A_{J,D,m} \leq 6
r^{-1}2^J p^{2/J} $ gives that $$p \geq \frac{\tilde{r}^{J/2}m}{2^{J^2/2} (\log D)^{J/2}} \geq m e^{-\tilde{K} (\log m)^{2/3}},$$
where $\tilde{K}$ is a universal constant. This completes the proof.
\end{proof}
We are now ready to prove Lemma \ref{BH2}.
\begin{proof}[Proof of Lemma \ref{BH2}]
Let $a \in \mathbb{Z}$ and let
$$\theta^a_k(t):= \max_{\substack{A \subset \mathbb{Z}/ m \mathbb{Z} \\ \vert A \vert \leq k}} \{\pr{aZ_t(n-J-1) \in A}\}.$$
Let $g=g.c.d.(a,m) $. If $g=1$ then $\theta^a_k(t)=\theta_k(t)$ and the statement follows by Corollary \ref{important3}.
If $g \neq 1$ then let $m'= m/g$ and $a'=a/g$. Then we can view $aZ_t(n-J-1)$ as a random walk on $\mathbb{Z}/m' \mathbb{Z}$. We denote this random walk by $\overline{Z}_t$.
If $m' \gg 1,$ then let $p'$ be such that $\frac{1}{20} A_{J,D,m'} \leq 6r^{-1}2^J (p')^{2/J} \leq \frac{1}{2} A_{J,D,m'}$. We therefore have that there exists a constant $r'$, that does not depend on $m'$, such that $p \geq 6^{-1} \tilde{r}^{J/2} 2^{-J^2/2}D^{-J/2} (\log D)^{-J/2}m' $. Corollary \ref{important3} says that
\[\pr{\vert a' \overline{Z}_t)\vert > p'/6} \geq 1/2.\]
and therefore
\[\pr{\vert aZ_t(n-J-1)\vert > (gp')/6} \geq 1/2.\]
Using the fact that $gp' \geq 6^{-1} \tilde{r}^{J/2} 2^{-J^2/2}D^{-J/2} (\log D)^{-J/2}m $ we get that
\[\pr{\vert aZ_t(n-J-1)\vert >6^{-1} \tilde{r}^{J/2} 2^{-J^2/2}D^{-J/2} (\log D)^{-J/2}m } \geq 1/2.\]
If $m'\ll 1$, then the walk $aZ_t(n-J-1)$ on $\mathbb{Z}/m' \mathbb{Z}$ mixes in finite steps. Similarly to before, we have that
\[\pr{\vert aZ_t(n-J-1)\vert > 6^{-1} \tilde{r}^{J/2} 2^{-J^2/2}D^{-J/2} (\log D)^{-J/2}m } \geq 1/2.\]
Using the fact that there is a universal constant $K$ such that
\[6^{-1} \tilde{r}^{J/2} 2^{-J^2/2}D^{-J/2} (\log D)^{-J/2}m \geq m e^{-K (\log m)^{2/3}},\]
we conclude the proof.
\end{proof}
\subsection{The eigenvalues $y\in W_I$}
We are now going to consider the decomposition proved in Lemma \ref{coordinate}. For the definition of the $y$--good intervals, we are going to consider $I=1+J$, where $J= \lfloor (\log m)^{1/3} \rfloor$. Let $L_1=d A_{J,D,m}$ be the length of each $y$--good interval, where $d$ is a suitable constant.
Let $W_I=\left( \mathbb{Z}/ m \mathbb{Z}\right)^{n-1} \setminus \langle e_1, \ldots, e_{I-2} \rangle$. The following lemma is one of the main tools for proving Lemma \ref{q}.
\begin{lemma}\label{deg}
For $y \in W_I$, we have that
\begin{equation}
\prcond{ \vert Z_y^{t}(2) \vert \geq m e^{-K (\log m)^{2/3}} }{\mathcal{G}^y_{j}} \geq \frac{1}{8},
\end{equation}
for every $t \geq t_{j+1}$.
\end{lemma}
\begin{proof}
We just need to check that the assumptions of Lemma \ref{de20} hold. Using the following decomposition
$$Z_y^{t}(2) = Y_{0, t } Z_y^{0}(2)+ \sum_{k=1}^{\ell-1} a_k Y_{s_k, t} E(I-1,I) Y_{0, s_{k} } Z_y^{0}(2),$$
as presented in Corollary \ref{coordinate}, we notice that $N_t= Y_{s_1, t} E(I-1,I) Y_{0, s_{1} } Z_y^{0}(2)$ has the form $a Z_t(n-J-1)$.
By our choice of $L$ and $J$, Lemma \ref{BH2} gives that
$$\prcond{ \vert 2 N^t\vert \geq m e^{-K (\log m)^{2/3}} }{\mathcal{G}^y_{j} } \geq \frac{1}{2},$$
and
\[\pr{s_1 \leq t_{j+1}} \geq 1/2.\]
\end{proof}
\begin{lemma}\label{B12}
Let $B_{j,y}$ denote the event that $\vert Z^s_y(2) \vert > m e^{-K (\log m)^{2/3}}$ for at least one third of
$ [ t_{j+1}, t_{j+2}]$. For $y \in W_I ,$ we have that
$$\prcond{B_{j,y}} {\mathcal{G}_{j}} \geq \frac{1}{2},$$
for every $j \geq 1$.
\end{lemma}
\begin{proof}
Let $R_j$ count the number of $s \in [t_{j+1}, t_{j+2}]$, such that $\vert Z^{s}_y(2) \vert \leq m e^{-K (\log m)^{2/3}}.$ So $R_j= \int_{t_{j+1}}^{ t_{j+2}} 1_{ \lbrace \vert Z^{s}_y(2) \vert \leq m e^{-K (\log m)^{2/3}} \vert \mathcal{G}^y_{j} \rbrace}ds $. Lemma \ref{deg} gives that
\begin{align*}
&\expect{ R_j \vert \mathcal{G}^y_{j}}\leq \frac{1}{8} L.
\end{align*}
Markov's inequality then gives that
\begin{align*}
&\prcond{ R_j > \frac{1}{3} L }{\mathcal{G}^y_{j}} \leq \frac{3}{8}.\end{align*}
The constant $S$ is chosen so that
\begin{align*}
&\prcond{B_{j,y}} {\mathcal{G}_{j}} \geq \prcond{ R_j \leq \frac{1}{3} L }{\mathcal{G}^y_{j}} \geq \frac{5}{8}.
\end{align*}
\end{proof}
\begin{lemma}\label{r1}
Recall that $P^t$ is the indicator function that the clock of the second row rings at time $t$ and $A^t_{y,x}= \int_0^t 1_{\lbrace P^s=1 \rbrace } 1_{\lbrace \vert Z_y^s(2) \vert > x\rbrace } ds$. Let $x= m e^{-K (\log m)^{2/3}}$. For $y \in W_I$,
Consider $\tilde{B}_{y}= \lbrace A^t_{y,x}
\geq (2A)^{-1} t_{n,m} \rbrace $. We have that
\begin{align*}
\pr{\tilde{B}_{y}} \geq 1- \frac{b}{m^{gn}} e^{-c},
\end{align*}
where $b,g$ are suitable constants.
\end{lemma}
\begin{proof}
We have that there is a constant $R$ such that $\frac{t_{n,m}}{L_1} \geq R n \log m$. Lemmas \ref{r15} and \ref{B12} give the desired result.
\end{proof}
\subsection{The eigenvalues $y\in \langle e_1, \ldots, e_{I-2} \rangle \setminus \langle e_1 , e_2\rangle $}
Let $Q_I= \langle e_1, \ldots, e_{I-2} \rangle \setminus \langle e_1 , e_2\rangle $. Adjusting the proof of Corollary \ref{important3} for $J=3$ we get the following result.
\begin{corollary}\label{important4}
For $t \in [48r^{-1} p^{2/3}
, A_{3,D,m} ]$, we have that there is a universal constant $\tilde{K}$ such that
\[\pr{\vert Z_t(n-4)\vert > m / \tilde{K}} \geq 1/2.\]
\end{corollary}
Let $L_2
=d m$ be the length of each $y$--good interval, where $d$ is a suitable constant.
\begin{lemma}\label{deg2}
For $y \in Q_I$, we have that
\begin{equation}
\prcond{ \vert Z_y^{t}(2) \vert \geq m /\tilde{K} }{\mathcal{G}^y_{j}} \geq \frac{1}{8},
\end{equation}
for every $t \geq t_{j+1}$.
\end{lemma}
The proof of Lemma \ref{deg2} is similar to the proof of Lemma \ref{deg}. Similarly to Lemma \ref{B12}, we have the following lemma.
\begin{lemma}\label{B13}
Let $B_{j,y}$ denote the event that $\vert Z^s_y(2) \vert > m / \tilde{K}$ for at least one third of
$ [ t_{j+1}, t_{j+2}]$. For $y \in Q_I ,$ we have that
$$\prcond{B_{j,y}} {\mathcal{G}_{j}} \geq \frac{1}{2},$$
for every $j \geq 1$.
\end{lemma}
\begin{lemma}\label{r2}
Recall that $P^t$ is the indicator function that the clock of the second row rings at time $t$ and $A^t_{y,x}= \int_0^t 1_{\lbrace P^s=1 \rbrace } 1_{\lbrace \vert Z_y^s(2) \vert > x\rbrace } ds$. Let $x=m / \tilde{K}.$ For $y \in Q_I$,
Consider $\tilde{D}_{y}= \lbrace A^t_{y,x}
\geq (2A)^{-1} t_{n,m} \rbrace $. We have that
\begin{align*}
\pr{\tilde{D}_{y}} \geq 1- b \left( \frac{1}{n^{gm}}+ \frac{1}{m^{gn}}\right) e^{-c},
\end{align*}
where $b,g$ are suitable constants.
\end{lemma}
\begin{proof}
We have that there is a constant $R$ such that $\frac{t_{n,m}}{L_2} \geq R m \log n$. Lemmas \ref{r15} and \ref{B13} give the desired result.
\end{proof}
\subsection{The eigenvalues $y\in \langle e_1,e_2 \rangle \setminus \langle e_1 \rangle $}
Let $P_2= \langle e_1,e_2 \rangle \setminus \langle e_1 \rangle $. For $y \in P_2$, we write $y= ae_1 +b e_2$ with $b \neq 0$. Therefore, we observe that $Z_y^t(2)= a+ b S^t$, where $S^t$ is a simple random walk on the cycle starting at zero. In this section, we consider the length of the intervals to be $L_3= \delta \log m$, where $\delta $ is a suitable constant.
\begin{lemma}
Let $\mathcal{I }\subset \mathbb{Z} / m \mathbb{Z}$ with $\vert \mathcal{I} \vert= \sqrt{\log m}$. For every $y \in P_2$ and $t\geq t_{j+1}$, we have that
\[\pr{ Z_y^t(2)-Z_y^{t_j}(2) \notin \mathcal{I} } \geq 1/2 .\]
\end{lemma}
\begin{proof}
Writing $y= ae_1 +b e_2$, we have that $Z_y^t(2)-Z_y^{t_j}(2)= b( S^t- S^{t_j})$.
Assume that
\begin{equation}\label{opp}
\pr{ b(S^t- S^{t_j}) \notin \mathcal{I} } < 1/2 .
\end{equation}
Let $Q$ be the transition matrix of $b S$. We have that
\begin{align}
\label{qel2}\Vert Q^{t-t_j}- \pi \Vert_2^2 &= \sum_z \frac{1}{m}\vert m Q^{t-t_j}_0(z)-1 \vert^2 \cr
&= \sum_z m ( Q^{t-t_j} (z))^2-1 \cr
& \geq \sum_{z \in \mathcal{I} } m ( Q^{t-t_j} (z))^2-1
\end{align}
Cauchy Schwartz leads to
\begin{align}
\eqref{qel2} & \geq \frac{ m}{\vert \mathcal{I} \vert} \left( \sum_{ z \in \mathcal{I} } Q^{t-t_j} (z)\right)^2-1 \cr
& \label{f12} \geq \frac{ m}{4\vert \mathcal{I} \vert}-1\\
& \label{dr} = \frac{ m}{4\sqrt{\log m}}-1,
\end{align}
where \eqref{f12} occurs by applying \eqref{opp}. Let $g$ be the gcd of $b$ amd $m$ Given that $bS^t$ can be viewed as SRW on $\mathbb{Z}/ g \mathbb{Z}$, we have that
\begin{align}
\Vert Q^{t-t_j}- \pi \Vert_2^2 & \leq \sum_{y=1}^{m/g -1} e^{ -2 \left( t- \sum_{i=1}^t \cos \frac{2 \pi g y }{m} \right) }\cr
& \label{expl} \leq \frac{m}{g} e^{-2t} + \frac{\sqrt{3} m}{ 2g\sqrt{2 \pi (t-t_j)}}\\
&\label{he} \leq \frac{m}{g} e^{-2L_3} + \frac{\sqrt{3} m}{ 2g\sqrt{2 \pi L_3}},
\end{align}
where \eqref{expl} is a straightforward application of Lemma \ref{integral}.
Equation \eqref{he} contradicts \eqref{dr} for a suitable choice of the constant $\delta$ and this completes the proof.
\end{proof}
This implies the following corollary.
\begin{corollary}
For every $y \in P_2$ and $t\geq t_{j+1}$, we have that
\[\pr{\vert Z_y^t(2)\vert \geq \sqrt{ \log m}/2} \geq 1/2 .\]
\end{corollary}
We follow the reasoning of the previous section to conclude the following Lemmas.
\begin{lemma}\label{B23}
For $y \in P_2$, let $B_{j,y}$ denote the event that $\vert Z^s_y(2) \vert > \sqrt{ \log m}/2$ for at least one third of
$ [ t_{j+1}, t_{j+2}]$, where $\beta$ is a suitable constant. For $y \in P_2 ,$ we have that
$$\prcond{B_{j,y}} {\mathcal{G}_{3}} \geq \frac{1}{2},$$
for every $j \geq 1$.
\end{lemma}
\begin{lemma}\label{r4}
Recall that $P^t$ is the indicator function that the clock of the second row rings at time $t$ and $A^t_{y,x}= \int_0^t 1_{\lbrace P^s=1 \rbrace } 1_{\lbrace \vert Z_y^s(2) \vert > x\rbrace } ds$ and let $x=\sqrt{\log m}/2$. For $y \in P_2$,
Consider $\tilde{D}_{y}= \lbrace A^t_{y,x}
\geq (2A)^{-1} t_{n,m} \rbrace $. We have that
\begin{align*}
\pr{\tilde{D}_{y}} \geq 1- \frac{b}{m^{gn}} e^{-c},
\end{align*}
where $b,g$ are suitable constants.
\end{lemma}
\begin{proof}
We have that there is a constant $R$ such that $\frac{t_{n,m}}{L_3} \geq R m \log n$. Lemmas \ref{r15} and \ref{B12} give the desired result.
\end{proof}
\section{The proof of Lemma \ref{q}.}
\begin{proof}[Proof of Lemma \ref{q}]
We will first consider the case where $m$ is not prime. Let $x=m e^{-K (\log m)^{2/3}},$ $w=m / \tilde{K} $ and $I=1+J$, where $J= \lfloor (\log m)^{1/3} \rfloor$. We also consider
$$t_{n,m}= D (m^2 \log n + n e^{C (\log m)^{2/3}}) + cnm^2 \log \log n,$$
which satisfies
\begin{equation}\label{fb}
t_{n,m} \geq D n L_1 e^{K (\log m)^{2/3}} \log m + cnm^2 \log \log n
\end{equation}
and
\begin{equation}\label{sb}
t_{n,m} \geq D m^2 n \log n + cnm^2 \log \log n \geq D (\log m)^{4/3}+ cnm^2 \log \log n.
\end{equation}
Looking at \eqref{y}, given that $k, w_1, \ldots w_k$ are such so that $E_{t_{n,m},x,w}$ is satisfied, we have that for every $y \in W_I$ it is the case that $\vert Z_y^t(2) \vert \geq m e^{-K (\log m)^{2/3}}$ for at least $A^{-1}t_{n,m}$ steps. Then, \eqref{y} says that
\begin{align*}
4 \Vert q^{t_{n,m}}_{\substack{k,w_1, \ldots w_k}}-u \Vert_{T.V.}^2 & \leq
\sum_{y \neq 0} e^{ -2 \left( k- \sum_{i=1,\ldots k} \cos \frac{2 \pi (y^T w_i(2))}{m} \right) } .
\end{align*}
The definition of $E_{t_{n,m},x,w}$ gives that
\begin{align}
4 \Vert q^{t_{n,m}}_{\substack{k,w_1, \ldots w_k}}-u \Vert_{T.V.}^2 & \leq \sum_{\substack{y \neq 0 \cr y \in \langle e_1 \rangle}} e^{ -2 \left( k- \sum_{i=1}^k \cos \frac{2 \pi y }{m} \right) } + \sum_{ y \in P_2} e^{ -2 k\left(1- \cos \frac{ \tilde{\beta}\sqrt{ \log m} }{ m} \right) } +\cr
& \quad \sum_{y \in W_I} e^{ -2 k\left(1- \cos \left(2 \pi e^{-K (\log m)^{2/3}} \right)\right)} +\sum_{y \in Q_I} e^{ -2 k\left(1- \cos \left(2 \pi/ \tilde{K} \right)\right)} \cr
\label{30}
\end{align}
Equation \eqref{spl} gives that
\begin{align}
\eqref{30}
& \leq m e^{-2k} + 2 \sum_{j=1}^{\infty} e^{- \frac{4j^2 \pi^2}{ m^2} k }
+ m^2 e^{ -2 k \frac{ \overline{\beta} \log m}{ m} } + m^n e^{ -\tilde{C}k e^{-K (\log m)^{2/3}} } + m^{I } e^{-K'k }
\end{align}
Since $k \geq A^{-1}t_{n,m}$ and choosing $D$ to be a suitable constant, we have that there is a constant $B$ such that
\begin{align}\label{mnk}
&\eqref{30} \leq B \frac{1}{n (\log n)^c}.
\end{align}
Combining \eqref{l} and \eqref{mnk}, we have that there is a universal, positive constant $D$, such that
\begin{align*}
&\mathbb{E}[ \Vert q^{t_{n,m}}-u \Vert_{T.V.} ] \leq D \frac{1}{n (\log n)^c}.
\end{align*}
For $m$ prime, we make a choice of $I $ that allows us to prove a sharper result.
Set $I= 2 + \sqrt{ \log m}$, $x= m/8 $ and $w= m/ \tilde{K}$.
To prove part (a), we assume that $E_{t_{n,m},I, x,w}$ is satisfied for a universal constant $A$ that will be determined later in the proof.
For the case of $m$ prime, we have
\begin{align}
4 \Vert q^{t_{n,m}}_{\substack{k,w_1, \ldots w_k}}-u \Vert_{T.V.}^2 & \leq \sum_{\substack{y \neq 0 \cr y \in \langle e_1 \rangle}} e^{ -2 \left( k- \sum_{i=1}^k \cos \frac{2 \pi y }{m} \right) } +\sum_{ y \in P_2} e^{ -2 k\left(1- \cos \frac{\beta \sqrt{\log m} }{ m} \right) } + \cr
& \label{tv} \quad + \sum_{ y \in W_I} e^{ -2 k\left(1- \cos \frac{ \pi }{4} \right) } + \sum_{y \in Q_I} e^{ -2 \left( k- \sum_{i=1}^k \cos \frac{2 \pi }{\tilde{K}} \right) }
\end{align}
Equation \eqref{spl} gives that
\begin{align}
\eqref{tv} & \leq m e^{-2k} +2 \sum_{j=1}^{m/4} e^{- \frac{4j^2 \pi^2}{ m^2} k } + m^2 e^{ -2 k \frac{ \overline{\beta} \log m}{ m} } \cr
&+ m^n e^{ -k\left(2- \sqrt{2} \right) } + m^{I} e^{ -2 k\left(1- \cos \frac{2 \pi \sqrt{\delta} \log m }{m} \right) }
\cr
& \leq m e^{-2k} + 2 \sum_{j=1}^{\infty} e^{- \frac{4j^2 \pi^2}{ m^2} k }
+ m^n e^{ -k\left(2- \sqrt{2} \right) }
+ m^{I} e^{ - \frac{2k \pi^2 }{\tilde{K}^2} }
\end{align}
Since $k \geq A^{-1}t_{n,m}$ and choosing $D$ to be a suitable constant, we have that there is a constant $B$ such that
\begin{align}\label{mnk}
&\eqref{tv} \leq B \frac{1}{n (\log n)^c}.
\end{align}
Combining \eqref{l} and \eqref{mnk}, we have that there is a universal, positive constant $D$, such that
\begin{align*}
&\mathbb{E}[ \Vert q^{t_{n,m}}-u \Vert_{T.V.} ] \leq D \frac{1}{n (\log n)^c}.
\end{align*}
For the second part of Lemma \ref{q}, we will only focus on the case of general $m$, since the case $m$ prime follows the same outline. Lemmas \ref{r1}, \ref{r2}, \ref{r4} and a union bound gives
\begin{align*}
\pr{E^c_{t_{n,m},x,w}} & \leq \pr{ \cup_{y \in W_I} \tilde{B}_{y}^c} + \pr{ \cup_{y \in Q_I} \tilde{D}_{y}^c}+ \pr{ \cup_{y \in P_2} \tilde{D}_{y}^c} \\
&\leq m^n b e^{-c} \frac{1}{m^{g n}} + m^I b e^{-c} \left(\frac{1}{m^{g n}} + \frac{1}{n^{g m}} \right)\\
& \leq \tilde{b} e^{-c} \left(\frac{1}{m^{\tilde{g} n}} + \frac{1}{n^{\tilde{g} m}} \right),
\end{align*}
which finishes the proof.
\end{proof}
\bibliographystyle{plain}
| ef1edbd89415b845fd5b396f79457b508a7433d7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Quantum spin liquids (QSLs) represent a novel state of matter in which spins are highly entangled, but do not order nor freeze even in the zero-temperature limit~\cite{Anderson1973153,nature464_199}. Such a state does not involve any spontaneous symmetry breaking, which is beyond Landau's paradigm for a phase and the associated transition~\cite{imai2016quantum}. They are proposed to host fractional excitations and emergent gauge structures, and thus are promising candidates for quantum computation~\cite{Kitaev20032,Barkeshli722}. Furthermore, high-temperature superconductivity may emerge from carrier doping a QSL~\cite{anderson1,Baskaran1987973,lee:17}. Thus, it has been a long-sought goal to achieve the QSL state. However, spins often tend to order at low temperatures~\cite{Neel1985}. One approach to the goal is to introduce geometrical frustration into a low-spin system to enhance quantum fluctuations, so magnetic exchange interactions cannot be satisfied simultaneously among different lattice sites and the static magnetic order is prohibited~\cite{Anderson1973153,arms24_453}. By now, a number of QSL candidates resulting from geometrical frustrations have been proposed and explored experimentally~\cite{0034-4885-80-1-016502,RevModPhys.89.025003,Wen2019Experimental}, and some typical examples include organic triangular-lattice systems $\kappa$-(ET)$_2$Cu$_2$(CN)$_3$~\cite{PhysRevLett.91.107001,PhysRevLett.95.177001,Ohira2006,np4_459,np5_44,Furukawa2018Quasi} and EtMe$_3$Sb[Pd(dmit)$_2$]$_2$~\cite{np6_673,Yamashita1246,nc2_275}, kagom\'e-lattice compound ZnCu$_3$(OH)$_6$Cl$_2$~\cite{prl98_107204,nature492_406,RevModPhys.88.041002}, inorganic triangular-lattice compound YbMgGaO$_4$~\cite{sr5_16419,prl115_167203,np13_117,nature540_559,PhysRevLett.117.097201} and its sister compound YbZnGaO$_4$~\cite{PhysRevLett.120.087201}, and more recently found triangular-lattice system delafossites~\cite{Liu_2018,PhysRevB.98.220409,np15_1058,PhysRevB.100.144432,PhysRevB.100.241116,PhysRevB.100.224417,doi:10.1021/acsmaterialslett.9b00464,PhysRevMaterials.3.114413,PhysRevB.100.220407}. The disorder-free delafossites with effective spin-1/2 moments provide an excellent platform to unveil the QSL nature in a clean system. However, for most of these QSL candidates, the magnetic or nonmagnetic disorder can be significant, and complicates the interpretation of the intrinsic physics of the investigated systems~\cite{PhysRevB.94.060409,PhysRevLett.119.157201,PhysRevB.97.184413,np13_117,PhysRevLett.118.107202,PhysRevLett.118.087203,PhysRevLett.120.087201,PhysRevLett.120.207203,nc_Itamar,PhysRevX.8.031001,PhysRevX.8.031028,PhysRevX.8.041040,PhysRevLett.123.087201,PhysRevB.96.174418,PhysRevLett.115.077001}. For this reason, how disorder affects the QSL state is still a controversial issue.
Recently, $RE_3$Sb$_3M_2$O$_{14}$ as a new family of two-dimensional kagom\'e-lattice compounds were synthesized~\cite{doi:10.1002/pssb.201600256,Sanders2016RE,PhysRevB.95.104439}, where $RE^{3+}$ represents rare-earth ions and $M^{2+}$ denotes nonmagnetic Zn$^{2+}$ or Mg$^{2+}$ ions. Among these compounds, Tm$_3$Sb$_3$Zn$_2$O$_{14}$ was proposed to be a QSL~\cite{PhysRevB.98.174404}. The Tm$^{3+}$ ion has an electron configuration of $4f^{12}$ and, according to the Hund's rule, it has a total angular momentum of $J = 6$ with a 13-fold degeneracy. Considering the crystal-electric-field~(CEF) effect, the degeneracy of $J$ will be lifted, which was suggested to give rise to a non-Kramers doublet ground state~\cite{PhysRevB.98.174404}. In \ref{subseccef}, we will discuss the ground state and CEF excitations in more details. There is no signature of magnetic phase transition by magnetic susceptibility, heat capacity, and muon-spin relaxation measurements, suggesting the possible realization of a gapless QSL ground state~\cite{PhysRevB.98.174404}. However, it was shown that there was a significant site mixing between the magnetic Tm$^{3+}$ and nonmagnetic Zn$^{2+}$ sites~\cite{PhysRevB.98.174404}, causing a strong disorder effect that could impact the proposed QSL state.
\begin{figure*}[htb]
\centerline{\includegraphics[width=0.8\linewidth]{fig1.pdf}}
\caption{
(a) Schematic crystal structure of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace with $R\bar3m$ space group. (b) Top view of the kagom\'e layer consisting of Tm$^{3+}$ ions at the corners and Zn$^{2+}$ ions at the centers. (c) and (d) Rietveld refinement results for the powder X-ray diffraction data for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, respectively. (e) and (f) Temperature dependence of magnetic susceptibility for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, respectively, measured under zero-field-cooling (ZFC) and field-cooling (FC) conditions with a magnetic field of 0.1~T from 2 to 300~K. The insets of (e) and (f) show the inverse susceptibility data at low temperatures. The dashed lines are the fits with the Curie-Weiss law.
\label{fig1}}
\end{figure*}
\begin{table*}[htb]
\begin{threeparttable}
\caption{Room-temperature XRD pattern refinements for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace.}
\label{tab:para}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}ccccccccccc}
\hline \hline
\begin{minipage}{2cm}\vspace{1mm} Compound \vspace{1mm} \end{minipage} & Atom & Wyckoff position & $x$ & $y$ & $z$ & Occ. & $U_{\rm iso}$~(\AA$^2$) & $a$~(\AA) & $c$~(\AA) & $\chi^2$\\
\hline
\begin{minipage}{2cm}\vspace{1mm} \vspace{1mm} \end{minipage} & Tm & 9e & 0.5 & 0 & 0 & 0.65(3) & & & & \\
& Zn(disorder) & 9e & 0.5 & 0 & 0 & 0.35(3) & & & & \\
& Zn(1) & 3a & 0 & 0 & 0 & 0.45(3) & & & & \\
& Tm(disorder) & 3a & 0 & 0 & 0 & 0.55(3) & & & & \\
Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace & Zn(2) & 3b & 0 & 0 & 0.5 & 0.44(4) & 0.025 & 7.3541(2) & 16.9956(5) & 3.84 \\
& Tm(disorder) & 3b & 0 & 0 & 0.5 & 0.56(4) & & & & \\
& Sb & 9d & 0.5 & 0 & 0.5 & 1 & & & & \\
& O(1) & 6c & 0 & 0 & 0.393(6) & 1 & & & & \\
& O(2) & 18h & 0.504(3) & -0.504(3) & 0.117(2) & 1 & & & & \\
& O(3) & 18h & 0.143(3) & -0.143(3) & -0.026(2) & 1 & & & & \\
\hline
\begin{minipage}{2cm}\vspace{1mm} \vspace{1mm} \end{minipage} & Tm & 9e & 0.5 & 0 & 0 & 0.93(1) & & & & \\
& Mg (disorder) & 9e & 0.5 & 0 & 0 & 0.07(1) & & & & \\
& Mg(1) & 3a & 0 & 0 & 0 & 0.82(1) & & & & \\
& Tm(disorder) & 3a & 0 & 0 & 0.5 & 0.18(1) & & & & \\
Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace & Mg(2) & 3b & 0 & 0 & 0.5 & 0.95(1) & 0.025 & 7.2772(5) & 17.2278(1) & 3.67 \\
& Tm(disorder) & 3b & 0 & 0 & 0.5 & 0.05(1) & & & & \\
& Sb & 9d & 0.5 & 0 & 0.5 & 1 & & & & \\
& O(1) & 6c & 0 & 0 & 0.615(2) & 1 & & & & \\
& O(2) & 18h & 0.522(2) & -0.522(2) & 0.139(1) & 1 & & & & \\
& O(3) & 18h & 0.132(2) & -0.132(2) & -0.058(1) & 1 & & & & \\
\hline \hline
\end{tabular*}
\begin{tablenotes}
\small
\item $T$ = 300~K, wavelength of the x-ray $\lambda$ = 1.54~\AA, and space group: $R\bar3m$.\\
\item $U_{\rm iso}$ denotes the isotropic actomic displacement from the equilibrium positions, and $\chi^2$ represents the goodness of fitting.\\
\end{tablenotes}
\end{threeparttable}
\end{table*}
In this work, in order to identify the role of disorder in the QSL candidates, we choose Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and its sister compound Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace with quantifiable disorder, that is, strong and weak disorder in the former and latter, respectively, and investigate how disorder affects the material's magnetic properties. On one hand, the absence of magnetic order and spin freezing down to $\sim$50~mK far below the Curie-Weiss temperature ($\Theta_{\rm CW}$), the presence of residual linear term in the specific heat, and the observation of broad gapless magnetic excitations resemble those of gapless QSLs. On the other hand, we find that the disorder resulting from the random mixing of magnetic Tm$^{3+}$ and nonmagnetic Zn$^{2+}$ or Mg$^{2+}$ in the Tm layers is intimately correlated with the strength of the low-energy magnetic excitations. In particular, in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace which has strong disorder, as identified from the structural refinement and CEF excitations, the value of the residual linear term in the specific heat is about 4 times larger, and the intensity of the gapless spin excitations in the INS spectra is greatly enhanced, compared to those of Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace with much less disorder. These results demonstrate conclusively that disorder in a geometrically frustrated compound can make it mimic a QSL.
\section{Experimental Details}
Polycrystalline samples of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace (La$_3$Sb$_3$Zn$_2$O$_{14}$\xspace) and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace (La$_3$Sb$_3$Mg$_2$O$_{14}$\xspace) were synthesized by conventional solid-state reactions with stoichiometric amounts of Tm$_2$O$_3$ (99.99\%) (Lu$_2$O$_3$, 99.99\%), Sb$_2$O$_3$ (99.99\%), ZnO (99.99\%), and MgO (99.99\%) powders. The mixtures of the precursor compounds of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace (La$_3$Sb$_3$Zn$_2$O$_{14}$\xspace) were carefully ground and sintered at 1200~$^{\circ}$C for 3 days. For Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace (La$_3$Sb$_3$Mg$_2$O$_{14}$\xspace) samples, a higher reaction temperature of 1350~$^{\circ}$C and longer reaction time of 5 days were required to obtain the pure phase. X-ray diffraction (XRD) data were collected in an x-ray diffractometer (X$^\prime$TRA, ARL) using the Cu-$K_\alpha$ edge with a wavelength of 1.54~\AA. Rietveld refinements on the data were run by the EXPGUI and GSAS programs~\cite{Toby:hw0089,Toby:aj5212}. The dc magnetic susceptibility was measured in a Quantum Design physical property measurement system (PPMS, EverCool). Specific heat above 2~K was measured on 5.6-mg Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and 4.7-mg Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace samples, respectively, in a PPMS EverCool. The data below 2~K were collected on 1.7-mg Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and 8.0-mg Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace in a PPMS DynaCool equipped with a dilution refrigerator.
INS experiments on 5.3-g Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and 4.5-g Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace polycrystalline samples were carried out on PELICAN equipped with a dilution refrigerator, a cold neutron time-of-flight (TOF) spectrometer located at ANSTO at Lucas Heights, Australia. The powders were loaded into a pure copper can in the dilution refrigerator, which was able to cool down to around 50~mK. The incident neutron wavelength was selected as $\lambda~\sim$ 4.69 \AA, corresponding to an incident energy of 3.69~meV and an energy resolution of $\Delta E$ = 0.067~meV (half width at half maximum, HWHM). For each temperature, we collected data for about 12 hours. The CEF experiments were performed on 8-g powders on a thermal triple-axis spectrometer TAIPAN at ANSTO. The powders were loaded into an aluminum can and then mounted onto a closed-cycle refrigerator which could reach 1.6~K. A pyrolytic graphite (PG) filter was placed after the sample to reduce contaminations from higher-order neutrons. The beam collimations were 0$^\prime$-40$^\prime$-sample-40$^\prime$-0$^\prime$. A fixed-final-energy ($E_{\rm{f}}$) mode with $E_{\rm{f}}$ = 14.87~meV was used in the measurements. The resulting energy resolution was about 0.41~meV (HWHM). Measurements were performed under vertical-focusing conditions for both the monochromator and analyzer on TAIPAN.
\section{Results}
\subsection{Structure and magnetic susceptibility}
Figures~\ref{fig1}(a) and~\ref{fig1}(b) show the schematics of the crystal structure and the two-dimensional kagom\'e layer of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace, respectively. Magnetic Tm$^{3+}$ ions forming corner-shared kagom\'e-lattice layers are well separated by nonmagnetic layers and have an ABC stacking arrangement along the $c$ axis~\cite{PhysRevB.95.104439,PhysRevB.98.174404}. We have performed Rietveld refinements on the XRD data of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and the results are shown in Fig.~\ref{fig1}(c). For the $R\bar3m$ space group with a perfect kagom\'{e} lattice, there should be some Bragg reflections below 30 degrees, which are absent in our data. In order to capture the absence of low-angle reflections, we have to allow some mixings between the Tm$^{3+}$ and Zn$^{2+}$(1) sites. Such a site-mixing model was used in Ref.~\onlinecite{PhysRevB.98.174404} to analyze the XRD data in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace as well. In order to keep the stoichiometry, we need to free the site mixing between the Tm$^{3+}$ and Zn$^{2+}$(2) sites as well, which was not considered in Ref.~\onlinecite{PhysRevB.98.174404}. The detailed refinement parameters are listed in Table~\ref{tab:para}. Our refinement results show that there are around 35\% Zn$^{2+}$ ions occupying Tm$^{3+}$ positions, and 55\% and 56\% Tm$^{3+}$ ions occupying Zn$^{2+}$(1) and Zn$^{2+}$(2) positions, respectively. The strong site mixing of Tm$^{3+}$ and Zn$^{2+}$ reduces the distinctness of their original positions and then increases the crystallographic symmetry, {\it i.e.}, from kagom\'{e} to triangular within the $a$-$b$ plane~\cite{PhysRevB.95.104439,PhysRevB.98.174404}. This naturally explains the absence of the low-angle Bragg reflections in Fig.~\ref{fig1}(c).
The random site mixing between the magnetic and nonmagnetic sites is expected to have a strong impact on the magnetic properties, which is indeed the case as will be discussed later. To address this issue, it is better to have a comparative compound that is isostructural to Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace but has less disorder resulting from the random mixing of the magnetic and nonmagnetic ions. For this purpose, we have replaced the nonmagnetic Zn in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace with Mg and synthesized Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, which has less disorder as we show below. We have performed similar refinements and the results are shown in Fig.~\ref{fig1}(d) and Table~\ref{tab:para}. The XRD pattern for Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace shown in Fig.~\ref{fig1}(d) is almost the same as that for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace shown in Fig.~\ref{fig1}(c), except for the additional reflections below 30 degree, which is expected for the $R\bar3m$ space group with perfect kagom\'{e} layers. As shown in Table~\ref{tab:para}, the crystal structure for both compounds is the same, but the amount of Tm$^{3+}$ in the Mg$^{2+}$ sites is significantly reduced. We believe that the different amounts of disorder in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace are due to the different radii of the Zn$^{2+}$ and Mg$^{2+}$ ions. Compared to the smaller Mg$^{2+}$, the larger Zn$^{2+}$ ions are closer to Tm$^{3+}$ ions in size, and thus it is easier to occupy each other randomly. In support of this point, it has been reported that when the radius of the rare-earth ion becomes large enough to have an obvious difference from that of Zn$^{2+}$ ion, such as Dy$^{3+}$ or larger ones in the 4$f$ row, the disorder effect will be weakened significantly~\cite{PhysRevB.95.104439}.
We further characterize both compounds by measuring the magnetic susceptibility ($\chi$) with a magnetic field of 0.1~T, and the results are shown in Figs.~\ref{fig1}(e) and~\ref{fig1}(f). There is no indication of long-range magnetic order down to 2~K. In addition, the absence of the bifurcation of zero-field-cooling (ZFC) and field-cooling (FC) data also indicates that there is no spin freezing at the lowest temperature measured. The inverse susceptibility is linear for most of the temperature range, except for the slight deviation at low temperatures, as shown in the inset of Figs.~\ref{fig1}(e) and~\ref{fig1}(f). Such a deviation may be associated with the development of short-range magnetic correlations, and is commonly observed in QSL candidates~\cite{PhysRevB.76.132411,PhysRevLett.120.087201,prl115_167203,PhysRevB.95.174414}. It can also be explained with the thermal population of CEF levels and we will discuss it later. From the Curie-Weiss fits, we obtain the Curie-Weiss temperature $\Theta_{\rm {CW}}$ of -17.4 and -28.7~K for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, respectively, implying dominating antiferromagnetic spin correlations in these materials. These results are consistent with the QSL state expected for a kagom\'{e}-lattice system with antiferromagnetic interactions and strong geometrical frustration~\cite{nature464_199,Wen2019Experimental,PhysRevB.98.174404}. On the other hand, the disorder resulting from the random mixing of the magnetic and nonmagnetic sites discussed earlier can be also responsible for these observations~\cite{PhysRevLett.120.087201,PhysRevLett.119.157201,PhysRevX.8.031028,PhysRevX.8.041040}.
\subsection{CEF excitations}\label{subseccef}
\begin{figure}[htb]
\centerline{\includegraphics[width=3.45in]{fig2.pdf}}
\caption{(a) and (b) Crystal-electric-field (CEF) excitations below 40~meV for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, respectively, measured at 1.6~K on TAIPAN spectrometer. Green and black arrows in (a) indicate the positions of CEF transitions at 1.6 and 18.5~meV caused by Tm$^{3+}$ at the Zn(1) sites, and 4.5, 24.8, and 33.5~meV caused by Tm$^{3+}$ at the original sites for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace. Black solid line represents the convolution with two Gaussian functions denoted by dashed lines. Arrows in (b) indicate the positions of CEF transitions at 4, 23.5, and 33.5~meV for Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace. (c) and (d) $\bm{Q}$-dependence intensities of the CEF levels for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, respectively. Solid lines indicate that the intensities of the CEF excitations follow the magnetic form factor of Tm$^{3+}$ ion well with $I~\propto~|f(\bm{Q})|^2$. $I$ and $|f(\bm{Q})|$ denote excitation intensity measured and magnetic form factor of Tm$^{3+}$ ion, respectively. Errors represent one standard deviation throughout the paper.
\label{fig2}}
\end{figure}
To further investigate the disorder effect in these two compounds, we measured the CEF excitations, and the experimental results are shown in Fig.~\ref{fig2}. The experimental CEF spectra of Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace are shown in Fig.~\ref{fig2}(b). There are three well-isolated CEF transitions centered at 4, 23.5, and 33.5~meV, respectively, with no asymmetry nor significant broadening observed. For Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace, there are two CEF levels observed at 1.6 and 4.5~meV. Both peaks are slightly broader than the instrument resolution of 0.41~meV but still relatively sharp. However, another one centered at 19~meV is asymmetric and much broader. To analyze the experimental results in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace, we need to consider the random site mixing between the Tm$^{3+}$ and Zn$^{2+}$ ions, which is strong as shown in Table~\ref{tab:para}. A Zn$^{2+}$(1) cation with surrounding six O$^{2-}$ anions produces a ZnO$_6$ polyhedra, while the Zn$^{2+}$(2) and Tm$^{3+}$ cations are both surrounded by eight O$^{2-}$ anions~\cite{doi:10.1002/pssb.201600256}. As a result, the Tm$^{3+}$ ions at the original and Zn(1) sites have different ligand environments and should give rise to two different sets of CEF excitations associated with these two sites. This indeed makes the CEF pattern of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace more complicated than that of Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace. By comparing their experimental results in Figs.~\ref{fig2}(a) and~\ref{fig2}(b), we believe the CEF transitions observed at 4.5, 24.8, and 33.5~meV of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace can be attributed to excitations of the Tm$^{3+}$ ions at the original sites, since these energies are almost the same as those of Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, although the one at 33.5~meV is weaker than that in the latter. In Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace, the remaining two CEF levels at 1.6 and 18.5~meV which are completely absent in Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace should result from the CEF excitations of Tm$^{3+}$ cations at the Zn(1) sites. In this case, the asymmetric and very broad peak around 19~meV for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace shown in Fig.~\ref{fig2}(a) is actually composed of two CEF levels centered at 18.5 and 24.8~meV resulting from two different sites. These results show clearly that the site mixing between the magnetic and nonmagnetic ions will have a strong impact on the CEF excitations.
In order to confirm the signals we have observed here indeed originate from CEF excitations, some $\bm{Q}$ scans located at representative energy levels were performed. In Figs.~\ref{fig2}(c) and~\ref{fig2}(d), the behavior of monotonic decrease of intensities with increasing $\bm{Q}$ follows the magnetic form factor of a Tm$^{3+}$ ion, distinct from phonon scatterings. Moreover, INS results of a nonmagnetic reference compound La$_3$Sb$_3$Mg$_2$O$_{14}$ show phonon scattering is only pronounced at a larger range of $\bm{Q}$ $\ge$ 10~\AA$^{-1}$~(Ref.~\onlinecite{PhysRevB.98.134401}), which is far away from our currently investigated area.
The rare-earth ions in the materials are strongly influenced by the electrostatic environment they occupy. Therefore, in order to quantitatively identify how the $(2J+1)$-fold ($J=6$) spin-orbital degeneracy is lifted by the CEF effect, a CEF analysis was performed, and the effective Hamiltonian was obtained by the point-charge model according to the point-group symmetry\cite{HUTCHINGS1964227},
\begin{equation}\label{CEF}
H_{\rm CEF}=\sum_{n,m}B_n^m O_n^m,
\end{equation}
where $O_n^m$ and $B_n^m$ are the Steven operators\cite{Stevens_1952} and CEF parameters, respectively. In the following, $B_n^m $ are calculated based on the 15 nontrivial CEF parameters shown in Table~\ref{CEFp} within the point-charge approximation\cite{HUTCHINGS1964227,PhysRevB.98.134401}.
Since the CEF of Tm$^{3+}$ at the original sites has a very low symmetry of the $C_{2h}$ point group with eight oxygen ligands, the $13$-fold degeneracy of the ground-state manifold $^3H_6$ is expected to be completely lifted. Fortunately, the diagonalization of Eq.~(\ref{CEF}) predicts that the splitting of two low-lying singlets are small enough to be regarded as a nearly degenerate non-Kramers doublet. The effective spin in the doublet is an easy-axis moment, and the components are mostly $J_z=\pm6$ along the easy axis.
If we choose our local axes such that the two-fold rotation symmetry is about the $y$ axis and the easy axis is the $z$ axis, the doublet shown in Table~\ref{doublet} can be well described by symmetric and antisymmetric combinations of $J_z=\pm6$ states:
\begin{equation}\label{state}
|+\rangle\approx\frac{\sqrt{2}}{2}(|6\rangle+|-6\rangle),|-\rangle\approx\frac{\sqrt{2}}{2}(|6\rangle-|-6\rangle).
\end{equation}
Due to the low symmetry, the CEF peaks observed experimentally are not sufficient to simulate all the CEF parameters. In addition, the disorder would further influence the simulation. In this case, we simply do the symmetry analysis and obtain the effective CEF Hamiltonian by point-charge approximation. Since the lifting of the degeneracy is mainly determined by the point-group symmetry, we think the analysis is sufficient to identify the low-lying states, which are very far away from other higher energy levels. Further fitting would revise the higher energy levels, but the components of the non-Kramers doublet would not change significantly. A similar non-Kramers doublet was also reported for Ho$_3$Sb$_3$Mg$_1$O$_{14}$~(Ref.~\onlinecite{PhysRevX.10.031069}) and Tb$_3$Sb$_3$Mg$_2$O$_{14}$~(Ref.~\onlinecite{dun2020effective}), which share the same structure. The fitting analysis of the same structure of other crystals also supported that the symmetry analysis is qualitatively efficient for the low-lying states\cite{PhysRevB.98.134401}.
For the Tm$^{3+}$ cations at the Zn($1$) sites in Tm$_3$Zn$_2$Sb$_3$O$_{14}$, the ligand environment is a squashed oxygen octahedron, whose symmetry is higher than the original sites. The CEF with $D_{3d}$ point group symmetry will split the $13$-degenerate states into five singlets and four doublets. It has been revealed that two low-lying singlets would compose the nearly degenerate non-Kramers doublet\cite{PhysRevX.10.011007}. The components of the nearly degenerate non-Kramers doublet ground state are primarily $J_z=\pm6$ and $\pm3$ states. Therefore, an effective spin-$1/2$ can be defined in this doublet\cite{PhysRevX.10.011007}. Due to the occupations of the Tm$^{3+}$ cations at the original and Zn(1) sites, the two different ligand environments lead to different CEF excitations, which results in a very different CEF pattern shown in Fig.~\ref{fig2}(a).
\begin{table}[htb]
\caption{The CEF parameters obtained from the point-charge approximation.}
\label{CEFp}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\fill}}cc}
\hline\hline
\begin{minipage}{2cm}\vspace{1mm} $B_n^m$ \vspace{1mm} \end{minipage} & \begin{minipage}{2cm}\vspace{1mm} meV \vspace{1mm} \end{minipage} \\
\hline
\begin{minipage}{2cm}\vspace{0.5mm} $B_2^0$ \vspace{0.5mm} \end{minipage} & -1.40866 \\
$B_2^1$ & 2.02733 \\
$B_2^2$ & 0.79967 \\
$B_4^0$ & -0.00252 \\
$B_4^1$ & -0.01587 \\
$B_4^2$ & 0.00610 \\
$B_4^3$ & 0.03075 \\
$B_4^4$ & 0.02983 \\
$B_6^0$ & -0.00002 \\
$B_6^1$ & -0.00007 \\
$B_6^2$ & 0.00004 \\
$B_6^3$ & 0.00016 \\
$B_6^4$ & -0.00017 \\
$B_6^5$ & -0.00054 \\
$B_6^6$ & -0.00005 \\
\hline\hline
\end{tabular*}
\end{table}
\begin{table*}[htb]
\begin{threeparttable}
\caption{Eigenvalues and eigenvectors for the non-Kramers doublet ground state of Tm$^{3+}$ at the original sites. The first column indicates the energies and the rest indicate the antisymmetric and symmetric states of the doublet.}
\label{doublet}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}ccccccccccccccc}
\hline \hline
\begin{minipage}{2cm}\vspace{1mm} $E$ (meV) \vspace{1mm} \end{minipage} & & $|-6\rangle$ & $|-5\rangle$ & $|-4\rangle$ & $|-3\rangle$ & $|-2\rangle$ & $|-1\rangle$ & $|0\rangle$ & $|1\rangle$ & $|2\rangle$ & $|3\rangle$ & $|4\rangle$ & $|5\rangle$ & $|6\rangle$ \\
\hline
0.000 & $|-\rangle=$ & (0.701 & 0.010 & -0.079 & 0.042 & 0.008 & -0.024 & 0.000 & -0.024 & -0.008 & 0.042 & 0.079 & 0.010 & -0.701)\\
0.018 & $|+\rangle=$ & (0.701 & 0.008 & -0.086 & 0.031 & 0.001 & -0.017 & 0.024 & 0.017 & 0.001 & -0.031 & -0.086 & -0.008 & 0.701)\\
\hline \hline
\end{tabular*}
\end{threeparttable}
\end{table*}
\subsection{Specific heat results}
\begin{figure}[htb]
\centerline{\includegraphics[width=3.45in]{fig3.pdf}}
\caption{(a) and (b) Specific heat results of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace at ultralow temperatures. The specific heat of nonmagnetic references La$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and La$_3$Sb$_3$Mg$_2$O$_{14}$\xspace are also shown. Solid lines are the fits with Debye model as $C_{\rm p}~\sim~T^3$ for the nonmagnetic compounds. Insets show the low-temperature $C_{\rm p}/T$ $vs.$ $T^2$. Dashed lines are the linear fits. (c) and (d) Magnetic specific heat~($C_{\rm m}$) results of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace after subtracting the contribution from the lattice using nonmagnetic reference compounds La$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and La$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, respectively.
\label{fig3}}
\end{figure}
We performed ultralow-temperature specific heat ($C_{\rm p}$) measurements of these two compounds and the results are shown in Fig.~\ref{fig3}. Figure~\ref{fig3}(a) shows the specific heat results of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace down to 70~mK. There is no sharp $\lambda$-type peak expected for a well-defined phase transition. Instead, there is a kink around 7~K. This kink temperature almost coincides with the temperature when the susceptibility deviates from the Curie-Weiss behavior shown in Fig.~\ref{fig1}(e). In some other QSL candidates, there is a more obvious hump which probably corresponds to the establishment of short-range spin correlations~\cite{prl98_107204,np4_459,sr5_16419,PhysRevLett.120.087201}. We conjecture that the underlying physics for the kink observed here is similar. Another possible explanation is that it results from the thermal population of the low-lying CEF level located at 1.6~meV as shown in Fig.~\ref{fig2}(a). As for a two-level system with the energy splitting of 1.6~meV, it gives rise to a maximum at 7.7~K which is close to the kink temperature observed in both $\chi(T)$ and $C(T)$. Generally speaking, the specific heat at low temperatures can be fitted as $C_{\rm p} = \gamma T + \beta T^3$ for a system with gapless fermionic excitations~(Refs.~\onlinecite{np4_459,nc2_275}), where the linear $T$ term and $T^3$ term denote electronic and phononic contributions, respectively. Compared with the nonmagnetic reference compound La$_3$Sb$_3$Zn$_2$O$_{14}$\xspace behaving well as $C_{\rm p}/T \sim T^2$, which is reasonable, since it is an insulator~\cite{PhysRevB.98.174404}, Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace has a large residual linear term. In the inset of Fig.~\ref{fig3}(a), we plot $C_{\rm p}/T$ as a function of $T^2$ to focus on the low-temperature part, and the linear extrapolation to absolute zero temperature yields a finite linear term coefficient $\gamma \sim$ 31.5(6)~mJ~mol$^{-1}$~K$^{-2}$. Such an observation is quite unusual for an insulator, and is often interpreted to be due to the fermionic fractional excitations such as spinons of a QSL~\cite{np4_459,nc2_275,sr5_16419}. The specific heat results of Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace depicted in Fig.~\ref{fig3}(b) exhibit similar behaviors. The kink shifts to a higher temperature of around 17~K, and it may reflect the fact that the exchange interaction is stronger than that in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace, which is consistent with the higher $\Theta_{\rm CW}$ in Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace. We believe the kink should have the same origin as discussed earlier for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace. The specific heat also exhibits a linear term, as shown in the inset of Fig.~\ref{fig3}(b). However, the $\gamma$ value of 8.9(7)~mJ~mol$^{-1}$~K$^{-2}$ is more than three times smaller than that of 31.5(6)~mJ~mol$^{-1}$~K$^{-2}$ of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace, indicating much less density of states at low energies in Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace. Considering the substantial amount of disorder in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace, we believe that the enhancement of the $\gamma$ is due to the disorder resulting from the strong site mixing of Tm$^{3+}$ and Zn$^{2+}$ ions. We also present the magnetic specific heat~($C_{\rm m}$) results of these two compounds in Figs.~\ref{fig3}(c) and~\ref{fig3}(d). The kinks around 7~K in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and 17~K in Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace are more clearly shown after subtracting the phonon contributions.
\subsection{INS spectra}
\begin{figure*}[htb]
\centering
\includegraphics[width=0.9\linewidth]{fig4.pdf}
\caption{Inelastic neutron scattering spectra of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace (a)-(c) and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace (e)-(g) measured on PELICAN spectrometer~\cite{doi:10.7566/JPSJS.82SA.SA027}. (d) shows the wave-vector $\bm{Q}$-dependence of the energy-integrated intensity of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace. The energy and $\bm{Q}$ range are marked with the dashed rectangular in (a) and (e). Because the signals for Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace are too weak, to visualize the $\bm{Q}$-dependence of the intensity in (d), we used the 60-K data as the background and subtracted it. Solid lines are guides to the eye. }
\label{fig4}
\end{figure*}
\begin{figure}[htb]
\centerline{\includegraphics[width=3.4in]{fig5.pdf}}
\caption{Energy dependence of the integrated neutron scattering intensity in the $\bm{Q}$ range between 0.6 to 1.4~\AA$^{-1}$ for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace (a) and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace (b). Solid lines are guides to the eye. Dashed lines represent the instrumental resolutions. Elastic neutron scattering data for (c) Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and (d) Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, obtained by integrating the intensity in an energy window of [-0.1,0.1]~meV. The data were collected on a time-of-flight spectrometer PELICAN at various temperatures.
\label{fig5}}
\end{figure}
We now turn to the INS measurements at low energies to gain further insights into the magnetic state. Figures~\ref{fig4}(a)-\ref{fig4}(c) show the magnetic excitation spectra (raw data) collected for the polycrystalline sample of Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace at three characteristic temperatures. The INS results contain two pronounced features. First is the flat excitation band between $E= 1.1$ and 2.1~meV. These excitations are the low-lying CEF excitations of the Tm$^{3+}$ ions as also shown in Fig.~\ref{fig2}(a). Second is another broad and nearly flat excitation band below the CEF exciations, approximately in the energy range of 0.3 to 0.6~meV. The intensity weakens at 5~K [Fig.~\ref{fig4}(b)] and disappears at 30~K [Fig.~\ref{fig4}(c)]. To investigate the low-energy excitations in detail, we integrate the intensities with energy ranging from 0.3 to 0.6~meV to avoid contaminations from the CEF excitations and elastic scattering, and plot the integrated intensity as a function of $\bm{Q}$ in Fig.~\ref{fig4}(d). It is clear that there is a broad peak centered at $\bm{Q}~\sim$~1~\AA$^{-1}$ which corresponds to the $\Gamma$ point in the second Brillouin zone of a kagom\'{e} lattice. The $\bm{Q}$ and temperature dependence of these excitations indicate that they are of magnetic origin~\cite{PhysRevLett.120.087201,np15.262}. Figures~\ref{fig4}(e)-\ref{fig4}(g) show the INS spectra of Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace. As discussed earlier and shown in Fig.~\ref{fig2}(b), the first CEF level of Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace is 4~meV, which is beyond the energy range in Fig.~\ref{fig4}. As a result, we do not observe the CEF excitations in Figs.~\ref{fig4}(e)-\ref{fig4}(g) for Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace. The INS measurements were performed on the samples with similar weights and thus similar amount of magnetic Tm$^{3+}$ ions. Furthermore, we used the same experimental setup and equal counting time. Therefore, Figs.~\ref{fig4}(a)-~\ref{fig4}(c) and~\ref{fig4}(e)-\ref{fig4}(g) can be compared directly. Different from that in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace, the low-energy excitations are barely visible in Fig.~\ref{fig4}(e) for Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace. We integrate the intensity between 0.3 and 0.9~meV, but since the intensity is too weak, we need to subtract the background data at 60~K to make it visible. The so-obtained intensity as a function of $\bm{Q}$ is shown in Fig.~\ref{fig4}(d). There also appears to be a broad peak centered at $\bm{Q}~\sim$~1~\AA$^{-1}$, and the intensity increases as the temperature decreases. However, the overall intensities are significantly weaker than those in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace.
To further investigate the broad excitations centered at $\bm{Q}~\sim$~1~\AA$^{-1}$, we integrated the intensities with $\bm{Q}$ ranging from 0.6 and 1.4~\AA$^{-1}$ at various temperatures, and the integrated intensities are plotted as a function of energy in Fig.~\ref{fig5}. The intensities follow similar temperature dependence for both compounds: on the energy-loss side ($E>0$), intensities increase with decreasing temperature; on the energy-gain ($E<0$) side, the intensities are suppressed at low temperatures due to the detail balance. These results indicate that the intensities are resulting from intrinsic magnetic excitations. Moreover, the magnetic signals are dominated by the inelastic scattering, while the quasi-elastic scattering has no temperature dependence. It is reasonable that the spectral weight is mainly distributed in the inelastic channels for a system with moderate spin interactions in which $\Theta_{\rm {CW}}$ is -17.4 and -28.7~K for Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace and Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, respectively. We integrate the elastic channels and plot the data collected at $\sim$50~mK, and there is no magnetic Bragg peak observed in Figs.~\ref{fig5}(c) and~\ref{fig5}(d), which also proves the absence of magnetic order for both compounds down to $\sim$50~mK.
\section{Discussions and Conclusions}
How do we understand the INS results given the strong site mixing demonstrated from the XRD, CEF, and specific heat results? At first glance, the low-energy magnetic excitations in Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace are distinct from conventional spin-wave excitations, but similar to the spectrum expected from the deconfined spinons in QSLs~\cite{nature492_406,nature540_559,np12_942,np15.262,np15_1052}. However, we believe the disorder-induced low-energy excitations will be a more natural explanation. Since the low-lying state of Tm$^{3+}$ at the original sites is a nearly degenerate non-Kramers doublet and dominated by $J_z=\pm6$ components by our CEF analysis, the effective spin residing in the doublet is more likely to behave as a multipole, which is not directly accessible for neutron scattering that has a selection rule of $\Delta S = \pm1$~(Ref.~\onlinecite{RevModPhys.85.367}). This explains nicely why the low-energy excitations are so weak in Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace, which has much less disorder. We need to point out that a well-isolated singlet ground state of Tm$^{3+}$ at the original sites in Tm$_3$Sb$_3$Mg$_2$O$_{14}$\xspace was also proposed\cite{dun2020effective}. Although we cannot rule out this possibility, it will not affect the conclusion for the absence of low-energy magnetic excitations resulting from the original sites, as there will be no extra magnetic excitations below the first CEF level from the isolated singlet ground state. We believe the spin excitations resulting from the effective spin of Tm$^{3+}$ in the Zn$^{2+}$(2) sites will not give rise to the intensities observed in the INS experiment either as these two sites have similar CEF environments with eight O$^{2-}$ anions surrounding them~\cite{doi:10.1002/pssb.201600256}. For Tm$_3$Sb$_3$Zn$_2$O$_{14}$\xspace with a strong disorder effect, a large amount of Tm$^{3+}$ ions leave their original sites and occupy Zn$^{2+}$(1) position which is surrounded by six O$^{2-}$ anions, different from the situation in their original sites~\cite{doi:10.1002/pssb.201600256}. The CEF environment with a higher symmetry may preserve the effective spin-1/2 resulting from the dominating $J_z=\pm6$ and $\pm3$ components in the non-Kramers doublet, and in this case the dipole moments of the non-Kramers doublet are effective and can give rise to the low-energy magnetic excitations observed in the INS
experiment\cite{PhysRevX.10.011007}. Furthermore, the Zn$^{2+}$(1) sites are at the center of the hexagon. This will tune the kagom\'e lattice to triangular lattice effectively. In this case, the peak center of the excitation intensity $\bm{Q}\sim$1~\AA$^{-1}$ corresponds to the M point of the twice-enlarged Brillouin zone in the triangular lattice due to the reduction of the $a$ axis by half. This is fully consistent with the observation of negative Curie-Weiss temperature for the dominant antiferromagnetic interactions. Of course, INS measurements on single crystals are desirable in order to better reveal the momentum distribution of the spectral weight. Nevertheless, our observation of the site-mixing-induced broad gapless excitations is fully in line with previous reports in YbZnGaO$_4$~(Ref.~\onlinecite{PhysRevLett.120.087201}), Yb$_2$TiO$_5$~(Ref.~\onlinecite{PhysRevB.96.174418}), and $\kappa$-(ET)$_2$Cu[N(CN)$_2$]Cl~(Ref.~\onlinecite{PhysRevLett.115.077001}).
At this point, we conclude that the disorder resulting from the site mixing of the magnetic Tm$^{3+}$ ions and the nonmagnetic ones is responsible for the absence of magnetic order, CEF excitations, specific heat, and the INS spectra. Intriguingly, the amount of disorder is manifested in the value of the residual linear term in the specific heat and the strength of the low-energy magnetic excitations. Then a natural question is what is the ground state with no disorder? One possibility is the valence-bond-solid state, in which two nearest-neighbor antiparallel spins form a singlet---such a state will not exhibit a magnetic order either~\cite{nc_Itamar,PhysRevX.8.031028,PhysRevX.8.041040,nc10_2561,PhysRevB.101.140401}. Another possibility is the gapped QSL state. In both cases, disorder will induce the low-energy magnetic excitations that contribute to the specific heat and magnetic neutron scattering. Although to distinguish these two states requires further experimental efforts, our results here demonstrate directly that by bringing disorder into such a geometrically frustrated system, a spin-liquid-like state can be induced. We believe this conclusion holds for a broad class of frustrated magnets including both the organic and inorganic compounds in the presence of either magnetic or nonmagnetic disorder. With many efforts, both experimental~\cite{np13_117,PhysRevLett.118.107202,PhysRevB.94.060409,PhysRevLett.120.087201,PhysRevX.8.031001,PhysRevB.96.174418,PhysRevLett.115.077001} and theoretical~\cite{PhysRevLett.119.157201,PhysRevB.97.184413,PhysRevLett.118.087203,PhysRevLett.120.207203,nc_Itamar,PhysRevX.8.031028,PhysRevX.8.041040,PhysRevLett.123.087201}, the role of disorder in the realization towards the spin-liquid-like behaviors has been made prominent. This also poses a great challenge for experimentalists in identifying a QSL~\cite{Wen2019Experimental}.
\section{Acknowledgements}
The work was supported by the National Natural Science Foundation of China with Grant Nos~11822405, 12074174, 11674157, 11674158, 11774152, 11904170, 12004249, and 12004251, National Key
Projects for Research and Development of China with Grant No.~2016YFA0300401, Natural Science Foundation of Jiangsu Province with Grant Nos~BK20180006 and BK20190436, Shanghai Sailing Program with Grant No.~20YF1430600, Fundamental Research Funds for the Central Universities with Grant No.~020414380117, and the Office of International Cooperation and Exchanges of Nanjing University. Hai-Feng Li acknowledges financial support from the Science and Technology Development Fund, Macau SAR (File Nos~028/2017/A1 and 0051/2019/AFJ). We acknowledge stimulating discussions with Itamar Kimchi. We thank the ACNS at ANSTO for the access of neutron-scattering facilities through Proposals 7312 and 7314, and the excellent support from Gene Davidson from the Sample Environment Group for setting up the dilution refrigerator on PELICAN.
| 2ec8da1cd674093947caf7bf36f35585aecee80e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $G$ be a group, and let $I=\{\mathcal{C}_1,\dots, \mathcal{C}_k\}$ be a set of conjugacy classes of $G$. We say that $I$ \emph{invariably generates} $G$ if $\langle x_1,\dots, x_k \rangle=G$ for all $x_1\in\mathcal{C}_1,\dots, x_k\in\mathcal{C}_k$. The set $I$ is a \emph{minimal invariable generating set} (or \emph{MIG-set} for short) if $I$ invariably generates $G$ and no proper subset of $I$ invariably generates $G$. We let $m_I(G)$ denote the largest possible cardinality of a MIG-set for the group $G$.
We state our main theorem.
\begin{theorem}\label{t: main}
Let $n\geq 2$ be an integer, and let $G=S_n$ be the symmetric group on $n$ letters. Then
\[
\frac{n}{2}-\log n < m_I(G) < \frac n2 +\Delta(n)+O\left(\frac{\log n}{\log \log n}\right),
\]
where $\Delta(n)$ is the number of divisors of $n$.
\end{theorem}
(Logarithms are in base $2$.) It is well known that $\Delta(n)= n^{o(1)}$, therefore Theorem \ref{t: main} implies that $m_I(S_n)$ is asymptotic to $n/2$ as $n\rightarrow \infty$.
The parameter $m_I(G)$ is the ``invariable'' analogue of the parameter $m(G)$, which denotes the largest possible cardinality of a minimal generating set of $G$ (with analogous definition). See Subsection \ref{subsec: context_mIm} for more context.
In \cite{GarLuc}, it was asked whether $m_I(G)\leq m(G)$ holds for every finite group $G$, and it was proved that $m_I(G)=m(G)$ in case $G$ is soluble. In particular, $m_I(S_n)=m(S_n)$ for $n\leq 4$.
In the upper bound in Theorem \ref{t: main}, we will in fact prove a more explicit estimate, which will have the following consequence.
\begin{corollary}\label{c: main}
If $G=S_n$ with $n\geq 5$, then $m_I(G)<m(G)$.
\end{corollary}
As we shall explain in the next subsection, this was known for large enough $n$.
\subsection{Methods of proof} For a finite group $G$, we denote by $k(G)$ the number of conjugacy classes of $G$. We begin with an elementary observation.
\begin{lemma}
\label{l: basic_obs_mI}
Suppose that $I=\{\mathcal{C}_1,\dots, \mathcal{C}_t\}$ is a set of conjugacy classes of a non-trivial finite group $G$. Then, $I$ is a MIG-set if and only if the following conditions are both satisfied:
\begin{itemize}
\item[(a)] There exists a set of maximal subgroups $J=\{M_1,\dots, M_t\}$ of $G$ such that, for every $i\neq j$, $\mathcal{C}_i \cap M_j\neq \emptyset$.
\item[(b)] No proper subgroup of $G$ has non-empty intersection with $\mathcal{C}_i$ for all $i=1,\dots, t$.
\end{itemize}
\end{lemma}
In particular, with notation as in the previous lemma, we have the following two consequences:
\begin{itemize}
\item[$\diamond$] For every $j=1, \ldots, t$, $M_j$ has non-empty intersection with at least $t-1$ non-trivial $G$-conjugacy classes, and therefore $k(M_j)\geq t$.
\item[$\diamond$] For every $i\neq j$, $M_i$ and $M_j$ are not $G$-conjugate.
\end{itemize}
\noindent {\bf The lower bound.} Let us see how Lemma \ref{l: basic_obs_mI} can be exploited to give a lower bound for $m_I(S_n)$. Each conjugacy class $\mathcal{C}$ of $S_n$ corresponds to a particular partition, $X_\mathcal{C}$, of the integer $n$. On the other hand, if $M$ is an \emph{intransitive} subgroup of $S_n$, then $M$ is the stabilizer of some $i$-subset of $\{1,\dots, n\}$.
We say that \emph{the integer $i$ is a partial sum of the partition $\mathfrak p=(a_1,\dots, a_t)$} if we can write $i=a_{j_1}+a_{j_2}+\cdots + a_{j_\ell}$ for some $1\leq j_1 < \cdots < j_\ell \leq t$. It is clear that the intersection $\mathcal{C}\cap M$ is non-empty if and only if $i$ is a partial sum of $X_\mathcal{C}$. We will prove the following:
\begin{proposition}\label{p: partitions}
Let $n\geq 5$ be an integer. There is a set $X$ of partitions of $n$ with the following properties:
\begin{enumerate}
\item There is no integer $1 \leq i \leq n/2$ which is a partial sum in $x$ for every $x \in X$;
\item For every $x \in X$, there exists an integer $1 \leq i \leq n/2$ which is a partial sum in $y$ for every $y \in X \setminus \{x\}$;
\item $|X|>\frac 12n - \log n$.
\end{enumerate}
\end{proposition}
Proposition~\ref{p: partitions} is almost enough to yield the lower bound in Theorem~\ref{t: main} straight away. To complete the proof of that lower bound, we must take care of Lemma \ref{l: basic_obs_mI}(b) for proper transitive subgroups of $S_n$.
Our feeling is that the construction we give in our proof of Proposition~\ref{p: partitions} is pretty close to being as large a set $X$ as is possible.
\begin{question}
\label{q: large_partitions}
Is it true that the largest cardinality of a set $X$ of partitions of $n$ satisfying properties (1) and (2) of Proposition~\ref{p: partitions} is at most $\frac 12n-\log n+O(1)$?
\end{question}
Note that we certainly have $|X|\leq \frac 12n$.
\vspace{3pt}
\noindent {\bf The upper bound.}
\begin{comment}
{\color{red}********************************************************}
The part between red ********** (which is a comment for you) has been added. Removing it we get the previous version.
Our argument, in fact, gives an upper bound to $\iota(G)$, as defined in here. I used this parameter in another paper with Lucchini, and in my thesis as well. I wonder whether it is worth introducing it here, or whether it would only load the exposition, with no benefit. Your opinion?
\begin{lemma}
\label{iota}
$m_I(G)\leq \iota(G)$.
\end{lemma}
\begin{proof}
Let $m=m_I(G)$ and let $\{C_1,\dots,C_m\}$ be a MIG-set for $G.$ For every $1\leq i \leq m,$ there exists a maximal subgroup $M_i$ of $G$ such that
$\{C_1,\dots,C_{i-1},C_{i+1},\dots,C_m\} \subseteq M_i^*$ but $C_i\notin M_i^*.$ It follows that $\{M_1^*,\dots,M_m^*\}$ is an independent subset of $\mathcal C(G),$ and therefore $m\leq \iota(G).$
\end{proof}
{\color{red}**************************************************************}
\end{comment}
In view of Lemma \ref{l: basic_obs_mI} and of the considerations following it, it is clear that the upper bound in Theorem \ref{t: main} will be established once we prove the following result.\footnote{The idea we are using here is laid out explicitly in \cite{GarLuc}: For every maximal subgroup $M$ of $G$, denote by $M^*$ the set of $G$-conjugacy classes having non-empty intersection with $M$. Let
\[
\mathcal M(G)=\{M^*\mid M \text { maximal subgroups of $G$}\}.
\]
We say that a subset $\{X_1,\dots,X_t\}$ of $\mathcal M(G)$ is \textit{independent} if, for every $1\leq i\leq t$, the intersection $\cap_{j\neq i}X_j$ properly contains $\cap_j X_j$. We denote by $\iota(G)$ the largest cardinality of an independent subset of $\mathcal M(G)$. It is not hard to see, first, that $m_I(G)\leqslant\iota(G)$ (\cite[Lemma 4.2]{GarLuc}) and, second, that Proposition~\ref{p: upper} yields an upper bound for $\iota(S_n)$.}
\begin{proposition}\label{p: upper}
Suppose that $\{M_1,\dots, M_t\}$ is a set of maximal subgroups of the symmetric group $S_n$ such that (a) $k(M_i)\geq \frac 12n$ for every $i$; (b) if $i\neq j$, then $M_i$ and $M_j$ are not $S_n$-conjugate. Then
\begin{equation}\label{e: pp}
t\leq \frac{n}{2}+\Delta(n)+O\left(\frac{\log n}{\log \log n}\right),
\end{equation}
where $\Delta(n)$ is the number of divisors of $n$.
\end{proposition}
The main point in the proof of Proposition~\ref{p: upper} is to deal with the family of almost simple primitive subgroups of $S_n$; see Theorem \ref{p: locally valid}. The key ingredient is \cite[Theorem 1.2]{garzonigill}, which determines the almost simple primitive subgroups $G$ of $S_n$ such that $k(G)\geq \frac 12n$.
We remark that a theorem of Liebeck and Shalev gives a general upper bound for the number of conjugacy classes of maximal subgroup of $S_n$ of the form $\frac n2+o(n)$ \cite{lisha}. This immediately gives an upper bound for $m_I(S_n)$ and, in light of the easy fact that $m(S_n)\geq n-1$, yields Corollary~\ref{c: main} provided $n$ is large enough (this was observed also in \cite{GarLuc}). We note that in Proposition \ref{p: upper} we do not use \cite{lisha}, but we use \cite{garzonigill}, which relies on upper bounds for the number of conjugacy classes of almost simple groups of Lie type by Fulman--Guralnick \cite{FG1}.
We remark, moreover, that although Proposition~\ref{p: upper} only states an upper bound for the number of maximal subgroups with at least $\frac12 n$ conjugacy classes, results in \textbf{S}\ref{s: upper} outline specific families of maximal subgroups. In particular, the first two terms of \eqref{e: pp} correspond to the intransitive and imprimitive subgroups of $S_n$, respectively.
This is important because our original aim in this paper was to prove that $|m_I(G)-\frac n2| =O(\log n)$. We have managed this with the lower bound but not with the upper, precisely because $\Delta(n)-2$, which is the number of conjugacy classes of maximal imprimitive subgroups of $S_n$, is not $O(\log n)$. To achieve our original aim, it would be sufficient to establish that, in the following question, $t\leq \frac n2+O(\log n)$. We state the question in terms of properties of $S_n$ -- it is easy enough to recast it as a number-theoretic question concerning partitions, similar to Question \ref{q: large_partitions} above.\footnote{Yet another way to think of this question uses the terminology of the previous footnote. We are effectively asking the following: Let $t$ be a positive integer and let $M_1,\dots, M_t$ be maximal subgroups of $S_n$ that are either intransitive or imprimitive. If $\{M_1^*, \ldots, M_t^*\}$ is independent, then how large can $t$ be?
}
\begin{question}
\label{q: partition_imprimitive}
For a positive integer $n$, how large can $t$ be such that we can find sets with the following properties?
\begin{enumerate}
\item $\{\mathcal{C}_1,\dots, \mathcal{C}_t\}$ is a set of conjugacy classes of $S_n$;
\item $\{M_1,\dots, M_t\}$ is a set of maximal subgroups of $S_n$, all of which are intransitive or imprimitive;
\item For $i=1,\dots, t$, $\mathcal{C}_i\cap M_i=\emptyset$;
\item For $i,j=1,\dots, t$, if $i\neq j$, then $\mathcal{C}_i\cap M_j\neq \emptyset$.
\end{enumerate}
\end{question}
Proposition \ref{p: partitions} shows that $t>n/2-\log n$. In truth, we believe that, at least for large enough $n$, a MIG-set of $S_n$ of size $m_I(S_n)$ should involve only intransitive subgroups (in the sense that the set $J$ from Lemma \ref{l: basic_obs_mI} should contain only intransitive subgroups). This would imply that $m_I(S_n)\leq \frac n2$, and the problem of determining $m_I(S_n)$ would be reduced to the purely combinatorial problem addressed in Proposition \ref{p: partitions} and Question \ref{q: large_partitions}.
\subsection{Context}
\label{subsec: context_mIm}
The concept of invariable generation was introduced by Dixon, with the motivation of recognizing $S_n$ as the Galois group of polynomials with integer coefficients \cite{Dix92}. See for instance Kantor--Lubotzky--Shalev \cite{KLS} for interesting results related to invariable generation of finite groups.
In \cite{GarLuc}, the invariant $m_I(G)$ was introduced and studied. This is the ``invariable'' version of the invariant $m(G)$, which is the largest possible cardinality of a minimal generating set of $G$.
See Lucchini \cite{min, min2} for results concerning $m(G)$ where $G$ is a general finite group.
Assume now $G=S_n$. It is easy to see that $m(S_n)\geq n-1$, by considering the $n-1$ transpositions $(1,2), \ldots, (n-1,n)$. Using CFSG, Whiston \cite{whi} proved that in fact $m(S_n)=n-1$. But more is true. Cameron--Cara \cite{Caca} showed that a minimal generating set of $S_n$ of size $n-1$ is very restrictive: either it is made of $n-1$ transpositions, or it is made of a transposition, some $3$-cycles, and some double transpositions (see \cite[Theorem 2.1]{Caca} for a precise statement).
One can hardly hope for a similar ``elegant'' result for $m_I(S_n)$, for the simple reason that a minimal invariable generating set of $S_n$ of size $t$ must contain $t$ distinct partitions which do not have a common partial sum. Still, it is true that in the proof of the lower bound in Theorem \ref{t: main}, we feel somewhat restricted about the choice of the relevant partitions -- but we are not able to make any precise statement in this direction.
In \cite{GarLuc}, it was shown that, if $G$ is a finite soluble group, then $m_I(G)=m(G)$, which in turn is equal to the number of complemented chief factors in a chief series of $G$. Moreover, it was asked whether $m_I(G)\leq m(G)$ is true for every finite group. It seems that, ``often'', for a finite non-soluble group $G$, the strict inequality $m_I(G)<m(G)$ holds. Corollary \ref{c: main} confirms this feeling in case $G=S_n$.
We recall, however, that $m_I(\text{PSL}_2(p))=m(\text{PSL}_2(p))$ for infinitely many primes $p$ (see \cite[Section 5]{GarLuc}).
\subsection{Structure of the paper and notation} In \textbf{S}\ref{s: lower} we prove the lower bound on $m_I(G)$ given in Theorem~\ref{t: main}. In \textbf{S}\ref{s: upper} we prove the upper bound on $m_I(G)$ given in Theorem~\ref{t: main}, along with Corollary~\ref{c: main}.
We will use exponential notation for partitions, so the partition $(a_1^{n_1}, a_2^{n_2}, \ldots, a_t^{n_t})$ has $n_1$ parts of length $a_1$, $n_2$ parts of length $a_2, \ldots$, and $n_t$ parts of length $a_t$. For a positive real number $x$, $\log(x)$ denotes a logarithm in base $2$. For a positive integer $x$, $\Delta(x)$ denotes the number of divisors of $x$.
\section{The lower bound}\label{s: lower}
In this section we prove the lower bound in Theorem \ref{t: main}. We first prove a lemma, then we prove Proposition ~\ref{p: partitions}, and finally we give a proof for the lower bound.
\begin{lemma}
\label{cruc}
Let $n$ and $i$ be positive integers, with $i < n/3$. Then there exist a partition $\mathfrak p_{i,n}$ of $n$ with the following properties:
\begin{enumerate}
\item If $n \neq 4i+2$ and $(n,i) \neq (8,1)$, then $\mathfrak p_{i,n}$ does not have $i$ and $n-i$ as partial sums, and everything else is a partial sum.
\item If $n=4i+2$ or $(n,i)=(8,1)$, then $\mathfrak p_{i,n}$ does not have $i, n-i$ and $\frac{n}{2}$ as partial sums, and everything else is a partial sum.
\end{enumerate}
\end{lemma}
\begin{proof}
Define
\[
\mathfrak p_{i,n} = (1^{i-1}, i+1, (i+2)^j, (i+1)^k, c)
\]
where $j \in \{0,1\}$, $k \geqslant 0$, and $i+1 \leqslant c \leqslant 2i+1$. To complete the definition we must specify $j,k$ and $c$. To do this we consider a partial sum $q$, adding from left to right: we first sum the $1$'s and $(i+1)$ to obtain $q=2i$. Now there are three cases:
\begin{enumerate}
\item If $n-q\leqslant 2i+1$, then we set $c=n-q$.
\item If $n-q=2i+2$, then we set $k=1$, $j=0$ and $c=i+1$.
\item If $n-q\geqslant 2i+3$, then we set $j=1$, $k=0$ and set $q=2i+(i+2)=3i+2$.
\end{enumerate}
In the first and second cases, we are done; notice that the partition has the stated properties (in the first case we use the fact that $i<n/3$ to obtain $i+1\leqslant c\leqslant 2i+1$ as required). If we are in the third case, then we proceed in a loop as follows:
\begin{enumerate}
\item If $n-q\leqslant 2i+1$, then we set $c=n-q$.
\item If $n-q\geqslant 2i+2$, then we set $k=k+1$ and set $q=q+(i+1)$.
\end{enumerate}
It turns out that there is one situation -- when $n=4i+4$ and $(n,i) \neq (8,1)$ -- where our definition needs to be adjusted. In this case, we make the following definition:
\[
\mathfrak p_{i,n}=(1^{i-1}, i+1, i+3, i+1).
\]
Now our definition is complete. We now let $m$ be an integer such that $1 \leqslant m \leqslant n/2$ and we study when $m$ is a partial sum of $\mathfrak{p}_{i,n}$ with a view to proving items (1) and (2) of the lemma.
Both items are clear for $m \leqslant 2i$,
thus we may assume that $m \geqslant 2i+1$. In particular this means that $n\geq 4i+2$.
If $n=4i+2$, then we are in item (2) of the lemma, $\mathfrak p_{i,n}=(1^{i-1}, (i+1)^3)$, and the statement holds.
If $n\geqslant 4i+3$, note first that $j=1$. Suppose, first, that $k=0$. There are two possibilities: first, if $c\neq i+2$, then $\mathfrak p_{i,n} = (1^{i-1}, i+1, i+2, c)$, and the statement holds. If instead $c=i+2$, then $n=4i+4$. If $(n,i)=(8,1)$, then $\mathfrak p_{1,8}=(2,3,3)$ and we are in item (2) of the lemma. Otherwise, we are in the exceptional case in our definition where $\mathfrak p_{i,n}=(1^{i-1}, i+1, i+3, i+1)$, and the result holds.
We are left with the case in which $k\geqslant1$, i.e. $\mathfrak p_{i,n}$ contains at least two $(i+1)$'s (excluding $c$ which may also equal $i+1$).
We work here by induction on $m$: assuming that some $2i \leqslant m < n/2$ can be written \textit{without using c}, we want to show that the same holds for $m+1$. Since $m \geqslant 2i \geqslant i+1$, in writing $m$ without $c$ we have certainly used at least one of $i+1$ and $i+2$. We now divide into three cases.
\begin{enumerate}
\item In writing $m$ we have not used all $1$'s. Then add a $1$.
\item In writing $m$ we have not used $i+2$. Then remove an $i+1$ and add $i+2$.
\item In writing $m$ we have used all $1$'s and $i+2$. Suppose, first, that at least two $(i+1)$'s have not been used; then remove all $1$'s, remove $i+2$ and add two $(i+1)$'s and we are done. On the other hand, suppose (for a contradiction) that in writing $m$ as a partial sum all but one of the $i+1$'s have been used. Then $c+(i+1)> n/2$ and, since $c\leqslant 2i+1$, we obtain that $n/2< 3i+2$. However the partial sum $m$ has used all $1$'s, one $i+1$ and one $i+2$, so $m\geq 3i+2$. Since $m<n/2$, we get $n/2>3i+2$, which is a contradiction.\qedhere
\end{enumerate}
\end{proof}
\subsection{Proof of Proposition \ref{p: partitions}} Now we prove Proposition \ref{p: partitions}. The proof we give below is constructive -- we define an explicit set $X$ with the given properties.
We have decided not to define the set $X$ outside of this proof, as the construction is built up in pieces as the proof proceeds.
In deducing the lower bound in Theorem \ref{t: main}, we will be interested in the properties of the partitions of $X$ listed in the statement of Proposition \ref{p: partitions}, rather than their explicit construction. The paragraphs involving exceptions to this are labelled {\bf (C1)}, {\bf (C2)}, {\bf (C3)} and {\bf (C4)} in the following proof.
\iffalse
\begin{lemma}
\label{n_over_two}
Let $n \geq 11$ be a positive integer. There exists a set $X$ of partitions of $n$ with the following properties:
\begin{enumerate}
\item There is no integer $1 \leq i \leq n/2$ which is a partial sum in $x$ for every $x \in X$. On the other hand, for every $x \in X$, there exists an integer $1 \leq i \leq n/2$ which is a partial sum in $y$ for every $y \in X \setminus \{x\}$.
\item $|X| > n/2 - \log n$.
\end{enumerate}
\end{lemma}
\fi
\begin{proof}[Proof of Proposition \ref{p: partitions}]
Throughout the proof, we will use the notation $\mathfrak p_{t,n}$ to refer to the partitions in the statement of Lemma \ref{cruc} (so we allow any partition having the properties of the statement).
If $5\leq n\leq 10$, we have $n/2-\log n <2$, and the statement is easy to check. Therefore assume $n\geq 11$.
{\bf (C1)} For $n=11$, we set $x_1=(2^2,3,4)$, $x_2=(1,3^2,4)$, $x_3=(1^2,9)$. For $n=12$, we set $x_1=(2^2,3,5)$, $x_2=(1,3,4^2)$, $x_3=(1^2,10)$. The statement holds by setting $X=\{x_1, x_2, x_3\}$.
From now on we assume that $n > 12$. This has the advantage that in the proof that follows, all partitions of the form $\mathfrak p_{i,\ell}$ that we consider will have $\ell>8$ or $i >1$, and so we need not worry about the case $(\ell,i)=(8,1)$ mentioned in Lemma~\ref{cruc}.
For $1 \leqslant t < n/3$, define
\[
x_t=\mathfrak p_{t,n}.
\]
{\bf (C2)} We want to modify partition $x_1$. Namely, define
\[
x_1=
\begin{cases}
(3,5,2^k,4^j) & \text{if $n=14,16$}\\
(3,2^k,4^j) & \text{if $n=13,15,17$}\\
(3^\ell,7,2^k,4^j) & \text{if $n\geq 18$}
\end{cases}
\]
where $\ell=1$ or $2$ according to whether $n$ is even or odd, and $j\in \{0,1\}$ is defined by the condition that $x_1$ has an odd number of cycles of even length (and $k$ is consequently uniquely defined). It is easy to see that, in every case, every $2 \leq i \leq n/2$ is a partial sum in $x_1$.
Now, in order to go further, we will use a slightly different method. The partitions we are going to define will depend on a parameter $j$. We could define all of them at once, but to give an idea of the overall strategy, let us go through the first step explicitly.
We set $\alpha_1= \ceil{n/6-1}$ and, for every integer $\ceil{n/3} \leqslant t \leqslant t_1=5n/12$ we define
\[
x_{t} = (\mathfrak p_{\alpha_1, \alpha_1+t}, c_{t}).
\]
where $c_{t} = n-\alpha_1-t$. Let us justify this definition:
\begin{enumerate}
\item[(a)] Observe first that $1 \leq \alpha_1<(\alpha_1+t)/3$ hence the partition $\mathfrak p_{\alpha_1, \alpha_1+t}$ is well-defined.
\item[(b)] Next note that
\begin{align*}
&c_{t}=n-\alpha_1-t \\
&\quad > n-\left\lceil{\frac{n}{6}-1}\right\rceil-\frac{5n}{12} \\
&\quad\quad\quad >\frac{5n}{12}\geq t > \alpha_1,
\end{align*}
and so $\alpha_1$ and $t$ are not partial sums of $x_{t}$.
\item[(c)] We can easily check that either $4\alpha_1+2 > \alpha_1+t$ or else $(\alpha_1, t, n) = (1,5,12)$. The second possibility is excluded by our assumption $n > 12$. The first possibility implies that Lemma~\ref{cruc}(1) holds, and so all numbers up to $\alpha_1+t$ are partial sums, apart from $\alpha_1$ and $t$.
\item[(d)] Finally observe that
\begin{align}\label{n2}
\alpha_1+t &\geq \left\lceil{\frac{n}{6}-1}\right\rceil + \left\lceil{\frac{n}{3}}\right\rceil \\
&\geqslant \left\lceil{\frac{n}{2}-1}\right\rceil.\nonumber
\end{align}
We conclude that all numbers up to $n$ are partial sums in $x_t$ apart from $\alpha_1$, $t$ and (possibly) $n/2$. In fact, checking \eqref{n2} more carefully, it is clear that $\alpha_1+t\geq \lfloor n/2 \rfloor$ unless $n\equiv 0\pmod 6$ and $t=n/3$.
\end{enumerate}
{\bf Conclusion~1}: For $n/3<t\leq 5n/12$, the partition $x_t$ admits all partial sums up to $n$ except $\alpha_1$ and $t$.
{\bf Conclusion~2}: For $t=n/3$, the partition $x_t$ admits all partial sums up to $n$ except $\alpha_1$ and $t$ and (if $n$ is even) $n/2$.
Now our aim is to extend this definition to other parameters $t_i$ that are larger than $t_1$. More precisely, for integer $1 \leqslant j \leqslant \log(n/6)$ we define
\[
\alpha_j = \left\lceil{\frac{n}{2^{j-1}\cdot 6} -1}\right\rceil, \hspace{20pt}
t_j = \frac{(2^{j-1} \cdot 6 -1)n}{2^j \cdot 6}.
\]
For $j=1$ this is consistent with the previous definition. Now for integers $2 \leqslant j \leqslant \log(n/6)$ and $\floor{t_{j-1}} +1 \leqslant t \leqslant \floor{t_j}$ we define
\[
x_t = (\mathfrak p_{\alpha_j, \alpha_j+t}, c_t)
\]
where $c_t = n-\alpha_j-t$. Now, similarly to before, we must check four properties. Assume that $j\geqslant 2$.
\begin{enumerate}
\item[(a)] Observe that $1 \leq \alpha_j < (\alpha_j + \floor{t_{j-1}} +1) / 3$ and so the partition $\mathfrak p_{\alpha_j, \alpha_j+t}$ is well-defined.
\item[(b)] Next note that $c_t=n-\alpha_j - t > t > \alpha_j$ and so $\alpha_j$ and $t$ are not partial sums of $x_t$.
\item[(c)] Notice that the second case of Lemma \ref{cruc} does not occur. Indeed, $4\alpha_j+2$ is strictly smaller than $\alpha_j +\floor{t_{j-1}}+1$. Moreover, it is easy to check that the case $(\ell,i)=(8,1)$ cannot occur.
\item[(d)] Finally observe that $\alpha_j+\floor{t_{j-1}}+1 \geq \floor{n/2}$, and we conclude that all numbers up to $n$ are partial sums in $x_t$ apart from $\alpha_1$ and $t$.
\end{enumerate}
{\bf Conclusion~3}: Set $m=\floor{\log(n/6)}$. For $j=2,\dots, m$ and $\floor{t_{j-1}} +1 \leqslant t \leqslant \floor{t_j}$, the partition $x_t$ admits all partial sums up to $n$, except $\alpha_j$ and $t$.
We have now constructed $\floor{t_m}$ partitions of $n$; set $X_0=\{x_1, \ldots, x_{\floor{t_m}}\}$. Notice that, by the choice of $m$, $2^{m+1}\cdot 6 > n$. Then
\begin{align}\label{e: tm}
|X_0|=\floor{t_m} &> \frac{(2^{m-1} \cdot 6 -1)n}{2^m \cdot 6}-1 \\
&= \frac{n}{2} - \frac{n}{2^m \cdot 6} - 1 \nonumber \\
&> \frac n2 -3.\nonumber
\end{align}
Now we will remove some elements from $X_0$. First, observe that $\alpha_j<n/6$ for every $j$; we start by taking the subset $X$ obtained by removing $x_{\alpha_j}$ for every $j \geqslant 1$.
{\bf (C3)} Lemma~\ref{cruc}, and the three conclusions listed above imply that, for each $t$ satisfying $1\leq t\leq t_m$, the partition $x_t$ is the unique partition in $X$ which does not admit $t$ as a partial sum. Now we divide into two cases.
\begin{enumerate}
\item There exists an integer belonging to the interval $(t_m, n/2]$ which is a partial sum for all $x_t\in X$. Then, the minimum such integer is $\floor{t_m}+1$; we add one further partition to $X$:
\[
z = (1^{\floor{t_m}}, n-\floor{t_m}).
\]
\item No integer in $(t_m, n/2]$ is a partial sum in all $x_t$'s. Observing that $t_m\leq n/2-1$, Lemma~\ref{cruc} and the three conclusions above imply that $t_m = n/2-1$, i.e., $n=2^m \cdot 6$. In this case we could leave $X$ unchanged, and the statement would be proved. However, we prefer to immediately modify the set $X$.
Notice that $\alpha_m=1$ and $t_{m-1}=n/2-2$; it follows that $x_1=\mathfrak p_{1, n} \notin X$, and $X$ contains a unique partition, namely $x_{t_m}$, of the form $(\mathfrak p_{1, 1+t}, c_t)$. Then, we remove such partition and we reintegrate the partition $x_1$ in $X$. Moreover, we add to $X$ one further partition
\[
z = (1^{n/2-2}, n/2+2).
\]
\end{enumerate}
{\bf (C4)} Our construction is finished. Let us make one observation, before concluding the proof. In case (2) above, by construction $x_1 \in X$. We claim that the same holds in case (1). Indeed, one can easily check that $\alpha_m = 1$ if and only if $n=2^m \cdot 6$, and otherwise $\alpha_j > 1$ for every $j$. Therefore, in case (1), in our procedure we did not remove $x_1$ from $X_0$, hence clearly $x_1 \in X$.
We are now ready to conclude the proof of the statement. The considerations above imply that items (1) and (2) of the statement hold.
Regarding item (3),
\begin{align*}
|X| &\geq |X_0|+1 - \log(n/6) \\
&\geq \frac{n}{2}-2-\log n + \log 6 \\
&> \frac n2 - \log n.
\end{align*}
The proposition is now proved.
\end{proof}
We now deduce the lower bound of Theorem~\ref{t: main} from Proposition \ref{p: partitions}.
\begin{proof}[Proof of the lower bound of Theorem~\ref{t: main}]
For $n=2$, $n/2-\log n=0$ and the statement is trivial. For $3\leq n\leq 10$, we have $n/2-\log n<2$. Since certainly $m_I(S_n) \geq 2$, the statement holds and we may assume $n \geq 11$.
Consider the set $X$ of partitions constructed in the proof of Proposition \ref{p: partitions}. We will consider the elements of $X$ as conjugacy classes of $S_n$. We want to show that $X$ is a MIG-set for $S_n$.
It is easy to check the statement for $n=11,12$ (see the paragraph {\bf (C1)} in the proof of Proposition \ref{p: partitions}). Assume now $n \geq 13$.
By Proposition \ref{p: partitions}(1), the classes of $X$ cannot have non-empty intersection with an intransitive subgroup of $S_n$. On the other hand, by Proposition \ref{p: partitions}(2), if we drop one class from $X$, then the remaining classes have non-empty intersection with some intransitive subgroup.
Now we deal with transitive groups. Note that $X$ is not contained in $A_n$, since $x_1$ corresponds to an odd permutation (see the paragraphs {\bf (C2)} and {\bf (C4)}). Moreover, a power of $x_1$ corresponds to a cycle of prime length fixing at least $3$ points, which belongs to no primitive group different from $A_n$ and $S_n$ by a classical theorem of Jordan. Assume now the classes of $X$ preserve a partition of $\{1, \ldots, n\}$ made of $r>1$ blocks of size $k>1$. Recall that $X$ contains a partition $z=(1^{n-\ell}, \ell)$, with $\ell=n/2+2$ or $\ell=n-\floor{t_m}$ (see the paragraph {\bf (C3)}). By eq. \eqref{e: tm} in the proof of Proposition \ref{p: partitions}, we have $n/2-3 < \floor{t_m}< n/2$, and in particular $n/2<\ell < n/2+3$. We have that $k$ must divide $\ell$. Since $k$ also divides $n$, we get $k < 6$. If $n\neq 15$, then $x_1$ cannot preserve blocks of size at most $5$. If $n=15$, we note that $X$ contains $x_4=\mathfrak p_{4,15}$, and we may take $\mathfrak p_{4,15}=(1^3,5,7)$, which does not preserve any nontrivial partition of $\{1,\dots, 15\}$. The proof is now concluded.
\end{proof}
\section{The upper bound}
\label{s: upper}
In this section we prove the upper bound in Theorem \ref{t: main}, and we prove Corollary \ref{c: main}. Our main tool is the following result, which follows quickly from \cite{garzonigill}. Recall that $k(G)$ denotes the number of conjugacy classes of a finite group $G$.
\begin{theorem}\label{p: locally valid}
Let $G$ be a maximal almost simple primitive subgroup of $S_n$, and assume $k(G)\geq \frac n2$. Then one of the following occurs:
\begin{enumerate}
\item $G$ is listed in Table~\ref{tab: max};
\item $G=A_n$, or $G=S_d$ and the action of $G$ on $n$ points is isomorphic to the action on the set of $k$-subsets of $\{1, \ldots, d\}$ for some $2\leq k<d/2$.
\item $G=\mathrm{P\Gamma L}_d(q)$ and the action of $G$ on $n$ points is isomorphic to the action on the set of $1$-subspaces of $\mathbf F_q^d$.
\end{enumerate}
\end{theorem}
Note that the subgroups mentioned at item (2) satisfy $n=\binom{d}{k}$ for some integer $k$ with $1\leq k<d/2$; and the subgroups mentioned at item (3) satisfy $n=(q^d-1)/(q-1)$.
\begin{table}
\centering
\begin{tabular}{ccc}
\hline\noalign{\smallskip}
$n$ & $G$ & $k(G)$ \\
\noalign{\smallskip}\hline\noalign{\smallskip}
22 & $M_{22}.2$ & 21\\
40 & $\text{SU}_4(2).2$ & 25 \\
45 & $\text{SU}_4(2).2$ & 25 \\
\noalign{\smallskip}\hline\noalign{\medskip}
\end{tabular}
\caption{Some maximal almost simple primitive subgroups, $G$ of $S_n$, for which $k(G)\geq \frac{n}{2}$. In every case there is a single $S_n$-conjugacy class of primitive subgroups isomorphic to $G$
}\label{tab: max}
\end{table}
\begin{proof}
The statement follows from \cite[Theorem 1.2]{garzonigill}, by checking with \cite{GAP} which of the entries in \cite[Table 1]{garzonigill} correspond to maximal subgroups of $S_n$.
\end{proof}
As we observed in the introduction, the upper bound in Theorem \ref{t: main} follows immediately from Proposition~\ref{p: upper}, which we prove now.
\begin{proof}[Proof of Proposition~\ref{p: upper}]
We make use of the families of maximal subgroups given in the Aschbacher--O'Nan--Scott theorem, in particular the description given in \cite {LPS}.
\begin{enumerate}
\item {\bf Intransitive subgroups}: There are exactly $\lfloor\frac{n}{2}\rfloor$ conjugacy classes of these.
\item {\bf Imprimitive subgroups}: There are $\Delta(n)-2$ of these, where $\Delta$ is the divisor function.
\item {\bf Affine subgroups}: There is at most $1$ conjugacy class of these.
\item {\bf Almost simple subgroups}: If $M$ is almost simple and $k(M)\geq \frac n2$, then it is among the possibilities listed by Theorem~\ref{p: locally valid}, as follows.
\begin{enumerate}
\item There are three possibilities for degrees $22, 40, 45$ listed in Table~\ref{tab: max}.
\item There is at most one conjugacy class of maximal subgroups isomorphic to $\mathrm{P\Gamma L}_d(q)$ whenever $n=\frac{q^d-1}{q-1}$; we let $a_n$ be the number of pairs $(q,d)$ where $q$ is a prime power, $d$ is a positive integer, and $\frac{q^d-1}{q-1}=n$.
\item There is at most one conjugacy class of maximal subgroups with socle $A_d$ whenever $n=\binom{d}{k}$ for some $k$; we let $b_n$ be the number of pairs $(d,k)$ where $d$ and $k$ are positive integers with $k\leq d/2$ and $\binom{d}{k}=n$.
\end{enumerate}
\item {\bf Diagonal subgroups}: \cite[Theorem 1.1]{garzonigill} states that $k(M)<\frac n2$ in this case, so we can ignore these subgroups.
\item {\bf Product action subgroups}: In this case we have maximal subgroups isomorphic to $S_d\wr S_k$, where $n=d^k$ and $k>1$. For fixed values of $d$ and $k$, there is one conjugacy class, thus the number of conjugacy classes in $S_n$ is equal to the number of pairs $(d,k)$ where $d$ and $k$ are positive integers with $k>1$ and $n=d^k$; we write this number as $c_n$.
\item {\bf Twisted wreath subgroups}: These are never maximal, as they are defined to be subgroups of groups with a product action \cite{LPS} and so can be ignored (and in any case, $k(M)<\frac n2$ by \cite[Theorem 1.1]{garzonigill}).
\end{enumerate}
Observe that the number of conjugacy classes of maximal subgroup in $S_n$ that are either imprimitive, affine, or given in Table \ref{tab: max} is at most $\Delta(n)-1$. Therefore, if $\{M_1, \ldots, M_t\}$ is a set of maximal subgroups as in the statement, we have
\begin{equation}
\label{eq: upper bound}
t\leq \floor{\frac n2} + \Delta(n) + a_n + b_n + c_n -1.
\end{equation}
(We will use this in the proof of Corollary \ref{c: main}.) In order to prove Proposition~\ref{p: upper}, it is clearly enough to show that
\[a_n+b_n+c_n =O\left(\frac{\log n}{\log \log n}\right).\]
To bound $a_n$, observe that if
\[
\frac{q_1^{d_1}-1}{q_1-1} = \frac{q_2^{d_2}-1}{q_2-1},
\]
then $q_1$ and $q_2$ must be coprime. We obtain that $a_n$ must be bounded above by the number of distinct prime divisors of $n-1$.
In \cite{robin} it is proved that this number is
\iffalse
\begin{equation}\label{e: an}
1.3841\frac{\ln(n-1)}{\ln \ln(n-1)}
\end{equation}
\fi
\[
O\left(\frac{\log(n-1)}{\log \log(n-1)}\right),
\]
whence the same upper bound holds for $a_n$.
To bound $b_n$ we refer to a result of Kane \cite{kane2}, which asserts that\footnote{Singmaster's conjecture \cite{singmaster} asserts that $b_n$ is bounded above by an absolute constant; de Weger proposes that in fact this constant can be taken to be $4$ \cite{dew}, and evidence for the veracity of this conjecture is given in \cite{bbd}; in particular this is known to be true if $n\leq 10^{60}$.}
\[
b_n = O\left(\frac{\log n \log \log \log n}{(\log\log n)^3}\right).
\]
To bound $c_n$, we first recall (see \cite[Theorem 13.12]{Apo})
that, for a positive integer $x$,
\iffalse
\[
\text{exp}_2\left\{\frac{1.5379\ln x}{\ln \ln x}\right\}.
\]
\fi
\[
\Delta(x)\leq \text{exp}_2\left\{\frac{(1+o(1))\log x}{\log \log x}\right\}.
\]
Now consider the prime factorization of $n$: $n=p_1^{a_1}\cdots p_t^{a_t}$.
If $n=d^k$ then $p_1\cdots p_t$ divides $d$ and $k$ divides $a:=\text{gcd}\{a_1, \ldots, a_t\}$. Therefore, the number of choices for $k$ is at most the number of divisors of $a$ different from $1$. Now note that $a\leq \log n$,
and therefore
\[
c_n \leq \text{exp}_2\left\{\frac{(1+o(1))\log \log n}{\log \log \log n}\right\}.
\]
In particular we see that each of $a_n$, $b_n$, $c_n$ is $O(\log n/\log \log n)$. This proves the proposition.
\end{proof}
We conclude with the proof of Corollary~\ref{c: main}.
\begin{proof}[Proof of Corollary~\ref{c: main}]
It is easy to see that $\{(1,2), (2,3), (3,4), \dots, (n-1, n)\}$ is a minimal generating set of size $n-1$. It is, therefore, enough to show that $m_I(S_n)< n-1$.
From eq. \eqref{eq: upper bound} in the proof of Proposition \ref{p: upper}, we deduce that
\[
m_I(S_n)\leq \floor{\frac n2} + \Delta(n) + a_n + b_n + c_n -1,
\]
therefore it is sufficient to show that $\Delta(n)+a_n+b_n+c_n<n/2$.
Very weak estimates are enough here. First assume that $n\geq 71$.
As remarked in the proof of Proposition \ref{p: upper}, $a_n$ is bounded above by the number of distinct prime divisors of $n-1$, which is at most $\log n$. Moreover, $c_n$ is bounded above by $\max \{\Delta(x) : x\leq \floor{\log n}\}$, which is at most $\log n$. Let us consider $b_n$. Let $(d_1,k_1), \dots, (d_b, k_b)$ be pairs such that $\binom{d_i}{k_i}=n$ for all $i=1,\dots, b$. Order so that $i<j$ implies that $k_i< k_j$ and observe that then $k_b \geq b$ and $d_b\geq 2b$. This implies that $n\geq \binom{2b}{b}> 2^b$. In particular $b<\log n$.
Finally we need to bound $\Delta(n)$. For every real number $a\in (0,n]$, we have $\Delta(n)< n/a +a$. By choosing $a=\sqrt n$, we deduce $\Delta(n)< 2\sqrt n$.
Therefore $\Delta(n)+a_n+b_n+c_n<2\sqrt n +3\log n$, and it is sufficient to show that $2\sqrt n +3\log n\leq n/2$. Since $n\geq 71$, this is indeed the case.
\iffalse
We know that $a_n$ is bounded above by the number of prime divisors of $n-1$ and, since $n>211$, it is easy to conclude that $a_n<\frac12 \log n$.
Let us consider $b=b_n$. Let $(d_1,k_1), \dots, (d_b, k_b)$ be pairs such that $\binom{d_i}{k_i}=n$ for all $i=1,\dots, b$. Order so that $i<j$ implies that $k_i< k_j$ and observe that then $k_b \geq b$ and $d_b\geq 2b$. This implies that $n\geq \binom{2b}{b}> 2^b$. In particular $b<\log n$.
For $c_n$, we know that $c_n$ is bounded above by $\max\limits_{x\leq \lfloor \log n \rfloor}\Delta(x)$ and, since $n>211$, we know that $\log n >7$ and so $c_n\leq \frac 12 \log n$. We conclude that
\[
a_n+b_n+c_n < 2 \log n.
\]
Thus it is sufficient to prove that $\Delta(n)< \frac{n}{2}-2\log n-1$. Since $n>211$ it is sufficient to prove that $\Delta(n)<\frac25 n$. Using the obvious fact that $\Delta(n)\leq\frac{n}{3}+3$ and the fact that $n>211$ we are done.
\fi
For $n\leq 70$ we use \cite{GAP} to find that, except when $n\in\{5,6,8,12\}$, $S_n$ has less than $n-1$ conjugacy classes of maximal subgroup and the result follows immediately.
For the remaining cases, say that two cycle types are \textit{equivalent} if one is a power of the other one; e.g., $(2,2)$ is equivalent to $(4)$, $(2,3,3)$ is equivalent to $(2,1^6)$, etc. Note that a MIG-set of $S_n$ of size $t$ must contain $t$ pairwise non-equivalent cycle types.
For $n\in \{5,8,12\}$, there are exactly $n-1$ conjugacy classes of maximal subgroups of $S_n$. However, in each case, there is one which does not intersect at least $n-3$ pairwise non-equivalent cycle types, and the result follows. (For $n=5$ we may take $\mathrm{AGL}_1(5)$, for $n=8$ we may take $\mathrm{PGL}_2(7)$, and for $n=12$ we may take $\mathrm{PGL}_2(11)$.)
For $n=6$, we note that a MIG-set of size $t\geq 5$ must contain $5$ distinct non-trivial cycle types, each of which intersects non-trivially $4$ pairwise non-conjugate maximal subgroups. A direct check shows that the cycle types with this property are
\[
(2), (2^2), (2^3), (3), (3^2), (4), (4,2). \tag*{$(\star)$}
\]
Now, a set of $4$ pairwise non-equivalent cycle types, each intersecting non-trivially $\mathrm{PGL}_2(5)$, must contain the cycle type $(5)$, which does not appear in $(\star)$. Therefore, we deduce that a MIG-set of size $t\geq5$ must contain $5$ distinct cycle types, each of which intersects non-trivially $4$ pairwise non-conjugate maximal subgroups, not isomorphic to $\mathrm{PGL}_2(5)$. We see that $(4)$ does not have this property. All other cycle types appearing in $(\star)$ intersect non-trivially $S_3\wr S_2$, and we deduce that $m_I(S_6)<5$, as wanted.
\iffalse
{\color{red}**********************} Alternative argument for $n=6$:
For $n=6$, there are $6$ conjugacy classes of maximal subgroups. One of these, with representative $\mathrm{PGL}_2(5)$, does not intersect $5$ pairwise non-equivalent cycle types. In particular, we deduce that the only independent set of size $5$ can possibly be
\begin{equation}
\label{eq: iota_independent}
\mathrm{\Omega}=\{A_6^*, (S_2\wr S_3)^*, (S_3\wr S_2)^*, S_5^*, (S_4\times S_2)^*\}.
\end{equation}
We easily verify that the intersection of the first four $M^*$ in \eqref{eq: iota_independent} consists of the conjugacy classes $(1^6)$ and $(2^2, 1^2)$, both of which are contained in $(S_4\times S_2)^*$. This shows that $\mathrm{\Omega}$ is not independent, and therefore $\iota(S_6)<5$, as wanted.
{\color{red}**********************}
\fi
\end{proof}
| bfea5b8a8fd35364a6bf60d26829f1ad4d80131d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec:intro}
Apollo-type near-Earth asteroid 2012~TC4 was discovered in October 2012 by the Pan-STARRS1 survey, few days before its closest approach to the Earth (having a geocentric distance of about 95,000\,km). It was observed photometrically and its rotation period of about 12\,min \citep{Pol:13, War:13b, Odd.ea:13, Car:14} and effective diameter of 7-34 m \citep{Pol:13} were determined. Later, \cite{Rya.Rya:17} noticed also a second period in the data and interpreted it as a manifestation of a tumbling rotation state.
The next close approach in October 2017 was at a geocentric distance of about 50,000\,km and an even more extensive observing campaign (including spectroscopic and radar observations) was coordinated by the NASA Planetary Defense Coordination Office (PDCO) at that time. This campaign served also as a planetary defense exercise and its results were summarised by \cite{Red.ea:19}. Additionally, \cite{Ura.ea:19} also conducted the observing campaign of this asteroid in the same apparition independently and they attempted to reproduce their light curves with a model of a tumbling triaxial ellipsoid. Besides these observing campaigns, a few photometric observations of 2012 TC4 were carried out in this apparition \citep{Son.ea:17, War:18, Lin.ea:19}. All photometric data observed in 2017 confirmed the excited rotation state of 2012~TC4 with the main period of 12.2\,min.
Here, we revisit the situation using more sophisticated methods and tools. We reconstruct the convex shape model and spin state of 2012 TC4 from the available light curves that include the published data in the literature and our own new observations. We show that the rotation state must have changed between 2012 and 2017 apparitions and we propose a YORP-driven spin evolution as the most likely explanation. The data are described in Sect.~\ref{sec:data}, the physical model of the body is developed in Sect.~\ref{sec:physical_model}, and the theoretical analysis of rotation dynamics is in Sect.~\ref{sec:theory}. Mathematical methods and numerical set-up of the theoretical model are summarized in the Appendix~\ref{methods}. The best fit of our physical model to the available light curves is shown in the Appendix~\ref{sec:lc_fits}.
\section{Optical Photometry Data}
\label{sec:data}
To reconstruct the spin state of 2012~TC4, we collected its light curves observed during both close approaches. Photometric observations from 2012 and 2017 were made using a variety of telescopes having apertures between 0.35~m and 5~m and equipped with CCD cameras. Observational circumstances with references to original sources are listed in Table~\ref{table:observation}.
Apart from previously published light curves, our dataset
also includes several new observations (indicated by coauthor names in the last column).
In particular, we obtained four light curves using Pistoiese 0.6~m telescope (MPC code: 104) with CCD Apogee U6 which has a $35' \times 35'$ field of view corresponding to a pixel scale of 2 arcsec/pixel in both apparitions. The raw frames were processed for the dark and flat-field correction and the light curves of this observatory were constructed using Canopus software \citep{Warner2006}.
The pre-processing and the photometry of the light curve observed at Wildberg Observatory (MPC code: 198) using a 0.35~m telescope equipped with SXVF-H16 $2048 \times 2048$ CCD Camera was conducted using Astrometrica \citep{Raab_2012}. In the pre-process for this data, both dark and flat-field correction was carried out. All photometric data were calibrated referenced the PPMXL Catalog \citep{Roeser_et_al_2010}.
The Panchromatic Robotic Optical Monitoring and Polarimetry Telescopes (PROMPT) located at the Cerro Tololo Inter-American Observatory (CTIO) in Chile consist of six 0.41-m reflectors equipped with the Apogee Alta U47+ E2V camera. The field of view is $10'\times 10'$ with 0.59 arcsec/pixel. All raw image frames were processed (master dark, master flat, bad pixel correction) using the software package MIRA. Aperture photometry was then performed on the asteroid and three comparison stars. A master image frame was created to identify any faint stars in the path of the asteroid. Data from images with background contamination stars in the asteroid's path were then eliminated.
A part of the published light curves, we obtained from the Asteroid Lightcurve Data Exchange Fromat database (ALCDEF\footnote{http://alcdef.org/}, \citealp{War.ea:11}).
We also acquired the light curve published by \citep{Red.ea:19} from International Asteroid Warning Network (IAWN) 2012 TC4 Observing Campaign homepage\footnote{https://2012tc4.astro.umd.edu/Lightcurve/supplement/2012TC4\_Lightcurve\_Observations\_Summary.html}.
Since the corrected light curves were observed using various filters, and include both the relative and the absolutely calibrated observations, they have a magnitude offset between each other. As a result, the whole dataset is primarily treated as an ensemble of relative light curves.
\subsection{Two-period Fourier series analysis}
In the first step, we analyzed the photometry data from 2012 and 2017 using the 2-period Fourier series method \citep{Pra.ea:05, Pra.ea:14}.
Concerning the 2012 observations, we used all but one photometric light curve series taken from 2012-10-09.9 to 2012-10-11.2 (see Table~\ref{table:observation}). In particular, we excluded the data taken on 2012-10-10.7 by \cite{Pol:13}, where we found a possible timing problem.\footnote{Indeed, David Polishook (personal communication) checked his data upon our request and confirmed that there was an issue with the times in his 2012 observations. We then used the corrected data for physical model reconstruction in Sect.~\ref{sec:physical_model}.}
Concerning the 2017 observations, we used a selected subset of the data taken between 2017-10-09.1 and 2017-10-11.1. This choice was motivated by noting that the observing
geometry during this time interval in 2017 was very similar (due to the fortuitous resonant return of the asteroid to Earth after the five years) to the geometry of the time interval of the 2012 observations. This choice minimizes possible (but anyway small) systematic effects due to changes in observing geometry when comparing results of our analysis of the data from the two apparitions (see below).
We reduced the data to the unit geo- and heliocentric distances and to a consistent solar phase using
the $H$-$G$ phase relation, assuming $G = 0.24$, converted them to flux units and fitted them
with the 4th order 2-period Fourier series. A search for periods quickly converged and we found the two main periods
$P_1 = 12.2183 \pm 0.0002$\,min and $P_2 = 8.4944 \pm 0.0002$\,min in 2012 and
$P_1 = 12.2492 \pm 0.0001$\,min and $P_2 = 8.4752 \pm 0.0001$\,min in 2017.
The phased data and the best-fit Fourier series, together with the post-fit residuals, are plotted in Fig.~\ref{fig:Fourier_fit}.
We note that the smaller formal errors of the periods determined from the 2017 data were due to a higher quality of the 2017 observations
(the best-fit rms residuals were 0.081 and 0.065~mag for
the 2012 and 2017 data, respectively; see also the post-fit residuals plotted at the bottom part of Fig.~\ref{fig:Fourier_fit}). As for possible systematic errors of the determined periods, the largest could be due to the so-called synodic effect. It is caused by the change of position of a studied asteroid with respect to the Earth and Sun in the inertial frame during the observational time interval. An estimate of the magnitude of the synodic effect can be obtained using the phase-angle-bisector approximation, for which we used eq.~(4) from \cite{Pra.ea:96}.
Using this approach, we estimate the systematic errors of the determined periods
$\Delta P_1 = 0.0005$ and 0.0002\,min and $\Delta P_2 = 0.0002$ and 0.0001\,min
in 2012 and 2017, respectively. These systematic errors are only slightly larger than the formal errors given above. A caveat is that the formula eq.~(4) in \cite{Pra.ea:96} was determined for the case of a principal-axis rotator. An exact estimate of the systematic period uncertainties due to the synodic effect for a tumbling asteroid would require an analysis of its actual non-principal-axis rotation in given observing geometry, but we did not pursue it here as the effect was naturally surmounted by the physical modeling presented in the next section.
As will be shown in the next section, the strongest observed frequency in the light curve, $P_1^{-1}$, is actually a difference between the precession and the rotation frequency of the tumbler: $P_1^{-1} = P_\phi^{-1} - P_\psi^{-1}$. The second strongest frequency $P_2^{-1}$ is then the precession frequency $P_\phi^{-1}$. This is a characteristic feature of light curve of a tumbling asteroid in short-axis mode (SAM; see below).
We note that the same behavior was observed for (99942) Apophis and (5247) Krylov that are also in SAM \citep{Pra.ea:14,Lee_et_al_2020}.
The principal light curve periods $P_1$ and $P_2$ determined from the 2012 and 2017 observations differ at a high level of significance, formally on an order of about $100\,\sigma$.
While the systematic errors due to the synodic effect could decrease the formal significance by a factor of a few, the significance
of the observed period changes would still remain large, on an order of several tens of $\sigma$. To interpret these findings in more depth, we turned to construct a physical model of 2012~TC4 as a tumbling object in the next section.
\begin{figure*}[t]
\includegraphics[width=0.49\textwidth]{2012tc4_2012complc.png}
\includegraphics[width=0.49\textwidth]{2012tc4_2017complc.png}
\caption{The blue open circles are the photometric data of 2012~TC4 taken from 2012-10-09.9 to 2012-10-11.2 (left) and from 2017-10-09.1 to 2017-10-11.1 (right) reduced to the unit geo- and heliocentric distances and to a consistent solar phase (see text for details), folded with the respective main periods $P_1$. The red curve is the best-fit 4th order Fourier series with the two periods. The black plusses are the post-fit residuals (see the right ordinates).}
\label{fig:Fourier_fit}
\end{figure*}
\begin{table*}[t]
\caption{Observation details (new data denoted by name of a coauthor)}
\label{table:observation}
\begin{tabular}{l l l l}
\hline\hline
Telescope & Date (UT) & Filter & Ref. \\
\hline
\multicolumn{4}{c}{{\bf -- 2012 --}} \\
OAVdA 0.81-m (Italy) & 2012 10 09.9 & C & \cite{Car:14} \\
Pistoiese 0.6-m (Italy) & 2012 10 09.9 & R & Bacci \\
Pistoiese 0.6-m (Italy) & 2012 10 10.0 & R & Bacci \\
MRO 2.4-m (USA) & 2012 10 10.2 & V & \cite{Rya.Rya:17} \\
Wise observatory 0.72-m (Israel) & 2012 10 10.8 & V & \cite{Pol:13} \\
OAVdA 0.81-m (Italy) & 2012 10 10.8 & C & \cite{Car:14} \\
PROMPT1 0.41-m (Chile) & 2012 10 11.1 & Lum & Pollock \\
MRO 2.4-m (USA) & 2012 10 11.1 & V & \cite{Rya.Rya:17} \\
PDO 0.35-m (USA)$^a$ & 2012 10 11.2 & V & \cite{War:13b} \\ [4pt]
\multicolumn{4}{c}{{\bf -- 2017 --}} \\
Kitt Peak Mayall 4-m (USA) & 2017 09 13.2 & R & \cite{Red.ea:19} \\
Kitt Peak Mayall 4-m (USA) & 2017 09 14.1 & R & \cite{Red.ea:19} \\
Palomar Hale 5-m (USA) & 2017 09 17.4 & SR & \cite{Red.ea:19} \\
Palomar Hale 5-m (USA) & 2017 09 20.2 & SR & \cite{War:18} \\
SOAR 4.1-m (Chile) & 2017 10 06.2 & SR & \cite{Red.ea:19} \\
PDO 0.35-m (USA) & 2017 10 09.2 & V & \cite{War:18} \\
MRO 2.4-m (USA) & 2017 10 09.2 & V & \cite{Red.ea:19} \\
Kiso 1.05-m (Japan) & 2017 10 09.5 & SG & \cite{Ura.ea:19} \\
Wise observatory 0.72-m (Israel) & 2017 10 09.8 & V & \cite{Red.ea:19} \\
LCO-C 1-m (Chile) & 2017 10 10.1 & SR, SI & \cite{Red.ea:19} \\
LCO-A 1-m (Chile) & 2017 10 10.1 & SR & \cite{Red.ea:19} \\
PDO 0.35-m (USA) & 2017 10 10.2 & V & \cite{War:18} \\
Nayoro 0.4-m (Japan) & 2017 10 10.4 & V & \cite{Ura.ea:19} \\
BSGC 1-m (Japan) & 2017 10 10.6 & SG, SR, SI, SZ & \cite{Ura.ea:19} \\
Lulin 1-m (Taiwan)$^b$ & 2017 10 10.6 & BVRI (diff.) & \cite{Lin.ea:19} \\
Kiso 1.05-m (Japan) & 2017 10 10.5 & SG & \cite{Ura.ea:19} \\
Wise observatory 0.72-m (Israel) & 2017 10 10.8 & V & \cite{Red.ea:19} \\
Pistoiese 0.6-m (Italy) & 2017 10 10.9 & R & Bacci \\
KMTNet 1.6-m (South Africa) & 2017 10 10.9 & V & \cite{Red.ea:19} \\
Pistoiese 0.6-m (Italy) & 2017 10 11.0 & R & Bacci \\
USNA 0.51-m (USA) & 2017 10 11.0 & V & \cite{Red.ea:19} \\
MRO 2.4-m (USA) & 2017 10 11.1 & R & \cite{Red.ea:19} \\
PDO 0.35-m (USA) & 2017 10 11.2 & V & \cite{War:18} \\
Kiso 1.05 m (Japan)$^c$ & 2017 10 11.5 & SG & \cite{Ura.ea:19} \\
Lulin 1-m (Taiwan)$^b$ & 2017 10 11.6 & BVRI (diff.) & \cite{Lin.ea:19} \\
Anan Science Center 1.13-m (Japan)$^c$ & 2017 10 11.6 & V & \cite{Ura.ea:19} \\
Wise Observatory 0.72-m (Israel) & 2017 10 11.8 & V & \cite{Red.ea:19} \\
AIRA 0.38-m (Romania) & 2017 10 11.8 & V & \cite{Son.ea:17} \\
Wildberg Observatory 0.35-m (Germany)$^d$ & 2017 10 11.8 & & Apitzsch \\
KMTNet 1.6-m (South Africa) & 2017 10 11.9 & V & \cite{Red.ea:19} \\
MRO 2.4-m (USA) & 2017 10 12.1 & R & \cite{Red.ea:19} \\ [4pt]
\hline
\end{tabular}
\tablecomments{\scriptsize OAVdA: Astronomical Observatory of the Autonomous Region of the Aosta Valley, MRO: Magdalena Ridge Observatory, PROMPT1: Panchromatic Robotic Optical Monitoring and Polarimetry Telescopes, PDO: Palmer Divide Observatory, SOAR: Southern Astrophysical Research, LCO: Las Cumbres Observatory, BSGC: Bisei Spaceguard Center, KMTNet: Korea Microlensing Telescope Network \citep{Kimetal2016}, USNA: The United States Naval Observatory, AIRA: Astronomical Institute of the Romanian Academy \\
$^a$ Split into six parts, each using different comparison stars. \\
$^b$ These data were estimated by subtracting the average magnitudes of the comparison stars. \\
$^c$ Split into two parts because we noted a possible calibration issue. \\
$^d$ Split into two parts, each using different comparison stars.}
\end{table*}
\section{Physical Model}
\label{sec:physical_model}
\subsection{Model from 2017 data}
\label{sec:model_2017}
The light curve data set from 2017 is much richer than that from 2012, so we started with the inversion of 2017 data. We investigated possible frequency combinations based on the fact that the main frequencies $f_1$ and $f_2$ of a tumbling asteroid light curve are usually found at 2$f_\phi$ and 2($f_\phi \pm f_\psi$) or low harmonics and combination, where $f_\phi$ is the precession and $f_\psi$ the rotation frequency, respectively, and the plus sign is for a long-axis mode (LAM) and the minus sign for a short-axis mode (SAM) \citep{Kaa:01}. Using the values $f_1 = 4.898\,\text{hr}^{-1}$ and $f_2 = 7.079\,\text{hr}^{-1}$ from the previous section, we found
eight possible frequency combinations: $f_1 = f_{\phi}$, $f_2 = 2 (f_{\phi} - f_{\psi})$ (SAM1); $f_1 = 2(f_{\phi} - f_{\psi}), f_2 = 2 f_{\phi}$ (SAM2); $f_1 = 2(f_{\phi} - f_{\psi}), f_2 = f_{\phi}$ (SAM3); $f_1 = f_{\phi} - f_{\psi}, f_2 = f_{\phi}$ (SAM4); $f_1 = 2 f_{\phi}, f_2 = 2 (f_{\phi} + f_{\psi})$ (LAM1); $f_1 = 2 f_{\phi}, f_2 = f_{\phi} + f_{\psi}$ (LAM2); $f_1 = f_{\phi}, f_2 = (f_{\phi} + f_{\psi})$ (LAM3); $f_1 = f_{\phi} + f_{\psi}, f_2 = f_{\phi}$ (LAM4). Then we conducted the shape and spin optimization for these combinations with the same way as in \citet{Lee_et_al_2020}. It was found that only the SAM4 solution provided an acceptable fit to the data and was physically self consistent.
We used 34 light curves from October 2017. Four light curves from September 2017 were very noisy and did not further constrain the model, so we did not include them in our analysis. We inverted the light curves with the method and code developed by \cite{Kaa:01} combined with Hapke's light-scattering model \citep{Hapke_1993}. According to \cite{Red.ea:19}, colors of TC4 are consistent with C or X complex and also its spectrum is X type, with Xc type being the best match. The X complex contains low and high albedo objects but the E type seems most likely because the high circular polarization ratio suggests that TC4 is optically bright \citep{Red.ea:19}. This is in agreement with \cite{Ura.ea:19} who also report X type colors. As a result we used Hapke's model with parameters derived for an E type asteroid (2867)~\v{S}teins \citep{Spj.ea:12}: $\varpi = 0.57$, $g = -0.30$, $h = 0.062$, $B_{0} = 0.6$, and $\bar{\theta} = 28^\circ$. Because our data did not cover low solar phase angles, $h$ and $B_{0}$ parameters for opposition surge and also roughness $\bar{\theta}$ were fixed. We optimized $\varpi$ and $g$ parameters and they converged to values $\varpi = 0.69$, $g = -0.20$, which gave geometric albedo of 0.34. An alternative solution with fixed values $\varpi = 0.57$, $g = -0.30$ provided only marginally worse fit and the kinematic parameters were not affected. In general, solution of our inverse problem was not sensitive to Hapke's parameters likewise in previous studies \citep[e.g.,][]{Scheirich_et_al_2010, Pra.ea:14, Lee_et_al_2020}.
The rotation and precession periods had the following values: $P_\psi = 27.5070 \pm 0.002$\,min and $P_\phi = 8.47512 \pm 0.0002$\,min, respectively. The $1 \sigma$ uncertainties were estimated from the increase of $\chi^2$ when varying the solved-for parameters -- given the number of measurements about 8600, $3\sigma$ uncertainty interval corresponds to about 5\% increase in $\chi^2$ \citep[e.g.,][]{Vok.ea:17}. Direction of the angular momentum vector in ecliptic coordinates was $\lambda = 92^\circ$, $\beta = -89.6^\circ$, practically oriented toward the south ecliptic pole. Normalized moments of inertia were $I_1 = 0.41$, $I_2 = 0.81$, but the model was not too sensitive to their particular values. The inertia moments computed from the 3D shape (assuming constant density) were $I_1 = 0.435$, $I_2 = 0.831$, indicating consistency with the kinematic parameters above.
The dark facet, which is always introduced into a convex shape model to regularize the solution \citep{Kaa.Tor:01}, represented about a few percent of the total surface area and forcing it to smaller values led to a worse fit. This might mean that there is some albedo variegation on the surface of TC4 or that its real shape is highly nonconvex and a convex-shape approximation represent its limits.
\subsection{Model from 2012 data}
\label{sec:model_2012}
For this model, we used 14 light curves from 2012. The rotation and precession periods were now $P_\psi = 27.873 \pm 0.005$\,min and $P_\phi = 8.4945 \pm 0.0003$\,min, respectively. The $1 \sigma$ uncertainties were estimated in the same way as for the 2017 model, namely from the increase of $\chi^2$.
The direction of the angular momentum vector was $\lambda = 89^\circ$, $\beta = -90^\circ$, moments of inertia $I_1 = 0.43$, $I_2 = 0.80$. The 3D shape model was similar to that reconstructed from 2017 data and also the direction of the angular momentum vector was practically the same.
In spite of consistency in all other solved-for parameters, we thus observed that the periods $P_\phi$ and $P_\psi$ for the two apparitions were significantly different. Attempts to use the 2017 values to fit the 2012 data, or vice versa, led to a dramatically worse fit. We scanned period parameter space around the best-fit values and plotted the relative $\chi^2$ values -- see Fig.~\ref{fig:period_scan}. The minima in $\chi^2$ for 2012 and 2017 periods are clearly separated.
\begin{figure*}[t]
\includegraphics[width=0.49\textwidth]{period_scan_P_phi.eps}
\includegraphics[width=0.49\textwidth]{period_scan_P_psi.eps}
\caption{Period scan for data from 2017 (October only) and 2012. Each point represents one trial model that converged to given $P_\phi$, $P_\psi$, and $\chi^2$ values. Minima for 2012 and 2017 data are clearly separated. Relative $\chi^2$ was normalized to have the minimum value of 1.}
\label{fig:period_scan}
\end{figure*}
\subsection{Model from 2012 and 2017 data}
\label{sec:model_2app}
\begin{figure*}[t]
\centering
\includegraphics[width=0.7\textwidth, trim=2.5cm 8.5cm 1cm 1cm, clip]{model_2012.png}\\
\includegraphics[width=0.7\textwidth, trim=2.5cm 8.5cm 1cm 1cm, clip]{model_2017_October.png}
\caption{Shape models reconstructed independently from 2012 (top) and 2017 (bottom) light curves. The 2017 model is also almost identical to that reconstructed from joint inversion of 2012 and 2017 data.}
\label{fig:shape_models_comparison}
\end{figure*}
The 3D models reconstructed from the two apparitions independently have similar global shapes but their details are different (Fig.~\ref{fig:shape_models_comparison}). If we take the shape reconstructed from the 2017 data and use it to fit the 2012 data, it gives a satisfactory fit to light curves. However, a better way how to use all data together is not to treat them separately but to instead invert both apparitions together with a common shape model that would differ only in kinematic parameters. Therefore, we modified the original inversion code of \cite{Kaa:01} to enable including two independent light curve sets. We assumed that the only parameters that were different for the two apparitions were the rotation and precession periods $P_\psi$, $P_\phi$ and initial Euler angles $\phi_0$, $\psi_0$. Parameters describing the shape and the direction of the angular momentum vector were the same. So the full set of kinematic parameters for a two-apparition model was: $(\lambda, \beta, \phi_0^{(1)}, \phi_0^{(2)}, \psi_0^{(1)}, \psi_0^{(2)}, P_\psi^{(1)}, P_\psi^{(2)}, P_\phi^{(1)}, P_\phi^{(2)},I_1, I_2)$, where the superscript $(1)$ is for 2012 apparition and $(2)$ is for 2017 apparition. The two light curve data sets were independent in the sense that the integration of kinematic equations \citep[eq.~A.3 in][]{Kaa:01} was done separately for 2012 and 2017 data, the two epochs were not directly connected. The final model shape is almost identical to that reconstructed from only 2017 data shown in Fig.~\ref{fig:shape_models_comparison}. The fit of the final model to individual light curves is shown in Figs.~\ref{fig:lc_fit_2012_1}--\ref{fig:lc_fit_2017_5}. Rotation and precession periods converged to practically the same values as with the independent treatment of each of two apparitions. The best-fit parameters for the 2012 and 2017 apparitions are listed in Table~\ref{tab:parameters}. The physical models of 2012 TC4 from 2012 and 2017 and the light curve data set used to reconstruct these models are available from the DAMIT database\footnote{https://astro.troja.mff.cuni.cz/projects/damit/} \citep{Dur.ea:10}.
\begin{table*}[t]
\caption{Parameters of the model in Fig.~\ref{fig:shape_models_comparison} reconstructed from 2012 and 2017 light curves. The reported errors correspond to standard deviations of parameters computed from bootstrap models. Parameters without errors were fixed (Hapke's parameters) or did not change ($I_1$ and $I_2$). Formally small uncertainty of $\psi_0$ means that this initial orientation angle is correlated with the shape and does not change significantly with different bootstrap data sets. Parameter $\lambda$ is practically unconstrained because the angular momentum direction is very close to $\beta = -90^\circ$. Values $\delta L/L$ and $\delta E/E$ are relative changes of angular momentum $L$ and energy $E$ between 2012 and 2017, i.e., $\delta L/L = (L_{2017} - L_{2012}) / L_{2017}$ and $\delta E/E = (E_{2017} - E_{2012}) / E_{2017}$.}
\label{tab:parameters}
\centering
\begin{tabular}{l c c}
\hline\hline
& 2012 & 2017 \\
\hline
$P_\psi$ [min] & $27.8720 \pm 0.0007$ & $27.5070 \pm 0.0002\ $ \\
$P_\phi$ [min] & $\ 8.4944 \pm 0.0005$ & $\ 8.47511 \pm 0.00008$ \\
$\phi_0$ [deg] & $322 \pm 8$ & $74 \pm 9$ \\
$\psi_0$ [deg] & $198 \pm 0.02$ & $180 \pm 0.2$ \\
JD$_0$ & 2456210.00 & 2458032.69 \\
\hline
$\lambda$ [deg] & \multicolumn{2}{c}{$103 \pm 78$} \\
$\beta$ [deg] & \multicolumn{2}{c}{$-88.5 \pm 0.7$} \\
$I_1$ & \multicolumn{2}{c}{0.42} \\
$I_2$ & \multicolumn{2}{c}{0.81} \\
$w$ & \multicolumn{2}{c}{0.67} \\
$g$ & \multicolumn{2}{c}{$-0.20$} \\
$h$ & \multicolumn{2}{c}{0.062} \\
$B_0$ & \multicolumn{2}{c}{0.6} \\
$\bar{\theta}$ [deg] & \multicolumn{2}{c}{28} \\
$\delta L/L$ & \multicolumn{2}{c}{$0.00078 \pm 0.00006$} \\
$\delta E/E$ & \multicolumn{2}{c}{$-0.0035 \pm 0.0001$} \\
\hline
\end{tabular}
\end{table*}
\subsection{Bootstrap} \label{boo}
To estimate uncertainties of physical periods, and to further robustly demonstrate that their change between 2012 and 2017 is significant, we created bootstrapped data samples and repeated the light curve inversion for them. From both 2012 and 2017 data sets, we created 1000 bootstrap samples by randomly selecting the same number of light curves from the original data set. For October 2017 bootstrap, 279 final shape models had clearly wrong inertia tensor that was not consistent with the kinematic $I_1, I_2$ parameters, so we removed them from the analysis. For all remaining bootstrap models, we plot histograms of $P_\phi$ and $P_\psi$ distribution in Fig.~\ref{fig:histogram_periods}. The standard deviations of the precession period $P_\phi$ are $0.0005\,$min and $0.00009\,$min for 2012 and 2017 data, respectively, which are similar to uncertainty values derived in Sects.~\ref{sec:model_2017} and \ref{sec:model_2012}. For the rotation period $P_\psi$, these standard deviations are $0.0008$ and $0.0002$\,min, which is significantly smaller than our $\chi^2$-based estimate. Nevertheless, the difference between periods determined from 2012 and 2017 apparitions is much larger than their uncertainty intervals and we are not aware of any random or model errors that could cause such difference. Our conclusion is that the spin state of TC4 has changed from 2012 to 2017, rotation and precession periods have decreased. In what follows, we try to interpret this change using a theoretical model of TC4 spin evolution with the relevant torques.
\begin{figure*}[t]
\includegraphics[width=0.49\textwidth]{bs_hist_P_phi.png}
\includegraphics[width=0.49\textwidth]{bs_hist_P_psi.png}
\caption{Distribution of periods $P_\phi$ (left) and $P_\psi$ (right) for bootstrapped light curves.}
\label{fig:histogram_periods}
\end{figure*}
\input{dv}
\section{Conclusions}
Photometric data of 2012~TC4 collected during its two close approaches to the Earth in 2012 and 2017 clearly show that this asteroid is in an excited rotation state. Fourier analysis of 2012 and 2017 data sets finds two unique periods in the signal and these two periods are significantly different, which means that the rotation state of TC4 must have changed slowly between 2012 and 2017 or suddenly during the 2012 flyby (the 2017 flyby was after photometric observations). This detection of period change is robust and not model-dependent.
The periods detected by Fourier analysis were used to constrain physical rotation and precession periods of the tumbling rotation state. We found only one physically acceptable solution that fits photometric data, namely free-tumbling situation about the shortest axis of the inertia tensor (SAM mode). When modeling light curve sets from 2012 and 2017 separately, the shape models are similar with about the same direction of angular momentum vector but the rotation and precession periods are significantly different and there is no combination of parameters that would provide an acceptable fit to the whole data set. The change of physical periods of tumbling is consistent with the change revealed by Fourier analysis. Including this period change into our model, we were able to fit all available photometry from both apparitions. The difference in periods for 2012 and 2017 apparitions is much larger than any possible random or model error, so this model-based detection of periods change is significant and robust.
Having detected the period change and creating a physical model of TC4, we looked for a possible explanation for this change. First, we show that the effect of close encounter in 2012 on the rotation state was negligibly small compared to the detected change of the rotation state. Second, we show that a plausible explanation is the YORP effect -- the numerical simulation of the rotation dynamics based on our shape model of TC4 gives a general agreement with observed periods change. Although the match is not ideal, we believe that the discrepancy is caused by simplification in our YORP model and uncertainties in the shape model and other parameters. We also show that the other two possible mechanisms that could affect the rotation state -- namely the internal energy dissipation and impacts of interplanetary particles -- are too small to cause the measured effect, so YORP remains the only plausible explanation of the observed change of the rotation state of 2012~TC4. Accepting this explanation, this is the first detection of YORP acting on a tumbling asteroid.
\acknowledgements
This research is supported by Korea Astronomy and Space Science Institute (KASI).
The work at Charles University and Ond\v{r}ejov Observatory was supported by the Czech Science Foundation (grant 20-04431S).
The work by C.-H. Kim was supported by the Basic Science Research Program through the National Research Foundation of Korea funded by the Ministry of Education (2018R1D1A1A09081827, 2020R1A4A2002885).
This research has made use of the KMTNet system operated by the Korea Astronomy and Space Science Institute (KASI) and the data were obtained at SAAO in South Africa.
We appreciate all astronomers who uploaded or shared their previously published light curves of 2012 TC4 used in this paper.
\section{Rotational dynamics of 2012~TC4} \label{methods}
In this Appendix, we summarize variables and mathematical approach
used for propagation of 2012~TC4 rotation state in between the 2012 and 2017
epochs (Section~\ref{sec:theory}). This is obviously a classical piece of mechanics which can be found
in many textbooks \citep[e.g.,][]{Lan.Lif:69, Gol:80}. For that
reason we keep our description to a very minimum.
The kinematical part of the problem describes orientation of the
asteroid in the inertial frame. For simplicity we assume the asteroid is a rigid
body, allowing us to define unambiguously a proper body-fixed frame. The
easiest choice has (i) the origin in the asteroid's center-of-mass, and (ii)
the axes coinciding with the principal axes of the inertia tensor $\mathbf{I}$
(therefore $\mathbf{I}={\rm diag}(A,B,C)$, with $A\leq B\leq C$). Transformation
between the inertial and body-fixed frames is conventionally parametrized by a set
of Euler angles, most often the 3-1-3 sequence of the precession angle $\phi$,
the nutation angle $\theta$, and the angle of proper rotation $\psi$. However,
instead of the three Euler angles $(\phi,\theta,\psi)$ we use here the
Rodrigues-Hamilton parameters $\mbox{\boldmath$\lambda$}=(\lambda_0,\lambda_1,
\lambda_2,\lambda_3)$ \citep[e.g.,][]{Whi:1917}. Their relation to the Euler angles
is given by: (i) $\lambda_0+\imath \lambda_3 = \cos\frac{\theta}{2}\,\exp\left[
\frac{\imath}{2} \left(\psi+\phi\right)\right]$, and (ii) $\lambda_2+\imath \lambda_1
= \imath \sin\frac{\theta}{2}\,\exp\left[\frac{\imath}{2} \left(\psi-\phi\right)\right]$
($\imath$ is a complex unit). One can easily verify a constraint: $\lambda_0^2+
\lambda_1^2+\lambda_2^2+\lambda_3^2=1$ (in our numerical runs satisfied with a
$\leq 10^{-13}$ accuracy). The sacrifice of using four instead of three parameters
pays off in at least two advantages. First, parametrization by Euler angles is
unstable when, or near, $\sin\theta=0$ state. No such problem occurs when using
the Rodrigues-Hamilton parameters which provide a uniformly non-singular description
of the rotation. Second, Euler-angle parametrization necessarily
requires use of trigonometric functions. Instead, manipulation with the
Rodrigues-Hamilton parameters is limited to simple algebraic functions, in fact
quadratic at maximum, as shown below in Eqs.~(\ref{e1}) and (\ref{e2}). For that
reason the use of the four Rodrigues-Hamilton parameters does not even slow down
the computations in a noticeable way.
The rotation matrix $\mathbf{A}$ needed for the vector transformation from the
inertial frame to the body-fixed frame is a simple quadratic form of
$\mbox{\boldmath$\lambda$}$, namely
\begin{equation}
\mathbf{A} = \left(
\begin{array}{ccccc}
\lambda_0^2+\lambda_1^2-\lambda_2^2-\lambda_3^2 & , & 2\left(\lambda_0\lambda_3+
\lambda_1\lambda_2\right) & , & 2\left(\lambda_1\lambda_3-\lambda_0\lambda_2\right) \\
2\left(\lambda_1\lambda_2-\lambda_0\lambda_3\right) & , & \lambda_0^2+\lambda_2^2-
\lambda_1^2-\lambda_3^2 & , & 2\left(\lambda_0\lambda_1+ \lambda_2\lambda_3\right) \\
2\left(\lambda_0\lambda_2+ \lambda_1\lambda_3\right) & , & 2\left(\lambda_2\lambda_3-
\lambda_0\lambda_1\right) & , & \lambda_0^2+\lambda_3^2-\lambda_1^2-\lambda_2^2 \\
\end{array}
\right) \; . \label{e1}
\end{equation}
The inverse transformation is represented by a transposed matrix $\mathbf{A}^{\rm T}$.
Asteroid's rotation is represented with the angular velocity vector $\mbox{\boldmath$
\omega$}$, whose components in the body-fixed frame are $(\omega_1,\omega_2,\omega_3)$.
Their relation to the time derivatives of the Rodrigues-Hamilton parameters is
simply
\begin{equation}
\frac{d \mbox{\boldmath$\lambda$}}{dt} = \frac{1}{2}\,\mathbf{P}\cdot
\mbox{\boldmath$\lambda$} \; , \label{e2}
\end{equation}
where
\begin{equation}
\mathbf{P} = \left(
\begin{array}{ccccccc}
0 & , & -\omega_1 & , & -\omega_2 & , & -\omega_3 \\
\omega_1 & , & 0 & , & \phantom{-}\omega_3 & , & -\omega_2 \\
\omega_2 & , & -\omega_3 & , & 0 & , & \phantom{-}\omega_1 \\
\omega_3 & , & \phantom{-}\omega_2 & , & -\omega_1 & , & 0 \\
\end{array}
\right) \; . \label{e3}
\end{equation}
This explicitly linear differential equation for $\mbox{\boldmath$\lambda$}$ cannot
be solved in a trivial way, because $\mathbf{P}=\mathbf{P}\left(\mbox{\boldmath$
\omega$}\right)$, and the angular momentum vector is a time-dependent variable.
The antisymmetry of $\mathbf{P}$ implies conservation of the above mentioned
quadratic constraint of $\mbox{\boldmath$\lambda$}$.
The dynamical part of the problem expresses the Newton's principle that a change
of the rotational (intrinsic) angular momentum $\mathbf{L} = \mathbf{I}\cdot\mbox{
\boldmath$\omega$}$ is given by the applied torque $\mathbf{M}$. Tradition has it to
state this rule in the body-fixed frame, where $\mathbf{I}$ is constant and even
diagonal in our choice of axes, such that
\begin{equation}
\frac{d\mathbf{L}}{dt} + \mbox{\boldmath$\omega$}\times \mathbf{L} = \mathbf{M}
\; . \label{e4}
\end{equation}
Equations (\ref{e2}) and (\ref{e4}) define the problem of asteroid's rotation in our
set of seven parameters $(\mbox{\boldmath$\lambda$},\mbox{\boldmath$\omega$})$. Once
the torques $\mathbf{M}$ are specified, we numerically integrate this system of
differential equations with the initial data determined from the set of observations (either forward in time if the 2012 data are used, or backward in time if the 2017 data are used). We use Burlish-Stoer integration scheme with tightly controlled accuracy.
We also note another useful quantity, namely the energy of rotational motion about
the center given by $E=\frac{1}{2}\,\mbox{\boldmath$\omega$}\cdot \mathbf{L}$. In a
classical problem of a free top (i.e., $\mathbf{M}=0$), both $E$ and $\mathbf{L}$
in the inertial frame are conserved. In the body-fixed frame only $L=|\mathbf{L}|$
is constant. Nevertheless, conservation of $E$ and $L$ (together with the principal values of the inertia tensor $A$, $B$ and $C$) uniquely determines the
wobbling trajectory of $\mathbf{L}$ in the body-fixed frame \citep[e.g.,][]{Lan.Lif:69, Dep.Eli:93}.
There are two options for this motion: (i)
short-axis mode (SAM),
when $\mathbf{L}$ circulates about $+z$ or $-z$ body axis, or (ii) long-axis (LAM),
mode $\mathbf{L}$ circulates about $+x$ or $-x$ body axis. A useful discriminator
of the two is yet another conserved and non-dimensional quantity in the free-top
problem, namely $p=2BE/L^2$: (i) SAM is characterized by $p$ values in between
$\beta=B/C$ and $1$, while (ii) LAM is characterized by $p$ values in between $1$
and $\alpha=B/A$. Note that $\Delta=B/p$ plays an important role in description
of the free-top problem using Hamiltonian tools \citep[e.g.,][]{Dep.Eli:93,
Bre.ea:11}. The free-top motion of $\mathbf{L}$ in the body fixed frame
is easily integrable using Jacobi elliptic functions. When plugged in the
kinematical equations (\ref{e3}), one can also obtain solution for $\mbox{\boldmath$
\lambda$}$ or the Euler angles $(\phi,\theta,\psi)$ \citep[e.g.,][]{Lan.Lif:69}.
Those of $\psi$ and $\theta$ are strictly periodic with a period (SAM mode relevant
for 2012~TC4 assumed here)
\begin{equation}
P_\psi = \frac{C}{L}\, \frac{4\beta K\left(k\right)}{\sqrt{\left(1-\beta\right)
\left(\beta/\alpha-p\right)}}\; , \label{e5}
\end{equation}
where $K(k)$ is complete elliptic integral of the first kind with the modulus $k$
given by
\begin{equation}
k^2 = \frac{\left(\beta/\alpha-1\right)\left(p-\beta\right)}{\left(1-\beta\right)
\left(\beta/\alpha-p\right)} \label{e6}
\end{equation}
(the motion of $\theta$ has a periodicity $P_\psi/2$). The motion of the precession
angle $\phi$ is not periodic. Nevertheless a fully analytical solution still exists
and it is composed of two parts, the first of which has periodicity $P_\psi$ and a
second has another periodicity, generally incommensurable with $P_\psi$ \citep[e.g.,][]{Lan.Lif:69}.
Yet it is both practical and conventional to define an approximate periodicity
$P_\phi$ of $\phi$. We use definition of \citet{Kaa:01}, Eq.~(A.11), namely numerically determining an advance in $\phi$ angle over $P_\psi$ period (this, in principle, averages out contribution of the $P_\psi$-periodic part in the $\phi$ solution). When weak torques are applied,
the free-top solution represents still a very useful (osculating) template with all
above-discussed variables such as $E$, $L$, $p$, $P_\psi$ or $P_\phi$ adiabatically
changing in time.
Finally, we discuss the torques used in our analysis. The first class is due to
gravitational tidal fields of the Sun and the Earth. Assume a point mass source
$M$ specified in the body-fixed frame of the asteroid with a position vector
$\mathbf{R}$. Using the quadrupole part of the exterior perturber tidal field
we have \citep[e.g.,][]{Fit:70, Tak.ea:13}
\begin{equation}
\mathbf{M}_{\rm grav} = \frac{3GM}{R^5}\,\mathbf{R}\times \left(\mathbf{I}
\cdot\mathbf{R} \right)\; . \label{e7}
\end{equation}
We neglect the formally dipole part of the tidal field, which could only
occur if the true center-of-mass of the asteroid is slightly displaced from the
assumed location \citep[determined by using the assumption of homogeneous density;
see, e.g.,][]{Tak.ea:13}. Note the position of all bodies, the
asteroid, the Sun and the Earth, are primarily determined using the numerical
integration of the orbital problem in the inertial frame (or, actually, displaced
heliocentric frame). In our case, we numerically integrated planetary orbits,
including the Earth, and 2012~TC4 in the heliocentric system by taking initial
data from NEODyS website (\url{https://newton.spacedys.com/neodys/}). We output the
necessary positions every $170$~s, enough for the purpose of the 2012~TC4's
rotation dynamics. We also compared our solution with that available at the JPL
Horizons system (\url{http://ssd.jpl.nasa.gov/?horizons}), and found a very good
correspondence with tiny differences, not meaningful for our application.
The relative position $\mathbf{R}$ in (\ref{e7}) is determined by (i) the
difference of the corresponding bodies in our orbital solution, and (ii)
transformation to the body-fixed
frame. As a result $\mathbf{M}_{\rm grav} = \mathbf{M}_{\rm grav}\left(t,
\mbox{\boldmath$\lambda$}\right)$. The steep dependence $M_{\rm grav}\propto
R^{-3}$ implies that the Earth effect is non-negligible only during the close
encounters to this planet (e.g., Fig.~\ref{f1}).
Our analysis also includes the radiation torque known as the YORP effect
\citep[e.g.,][]{Bot.ea:06, Vok.ea:15}. Because of the tumbling
rotation state of 2012~TC4, we resort to the simplest variant, namely a limit
of zero thermal inertia \citep[the effects of finite thermal inertia were studied
only for objects rotating about the shortest axis of the inertia tensor so far;
e.g.,][]{Cap.Vok:04, Gol.Kru:12}. In this
approximation the radiation torque is given by \citep[e.g.,][]{Rub:00, Vok.Cap:02}
\begin{equation}
\mathbf{M}_{\rm YORP}= -\frac{2 F}{3c}\int H\left[\mathbf{n}\cdot\mathbf{n}_0\right]
\left(\mathbf{n}\cdot\mathbf{n}_0\right)\,\mathbf{r}\times d\mathbf{S}\; , \label{e8}
\end{equation}
where $F$ is the solar radiation flux at the location of the asteroid and $c$
is the light velocity. The integral in (\ref{e8}) is performed over the surface
of the body characterized by an ensemble of outward-oriented surface elements
$d\mathbf{S}=\mathbf{n}\,dS$, $\mathbf{n}$ is the normal to the surface and
$\mathbf{r}$ is the position of the surface element with respect to the origin
of the body-fixed frame. The unit vector of the solar position in the body fixed
frame is denoted with $\mathbf{n}_0$, and $H[x]$ is the Heaviside step function
(its presence in the integrand of (\ref{e8}) implies that a non-zero contribution
to the radiation torque is provided by surface units for which the Sun is
above local horizon). In fact, our code includes even more complex feature
of self-shadowing of surface units, but this is not active in the case of
2012~TC4 whose resolved shape is convex. The factor $2/3$ on the right hand
side of Eq.~(\ref{e8}) is due to the assumption of Lambertian reradiation
from the surface. More complicated assumptions about directionality of the thermally emitted radiation from the surface, such as the beaming effects \citep[e.g.,][]{Roz.Gre:12}, are presently not implemented in the code. The lightcurve inversion obviously allows only a finite
accuracy in shape determination of the body, typically a convex polyhedron
with little more than thousand surface facets. It is known that already this fact is an obstacle to an accurate evaluation of the YORP torque, which may sensitively depend on smaller-scale surface irregularities, not resolved by our shape model. The formal integration in
Eq.~(\ref{e8}) is therefore represented with a summation over the surface units
of the resolved shape model. We use algebra from \cite{Dob:96} to
determine all necessary variables. This also means we assume constant density
distribution in the body.
\section{Theory}
\label{sec:theory}
\subsection{Orbital dynamics}
The unique observational opportunities of 2012~TC4 are directly related to its
exceptional orbit. The asteroid had a deep encounter with the Earth on October~12, 2012,
during which the closest distance to the geocenter was approximately $95,000$~km
(Fig.~\ref{f1}).
However, the more unusual circumstance was that the 2012 close encounter resulted
in a change of the 2012~TC4's orbital semimajor axis which placed it nearly
exactly to the 5:3 resonance with the Earth heliocentric motion. As a result,
in five years after the first close encounter, i.e. on October~12, 2017, the relative
configuration of the asteroid and the Earth nearly exactly repeated, placing it again
in a deep encounter configuration. This time the closest approach to the geocenter
had even closer distance of $50,200$~km. Astrometric observations during
the two close approaches, including Arecibo and Green Bank radar data taken in 2017,
allowed a very
accurate orbital solution over the five year period of time in between 2012
and 2017. In the context of this paper, we note that it also provided an interesting
information about nongravitational effects needed to be empirically included in orbit
determination. Adopting methodology from cometary motion (e.g., \cite{Mars.ea:73} and
\cite{Far.ea:13} or \cite{Mom.ea:14} for the asteroidal context),
we note the following values of radial $A_1$ and transverse $A_2$ accelerations:
(i) $A_1 = (2.17\pm 0.80)\times 10^{-11}$ au~d$^{-2}$, or $(4.35\pm
1.60) \times 10^{-10}$ m~s$^{-2}$, and (ii) $A_2 = -(2.73\pm 0.65)\times
10^{-13}$ au~d$^{-2}$ (both assume $\propto r^{-2}$ heliocentric decrease;
see JPL/Horizons web page \url{https://ssd.jpl.nasa.gov/sbdb.cgi}). At a first sight,
these values appear very reasonable \citep[compare, e.g., with a $A_1$ and
$A_2$ fits for a $\simeq 4$~m body 2009~BD,][]{Mom.ea:14,Vok.ea:15}.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=12cm]{f1.eps}
\end{center}
\caption{Geocentric distance (top) and relative velocity (bottom)
of 2012~TC4 during its close encounter with Earth on October~12,
2012 (nominal minimum configuration at MJD56212.229, gray vertical
line); the abscissa shows time in days with respect to MJD56212.
The dashed horizontal line at the top panel shows the minimum distance
of $\simeq 95,000$~km. The asymptotic value of the relative velocity
with respect to the Earth, $\simeq 6.5$~km/s, increases to more than
$7.1$~km/s by the Earth gravity.}
\label{f1}
\end{figure}
If we were to interpret both components as a result of radiation
forces, we may further obtain useful information about the body. The
radial component would represent the direct solar radiation pressure.
In a simple model, where we assume a spherical body of size $D$ and bulk density
$\rho$, we have $A_1\simeq 3C_{\rm R} F_0/(2\rho D c)$, with $C_{\rm R}$
the radiation pressure coefficient (dependent on sunlight scattering
properties on the surface), $F_0\simeq 1367$ W~m$^{-2}$ the solar constant and
$c$ the light velocity. Adopting $C_{\rm R}\simeq 1.2$ and $D\simeq 10$~m,
we obtain a very reasonable bulk density $\rho\simeq (1.9\pm 0.7)$ g~cm$^{-3}$.
The transverse component of nongravitational orbital effects makes sense
when interpreted as the Yarkovsky effect \citep[e.g.,][]{Vok.ea:15}.
Note that the above mentioned value of $A_2$ translates to a secular
change of the semimajor axis $da/dt = -(110\pm 26)\times 10^{-4}$ au~My$^{-1}$ \citep[see,
e.g.,][]{Far.ea:13}. Given the near extreme obliquity of the
rotational angular momentum, we may safely restrict to the diurnal component
of the Yarkovsky effect. The negative value of $A_2$, or $da/dt$, corresponds
well to the retrograde sense of 2012~TC4 rotation (implied by the direction of
the rotational angular momentum vector, Sec.~\ref{sec:physical_model}). Next, borrowing the simple model for
a spherical body from \cite{Vok:98}, and fixing the size $D=10$~m, the
inferred bulk density $\rho\simeq 1.9$ g~cm$^{-3}$ and the surface thermal
conductivity $K=0.05$ W~m$^{-1}$~K$^{-1}$, we estimate that the
corresponding surface thermal inertia is $\Gamma\simeq 490^{+270}_{-250}$ in SI units \citep[though, we note there is also a lower-inertia solution possible like in][]{Mom.ea:14}. This is a very adequate value too \citep[e.g.,][]{Del.ea:15}.
Obviously, due to many frozen parameters in the
model (and its simplicity), the realistic uncertainty in $\Gamma$ would be
larger. However, it is not our intention to fully solve this problem. We
satisfy ourselves with observation that the needed empirical nongravitational
accelerations in the orbital fit may be very satisfactorily interpreted as
radiation-related effects. In the next sections, we show that also the change
in the rotation state in between 2012 and 2017 observation epochs may be
very well explained by the radiation torque known as the
Yarkovsky-O'Keefe-Radzievskii-Paddack effect \citep[e.g.,][]{Vok.ea:15}.
\subsection{Rotational dynamics}
Our methods and mathematical approach used to describe evolution of the rotation
state of 2012~TC4 are presented in the Appendix~\ref{methods}, therefore here we provide just a general outline. We numerically integrated Euler equations (\ref{e2}) and
(\ref{e4}) describing spin state evolution of 2012~TC4 in between the
observation runs in 2012 and
2017. The kinematical part, describing transformation between the
inertial frame and the body-frame defined by principal axes of the
tensor of inertia, was parametrized by the Rodrigues-Hamilton
parameters $\mbox{\boldmath$\lambda$}=(\lambda_0,\lambda_1,\lambda_2,\lambda_3)$
\citep[see, e.g.,][]{Whi:1917}. This choice helps to remove problems
related to coordinate singularity given by zero value of the nutation
angle. The dynamical part is represented by evolution of the angular velocity
$\mbox{\boldmath$\omega$}$ in the body-frame. Note that in most cases of
asteroid lightcurve interpretation, a simplified model of a free-top would
be sufficient (used also in Sec.~\ref{sec:physical_model} to fit the 2012 and 2017 data separately). However, the evidence of change of the rotation state of
2012~TC4 in between the 2012 and 2017 epochs, discussed above, requires appropriate
torques to be included in the model. We addressed two effects:
\begin{itemize}
\item gravitational torques due to the Sun and the Earth, and
\item radiation torques due to the sunlight scattered by the surface
and thermally re-radiated \citep[the YORP effect; e.g.,][]{Bot.ea:06,Vok.ea:15}.
\end{itemize}
The tidal gravitational fields of the Sun and the Earth were represented
in the body-frame using the quadrupole approximation (\ref{e7}) \citep[e.g.,][]{Fit:70,Tak.ea:13}. The nature of the perturbation is different
for the Sun and the Earth. In the solar case, the gravitational torque results in
a small tilt by less than $1$ arcminute describing a small segment on the
precession cone. The effect of the Earth-induced gravitational torques
manifests as an impulsive effect only during close encounters. Our main
goal was to verify that, due to fast rotation of 2012~TC4, the effect
averages out and cannot contribute to the observed change in rotational
frequencies. Indeed, Fig.~\ref{f2} shows change of the osculating rotational
(intrinsic) angular momentum and energy during the 2012 close encounter
with the Earth (only about six times larger effects are observed during the closer
encounter in 2017, but this is not relevant for our analysis anyway, because
observations preceded this approach). The initial data of the simulation were taken
from the observations fit in 2012, namely before the encounter. Recall that
the characteristic timescale
of the encounter is about half a day (determined, for instance, as a width
of characteristic Earth relative velocity increase at the bottom panel of
Fig.~\ref{f1} at $\simeq 6.8$~km s$^{-1}$ level, half value between the asymptotic and
peak velocities). In comparison, the characteristic rotational periods $P_\psi$
and $P_\phi$ are of the order of minutes, i.e. much shorter. As a result,
the effect of the Earth gravitational torque efficiently averages out during
each of the rotation cycles \citep[a significant effect may be expected only for
very slowly rotating bodies and sufficiently deep encounters, such as
seen for 4179 Toutatis during its 2004 close encounter with the Earth; e.g.,][]{Tak.ea:13}. Therefore, we may conclude that the gravitational
torques cannot explain the observed change in the rotation state of 2012~TC4
in between the 2012 and 2017 epochs. Still, we keep them in our model for
the sake of completeness.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=12cm]{f2.eps}
\end{center}
\caption{The effect of the Earth gravitational torque in quadrupole
approximation on rotation state parameters of 2012~TC4 during its close
encounter on October~12, 2012 (gray line denotes the nominal minimum
distance configuration). The abscissa shows time in days with respect
to MJD56212 (as in Fig.~\ref{f1}). The upper panel shows a fractional
change of the rotational angular momentum $\delta L=L-L_0$, normalized by
the initial value $L_0$, the bottom panel panel shows a fractional change
of the rotational energy $\delta E = E - E_0$, normalized by the initial
value $E_0$ (note that
both ordinate scales are in $10^{-6}$). The resulting change of both
parameters after the encounter is $\leq 5\times 10^{-9}$.}
\label{f2}
\end{figure}
\begin{figure}[tp]
\begin{center}
\includegraphics[width=12cm]{f3.eps}
\end{center}
\caption{The effect of the radiation (YORP) torque on rotation state parameters
of 2012~TC4 in the time interval between the two recent close encounters with
the Earth (i.e., October~12, 2012 and October~12, 2017). Nominal model is used here to
obtain a first insight of the expected order of magnitude of the perturbation.
The abscissa shows time in years. The upper panel shows a fractional
change of the rotational angular momentum $\delta L=L-L_0$, normalized by
the initial value $L_0$, the bottom panel panel shows a fractional change
of the rotational energy $\delta E = E - E_0$, normalized by the initial
value $E_0$ (note that both ordinate scales are in $10^{-3}$). Both reference
values taken at MJD58032.19, mean epoch of 2017 observations, and the asteroid's
rotation state was propagated backward in time to October 2012.}
\label{f3}
\end{figure}
The radiation torques are of a quite different importance. It is well-known
that the YORP effect is able to secularly change rotational frequency and
tilt the rotational angular momentum in space. While mostly studied in the
limit of a rotation about the shortest axis of the inertial tensor, generalizations
to the tumbling situation were also developed. Both numerical \citep{Vok.ea:07}
and analytical \citep{Cic.Sch:10, Bre.ea:11} studies
confirmed that YORP effect is able to change rotational angular momentum, its
orientation in both the inertial space and body-frame in an appreciable manner.
As usual, the effect is more important on small bodies. Here we included the
simple, zero-inertia limit developed in \cite{Rub:00} and \cite{Vok.Cap:02},
see Eq.~(\ref{e8}). A first look into importance of the radiation
torques may be obtained by taking nominal solution of the rotation state from
the 2017 observations, including the appropriate shape model, and propagate it
backwards in time to early October 2012 (epoch of the first set of observations).
We use the more accurate 2017 model as reference rather than the one constructed
from poorer 2012 observations. The length-scale of the model was adjusted to
correspond to an equivalent sphere of diameter $D=10$~m and density was
$\rho=1.4$ g~cm$^{-3}$. We verified that results have the expected invariance to
rescaling of both $\rho$ and $D$ such that $\rho\,D^2=$~const. Our nominal choice of the $1.4$ g~cm$^{-3}$ bulk density, albeit conflicting with the suggested E-type specral classification of TC4, is therefore linked to the assumed equivalent size of $10$~m, but it might
be redefined according to the rescaling principle. This combination of parameters provides a very nice match of the $P_\phi$ period change due to our YORP model (see Sec.~\ref{sec:theory_res}).
Figure~\ref{f3} provides
information about secular change in rotational angular momentum $L$ (top) and
energy $E$ (bottom) in that simulation. Here we see a long-term change in both
quantities. The wavy pattern is due to eccentricity of the 2012~TC4 orbit
and a stronger YORP torque at perihelion. The accumulated fractional change in both $L$
and $E$ is few times $10^{-3}$. This is promising, because the observed change in these quantities in of the same order of magnitude (see Table~\ref{tab:parameters}), and makes us believe that the change in the directly observable $P_\psi$ and $P_\phi$ periods will also be as needed.
Nevertheless, we also note a difference. The simulation results shown in Fig.~\ref{f3} indicate both angular momentum and energy increased from 2012 to 2017. Such a behaviour is perhaps expected at the first place (for instance, in the case of a body in a principal-rotation state, YORP would necessarily affect both $E$ and $L$ in the same way). However, rotation state solutions from observations in 2012 and 2017 tell us something else (see Table~\ref{tab:parameters}): the rotational angular momentum $L$ increased in between 2012 and 2017, while the energy $E$ decreased. Before commenting more on this difference, we first provide more detailed analysis of the radiation torque effects for 2012~TC4 in our modeling, this time using the whole suite of acceptable initial data and shape models (all compatible with the observations; Sec.~\ref{boo}). This will allow a statistical assessment of the predicted values.
\subsection{Results} \label{sec:theory_res}
An ideal procedure of proving that the observed changes in tumbling-state periods
$P_\phi$ and $P_\psi$ are due to the radiation (YORP) torques would require a
highly-reliable theoretical model (numerical propagation of 2012~TC4's rotation
state with appropriate torques included) employed to fit all available
observations (in our case data from October 2012 and 2017). Obviously, the
only ``comfort'' of this analysis would be to possibly adjust some
free (unknown) parameters. Unfortunately, such a plan is presently too ambitious,
and thus we resort to a simpler way.
Recall the much easier situation when the YORP effect has been searched
(and detected) for asteroids rotating in the lowest-energy mode, namely about the
shortest principal axis of the inertia tensor. In this case, YORP results in
a secular change of the unique rotation period $P$ (as in $P_\phi$ and $P_\psi$,
when the body tumbles). The measurements are rarely precise enough, and the
effect strong enough, to directly reveal the change in the period $P$
\citep[see, though, an exception for 54509 YORP,][]{Low.ea:07}. More often, one
uses the fact that the linear-in-time change in $P$ produces a quadratic-in-time
effect in the rotation-phase. Linking properly the asteroid rotation phase
over many observation sessions with an empirical quadratic term helps to characterize
changes of $P$ which are individually too small to be determined from
one-apparition observations \citep[see, e.g.,][]{Vok.ea:15}. This approach
adopts an empirical magnitude of the quadratic term in the rotation phase and simply
solves it as a free parameter (not combining it with a theoretical model at that
stage). Interpretation in terms of the YORP effect is done only aposteriori, when
the fitted amplitude of the phase-quadratic term is compared with a prediction
from the YORP model. And even then, the comparison is often not simple,
because the model prediction for YORP is known to depend on unresolved
small-scale irregularities of the body shape. At several occasions one had
to satisfy with a factor of few difference accounting for the model inaccuracy.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=15cm]{f4.eps}
\end{center}
\caption{Distribution of periods $P_\phi$ (left) and $P_\psi$ (right) for 687
models of 2012~TC4 from our numerical simulation containing both gravitational
and radiation (YORP) torques. The individual models sample possible initial
orientation of the angular momentum vector $\mathbf{L}$ in the inertial space,
orientation of the body-frame in the inertial space and
slight shape variants of the body. All models were constructed using the October 2017
observations and therefore they are referred to the common epoch MJD58032.19.
Blue histograms are computed from these initial data and they are identical to
those in Fig.~\ref{fig:histogram_periods}. The red histograms correspond to the rotation state of 2012~TC4 at MJD56209.88, nominal epoch of the October 2012 observations (the mean $P_\phi$ value from the observations is shown with the vertical dashed line). Unlike in
Fig.~\ref{fig:histogram_periods}, the 2012 data here were computed from the spin
state vectors obtained from our numerical propagation of the 2012~TC4 rotation starting
in October 2017.}
\label{f4}
\end{figure}
Things are quite more complex when the body tumbles. First of all, it is not clear how
to set the empirical approach from above and apply it in this situation. In the
same time, direct modeling approach is probably even less accurate than in the
case of rotation about the principal axis of the inertia tensor. Not only
the worry about the role of unresolved small-scale irregularities remains, but the present
YORP model is restricted to the zero thermal conductivity limit (see, e.g.,
\cite{Vok.ea:07} for numerical approach, and \cite{Cic.Sch:10,
Bre.ea:11} for analytical studies). Yarkovsky acceleration for tumblers
was evaluated with thermal models \citep[e.g.,][]{Vok.ea:05, Vok.apophis:15},
but in these cases $P_\phi$ and $P_\psi$ were slightly
tweaked to make them resonant (an approach we cannot afford here).
In this situation, we adopted the following simple procedure.
\subsubsection{Model based on 2017 data}
We start with a set of models based uniquely on the most reliable and accurate observations from the 2017 apparition.
In particular, we constructed
687 variants of the 2012~TC4 physical model (Sec.~\ref{sec:model_2017}). They are all
very similar, because they sample tight parameter variations
all resulting in acceptable fits of the data. These represent (i) slightly modified
initial rotation parameters (Euler angles and their derivatives, as well as
inertial space direction of the rotational angular momentum vector), and (ii)
slight shape variants of the body. The initial epoch MJD58032.19 was common to all
variant models. Using these initial data and shape models, we propagated all 687
clone realizations of 2012~TC4 backward in time to the epoch MJD56209.88,
characteristic of the October~2012 observations. We used our numerical approach
described above with both gravitational and radiation (YORP) torques included.
For the latter, we assumed an effective size $D=10$~m (corresponding to a sphere
of the same volume of the models) and bulk density $\rho=1.4$ g~cm$^{-3}$. The above-mentioned rescaling rule, namely invariance to $\rho\,D^2$, may allow us to transform our results to other combinations of $D$ and $\rho$ values.
The evolutionary tracks of angular momentum $L$ and energy $E$ secular changes due to
the YORP torques mostly resemble those from Fig.~\ref{f3}. For some model variants,
which were more different from the nominal one, the slope of the overall secular
change in $L$ and/or $E$ was shallower or steeper. At the initial and final epochs
of our simulations we determined $P_\phi$ and $P_\psi$ periods from a short numerical simulation of a free-top model. We also verified that the
$P_\psi$ values correspond exactly to those provided by the analytical formula
(\ref{e5}). Figure~\ref{f4} shows our results. The blue histograms are for the
initial data, i.e., the October 2017 rotation state. Because the observations
were numerous and of a good quality, the model variants differ only slightly and
the $P_\phi$ and $P_\psi$ distributions are tight (they also match
those from the Fig.~\ref{fig:histogram_periods}). The red histograms in Fig.~\ref{f4} were determined from
the last epoch of our numerical runs and correspond to the predicted rotation
state of 2012~TC4 in October~2012. We note both period appreciably changed, as
already anticipated from preliminary simulation shown in Fig.~\ref{f3}. These
$P_\phi$ and $P_\psi$ distributions are obviously less tight than the initial ones,
reflecting different evolutionary tracks of the individual clone variants. Our
prime interest is to compare the red distribution in Fig.~\ref{f4} (from simulations)
to the red distributions in Fig.~\ref{fig:histogram_periods} (from the 2012 observations, also highlighted withe dashed line for $P_\phi$ period). First,
we note that the match of the $P_\phi$ periods is surprisingly good. The mean value
$8.495$~min of the observations rather well corresponds to the mean value
$8.497$~min of the simulated data, which have comfortably large dispersion of
$0.003$~min to overlap with the observed data; in fact, the shift in $P_\phi$ is even larger than required. Interestingly, the comparison
is not as good in the $P_\psi$ period. The model-predicted value $27.59\pm 0.02$~min
(formal uncertainty) is short to explain the observations which provide on average
$27.87$~min. Still, the model indicates a significant shift from the initial
value $27.5070\pm 0.0002$~min. Nevertheless, to reach the value from the 2012
observations the shift would need to be about $3.7$ times larger. We do not know
the reason for this difference, at all likelihood also related to the misbehaviour in the rotational energy evolution (see above). We suspect that the overly simple modeling
of the YORP effect, such as the surface thermal inertia and/or the unresolved small-scale shape irregularities, play
and important role here (note that a factor of $3$ between the observations and
model-prediction was also seen in the cases when YORP was detected for asteroids
rotating about the principal axis of the inertia tensor). The fact that some deeper aspects of the model are not characterized well, as witnessed by the opposite sign of the energy evolution, implies that our results cannot be
easily reconciled with the observations by a simple rescaling of size $D$ and
bulk density $\rho$. We have verified that the accumulated shifts in both
$P_\phi$ and $P_\psi$ periods are proportional to $\rho\,D^2$. Thus, the $P_\psi$
mismatch could be explained by assuming the 2012~TC4's size is $\simeq 5.2$~m,
but this would produce factor $\simeq 3.7$ inconsistency in $P_\phi$ period
(making the modeled value larger than observed). Instead, we believe that some
missing details of the YORP modeling, which result in different effects on
$P_\phi$ and $P_\psi$, are responsible for the difference.
\subsubsection{Model based on a combination of 2012 and 2017 data}
For sake of
comparison, we also repeated our analysis using models based on a combination of the
observations in 2012 and 2017 (Sec.~\ref{sec:model_2app}). Obviously at each of these epochs we considered different
parameters of the rotation state, but now we enforced the same shape model is used for both
data-sets. This solution gives us an opportunity to consider two sets of
initial conditions for our simulation, both in 2012 and 2017, and analyses
predictions in the complementary epoch (integrating the rotation model once
backward in time and once forward in time). Obviously, in all cases our model
includes the gravitational and radiation torques, $D=10$~m effective size and
$\rho = 1.4$ g~cm$^{-3}$ bulk density as before (needed for the evaluation of
the radiation torques).
\begin{figure}[tp]
\begin{center}
\includegraphics[width=15cm]{f5.eps}
\end{center}
\caption{The same as in Fig.~\ref{f4}, but now for nearly a thousand variant
shape models of 2012~TC4 constructed from a combination of the 2012 and 2017 data.
Data at the initial epoch MJD58032.19 (October 2017) are in blue. They have
been propagated using our dynamical model to MJD56209.88 (October 2012), and
the endstates of these runs served to compute $P_\phi$ and $P_\psi$ shown by
the red histograms.}
\label{f5}
\end{figure}
We start with the case of the initial data in October 2017 and backward-in-time
integration. This is directly comparable with results above, when only observations
in 2017 were used. However, the models are slightly different in all their
aspects (initial conditions and shape), because now the 2012 observations play
the role in their construction. Results are shown in Fig.~\ref{f5}. While
slightly different then in Fig.~\ref{f4}, the principal outcome is the same:
(i) fairly satisfactory prediction for the $P_\phi$ period, while (ii) too small
change in the $P_\psi$ period.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=15cm]{f6.eps}
\end{center}
\caption{The same as in Fig.~\ref{f5} (shape models constrained by both 2012 and
2017 observations), but now propagation was performed from MJD56209.5 epoch
in 2012 (red) to MJD58032.19 in 2017 (blue). The 2012 data constrain the rotation
state solution less accurately and thus result in a larger scatter of the predicted
periods $P_\phi$ and $P_\psi$ in 2017. The vertical dashed line shows mean value
of the $P_\phi$ period from the October 2017 observations.}
\label{f6}
\end{figure}
We next consider the opposite case, namely initial condition in October 2012 and
model propagation forward in time to October 2017. Results are shown in Fig.~\ref{f6}.
Obviously, here the red histograms (from initial data in October 2012) are more
constrained than the blue histograms (resulting from rotation state vectors propagated
using our model to October 2017). The latter are more dispersed than the red histograms
in Fig.~\ref{f5}, because the less numerous and accurate observations in 2012 constrain
the models with lower accuracy. Nevertheless, the principal features of the solution
are still present: (i) the $P_\phi$ period changed adequately for majority of cases,
while (ii) the $P_\psi$ period changed too little.
\subsubsection{In summary}
So while we are not able to provide an exact proof that the
change in $P_\phi$ and $P_\psi$ periods between the 2012 and 2017 apparitions of 2012~TC4
are due to the YORP effect, we consider the difference between the observations and model
predictions can be accounted for the model inaccuracy.
\section{Discussion}
In the previous section, we demonstrated that the observed change of the rotation
state of 2012~TC4 in between the two apparitions in 2012 and 2017 may be possibly
explained as a result of the YORP effect. We also showed that the effect of
the close encounter in October~12, 2012 on the rotation state was minimum, at
least in a rigid-body approximation. However, since the YORP model --for the reasons explained-- did
not provide an exact match of the observations, and even left unresolved the
issue of the observed energy change versus the model prediction, it is
both useful and necessary to also briefly analyse possible alternative
explanations. Here we discuss in some detail two plausible processes. We leave
aside a third possibility, notably a mass-loss from the surface
of 2014~TC4 sometime in the period between the two observation campaigns (or
during the 2012 Earth encounter). At the first sight, this may look an attractive
explanation because the fast rotation of this body implies formally negative
gravitational attraction at the surface. Therefore, it takes only the effect
of breaking cohesive bonds near the surface to make part of the body escape.
This event would have an influence on rotational energy and angular momentum,
both directly by the quanta carried away by escaping mass but also by a change
of the TC4's tensor of inertia. While possible, caveats of this model are twofold: (i) first, the observational data do not have resolution to provide a conclusive information about a shape change of the body (which may not be large for the effect to work at the one-per-mile level), and, (ii) more importantly, in most scenarios the process would lead to a decrease of the rotational angular momentum of TC4. Unlike in the two processes discussed below, we are not able to simply estimate likelihood of this process.
\subsection{Internal energy dissipation effects}
Individual solutions of
2012~TC4's rotation parameters in 2012 and 2017 indicate that periods $P_\psi$
and $P_\phi$ decreased in the latter epoch (Fig.~\ref{fig:histogram_periods}). In terms of osculating
approximation with a free-top model it implies that the wobbling motion of the
angular momentum vector ${\bf L}$ in the body-fixed frame moved toward the fundamental
mode of its direction along the shortest axis of the inertia tensor. Note the trajectory
of ${\bf L}$ in the body-fixed frame is uniquely parametrized with $p=2BE/L^2$
\citep[see, e.g., the Appendix~\ref{methods} and][]{Lan.Lif:69}. In quantitative terms, the
change from 2012 to 2017 state is expressed by $\delta p\simeq -6.2\times 10^{-3}$ (Table~\ref{tab:parameters}).
The average rate over $\delta t\simeq 5$~year interval would thus be $\delta p /
\delta t \simeq -3.9\times 10^{-11}$~s$^{-1}$.
In the YORP model presented above, the change in $p$ was a composition of changes
both in the energy $E$ and angular momentum $L$. In fact, both $E$ and $L$ increased
from 2012 to 2017 (Fig.~\ref{f3}), but the composite effect was a decrease in $p$, in this case $\delta p\simeq -4.6\times 10^{-3}$, a similar value to that directly determined from osculating $L$ and $E$ above. Another
processes may lead to approximately the same results by producing a different
combination of energy and angular momentum changes. For instance, effects of
material inelasticity result directly in energy dissipation while preserving
angular momentum. In this case, both $E$ and $p$ decrease with a direct relation
$\delta E \simeq (L^2/2B)\,\delta p$.
In order to explore whether the observed effect of the 2012~TC4's spin change
could even be plausibly matched by internal energy dissipation we used the model presented
by \cite{Bre.ea:12}. These authors assumed a fully triaxial geometry of
the body, but restricted their analysis of energy dissipation to the empirical
description with a quality factor $Q$ (see an alternative model of \cite{Fro.Efr:17},
where the authors describe the energy dissipation using a Maxwell
viscous liquid but allow only a biaxial shape of the body). With these assumptions,
they expressed the secular (i.e., wobbling-cycle-averaged) rate of energy change
in the following form
\begin{equation}
\frac{\delta E}{\delta t}\simeq - \frac{a^4 \rho m \Omega^5}{\mu Q}\,\Psi\; ,
\label{diss}
\end{equation}
where $a$ is semimajor axis of the body's triaxial approximation, $\rho$ its
density, $m$ its mass, $\Omega = L/C$ and $\Psi$ a rather complicated factor
depending on nutation angle $\theta_{\rm nut}$ (i.e., tilt between ${\bf L}$ and
the shortest body axis; we find $\theta_{\rm nut}$ oscillates between $\simeq 16^\circ$
and $\simeq 46^\circ$, with a mean $\simeq 30^\circ$), body-axes ratios and Lame coefficients
\citep[see][Sec. 4.3]{Bre.ea:12}. Finally, $\mu$ is the Lam\'e shear modulus
(rigidity) and $Q$ the quality factor, expressing empirically the energy
dissipation per wobbling cycle. The product $\mu Q$ is characteristic to studies
involving energy dissipation in planetary science since the pioneering work of
\cite{Bur.Saf:73} \citep[though, see already][]{Pre:58}. While highly
uncertain, typical values of this parameter for asteroids range between $10^{11}$
and $5\times 10^{12}$~Pa \citep[e.g.,][]{Har:94}. Using $\delta E/\delta t
\simeq (L^2/2B)\,\delta p / \delta t$, with the above mentioned $\delta p / \delta t$
value, we can now use (\ref{diss}) to infer what values of $\mu Q$ would be needed
to explain the change in 2012~TC4's tumbling state in between 2012 and 2017
close approaches. The remaining unknown parameter is $\Psi$, which we
estimate to be $\simeq (1-5)\times 10^{-3}$. Plugging this value to (\ref{diss}),
we find $\mu Q$ ranging from $4\times 10^5$ to $4\times 10^6$~Pa. Remarkably, such
values are four to five orders of magnitude smaller than the usually adopted
estimates. Therefore, unless the energy dissipation by internal friction is
extraordinarily high (and thus the $\mu Q$ value very small), this process
cannot explain the observed rotation change of 2012~TC4. Note additionally, that we considered derivation of the needed $\mu Q$ value within the energy dissipation model as a useful exercise to match the energy change. Such a model, however, would not explain the observed angular momentum change.
\subsection{Impact by an interplanetary particle}
Another alternative process
to the radiation torques is that of an impact by interplanetary meteoroid. However,
in the following we provide an argument that the likelihood of this to happen at
the level needed to explain the 2012~TC4 data is again very small. To that end we
used information from \cite{Bot.ea:20}. This paper determined meteoroid flux
on a small asteroid (101955) Bennu using a state-of-art model MEM-3 allowing to
predict parameters of meteoroids impacting a target body orbiting between Mercury
and the asteroid belt \citep[e.g.,][]{Moo.ea:20}. Note that the orbit of Bennu is similar
to 2012~TC4 and we shall neglect the small flux differences that could result
from a small orbital dissimilarity of these two objects (if anything, the flux would
be slightly larger on Bennu, because of its orbit closer to the Sun). In their Fig.~2,
\cite{Bot.ea:20} show that 5~mg interplanetary particles, approximately
$2$~mm in size, strike Bennu with frequency of $\simeq 60$ per year with a
median impact velocity little less than $\simeq 30$~km/s.
For 2012~TC4 we only need to rescale this number to account for a much smaller
size: Bennu is a $\simeq 500$~m size asteroid, while a characteristic size
of 2012~TC4 is only $\simeq 10$~m. Therefore, the 5~mg flux on 2012~TC4 is
about $\simeq 2.4\times 10^{-2}$ per year. The chances to be hit by such a particle
in $5$ years is therefore merely $\simeq 0.12$. However, even if it
has happened, the dynamical effect would be minimum. Estimating the change in
rotational angular momentum $L$ plainly by $\delta L/L \simeq (m/M)(v_{\rm imp}/
R\omega)$, where $m$ and $M$ are masses of the particle and 2012~TC4, $v_{\rm imp}$ is
the impact velocity, and $R$ and $\omega$ are 2012~TC4's characteristic radius
and rotational frequency, we would obtain $\delta L/L \simeq 2\times 10^{-7}$.
At least ten thousand times larger effect would be needed to approach the level
observed for 2012~TC4, and this would require an impacting particle at least
twenty times larger, i.e. $5$~cm or more. Because MEM-3 incorporates flux
dependence on mass from \cite{Gru.ea:85}, thus $\propto m^{-4/3}$, the
chances that 2012~TC4 was hit by a $5$~cm meteoroid in between 2012 and 2017 is
only $\simeq 6\times 10^{-7}$. We thus conclude that chances that the observed
effect was produced by an impact of interplanetary particle is negligibly
small and may be discarded. | 6a6b8a7c376684ea03358701478dd4ccc5bbb361 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Object detection, as a hot topic of computer vision, is to find out the objects of interest on the image and to determine their locations and categories.
In the era of deep learning~\cite{NIPS2012_4824}, many deep convolutional neural network (DNN)-based methods~\cite{girshick2015fast,ren2015faster,Redmon_2016_CVPR,Redmon_2017_CVPR,redmon2018yolov3,liu2016ssd} have been proposed to make the detection models with excellent performance in natural image benchmark datasets~\cite{everingham2010pascal, lin2014microsoft}.
While in reality, there is an obvious distributional gap between the training set (source domain) and the testing set (target domain), which may lead to a significant reduction in the generalization ability of models in target domain.
Therefore, many unsupervised domain adaptive object detection~(DAOD) methods~\cite{DAfaster, saito2019strong, Xinge2019Adapting, kim2019diversify, cai2019exploring, Chen2020Harmonizing} are proposed to solve the problem of domain distributional shifts.
Most unsupervised DAOD methods are nested adversarial training module within advanced object detection frameworks, such as Faster R-CNN~\cite{ren2015faster}. By minimizing the domain shifts of multi-level feature maps, domain adaption is achieved to improve the detection accuracy in the target domain.
Numerous DAOD methods are proposed and most of them focus on image-level~\cite{DAfaster}, instance-level~\cite{DAfaster,Chen2020Harmonizing}, pixel-level~\cite{kim2019diversify, Chen2020Harmonizing}, region-level~\cite{Xinge2019Adapting} and strong-local \& weak-global-level~\cite{saito2019strong} alignments.
Whereas, the existing multi-level alignment modules may cause three issues:
Firstly, they only focus on the mapping of the source and target domain, and less consider other distractive information that is unnecessary to be aligned, \eg, background information.
Secondly, we should concern over the regions where the object needs to be aligned on the image, and ignore background noise that is useless for detection.
However, the current region searching algorithm \cite{Xinge2019Adapting} in region-level alignment
restrict the category numbers of candidate region. Since there is no label for the target domain, we cannot exactly determine the object region numbers.
Lastly, for the existing instance alignment modules,
they are all based on ROI-Pooling~\cite{ren2015faster} feature alignment. While the training of region proposal network (RPN)~\cite{ren2015faster} mainly depends on the labels of the source domain, and there may be abundant redundancy or noise in region of interests (ROIs) on the target domain. Resultly, when the instance is aligned, there will be an extra alignment of background features or multiple alignments of instance-level features for some certain objects.
In order to solve the above mentioned issues, we propose a DAOD method based on the architecture of Faster R-CNN~\cite{ren2015faster}, called feature separation and alignment network (FSANet). The network consists of a gray-scale feature separation (GSFS) module, a local-global feature alignment module (LGFA) and a region-instance-level alignment (RILA) module. The feature separation module separates the distractive information containing the features that do not need to be aligned and the shared information related to the object detection through the difference loss, so that the feature alignment module can concentrate on the object information. LGFA and RILA are domain adaptation components, which reduce the distributional gaps of the multi-level features. Our contributions can be divided into two-fold:
(1) The dual-stream network~\cite{bousmalis2016domain, zhaoijcai2020} can extract multi-model information of different domain images.
Inspired by this, we design a dual-stream auto-encoder network for the GSFS. The distractive information of the image is separated by a private encoder, and the object detection backbone network acts as a shared information extractor to obtain useful features for detection.
When aligning high-dimensional features, we use shared features instead of the whole image features to reduce the impact of distractive information that is not related to detection, so as to improve domain adaptability.
Particularly, in order to reduce the impact of image color differences and the difficulty of reconstruction, we use the gray-scale image to extract the distractive features. This operation can make the private encoder focus on detection-related/irrelevant features of the objects rather than the image color, and while simplifying the reconstruction task.
(2) We design the RAIL module, which can effectively solve the negative effects caused by the redundancy of the region proposals and background noise, and determine the category number for candidate region adaptively. Initially, by exploiting Scale-Space Filtering algorithm (SSF)~\cite{895974}, we cluster the center coordinate of the bounding boxes predicted by RPN, to implement an adaptive region selection. Thus the traditional clustering region number uncertainty problem is well solved.
Then, after obtaining the cluster centers and scale by SSF, the outliers of ROIs can be found. We believe that these abnormal ROIs have a high probability of containing background information. So we exclude them before region grouping, thereby reducing the effect of ROIs noise. Finally, through global pooling layer, the instance-level features of each group (\ie, candidate region) are refined and instance alignment is performed to figure out the problem of ROIs redundancy.
In summary, the adaptability and transferability of adversarial-based adaptive detection methods are enhanced by separating the shared and distractive information of the source/target domain and aligning the region-instance-level features.
Extensive experiments illustrate that the proposed FSANet has reached an outstanding level of cross-domain detection performance on multiple benchmark datasets.
For example, the FSANet achieves $42.7\%$ mAP for transfer task on PASCAL $\to$ Clipart1k, which surpassing the state-of-the-art (SOTA) adversarial-based adaptive methods~\cite{DAfaster,Xinge2019Adapting,saito2019strong,kim2019diversify,Chen2020Harmonizing,sindagi2019prior,he2020domain}.
\section{Related Work}
\noindent
{\bf Object Detection.}
Nowadays, with the rapid development of deep neural convolutional networks (DNN), lots of methods have been proposed to solve the object detection task. They can be divided into two-stage object detection methods based on region proposals and end-to-end one-stage methods.
The pioneering work of the two-stage methods is R-CNN~\cite{girshick2014rich}, which first utilizes a selective search method~\cite{uijlings2013selective} to extract region proposals from images, and then trains a network to classify each ROIs. SPP-Net~\cite{he2015spatial} and Fast R-CNN~\cite{girshick2015fast} exploited DNNs to propose region proposals and significantly improved detection accuracy and speed. As DNNs were extended to share convolutional feature maps among all ROIs~\cite{ren2015faster} , the end-to-end two-stage methods (\eg, Faster R-CNN~\cite{ren2015faster} and RetinaNet~\cite{lin2017focal}) were proposed.
On the other hand, typical methods of one-stage object detection algorithms are YOLO~\cite{Redmon_2016_CVPR, Redmon_2017_CVPR, redmon2018yolov3} and SSD~\cite{liu2016ssd}, which extract the feature maps from the original image based on the convolutional network, and then directly perform object classification and bounding box regression. \medskip\\
{\bf Domain Adaptive Object Detection.} The existing unsupervised DAOD methods can be divided into three categories~\cite{li2020deep}: adversarial-based, reconstruction-based and hybrid adaptive methods.
For the adversarial-based adaptive methods,
Chen \etal~\cite{DAfaster} first applied the unsupervised domain adaptive method to object detection. By combining with the gradient reversal layer and domain classifier, an adversarial-based method~\cite{ganin2015unsupervised} is proposed to solve the domain shift in the real scene. At the same time, it is pointed out that the core issue of unsupervised DAOD is to solve the domain gaps in image and instance level.
For example, Zhu \etal~\cite{Xinge2019Adapting} indicated that the object detection needs to concern with the object rather than the entire image features, and then a regional-level domain adaptive network based on GAN~\cite{goodfellow2014generative} is established to solve the region proposal redundancy problem.
Saito \etal~\cite{saito2019strong} proposed a local domain classifier network based on FCN~\cite{Long2017Fully} to achieve the alignment of low-level features.
For the reconstruction-based adaptive methods,
Arruda \etal~\cite{Arruda2019Cross} firstly generated pseudo samples similar to the target domain from the source samples by CycleGAN~\cite{Zhu2017Unpaired}, and added them to the training set to improve detection accuracy.
The hybrid adaptive methods aim at combining the above two categories to improve the model performance.
In order to realize the adaptative object detection from natural images to artistic images, Inoue \etal~\cite{Inoue2018Cross} initially trained a detection model from the source domain samples, then fine-tuned the model on the pseudo samples of source domain, and finally implemented the weak training on the target domain samples.
Based on pseudo samples, Kim \etal~\cite{kim2019diversify} proposed an adaptive method for domain diversification and multi-domain invariant representation.
Chen \etal~\cite{Chen2020Harmonizing} added the attention mechanism to the network based on~\cite{saito2019strong}.
Although the above methods have achieved good performance, they do not concern for separating the distractive information and solving the region proposals redundancy and background noise. So in our paper, we propose a novel DAOD method called FSANet to further figure out the above mentioned issues.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.77\linewidth]{net.png}
\end{center}
\caption{The overall structure of the proposed FSANet.}
\label{fig:net}
\end{figure*}
\section{Method}
In this section, we introduce the details of FSANet. The framework is illustrate in Figure~\ref{fig:net}, which consists of four parts, namely the \textit{object detection} module, the \textit{LGFA} module, the \textit{RILA} module and the \textit{GSFS} module.
Initially, the paired source/target domain RGB images are input to the detection network, and the detection loss is calculated with the label of the source domain images, and the instance-level features alignment is realized by the RILA module. For the multi-level features $f_{\tau}(\tau=1,2,3)$ extracted by the backbone network $F_{\tau}$, they are input into the LGFA module to complete the multi-level features alignment. In addition, the gray-scale paired images are input into the GSFS module, and the useful/useless features for detection are separated through a specific loss.
Here, we use Faster R-CNN~\cite{ren2015faster} as the object detection backbone architecture, and the LGFA module follows the design of~\cite{saito2019strong,Chen2020Harmonizing}.
Especially, we focus on the GSFS module and the RILA module.
For the notations in this paper, the source domain dataset with labeled samples are denoted by $D_s=\{(x_i^s, b_i^s, y_i^s)\}_{i}^{n_s}$, where $x_i^s$ is the input source image, $b_i^s$ is the coordinates of the bounding box, $y_i^s$ is the object category label, and $n_s$ is the number of samples in the source domain dataset. Meanwhile, unlabeled target domain dataset is represented as $D_t=\{(x_i^t)\}_i^{n_t}$, where $n_t$ is the number of samples in the target dataset. Our task is to transfer the model learned from $D_s$ to $D_t$.
\subsection{Gray-Scale Feature Separation Module}
As is known to all, domain adversarial loss can reduce the inconsistency between the features distribution of the two domains. However, there will still be some distractive information that cannot be reduced, \eg, the background distractive information of the source/target domain.
So we propose a GSFS module based on a dual-stream network, and employ it to separate the distractive/shared information which is useless/useful for object detection with the high-level features through the private encoder $E_s$ and $E_t$, the shared decoder $S_D$ and the features extraction module $F_\tau$, as shown in Figure~\ref{fig:net}. Firstly, $F_3$ extracts high-level features related to object detection via the detection loss. Then domain adversarial loss of LGFA module makes the features $f_\tau$ of the source/target domain aligned. As a result, $f_3$ is the shared high-level features useful for detection.
For the GSFS module, we define the different loss to limit the similarity of $\left\lbrace d^s,f_3^s\right\rbrace $ and $\left\lbrace d^t,f_3^t\right\rbrace$:
\begin{equation}
\mathcal{L}\!_\mathit{diff}\!=\!\frac{1}{n_s}\!\sum_{i=1}^{n_s} \lVert g(d^s_i)^\top\!g(f_{3,i}^s) \rVert _F^2 + \frac{1}{n_t}\!\sum_{i=1}^{n_t}\!\lVert g(d^t_i)^ \top\!g(f_{3,i}^t) \rVert _F^2,
\end{equation}
where $\lVert \cdot \rVert _F^2$ is the squared Frobenius norm , $g(\cdot)$ is the global pooling layer, $d_i^s$, $d_i^t$, $f_{3,i}^s$, $f_{3,i}^t$ denote distractive information output by private encoder $E_s$/$E_t$ and the shared information $f_3$ of source/target domain for $i$th image, respectively.
In addition, considering that simple restrictions will make $d^s$ and $d^t$ meaningless, and the information related to objects cannot include all the information of the input image, we use the fusion features $f_\mathit{fus}^s\!=\![d^s, f_{3}^s]$, $f_\mathit{fus}^t\!=\![d^t, f_{3}^t]$ as the input of the shared decoder $S_D$ to reconstruct the input image, where $[\cdot,\cdot]$ denotes the concatenation of two feature maps in the channel dimension. Through the restriction of image reconstruction, the separated features can possess almost all the information of source/target images. So that the private encoder extracts the distractive information irrelevant to detection in the image.
Notably, the private encoder inputs two-domain images in gray-scale, and extract the distractive information which is irrelevant to the color of the two-domain images.
If the input is an RGB image, it will not only increase the difficulty of reconstruction, but also make the private encoder disturb by the color difference of the image, while ignoring other distractive information on the two data domains.
Finally, the reconstruction loss of the gray-scale images is defined as:
\begin{equation}
\mathcal{L}_\mathit{rec} = \frac{1}{n_s}\sum_{i=1}^{n_s} \lVert x_{g,i}^s - \hat{x}_{g,i}^s \rVert _1^1 +
\frac{1}{n_t}\sum_{i=1}^{n_t} \lVert x_{g,i}^t - \hat{x}_{g,i}^t \rVert _1^1,
\end{equation}
where $\lVert \cdot \rVert _1^1$ is the $\ell_1$-norm, ${x}_{g,i}^s$, ${x}_{g,i}^t$, $\hat{x}_{g,i}^s$ and $\hat{x}_{g,i}^t$ are the $i$th gray-scale input images and reconstructed images by the shared decoder $S_D$ in source/target domain.
\subsection{Region-Instance-Level Alignment}
The RILA module consists of a \textit{grouping} component and a \textit{context-aware RILA} component. Please refer to supplemental material to see illustration and more details.\\
\noindent
{\bf Grouping.} Some alignment module~\cite{DAfaster, he2019multi, Chen2020Harmonizing} implement the instance-level feature alignment of all region proposals. However, the absence of labels on the target domain leads to the training of RPN mainly guided by the labeled data on the source domain~\cite{ren2015faster}. Therefore, we cannot ensure that the region proposals on the target domain contain the objects, where may contain a lot of background noise.
Meanwhile, objects are generally detected by multiple region proposals, which means that some region proposals are redundant.
We should perform feature alignments for the instance-level features in the region proposals where the object is most likely to exist, instead of aligning all features.
As we all know, the locations of region proposals where contain the same object at multiple scales should be similar if they are predicted by a reliable RPN.
A natural idea is to cluster the similar region proposal coordinates to reduce redundancy for ROIs and eliminate the background noises by excluding outliers.
In particular, we use RPN to get the predicted bounding box $\{b_{i,x},b_{i,y},w_i,h_i\}_{i=1}^N$ of the region proposals, where $(b_{i,x},b_{i,y})$ is the center coordinate of the $i$th predicted bounding box, $w_i$ is its width, and $h_i$ is its height. Then we plan to exploit the scale-space filtering~(SSF)~\cite{895974} algorithm to cluster the center coordinates to adaptively obtain $K$ cluster centers, which means that ROIs can be divided into $K$ categories, and thereby, the instance-level features can be indirectly divided into $K$ categories. As a result, we only need to align the instance-level features in each category. \\
\noindent
{\bf Adaptive candidate region searching.}
SSF is a clustering method that adaptive to the category numbers without a manual setting. It is very suitable for the grouping module due to the unknown object numbers in the target domain. Therefore, inspired by~\cite{895974}, we decided to use SSF to improve the traditional K-means algorithm to accomplish the clustering of bounding boxes.
By regarding each sample as a ``light point'', the SSF algorithm formulates the clustering issue as finding the ``lighting blob center'' under different blur scales, \ie, different scale-space maps. The specific workflow can be divided into the following steps:\\
\indent
(1) \textit{Clustering center iteration.} For a given dataset $\mathcal{X}=\left\{x_{i} \in \mathbb{R}^{2}: i=1, \ldots, N\right\}$ (in this paper, this is a set of center coordinates of predicted bounding boxes by RPN), we define $P(x, \sigma_j)$ as the scale-space map for $\mathcal{X}$ under the blur scale $\sigma_j$. It can be calculated by $P(x, \sigma_j)=p(x) * \phi(x, \sigma_j)$, where $p(x)$ means the scatter image of data samples in $\mathcal{X}$, $\phi(\cdot, \sigma_j)$ is the Gaussian blur kernel with the blur scale $\sigma_j$, and $*$ represents the convolution operator. So the set of clustering centers $\mathcal{C}(\sigma_j)$ under different scales $\sigma_{j}$ is obtained by $\nabla_{x} P(x, \sigma_j)=0$.
In iteration steps, $\sigma_j$ is firstly updated, then after $\mathcal{C}(\sigma_j)$ has converged, a larger $\sigma_{j+1}$ is obtained by $\sigma_{j+1}\!=\!k\sigma_{j}$ ($k$ is a hyperparameter and $k\!>\!1$), and $\mathcal{C}(\sigma_{j+1})$ are iterated again until all samples belong to one category. More details can be seen in the supplemental material.
(2) \textit{Adaptive category number selection.} With the completion of the above iterations, we have obtained a serious of clustering center, \ie, $\{\mathcal{C}(\sigma_1), \cdots,\mathcal{C}(\sigma_j),\cdots,\mathcal{C}(\sigma_J)\}$. We hope to select the clustering model that best fit for the data distribution. So the ``lifetime''~\cite{895974} is used as a measure of the stability for category number, which is calculated by
\begin{equation}\label{eqs:pi}
\pi(\sigma_j)=c \log (\sigma_j / \varepsilon),
\end{equation}
where $\varepsilon=0.01, c=1 / \log (1.05)$.
Notably, the interval $ \mathbf{\Sigma}=\left\lbrace \sigma_{J_{inf}}, \sigma_{J_{inf}+1},\dots,\sigma_{J_{sup}} \right\rbrace $ with the longest ``lifetime'' represent the relative appropriate scales.
Finally, the most appropriate clustering model $\{\mathcal{C}(\sigma^*),\sigma^*\}$ equals to median of $\mathbf{\Sigma}$, and the adaptive category number $K$ equals to the element number of $\mathcal{C}(\sigma^*)$.
(3) \textit{Remove outliers.}
After obtaining the optimal $\{\mathcal{C}(\sigma^*),\sigma^*\}$, we can use the model to group the centers $\{(b_{i,x},b_{i,y})\}_{i=1}^N$ of the predicted bounding box of all region proposals.
For each bounding box center $b_i=(b_{i,x},b_{i,y})$, if $\lVert b_i-c_k^* \rVert _2>\sigma^*$, we consider $b_i$ to be an outlier, and its corresponding region proposal probably contains the background noise, which should be excluded from grouping.
Through this method, we can restrict certain outlier region proposals from participating in instance-level alignment, and find the instance features that need to be aligned in the image more robustly.\\
\noindent
{\bf Context-Aware Region-Instance-Level Alignment.}
After clustering the proposals, we need to refine the instance-level features of each category for instance alignment.
For simplicity, we reshape the features of instances belonging to the same category into the same size and then concatenate them into a feature map as the input of the domain classifier.
But the effect of redundant or noisy region proposals is not fundamentally solved.
Therefore, we use the global pooling layer to refine the instance-level features in the same category.
Let $\Theta_k\in \mathbb{R}^{m_k\times d}$ denote the instance-level features of the $k$th category obtained by SSF, where $m_k$ is the number of the $k$th category , $d$ is the dimension of the instance-level features.
Through the global pooling layer, we can get the feature of the $k$th category $\hat{\Theta}_k\in \mathbb{R}^{1\times d}$, which can fully reflect the instance-level features of the $k$ category and thereby reduce the influence of redundancy and noise.
Then, we use the context fusion instance-level features~\cite{saito2019strong} as the input of bounding box regression and classification prediction in the instance alignment and object detection module. We acquire three different levels of context vectors $f_l, f_m, f_g$ from the domain classifier $D_l, D_m, D_g$, and get the instance-level features $f_r$ of each region proposals based on ROI-Pooling~\cite{he2017mask}. Through feature concatenation, the fused context instance-level features $f_{ins}=[f_l,f_m,f_g,f_r]$ is input into the global pooling layer and obtain refined feature of each category $f_{ins_e}$, as shown in Figure~\ref{fig:net}.
Subsequently, we input the global pooling instance-level features into the domain classifier $G_{ri}$, and output the domain labels to achieve regional instance alignment. For the loss function of this module, we use Focal Loss~\cite{lin2017focal,saito2019strong} as follows:
\begin{equation*}
\mathcal{L}_\mathit{ri_s}\!=\!-\frac{1}{n_s} \sum_{i=1}^{n_s} \frac{1}{K_i^s}\sum_{k=1}^{K_i^s}(1\!-\!\mathit{D}_\mathit{ri}(\mathit{f}_\mathit{ins_e,i,k}^{s}))^\gamma\log(\mathit{D}_\mathit{ri}(\mathit{f}_\mathit{ins_e,i,k}^{s})),
\end{equation*}
\begin{equation*}
\mathcal{L}_\mathit{ri_t}\!=\!-\frac{1}{n_t} \sum_{i=1}^{n_t} \frac{1}{K_i^t}\sum_{k=1}^{K_i^t}(\mathit{D}_\mathit{ri}(\mathit{f}_\mathit{ins_e,i,k}^{t}))^\gamma\log(1\!-\!\mathit{D}_\mathit{ri}(\mathit{f}_\mathit{ins_e,i,k}^{t})),
\end{equation*}
\begin{equation}
\mathcal{L}_{ri}=\frac{1}{2}(\mathcal{L}_{ri_s}+\mathcal{L}_{ri_t}),
\label{kmeansloss}
\end{equation}
where $K_i^s$ is the number of adaptive categories for $i$th source sample, $f_\mathit{ins_e,k}^{s}$ is the instance-level features of the $k$th region of the source domain after global pooling, and $\gamma$ is the parameter of Focal Loss.
In short, this module can solve the redundancy of the region proposals and the influence of the background noise, so that the instance alignment is focused on the features where the object is existing.
\subsection{Traing Loss}
The loss of object detection includes the classification loss $\mathcal{L}_{c}$ on the source domain and the regression loss $\mathcal{L}_{r}$ of the bounding boxes~\cite{ren2015faster}.
For the LGFA module, we use the Local Feature Masks and IWAT-I frameworks mentioned in~\cite{Chen2020Harmonizing}, and the corresponding adversarial loss is $\mathcal{L}_{lg}\!=\!\mathcal{L}_{adv_1}\!+\!\mathcal{L}_{adv_2}\!+\!\mathcal{L}_{adv_3}$ in Figure~\ref{fig:net}.
By combining all modules mentioned above, the overall objective function of the model is
\begin{equation}\label{loss}
\max_{D} \min_{F, E, S} \mathcal{L}_{c} + \mathcal{L}_{r} +
\beta(\mathcal{L}_{rec} + \mathcal{L}_{diff}) -
\lambda(\mathcal{L}_{lg} + \mathcal{L}_{ri}),
\end{equation}
where $D,F,E,S$ denote the domain classifier, object detection framework, private encoder and shared decoder, respectively. $\lambda$ and $\beta$ are turning parameters. The sign of gradients for the last item is flipped by a gradient reversal layer proposed by~\cite{ganin2015unsupervised}.
\section{Experiments}
In this section, we show the effectiveness of FSANet in adaptive object detection experiments on three benchmark domain shifts datasets including real $\to$ artistic, normal $\to$ foggy and synthetic $\to$ real, \ie, PASCAL~\cite{everingham2010pascal} to Clipart1k~\cite{Inoue2018Cross}, Cityscapes~\cite{cordts2016cityscapes} to Foggy-Cityscapes~\cite{sakaridis2018semantic} and Sim10k~\cite{johnson2016driving} to Cityscapes.
\begin{table*}[t]
\setlength{\tabcolsep}{1.5pt}
\begin{center}
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccccccccccccccccccc}
\toprule
Methods & aero & bike & bird & boat & bot & bus & car & cat & chair & cow & table & dog & horse & mbike & persn & plant & sheep & sofa & train & tv & mAP \\
\midrule
Source Only & 35.6 & 52.5 & 24.3 & 23.0 & 20.0 & 43.9 & 32.8 & 10.7 & 30.6 & 11.7 & 13.8 & 6.0 & 36.8 & 45.9 & 48.7 & 41.9 & 16.5 & 7.3 & 22.9 & 32.0 & 27.8 \\
SWDA \cite{saito2019strong} & 26.2 & 48.5 & 32.6 & 33.7 & 38.5 & 54.3 & 37.1 & 18.6 & 34.8 & 58.3 & 17.0 & 12.5 & 33.8 & 65.5 & 61.6 & \textbf{52.0} & 9.3 & 24.9 & 54.1 & 49.1 & 38.1 \\
HTCN \cite{Chen2020Harmonizing} & 33.6 & 58.9 & 34.0 & 23.4 & 45.6 & 57.0 & 39.8 & 12.0 & 39.7 & 51.3 & 21.1 & 20.1 & 39.1 & 72.8 & 63.0 & 43.1 & 19.3 & 30.1 & 50.2 & 51.8 & 40.3 \\
DDMRL \cite{kim2019diversify} & 25.8 & 63.2 & 24.5 & \textbf{42.4} & \textbf{47.9} & 43.1 & 37.5 & 9.1 & \textbf{47.0} & 46.7 & 26.8 & 24.9 & \textbf{48.1} & 78.7 & 63.0 & 45.0 & 21.3 & \textbf{36.1} & 52.3 & \textbf{53.4} & 41.8 \\
ATF \cite{he2020domain} & \textbf{41.9} & 67.0 & 27.4 & 36.4 & 41.0 & 48.5 & 42.0 & 13.1 & 39.2 & \textbf{75.1} & \textbf{33.4} & 7.9 & 41.2 & 56.2 & 61.4 & 50.6 & \textbf{42.0} & 25.0 & 53.1 & 39.1 & 42.1 \\
\midrule
FSANet & 31.0 & 63.7 & \textbf{34.8} & 29.4 & 43.0 & \textbf{70.7} & 40.8 & 18.7 & 39.6 & 57.4 & 22.2 & \textbf{27.0} & 33.3 & \textbf{85.6} & 63.3 & 45.7 & 21.9 & 24.7 & \textbf{56.7} & 44.5 & \textbf{42.7} \\
\bottomrule
\end{tabular}
}
\end{center}
\caption{Results on adpatation from PASCAL VOC to Clipart Dataset (real $\to$ artistic). Source only stands for Faster R-CNN that is trained only using source domain without adaptation. Average precision ($\%$) is evaluated on target images. The backbone network is ResNet-101.}
\label{Clipart}
\end{table*}
\subsection{Datasets}
\noindent
{\bf PASCAL $\to$ Clipart1k.} We conduct domain adaptive experiments from real images to artistic images to show that FSANet is effective in dissimilar domains, where PASCAL VOC dataset~\cite{everingham2010pascal} and Clipart1k~\cite{Inoue2018Cross} are set as the real source domain and the target artistic domain, respectively. The PASCAL dataset contains 20 categories of images and their bounding boxes. In this experiment, the training and validation splits of PASCAL VOC 2007 and 2012 are used for training, resulting in about 15K images. Clipart1k has a total of 1K images, which contains the same class as PASCAL. All images in it are employed for training (w/o labels) and testing. \\
{\bf Cityscape $\to$ Foggy-Cityscapes.} For experiments in similar domains, we use Cityscapes and Foggy-Cityscapes as the source and target domain dataset, respectively.
The Cityscapes~\cite{cordts2016cityscapes} dataset contains street scenes in different cities under normal weather conditions captured by a vehicle-mounted camera. There are 2975 images in the training set and 500 images in the test set. The label data are acquired by~\cite{DAfaster}. In this experiment, we employed the training set of Cityscapes for training. Foggy-Cityscapes~\cite{sakaridis2018semantic} is obtained by adding fog noise to the Cityscapes dataset, and its label is the same as Cityscapes.
We set the training set of Foggy-Cityscapes for training and the test set for calculating average accuracy. \\
{\bf Sim10k $\to$ Cityscapes.} In the experiments from synthetic images to real images, we conduct the experiment for Sim10k $\to$ Cityscapes. Sim10k~\cite{johnson2016driving} is acquired in a game scene of a computer game Grand Theft Auto V. It has 10k computer-synthesized driving scene images. According to the protocol of~\cite{DAfaster}, we only evaluated detection accuracy on cars. All images of Sim10k are set for training.
\subsection{Implementation Details}
In all experiments, the object detection model follows the settings of~\cite{saito2019strong,Xinge2019Adapting,Chen2020Harmonizing}, using Faster R-CNN~\cite{ren2015faster} with ROI-alignment~\cite{he2017mask}, where the shorter side of the images equal to 600 and the hyper-parameters in the object detection are set based on the setting of~\cite{ren2015faster}. For different dataset, the backbone network utilize VGG-16~\cite{simonyan2014very} or ResNet-101~\cite{he2016deep} with the parameter initialization weights following the pre-trained on ImageNet~\cite{deng2009imagenet}. A stochastic gradient descent training model with a momentum of 0.9 is used, and the learning rate for the first 50K iterations is set to 0.001 and then decays to 0.0001. In each iteration, we input a pair of source and target domain images. After 70K iterations, we calculate the mean average precision (mAP) where the IoU threshold is 0.5. We set $\gamma\!=\!5$ in Eq.~(\ref{kmeansloss}) and $\beta\!=\!0.1$ in Eq.~(\ref{loss}) in all experiments. $\lambda$ is set to 1 for PASCAL $\to$ Clipart1k and Cityscapes $\to$ Foggy-Cityscapes while set to 0.1 for Sim10k $\to$ Cityscapes. Our experiments are implemented by the Pytorch~\cite{paszke2017automatic}.
\subsection{Performance Comparison}
We compare our FASNet with various SOTA DAOD methods, including Domain adaptive Faster-RCNN ({DA-Faster}) \cite{DAfaster}, Selective Cross-Domain Alignment ({SCDA} )~\cite{Xinge2019Adapting}, Strong-Weak Distribution Alignment ({SWDA})~\cite{saito2019strong}, Domain Diversification and Multi-domain-invariant Representation Learning ({DDMRL})~\cite{kim2019diversify}, Hierarchical Transferability Calibration Network ({HTCN})~\cite{Chen2020Harmonizing}, Prior-based Domain Adaptive Object Detection ({PBDA}~\cite{sindagi2019prior}) and Asymmetric Tri-way Faster-RCNN ({ATF})~\cite{he2020domain}.
We cite the quantitative results in their original papers for comparison.
\noindent
{\bf Results on real $\to$ artistic. } The performance of FASNet in the target domain is significantly better than all comparison methods, as displayed in Table~\ref{Clipart}. Compared with SOTAs, mAP is improved by $+0.6\%$ (from $42.1\%$ to $42.7\%$), which fully illustrates that the proposed method can improve the transferability for real $\to$ artistic.
\noindent
{\bf Results on normal $\to$ foggy.} According to Table~\ref{cstocs_fg}, FASNet achieves the best performance and approaches the upper bound of the transfer task on Cityscape $\to$ Foggy-Cityscapes.
\noindent
{\bf Results on synthetic $\to$ real.} As shown in Table~\ref{sim2cs}, Our FSANet exceeds all comparison methods in the adaptive task on Sim10k $\to$ Cityscape, which further demonstrates that our method has better transferability and adaptivity.
\begin{table}[t]
\small
\setlength{\tabcolsep}{1pt}
\begin{center}
\begin{tabular}{lccccccccc}
\toprule
Method & persn & rider & car & truck & bus & train & mbike & bcycle & mAP \\
\midrule
Source Only & 24.1 & 33.1 & 34.3 & 4.1 & 22.3 & 3.0 & 15.3 & 26.5 & 20.3 \\
DA-Faster \cite{DAfaster} & 25.0 & 31.0 & 40.5 & 22.1 & 35.3 & 20.2 & 20.0 & 27.1 & 27.6 \\
SCDA \cite{Xinge2019Adapting} & 33.5 & 38.0 & 48.5 & 26.5 & 39.0 & 23.3 & 28.0 & 33.6 & 33.8 \\
SWDA \cite{saito2019strong} & 29.9 & 42.3 & 43.5 & 24.5 & 36.2 & 32.6 & 30.0 & 35.3 & 34.3 \\
DDMRL \cite{kim2019diversify} & 30.8 & 40.5 & 44.3 & 27.2 & 38.4 & 34.5 & 28.4 & 32.2 & 34.6 \\
ATF \cite{he2020domain} & 34.6 & 47.0 & 50.0 & 23.7 & 43.3 & 38.7 & 33.4 & \textbf{38.8} & 38.7 \\
PBDA \cite{sindagi2019prior} & \textbf{36.4} & 47.3 & \textbf{51.7} & 22.8 & 47.6 & 34.1 & \textbf{36.0} & 38.7 & 39.3 \\
HTCN \cite{Chen2020Harmonizing} & 33.2 & 47.5 & 47.9 & 31.6 & 47.4 & \textbf{40.9} & 32.3 & 37.1 & 39.8 \\
\midrule
FSANet & 33.5 & \textbf{47.6} & 45.8 & \textbf{32.0} & \textbf{50.1} & 37.7 & 35.7 & 37.7 & \textbf{40.0} \\
\midrule
Oracle & 33.2 & 45.9 & 49.7 & 35.6 & 50.0 & 37.4 & 34.7 & 36.2 & 40.3 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Results on adpatation from Cityscapes to Foggy-Cityscapes (normal $\to$ foggy). The backbone network is VGG-16.}
\label{cstocs_fg}
\end{table}
\begin{table}[t]
\begin{center}
\begin{tabular}{lclc}
\toprule
Method & AP on car & Method & AP on car \\
\midrule
Source Only & 34.6 & DA-Faster \cite{DAfaster} & 38.9 \\
SWDA \cite{saito2019strong} & 40.1 & HTCN \cite{Chen2020Harmonizing} & 42.5 \\
ATF \cite{he2020domain} & 42.8 & SCDA \cite{Xinge2019Adapting} & 43.0 \\
\midrule
FSANet & \textbf{43.2} & & \\
\bottomrule
\end{tabular}
\end{center}
\caption{Results on adpatation from Sim10k to Cityscapes (synthetic $\to$ real). The backbone network is VGG-16.}
\label{sim2cs}
\end{table}
\begin{figure*}[t]
\begin{center}
\subfigure{\includegraphics[width=0.3\linewidth]{source_zurich_000033_000019_leftImg8bit.jpg}}
\subfigure{\includegraphics[width=0.3\linewidth]{target_zurich_000033_000019_leftImg8bit_foggy_beta_s_o.jpg}}
\subfigure{\includegraphics[width=0.3\linewidth]{target_zurich_000033_000019_leftImg8bit_foggy_beta.jpg}}\vspace{-2.5mm}\\
\setcounter{subfigure}{0}
\subfigure[Original Image]{\includegraphics[width=0.3\linewidth]{source_zurich_000032_000019_leftImg8bit.jpg}}
\subfigure[Source Only]{\includegraphics[width=0.3\linewidth]{target_zurich_000032_000019_leftImg8bit_foggy_beta_s_o.jpg}}
\subfigure[FSANet]{\includegraphics[width=0.3\linewidth]{target_zurich_000032_000019_leftImg8bit_foggy_beta.jpg}}\vspace{-2.5mm}\\
\subfigure[Clipart1k]{\includegraphics[width=0.3\linewidth]{clipart.png}}
\subfigure[Cityscapes]{\includegraphics[width=0.6\linewidth]{sim.png}}
\end{center}
\caption{Examples of detection results on the target domain. From first to second rows (a): there is not foggy noise in the original images, (b-c): Cityscapes $\to$ Foggy-Cityscapes. (d): PASCAL $\to$ Clipart1k. (e): Sim10k $\to$ Cityscapes.}
\label{fig:demo}
\end{figure*}
\begin{table*}[t]
\setlength{\tabcolsep}{1pt}
\begin{center}
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccccccccccccccccccc}
\toprule
Method & aero & bike & bird & boat & bot & bus & car & cat & chair & cow & table & dog & horse & mbike & persn & plant & sheep & sofa & train & tv & mAP \\
\midrule
Source Only & 35.6 & 52.5 & 24.3 & 23.0 & 20.0 & 43.9 & 32.8 & 10.7 & 30.6 & 11.7 & 13.8 & 6.0 & 36.8 & 45.9 & 48.7 & 41.9 & 16.5 & 7.3 & 22.9 & 32.0 & 27.8 \\
Only LGFA & 25.1 & 45.5 & 26.1 & 26.3 & 37.8 & 48.7 & 39.5 & 13.8 & 34.3 & 43.1 & \textbf{22.4} & 11.2 & 36.5 & 62.4 & 57.1 & 50.4 & 16.0 & 26.4 & 48.9 & 42.3 & 35.7 \\
FSANet w/o CSFS & \textbf{33.6} & 60.9 & 33.5 & \textbf{30.9} & \textbf{43.7} & 56.3 & 39.0 & \textbf{20.7} & 35.5 & 58.2 & 13.7 & 22.4 & 35.6 & 81.2 & 60.8 & 48.0 & \textbf{28.2} & 20.2 & 51.8 & 41.9 & 40.8 \\
FSANet w/o RILA & 32.5 & 62.2 & 32.4 & 30.5 & 42.3 & 53.6 & \textbf{42.8} & 17.0 & 38.3 & \textbf{62.4} & 20.2 & 19.9 & \textbf{36.9} & 79.7 & 62.3 & 48.9 & 19.0 & 24.4 & 53.2 & 41.7 & 41.0 \\
FSANet w/o LGFA & 29.8 & 40.2 & 28.6 & 22.7 & 32.0 & 51.2 & 35.6 & 12.9 & 34.7 & 17.2 & 19.8 & 12.1 & 33.5 & 43.0 & 42.5 & 45.1 & 10.3 & \textbf{27.9} & 42.9 & 40.9 & 31.1 \\
FSANet w/o diff loss & 32.8 & 60.4 & 32.1 & 30.5 & 38.7 & 70.6 & 40.0 & 19.9 & 37.2 & 56.4 & 22.1 & 22.0 & 34.9 & 71.9 & \textbf{63.9} & 46.6 & 23.8 & 27.2 & 52.5 & 45.0 & 41.4 \\
FSANet (RGB S) & 26.8 & 53.4 & 33.0 & 30.6 & 37.1 & \textbf{72.3} & 40.2 & 19.6 & \textbf{40.0} & 59.3 & 21.6 & 13.9 & 32.8 & \textbf{85.7} & 61.0 & \textbf{50.6} & 19.7 & 26.1 & 53.3 & \textbf{45.6} & 41.1 \\
\midrule
FSANet & 31.0 &\textbf{63.7} & \textbf{34.8} & 29.4 & 43.0 & 70.7 & 40.8 & 18.7 & 39.6 & 57.4 & 22.2 & \textbf{27.0} & 33.3 & 85.6 & 63.3 & 45.7 & 21.9 & 24.7 & \textbf{56.7} & 44.5 & \textbf{42.7} \\
\bottomrule
\end{tabular}
}
\end{center}
\caption{Ablation experiments of FSANet on PASCAL $\to$ Clipart1k. More results on different datasets are shown in supplemental material.}
\label{Ablation}
\end{table*}
\begin{table}[t]
\begin{center}
\begin{tabular}{lclc}
\toprule
Method & mAP & Method & mAP \\
\midrule
Source Only & 27.8 & K-Means ($K=2$) & 35.6 \\
K-Means ($K=4$) & 41.6 & K-Means ($K=8$) & 39.8 \\
\midrule
SSF & \textbf{42.7} & & \\
\bottomrule
\end{tabular}
\end{center}
\caption{Results of domain adaptation for object detection from PASCAL VOC to Clipart Dataset with K-Means or SSF. The $K$ is the numbers of clustering of K-Means.}
\label{kmeans}
\end{table}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.6\linewidth]{Figure_4}
\end{center}
\caption{The mAP with the variation of IoU thresholds on transfer task PASCAL$\to$Clipart1k. $\mathrm{HTCN}^*$ stands for HTCN trained without interpolation (w/o pixel-level alignment via CycleGAN~\cite{Chen2020Harmonizing}).}
\label{fig:instance}
\end{figure}
\begin{figure*}[!]
\begin{center}
\includegraphics[width=0.8\linewidth]{kmeans}
\end{center}
\caption{Samples in PASCAL VOC $\to$ Clipart.
(a): The change curve of the ``Lifetime'' with the number of iterations. It can be seen that the ``Lifetime'' is the longest when $K=3$, so the number of adaptive categories is 3.
(b): Predicting results of cluster centers and scales by SSF. The yellow circled area indicates the cluster center that is used for clustering, and the points outside the circle will be deleted as outliers. Obviously, the category center basically coincides with the regions where the objects are located.
(c-e): The clustering results of K-Means method in different $K$. From left to right: $K$=2, 4 and 8, respectively. Compared with the clustering result of SSF, the clustering centers of K-Means are away from the object regions.}
\label{fig:kmeans}
\end{figure*}
\begin{figure}[!]
\begin{center}
\includegraphics[width=0.65\linewidth]{tsne.png}
\end{center}
\caption{Visualization of features obtained by transfer task Cityscapes $\to$ Foggy-Cityscapes. Blue/Red dots: source/target samples. (a-b): The distractive and shared features by FSANet, (c): The shared features by FSANet (w/o GSFS).}
\label{fig:tsne}
\end{figure}
\subsection{Further Empirical Analysis}
In this section, some ablation experiments are established to show the rationality for different components in FSANet. Furthermore, we contrast the effectiveness between RILA and traditional instance-level alignment module and explain the advantage of SSF in RILA compared with other different clustering methods.
\noindent
{\bf Example of Detection Results. }
Examples of various adaptive object detection tasks on the target domain are exhibited in Figure~\ref{fig:demo}.
We find that FSANet has good transferability and adaptivity on different datasets.
For instance, from normal to foggy, FSANet can detect the indiscoverable objects in the foggy noise.
As for the detection of artistic images, we can also get wonderful object and accurate bounding box predictions.
\noindent
{\bf Ablation Study.} We design various ablation experiments to prove the contributions of different modules in FSANet. They are shown as follows: {Only LGFA} denotes that we remove the RILA module and the Feature Separation module. {FSANet w/o GSFS} removes the GSFS module. {FSANet w/o RILA} eliminates the RILA module. {FSANet w/o LGFA} removes the LGFA module. {FSANet w/o diff loss} denotes that we do not use different loss in the GSFS module. {FSANet (RGB S)} denotes that we train the feature separation module by RGB images instead of gray-scale images.
\indent The results are reported in Table~\ref{Ablation} on PASCAL $\to$ Clipart1k. Compared with Only LGFA, adding RILA module or feature separation module can improve the performance of the model.
FSANet w/o GSFS increases from $35.7\%$ to $40.8\%$ and the FSANet w/o RILA increases by $+5.3\%$.
As shown in Figure~\ref{fig:tsne}, the GSFS module effectively promotes the features alignment. In addition, the separated distractive information is highly differentiated, demonstrating the rationality of feature separation.
While, mAP is only $31.1\%$ in the FSANet w/o LGFA, indicating that high-dimensional feature alignment is essential to avoid excessive separation of high-dimensional features.
Then, network performance will decline if different loss is not utilized, indicating that different loss is necessary to limit feature separation component.
Moreover, experiment FSANet (RGB S) shows that RGB images will reduce the performance of feature separation module.
In short, these experimental results fully demonstrate that the proposed module is more conducive to improving the transferability of the model.\\
\noindent
{\bf Region-Instance-level $\boldsymbol{v}$.$\boldsymbol{s}$. Instance-level. }
The mAP of different methods with the variation of IoU thresholds is displayed in Figure~\ref{fig:instance}.
Noticeably, the mAP reduces as the IoU threshold decreases and that of RILA is significantly improved compared with instance-level alignment on the IoU range of 0.50-0.95, which indicates that RILA can promote the adaptability and make bounding box regression more robustly.\\
{\bf Comparisons with K-Means.} We compare K-Means with our SSF algorithm to test whether the adaptively selecting region is efficient. In Table~\ref{kmeans}, it is obvious that mAP has huge fluctuations when $K$ is changed. In addition, in Figure~\ref{fig:kmeans}, the fixed category number reduces the flexibility for region selection, which shows the superiority for adaptive determination of region categories.
\section{Conclusion}
In this paper, we propose a feature separation and alignment network (FSANet) for domain adaptive object detection. The FSANet can effectively decompose distractive information which is useless for detection by a feature separation module, and can restrain background noise and redundancy information by a region-instance-level alignment module, which can adaptively extract the regions to be aligned. Compared with existing methods, our novel FSANet can separate the distractive features and make our model focus on the features useful for detection. At the same time, region proposal redundancy and the background noise in feature alignment can be avoided.
Extensive experiments demonstrate that our method surpasses other existing models for adaptive object detection and the modules we proposed greatly improve the transferability and adaptability on several benchmark datasets.
{\small
\bibliographystyle{ieee_fullname}
| 28bf64f29c6d5c88b8f6fff67870168c06b5dbc4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
XB 1916-053 is a low-mass X-ray binary system (LMXB) showing dips and type I X-ray bursts. The source was the first LMXB in which periodic absorption dips were detected
\citep{Walter_82, White_82}, the analysis of which allowed an orbital period
of 50 min and its compact nature to be estimated.
The optical counterpart of the source (a V=21 star) was discovered by
\cite{Grindlay_88}.
\cite{Swank_84} discussed that the companion star (CS) is not hydrogen exhausted by analyzing the thermonuclear flash models of X-ray bursts, while \cite{Pac_81}
showed that X-ray binary systems with orbital periods shorter than 81 min cannot contain hydrogen-rich CS.
\cite{Galloway_08}, by studying two consecutive
type I X-ray bursts temporally separated by 6.3 hours,
estimated that the CS contains a 20\% fraction of hydrogen under the hypothesis of a NS mass of 1.4 M$_{\odot}$.
The same authors estimated a distance to the source of 8.9 kpc.
The optical light curve of the source shows a periodicity of
$3027.4 \pm 0.4$ \citep{Grindlay_88}. \cite{Callanan_95}
found that the optical period remained stable over a baseline of seven years, and refined the
period estimate to $3027.551\pm0.004$ s.
The most recent X-ray orbital ephemeris of the source gives an orbital period of
$3000.6496\pm 0.0008$ s and an orbital period derivative of
$1.44(6) \times 10^{-11}$ s s$^{-1}$ \citep{iaria_15}.
The 1\% discrepancy between the optical and X-ray periods was explained by \cite{Grindlay_88}, who invoked the presence of a third body with a period of 2.5 d and a retrograde orbit. This scenario predicts that the optical period reflect the real period of the binary system, while
the observed X-ray period is modified by an increase in mass transfer influenced by the presence of the third body.
An alternative scenario suggests that the real period of XB 1916-053 is actually the observed X-ray period. The SU Ursae Majoris (SU UMa) superhump scenario was initially adopted by \citet{White_89}.
The superhump phenomenon is the appearance of a periodic or quasi-periodic modulation in the light curves of the SU UMa class of dwarf novae while showing superoutburst activity. The period of the superhumps is longer than the orbital period of the system in which it is observed. The commonly accepted interpretation of the phenomenon ties the longer modulations to the beat between the binary period and the apsidal precession period of the elliptical accretion disk.
Hydrodynamic simulations show that accretion disks in cataclysmic variable (CV) systems with a low-mass secondary are tidally unstable with a high probability of forming
a precessing eccentric disk \citep{whur_88}. \cite{Hirose_90}
found the dependence of the dynamical term from the apsidal precession period of the disk. \cite{Chou_01} analyzed optical and X-ray data estimating the
superhump period and the associated apsidal precession period of $3.9087 \pm 0.0008$ days. Adopting the SU SMa scenario and the relation discussed by \cite{Hirose_90},
\cite{Chou_01} inferred a mass ratio $q=M_2/M_1$ of 0.022, where $M_1$ and $M_2$ are the NS and CS mass, respectively.
\cite{Retter_02} detected a further X-ray periodicity at 2979 s in the RXTE light curves of XB 1916-053.
A similar periodicity had already been observed by \cite{Smale_89} analyzing {\it Ginga} data.
\cite{Retter_02} interpreted the 2979 s period invoking the presence of a negative superhump (also called an infrahump) where the X-ray period is the orbital period of the system that beats with
the nodal precession period of 4.86 days. This means that the disk is tilted with respect to the equatorial plane.
Finally, \cite{Hu_08}, by adopting the relation between the nodal angular frequency and the mass ratio $q$ proposed by \cite{Larwood_96} and assuming a nodal precession period of 4.86 days, inferred $q \simeq 0.045$.
\cite{iaria_15} used a baseline of 37 years to
infer the accurate orbital ephemeris of the source
from the dip arrival times. Although they proposed several models to describes the dip arrival times, the authors focused their attention on the ephemeris containing a quadratic term plus a long sinusoidal modulation. The quadratic term
implies the presence of an orbital period derivative of
$1.44(6) \times 10^{-11}$ s s$^{-1}$, while the long modulation of 25 years was explained with the presence of third body with mass of 0.055 M$_{\odot}$ orbiting around the binary system. \cite{iaria_15} showed that
the large orbital period derivative can be explained only assuming a high non-conservative mass transfer rate with only 8\% of the mass transfer rate accreting onto the NS.
The spectrum of XB 1916-053 shows absorption lines superimposed on the continuum emission. The \ion{Fe}{xxv} and \ion{Fe}{xxvi} absorption lines were detected for the first time by \cite{Boirin_04} from the analysis of {\it XMM-Newton} data.
\cite{Iaria_06}, using {\it Chandra} data, detected prominent absorption lines in the spectrum associated with
the presence of \ion{Ne}{x}, \ion{Mg}{xii}, \ion{Si}{xiv},
\ion{S}{xvi,} and \ion{Fe}{xxvi} ions. From the
plasma diagnostic the authors proposed that the lines should be produced at the outer rim of the accretion disk.
Analyzing the same data, \cite{Juett_06} suggested
a thickness for the X-ray absorber of $<3.2\times 10^9$ cm, assuming that it is located at the outer edge of the accretion disk.
Finally, analyzing {\it Suzaku} data, \cite{Gambino_19} set an upper limit of $<1 \times 10^{10 }$ cm
on the distance of the the ionized absorber from the NS,
placing it at the innermost region of the accretion disk.
In this work we update the orbital ephemeris of XB 1916-053
by expanding the available baseline to 40 years using ten {\it Chandra} observations and one {\it Swift/XRT} observation. We revise
the estimation of the NS mass suggesting that a mass of 1.4 M$_{\odot}$ can also explain the observed orbital period derivative and luminosities.
From theoretical considerations we discuss that a mass ratio of 0.048 can explain both the apsidal and the nodal precession period of the accretion disk taking into account the presence of a spiral wave in the accretion disk with a pitch angle lower than 30$^{\circ}$.
From the spectroscopic analysis we study the absorption lines detected in the spectrum during the persistent emission at different source luminosities and during the dips.
We note that the same {\it Chandra} data have recently been analyzed by \cite{Trueba_20}. The authors observed a redshift in the absorption lines that can be interpreted as a gravitational redshift. We reached similar results by assuming the same scenario, but adopting a different analysis.
\section{Observations}
\begin{table}
\centering
\begin{threeparttable}
\caption{List of observations}
\begin{tabular}{l l c c c }
\hline
\hline
ObsID. & Start Time (UT) & Exposure & Type I \\
& & Time (ks) & bursts & \\
\hline
20171 & 2018 June 11 20:11:59 & 22 & 1 \\
21103 & 2018 June 12 09:08:20 & 29 & 1 \\
21104 & 2018 June 13 17:03:34 & 23 & 1 \\
21105 & 2018 June 15 23:35:18 & 21.7 & 1 \\
20172 & 2018 July 31 08:02:08 & 30.5 & no \\
21662 & 2018 August 01 06:48:53 & 29.5 & 1 \\
21663 & 2018 August 02 19:47:53 & 30.5 & no \\
21664 & 2018 August 05 05:09:45 & 22.4 & no\\
21106 & 2018 August 06 03:26:33 & 21 & no\\
21666 & 2018 August 06 18:04:31 & 19 & no\\
4584{\textit{$\rm ^a$}} & 2004 August 07 02:34:45 & 46 & 2 \\
\hline
\hline
\end{tabular}
\begin{tablenotes}
\item[a] Observation 4584 was already analyzed by \cite{Iaria_06}.
\end{tablenotes}
\label{tab:1}
\end{threeparttable}
\end{table}
\begin{figure*}
\centering
\includegraphics[scale=.50]{fig1a.eps}\\
\includegraphics[scale=.50]{fig1b.eps}
\caption{First-order MEG+HEG light curves of the observations in the 0.3-10 keV energy range. The bin time is 128 s.}
\label{figure:1}
\end{figure*}
We analyzed ten new observations of XB 1916-053 collected by the \textit{Chandra} observatory in 2018 and the Swift/XRT observation (ObsID. 87248021) taken on 2017 July 30; furthermore, we reanalyzed the Chandra observation taken in 2004 August 7 \citep[see][]{Iaria_06}.
The new {\it Chandra} observations were performed from 2018 June 11 to August 6 using the onboard High Energy Transmission
Grating Spectrometer (HETGS) in timed graded mode; the detailed list of the observations is shown in Table \ref{tab:1}. HETGS consists of two types of transmission gratings, the medium energy grating (MEG) and the high-energy grating (HEG). The HETGS provides high-resolution spectroscopy from 1.2 to 31 $\AA$ (0.4-10 keV) with a peak spectral resolution of $\lambda/\Delta\lambda \sim$1000 at 12 $\AA$ for first-order HEG. The dispersed spectra were recorded with an array of six charge-coupled devices (CCDs) that are part of the Advanced CCD Imaging Spectrometer-S.
We reprocessed the data using the FTOOLS ver. 6.26.1 and the CIAO ver. 4.11 packages.
The brightness of the source required additional efforts to mitigate pileup issues. A 512 row subarray (the first row = 1) was applied during the observations, reducing the CCD frame time to 1.7 s. We ignored the zeroth-order events in our analysis focusing on the first-order HEG and MEG spectra.
We reprocessed the data adopting a scale factor (width$\_$factor$\_$hetg) that multiplies the approximate one sigma width of the HEG/MEG mask in the cross-dispersion direction of 15, instead of the default value of 35, in order to avoid that HEG and MEG overlap in the Fe K region of the spectrum. We obtained HEG and MEG region widths of 33 and 44 pixels, respectively.
We applied barycentric correction to the events adopting the source coordinated obtained by \cite{Iaria_06} and extracted the light curve for each observation retaining
only the first-order MEG and HEG data in the 0.3-10 keV energy range. We show the light curves of the ten observations in Fig. \ref{figure:1}. During the persistent emission, the observations
taken in 2018 June (ObsID. 20171, 21103, 21104, 21105) have a higher count rate than those taken in July-August (ObsID. 20172, 21662, 21663, 21664, 21106, 21666). Observations
21103, 21104, and 21105 show a count rate close to 7 c s$^{-1}$ (20171
shows a count rate of 5 c s$^{-1}$); observations 20172, 21662, 21663, and 21664 show a count rate of 4.5 c s$^{-1}$; and observations 21106 and 21666 show a count rate of 3 c s$^{-1}$. We observed five type I X-ray bursts, four during the high-flux state and one when the flux of the source dropped below 5 c s$^{-1}$. Furthermore, the light curves corresponding to observations 21104 and 21105 (when XB 1916-056 showed a high flux) do not show dips, while the dipping activity is intense at lower fluxes.
A similar behavior was observed during a long Suzaku observation of the source \citep[see][]{Gambino_19}.
XB 1916-053 was monitored by the {\it X-Ray Telescope}
\citep[XRT;][]{Burrows_05}
on board the Swift Observatory \citep{Ger_04}. We analyzed the Swift/XRT observation 87248021 taken on 2017 July 30
10:54:17, with elapsed and exposure times of
87 ks and 8.9 ks, respectively. The observation was performed in photon counting (PC) mode. The data were locally reprocessed by the UK Swift Science Data Center
using HEASOFT v6.26.1. We applied barycentric correction to the events using the ftool {\tt barycorr}\footnote{https://www.swift.ac.uk/analysis/xrt/barycorr.php}. Using XSELECT we extracted the source events in the 1-4 keV energy range adopting a circular region of 100$\arcsec$; the background events in the same energy range were extracted using a same size circular region
free from the source. Finally, we applied the exposure-correction to the source and background-light curves and extracted the background-subtracted light curve using the ftools {\tt xrtlccorr} and {\tt lcmath}\footnote{https://www.swift.ac.uk/analysis/xrt/lccorr.php}.
\section{Updated orbital ephemeris of XB 1916-053}
First, we filtered the light curves to exclude the type I X-ray bursts; then we selected the events in the 1-4 keV energy range to estimate the dip arrival times.
We estimated one dip arrival time from the Chandra observations taken in June, one from those taken in July-August, and one from the XRT observation.
The light curves were folded adopting $P_0=$3000.6511 s as the orbital period.
For the first dip arrival time we folded the light curve obtained from observation 20171, where the dips are the deepest, adopting a folding time
$T_{fold}=$58281 MJD; to obtain the second we folded together the light curves of observations 20172, 21664, 21106, and 21666, adopting
$T_{fold}=$58333.65 MJD. Finally, the XRT light curve was folded assuming $T_{fold}=$57965 MJD.
We fitted the dips with a simple model consisting of a step-and-ramp function, where the count rates before,
during, and after the dip are constant and the intensity changes
linearly during the dip transitions \citep[see][for a description]{iaria_15}. We show
the X-ray dip arrival times in Table \ref{tab:2}. \begin{table}
\centering
\begin{threeparttable}
\caption{Journal of the estimated dip arrival times}
\begin{tabular}{l@{\hspace{2pt}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}l@{\hspace{\tabcolsep}}}
\hline
\hline
Dip Time & Cycle & Delay (s) & Observation\\
(MJD;TDB) & & \\
\hline
57965.0004(2)& 225800& $1062 \pm 14$ & Swift/XRT\\
58281.0075(2) & 234899 & $1155 \pm 21$ & 1st ord. HEG+MEG\\
58333.65781(6) & 236415 & $ 1151 \pm 8$& 1st ord. HEG+MEG\\
\hline
\hline
\end{tabular}
Epoch of reference 50123.00873 MJD, orbital period 3000.6511 s.
\label{tab:2}
\end{threeparttable}
\end{table}
To obtain the delays with respect to a constant period reference, we used the period $P_0= 3000.6511$ s and reference epoch $T_0= 50123.00873$ MJD. Hence, we added the three dip arrival times to the 27 values reported by \cite{iaria_15}.
We initially fitted the delays with a quadratic function
$
y(t) = a+b t+ c t^2$,
where $t$ is the time in days with respect to $T_0$, $a=\Delta T_0$ is
the correction to $T_0$ in units of seconds, $b=\Delta P/P_0$ in units
of s d$^{-1}$ with $\Delta P$ the correction to the orbital period, and
$c= 1/2 \;\dot{P}/P_0$ in units of s d$^{-2}$ with
$\dot{P}$ is the orbital period derivative. The quadratic form
does not fit the data well, returning a $\chi^2({\rm {d.o.f.}})$ of
182(27). We show the delays (red points) and the quadratic curve (black curve)
in the top panel of Fig. \ref{figure:3}.
The best-fit values of the parameters are shown in the second column of
Table \ref{table:fit_result}.
\begin{figure}
\centering
\includegraphics[scale=.5]{fig2.eps}
\caption{Top panel: Delays vs time (red points). The black curve is the best fit of the LQ function. The brown curve is the best fit of the LQS function. Bottom panel: Residuals in units of $\sigma$ vs time, corresponding to the LQS function}
\label{figure:3}
\end{figure}
\begin{table}
\setlength{\tabcolsep}{3pt}
\caption{Best-fit values of the functions adopted to fit the
delays.\label{table:fit_result}}
\begin{center}
\begin{tabular}{l@{\hspace{2pt}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}}
\hline
\hline
Parameters & LQ & LS & LQS \\
\hline
$a$ (s) &
$93 \pm 22$ &
$5495\pm300$ &
$14 \pm 14$ \\
$b$ ($\times 10^{-3}$ s d$^{-1}$) &
$-4 \pm 4$ &
$-408\pm144$ &
$5 \pm 3$ \\
$c$ ($\times 10^{-5}$ s d$^{-2}$)&
$1.64 \pm 0.07$ &
-- &
$1.82 \pm 0.04$ \\
$A$ (s) &
-- &
$7343\pm1380$ &
$130 \pm 14$ \\
$t_\phi$ (d) &
-- &
$9826\pm1795$ &
$1262 \pm 133$ \\
$P_{mod}$ (d) &
-- &
73574 (fixed) &
$9099 \pm 302$ \\
$\chi^2$(d.o.f.) &
182(27)&
132(26) &
37(24) \\
\hline
\hline
\end{tabular}
\end{center} {\small \sc Note} \footnotesize--- The reported errors
are at 68\% confidence level. The best-fit parameters of the delays
are obtained using the functions LQ (column 2), LS (column 3), and
LQS (column 4).
\end{table}
To improve the modeling of the delays we added
a cubic
term to the previous parabolic function,
$
y(t) = a +bt+ct^2+dt^3$,
where $d$ is
defined as $\ddot{P}/(6 P_0)$ and $\ddot{P}$ indicates the
temporal second derivative of the orbital period. Fitting with this
cubic function, we obtained a $\chi^2({\rm {d.o.f.}})$ of 129(26) with an F-test probability of chance
improvement of only $3 \times 10^{-3}$ with respect to the
quadratic form (hereafter LQ function).
We also tried to fit the delays using a linear
plus a sinusoidal function \citep[the LS function in][]{iaria_15},
$y(t) = a+b t+ A \sin\left[\frac{2 \pi}{P_{mod}}(t-t_\phi)\right],$ where $a$ and $b$ are defined as above, while $A$ and $P_{mod}$ are the amplitude in seconds and the period in days of the sinusoidal function, respectively. Finally, $t_\phi$ is the time in days
referred to $T_0$ at which
the sinusoidal function is null. Since the best-fit value of $P_{mod}$ is larger then the available time span, we fixed it to the best-fit value $P_{mod}=73574$ d.
We found a $\chi^2({\rm {d.o.f.}})$ of 132(26); the best-fit parameters are shown in the third column of Table \ref{table:fit_result}.
We added a quadratic term to the LS function to take into account
the possible presence of an orbital period derivative
$y(t) = a+b t+ c t^2 + A \sin\left[\frac{2 \pi}{P_{mod}}(t-t_\phi)\right]$ (hereafter LQS function).
We obtained a value of $\chi^2({\rm {d.o.f.}})$ of
37(24) and an F-test probability of chance improvement
with respect to the LQ function of $1.8 \times 10^{-8}$.
We show the best-fit curve of the LQS model (brown curve)
and its relative residuals in units of sigma in the top and bottom panels of Fig. \ref{figure:3}, respectively.
The best-fit values are shown in the forth column of Table \ref{table:fit_result}.
The obtained best-fit values are compatible with the
previous ones reported by \cite{iaria_15}.
The corresponding LQS ephemeris is
\begin{equation}
\label{linear_quad_sin_eph}\begin{split}
T_{dip}(N) = {\rm MJD(TDB)}\; 50\,123.0089(2) + \frac{3\,000.65129(8)}{86\,400} N +\\
+2.54(6) \times 10^{-13} N^2 + \frac{A }{86\,400} \sin\left[\frac{2 \pi}{P_{mod}}\left(t -t_{\phi}\right)\right],
\end{split}\end{equation}
with $P_{mod} = 9099 \pm 302 $ d ($24.9\pm0.8$ yr), $t_{\phi} = 1262 \pm 133$ d, and $A=130 \pm 14$ s.
The corresponding orbital period derivative is $\dot{P} = 1.46(3) \times 10^{-11}$ s s$^{-1}$.
Our analysis suggests that a quadratic or a quadratic plus a cubic
term do not adequately fit the delays, as already shown by \cite{iaria_15}.
We folded the 0.3-10 keV first-order MEG+HEG light curves
obtained from the observations (type I X-ray bursts excluded), adopting the orbital period inferred by the LQS ephemeris and changing the reference time so that the dip falls at the orbital phase 0.5. The epoch-folded light curves are shown
in Fig. \ref{figure:5}.
\begin{figure}
\centering
\includegraphics[scale=.52]{fig3.eps}
\caption{First-order MEG+HEG folded light curves in the 0.3-10 keV energy range adopting the LQS ephemeris. An ad hoc phase shift of 0.5 is imposed. The label in the plot shows the correspondence
between colors and observations. The blue dashed horizontal lines indicate the arbitrary count rate threshold of 3.5 c s$^{-1}$ and 5.5 c s$^{-1}$ adopted to select the four sets of observations (see text).}
\label{figure:5}
\end{figure}
We distinguish four different sets of epoch-folded light curves at different count rates: the first set (hereafter set A) associated with observation 4584 has a persistent count rate higher than 10 c s$^{-1}$ (gold curve in Fig. \ref{figure:5}); the folded curves of the second set (hereafter set B),
have a persistent count rate between 6 and 8 c s$^{-1}$ and are associated with observations 21103, 21104, and 21105 (orange, blue, and magenta curve in Fig. \ref{figure:5}); the third set of curves (hereafter set C) has a persistent count rate between 3.5 and 5.5 c s$^{-1}$ and is associated with observations 20171, 20172, 21662, 21663, and 21664 (black, red, brown, gray, and purple curve in Fig. \ref{figure:5});
the fourth set (hereafter set D) shows a count rate close to 3.5 c s$^{-1}$ outside the dip associated with observations 21106 and 21666 (yellow and green curve in Fig. \ref{figure:5}).
\begin{table}
\setlength{\tabcolsep}{3pt}
\caption{List of the combined spectra\label{table:4}}
\begin{center}
\scriptsize
\begin{tabular}{l@{\hspace{2pt}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}}
\hline
\hline
Combined first-order & Exp. time & ObsID. & Phase interval\\
HEG+MEG spectrum & (ks) & & excluded \\
\hline
spectrum A & 38.8 & 4584 & 0.45-0.60 \\ \\
\multirow{3}{*}{spectrum B} & \multirow{3}{*}{67.3} & 21104 & \\
& & 21105 & \\
& & 21103 & 0.45-0.60 \\\\
\multirow{5}{*}{spectrum C}& \multirow{5}{*}{104.2} & 20171 & 0.40-0.65 \\
& & 21662 & 0.45-0.60 \\
& & 21663 & 0.55-0.66 \\
& & 21664 & 0.40-0.70 \\
& & 20172 & 0.40-0.65 \\\\
\multirow{2}{*}{spectrum D} & \multirow{2}{*}{31.4} & 21106 & 0.40-0.60 \\
& & 21666 & 0.40-0.60 \\
\hline
\hline
\end{tabular}
\label{tab:excluded_phases}
\end{center}
\end{table}
\begin{table*}
\centering
\begin{threeparttable}
\caption{Best-fit values of {\tt Model 1} and {\tt Model 2}}
\scriptsize
\begin{tabular}{l@{\hspace{2pt}}l@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}}
\hline
\hline
Model & Component & \multicolumn{4}{c}{{\tt Model 1{\textit{$\rm ^a$}} }} &
\multicolumn{4}{c}{{\tt Model 2{\textit{$\rm ^b$}}}}\\
& & A & B & C &D & A & B & C & D\\
\hline
{\sc Constant} &C$_{\rm HEG}$ &
$0.972\pm0.006 $ &
$0.970 \pm 0.006$ &
$0.966 \pm 0.006$ &
$0.956 \pm 0.015$ &
$0.972\pm0.006 $ &
$0.969 \pm 0.006$ &
$0.966 \pm 0.006$ &
$0.956 \pm 0.015$ \\\\
{\sc Edge} & E (keV) &
\multicolumn{4}{c}{--} &
\multicolumn{4}{c}{0.871 (fixed)} \\
& $\tau$ &
\multicolumn{4}{c}{--} &
\multicolumn{4}{c}{$0.14\pm0.05$} \\ \\
{\sc TBabs} & N$_{\rm H}${\textit{$\rm ^c$}} &
$0.56 \pm 0.07$ &
$0.63\pm0.08 $&
\multicolumn{2}{c}{$0.53 \pm 0.09$} &
$0.55 \pm 0.04$ &
$0.60\pm0.05$&
\multicolumn{2}{c}{$0.50\pm0.05$} \\\\
{\sc partcov} & $f_{\rm cabs}${\textit{$\rm ^d$}} &
\multicolumn{4}{c}{--} &
\multicolumn{4}{c}{$>0.86$}\\
{\sc Cabs} & N$_{{\rm H}_{\rm cabs}}${\textit{$\rm ^e$}} &
\multicolumn{4}{c}{--} &
$19^{+7}_{-3}$ &
$14^{+5}_{-2}$ &
$17^{+6}_{-3} $&
$10\pm7$ \\\\
{\sc zxipcf} & N$_{{\rm H}_{\xi}}${\textit{$\rm ^c$}} &
\multicolumn{4}{c}{--} &
$19^{+7}_{-3}$ &
$14^{+5}_{-2}$ &
$17^{+6}_{-3} $&
$10\pm7$ \\\\
& log($\xi$) &
\multicolumn{4}{c}{--} &
\multicolumn{4}{c}{$4.33\pm0.02$}\\
& $f$ &
\multicolumn{4}{c}{--} &
\multicolumn{4}{c}{$>0.86$}\\
& $z$ ($\times 10^{-3}$)&
\multicolumn{4}{c}{--} &
$1.3\pm0.2$ &
$1.1\pm0.2 $&
$1.1 \pm 0.2$&
$1.0^{+1.3}_{-1.6}$\\\\
{\sc nthcomp} & $\Gamma$ &
$1.613\pm0.012$ &
$1.721\pm0.015 $ &
$1.841\pm 0.015 $ &
$1.80 ^{+0.02}_{-0.03}$ &
$1.620\pm0.009$ &
$1.730\pm0.013 $&
$1.848\pm0.014 $&
$1.803^{+0.014}_{-0.010} $\\
& kT$_{bb}$ (keV) &
\multicolumn{4}{c}{$0.17\pm 0.05$} &
\multicolumn{4}{c}{$0.16\pm0.04 $} \\
& kT$_{e}$ (keV) &
$2.9\pm 0.3$ &
$2.6 \pm 0.2$ &
$3.8^{+1.4}_{-0.9}$&
$3.3^{+4.9}_{-0.6}$&
$3.2^{+0.5}_{-0.3}$&
$2.6\pm0.3$&
$>3.6$&
$>2.8$\\
& Norm ($\times 10^{-2}$) &
$11.7\pm 1.3$ &
$10.9 \pm 1.3$ &
$7.2\pm1.0$&
$4.8\pm0.7$&
$14.3\pm1.0$&
$12.8\pm1.0$&
$8.7\pm0.8$&
$5.4\pm0.5$\\\\
& $\chi^2/dof$ &
\multicolumn{4}{c}{4957/4553} &
\multicolumn{4}{c}{3944/4542}\\
\hline
\hline
\end{tabular}
\begin{tablenotes}
\item[] The associated errors are at 90\% confidence level.
\item[a] Model 1 = {\sc Const*TBabs*nthcomp}.
\item[b] Model 2 = {\sc Const*TBabs*(partcov*Cabs)*zxipcf*nthcomp}.
\item[c] \textrm{Equivalent hydrogen column density in units of 10$^{22}$ atoms cm$^{-2}$}.
\item[d] The value of the parameter is tied to the value of $f$.
\item[e] The value of the parameter is tied to the value of N$_{{\rm H}_{\xi}}$.
\end{tablenotes}
\label{tab:mod1mod2}
\end{threeparttable}
\end{table*}
\section{Spectral analysis}
We extracted the combined spectrum from each distinguished set, calling the spectra A, B, C, and D, using the CIAO script {\tt combine$\_$grating$\_$spectra} and
combining the corresponding first-order MEG and HEG spectra.
Since our aim is the spectral analysis out of the dips, we excluded
the orbital phases in which the dip falls when it is present. The excluded phases are shown in Table \ref{table:4}.
Spectrum B was obtained by excluding the phases between 0.45 and 0.6 from observation 21103; the other two observations do not show dipping activity, and we included all the phases.
We rebinned the resulting first-order MEG and HEG spectra of selections A, B, C, and D to have at least 250 counts per energy channel.
Such a large rebinning could affect the spectroscopic study of narrow discrete features because it risks losing energy resolution. However, we tried to fit these spectra using a rebinning so as to have 25 counts per energy channel, verifying that we obtained results that are compatible with those reported in the following.
For the analysis we adopted the energy ranges 0.7-7 keV and 0.9-10 keV for the first-order MEG and HEG spectrum, respectively.
To fit the spectra we used XSPEC v12.10.1p; we adopted the cosmic abundances and the cross sections
derived by \cite{Wilms_00} and \cite{Verner_96}, respectively.
To take into account the interstellar absorption we adopted the T{\"u}bingen-Boulder model ({\sc TBabs} in XSPEC). We fitted the spectrum using a Comptonized component \cite[{\sc nthcomp} in XSPEC;][]{Zdi_96}.
Finally, we added a constant to take into account the different normalizations of the two instruments.
The initial model, called {\tt Model 1}, is defined as
$$
\texttt{Model 1} = \textsc{Const*TBabs*nthcomp}.
$$
We fixed at 0 the value of the parameter {\tt inp$\_$type} in {\sc nthcomp}, assuming the seed-photon spectrum injected in the Comptonizing cloud to be a blackbody.
The spectra A, B, C, and D were fitted simultaneously.
The value of the seed-photon temperature k$T_{bb}$ was tied
at the same value in each spectrum and, finally, the value of the interstellar equivalent neutral hydrogen column density was tied at the same value in spectra C and D.
We show the unfolded spectra and the residuals in the left panels of Fig. \ref{figure:6}. The best-fit parameters
are shown in Table \ref{tab:mod1mod2}. Strong residuals compatible with absorption lines are visible at 1 keV, 1.47 keV, 2 keV, 2.62 keV, and 6.97 keV.
Furthermore, below 0.9 keV there is a clear excess in the residuals.
\begin{figure*}
\centering
\includegraphics[scale=.6]{fig4a.eps}
\includegraphics[scale=.6]{fig4b.eps}\\
\caption{Unfolded spectrum and residuals adopting {\tt Model 1} (left panels) and {\tt Model 2} (right panels).
The black, red, green, and blue data correspond to the
first-order MEG of spectra A, B, C, and D, respectively. The cyan, purple, yellow, and orange data correspond to first-order HEG spectrum of spectra A, B, C, and D, respectively. The prominent absorption lines in residuals of {\tt Model 1} (bottom left panel) are associated with the presence of \ion{Ne}{x}, \ion{Mg}{xii}, \ion{Si}{xiv}, \ion{S }{xvi,} and \ion{Fe}{xxvi} ions.}
\label{figure:6}
\end{figure*}
\begin{table}
\centering
\begin{threeparttable}
\caption{Identified Gaussian absorption lines.}
\scriptsize
\begin{tabular}{l@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}} c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}}
\hline
\hline
Component & Spectrum A & Spectrum B& Spectrum C&Spectrum D \\
\hline\\
& \multicolumn{4}{c}{\ion{Ne}{x} K$\alpha$ 1s-2p (1.0218 keV)} \\\\
E (keV) &
$1.020\pm 0.002$&
$1.014\pm0.009 $ &
$1.019^{+0.009}_{-0.019}$&
-- \\
$\sigma$ (eV) &
$<2.0$ &
$<17$ &
$<31$&
-- \\
Intensity{\textit{$\rm ^a$}} &
$-8\pm2$&
$-8\pm5$&
$>-6$&
--\\
Eq. width{\textit{$\rm ^b$}} &
$-1.8\pm0.5$ &
$-2.1\pm0.8$ &
$-1.1\pm0.8$&
--\\
Bin size{\textit{$\rm ^c$}} (eV)&
6 &
10&
7&
--\\\\
& \multicolumn{4}{c}{\ion{Mg}{xii} K$\alpha$ 1s-2p (1.4723 keV)} \\\\
E (keV) &
$1.4715\pm0.0004$&
$1.4720\pm0.0005$ &
$1.4720^{+0.0005}_{-0.0009}$ &
--\\
$\sigma$ (eV) &
$1.0\pm0.5$ &
$<1.8$ &
$<2.2$&
-- \\
Intensity &
$-5.9\pm0.7$&
$-5.0\pm1.0$&
$-2.9\pm0.6$&
--\\
Eq. width &
$-1.3\pm0.2$ &
$-1.1\pm0.2$ &
$-1.1\pm0.2$ &
-- \\
Bin size (eV)&
2 &
2&
2&
--\\\\
& \multicolumn{4}{c}{\ion{Si}{xiv} K$\alpha$ 1s-2p (2.0055 keV)} \\\\
E (keV) &
$2.0044\pm0.0004$ &
$2.0040^{+0.0007}_{-0.0002}$ &
$2.0039\pm0.0010$ &
$2.005^{+0.003}_{-0.014} $\\
$\sigma$ (eV) &
$1.4^{+0.5}_{-1.2}$ &
$0.6^{+1.1}_{-0.6}$ &
$3.4^{+0.7}_{-1.2}$ &
$<29$\\
Intensity &
$-9.0\pm 0.8$&
$-6.7\pm0.8$&
$-5.2\pm0.8$&
$-2.2\pm1.5$\\
Eq. width &
$-2.6\pm0.3$ &
$-2.2\pm0.2$ &
$-2.8\pm0.3$ &
$-1.8\pm0.6$ \\
Bin size (eV)&
3 &
3&
3&
20\\\\
& \multicolumn{4}{c}{\ion{Si}{xiv} K$\alpha$ 1s-3p (2.376 keV)} \\\\
E (keV) &
$2.372\pm0.002$ &
-- &
-- &
-- \\
$\sigma$ (eV) &
$1.3^{+3.6}_{-1.3}$ &
-- &
-- &
-- \\
Intensity &
$-6\pm 2$&
--&
--&
--\\
Eq. width &
$-1.9\pm0.5$ &
-- &
-- &
-- \\
Bin size (eV)&
9 &
--&
--&
--\\
& \multicolumn{4}{c}{\ion{S}{xvi} K$\alpha$ 1s-2p (2.6217 keV) } \\\\
E (keV) &
$2.6157^{+0.0002}_{-0.0027}$&
$2.621^{+0.002}_{-0.008}$ &
$2.617^{+0.006}_{-0.004}$ &
-- \\
$\sigma$ (eV) &
$<3.0$ &
$7\pm6$ &
$5\pm5$
&--\\
Intensity &
$-9\pm 2$ &
$-7\pm4$&
$-3.3\pm1.5$&
--\\
Eq. width &
$-3.4\pm0.7$ &
$-3.3\pm0.9$ &
$-2.7\pm0.7$ &
--\\
Bin size (eV)&
8 &
8&
8&
--\\\\
& \multicolumn{4}{c}{\ion{Ca}{xx} K$\alpha$ 1s-2p (4.1050 keV)} \\\\
E (keV) &
$4.102\pm0.006$ &
-- &
-- &
--\\
$\sigma$ (eV) &
$8^{+7}_{-8}$ &
-- &
--&
-- \\
Intensity &
$-6\pm2$ &
--&
--&
--\\
Eq. width &
$-2.9 ^{+0.3}_{-2.5}$ &
-- &
-- &
-- \\
Bin size (eV)&
20 &
--&
--&
--\\\\
& \multicolumn{4}{c}{\ion{Fe}{xxvi} K$\alpha$ 1s-2p (6.9662 keV)} \\\\
E (keV) &
$6.98^{+0.04}_{-0.09}$ &
$6.956^{+0.028}_{-0.003}$ &
$6.96\pm0.04$ &
-- \\
$\sigma$ (eV) &
$<42$ &
$<60$ &
$<35$ &
-- \\
Intensity &
$-16\pm6$ &
$-18\pm6$ &
$-7\pm2$&
--\\
Eq. width &
$-13.1^{+1.0}_{-26.3}$ &
$-33\pm7$ &
$-43^{+23}_{-4}$ &
-- \\
Bin size (eV)&
100 &
100&
120&
--\\
\hline
\hline
\end{tabular}
\begin{tablenotes}
\item[] The associated errors are at 68\% confidence level.
\item[a] The line intensity is in units of $\times 10^{-5}$ photons cm$^{-2}$ s$^{-1}$.
\item[b] The equivalent width is in units of eV. \item[c] The size of the energy channel at the detected lines is estimated from MEG spectra except for the \ion{Ca}{xx} and \ion{Fe}{xxvi} lines.
\end{tablenotes}
\label{tab:line}
\end{threeparttable}
\end{table}
We fitted each observed absorption line with a {\sc Gaussian} component. The energies, widths, intensities, and equivalent widths
of each identified line are shown in Table \ref{tab:line} for each spectrum. We detected the absorption lines
associated with the presence of ions of \ion{Ne}{x}, \ion{Mg}{xii}, \ion{Si}{xiv}, \ion{S}{xvi}, \ion{Ca}{xx,} and \ion{Fe}{xxvi}. The line associated with the
\ion{Ca}{xx} ions is detected only in spectrum A. Spectrum D only showed the line associated with \ion{Si}{xiv}, probably because of the low flux of the spectrum. For the sake of clarity the sizes of the energy channels at which we identified the absorption lines are shown in Table \ref{tab:line}.
To account for the presence of the prominent absorption lines, to the continuum of {\tt Model 1}
we added the multiplicative
component {\sc zxipcf}\footnote{\scriptsize{https://heasarc.gsfc.nasa.gov/xanadu/xspec/models/zxipcf.html}}, which
takes into account a partial covering of ionized absorbing material. The component reproduces the absorption from photoionized matter illuminated by a power-law source with spectral index $\Gamma =2.2$, and it
assumes that the photoionized absorber has a microturbulent
velocity of 200 km s$^{-1}$ \citep[see][for applications to AGN and Seyfert 1 galaxies]{Reeves_08,Miller_07}.
Recently, this component was adopted to fit the absorbing features in the Fe-K region associated with highly ionized matter surrounding the eclipsing NS-LMXB AX J1745.6-2901 \citep{Ponti_15}, the dipping source XB 1916-053 \citep{Gambino_19}, and the eclipsing source
MXB 1659-298 \citep{Iaria_19}.
The parameters of this component are N$_{{\rm H}_{\xi}}$, log($\xi$), $f$, and $z$: N$_{{\rm H}_{\xi}}$ describes the equivalent hydrogen column density associated with the ionized absorber, log($\xi$) describes the ionization degree of the absorbing material, $f$ is a dimensionless covering fraction of the emitting region, and $z$ is the redshift associated with the absorption features. The parameter $\xi$ is defined as $\xi = L_x/(n_H r^2)$, where $L_x$ is the X-ray luminosity incident on the absorbing material, $r$ is the distance of the absorber from the X-ray source, and $n_H$ is the hydrogen atom density of the ionized absorber.
To take into account that the ionized absorber can scatter the radiation out of the line of sight via Thomson or Compton scattering, we added the multiplicative component {\sc cabs} to the model. The only parameter of the component {\sc cabs} is $N_{H_{\rm cabs}}$, which describes the equivalent hydrogen column density associated with the scattering cloud. Finally,
to account for a partial covering of the scattering material, we multiplied {\sc cabs} by the component {\sc partcov}. This is a convolution model that allows us to convert an absorption component into a partially covering absorption component, quantified by the parameter $f_{\rm cabs}$. In order to make the model self-consistent, we tied $f_{\rm cabs}$ to the parameter $f$ of the ionized absorber. Moreover, we linked
the equivalent hydrogen column density parameter of the scattering cloud, $N_{H_{\rm cabs}}$, to the value of the equivalent hydrogen column density of the ionized absorber, N$_{{\rm H}_{\xi}}$.
While fitting simultaneously spectra A, B, C and D, we imposed $f$, log($\xi$), and the seed-photon temperature to assume the same value for each spectrum. The model closely fits the absorption lines; however, the presence of an excess in the residuals below 0.9 persists. For this reason we added an absorption edge with the energy threshold fixed at 0.871 keV to account for the presence of \ion{O}{viii} ions along the line of sight, recently observed by \cite{Gambino_19} analyzing the {\it Suzaku} spectra of the source.
The adopted model, hereafter called {\tt Model 2}, is defined as
\begin{equation*}
\begin{split}
\texttt{Model 2} = \textsc{Const*Edge*TBabs*(partcov*cabs)*}\\
\textsc{{zxipcf*nthcomp}}.
\end{split}
\end{equation*}
The addition of {\sc zxipcf} and the absorption edge improves the fits, with a $\chi^2$(d.o.f.) of 3944(4542) and a $\Delta \chi^2$ of 1013 with respect to {\tt Model 1}. We show
the best-fit values in Table \ref{tab:mod1mod2} for each spectrum.
The unfolded spectra and the corresponding residuals are shown in the right panels of Fig. \ref{figure:6}; the residuals associated with the absorption lines and the excess below 0.9 keV observed adopting {\tt Model 1} are now absent.
We found that the ionized absorber covers more than 86\% of the emitting source and the ionization parameter log($\xi$) is $4.33\pm0.02$. The equivalent hydrogen column density associated with the ionized absorber
N$_{{\rm H}_{\xi}}$ is $(19^{+7}_{-3})\times 10^{22}$,
$(14^{+5}_{-2}) \times 10^{22}$, $(17^{+6}_{-3})\times 10^{22} $,
and $(10\pm7) \times 10^{22}$ cm$^{-2}$ for spectra A, B, C, and D, respectively. Furthermore, we observed a similar redshift value for each spectrum (between $1.0\times10^{-3}$ and $1.3\times10^{-3}$).
The $\Gamma$ parameter of {\sc nthcomp}
goes from 1.6 for spectrum A to 1.8 for
spectrum C and D, indicating that the spectral shape softens
going from spectrum A to spectrum D. The electron temperature of the Comptonizing cloud is compatible with 3-4 keV for all the spectra even if it is not well constrained for spectra C and D, while the seed-photon temperature is
$0.16 \pm 0.04$ keV.
Finally, the best-fit value of the optical depth of the absorption edge is $0.14\pm0.05$ and the equivalent hydrogen column density of the interstellar matter is compatible with $ 0.5 \times 10^{22}$ cm$^{-2}$ for each spectrum.
Assuming a distance to the source of 8.9 kpc \citep{Galloway_08}, the unabsorbed luminosity in the 0.1-100 keV energy range is $1.46 \times 10^{37}$ erg s$^{-1}$, $1.0 \times 10^{37}$ erg s$^{-1}$, $0.67 \times 10^{37}$ erg s$^{-1}$, and
$0.41 \times 10^{37}$ erg s$^{-1}$ for A, B, C, and D, respectively.
\section{Spectral analysis of the dips}
To analyze the average spectrum during the dips we
extracted the events from observations 20171, 20172, 21106, 21662, 21663, 21664, and 21666. The dip events were selected from observation 20171 taking into account
the good times intervals (GTIs) obtained from the first-order MEG+HEG light curve in the 0.3-10 keV energy range where the count rate was lower than 3 c s$^{-1}$. Similarly, we used the threshold of 2 c s$^{-1}$ for ObsID. 20172, 2.4 c s$^{-1}$ for ObsID. 21662, 2 c s$^{-1}$ for ObsID. 21663 and 21664, 1.5 c s$^{-1}$ for ObsID. 21106, and 1.1 c s$^{-1}$ for ObsID. 21666. We extracted
the first-order HEG and MEG spectrum from each observation and
combined them using the CIAO script {\tt combine$\_$grating$\_$spectra}.
Each spectrum was then grouped to have at least 100 counts per energy channel; the resulting spectrum has an exposure time of
9.5 ks. The energy range adopted for the spectral analysis is 0.9-7 keV and 1-9 keV for MEG and HEG, respectively.
We fitted the spectrum adopting {\tt Model 2};
however, we excluded the {\sc edge} component because of the low flux of the spectrum. does not allow us to constrain it. We kept the photon-index value,
the electron temperature, and the seed-photon temperature of the {\sc nthcomp} component fixed to 1.8, 4.3 keV, and 0.16 keV, respectively. We did this so that the dip-spectrum was extracted from the observations belonging to set C and D, under the assumption that the spectral shape of the continuum emission is similar both inside and outside the dips.
We show the best-fit parameters in Table \ref{tab:dip}; the unfolded spectrum and the corresponding residuals are shown in Fig. \ref{figure:7}.
\begin{figure}
\includegraphics[scale=.64]{fig5.eps}
\caption{Unfolded spectrum and residuals of the dip spectrum
adopting {\tt Model 2}. The black and red data correspond to the first-order MEG and HEG spectra, respectively.}
\label{figure:7}
\end{figure}
\begin{table}
\centering
\begin{threeparttable}
\caption{Best-fit parameters of the spectrum during the dip.}
\begin{tabular}{l@{\hspace{2pt}}l@{\hspace{\tabcolsep}}l@{\hspace{\tabcolsep}}}
\hline
\hline
& & Dip spectrum\\
\hline
Model & Component & \\
\hline
{\sc Constant} &C$_{\rm HEG}$ &
$0.92\pm0.08$ \\\\
{\sc partcov} & $f_{\rm cabs}${\textit{$\rm ^a$}} & $>0.68$ \\
{\sc cabs} & N$_{\rm H_{\rm cabs}}${\textit{$\rm ^b$}} ($\times 10^{22}$) &
$61^{+56}_{-10}$ \\\\
{\sc zxipcf} & N${_{\rm H}}_{\xi}$ ($\times 10^{22}$) & $61^{+56}_{-10}$ \\
&log($\xi$)& $2.8^{+0.2}_{-0.4} $\\
& $f$ & $>0.68$ \\
& Redshift& 0 (fixed) \\ \\
{\sc TBabs} & N$_{\rm H}$($\times 10^{22}$) & $0.9\pm0.2$ \\\\
{\sc nthcomp} & $\Gamma$ &
1.8 (fixed) \\
& kT$_{bb}$ (keV) & 0.16 (fixed) \\
& kT$_{e}$ (keV) & 4.3 (fixed) \\
& Norm ($\times 10^{-2}$) &
$7.2\pm1.4$ \\\\
& $\chi^2/dof$ &
18.3/76 \\
\hline
\hline
\end{tabular}
\begin{tablenotes}
\item[a] The value of the parameter is tied to the value of $f$.
\item[b] The value of the parameter is tied to the value of N$_{{\rm H}_{\xi}}$.
\end{tablenotes}
\label{tab:dip}
\end{threeparttable}
\end{table}
We found that during the dip the ionized absorber is characterized by log($\xi$)$= 2.8^{+0.2}_{-0.4}$, suggesting a low ionization of the absorber. The equivalent hydrogen column density
N${_{\rm H}}_{\xi}$ associated with the ionized absorber is $(61^{+56}_{-10})\times 10^{22}$ cm$^{-2}$, which is larger than the values obtained during the persistent emission by a factor between 3 and 6.
The covering fraction is $>0.68$, compatible with the values obtained for spectra A, B, C, and D.
The equivalent hydrogen column density associated with the interstellar matter is $(0.9\pm0.2)\times 10^{22}$ cm$^{-2}$. This value is larger than the observed value during the persistent emission. We cannot exclude that a partial contribution to this value comes from local neutral hydrogen in the hypothesis that during the dip the absorber is composed of neutral and mildly ionized matter.
Finally, the 0.1-100 keV unabsorbed luminosity, assuming a distance to the source of 9 kpc, is $0.57 \times 10^{37}$ erg s$^{-1}$.
\section{Discussion}
From the analysis of the {\it Chandra} and {\it Swift/XRT} observations we inferred three new dip arrival times that, added to the previous 27 reported by
\cite{iaria_15}, allowed us to update the orbital ephemeris of XB 1916-053 by extending the temporal baseline from 36 to 40 years. The largest baseline excludes some solutions previously
plausible with the available data, such as the LQ ephemeris shown by \cite{iaria_15}.
We find that the LQS ephemeris better models the dip arrival times,
improving the constraints on
the orbital period derivative and the sinusoidal modulation.
The orbital period derivative is $\dot{P}=1.46(3)\times 10^{-11}$ s s$^{-1}$ and the sinusoidal modulation has a period of $P_{\rm {mod}}=9099\pm302$ d.
\cite{iaria_15} discussed that such a large orbital period derivative
can be explained only assuming that the mass transfer rate
from the CS to the NS is highly non-conservative. The authors estimated that more that 90\% of the mass transfer rate from the CS is lost by the system, the mass ratio $q=m_2/m_1$
($m_1$ and $m_2$ are the NS and CS mass in units of solar masses) is close to 0.013 and the NS mass should be larger than 2.1 M$_{\odot}$.
Here we show that the conclusions of \cite{iaria_15} were strongly constrained by the hypothesis that the matter
leaves the system from the inner Lagrangian point. To discuss this point we have to estimate the mass ratio $q$ of XB 1916-053.
\subsection{Mass ratio $q$ of XB 1916-053}
XB 1916-053 shows a superhump period $P_{sh}$ \citep{Chou_01} and a negative superhump period, also called infrahump period, $P_{ih}$ \citep{Retter_02, Hu_08}.
\cite{Chou_01} discussed that the optical modulation close to 3028 s is likely caused by the coupling of the orbital motion with a 3.9-day disk apsidal precession period, like the superhumps in SU UMa-type dwarf novae, obtaining from this assumption $q \simeq 0.022$, while \cite{Hu_08} discussed the infrahump period of 2979.3(1.1) s detected by \cite{Retter_02} as
being due to the nodal precession period of the tilted accretion disk, and finding $q \simeq 0.045$.
Considering that the orbital period of the system is close to 3000.66 s, the apsidal precession period of the disk should be 3.9087(8) d, while the retrograde nodal precession period 4.86 d. However, the reported values of $q$ do not depict a self-consistent scenario.
We show below that the value of $q=0.048$ is consistent with the detected values of $P_{sh}$ and $P_{ih}$ only assuming an outer radius of the disk truncated at a 3:1 resonance.
\cite{Hirose_90}, performing hydrodynamic simulations of accretion disks to study the superhump phenomenon in SU UMa stars, showed that the apsidal precession frequency $\omega_p$ can be written as $\omega_p/\omega_{\rm orb}= Z(r) q/(1+q)^{1/2}$, where
\begin{equation*}
\begin{split}
Z(r)=\frac{1}{2}\frac{1}{r^{1/2}}\frac{d}{dr}\left[ r^2\frac{d}{dr}B_0(r)\right],
\end{split}
\label{eq:a}
\end{equation*}
and
\begin{equation*}
\begin{split}
B_0(r)=\frac{1}{2}b_{1/2}^{(0)}=1+\frac{1}{4}r^2+\frac{9}{64}r^4+...,
\end{split}
\label{eq:laplace}
\end{equation*}
which is the Laplace coefficient of order 0 in celestial mechanics
\citep[see ][chapter 15, eq. 42]{Brouwer_61} and $r$ is the ratio of the accretion disk radius $r_{\rm disk}$ to the orbital separation $a$. Combining the expressions
we find that
\begin{equation}
\begin{split}
\frac{\omega_p}{\omega_{\rm orb}} =\frac{q}{(1+q)^{1/2}}r^{3/2}\left[ 0.75+\frac{45}{32}r^2\right].
\end{split}
\label{eq:osaki}
\end{equation}
\cite{Hirose_90} demonstrated that the tidal instability in the accretion disks is caused by the resonance between the particle orbits in the disk and the companion star with a 3:1 period ratio in SU UMa dwarf novae. Since $r=r_{\rm disk}/a$ can be expressed by the equation
\begin{equation}
\begin{split}
\frac{r_{\rm disk}}{a} =\frac{R_{jk}}{a} =\left(\frac{j-k}{j}\right)^{2/3} (1+q)^{-1/3},
\end{split}
\label{eq:rdisk}
\end{equation}
with $j=3$ and $k=2$ for a 3:1 resonance \citep[see][eq. 5.125]{Frank_02}, we can combine
the two last equations to obtain
\begin{equation}
\begin{split}
\frac{\omega_p}{\omega_{\rm orb}} = \frac{1}{3}\frac{q}{1+q}[0.75+0.325(1+q)^{-2/3}];
\end{split}
\label{eq:ompomorb}
\end{equation}
using the value $\omega_p/\omega_{\rm orb}\simeq 0.0089$
inferred by \cite{Chou_01} we find $q\simeq 0.0256$.
However, \cite{Hirose_90} estimated the apsidal precession period of the disk taking into account only the dynamical effects of the matter.
\cite{Lubow_91a,Lubow_91b,Lubow_92} showed that the apsidal precession rate $\omega$ for an eccentric disk is given
by three terms $\omega=\omega_{\rm dyn}+\omega_{\rm press}+\omega_{\rm tr}$,
where $\omega_{\rm dyn}$ is the dynamical precession frequency discussed by \cite{Hirose_90} and shown above, $\omega_{\rm press}$ is a term due to the pressure, and
$\omega_{\rm tr}$ is a transient term related to the time-derivative of the mode giving rise to the dynamical precession that
can be neglected in steady state. Since XB 1916-053 is a persistent source that does not show outburst, and since \cite{Callanan_95}
verified the stability of the optical period over seven years, we do not consider the last term as we assume
that the disk precesses in a steady state.
\cite{Lubow_92} showed that the pressure term can be written as $\omega_{\rm press}$=$-k^2c_s^2/(2\Omega)$, where
$k$ is the radial wavenumber of the mode, $c_s$ the sound speed of the gas, and $\Omega$ is the frequency of a particle in the disk at a given radius. The negative sign
is due to the pressure
term associated with a spiral arm in the disk, which acts in the opposite sense with respect to the dynamical term.
For a spiral wave, the pitch angle $\theta$ is related to $k$ by the relation $\tan \theta = (kR)^{-1}$, where $R$ is the distance from the central source.
\cite{Pearson_06}, considering a 3:1 resonance and adopting eq. \ref{eq:rdisk}, expressed the term associated with the pressure of the spiral wave $\omega_{\rm press}$ as
\begin{equation}
\begin{split}
\frac{\omega_{\rm press}}{\omega_{\rm orb}} = -\frac{j^{1/3}}{2}(1+q)^{2/3}\left( \frac{c_s}{\omega_{\rm orb}\; a} \tan^{-1}\theta \right)^2 = - j^{1/3} \eta_A (1+q)^{2/3},
\end{split}
\label{eq:pearson}
\end{equation}
where $a$ is the orbital separation, $j=3$, and $\eta_A = 0.5 [c_s/(\omega_{\rm orb}\; a)]^2 \tan^{-2}\theta$. We combined eqs. \ref{eq:ompomorb} and \ref{eq:pearson} and obtained
\begin{equation}
\begin{split}
\frac{\omega}{\omega_{\rm orb}}=\frac{1}{3}\frac{q}{1+q}[0.75+0.325(1+q)^{-2/3}]- 3^{1/3} \eta_A (1+q)^{2/3},
\end{split}
\label{eq:mia}
\end{equation}
where $\omega$ is now the apsidal precession frequency of the disk taking into account the pressure term.
\cite{Hu_08}, assuming a 3:1 resonance, inferred a value of $q \simeq 0.045$ based on the negative superhump with a period of 2979.3 s observed by
\cite{Retter_02}, and interpreted as the beat period between the orbital period and the 4.86 d nodal precession period of the disk. The value of $q \simeq 0.045$ was obtained from the
expression proposed by \cite{Larwood_96} and \cite{Montgomery_09}
\begin{equation}
\begin{split}
\frac{\omega_n}{\omega_{\rm k}} = -\frac{15}{32}qr^3\cos\delta,
\end{split}
\label{eq:larwood}
\end{equation}
where $\omega_n$ is the angular frequency associated with
the nodal precession of the disk, $\omega_k$ is the Keplerian frequency of the matter at a given radius $r$ of the accretion disk
that is defined in units of orbital separation $a$, and the angle $\delta$ describes the orbit of the CS with respect to the NS. For $\delta \simeq 0$ the orbits of the two bodies are coplanar \citep{papa_95}.
The minus sign is present because the nodal frequency is retrograde with respect to the orbital frequency \citep[see Fig. 6.18 in][]{Hellier_01}.
Rewriting the Keplerian frequency in the known form
$\omega_k=(GM_1/R^{3})^{1/2}$ and combining it with Kepler's third law $a^3=GM_1(1+q)/\omega_{orb}^2$ we find
$\omega_k=\omega_{orb}(1+q)^{-1/2}r^{-3/2}$. We combined the last expression with eq. \ref{eq:larwood}, obtaining
\begin{equation}
\begin{split}
\frac{\omega_n}{\omega_{\rm orb}} = -\frac{15}{32}q(1+q)^{-1/2}r^{3/2}\cos\delta.
\end{split}
\label{eq:larwood2}
\end{equation}
Assuming that $r=r_{disk}/a$ \citep[the same assumption made by][]{Hu_08} and combining the last equation with eq. \ref{eq:rdisk} for a 3:1 resonance, we infer
\begin{equation}
\begin{split}
\frac{\omega_n}{\omega_{\rm orb}} = -\frac{5}{32}\frac{q}{1+q}\cos\delta.
\end{split}
\label{eq:larwood3}
\end{equation}
Because $\omega_n/\omega_{orb} \simeq 7.146 \times 10^{-3}$, we adopt the nodal period of 4.86 days and find that $q\simeq 0.048$ for $\cos\delta\simeq1$.
We show in Fig. \ref{figure:inconsisyency}
\begin{figure}
\includegraphics[scale=.50]{fig6.eps}
\caption{Characteristic radii of XB 1916-053 in units of orbital separation vs the mass ratio q. Show are the NS Roche-lobe radius (blue curve), the radius of the disk at which a 2:1 and 3:1 resonance could occur (yellow and black curves), the radius associated with the apsidal rate of the disk for a 3:1 resonance without and with the pressure term due to the presence of a spiral wave in the disk (green and brown curve), and the radius associated with the nodal precession period for a 3:1 resonance (red curve).}
\label{figure:inconsisyency}
\end{figure}
the Roche-lobe radius of the NS as a function of $q$ (blue curve),
the outer radius of the accretion disk for a 2:1 resonance (yellow curve) and for a 3:1 resonance (black curve).
The green curve represents the outer radius of the disk
adopting the value of $\omega/\omega_{\rm orb} \simeq 8.9 \times 10^{-3}$
obtained by \cite{Chou_01}, who neglect the pressure term due to the spiral wave in the apsidal precession rate (see eq. \ref{eq:ompomorb}).
The red curve represents the outer radius of the disk assuming a nodal precession of 4.86 days obtained from eq. \ref{eq:larwood3}. All the radii are in units of orbital separation $a$.
Initially, we note that the green and red curve do not have intersections that exclude any self-consistent scenario; in other words, the outer radius of the disk estimated from the observed apsidal precession period is not compatible with that obtained from the observed nodal precession period if we ignore the pressure term. Taking into account $\omega_{\rm press}$, the outer radius of
the disk obtained from the apsidal precession period can intersect the red curve. It is possible to obtain a pair of $r$ and $q$ values for a 3:1 resonance assuming that the parameter $\eta_A\simeq 0.0049$. For that value we find
$q \simeq 0.048$ and an outer disk radius $r\simeq 0.474$ in units of orbital separation. It could be possible
a solution for a 2:1 resonance for which $q \simeq 0.0315$, $r\simeq 0.62$, and $\eta_A\simeq 0.0089$; however, since the truncation radius of the disk due to the tidal interaction with the CS is $r_T=R_T/a=0.6/(1+q)$ \citep[for $q$ between 0.03 and 1;][]{Pac_77}, we obtain $r>r_T$, making the solution unrealistic.
A further indication that the mass ratio $q$ could be close to 0.048 is given by the empirical expressions inferred for the CVs, in which the period excess $\epsilon$ depends on $q$.
Using the values of the superhump period $P_{\rm sh}=3027.5510(52)$ s and the orbital period $P_{\rm orb}=3000.6508(9)$ \citep[see Table 1 in][and references therein]{Retter_02} we can define the period excess as
\begin{equation}
\begin{split}
\epsilon = \frac{P_{\rm sh}-P_{\rm orb} }{P_{\rm orb}}=(8.965\pm0.002) \times 10^{-3}.
\end{split}
\label{eq:mia}
\end{equation}
Using the empirical form $\epsilon = 0.18q+0.29^2$ \citep{Patterson_05} we obtain $q\simeq 0.045$. \cite{Goodchild_06} inferred $\epsilon =0.2076(3)q -4.1(6) \times 10^{-4} $, where the small offset from the origin is due to pressure-induced retrograde precession of a stable eccentric mode in systems with very low $q$. Using the last relation we infer that $q = 0.0452 \pm 0.0004$.
We obtain $q \simeq 0.048$ for
\begin{equation}
\begin{split}
\eta_A= \frac{1}{2} \left(\frac{c_s}{\omega_{\rm orb}\; a}\right)^2 \tan^{-2}\theta=0.0049.
\end{split}
\label{eq:eta}
\end{equation}
\cite{Lubow_92} gave a range of 0.01-0.05 for the normalized speed of sound,
$ c_s/(\omega_{\rm orb}\; a),$ and inferred that for a tightly wrapped spiral arm
the pitch angle $\theta$ is between 5.7$^{\circ}$ and 31$^{\circ}$. We show in Fig. \ref{figure:eta} that for $\eta_A\simeq0.0049$ this condition
\begin{figure}
\includegraphics[scale=.50]{fig7.eps}
\caption{Pitch angle of the spiral wave vs the normalized sound of speed (black curve). The red horizontal lines correspond to the angles of pitch between 13$^{\circ}$ and 21$^{\circ}$, as obtained by \cite{Montgomery_01} from simulations.}
\label{figure:eta}
\end{figure}
is verified, and we obtain that the pitch angle is between
5.8$^{\circ}$ and 26.7$^{\circ}$. From simulations \cite{Montgomery_01} restricted the range of the pitch angle
between 13$^{\circ}$ and 21$^{\circ}$; these values restrict the range
of the normalized sound of speed between 0.023 and 0.038.
Using Kepler's third law and assuming $q=0.048$ we obtain that the sound speed $c_s$ is between
$15.2 m_1^{1/3}$ km s$^{-1}$ and $25.2 m_1^{1/3}$ km s$^{-1}$. This range of values is compatible
with a disk temperature in the outer region between $3\times10^4$ K and $7\times10^4$ K \citep[see eq. 2.21 in][]{Frank_02}, which is compatible with the results obtained by \cite{Nelemans_06} from the analysis of the optical band of XB 1916-053. The authors found that
a LTE model consisting of pure helium plus overabundant nitrogen closely fits the observed spectrum finding that the model has a temperature of $\sim 3\times10^4$ K.
\subsection{Neutron star mass $m_1$}
Assuming that the CS fills its Roche lobe, we can obtain an indication of the NS mass. The CS is a degenerate star and its radius $R_2$
is given by $R_2/R_{\odot} = 0.0126 (1+X)^{5/3} m_2^{-1/3}$,
where $X$ is the fraction of hydrogen in the star. This equation has to be corrected for the thermal bloating factor $b_f\ge1$, which is the ratio of the CS radius to the radius of a star with the same mass and composition; this star is completely degenerate and supported only by the Fermi pressure of the electrons.
Adopting
the Roche-lobe prescription of \cite{Eggleton_83},
using Kepler's third law and imposing that $R_{\rm l2}= b_f R_2$, where $R_{\rm l2}$ is the Roche-lobe radius of the CS, we obtain
\begin{equation}
\begin{split}
m_1
=
14.17 \; q^{-1/2}(1+q)^{-1/2} \left[\frac{0.49\;q^{2/3}}{0.6 q^{2/3}+\ln(1+q^{1/3})} \right]^{-3/2}\\P^{-1} (1+X)^{5/2} b_f^{3/2},
\end{split}
\label{eq:m1}
\end{equation}
where $P$ is the orbital period in seconds.
\cite{Heinke_13} inferred $X \sim 0.14$ for a NS mass of 1.4 M$_{\odot}$, compatible with the measurements of \cite{Nelemans_06}.
We can predict how $X$ varies with respect to $m_1$.
Even though the CS is almost a pure helium dwarf, the small percentage of hydrogen that can be transferred onto the NS via Roche-lobe accretion can significantly change the energy released due to hydrogen’s larger energy release per nucleon \citep[energy released per nucleon $Q_{\rm nuc} = 1.6+4.0 X$ MeV nucleon$^{-1}$;][]{Cumming_03}. \cite{Galloway_08}, using {\it RXTE} data, measured the ratio $\alpha$ of burst to persistent flux
between two consecutive type I X-ray bursts temporally spaced by 6.3 hours obtaining $\alpha=78.8\pm0.3$. We rewrite eq. 6 shown
by \cite{Galloway_08} as
\begin{equation}
\begin{split}
\alpha
= 34.57\; m_1 (0.4+X)^{-1},
\end{split}
\label{eq:burst}
\end{equation}
in which we assumed a NS radius of 10 km. Adopting $\alpha=78.8\pm0.3$ we find that $X$ is 0.21, 0.3, and 0.39 for a NS mass of 1.4, 1.6, and 1.8
M$_{\odot}$ respectively, while for a NS mass of 2.1 M$_{\odot}$ we find that
$X \simeq 0.52$, in contrast with the pure helium dwarf nature of the CS. Furthermore, assuming an orbital period value $P=3000.66$ s and a mass ratio $q=0.048$, and combining
Eqs. \ref{eq:m1} and \ref{eq:burst}, we find
\begin{equation}
\begin{split}
b_f
= m_1^{2/3} (0.376+0.275 m_1)^{-5/3}.
\end{split}
\label{eq:bf}
\end{equation}
We obtain that $b_f$ is 1.97, 1.92, and 1.86 for a NS mass of 1.4, 1.6, and 1.8 M$_{\odot}$, respectively.
In the following we explore the NS mass between 1.4 and 1.8 M$_{\odot}$ in order to have a fraction of hydrogen $X$ in the CS lower than 0.4.
Finally, we can express $m_2$ from eq. \ref{eq:m1} remembering
that $m_1=m_2q^{-1}$. We find
\begin{equation}
\begin{split}
m_2
=
41.31 \left(\frac{q}{1+q}\right)^{1/2} [0.6+q^{-2/3}\ln(1+q^{1/3})]^{3/2}\\P^{-1} (1+X)^{5/2} b_f^{3/2}.
\end{split}
\label{eq:m2}
\end{equation}
\begin{figure*}
\centering
\includegraphics[scale=.34]{fig8a.eps}
\includegraphics[scale=.34]{fig8b.eps}
\includegraphics[scale=.34]{fig8c.eps}\\
\caption{Constraint of the matter-ejection point position and $\beta$ for $q=0.048$. The left, middle, and right panels describe the constraints for $m_1$ values 1.4, 1.6, and 1.8 M$_{\odot}$. The red curve
represents the orbital period derivative of
$1.46 \times 10^{-11}$ s s$^{-1}$ for the pairs $\beta$-$d/a$. The black and blue horizontal lines represent the minimum and maximum luminosity of the source during the observations. The minimum and maximum value of $d/a$ correspond to the inner Lagrangian point and the CS position with respect to the center of mass. The observed luminosities suggest that the mass transfer is highly non-conservative and that the ejected matter leaves the system close to the inner Lagrangian point.}
\label{figure:9}
\end{figure*}
\subsection{Orbital period derivative and luminosity of XB 1916-053}
Since XB 1916-053 has an orbital period close to 50 min we assume that the angular momentum lost via magnetic braking
is negligible with respect to the angular momentum lost via gravitational waves. Combining eqs. 1, 2, and 3 shown by
\cite{Rappaport_87} with the assumption that the stellar adiabatic index $\xi_{\rm ad}$ is -1/3, as suggested by the authors, we obtain
\begin{equation}
\begin{split}
-\frac{\dot{m_2}}{m_2} = \frac{6.046\times 10^{-16}}{\Lambda(q,\alpha,\beta)} m_1^{5/3} q (1+q)^{-1/3}P_h^{-8/3} {\rm \; s^{-1}},
\end{split}
\label{eq:m2dotonm2_1}
\end{equation}
where $P_h$ is the orbital period in hours,
$\beta$ is the fraction of the mass that transferred
from the CS accretes onto the NS and $\alpha$ is linked to
the rate of specific angular momentum $\dot{J}_{\rm mass}$ lost via mass loss by the binary system. Defining
$\dot{J}_{\rm mass}=2\pi\alpha(1-\beta)\dot{M}_2a^2/P$,
where $a$ is the orbital separation and $P$ the orbital period, we can define $\alpha$ as the ratio $(d/a)^2$,
where $d$ is the distance from the center of mass (CM) of the binary system from which the matter leaves.
Finally, the function $\Lambda(q,\alpha,\beta)$ is defined as
\begin{equation}
\begin{split}
\Lambda(q,\alpha,\beta)
= 1-\frac{3}{2}\beta q - \frac{1-\beta}{2} \frac{q}{1+q} -1.5\; \alpha (1-\beta)(1+q).
\end{split}
\end{equation}
\cite{iaria_15} assumed that a large fraction of the
mass transferred by the CS leaves the system at the inner Lagrangian point.
Since the masses of the two bodies are unknown, we
estimate the ratio $b_1/a$ as a function of the mass ratio $q$ between 0.005 and 0.1,
where $b_1$ is the distance of the inner Lagrangian point from the NS. We note that
\cite{Plavec_64} found an expression of $b_1/a$ for values of $q$ higher than 0.1 and we verified {a posteriori} that
it does not work for the range of $q$ of interest to us. We derived a more generic expression,
\begin{equation}
\frac{b_1}{a}=0.702-0.948\;q+ 2.77\;q^2-0.0825 \log(q),
\label{eq:L1}
\end{equation}
imposing that the sum of the gravitational forces of the two bodies and the centrifugal force is null at the inner Lagrangian point (L$_1$); the accuracy of this relation is 0.5\%
for $q$ between 0.005 and 0.1. As reported above the free parameter $\alpha$ is defined as $(d/a)^2$, where $d$ indicates the distance to the CM from where the binary system loses specific angular momentum due to mass loss. The minimum value of $d/a$
corresponds to the ejection point coinciding with the inner Lagrangian point (i.e.,
$d/a=b_1/a-q/(1+q)$, which gives 0.726 for $q=0.048$), while it is reasonable to assume that the maximum value of $d/a$ corresponds to the ejection point coinciding with the CS position (i.e., $d/a=1-q/(1+q)$, which gives 0.954 for $q=0.048$). Consequently, in the following we explore the range of $d/a$ between 0.726 and 0.954.
Since we expect $q<0.1$ we can approximate the term $q/(1+q)$ as $\simeq q$ in eq. \ref{eq:m2}. By differentiating eq. \ref{eq:m2} and some algebraic manipulation
we obtain
\begin{equation}
\begin{split}
\frac{\dot{m}_2}{m_2}
=
-\frac{\dot{P}}{P} +\frac{3}{2}\frac{\dot{b_f}}{b_f}+\\ +\frac{1}{2} \frac{0.6q^{2/3}+q^{1/3}(1+q^{1/3})^{-1}-\ln(1+q^{1/3})}{0.6q^{2/3}+\ln(1+q^{1/3})} (1+\beta q) \frac{\dot{m}_2}{m_2}.
\end{split}
\label{eq:m2dotonm2}
\end{equation}
As suggested by \cite{Rappaport_87}, the term $\dot{b}_f/b_f$ is much smaller than $\dot{m}_2/m_2$; neglecting
that term we can rewrite eq. \ref{eq:m2dotonm2} in the compact form
\begin{equation}
\begin{split}
\frac{\dot{m}_2}{m_2}
=
-\frac{\dot{P}}{P} \left[ 1-\frac{1}{2} \Gamma(\beta,q) \right]^{-1},
\end{split}
\label{eq:m2_seconda}
\end{equation}
where
\begin{equation*}
\Gamma(\beta,q)=
\frac{0.6q^{2/3}+q^{1/3}(1+q^{1/3})^{-1}-\ln(1+q^{1/3})}{0.6q^{2/3}+\ln(1+q^{1/3})} (1+\beta q).
\end{equation*}
Combining eqs. \ref{eq:m2dotonm2_1} and \ref{eq:m2_seconda}
we obtain
\begin{equation}
\begin{split}
\dot{P}_{-11}=0.218 \frac{1-\frac{1}{2} \Gamma(\beta,q)}{\Lambda(q,\alpha,\beta)} m_1^{5/3} q (1+q)^{-1/3}P_h^{-5/3} {\rm \; s \;s^{-1}},
\end{split}
\label{eq:pdot}
\end{equation}
where $\alpha$ only depends on $q$ and
$\dot{P}_{-11}$ is the orbital period derivative in units
of $10^{-11}$.
On the other hand, the source luminosity
can be written as $L_x = -\beta (G M_1 \dot{M}_2)/R_{NS}$,
where $R_{NS}$ is the NS radius. This relation can be opportunely rewritten as
$L_x \simeq 2.64\times10^{53} \beta m_1^2 q (\dot{m}_2/m_2)$ erg s$^{-1} $ for a NS radius of 10 km; combined with eq. \ref{eq:m2_seconda} it becomes
\begin{equation}
\begin{split}
L_{37}=73.33 \; \beta \; m_1^2 q \left[ 1-\frac{1}{2} \Gamma(\beta,q) \right]^{-1} \frac{\dot{P}_{-11}}{P_h} {\rm \; erg \;s^{-1}},
\end{split}
\label{eq:lumin}
\end{equation}
where $L_{37}$ is the luminosity in units of $10^{37}$.
\begin{figure*}
\centering
\includegraphics[scale=.344]{fig9a.eps}
\includegraphics[scale=.376]{fig9b.eps}\\
\caption{ Top view in scale of XB 1916-053 for a mass ratio $q=0.048$. The left and right panel shows the whole binary system and a zoom of it close to the inner Lagrangian point, respectively. The blue curve indicates the Roche lobes of the two stars. The dashed and red circle represent the NS Roche-lobe radius and CS Roche-lobe radius. The magenta, black, and green diamond points are the position of the center of mass (CM), the companion star, and the inner Lagrangian point $L_1$. The dotted circle is the radius of the companion star, without thermal inflation. The green circle represents the accretion disk truncated at $0.47 a$ (see text). The two black dashed segments indicate the distances from $L_1$ at which the matter leaves the binary system for a NS mass of 1.4 M$_{\odot}$ at a luminosity of $0.41 \times 10^{37}$ erg s$^{-1}$ and $1.46 \times 10^{37}$ erg s$^{-1}$ going from the closest to the farthest.}
\label{figure:9b}
\end{figure*}
From the study of the dip arrival times
we infer an orbital period derivative
$\dot{P} = (1.46 \pm 0.03) \times 10^{-11}$ s s$^{-1}$ and an
orbital period $P=3000.662(1)$ s. Studying eq. \ref{eq:pdot} for $q=0.048$,
we explored the pairs of
$\beta$ and $d/a$ for which it is possible to obtain the $\dot{P}$ value assuming
NS masses of 1.4, 1.6, and 1.8 M$_{\odot}$.
To explain a $\dot{P}$ value of $(1.46 \pm 0.03) \times 10^{-11}$ s s$^{-1}$,
a non-conservative mass transfer is required; we obtain that the parameter $\beta$ has to be smaller
than 0.35 for each value of the NS mass studied.
We find that the closer the ejection point is to the Lagrangian point, the smaller the fraction of accreted matter is.
The possible pairs of $\beta$ and $d/a$ are indicated with the red curve in Fig. \ref{figure:9}.
A further constraint on the ejection point and on the fraction of matter accreting onto the NS
is given by the luminosity determined in eq. \ref{eq:lumin}.
The distance of the source can be estimated using the flux during
the type I X-ray bursts observed using {\it Rossi-XTE} data, as done by \cite{Galloway_08}.
Using eq. 10 in \cite{iaria_15}, given that the flux measured during
the bursts showing photospheric radius expansion (PRE)
is $(2.9 \pm 0.4) \times 10^{-8}$ erg s$^{-1}$ cm$^{-2}$ and
that the NS photospheric radius is $r_{PRE}=1.1$ in units of 10 km for XB 1916-053, we find
that the distance $D$ to the source is
$8.9\pm0.6$ kpc, $9.3\pm0.6$ kpc, and $9.6\pm0.7$ kpc for a NS mass of 1.4, 1.6, and 1.8 M$_{\odot}$, respectively.
Adopting the 0.1-100 keV unabsorbed flux obtained from spectra A, B, C, and D,
we infer that the luminosity varies between $0.41 \times 10^{37}$ and $1.46 \times 10^{37}$ erg s$^{-1}$ for a NS mass of 1.4 M$_{\odot}$, between
$0.45 \times 10^{37}$ and $1.6 \times 10^{37}$ erg s$^{-1}$ for a NS mass of 1.6 M$_{\odot}$, and between
$0.48 \times 10^{37}$ and $1.7 \times 10^{37}$ erg s$^{-1}$ for a NS mass of 1.8 M$_{\odot}$ by taking into account the error associated with the distance and considering a 10\% relative error for the flux. A similar ranges of luminosities
were also observed by \cite{Boirin_00}, which confirmed that the source belongs to the Atoll class.
The 0.1-100 keV unabsorbed luminosities with the corresponding uncertainties are shown in Table \ref{tab:diagnostic_abcd}
for spectra A, B, C, and D.
We plotted the boundary values of luminosity for $q=0.048$ in Fig. \ref{figure:9}; since the luminosity depends on $\beta$ but not on the quantity $\alpha=(d/a)^2$ (see eq. \ref{eq:lumin}), it is represented with horizontal lines.
The observed luminosities constrain $\beta$
to be lower than 15\%, independently of the assumed NS mass. The variation in observed luminosity can be explained
with a change of the ejection point from which the matter leaves the binary system.
At low luminosity the matter leaves the system close to the inner Lagrangian point, and the fraction of mass accretion rate is
only 3\%, while at higher luminosity the matter leaves the system from a larger distance and $\beta=0.11$.
We show a sketch of the top-view geometry (in scale) of XB 1916-053 assuming
$q=0.048$ and a NS mass of 1.4 M$_{\odot}$ in Fig. \ref{figure:9b}. The left panel shows the whole binary system, while the right panel shows the companion star and the region close the inner Lagrangian point.
The two black dashed segments indicate the boundaries of the ejection region measured from $L_1$, the segment distant 0.065$a$ from $L_1$. The lower boundary corresponds to a luminosity of $0.41 \times 10^{37}$ erg s$^{-1}$, while the upper boundary (0.101$a$) corresponds to a luminosity of $1.46 \times 10^{37}$ erg s$^{-1}$. Similar results are obtained assuming a NS mass of 1.6 and 1.8 M$_{\odot}$.
Hence, we have shown that abandoning the stringent hypothesis that the matter could leave the system only from the inner Lagrangian point, as supposed by \cite{iaria_15}, a NS mass of 1.4 M$_{\odot}$ also describes the observed scenario well.
We note that this scenario
assumes that the observed luminosity $L_{\rm obs}$ is that emitted from the source $L_{\rm real}$; however, XB 1916-053 is a dipping source, and so we expect an inclination angle larger than $60^{\circ}$. Since we do not observe eclipses, we can estimate the upper limit on the inclination angle $i$ from the relation $i=90^{\circ}-\arctan(R_{L_2}/a)$, where $R_{L_2}$ is the Roche lobe radius of the CS, finding that $i<80^{\circ}$. If the
emission is not spherical from the innermost region of the system then $L_{\rm obs} = L_{\rm real}\cos{i}$ and we should take into account this amplification factor. By
assuming the arbitrary value of $i=70^{\circ}$, the real luminosity varies between $1.2\times 10^{37}$ erg s$^{-1}$ and $4.3 \times 10^{37}$
erg s$^{-1}$. These two values of luminosity combined with the necessity to reproduce
an orbital period derivative of $1.46 \times 10^{-11}$ s s$^{-1}$ gives a value of $\beta$ between 0.08 and 0.34. In this scenario the matter leaves the system
at a distance from $L_1$ close the CS position.
We note that the mass transfer rate is close to the Eddington limit in this evolutive stage of the source. However, the secular evolution predicts that this system should have a mass transfer rate that is two orders of magnitude lower \cite[see, e.g.,][]{Heinke_13} for a conservative mass transfer scenario.
\begin{table}
\centering
\scriptsize
\begin{threeparttable}
\caption{Parameters of XB 1916-053}
\begin{tabular}{l@{\hspace{2pt}}c@{\hspace{\tabcolsep}}}
\hline
\hline
Parameter & \\
\hline
\hline
Mass ratio $q$ & 0.048 \\\\
Orbital period $P$ (s) & 3000.66 \\
Orb. period derivative $\dot{P}$ (s s$^{-1}$) & $(1.46 \pm 0.03) \times 10^{-11}$\\\\
Orbital separation $a${\textit{$\rm ^a$}} (cm) & $3.54 \times 10^{10}$\\
Roche lobe radius of the CS, $R_{L_2}$ & $0.156a$\\
Roche lobe radius of the NS, $R_{L_1}$ & $0.594a$\\
Outer radius of the accretion disc, $r$ & $0.47a$ \\
Distance of the CM from the NS & $0.046a$\\
Distance of $L_1$ from the NS, $b_1$ & $0.772a$ \\
Inclination angle $i$ & $60^{\circ}<i<80^{\circ}$\\
Bloating thermal factor $b_f$ of the CS{\textit{$\rm ^a$}} & 1.92 \\
Observed luminosity{\textit{$\rm ^a$}} ($10^{37}$ erg s$^{-1}$) & 0.41-1.46\\
$\beta${\textit{$\rm ^b$}} & 0.03-0.11\\
Boundaries of the ejection region{\textit{$\rm ^c$}}&0.065$a$-0.101$a$ \\
\hline
\hline
\end{tabular}
\begin{tablenotes}
\item[a] Assuming a NS mass of 1.4 M$_{\odot}$.
\item[b] Fraction of the mass transferred
from the CS that accretes onto the NS.
\item[c] The boundaries are measured from the inner Lagrangian point $L_1$.
\end{tablenotes}
\label{tab:resume}
\end{threeparttable}
\end{table}
Finally, \cite{Lasota_08} discussed the stability of the helium accretion disk in ultracompact X-ray binary systems. From their estimations, the authors suggested that XB 1916-053 should be transient; however, the source is observed as persistent. The authors
showed three possible explanations of that incongruity: i) the adopted mass accretion rates (bolometric luminosities) could be underestimated, ii) the outer disk radius is overestimated, iii) the CS is not a pure-helium star, but still contains some hydrogen. From our analysis all the caveats highlighted by the authors are verified. \cite{Lasota_08} assumed a mass accretion rate of $7.6 \times 10^{-10}$ M$_{\odot}$ yr$^{-1}$, while we find from the estimated bolometric fluxes that
the mass accretion rate is between $3 \times 10^{-10}$ and $1.2 \times 10^{-9}$ M$_{\odot}$ yr$^{-1}$ for a NS mass of 1.4 M$_{\odot}$ and NS radius of 10 km. We show that the outer radius of the accretion disk is not truncated at $r_T=R_T/a=0.6/(1+q),$ which corresponds to $R_T=0.57\;a$ for $q=0.048$; it is smaller because of the 3:1 resonance, and we find
$r = 0.47\; a$. Finally, the fraction of hydrogen is close to 0.2 for a NS mass of 1.4 M$_{\odot}$, and increases for higher values of NS mass. We summarized the main parameters of the components of XB 1916-053 in Table \ref{tab:resume}.
In the following we arbitrarily assume a NS mass of 1.4 M$_{\odot}$. Using the mass ratio $q=0.048$, the CS mass
is 0.07 M$_{\odot}$, the bloating thermal factor of the CS is $b_f=1.92$, and the fraction of hydrogen of the CS is close to 20\%.
We find that the orbital separation of the binary system is $a\simeq 3.54 \times 10^{10}$ cm, while the outer radius of the accretion disk is $r_{\rm disk}= 1.7 \times 10^{10}$ cm.
\cite{iaria_15} suggest that the sinusoidal modulation obtained from the LQS ephemeris could due to be the presence of a third body gravitationally bound to the X-ray binary system. Assuming the existence of a third body of mass $M_3$, the binary system orbits around the new CM of the triple system.
The distance of the binary system from the new CM is given by $a_x = a_{\rm{bin}} \sin i = A\; c$, where $i$ is the inclination of the binary plane to the plane of the sky, $A$ is the amplitude of the sinusoidal function obtained from the ephemeris shown in Eq. \ref{linear_quad_sin_eph}, and $c$ is the speed of the light.
We obtain $a_x = (3.9\pm0.4) \times 10^{12}$ cm. To infer the mass of the third body we can write its mass function
as
\begin{equation*}
\frac{M_3 \sin i}{(M_3+M_{\rm{bin}})^{2/3}}=\left(\frac{4\pi^2}{G}\right)^{1/3} \frac{a_x}{P_{\rm{mod}}^{2/3}},
\end{equation*}
where $M_{\rm{bin}}$ is the mass of the binary system. Assuming arbitrarily $i = 70^{\circ}$, we found M$_3 \simeq 45$ M$_J$ under the hypothesis that the orbit of the third body is coplanar to the binary system.
\begin{figure*}
\centering
\includegraphics[scale=.51]{fig10a.eps}
\includegraphics[scale=.51]{fig10b.eps}
\caption{Distance $r$ of the absorber, in units of $10^9$ cm, from the NS with respect to its thickness $\Delta r$ (in units of $10^9$ cm). The gray area indicates the possible pairs of $r$ and $\Delta r$.
The horizontal dashed line indicates the best-fit value of $r$ obtained assuming that the observed redshift is due to gravitational redshift for a NS mass of 1.4 M$_{\odot}$; the green area takes into account the associated errors. The black, red, and blue lines are the constraint conditions (see text).}
\label{figure:10}
\end{figure*}
\subsection{Ionized absorber}
From the spectroscopic analysis we
investigated the nature of the ionized absorber around the compact object.
\begin{table}
\centering
\begin{threeparttable}
\caption{Plasma diagnostic of the spectra A, B, C, and D.}
\scriptsize
\begin{tabular}{l@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}c@{\hspace{\tabcolsep}}}
\hline
\hline
& Spectrum A & Spectrum B& Spectrum C & Spectrum D \\
\hline
N$_{\ion{Ne}{x}}${\textit{$\rm ^a$}} &
$0.39 \pm 0.09$ &
$0.37 \pm 0.11$ &
$0.31 \pm 0.11$ &
-- \\
N$_{\ion{Mg}{xii}}$ &
$0.28 \pm 0.04$ &
$0.24 \pm 0.04$&
$0.26\pm0.04$&
$0.24\pm0.11$\\
N$_{\ion{Si}{xiv}}$ &
$0.70 \pm 0.03$ &
$0.59 \pm 0.04$&
$0.72\pm0.07$&
$1.5\pm0.4$\\
N$_{\ion{S}{xvi}}$ &
$0.74 \pm 0.15$ &
$0.61 \pm 0.15$&
$1.0\pm0.2$&
--\\
N$_{\ion{Ca}{xx}}$ &
$1.0 \pm 0.3$ &
--&
$0.8\pm0.3$&
--\\
N$_{\ion{Fe}{xxv}}$ &
$1.5 \pm 0.3$ &
--&
--&
--\\
N$_{\ion{Fe}{xxvi}}$ &
$7.2 \pm 1.3$ &
$9.6 \pm 1.1$&
$8.5\pm1.5$&
--\\\\
L$_x${\textit{$\rm ^b$}} &
$1.5\pm0.2$ &
$1.0 \pm0.2$ &
$0.67\pm0.11$ &
$0.41\pm0.06$ \\\\
[N$_{\ion{Ne}{x}}$]/[N$_{\rm Ne}$]{\textit{$\rm ^c$}} &
$0.24\pm0.08$ &
$0.31\pm0.10$ &
$0.23\pm0.09$ &
-- \\
[N$_{\ion{Mg}{xii}}$]/[N$_{\rm Mg}$] &
$0.37\pm0.09$ &
$0.43\pm0.09 $&
$0.41\pm0.10$ &
$0.7\pm0.4$\\
[N$_{\ion{Si}{xiv}}$]/[N$_{\rm Si}$] &
$1.1\pm0.2$ &
$1.3\pm0.2 $&
$1.4\pm0.3$&
$5\pm3$\\
[N$_{\ion{S}{xvi}}$]/[N$_{\rm S}$] &
$3.0\pm0.9$ &
$3.3\pm0.9$&
$4.7\pm1.2$&
--\\
[N$_{\ion{Ca}{xx}}$]/[N$_{\rm Ca}$] &
$24\pm9$ &
--&
$23\pm10$&
--\\
[N$_{\ion{Fe}{xxv}}$]/[N$_{\rm Fe}$] &
$2.5\pm0.9$ &
--&
--&
--\\
[N$_{\ion{Fe}{xxvi}}$]/[N$_{\rm Fe}$] &
$12\pm3$ &
$22\pm4$&
$17\pm4$&
--\\ \\
\hline
\hline
\end{tabular}
\begin{tablenotes}
\item[a] The equivalent column density of the ions
shown are in units of 10$^{17}$ atoms cm$^{-2}$.
\item[b] Unabsorbed extrapolated luminosity in the 0.1-100 keV energy range. The luminosity is in units of $10^{37}$ erg s$^{-1}$ for a NS mass of 1.4 M$_{\odot}$.
\item[c] The ratio is in units of $10^{-2}$.
\end{tablenotes}
\label{tab:diagnostic_abcd}
\end{threeparttable}
\end{table}
Initially we estimated the equivalent column densities
of the ions producing the absorption lines in
spectra A, B, C, and D. To do this we used the relation shown by \cite{Spitzer_78}
\begin{equation*}
\frac{W_\lambda}{\lambda}=\frac{\pi e^2}{m_ec^2} N_j \lambda f_{ij} \simeq 8.85 \;10^{-13} N_j \lambda f_{ij},
\end{equation*}
where $ N_j$ is the equivalent column density for the relevant species,
$f_{ij}$ is the oscillator strength, $W_\lambda$ is the equivalent width of the
line, $\lambda$ is the wavelength in centimeters,
$e$ is the electron charge, $m_e$ is the electron mass, and $c$ is the speed of the light. A similar analysis was done for the dipping source XB 1254-690 \citep{Iaria_07}
and X 1624-690 \citep{Iaria_07_1624} and, recently, for XB 1916-053 using {\it Suzaku} data \citep{Gambino_19}.
To infer the column densities
of the ions we assumed $f_{ij} =0.416$ for all the
H-like ions and $f_{ij} =0.798$ for the \ion{Fe}{xxv} ion
\citep{Verner_96_fij}, and
we used the equivalent widths shown in Table
\ref{tab:line}; the inferred $ N_j$ values are shown in Table
\ref{tab:diagnostic_abcd}.
We find that the equivalent ion column density is in the range between $10^{16}$ and $(7.2-8.5) \times 10^{17}$ cm$^{-2}$.
We estimated the abundances of each ion with respect to the neutral element $N_{\rm ion}/N_{\rm el}$. To this end, we used the values of the neutral hydrogen equivalent column density of the ionized absorber, N$_{{\rm H}_{\xi}}$, shown in Table \ref{tab:mod1mod2} and we adopted the relation
$N_{{\rm H}_{\xi}}= N_{{\rm H}_{\xi}}/N_{\rm el}\; N_{\rm el}/N_{\rm ion} \;N_{\rm ion} $, where $N_{{\rm H}_{\xi}}/N_{\rm el}$ is the inverse of a given element with respect to the hydrogen and $N_{\rm ion}$ is the value
of the equivalent column density associated
with the ions shown in Table \ref{tab:diagnostic_abcd}. To infer the value of $N_{\rm ion}/N_{\rm el}$ we adopted the
solar abundance shown by \cite{asplund_09}. The values of $N_{\rm ion}/N_{\rm el}$ for spectra A, B, C, and D are shown in Table
\ref{tab:diagnostic_abcd}. We find that the ratio $N_{\rm ion}/N_{\rm el}$ is
0.2\%, 0.4\%, 1-2\%, 3-4\%, $(24\pm10)$\%, 2.5\%, and 10-20\% for \ion{Ne}{x}, \ion{Mg}{xii}, \ion{Si}{xiv}, \ion{S}{xvi}, \ion{Ca}{xx}, \ion{Fe}{xxv,} and \ion{Fe}{xxvi}, respectively. In a scenario where the NS mass is 1.4 M$_{\odot}$, the fraction of hydrogen of the CS is close to 20\%; because $X \simeq 0.7$ for solar abundance, we should roughly multiply the adopted abundances by a factor
of three. Consequently, the ion population with respect to a given neutral element should be reduced by a factor of $1/3$.
To estimate the region where the ionized absorber is located,
we used the definition of $\xi$ in {\sc zxipcf} from which we get $r^2=L_x/(n_H\xi)$, where $L_x$ is the unabsorbed luminosity in the 0.1-100 keV energy range, $r$ is the distance of the absorber from the central source, $\xi$ is the ionization parameter obtained by our fits, and $n_H$ is the hydrogen atom density of the absorber. By combining the last equation with N$_{{\rm H}_{\xi}}=n_H\Delta r$, where $\Delta r$ is the geometrical thickness of the absorber along the line of sight and N$_{{\rm H}_{\xi}}$
is the equivalent hydrogen column density associated with the absorber, and by
imposing $r>\Delta r$, we obtained an
upper limit on $r$ that is $r<L_x/(N_{{\rm H}_{\xi}}\xi)$. Furthermore, we expect that the optical depth $\tau=\sigma_T N_{{\rm H}_{\xi}}$
of the absorber is lower than one, where $\sigma_T$ is the Thomson cross section. By rearranging $r^2=L_x/(n_H\xi)$, we obtain $r^2=L_x\sigma_T \Delta r/(\tau \xi),$ and consequently $r>(L_x \sigma_T \Delta r/\xi)^{1/2}$. By using the best-fit values of N$_{{\rm H}_{\xi}}$ and $\xi$ shown in Table \ref{tab:mod1mod2} and the 0.1-100 keV unabsorbed luminosities shown in Table \ref{tab:diagnostic_abcd} we are able to constrain
the region where the ionized absorber is placed.
We show the distance of the absorber from the NS in units of $10^9$ cm with respect to the thickness $\Delta r$ in units of $10^9$ cm in Fig. \ref{figure:10}.
The condition $r<L_x/(N_{{\rm H}_{\xi}}\xi)$ is represented with a blue line, $r>\Delta r$ is indicated with a black line, and $r>(L_x \sigma_T \Delta r/\xi)^{1/2}$ with a red line. The gray area gives the pairs of $r$ and $\Delta r$ satisfying the conditions. Since the outer radius of the accretion disk is $1.7 \times 10^{10}$ cm and because the maximum value of $r$ is roughly $5 \times 10^9$ cm, we conclude that the ionized absorber is placed in the innermost region of the system.
After analyzing the same {\it Chandra} data,
\cite{Trueba_20} suggests that the observed redshift in the absorption lines is probably due to gravitational redshift, and finds the
distance of the absorber from the NS is $(2.7 \pm 1.7) \times 10^8$ cm and $(2.3 \pm 1.2) \times 10^8$ cm
for a NS mass of 1.4 M$_{\odot}$; these values refer to spectra B and $\Gamma$ reported by the authors. By adopting the same scenario and using the best-fit values of the observed redshift shown in Table \ref{tab:mod1mod2} we find that the absorber is placed
at $(1.1 \pm 0.1) \times 10^{8}$ cm,
$(1.3 \pm 0.2) \times 10^{8}$ cm, $(1.2 \pm 0.2) \times 10^{8}$ cm, and $(1.3 \pm 0.7) \times 10^{8}$ cm for spectra A, B, C, and D, respectively. Our estimations are almost a factor of two smaller than the value estimated by \cite{Trueba_20}, even if the values are compatible with each other. We show the best-fit value of $r$ and the associated errors with a dashed line and the green area in Fig. \ref{figure:10}. We find that the ionized absorber can reach the NS surface only in spectrum D when the source is dimmer; in the other cases the absorber forms a torus around the compact object. From N$_{{\rm H}_{\xi}}=n_H\Delta r$ we can infer the lower limit of $n_H$ assuming the maximum possible thickness of the absorber;
we obtain $n_H>5 \times 10^{15}$ cm$^{-3}$, $n_H>2 \times 10^{15}$ cm$^{-3}$, $n_H>1.5 \times 10^{15}$ cm$^{-3}$, and $n_H>0.4 \times 10^{15}$ cm$^{-3}$ for spectra A, B, C, and D.
We repeated the same procedure to estimate the distance of the absorber during the dip using the values in
Table \ref{tab:dip} and knowing that the 0.1-100 keV unabsorbed luminosity is $(0.6 \pm 0.1) \times 10^{37}$ erg s$^{-1}$.
We find a weak constraint for $r$; it
is between $2 \times 10^8$ and $1.7 \times 10^{10}$ cm (the outer radius of the disk). Because of the low
value of the ionization parameter it is possible that the absorber is placed at the outer radius;
however, we cannot exclude inner radii. We note that the circularization radius $r_{\rm circ} = 0.0859\; a\; q^{-0.426}$ \citep[valid for $q$ between 0.05 and 1;][]{Hessman_90} is $r_{\rm circ}\simeq 1.1\times 10^{10}$ cm for $q=0.048$, compatible with
the obtained distances.
Finally, we explored the nature of the \ion{O}{viii} absorption edge observed in the spectra during the persistent emission.
Since its optical depth is $0.16$, assuming that $n_e\sim n_i$, we can write $\tau = N_H \sigma_T$, where $\sigma_T$ is the Thomson cross section. We inferred an equivalent hydrogen column density of $\sim 2.3 \times 10^{23}$ cm$^{-2}$. This value is compatible with the equivalent hydrogen column associated with the ionized absorber.
We conclude that a small fraction of \ion{O}{viii} could be present in the absorber.
\section{Conclusions}
We updated the orbital ephemeris of the compact dipping source
XB 1916-053 by analyzing ten new {\it Chandra} observations and one {\it Swift/XRT} observation.
Three new dip arrival times were extracted, allowing us to extend the available baseline from 36 to 40 years. Our analysis definitively excludes several models of ephemeris suggested previously, leaving as a more likely candidate the quadratic orbital ephemeris including a large periodic modulation close to 25 years (LQS ephemeris in this work).
The quadratic term of the orbital ephemeris implies the
presence of an orbital period derivative of $1.46(3) \times 10^{-11}$ s s$^{-1}$, and allows us to refine the orbital period measurement to 3000.662(1) s. We confirm that such a fast orbital expansion implies that mass transfer should be highly non-conservative.
Moreover, we find that for a NS mass of $1.4$ M$_{\odot}$ it is possible to obtain a compatible orbital period derivative and luminosity values, releasing the stringent hypothesis that the mass leaves the system at the inner Lagrangian point. For a NS mass of $1.4$ M$_{\odot}$ we find that the ejection point is distant from the inner Lagrangian point between $3 \times 10^8$ and $1.5 \times 10^9$ cm.
We show that the mass ratio of the system is $q=0.048$. This value is obtained from the observed infrahump period, as already discussed in the literature by \cite{Hu_08}, and from the observed superhump period, which introduces the pressure term
due to a spiral wave to explain the apsidal precession period of the accretion disk.
Our study allows us to explain the infrahump and superhump periods \citep{Chou_01}, assuming that the accretion disk
has a prograde apsidal precession of 3.9 days and a retrograde nodal precession of 4.86 days. The outer radius of the
accretion disk is truncated where the 3:1 resonance occurs (i.e., $1.7 \times 10^{10}$ cm for a NS mass of 1.4 M$_{\odot}$).
Assuming a NS mass of 1.4 M$_{\odot}$, we obtain that the CS mass is 0.07 M$_{\odot}$. The thermal bloating of the CS is 1.92 and its percentage of content hydrogen is 20\%.
We explain the periodic modulation observed in the orbital
ephemeris as the presence of a third body orbiting around the binary system. Assuming a co-planar orbit with respect to the binary system, an inclination angle of $70^{\circ}$, and a hierarchical triple system, we find that the mass of the third body is $M_3 \simeq 45$ Jovian masses.
To perform our spectroscopic analysis we also included an old Chandra observation, previously analyzed by
\cite{Iaria_06}. We observed a large variation in the count rate for the different observations. Using
a Comptonized component for the continuum we estimated that the 0.1-100 keV unabsorbed luminosity varies from $0.41 \times 10^{37}$ erg $^{-1}$ to $1.46 \times 10^{37}$ erg $^{-1}$ during the persistent emission.
The spectra show the presence of prominent absorption lines associated with \ion{Ne}{x}, \ion{Mg}{xii}, \ion{Si}{xiv}, \ion{S}{xvi,} and \ion{Fe}{xxvi}. To fit these lines
we included in the model the component {\sc zxipcf}
that assumes the presence of an ionized absorber between the central object and the observer along the line of sight.
We find that the equivalent hydrogen column density associated with the ionized absorber ranges from $9 \times 10^{22}$ cm$^{-2}$ to $2 \times 10^{23}$ cm$^{-2}$, going from the dimmest to the brightest spectrum. The ionization parameter is close to log${(\xi)}=4.3$. From our diagnostic, under the assumption that the observed redshift of the line produced in the absorber is due to gravitational redshift, as recently discussed by \cite{Trueba_20},
we find that the absorber has a hydrogen atom density $n_H$ higher than
$10^{15}$ cm$^{-3}$ and it is located far from the central source at $1 \times 10^8$ cm for a NS mass of 1.4 M$_{\odot}$.
We extracted the dip spectrum finding that the
equivalent hydrogen column density associated with the ionized absorber is $6 \times 10^{23}$ cm$^{-2}$ and
the ionization parameter is close to log${(\xi)}=2.8$. Finally,
we find that during the dip the ionized absorber is placed at a greater distance, which could be compatible with the edge of the
accretion disk ($1.7 \times 10^{10}$ cm) even if we cannot exclude that it is placed at the circularization radius ($1.0 \times 10^{10}$ cm).
\section*{Acknowledgements}
This research has made use of data and/or software provided by the High Energy Astrophysics Science Archive Research Center (HEASARC), which is a service of the Astrophysics Science Division at NASA/GSFC and the High Energy Astrophysics Division of the Smithsonian Astrophysical Observatory.\\
The authors acknowledge financial contribution from the agreement
ASI-INAF n.2017-14-H.0, from INAF mainstream (PI: T. Belloni; PI: A. De Rosa)
and from the HERMES project financed by the Italian Space
Agency (ASI) Agreement n. 2016/13 U.O.
RI and TDS acknowledge the research
grant iPeska (PI: Andrea Possenti) funded under the INAF
national call Prin-SKA/CTA approved with the Presidential Decree
70/2016.
\bibliographystyle{aa}
| 703439548c8789fa672699e2d91c190424001784 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Introduction}
Materials crystallizing in the perovskite structure attract attention in many fields due to their high tunability and the resulting, broad spectrum of physical and chemical properties \cite{hwang2017perovskites,zhu2014perovskite,ha2011adaptive}. Most applications of these materials rely on their surfaces and interfaces. The ternary chemical composition of perovskites allows for many different surface terminations \cite{erdman2003surface, kienzle2011vacant} with dramatically different chemical and physical behaviors. It is mostly assumed, however, that perovskites prepared by wet chemical techniques are bulk-terminated (1$\times$1) \cite{kawasaki1994atomic}. Theoretical modeling of catalytic processes \cite{Suntivich2011, Norskov2011} and physical phenomena \cite{santander-syro2011two, meevasana2011creation} generally assume this pristine surface structure. The interfaces between two perovskites preserve the bulk crystal structure and the properties of the junction are strongly influenced by the local defect chemistry. Several emerging applications also profit from the bulk termination, such as employing ferroelectricity for optimizing or promoting catalytic reactions \cite{Kakekhani2015ACS, Kakekhani2016}.
Surprisingly, the available atomic-scale information on bulk-terminated cubic perovskites is quite limited. A prototypical example is SrTiO$_3$: The most common SrTiO$_3$(001) surface preparation method is wet-chemical treatment. Etching with buffered-HF \cite{kawasaki1994atomic} preferentially removes the SrO layer and leaves the surface fully covered with flat, TiO$_2$-terminated terraces \cite{koster1998quasi, koster1998influence}. This method was re-examined several times \cite{deak2006ordering,castell2002scanning,silly2006srtio3,baniecki2008photoemission,herger2007surfacePRL}; as was shown by x-ray photoelectron spectroscopy (XPS) it can induce an unintentional substitution of oxygen with F \cite{chambers2012unintentional}. This can be avoided by etching the surface with HCl--HNO$_3$ \cite{kareev2008atomic} or non-acidic solvents \cite{chambers2012unintentional,connell2012preparation, gerhold2014stoichiometry}. Etched surfaces must undergo annealing to at least 600\,$^\circ$C after introduction into vacuum to remove adsorbates. Up to this temperature, perfectly flat, TiO$_2$-terminated surfaces usually display a (1$\times$1) pattern in low-energy electron diffraction (LEED) or reflection high energy diffraction (RHEED) \cite{castell2002scanning,di2012observation}, while a series of surface reconstructions appear after annealing at higher temperatures \cite{herger2007surfacePRL,herger2007surface,silly2006srtio3,dagdeviren2016surface}. Some reconstructions have been atomically resolved using scanning tunneling microscopy (STM) including the (1$\times$2), (2$\times$2), \textit{c}(4$\times$2), \textit{c}(4$\times$4), and \textit{c}(6$\times$2) \cite{castell2002scanning,erdman2003surface,lanier2007atomic,gerhold2014stoichiometry}, and several more on surfaces that were sputtered and annealed in ultrahigh vacuum (UHV) \cite{deak2006ordering}.
So far, the structural characterization of SrTiO$_3$(001) and other cubic perovskites was mostly based on diffraction techniques. A simple (1$\times$1) diffraction pattern could also stem from the bulk underneath a disordered layer, however; a true proof of a crystalline top surface requires atomically resolved imaging. For perovskite oxides this has been demonstrated only recently \cite{sokolovic2019incipient,setvin2018polarity} using non-contact atomic force microscopy (nc-AFM, AFM hereafter) \cite{morita2015noncontact}. This technique provides clear, atomically resolved images of the (1$\times$1) termination, whereas STM shows no atomic corrugation \cite{sokolovic2019incipient}. It should be noted that the cleaving process that produces such unequivocally crystalline (1$\times$1) terminated SrTiO$_3$(001) surfaces relies on incipient ferroelectricity, and that the induced polarity necessarily results in a sizable density of charged point defects \cite{sokolovic2019incipient}. Such cleaved surfaces can have micron-sized domains with exclusively TiO$_2$ and SrO termination, and each contains 14$\pm$2\% Sr adatoms and Sr vacancies, respectively.
One could expect that thermal annealing heals such intrinsic point defects. Instead, this work shows that the as-cleaved SrTiO$_3$(001)-(1$\times$1) surface is unstable. Raising the temperature results in the lateral migration of the point defects, and an overlayer without long-range order forms above 400\,$^\circ$C. Having established that AFM is capable of providing a clear picture of ordered perovskite surfaces, it was applied to SrTiO$_3$(001) prepared by wet etching. These samples show no signs of an ordered surface in AFM, however, raising doubts that the commonly observed (1$\times$1) diffraction pattern indicates a crystalline, unreconstructed surface.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=1.0\columnwidth,clip=true]{Fig4.pdf}
\end{center}
\caption{
Cut-and-polished TiO$_2$-terminated SrTiO$_3$(001) surface prepared by \textit{ex-situ} wet-chemical treatment. (Top to bottom) LEED, large- and small-area empty-states STM, and AFM images after annealing at (a) 400\,$^\circ$C in UHV, (b) 500\,$^\circ$C and (c) 700\,$^\circ$C in 1$\times$10$^{-6}$\,mbar O$_2$. (d) XPS core-level spectra of the main elements at different annealing conditions, obtained with Al$K\alpha$ line in grazing emission.
}\label{fig4}
\end{figure}
SrTiO$_3$ samples doped with 0.7\,at.\% Nb were used. For cleaving, samples were held in a cleaving device \cite{sokolovic2019incipient} constructed from Mo. After insertion into a UHV chamber with a base pressure below 2$\times$10$^{-10}$ mbar, the device was thoroughly degassed and cooled down to room temperature (RT) prior to cleaving. Annealing was performed \textit{via} a resistive-heating wire below the sample mount. The precision of the reported annealing temperatures is estimated as $\pm$30\,$^\circ$C due to the limited thermal conductivity of the sample mount. Annealing times ranged from 30 to 60 min.
Scanning--probe measurements were performed in UHV with a base pressure below 2$\times$10$^{-11}$~mbar on a ScientaOmicron low-temperature STM/AFM at either $T$=77.7\,K or $T$=4.8\,K, using qPlus cantilevers \cite{giessibl2013sensor} with a separate wire for the tunneling current \cite{Majzik2012} and a differential preamplifier \cite{GiessiblPreamp}. Etched tungsten tips were glued on the tuning fork and cleaned by self-sputtering in Ar atmosphere \cite{setvin2012ultrasharp} prior to the experiment. The resonance frequency of the qPlus cantilevers ranged from 25 to 80\,kHz, with Q factors of $\approx$ 50000. The STM images were obtained by applying a positive bias to the sample. All presented AFM images were acquired with an oxygen-terminated tip \cite{SokolovicPNAS2020}, $i.e.$, surface cations appear attractive (bright) and anions repulsive (dark) \cite{Yurtsever2012}.
Figure\,\hyperref[fig1]{1} shows XPS and LEED results of SrTiO$_3$(001) prepared by the \emph{in-situ} cleaving \cite{sokolovic2019incipient} and after annealing at increasingly higher temperatures for 45 min each, up to 500\,$^\circ$C. In XPS the surfaces are free of contaminants such as carbon. The core-level spectra of the constituents do not change, indicating that the overall surface stoichiometry is preserved. Despite the reducing character of the UHV environment, the elements retain their oxidation state.
The as-cleaved surface exhibits a clear (1$\times$1) LEED pattern. The periodicity does not change various annealing steps. On the other hand, the sharpness of the LEED spots and the background intensity vary: The sharpest diffraction pattern was observed after annealing at 200\,$^\circ$C, while further annealing degraded the pattern somewhat. A distinct (1$\times$1) LEED pattern is also reported throughout literature for polished SrTiO$_3$(001) surfaces annealed at temperatures above 600\,$^\circ$C, necessary for degassing after introduction into UHV. Above 850\,$^\circ$C, spots originating from \textit{c}(4$\times$2) \cite{erdman2002structure,erdman2003srtio3}, or (1$\times$2) \cite{castell2002scanning} reconstructions start to appear.
The surfaces of cleaved samples have SrO and TiO$_2$ domains that span 10 to 100 $\mu$m \cite{sokolovic2019incipient}. This is below the resolution of either of the techniques applied in Fig.\,\hyperref[fig1]{1}. The STM/AFM results in Figs. \hyperref[fig2]{2} and \hyperref[fig3]{3} show the temperature evolution of the two terminations separately.
Figure\,\hyperref[fig2]{2} focuses on the SrO termination. STM images illustrate the large-scale surface morphology, while smaller-size AFM images provide details of the atomic structure. After cleaving at room temperature the surface is atomically flat \cite{sokolovic2019incipient}. The SrO (1$\times$1) surface is covered with the specific concentration of 14$\pm$2\% of Sr vacancies, V$_{\mathrm{Sr}}^{2-}$, apparent as black, `missing' atoms in Fig.\,\hyperref[fig2]{2(b)} \cite{sokolovic2019incipient}. The corresponding STM image shows a corrugation as high as a full unit cell ($\approx$0.4\,nm), which is a purely electronic effect originating from the band bending induced by the charged Sr vacancies.
Annealing at 250$^\circ$C results in the formation of pits at the flat SrO terraces, seen in both, STM and AFM. Their surroundings remain unreconstructed, with fewer V$_{\mathrm{Sr}}^{2-}$ defects. Intrinsic V$_{\mathrm{Sr}}^{2-}$ agglomerate into larger half-unit-cell-deep pits, that expose the underlying TiO$_2$ termination. No oxygen atoms are visible within the pits; they can be considered aggregates of Schottky-type defects \cite{walsh2015self}, formed by Sr vacancy diffusion and the formation of O vacancies \cite{walsh2011strontium}. STM images of the SrO domains annealed to 250\,$^\circ$C and 330\,$^\circ$C show significantly smaller corrugation compared to the as-cleaved surface, consistent with a reduced band bending. Such surfaces also show a LEED pattern [Fig.\,\hyperref[fig1]{1(b)}] with the lowest background intensity.
Further annealing results in the lateral growth of the pits, preferentially along the [100] and [010] directions. The STM images in Fig.\,\hyperref[fig2]{2(a)} appear considerably rougher (up to 3 layers). Annealing the SrO termination at 500\,$^\circ$C results in a loss of the (1$\times$1) ordering in AFM. Instead, images locally show a short-range (2$\times$2) periodicity, while LEED retains the (1$\times$1) symmetry. The surface corrugation deduced from the AFM images is less than half unit cell. STM images exhibit larger apparent height differences, but this can be partially attributed to electronic effects, when domains with different electronic properties form. At this stage of annealing, it is no longer possible to distinguish the previously SrO- and TiO$_2$-terminated areas: The entire cleaved SrTiO$_3$(001) surface shows the same morphology in STM and AFM images as in the rightmost panels of Fig.\,\hyperref[fig2]{2a} and Fig.\,\hyperref[fig2]{2b}, respectively. The (1$\times$1) LEED pattern of this surface is attributed to diffraction from the subsurface layers, while the disordered surface layer results in the increased background intensity.
The evolution of the TiO$_2$ termination with temperature is shown in Fig.\,\hyperref[fig3]{3}. After cleaving, the TiO$_2$ termination hosts Sr$_{\mathrm{ad}}^{2+}$ adatoms [bright dots in Fig.\,\hyperref[fig3]{3(b)}], complementary to the V$_{\mathrm{Sr}}^{2-}$ vacancies at the SrO termination. In STM, the two terminations appear very similar. Annealing above 250\,$^\circ$C results in the formation of small, disconnected islands. The clustering of Sr$_{\mathrm{ad}}^{2+}$ adatoms requires the presence of O$^{2-}$ to compensate their electric charge, and indeed AFM images show the presence of anions [dark dots in Fig.\,\hyperref[fig3]{3(b)}]. These SrO islands show tiny areas with a \textit{c}(2$\times$2) SrO structure on top of the TiO$_2$ termination. The intrinsic excess of Sr adatoms at the as-cleaved TiO$_2$ termination constitute an ideal seed for the crystal growth of the next perovskite SrTiO$_3$ layer, provided the temperature is sufficiently high for the diffusion of the adatoms and of oxygen to complete the SrO stoichiometry.
The \textit{c}(2$\times$2)-like areas spread with increasing temperature up to 430\,$^\circ$C. Islands grow more connected, while still not covering the entire surface. The maximum coverage of this SrO superstructure over the TiO$_2$ termination is limited by the initial 0.14~ML coverage of the Sr adatoms. When arranged in a \textit{c}(2$\times$2) superstructure, it can cover 28\% of the surface area, close to the maximum coverage observed in AFM. After annealing at 500\,$^\circ$C the previous TiO$_2$ termination appears disordered and becomes indistinguishable from what was the SrO termination [rightmost panel of Fig.\,\hyperref[fig2]{2(a)} and Fig.\,\hyperref[fig2]{2(b)}]. Further annealing of the mixed-termination morphology at 600\,$^\circ$C [rightmost panel of Fig.\,\hyperref[fig3]{3(a)}] does not improve the surface roughness, but instead increases the width of the pits and islands.
The two opposite terminations of the as-cleaved SrTiO$_3$(001)-(1$\times$1) surface experience a complementary evolution with annealing. The pit/island creation mechanism is induced by the presence of the intrinsic, polarity--compensating point defects, $i.e.$, Sr vacancies and adatoms \cite{sokolovic2019incipient}. Migration of these charged V$_{\mathrm{Sr}}^{2-}$ and Sr$_{\mathrm{ad}}^{2+}$ point defects is activated at temperatures as low as 200\,$^\circ$C. Moreover, when the surface is additionally enriched by Sr adatoms \textit{via} evaporation, the adatoms are mobile at both terminations and start to aggregate at temperatures as low as 150\,$^{\circ}$C, as shown in the Supplemental Material (SM). The disappearance/appearance of O likely originates from the exchange with the subsurface region, because lateral diffusion across the whole micrometers-wide domains is unlikely \cite{riva2019pushing,riva2019epitaxial}. The temperature range for migration of vacancies is slightly lower than reported in the literature for SrTiO$_3$ \cite{de2015oxygen, riva2018influence}, but can be rationalized by the presence of electric fields related to the charged defects.
Since the intrinsic point defects dominate the thermal behavior of the cleaved crystals, annealing excursions were also conducted on SrTiO$_3$ samples prepared by a wet chemical treatment that exposes only the TiO$_2$ termination. Full details are laid out in the SM. Two cut-and-polished SrTiO$_3$(001) crystals, again with 0.7\,at.\% Nb doping, were cleaned \textit{ex situ} and boiled in ultra-pure water to etch away the soluble SrO termination. One sample was baked in air at 950\,$^\circ$C prior to introduction to UHV (to create large, flat terraces) and turned out to be contaminated with carbon, alkali, and alkaline earth metals, see SM. The second sample was introduced to UHV directly after wet cleaning and was contaminated with carbon alone.
Figure\,\hyperref[fig4]{4} shows the temperature evolution of the latter sample. After annealing to mild temperatures, LEED shows a distinct (1$\times$1) pattern, and large-area STM images show flat terraces. In AFM, however, clumps are visible, likely due to contamination. A substantial C1s signal in XPS [Fig.\,\hyperref[fig4]{4(d)}], indicates that the contaminants are carbon-based organics. Annealing at 500\,$^\circ$C--650\,$^\circ$C in 1$\times$10$^{-7}$--1$\times$10$^{-6}$ O$_2$ back-pressure for 1--2 h gradually removed the C, but constant-height AFM still shows a surface covered by undetermined hillocks, with no hints of the underlying substrate. A significant reduction of C occurred only after annealing at 700\,$^\circ$C in 1$\times$10$^{-7}$\,mbar O$_2$ for 2 h. This treatment resulted in a reconstructed surface with a $\left(\sqrt{\mathrm{13}}\times \sqrt{\mathrm{13}}\right)\!\!R\mathrm{33.7}^\circ$ superstructure \cite{kienzle2011vacant,ohsawa2016negligible} clearly visible in LEED and the constant-height AFM image in Fig.\,\hyperref[fig4]{4(c)}.
In summary, the search for an SrTiO$_3$(001)-(1$\times$1) surface that can be considered 'pristine' --- crystalline and well-ordered, with a negligible amount of defects and contaminants --- was not successful so far. As-cleaved surfaces come closest, but necessarily contain both, SrO- and TiO$_2$-terminated domains (albeit of considerable size), and charged point defects. The temperature-induced transformation to an ill-defined, disordered top layer, indicates that the (1$\times$1) has a high surface energy and is only metastable. The techniques generally applied to judge surface quality (electron diffraction, XPS, large-area STM or ambient AFM) give results that would be consistent with a perfect (1$\times$1) termination; but are contradicted by atomically resolved nc-AFM. The same is true for the etched TiO$_2$-terminated surfaces that are used to great extent as substrates in the growth of heteroepitaxial oxide films. While such surfaces can display a flat morphology with a high-quality diffraction pattern, nc-AFM shows no signs of the unreconstructed surface. The temperature required for removing carbon lies above the stability region of the bulk-terminated surface.
The (001) surface of SrTiO$_3$ and other perovskites continue to be of great interest for both, probing their intrinsic properties and as substrates for heteroepitaxy. As the results presented here show, the assumption of an atomically clean, crystalline bulk-termination might not be warranted. This should be considered in the proper interpretation of experimental data.
This work was supported by the Austrian Science Fund (FWF) projects Wittgenstein Prize (Z-250), Solids4Fun (F-1234) and SuPer (P32148-N36). J.X. acknowledges the support from the National Natural Science Foundation of China (91634106), China Scholarship Council and Chongqing University. M.Se. acknowledges support from the Czech Science Foundation GACR 20-21727X and GAUK Primus/20/SCI/009.
| 9e8386c8b2d8ef7fb38a3da43ade44dedad91b03 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Experiments}
\label{sec:evaluation}
In this section we evaluate the proposed algorithm.
First, we present the datasets and evaluation metrics used (Section~\ref{sec:experim:datasets}).
Second, we provide both quantitative and qualitative results on several datasets %
and compare against state-of-the-art baselines (Section~\ref{sec:experim:quantitative}).
Third, we justify the choice of parameters (Section~\ref{sec:experim:ablation}).
Finally, we analyze the computational efficiency of our implementation (Section~\ref{sec:experim:computational})
and discuss its limitations (Section~\ref{sec:experim:limitations}).
\subsection{Datasets and Evaluation Metrics}
\label{sec:experim:datasets}
We evaluate Algorithm~\ref{alg: alternating strategy} extensively on datasets accompanying recent publications~\cite{Mitrokhin18iros,Stoffregen19iccv,Mitrokhin19iros,almatrafi2020distance}.
We use the same acronym to indicate the method and dataset from a publication, e.g.,
EMSMC denotes the ``Event-based Motion Segmentation by Motion Compensation'' method and its associated dataset, both presented in~\cite{Stoffregen19iccv}.
The distinction is clear from the context.
\begin{itemize}[noitemsep,nolistsep]
\item The Extreme Event Dataset (EED) \cite{Mitrokhin18iros} is one of the first open-source datasets used for the research of IMO detection and tracking.
Besides the camera's ego motion, there are other %
IMOs (up to three) in each sequence.
It provides manually-annotated bounding boxes for quantitative evaluation.
All sequences are collected in a laboratory environment, aiming to demonstrate the outstanding performance of event cameras in HDR scenarios.
\item EVIMO \cite{Mitrokhin19iros} is also collected in a lab environment but with better illumination.
Up to three IMOs %
appear in the sequences.
For evaluation, it provides dense segmentation masks of the IMOs.
Recently, an extension of the dataset, EVIMO2, has been made available online.
\item EMSMC \cite{Stoffregen19iccv} provides a set of real-world sequences captured in both indoor and outdoor environments.
Different from the above two datasets, it consists of a larger number of IMOs and even with non-rigid motions.
HDR scenarios are also captured in this dataset.
\item DistSurf \cite{almatrafi2020distance} is a dataset collected specifically for evaluating event-based optical flow estimation.
We find the data is in good quality and can be used for qualitative evaluation of our algorithm.
\end{itemize}
Table~\ref{talble: dataset summary} summarizes the key characteristics of the datasets.
\input{floats/dataset_summary_table}
\vspace{0.5ex}
\textbf{Evaluation Metrics}.
For quantitative evaluation we use two standard metrics.
The first one is \textit{detection rate} based on the overlap between the bounding boxes of detected and labeled objects, which was introduced in~\cite{Mitrokhin18iros} and used ever since.
It considers the detection result as successful if it meets the following conditions:
\begin{equation}
\label{eq:detectionrate}
\mathcal{B}_D \cap \mathcal{B}_G > 0.5\quad\text{ and }\quad(\mathcal{B}_D \cap \mathcal{B}_G) > (\mathcal{B}_D \cap \overline{\mathcal{B}_G}),
\end{equation}
where $\mathcal{B}_D$ refers to the estimated bounding box (or convex hull),
$\mathcal{B}_G$ the ground truth bounding box, and $\overline{\cdot}$ denotes the complement of a set.
The second metric is \emph{Intersection over Union} (IoU), which is the most commonly used metric to evaluate the performance of segmentation methods,
and was proposed for event data in \cite{parameshwara2020moms, Mitrokhin19iros}.
IoU is typically formulated as
\begin{equation}
\label{eq:IoUmetric}
\text{IoU} = (\cS_D \cap \cS_G) / (\cS_D \cup \cS_G),
\end{equation}
where $\cS_D$ refers to the resulting segmentation mask and $\cS_G$ the ground truth mask.
Note that the result of our algorithm consist of sparse warped events with specific labels.
To obtain a dense segmentation mask from a cluster of warped events, we label all pixels in its convex hull identically.
\subsection{Quantitative and Qualitative Evaluation}
\label{sec:experim:quantitative}
\textbf{EED Dataset}.
We first evaluate our algorithm on the EED~\cite{Mitrokhin18iros} dataset using the detection rate metric.
As reported in Table~\ref{tab: EED detection rate}, without having to specify in advance the number of clusters,
our algorithm outperforms other state-of-the-art solutions~\cite{Mitrokhin18iros,Stoffregen19iccv,parameshwara2020moms} in terms of average detection rate (97.45\%).
The numbers for the baseline methods are taken from the corresponding publications
in the absence of publicly available source code.
The numbers are close to those in~\cite{Stoffregen19iccv,parameshwara2020moms} partly because the detection rate~\eqref{eq:detectionrate} is a coarse evaluation metric.
Qualitative results are given in Fig.~\ref{fig: qualitative result EED}, where the red rectangles are ground truth bounding boxes.
\iflongversion %
Since the ground truth bounding boxes are manually annotated on raw images (from the DAVIS sensor) which are temporally close to the evaluation time, thus, a tiny spatio offset is witnessed sometimes.
\fi
\input{floats/EED_table}
\input{floats/EED_IWE_BoundingBox}
\textbf{EVIMO Dataset}.
A second quantitative evaluation is performed on the EVIMO dataset using the IoU metric~\eqref{eq:IoUmetric}.
As shown in Table~\ref{tab: IoU evaluation}, our algorithm performs the second best (only 0.19\% lower than the best one) among all four segmentation methods,
and about 2\% better than the next best method.
The numbers for the baseline methods are taken from~\cite{parameshwara2020moms}.
Note that Tables~\ref{tab: EED detection rate} and~\ref{tab: IoU evaluation} report results using two different metrics and datasets, so they are not directly related.
During the experiments we found several issues that may deteriorate the IoU score.
First, the ground truth masks are not perfectly aligned with objects in the raw images
because of inaccurate CAD models and IMO poses produced by the motion capture system.
Second, our labeled IWE and segmentation masks are computed using undistorted events (using camera calibration); then the masks are warped to the raw (distorted) image plane where the IoU score is calculated.
Undistortion leads to information loss near image boundaries. %
Third, some of the IMOs are hanged by the person who was holding the event camera, thus, the IMOs sometimes undergo the same motion as the camera.
In such a case, IMOs are labeled as background motion and no IoU score is calculated.
\input{floats/EVIMO_MOD_MASK}
\input{floats/EVIMO_IoU_table}
\input{floats/real_qualitative_comparison}
\input{floats/fig_real_more}
Exemplary results are given in Fig.~\ref{fig: qualtitative evaluation EVIMO}, where both labeled IWEs and their corresponding dense segmentation masks are visualized.
In the \emph{boxes} sequence, the toy car traverses from right to left and can be continuously detected as a moving object.
There are two IMOs in the \emph{table} sequence (toy car and plane).
The two move against each other and meet in the middle.
They are successfully detected even when they are partially overlapped.
The toy plane stays almost still at the end of the sequence, which explains why it is labeled as background.
The IWE around the area of the plane looks as sharp as the background, which also indicates that the plane stays still.
The figure also includes qualitative comparisons against~\cite{parameshwara2020moms} in terms of the labeled IWEs.
Despite the different coordinates used, our results look overall sharper and do not exhibit the rectangular segmentation boundaries present in~\cite{parameshwara2020moms} (caused by naively labeling events according to whether or not they are inside the convex hull of the cluster features).
In the middle of the \emph{table} sequence, our method still detects the toy plane as an IMO (before it slows down and merges with the background), whereas \cite{parameshwara2020moms} does not.
In addition, the authors of \cite{Mitrokhin19iros} just released a new dataset, EVIMO2, using event-based cameras of VGA resolution ($640\times 480$ pixels).
The dataset can be used for evaluation of event-based object segmentation, motion segmentation, and structure from motion.
The multi-camera system consists of a Samsung Gen3 DVS~\cite{Son17isscc},
two Prophesee CD Gen3 event cameras~\cite{propheseeevk} and a Flea3 RGB camera.
We test our method on the events from the Samsung DVS and provide qualitative results in Fig.~\ref{fig:real:more}.
We also compute the IoU score (64.38 \%), and the main reason for the less accurate segmentation than in EVIMO is the motion patterns:
the objects in EVIMO2 undergo more frequent 3D rotations than in EVIMO.
Consequently, the 2D appearance of the objects on the image plane continuously changes, causing self-occlusions that are difficult to model with current image-plane motion models.
To obtain best results in this scenario, knowledge of the 3D shape and appearance of the IMO would be required, which is left as future work.
In spite of this, the qualitative results in Fig.~\ref{fig:real:more} (columns 4 and~5) show good segmentation results.
\input{floats/fig_ablation_study}
\textbf{Other Datasets}.
Besides above quantitative results, we also provide an extensive qualitative evaluation on real-world data from EMSMC~\cite{Stoffregen19iccv} dataset, DistSurf~\cite{almatrafi2020distance} dataset, and our own collection.
These sequences cover a wide variety of scenes, ranging from indoor lab environments to outdoor traffic scenes with moving vehicles and pedestrians, including non-rigid motion and HDR scenarios.
We provide exemplary results in Figs.~\ref{fig:emsmc:compare} and~\ref{fig:real:more}.
The first sequence of Fig.~\ref{fig:emsmc:compare} shows a traffic scene captured from above, with three vehicles driving on the road.
The two cars moving to the left have very similar velocities, thus, they are clustered in the same group.
For visual comparison we provide the segmented, motion-compensated images from~\cite{Stoffregen19iccv}.
The second sequence of Fig.~\ref{fig:emsmc:compare} shows an outdoor HDR scene captured with the event camera facing the sun while a pedestrian and a skateboarder pass by.
Our algorithm preserves the motion discrepancy among different parts of the non-rigid human bodies while maintaining the compactness of the segmentation due to the applied MDL term.
The third sequence shows a vehicle passing by in front of buildings.
Our algorithm successfully distinguishes the car from the buildings in the background, which are compactly labeled together, thus identifying the panning camera motion.
In Fig.~\ref{fig:real:more}, the first column shows a fan (whose blades rotate at $\approx$\SI{1800}{\degree/\second})
and a free-falling coin, \textit{i}.\textit{e}., a high-speed scenario.
Since our algorithm supports multiple motion models, the fan blades and the coin are successfully detected and distinguished.
The second column shows a traffic scene captured at street level.
The motions of overlapping vehicles can be successfully distinguished.
The third column shows a scene with a pair of waving hands.
The motions of the hand and arm are sometimes segmented from each other when there is a small wrist motion.
The fourth and fifth columns are exemplary results of EVIMO2 dataset, already discussed.
The last column shows another sequence of our own data (different from the one in Fig.~\ref{fig:eyecatcher}), in which a nursing pillow is thrown.
The pillow is detected as an IMO detached from the hand as soon as the apparent in-plane rotation happens.
\subsection{Parameters of the Method}
\label{sec:experim:ablation}
Let us mention how the parameters of the method are set.
First, we process events in packets (\textit{i}.\textit{e}., sliding window fashion).
The number of events $N_e$ may be selected based on the scene dynamics or texture~\cite{Liu18bmvc}.
However, we found that $N_e \in [15000, 30000]$ is a sensible choice for sequences captured by the DAVIS240 or DAVIS346~\cite{Brandli14ssc}.
For sequences acquired by higher resolution sensors (e.g., $640 \times 480$ pixels), $N_e$ is increased proportionally.
As shown in the experiments, motion parameters can be assumed to be constant within each sliding window.
Second, we set the weights $\lambda_{\text{P}} = 40, \lambda_{\text{M}} = 8,000$.
The Potts model weight $\lambda_{\text{P}}$ is set according to the fact that an event typically has at least six spatio-temporal neighbours (At least two edges are linking to the event's pixel location due to the Delaunay triangulation).
The maximum discrepancy for an unary term is 255 according to the IWE negative.
Thus, $\lambda_{\text{P}} = 255 / 6 \approx 40$ so that local consistency would be sacrificed for spatial coherence.
For sequences with relatively small IMOs traversing a textured background, $\lambda_{\text{P}}$ has to be reduced to avoid over-smoothing effects.
The MDL weight $\lambda_{\text{M}}$ is set according to an ablation study, as shown in Fig.~\ref{fig:ablation study}.
Unsurprisingly, we see the resulting number of IMOs becomes smaller as $\lambda_{\text{M}}$ increases.
We observe $\lambda_{\text{M}}$ having a wide range of endurance ($[8000, 16000]$), which makes a good balance between enhancing spatial coherence and circumventing over-smoothing effect.
We find 8,000 a reasonable choice that returns the true number of IMOs in most cases.
Finally, the number of hierarchy levels $N$, which determines the size of the sub-volumes used during initialization, is set empirically.
A good choice (e.g., Fig.~\ref{fig: initialization procedure}) can effectively pick up small IMOs while circumventing cases of bad signal-to-noise ratio (Sec.~\ref{sec:experim:limitations}).
We find that $N=4$ is a good choice for sensors with similar spatial resolution to the DAVIS346, and $N$ may be increased to deal with very small IMOs and/or new devices with higher spatial resolution.
\subsection{Computational Performance}
\label{sec:experim:computational}
Algorithm~\ref{alg: alternating strategy} consists of three main steps:
($i$) initialization, ($ii$) creation of event graph,
and ($iii$) alternating discrete-continuous optimization.
The initialization is the most time-consuming step.
Using 15,000 events as input and a 4-level subdivision as an example, it takes about \SI{4}{\second} to compute all motion candidates of the model pool.
This time is spent on one motion model type; for multi-model proposals, \textit{e}.\textit{g}., $K$ types, the computation is $K$ times larger.
We apply hyper-threading to speed up the process.
Initialization may be expensive for the first set of events, but it can be propagated according to the motion and reutilized for the upcoming events.
The creation of the space-time event graph is efficient, taking \SI{45}{\milli\second}.
The optimization terminates within 3 iterations, and its time is proportional to the size of the motion pool; it takes about \SI{3}{\second}.
The discrete-labeling (graph-cut) sub problem takes the majority of the runtime compared to the negligible time spent on continuous model fitting.
The proposed method is implemented in C++ on ROS and runs on a laptop with an Intel Core i7-8750H CPU.
The code uses two cores and consumes about 400 MB of memory.
While there is room for improvement in runtime performance to speed up the method by an optimized implementation and/or dedicated hardware,
event cameras are data-driven sensors that produce more events per second the faster the motion, the smaller the contrast sensitivity and the more texture present in the scene.
Hence, it is conceivable to modify the above factors to increase the event rate in order to overwhelm any non-trivial event-based method so that it becomes non ``real-time''.
This could be mitigated by decreasing the contrast sensitivity and/or by dropping events (randomly or with some control strategy~\cite{Glover18icra}).
However, this is out of the scope of this work.
In contrast, we think that the most alluring feature of our proposal is the modeling idea of using graph cuts on event-based data, which ($i$) brings well-known principles to the realm of event cameras and ($ii$) allows us to address shortcomings of previous methods, such as the need to either specify the number of clusters in advance or over-segment the scene.
\subsection{Limitations}
\label{sec:experim:limitations}
A limitation observed during initialization entails the size of the IMOs.
Small IMOs, \textit{e}.\textit{g}., smaller than $1\%$ of the sensor's spatial resolution (as in the $\emph{multiple objects}$ sequence of~\cite{Mitrokhin18iros})
cast signal-to-noise ratio problems during motion proposal generation and are thus unlikely to be properly initialized.
The size of the IMOs in the other datasets is larger than in the EED dataset, hence detecting IMOs is not an issue in them.
\emph{Noise}. In addition, the performance of our method is affected by the signal-to-noise ratio of the input events because the topological structure of the ST graph deteriorates with increasing event noise (jitter, bursts, etc.).
To alleviate this issue, a pre-processing de-noising step~\cite{czech2016evaluating, wu2020probabilistic} is suggested. %
We apply a simple de-noising operation, as proposed in~\cite{zhou2020event}, which filters isolated events in the spatio-temporal domain.
To deal with the specific noise in the datasets that is due to ground truth acquisition by the motion capture system (disturbing light from a VICON system's emitters) additional de-noising methods, such as burst filters, are needed.
\section{Event-based Motion Segmentation}
\input{chapters/02_event_camera_intro}
\subsection{The Clustering Nature of the Problem}
\label{sec:problem:intro}
Assuming constant illumination, events are due to the relative motion between the camera and the scene,
and since events represent brightness changes~\eqref{eq:generativeEventCondition}, they are mostly produced by scene edges (contours, texture, etc.).
If the camera is stationary, events are solely due to moving objects, whereas if the camera moves, events are due to both:
edges of moving objects and moving edges induced by the camera's ego-motion.
The problem of \emph{event-based motion segmentation} consists of identifying which events in a given event stream are triggered by the same scene object,
and because events are caused by motion, the identification criterion primarily refers to grouping (i.e., \emph{classifying}) events by the type of motion.
Hence it is key to realize that event-based motion segmentation is a \emph{clustering problem by nature}:
splitting the event stream into different groups of events, with each one representing a coherent motion (classification criterion),
as shown in the input-output of Fig.~\ref{fig:eyecatcher}.
The problem is most challenging in the moving-camera scenario because events are triggered everywhere on the image plane, not just around the moving objects.
As we review next (Section~\ref{sec: related work}), several solution methods have been proposed, and due to the above nature of the problem, they all inherently perform some sort of clustering.
However, they differ in the clustering technique developed.
\subsection{Related Work}
\label{sec: related work}
Early works on event-based motion segmentation required prior knowledge on either the shape of IMOs (\textit{e}.\textit{g}., a circle \cite{Glover16iros}) or the correlation between the tracked geometric primitives and the motion of the event camera~\cite{Vasco17icar}.
Such prior knowledge is no longer required in recent works \cite{Mitrokhin18iros,Stoffregen17acra,Stoffregen19iccv,parameshwara2020moms,Mitrokhin19iros}.
Several of these pipelines follow a sequential strategy, which consists of analyzing the dominant events (\textit{e}.\textit{g}., background), then removing these (by empirical thresholding \cite{Mitrokhin18iros,Stoffregen17acra}) and analyzing the remaining events (\textit{i}.\textit{e}., the IMOs), \emph{greedily}.
Instead, using \cite{Stoffregen17acra} as initialization, \cite{Stoffregen19iccv} was the first to tackle the problem \emph{jointly}, analyzing all events while solving the two sub-problems of clustering: event-object association (classification) and object / cluster refinement (model-fitting).
By leveraging the idea of motion compensation~\cite{Gallego18cvpr}, it formulated the segmentation problem using an expectation-maximization (EM) approach, which iteratively updated the soft event-cluster associations and the motion model parameters.
It provided per-event segmentation rather than classical bounding-box results.
Recently, \cite{parameshwara2020moms} proposed a similar joint optimization method, but with differences:
($i$) initialization of IMO models was based on $K$-means clustering of event-based feature tracks,
and ($ii$) event-cluster assignments were based on morphological operations via empirical thresholding.
In addition to the above methods, \cite{Mitrokhin19iros} proposed a supervised end-to-end learning-based pipeline that simultaneously solved for optical flow, 3D motion and object segmentation.
Although \cite{Mitrokhin19iros} is not closely related to the above approaches (neither is to ours), it is mentioned because it provides a state-of-the-art dataset for segmentation evaluation (Section~\ref{sec:evaluation}).
\paragraph*{Similarities and Differences with Prior Work}
Like previous methods (\textit{e}.\textit{g}., \cite{Stoffregen19iccv}), our method also allows general parametric warps (motion models) and performs per-event segmentation.
Besides, we are also able to produce sharp, motion-compensated images as a by-product (Fig.~\ref{fig:eyecatcher}, Right).
However, we claim the following fundamental \emph{differences} compared to previous approaches.
First, we formulate the problem using Markov Random Fields (MRF) and defining a spatio-temporal graph through the events.
This leads us to efficiently solve the problem using \emph{graph cuts}, which is the first time that they are adapted to work on event data.
Second, we pose the problem as a joint optimization over the motion parameters and event associations, but in contrast to \cite{Stoffregen19iccv} we introduce two spatial regularizers.
In particular, we explicitly minimize the number of clusters and smooth their shape, which is pursued naturally via an energy formulation.
This allows us to solve the issue of not knowing the number of moving objects in the scene~\cite{Stoffregen19iccv}.
Third, digging into details of the event alignment data-fidelity terms, \cite{Stoffregen19iccv} is based on variance maximization, whereas graph cuts require a minimization formulation with non-negative terms.
For this non-trivial adaptation we build on our work~\cite{zhou2020event} and propose negative Images of Warped Events (IWEs), which have not been investigated for clustering.
Finally, we do not follow the greedy strategy nor do we need additional methods for initialization (\textit{e}.\textit{g}., feature tracking~\cite{parameshwara2020moms}).
We provide a new initialization method, based on a hierarchical subdivision of the volume of events to provide a pool of motion instances.
\section{Conclusion}
\label{sec:conclusion}
We presented a novel method for event-based motion segmentation.
Our approach is a multi-model fitting scheme that jointly clusters events and fits motion models to them.
The proposed graph-based (MRF) formulation with the additional MDL energy term leads to globally consistent and spatially coherent segmentation results using fewest labels.
As a by-product, the method produces labeled, motion-compensated images of warped events that may be used for further processing (\textit{e}.\textit{g}., recognition).
A thorough evaluation demonstrated the versatility of our method in scenes with different motion patterns and unknown number of independent motions.
We also showed that the method is able to bring the advantages of event-based cameras to tackle traditionally difficult scenarios for standard frame-based cameras, such as segmentation of fast moving objects (that would cause motion blur) or in HDR conditions.
Finally, we hope this work inspires new research in the topic of segmentation with event-based cameras, a paramount but rather unexplored topic.
\section{Joint Optimization of Motions and Segments}
\label{sec:optimization}
We now discuss how to minimize the proposed energy function \eqref{eq: overall energy function}.
This energy depends on both discrete labeling variables $L$ and continuous motion parameters $\mathcal{M}$, thus leading to a discrete-continuous optimization problem.
Inspired by efficient solvers used in classical multi-model fitting methods~\cite{delong2012fast,yang2015dense},
we employ a block-coordinate descent strategy to optimize $L$ and $\mathcal{M}$ in an alternating manner.
We present the solution to each sub problem, followed by the initialization strategy.
The overall method is summarized in Algorithm.~\ref{alg: alternating strategy}.
\input{floats/alg_float}
\subsection{Segmentation:\! update labels $L$ given motions~$\mathcal{M}$}
\label{subsec: discrete sub problem}
The overall energy~\eqref{eq: overall energy function} reduces to the sub-problem of discrete labeling when motion models $\mathcal{M}$ are fixed:
\begin{equation}
E(L) = E_{\text{D}}(L) + \lambda_{\text{P}}E_{\text{P}}(L) + \lambda_{\text{M}}E_{\text{M}}(L).
\label{eq: discrete sub problem}
\end{equation}
This energy describes a standard %
MRF problem plus an additional MDL term.
The graph-cut method is one of the most widely adopted techniques for solving MRF problems.
The simplest case of graph cut, also known as s-t cut \cite{kolmogorov2004energy}, is typically used for solving a binary classification problem.
A graph is defined connecting the unknowns of the problem, that is, the graph vertices, which must be assigned labels. Two additional vertices called source (s) and sink (t) are defined.
The minimum s-t cut partitions the vertices of the graph into two disjoint groups (i.e., labels) at the smallest energy cost, which is equivalent to computing the maximum flow from source to sink \cite{ford2015flows}.
The generalization of the minimum s-t-cut problem, such as~\eqref{eq: discrete sub problem}, involves more than two terminals (labels).
For energy functions consisting of a smoothness term with discontinuity-preserving property (e.g., the applied Potts model), the expansion move algorithm \cite{boykov2001fast} can be used as long as the smoothness term is a metric on the space of labels.
The $\alpha$-expansion algorithm loops through the labels $\alpha$ in some order and looks for the lowest energy.
In every iteration, it solves a minimum s-t-cut problem which partitions the vertices into the current-labeling group and the $\alpha$-label group.
An $\alpha$-expansion movement is made according to the s-t cut that leads to the lowest energy.
To minimize \eqref{eq: discrete sub problem}, we apply the $\alpha$-expansion based graph-cut method~\cite{boykov2001fast} combined with the method in~\cite{delong2012fast} to handle the label costs induced by the MDL term.
\iflongversion
From an implementation point of view, we use bilinear interpolation to count events at non-integer warped positions $\mathbf{x}'_k$ of the IWE~\cite{Rebecq18ijcv}.
Then, unary terms are floored before they are passed to graph cut, as it is standard.
\fi
To accelerate the algorithm, if a motion model is not assigned to any cluster of events, it is removed from the model pool.
The remaining models (label ID) are sorted according to the number of events that belong to the corresponding clusters.
\subsection{Model Fitting:\! update motions $\mathcal{M}$ given labels~$L$}
\label{subsec: continuous sub problem}
When the label variables are fixed, the energy~\eqref{eq: overall energy function} simplifies to the data term only.
The continuous variables of each motion model can be re-fitted independently from other motion models using the corresponding cluster of events.
The original motion compensation scheme \eqref{eq:argmaxMotionModel} %
is applied for model fitting.
\subsection{Initialization}
\label{subsec: initialization}
Let us show how to initialize the optimization procedure.
Unlike existing solutions, which either greedily initialize motion models $\mathcal{M}$~\cite{Stoffregen17acra, Mitrokhin18iros,Stoffregen19iccv} or apply the K-means method on computed event-based optical flow~\cite{Benosman14tnnls} or feature tracks~\cite{parameshwara2020moms},
we propose a simple, direct and effective initialization based on the original motion compensation scheme.
Given a space-time volume $\mathbf{V}$ of events $\cE$, we carry out an $N$-level subdivision operation.
At level $n \in [0, N-1]$ of the hierarchy, the volume $\mathbf{V}$ is divided evenly into $4^{n}$ sub-volumes.
Let us use $N = 4$ as an example, as shown in Fig.~\ref{fig: initialization procedure}, where the blue dashed rectangle illustrates a sub volume at level $n = 1$, while the green one shows a sub-volume at level $n = 3$.
For simplicity, the volume $\mathbf{V}$ is visualized in 2D, namely by accumulating events on the reference time slice.
After the division operations, we have %
$4^{0} + 4^{1} + \cdots + 4^{3} = 85$ sub-volumes (including the whole volume at the base level of the hierarchy).
By feeding the events in these sub-volumes to the motion compensation scheme, we get a pool of 85 motion model candidates~$\mathcal{M}$.
This strategy aims at capturing %
IMOs of different size.
As illustrated in Fig.~\ref{fig: initialization procedure}, the blue dashed sub-volume captures one of the boxes in the background, which leads to an IWE with high contrast at the background structures,
whereas the green dashed sub-volume senses part of a smaller IMO, which leads to an IWE with high contrast around that IMO.
The resulting model pool is used to compute the data term (Section.~\ref{subsec: data term}).
Computational complexity is proportional to the number of event warping operations.
Thus we compare our method against the greedy alternative in terms of the number of warped events.
Assume there are $N_e$ events involved totally and they are induced by the background motion (bg) as well as $m$ IMOs.
Thus, we have $N_e = N_e^{\text{bg}} + \sum_{i=1}^{m} N_e^{\text{IMO}_i}$.
The greedy solution assumes a dominance order of motion models, sorted by the number of events: $N_e^{\text{bg}} \gg N_e^{\text{IMO}_1} \gg ... \gg N_e^{\text{IMO}_m}$.
Under the reasonable assumption of a predominant background motion, $N_e \approx 2\, N_e^{\text{bg}}$, and IMOs accounting for the other half of the events $N_e \approx 2^{i+1}\, N_e^{\text{IMO}_i}$,
the computational complexity of the greedy method is $O((2 - 2^{-m})\,N_e)$, which is bounded by $O(2\, N_e)$.
The computational complexity of our initialization method depends on the number of subdivision levels $N$, and it is $ O(N\, N_e)$.
While this simplified complexity comparison favors the greedy approach, in practice we found out that our initialization works well by using only the finest level ($n=3$), which has the smallest complexity $O(N_e)$.
\input{floats/initialization_procedure}
\section{Designed Energy for Motion Segmentation}
\label{sec:energy-formualtion}
We propose an energy function that considers jointly the sub-problems of labeling (i.e., segmentation) and model fitting.
The energy function is defined on both unknowns (labeling configuration $L$ and cluster motions $\mathcal{M}$) as
\begin{equation}
E(L, \mathcal{M}) \doteq E_{\text{D}}(L, \mathcal{M}) + \lambda_{\text{P}}E_{\text{P}}(L) + \lambda_{\text{M}}E_{\text{M}}(L),
\label{eq: overall energy function}
\end{equation}
where $E_{\text{D}}$ denotes the data term and $E_{\text{P}}$, $E_{\text{M}}$ constitute the regularizer. %
Specifically, $E_{\text{P}}$ is the regularizer of the Potts functional \cite{potts1952some}
and $E_{\text{M}}$ is a label cost term representing the Minimum Descriptor Length (MDL) principle~\cite{delong2012fast}.
The weights $\lambda \geq 0$ balance the contribution of each term.
The energy is designed such that its minimizer achieves the best fit to the data while being spatio-temporally smooth and having the fewest labels (i.e., segments).
\iflongversion
It is thus a constrained optimization problem that is posed in an unconstrained weak optimization manner.
\fi
We detail the design of each energy term in the upcoming sections.
\subsection{Data Fidelity Term}
\label{subsec: data term}
The data term in~\eqref{eq: overall energy function} is defined on both the discrete labeling variables and the continuous model parameters.
If the labeling variables are fixed, the energy reduces to the data term and the optimal motion model for each cluster can be obtained using the motion compensation scheme (see Section~\ref{subsec: continuous sub problem}).
Hence, we focus on the design of the data term from the perspective of the discrete labeling variables.
To adapt motion compensation~\cite{Gallego18cvpr} to a graph-based energy formulation~\eqref{eq: overall energy function} we need to reformulate it so that:
($i$) energy is minimized instead of maximized,
($ii$) the fitting cost of a group of events is given by the sum of fitting costs of individual events, i.e., the so-called unary terms~\cite{boykov2001fast}.
The original motion compensation scheme creates an IWE~\eqref{eq:IWE}, on which image contrast is measured.
The IWE is created by warping events according to a certain motion model. %
The higher the IWE contrast, the sharper the IWE, and consequently, higher pixel values (i.e., event accumulation) appear around the edge patterns.
Therefore, the intensity value at each IWE pixel is a proxy for the consistency (goodness of fit) between the model and the events that are warped to that pixel.
To convert the maximization problem into a minimization one and to build the unary costs of the data term, we propose to use the IWE ``negative''~\cite{zhou2020event}.
Once an IWE $I$ is computed~\eqref{eq:IWE}, it is normalized to a fixed range, e.g., $\left[0, 255\right]$, and then its negative is calculated as $\bar{I} = 255 - \hat{I}$,
where $\hat{\cdot}$ denotes the normalization operation, and $\overline{\cdot}$ the negative operation.
We define the unary term for an event $e_k$ as the value at its warped location, $\mathbf{x}'_k$ in~\eqref{eq:warped-events},
on the IWE negative, namely $\bar{I}(\mathbf{x}'_k; \bm_l)$.
The data term is defined as the sum of all unary terms:
\begin{equation}
E_\text{D}(L) \doteq \sum_{l \in \mathcal{L}}\sum_{e_k \in \mathcal{C}_l} \bar{I}(\mathbf{x}'_{k}; \bm_l),
\end{equation}
where $\mathcal{C}_l$ denotes the cluster of events with label $l$,
and $\bar{I}(\cdot\,; \bm_l)$ the IWE negative created using motion model $\bm_l$.
\subsection{Spatially Coherent Labeling}
\label{subsec: potts model term}
As for the regularizer, we simply use a pairwise Potts model term $E_{\text{P}}$ to encourage spatially coherent labeling.
This term is defined only on the discrete labeling variables:
\begin{equation}
E_{\text{P}}(L) \doteq \sum_{e_i, e_j \in \mathcal{N}} \delta_{L(e_i),L(e_j)},
\end{equation}
where $\mathcal{N}$ denotes the event neighbourhood in the spatio-temporal graph (Section~\ref{sec:event-graph}),
and $\delta_{m,n}$ is the Kronecker delta (1 if the variables $m,n$ are equal, and 0 otherwise).
\subsection{Encouraging Few Number of Segments}
\label{subsec: MDL term}
To discourage redundancy of the assigned motion models we introduce an MDL term:
\begin{equation}
\label{eq:MDL}
E_{\text{M}}(L) \doteq \sum_{l=1}^{N} \psi(l),
\;\; \psi(l) \doteq \begin{cases} 1 &\mbox{if } \displaystyle{\sum_{e \in \mathcal{C}_l}\delta_{L(e),l} > 0} \\
0 & \mbox{otherwise.}
\end{cases}
\end{equation}
This term penalizes the total number of assigned (active) labels (i.e., segments),
which encourages it to converge to the actual number of IMOs. %
We describe the optimization procedure for solving the proposed energy in the next section.
\section{Acknowledgement}
We thank Dr. Timo Stoffregen and his co-authors for providing test data from~\cite{Stoffregen19iccv}.
We thank Chethan Parameshwara for providing the comparison images in Fig.~\ref{fig: qualtitative evaluation EVIMO}.
We thank Chuhao Liu and Hao Xu for their help during data collection.
\subsection{Event-based Camera Working Principle}
\label{sec:DVS_operation_application}
Event-based cameras~\cite{Lichtsteiner08ssc} have independent pixels that continuously monitor their incident light and respond to changes of predefined size $C$, which are called ``events''.
Specifically, if $L(\mathbf{x},t)\doteq \log I(\mathbf{x},t)$ is the logarithmic brightness at pixel $\mathbf{x} \doteq (x,y)^\top$ on the image plane, an event $e_k \doteq (\mathbf{x}_k,t_k,p_k)$ is generated at pixel $\mathbf{x}_k$ and time $t_k$ (with microsecond resolution) if the change in log-brightness reaches $C$ (typically 10-15\% relative change):
\begin{equation}
\label{eq:generativeEventCondition}
\Delta L \doteq L(\mathbf{x}_k,t_k) - L(\mathbf{x}_k,t_k-\Delta t_k) = p_{k}\, C,
\end{equation}
where $\Delta t_k$ is the time since the last event at the same pixel $\mathbf{x}_k$
and $p_k \in \{+1,-1\}$ is the event polarity (i.e., sign of $\Delta L$).
Hence, each pixel has its own sampling rate (which depends on the visual input)
and outputs data proportionally to the amount of motion or illumination variations in the scene.
An event-based camera does not produce images at a constant rate, but rather a stream of \emph{asynchronous}, sparse events in space-time domain.
\section*{Multimedia Material}
\noindent Project page: {\small \url{https://sites.google.com/view/emsgc}}\\
Code: {\small \url{https://github.com/HKUST-Aerial-Robotics/EMSGC.git}}\\
\section{Introduction}
\label{sec: introduction}
\input{floats/fig_eye_catcher_page2}
Event-based cameras, such as the Dynamic Vision Sensor (DVS) \cite{Lichtsteiner08ssc,Suh20iscas,Finateu20isscc}, are novel bio-inspired visual sensors.
Unlike standard cameras that acquire data at a fixed rate, event-based cameras report per-pixel intensity changes asynchronously at the time they occur, with microsecond resolution, called ``events''.
This working principle offers potential advantages (low latency, high temporal resolution, high dynamic range and low power consumption) to tackle challenging scenarios in computer vision such as high-speed and/or high dynamic range (HDR) optical flow estimation~\cite{Benosman14tnnls,Zhu18rss}, feature tracking~\cite{Lagorce15tnnls,Valeiras15tnnls,Zhu17icra,Gehrig19ijcv},
stereo depth estimation~\cite{Rogister12tnnls,Lee14tnnls,Osswald17srep,Zhou18eccv},
camera tracking~\cite{Mueggler15rss,Gallego17pami,Chamorro20bmvc},
control~\cite{Conradt09iscas,Delbruck13fns} and Simultaneous Localization and Mapping (SLAM)~\cite{Kim16eccv,Rebecq17ral,Rosinol18ral,Mueggler18tro,zhou2020event}.
However, due to the above principle of operation and unconventional output, algorithms designed for standard cameras cannot be directly applied.
Novel algorithms are needed to unlock event cameras' potential.
The recent survey \cite{Gallego20pami} provides a comprehensive review on event-based cameras, algorithms and applications.
In this paper, we consider the problem of event-based motion segmentation, which aims at classifying events occurred during a time interval into several groups that represent coherent moving objects.
We tackle the most general case: a possibly moving event camera observing a dynamic scene.
It is more challenging than the static camera case because events are no longer solely due to IMOs; they are also induced by the background.
We develop a method to jointly classify events and estimate their coherent motion, in an iterative and alternating manner.
\textbf{Contributions}. Our contributions can be summarized as:
\begin{enumerate}
\item A novel event-based motion segmentation method designed in the spirit of motion compensation~\cite{Gallego18cvpr} and built on top of classical multi-model fitting schemes.
We propose a space-time event graph representation to exploit the spatio-temporal nature of events, leading to globally consistent and locally coherent labeling results.
\item A simple and effective initialization method using the original motion compensation scheme on raw events.
\item A formulation that does not require prior knowledge in the form of scene geometry, motion patterns or number of IMOs.%
\item An extensive evaluation, qualitative and quantitative, on %
available datasets, showing better performance than directly competing baseline methods.
\end{enumerate}
\vspace{1ex}
The rest of the paper is organized as follows:
Section~\ref{sec:DVS_operation_application} briefly describes the working principle of event-based cameras.
Section~\ref{sec:problem:intro} discusses the nature of the event-based motion segmentation problem and the related work.
Section~\ref{sec:problem-statement} more formally states the problem and necessary preliminaries.
Sections~\ref{sec:energy-formualtion} and~\ref{sec:optimization} disclose our energy formulation and optimization procedure, respectively.
The method is evaluated in Section~\ref{sec:evaluation} and conclusions are drawn in Section~\ref{sec:conclusion}.
\section{Problem Statement Preliminaries}
\label{sec:problem-statement}
The goal of this work is to cluster the events produced by an event-based camera into groups that undergo coherent motions.
Such motions aim to represent the unknown number of IMOs in the scene.
Once clustered, events are warped according to the estimated motions and produce an overall %
IWE with the highest contrast (Fig~\ref{fig:eyecatcher}, Right).
In this section we briefly introduce multi-model fitting and discuss how to adapt its components to event data.
We introduce the notions of spatio-temporal graph of events and model-fitting metric(s) on the graph.
\subsection{Multi-Model Fitting}
\label{sec:multi-model-fitting}
Multi-model fitting is a category of computer vision problems that aim at explaining data using several model instances.
It is a chicken-and-egg problem, often split into two sub-problems: assignment of data to a model instance (i.e., classification or ``labeling'') and model fitting (i.e., parameter estimation).
Examples consist of recognizing and segmenting partially occluded objects in 2D~\cite{winn2006layout} or 3D~\cite{hoiem20073d}, optical
flow estimation~\cite{trobin2008continuous,roth2009discrete,yang2015dense}, and motion segmentation~\cite{isack2012energy}.
These problems aim at assigning a label $l_p$ to each data point $p$, where each label $l \in \mathcal{L}$ corresponds to a model $\mathcal{M}_{l}$ that is consistent with local observations.
The solution to the problem is a labeling configuration $L$ that is locally smooth and globally consistent.
To find such a solution, multi-model fitting problems are naturally formulated as the minimization of an energy $
E = E_{\text{data}} + E_{\text{reg}},$
comprising a data term $E_{\text{data}}$ that measures the inconsistency between the data and the models,
and a regularizer (smoothness term) $E_{\text{reg}}$ that enforces prior knowledge about the models.
A successful set of solvers consider a graph through the data points
and seek to partition the graph into the optimal labels \cite{Boykov01iccv,boykov2001fast}.
We also follow this approach.
\subsection{Space-Time Event Graph}
\label{sec:event-graph}
The data points in our problem consist of events $\{e_k\}$ produced by a DVS~\cite{Lichtsteiner08ssc}.
Because events are sparse in the spatio-temporal domain (time-evolving image plane), the underlying graph considered is in general unstructured
(as opposed to the regular graph of pixel intensities in an image).
To build a spatio-temporal graph for events while keeping a low complexity, we propose to use a Delaunay triangulation~\cite{shewchuk2009general} on the binary image of active events in the space-time volume $\mathbf{V}$ (Fig.~\ref{fig:spatio-temporal-graph}).
Specifically, given the events in a volume $\mathbf{V}$ of size $W \times H \times \delta t$ (where $W, H$ refer to the width and height of the image plane, and $\delta t$ denotes the time span),
we first compute a binary image %
of event activity (i.e., the pixel is 1 if it contains at least one event, illustrated by blue squares in Fig.~\ref{fig:active pixels}, and it is 0 otherwise).
Then we compute the Delaunay triangulation on the non-zero pixels of this binary image (black dots in %
Fig.~\ref{fig:delaunay triangulation}), which returns a 2D graph (mesh). %
This 2D graph is used to build a space-time (3D) graph for the events
(see the connection principle in Fig.~\ref{fig:connection principle}).
Each event typically has $2 + 2N$ neighbors in the resulting graph (Fig.~\ref{fig:resulting ST graph}), where $N$ denotes the number of edges that link to the event's pixel location in the binary image (black dots in Fig.~\ref{fig:spatio-temporal-graph}).
There are many possible graphs that can be used to connect the event data points.
The above proposal is inspired in graphs proposed for Markov Random Fields
built on sets of sparse 2D features \cite{russell2011energy} and is easy to implement from a data structure point of view.
\input{floats/spatio_temporal_graph}
\subsection{Goodness of Fit by Motion Compensation}
\label{sec:review-motion-compensation}
The specification of a metric on the data graph allows us to assess the goodness of fit between the data and the model(s) (\textit{i}.\textit{e}., $E_\text{data}$).
Building upon~\cite{Gallego19cvpr}, the goodness of fit is given by the alignment of the events along point trajectories on the image plane,
which constitute a motion model and are parametrized by a parameter vector $\bm$.
Such event alignment is assessed by the strength of the contours of an IWE. %
The contour strength (which is related to image contrast~\cite{gonzalez2004digital}) can be measured with various dispersion metrics, such as variance~\cite{Gallego17ral,Gallego18cvpr,Stoffregen19iccv}, RMS of mean timestamp per pixel~\cite{Zhu19cvpr}
We utilize the variance loss
\iflongversion
introduced by~\cite{Gallego18cvpr}, which has shown advantages against others in terms of accuracy and computational efficiency~\cite{Gallego19cvpr},
\fi
as metric for the edges of the event graph.
To make the paper self-contained, we briefly review the idea of motion compensation~\cite{Gallego17ral,Gallego18cvpr}, upon which motion models are fitted.
\textbf{Motion Model Fitting}.
The contrast maximization framework~\cite{Gallego18cvpr} allows us to fit a motion model to a group of events $\cE = \{e_k\}_{k=1}^{N_e}$.
First, events are geometrically transformed according to a warping function~$\Warp$,
\begin{equation}
\label{eq:warped-events}
e_k \doteq (\mathbf{x}_k, t_k) \;\mapsto\; e_k^{\prime} \doteq (\mathbf{x}_k^{\prime}, t_{\text{ref}}),
\end{equation}
leading to a set of warped events $\cE' = \{e_k^{\prime}\}_{k=1}^{N_e}$ at a reference time $t_{\text{ref}}$.
The warping function $\Warp(\mathbf{x}_k, t_k; \bm) \doteq \mathbf{x}_k^{\prime}$ implements the image-plane trajectories of a motion model and is parametrized by $\bm$.
Secondly, events $\cE'$ are aggregated into an image (or histogram) of warped events (IWE),
\begin{equation}
\label{eq:IWE}
I(\mathbf{x}; \bm) \doteq \sum_{k=1}^{N_e}\delta (\mathbf{x} - \mathbf{x}_{k}^{\prime}(\bm)),
\end{equation}
where each pixel $\mathbf{x}$ counts the number of warped events that fall within it.
In practice, the Dirac function $\delta$ is replaced by a Gaussian $\mathcal{N}(\mathbf{x}; \mathbf{0},\epsilon^2\text{Id})$ of $\epsilon=1$ pixel width.
Finally, the variance of the IWE
\iflongversion
(or other metrics~\cite{Gallego19cvpr}),
\fi
defines a criterion for model fitting: %
\begin{equation}
\label{eq:argmaxMotionModel}
\bm^\ast = \arg\max_{\bm} \sigma^2(I(\mathbf{x}; \bm)).
\end{equation}
Like~\cite{Gallego18cvpr}, our formulation supports any type of parametric motion model,
such as \mbox{2-DoF} \emph{(degrees-of-freedom)} flow~\cite{Zhu17icra}, \mbox{3-DoF} rotational motion~\cite{Gallego17ral}, \mbox{4-DoF} model~\cite{Mitrokhin18iros}, etc.
\vspace{0.5ex}
In summary, the above event graph representation and model-fitting criterion
allow us to formulate the problem of event-based multi-motion segmentation
as joint estimation over two sets of variables:
\begin{itemize}[noitemsep,nolistsep]
\item \textbf{Discrete labels}.
The segmentation (clustering) is represented by the labeling function $L(e) : \SI{}{\Omega} \times \mathcal{T} \rightarrow \mathcal{L} = \{1,...,N\}$,
which assigns to each event $e$ a label $l \in \mathcal{L}$ indicating which independently moving object the event belongs to.
\item \textbf{Motion models}.
The motion of the independently moving object is represented by a collection of warps with parameters
$\mathcal{M} = \{\bm_1, ..., \bm_N\}$.
Each warp represents the coherent motion of a group of events.
\end{itemize}
| e7f948c4849d5d554af0d2a9fbedb39c6ea1908c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Direct detection of gravitational waves and images of black hole shadows ushers in a golden era of black hole astronomy. At present, these extreme stellar objects serve as our best candidates for testing fundamental quantum gravity theories, such as loop quantum gravity (LQG). LQG, being a non-perturbative approach to quantum gravity, goes beyond general relativity to resolve classical singularities in cosmological and (non-rotating) black hole spacetimes and, in this work, we extend similar techniques to the case of the rotating Kerr-like black hole. Indeed, a consistent LQG black hole (LQGBH) model should not only provide a singularity-free description of the spacetime inside the horizon, it must also have a viable picture for the exterior region with verifiable consequences for these observations.
Due to technical difficulties in solving the LQG equations of motion, especially when using real-valued Ashtekar-Barbero variables, axisymmetric spacetimes have remained largely unexplored. Since this is the class of spacetimes to which the Kerr black hole belongs, therefore, a direct loop quantization of rotating black holes is yet to be achieved (see \cite{LQG_Axi1, LQG_Axi2} for previous attempts). However, from the point of view of phenomenology, this is the primary case of interest since most of the astrophysical black holes which have been observed are those with non-zero angular momenta.
On the other hand, LQG effective equations have been thoroughly investigated for static, spherically symmetric, and non-rotating spacetimes, resulting in quantum extensions of the Schwarzschild black hole (see \cite{Bodendorfer:2019nvy, Bodendorfer:2019jay, LQG1, LQG2, LQG3, LQG4, LQG5, LQG6, LQG7, LQG8, LQG9, LQG10, LQG11, LQG12, Ashtekar:2018lag,Barrau:2018rts} for an incomplete list of these models, \cite{LQGBH_review} for a critical review and \cite{SigChange1,SigChange2} for signature-changing solutions). In this letter, starting from a non-rotating LQGBH \cite{Bodendorfer:2019nvy,Bodendorfer:2019jay}, we construct a rotating spacetime using the Newman-Janis-Algorithm (NJA) \cite{Newman:1965tw}. As a solution-generating method, NJA is successful in constructing the Kerr (Kerr-Newman) solution from the Schwarzschild (Reissner-Nordstr\"om) black hole. We wish to follow a similar strategy for their (loop) quantum counterparts in the hope that such a solution will not only exhibit a non-singular geometry that one expects, but also tell us how LQG effects can be tested in a realistic manner. A priori, it might seem a little \textit{ad hoc} to construct LQG solutions of rotating black holes in this way. However, this is similar in spirit to the ``effective equations'' one typically employs in symmetry-reduced models of LQG (\textit{e.g.} for LQGBHs), which include non-perturbative corrections inspired from the full theory. Analogously, we derive an \textit{effective} rotating, singularity-free spacetime which captures key aspects of LQG.
Previous attempts at generating rotating spacetimes, using NJA, starting from a non-rotating LQGBH, suffer either from using the now-outdated self-dual variables formalism \cite{Liu:2020ola} or an incorrect implementation of NJA \cite{Caravelli:2010ff,AzregAinou:2011fq}. The non-rotating LQGBH \cite{Bodendorfer:2019nvy, Bodendorfer:2019jay} that we are going to consider as the seed metric has several attractive features: In addition to the resolution of classical singularities as is expected to happen in LQG, the quantum effects (quantified by a single parameter) rapidly die out when moving away from the center, with a well-defined asymptotic region in the exterior, a property not shared by all effective models of LQGBHs \cite{Bouhmadi-Lopez:2019hpp}. We will show that the rotating counterpart also retains these characteristics. However, note that our solution is more general and some of its crucial features do not depend on explicit details of the seed metric we have used, thereby capturing some universal properties of rotating LQGBHs.
As we will show, the inclusion of spin naturally enriches the spacetime structure. In particular, it is possible to generate a rotating wormhole without horizon, although this geometry is disfavored by the measurement of the shadow of M87* by the Event Horizon Telescope Collaboration (EHT) \cite{Akiyama:2019cqa}. The most intriguing geometry is a regular black hole containing two horizons, with a timelike transition surface inside the inner one. Such a geometry is observationally favored by the requirement that the quantum parameter is extremely small.
\section{Non-rotating LQGBH}
On solving the LQG effective equations, the quantum extension of the Schwarzschild metric reads \cite{Bodendorfer:2019nvy,Bodendorfer:2019jay}
\begin{equation}
ds^2=-\tilde{a}(x)d\tau^2+\frac{dx^2}{\tilde{a}(x)}+b(x)^2d\Omega_2^2\,.\label{BMMmetric}
\end{equation}
The metric functions are defined in terms of the radial variable $x\in(-\infty,\infty)$ as
\begin{align}
b(x)^2=&\,\frac{A_\lambda}{\sqrt{1+x^2}}\frac{M_B^2\left(x+\sqrt{1+x^2}\right)^6+M_W^2}{\left(x+\sqrt{1+x^2}\right)^3}\,,\label{BMM2metrica1}\\
\tilde{a}(x)=&\,\left(1-\sqrt{\frac{1}{2A_\lambda}}\frac{1}{\sqrt{1+x^2}}\right)\frac{1+x^2}{b(x)^2}\,,\label{BMM2metricfunctions}
\end{align}
where $M_B$ and $M_W$ correspond to two Dirac observables in the model. For convenience, we have defined a dimensionless parameter $A_\lambda\equiv(\lambda_k/M_BM_W)^{2/3}/2$, where the quantum parameter $\lambda_k$ originates from holonomy modifications \cite{Bodendorfer:2019nvy,Bodendorfer:2019jay}. In LQG, these non-perturbative corrections arise from regularizing the curvature components when considering holonomies around loops which can only be shrunk to the minimum non-zero eigenvalue of the area-operator (known as the \textit{area gap}), as opposed to taking the limit to zero as in classical general relativity. One of our main findings is that the quantum parameter, and thereby this fundamental `area-gap', is constrained by observations of shadows of rotating black holes.
The most important feature of this LQGBH \eqref{BMMmetric} is that inside the black hole, the areal radius $b$ reaches a minimum value, representing a \textit{spacelike transition surface} which smoothly connects an asymptotically Schwarzschild black hole to a white hole with mass $M_B$ and $M_W$, respectively. Specifically, we will focus on the physically interesting case of the symmetric bounce in which $M_B=M_W$, \textit{i.e.} the spacetimes are symmetric with respect to the transition surface ($x=0$). Rescaling the coordinates $(x,\tau)\rightarrow(y,t)$ as $y\equiv\sqrt{8A_\lambda}M_Bx$ and $t\equiv\tau/\sqrt{8A_\lambda} M_B$, the metric \eqref{BMMmetric} can be rewritten as
\begin{equation}
ds^2=-8A_\lambda M_B^2\tilde{a}(y)dt^2+\frac{dy^2}{8A_\lambda M_B^2\tilde{a}(y)}+b(y)^2d\Omega_2^2\,.\label{BMMmetricformal1}
\end{equation}
When $|y|\rightarrow\infty$, we have $|y|\rightarrow b$ and $8A_\lambda M_B^2\tilde{a}(y)\rightarrow1-2M_B/b$. Therefore, the metric \eqref{BMMmetricformal} reduces to the Schwarzschild one in the asymptotic limit ($b\rightarrow\infty$).
\section{Rotating LQGBH}
The rotating counterpart of \eqref{BMMmetric} is obtained using NJA, in which the spin parameter $a$ is included through a complex shift on the advanced null coordinates \cite{Newman:1965tw}. In particular, we use the revised NJA \cite{Azreg-Ainou:2014pra} which allows a valid representation of the resulting metric in the Boyer-Lindquist coordinate system $(t,y,\theta,\varphi)$. As a result, the metric of the rotating LQG compact object (rLQGO){\footnote{We use \textit{compact object} instead of {\textit{black hole} because, as will be shown later, the resulting spacetime could be without any trapping regions, in some parts of the parameter space.}} can be cast in a Kerr-like form (see the Appendix for details of the construction)
\begin{align}
ds^2=&-\left(1-\frac{2Mb}{\rho^2}\right)dt^2-\frac{4aMb\sin^2\theta}{\rho^2}dtd\varphi\nonumber\\&+\rho^2d\theta^2+\frac{\rho^2dy^2}{\Delta}+\frac{\Sigma\sin^2\theta}{\rho^2}d\varphi^2\,,\label{finalrot}
\end{align}
where $\rho^2=b^2+a^2\cos^2\theta$, $M=b\left(1-8A_\lambda M_B^2\tilde{a}\right)/2$, $\Delta=8A_\lambda M_B^2\tilde{a}b^2+a^2$, and $\Sigma=\left(b^2+a^2\right)^2-a^2\Delta\sin^2\theta$. Note that the functions $\tilde{a}$, $b$, $M$, and $\Delta$ are functions of $y$, as can be seen from Eqs.~\eqref{BMM2metrica1} and \eqref{BMM2metricfunctions}.
Firstly, note that the metric \eqref{finalrot} reduces to Kerr asymptotically for $|y|\rightarrow\infty$, recovering the expected classical limit, while in the limit $a\rightarrow0$, the static LQGBH \eqref{BMMmetricformal1} is regained. Furthermore, setting $(a, M)\rightarrow 0$ gives one the flat limit \cite{Gan:2020dkb}, satisfying an essential consistency check lacking in some quantum gravity inspired solutions \cite{Hossenfelder:2009fc}. Secondly, $\Delta=0$ defines the event horizon of rLQGO, where the variable $y_h$ satisfies (see Eqs.~\eqref{BMM2metrica1} and \eqref{BMM2metricfunctions}):
\begin{equation}
\sqrt{8A_\lambda+\left(y_h^2/M_B^2\right)}=1\pm\sqrt{1-\left(a^2/M_B^2\right)}\,,\label{ehEX}
\end{equation}
with the plus (minus) sign indicating the outer (inner) horizon on each side of the transition surface. The expression under the radical on the r.h.s. of Eq.~\eqref{ehEX} implies that there is a maximum spin for rLQGO: $|a|\le M_B$, which is the same as the Kerr bound. Evidently, the spacetime structure of rLQGO strongly depends on the values of the parameters $\{a,A_\lambda\}$ under consideration. As illustrated in Figure~\ref{fig.boundary}, the transition surface can either be outside the outer horizon (region I), or between the two horizons (region II), or inside the inner horizon (region III). These regions are split by the boundaries which denote the case when the transition surface is on the outer (red curve) and the inner (blue curve) horizons.
\begin{figure}[t]
\center\includegraphics[scale=0.8]{123boundary}
\caption{\label{fig.boundary}The spacetime structure of the rLQGO metric \eqref{finalrot} with respect to the parameter space $\{a,A_\lambda\}$. In regions I, II, and III, the transition surface $(y=0)$ is located outside the outer horizon, between the inner and outer horizons, and inside the inner horizon, respectively. On the red (blue) curve, the transition surface is located on the outer (inner) horizon.}
\end{figure}
\textbf{\textit{Region I}} -- In this region of parameter space, the rLQGO is a rotating wormhole (Figures~\ref{fig.emb}(a) and \ref{fig.penrose1}(a)) without horizon. Its spacetime structure resembles that of the phenomenological Kerr-like wormhole proposed in Ref.~\cite{Bueno:2017hyj}. However, the Arnowitt-Deser-Misner mass of rLQGO \eqref{finalrot} is always $M_B$ (see the Appendix), while that of the model in Ref.~\cite{Bueno:2017hyj} depends explicitly on the throat parameter. Note that the ringdown signals generated by this type of wormholes are characterized by echos \cite{Bueno:2017hyj}.
\textbf{\textit{Region II}} -- The transition surface is hidden behind the outer event horizon and becomes spacelike (Figure~\ref{fig.emb}(b)). The green region is inside the event horizon where $t$ and $y$ exchange roles and the transition surface is located at the narrowest point in the middle. As expected, the Penrose diagram for this type of rotating black holes is similar to that of its non-rotating counterpart \cite{Bodendorfer:2019nvy,Bodendorfer:2019jay} (Figure~\ref{fig.penrose1}(b)), rendering the inner horizon irrelevant. This is because as $a/M_B \rightarrow 0$, the rLQGO tends to be in region II, as long as the transition surface is hidden by the outer horizon.
\textbf{\textit{Region III}} -- Given a non-zero finite value of $a/M_B$, this region is characterized by a small $A_\lambda$ and, thus, is the most physically relevant one for considering rotating black holes. The classical ring singularity behind the Cauchy horizon of the Kerr black hole is replaced by a timelike transition surface. As shown in its Penrose diagram (Figure~\ref{fig.penrose3}), a timelike trajectory (the black dashed curve) entering the black hole crosses the inner horizons (blue solid lines). Thereafter, this trajectory can be extended into another universe either by going upwards (trajectory A) without touching the transition surface (blue dashed lines), or by crossing the transition surface into another interior patch (trajectory B). Its embedding diagram (Figure~\ref{fig.emb}(c)) has to terminate at a timelike surface outside the transition surface since it cannot be extended vertically downward any further (the surface of the cone becomes horizontal).
\begin{figure}[t]
\center(a)\includegraphics[scale=0.35]{embregion1}
(b)\includegraphics[scale=0.35]{embregion2}
(c)\includegraphics[scale=0.35]{embregion3}
\caption{\label{fig.emb}The embedding diagram of rLQGO. (a): A timelike wormhole without horizon (region I). (b): A spacelike transition surface inside the event horizon (region II). (c): The transition surface is inside the inner horizon (region III).}
\end{figure}
\begin{figure}
\center(a)\includegraphics[scale=0.3]{figWHpen}
(b)\includegraphics[scale=0.2]{figBHpen}
\caption{\label{fig.penrose1}The Penrose diagram of rLQGO in (a): region I and (b): region II. Blue dashed lines denote the transition surface and slanted red lines, at an angle of $45^{\circ}$, are event horizons.}
\end{figure}
\begin{figure}
\center\includegraphics[scale=0.18]{figINWHpen}
\caption{\label{fig.penrose3}The Penrose diagram of rLQGO in region III. The colored regions are inside the outer event horizon (red lines). After entering inner horizons (solid blue lines), a timelike trajectory can be extended to another universe either by going upward (trajectory A) without touching the transition surface (dashed blue lines), or by crossing the transition surface to another interior patch (trajectory B). Note that the exterior regions of the two adjacent interior patches can be causally disconnected.}
\end{figure}
Importantly, the rLQGO is free from spacetime singularities. As shown in Figure~\ref{fig.Ricci}, the Ricci scalar is finite everywhere on the $(y,\cos\theta)$ plane, and rapidly vanishes when moving away from the transition surface. In this figure, the solid red and blue lines represent the outer and inner horizons, respectively. The dashed curves are the ergosurface. Interestingly, since the areal radius $b\neq0$ for rLQGO, there is no closed timelike curve which usually appears near the ring singularity inside the Kerr black hole (see the Appendix for details). Even though the event horizon disappear when $|a|>M_B$, the absence of singularity naturally preserves the Cosmic censorship hypothesis and the configuration looks like a ``naked bounce'' (similar to the rLQGO solution in Region I).
\begin{figure*}[t]
\center\includegraphics[scale=0.35]{Ricci91}(a)
\includegraphics[scale=0.35]{Ricci92}(b)
\includegraphics[scale=0.35]{Ricci93}(c)
\caption{\label{fig.Ricci}The Ricci scalar $R$ of the rLQGO spacetime expressed in the $(y,\cos\theta)$ plane. In the figure, we set $M_B=1$ and $a/M_B=0.9$. The solid red and blue lines represent the outer and the inner event horizons. The dashed curves represent the ergosurface. (a): A wormhole without horizon (region I with $A_\lambda=0.4$). (b): The transition surface is covered by the outer horizon (region II with $A_\lambda=0.1$) (c): The transition surface is inside the inner horizon (region III with $A_\lambda=0.01$).}
\end{figure*}
\section{Astrophysical implications}
In addition to having properties of, \textit{e.g.} asymptotic flatness and regularity, we find that both the geodesic equations and the Klein-Gordon equation of the rLQGO allow for a complete separation of variables (see the Appendix), following the criteria of Refs.~\cite{Shaikh:2019fpu,Chen:2019jbs}. The separability of the geodesic equations is useful in testing the rLQGO spacetime with its shadow and the orbital motion of surrounding particles while the separability of the Klein-Gordon equation is helpful for studying the scattering problem and the quasinormal modes \cite{prepare}.
As an example, let us demonstrate that, it is possible, in principle, to constrain the quantum parameter $A_\lambda$ using the shadow image cast by the M87*. In particular, we find that the effects made by the parameter $A_\lambda$ on the shadow size $R_S/M_B$ are more significant than those on the non-circularity of the shadow contour. Provided that the shadow size cast by M87* is consistent with that of Kerr black hole within $17\%$ at $1\sigma$ level \cite{Akiyama:2019eap,Psaltis:2020lvx}, one can see from Figure~\ref{fig.shadowRS} that the parameter space corresponding to the wormhole geometry (the region on the right of the red curve) is disfavored by the bound from the $R_S$ measurement (black curve). Since the quantum parameter which enters the effective equations in LQG is directly related to the fundamental \textit{area gap} in the theory, shadows of rotating black holes give us a new way to constrain this parameter from observations. (Note that the quantum parameter is more tightly constrained by Solar System tests, $A_\lambda < 7.7\times 10^{-5}$ \cite{Williams:2004qba}. However, this assumes the validity of Birkhoff's theorem which need not hold in LQG.) This paves a novel method for deriving state-of-the-art constraints on LQG by examining observational consequences of rLQGO \cite{prepare}.
\begin{figure}[t]
\center\includegraphics[scale=0.71]{shadowRS2}
\caption{\label{fig.shadowRS}The apparent size $R_S/M_B$ of the shadow cast by rLQGO is shown with respect to the parameter space $\{a,A_\lambda\}$. The $17\%$ bound of $R_S/M_B$ (black curve) inferred from the M87* shadow disfavors parameters corresponding to the wormhole geometry (the region to the right of the red curve). Here we have taken into account the spin measurement obtained using the radio intensity data \cite{Tamburini:2019vrf} (cyan lines), and the inclination angle measured by the jet direction \cite{Walker:2018muw}.}
\end{figure}
\section{Universal features}
The most obvious limitation of our approach is that the resulting rLQGO metric is not derived by a direct loop quantization of the Kerr (or, more generally, axisymmetric) spacetime. How much of our results do we expect to generalize to such a scenario, and not be tied to the seed metric that we have chosen? Firstly, note that the existence of a \textit{spacelike} transition surface is very common for non-rotating LQGBHs, irrespective of quantization ambiguities (such as choosing the $\mu_0$ or the $\bar\mu$ scheme). Since this is the most crucial feature of the seed metric we have used in our construction, it is natural to expect that our rLQGO solution correctly captures the effective spacetime description of rotating LQGBHs, as long as we expect LQG to provide singularity-resolution of rotating black holes in a way such that there is a smooth bouncing geometry bridging black and white holes. Furthermore, our results indicate that such a geometry observationally favors having the transition surface inside the inner horizon, and is automatically consistent with the expectation that the quantum parameter is small (it inherits this property from the tiny area gap $\propto\ell_{\rm Pl}^2$). Interestingly, observations also seem to rule out models of non-rotating LQGBH spacetimes which describe a bounce \textit{outside} the event horizon \cite{Haggard:2014rza}, since their rotating counterparts are at odds with observations, as well as prefer non-rotating models which allow for an inner horizon \cite{LQG2, LQG4}.
To make our point more explicit, we present the result of the NJA analyses on another non-rotating LQG metric proposed in \cite{LQG2,LQG3} (see the Appendix and \cite{prepare} for details). In Figure~\ref{fig.edshadow}, we show the apparent size $R_S/M_B$ of the shadow cast by the rotating metric, which is obtained from \cite{LQG2,LQG3} using NJA, in the parameter space of $\{a,\tilde{\Delta}\}$. Firstly, we note that the spin and the quantum parameter $\tilde{\Delta}$ both shrink the shadow size, exactly similar to the rLQGO case. Secondly, the black boundary represents a spin-dependent upper bound of the quantum parameter, above which the object cannot cast shadows. In the non-rotating limit, this upper bound can be explicitly derived as $\gamma^2\tilde{\Delta}/M_B^2<3^6/2^{10}\approx0.71$, where $\gamma$ is the Immirzi parameter and $\tilde{\Delta}$ is the area-gap in LQG which is directly constrained in this model. Thus, we find other models of regular LQGBHs also support our general finding that the area-gap is constrained to be small from observations.
\begin{figure}[h]
\center\includegraphics[scale=0.71]{edshadow2}
\caption{\label{fig.edshadow}The apparent size $R_S/M_B$ of the shadow cast by the rotating black hole metric, corresponding to \cite{LQG2,LQG3}, in the parameter space of $\{a,\tilde{\Delta}\}$. The black boundary represents a spin-dependent upper bound of the quantum parameter, above which shadow contours disappear.}
\end{figure}
\section{Discussions}
The construction of rotating LQGBHs from holonomy-corrected effective equations in LQG is still an open problem. In order to catch up with the rapidly developing astronomical observations of spinning black holes in the coming years, there is an urgent need for a model of rotating LQGBH. To derive this, an alternate path is to use a viable solution-generating method to generate a rotating solution from a non-rotating LQGBH seed metric.
The resulting rLQGO spacetime \eqref{finalrot}, based on the seed metric \cite{Bodendorfer:2019jay, Bodendorfer:2019nvy}, possesses a rather simple expression and has several interesting properties. It is everywhere non-singular and it reduces to Kerr solution asymptotically. The geodesic equations and the Klein-Gordon equation both allow complete separations of variables. Most importantly, as in the static LQGBH, the rLQGO is characterized by the existence of a transition surface induced from non-perturbative quantum corrections. Depending on the relative location of the transition surface with respect to the two horizons, the rLQGO can represent a wormhole, a regular black hole with an interior spacelike transition surface, or a regular black hole with a timelike transition region inside the inner horizon. We show that the possibility of rLQGO being a wormhole without horizon has been almost \textit{ruled out} by the shadow size of M87* measured by EHT.
Most significantly, our works fills a lacuna between theoretical quantum gravity extensions of black holes, which have been mostly applied to non-rotating spacetimes, and experimental observations which have been of spinning black holes. Remarkably, we find that not only is it possible to find a regular effective description of rotating black holes from LQG, an extension to such backgrounds leads to observable effects which can rule out some proposals of loop quantization for non-rotating black holes while providing support for other more generic ones which capture some universal features of LQG.
\section*{Acknowledgements} SB is grateful to Jibril Ben Achour, Norbert Bodendorfer and Johannes M\"unch for comments on an earlier version of this draft. SB is supported in part by the NSERC (funding reference \#CITA 490888-16) through a CITA National Fellowship and by a McGill Space Institute fellowship. CYC is supported by Institute of Physics in Academia Sinica, Taiwan. DY is supported by the National Research Foundation of Korea (Grant no. 2018R1D1A1B07049126).
| 5c5df2cb8f4164ee9fd0c0d9a31ca20353622e2a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
GNNs \cite{scarselli2008graph}, especially graph convolutional networks \cite{bruna2013spectral,henaff2015deep} have demonstrated remarkable performance in modeling structured data in a wide variety of fields, such as social networks \cite{kipf2016semi,hamilton2017inductive, li2019semi} and graph-based representations of molecules \cite{gilmer2017neural,rong2020self}.
The common practice is to recursively update node embeddings by aggregating (or message passing) information from topological neighbors such that the GNNs can capture the local structure of nodes.
Subsequently, the learned embeddings can be used in downstream analyses, e.g., node classification \cite{kipf2016semi,hamilton2017inductive,huang2018adaptive,xu2018representation,rong2019dropedge}, link prediction \cite{zhang2018link}, and graph classification \cite{duvenaud2015convolutional,dai2016discriminative,gilmer2017neural}.
However, those GNNs fail to capture the hierarchical representations of graphs \cite{ying2018hierarchical}, which is essential for many scenarios. For instance, in order to predict the properties of a given molecule, it would be highly desirable to infer the sub-parts which are important for the molecular properties hierarchically.
To this end, various graph pooling methods are recently proposed, aiming to learn the coarse-grained graph structure by either reserving the most informative nodes \cite{gao2019graph,lee2019self} or aggregating nodes belonging to the same cluster \cite{ying2018hierarchical,yuanstructpool,khasahmadi2020memory,wang2019haarpooling,Bianchi2020spectral}.
In particular, the latter attracts considerable attention mainly attributed to its remarkable performance.
Such kind of methods learn a cluster assignment matrix to map each node to a set of clusters that may correspond to strongly connected communities within a social network or functional groups within a molecule.
However, their limitations lie in that (i) simply grouping node features fails to effectively model the part-whole relationship that is crucial in characterizing the hierarchical structure, and (ii) they ignore the entanglement of the latent factors behind node embeddings, resulting in limited capacity in preserving detailed node/graph properties and modeling graph hierarchy.
For example, it is of particular importance to consider the interaction of heterogeneous factors (e.g., work, hobby) underlying each node, in order to identify the communities in a social network.
Capsule neural networks (CapsNets) have proved its effectiveness in modeling hierarchical relationships on image data by exploiting the fact that while viewpoint changes have complicated effects on pixel intensities, they have linear effects at the part/object level \cite{sabour2017dynamic,hinton2018matrix,kosiorek2019stacked}.
In contrast to convolutional neural networks, CapsNets use activity vectors or pose matrices to represent the entities.
Moreover, the viewpoint-invariant relationship between the part and the whole is characterized by trainable transformation matrices, which is under the assumption that the human visual system relies on parse tree-like structure to recognize objects.
Such representations make CapsNets especially appealing in reasoning the part-whole hierarchy and robust to adversarial attacks \cite{hinton2018matrix,qin2019detecting}.
However, how to effectively take advantage of CapsNets to benefit graph classification remains largely unexplored.
In this work, we present the hierarchical graph capsule network (HGCN) that is able to jointly learn node embeddings and extract the hierarchical structure of the input graph.
Specifically, to preserve detailed node/graph information, we build graph capsules by disentangling heterogeneous factors behind the node embeddings such that each capsule encodes different properties of the same entity.
In order to capture the graph hierarchy, multiple graph capsule layers are stacked to get coarser and coarser representations.
In each layer, (i) to infer the votes of instantiation parameters of higher-level capsules (wholes), we propose transformation GNNs to reason about the part-whole relationship by explicitly considering the structure information among lower-level capsules (parts);
(ii) each of these votes that are weighted by a routing weight, are iteratively routed to the potential wholes that correspond to tight clusters in the votes.
We further introduce the auxiliary graph reconstruction to enhance the representation capacity of graph capsules and the training stability.
As a consequence, HGCN is capable of modeling hierarchical representations of the input graph and benefits the goal of graph classification.
Our main contributions can be summarized as:
(i) we propose a novel capsule-based graph neural network to learn node embeddings and hierarchical graph representations simultaneously,
(ii) we demonstrate the effectiveness of considering the entanglement of latent factors and the structure information within the parts in modeling part-whole relationships on the graph data;
(iii) comprehensive empirical studies demonstrate that our method achieves remarkably superior improvement over the state-of-the-art approaches on 11 commonly used benchmarks.
\section{Related Work}
\iffalse
EigenPool introduced a graph pooling that used the local graph
Fourier transform to extract subgraph information. Its potential drawback lies in the inherent computing bottleneck for the Laplacian-based graph Fourier transform, given the high
computational cost for the eigendecomposition of the graph Laplacian.
EigenPooling \cite{ma2019graph} summarize the subgraph information by also considering the subgraph structure; e local graph Fourier transform
MinCutPool \cite{Bianchi2020spectral} spectral clustering; is a well-known clustering technique that leverages the Laplacian spectrum to find strongly connected communities on a graph; eigendecomposition of the Laplacian is expensive; does not require to compute the spectral decomposition;
Haar graph pooling \cite{wang2019haarpooling} (ICML 2020): compressive Haar transforms; it is computed by following a sequence of clusterings of the input graph; the compressive Haar transform filters out fine detail information in the Haar wavelet domain.
GMN \cite{khasahmadi2020memory}: these models are not efficient as they require an iterative
process of message passing after each pooling layer; cluster-based; assignment matrix-based
\fi
\subsubsection{Graph Neural Networks}
GNNs attempt to exploit the structure information underlying graph structured data in order to benefit various downstream tasks \cite{li2015gated,kipf2016semi,hamilton2017inductive,li2018adaptive,xu2018representation,luan2019break,rong2019dropedge,you2019position}.
Recent studies have proved GNNs' wide applicability in, for example, drug discovery \cite{yan2020retroxpert,gilmer2017neural,ma2020multi}, protein interface prediction \cite{fout2017protein}, and recommendation system \cite{ying2018graph}.
Current GNNs can be mainly summarized as two streams: spectral-based methods and spatial-based methods.
The spectral-based methods largely rely on the convolution operation defined in the Fourier domain with spectral filters \cite{bruna2013spectral,henaff2015deep}.
This kind of method is further simplified and extended by introducing polynomial spectral filters \cite{defferrard2016convolutional} or linear filters \cite{kipf2016semi}.
To deal with arbitrarily structured graphs, the spatial-based methods define convolutions directly on the graph by aggregating features from topological neighbors \cite{atwood2016diffusion,niepert2016learning,hamilton2017inductive,vaswani2017attention}.
\subsubsection{Graph Pooling}
It is widely recognized that pooling operation plays an important role in graph classification \cite{errica2019fair} which requires the graph level representation.
The most straightforward way is, to sum up, or take an average of all node features \cite{hamilton2017inductive,xu2018powerful}.
The limitation of such strategy is that the hierarchical information which is crucial in capturing graph structure is not considered.
Inspired by the downsampling in convolutional neural networks, recent studies propose to adaptively keep the most informative nodes in a hierarchical manner \cite{gao2019graph,lee2019self}, or aggregate maximal cliques by only using topological information \cite{luzhnica2019clique}.
Another line of work focuses on finding strongly connected communities on a graph.
This is typically achieved by learning a cluster assignment matrix in order to map each node to a set of clusters \cite{ying2018hierarchical,ranjan2019asap,yuanstructpool,khasahmadi2020memory}.
Most recent studies approach this problem by leveraging more advanced clustering techniques, such as local graph Fourier transform \cite{ma2019graph}, spectral clustering \cite{Bianchi2020spectral}, and compressive Haar transforms \cite{wang2019haarpooling}.
However, simply grouping node features has limited capacity in modeling the part-whole relationships, especially for biological data.
In this work, we reason about the part-whole hierarchy by exploring the interaction of underlying latent factors and structure information among the parts, then use an routing mechanism to assign the parts to wholes.
\subsubsection{Capsule Networks}
A capsule \cite{hinton2011transforming} is a group of neurons whose orientation represents the instantiation parameters such as pose (position, size) of an entity (e.g., an object).
The probability that the entity exists can be represented by the capsule length \cite{sabour2017dynamic} or a logistic unit \cite{hinton2018matrix,kosiorek2019stacked}.
Compared to a single neuron, a capsule contains different properties of the same entity and can preserve hierarchical relationships between lower-level capsules (e.g., eyes, mouth) and higher-level capsules (e.g., face).
Such part-whole relationships are described by trainable transformation matrices which are viewpoint-invariant.
Concretely, a lower-level capsule (part) makes predictions for the pose of each higher-level capsule (whole) by multiplying its own pose by the transformation matrices.
Routing-by-agreement is then performed between two adjacent capsule layers to update the probability with which a part is assigned to a whole \cite{sabour2017dynamic,hinton2018matrix}.
Inspired by this, GCAPS-CNN \cite{verma2018graph} builds capsules on graphs by considering higher-order statistical moments as instantiation parameters.
Normal graph convolution is then carried out to aggregate information from neighbors and the covariance is computed as the permutation invariant feature for graph classification.
Its most obvious drawback lies in that the hierarchical structure of the graph is not considered.
Different from GCAPS-CNN, we explicitly take into account the hierarchy between two consecutive capsule layers through trainable transformation GNNs.
CapsGNN \cite{xinyi2018capsule} uses multiple GNN channels to build graph capsules and follow the same voting strategy as \cite{sabour2017dynamic} to predict higher-level capsules.
However, simply using the transformation matrix ignores the local structure information among lower-level capsules and fails to describe part-whole relationships in graphs.
Our method introduces transformation GNNs to reason about the pose of each whole in the layer above, which is in contrast to each individual part making its own prediction.
A further advantage of transformation GNNs is that they save orders of magnitude of memory compared to transformation matrices used in previous work \cite{xinyi2018capsule,sabour2017dynamic,hinton2018matrix}.
Furthermore, different from CapsGNN that reconstructs the histogram of input nodes, we reconstruct the adjacency matrix of the input graph to ensure the quality of graph capsules and enhance the training stability.
\iffalse
\paragraph{Disentangled Representation Learning}
We dynamically disentangle the node representations
1. PAYLESS ATTENTION WITH LIGHTWEIGHT AND DYNAMIC CONVOLUTIONS
2. dynamic convolution: attention over convolution kernels
3. Disentangled graph convolutional networks \cite{ma2019disentangled}
\fi
\section{Preliminaries}
\subsubsection{Graph Classification}
A graph $G$ with $N$ nodes is represented as $(\mathbf{A},\mathbf{X})$, where $\mathbf{A}\in\{0,1\}^{N \times N}$ is the adjacency matrix, and $\mathbf{X} \in \mathbb{R}^{N \times d}$ is the node feature matrix with feature dimension $d$.
Given a set of labeled graphs $\mathcal{D}=\{(G_1, y_1),(G_2, y_2),...\}$, the goal of graph classification is to learn a mapping $f:\mathcal{G} \rightarrow \mathcal{Y}$, where $G_i \in \mathcal{G}$ and $y_i \in \mathcal{Y}$.
For example, each graph is a molecule, and its label indicates whether it is toxic.
\subsubsection{Graph Neural Networks}
To extract useful information from local neighborhoods, our method is built upon GNNs by following the general "message-passing" paradigm, which is formulated as:
\begin{equation}
\mathbf{H}^{(l+1)}=\mathcal{M}(\mathbf{A},\mathbf{H}^{(l)};\mathbf{W}^{(l)}),
\end{equation}
where $\mathcal{M}$ indicates the message passing function with various possible implementations \cite{kipf2016semi, hamilton2017inductive}, $\mathbf{W}^{(l)}$ is learnable weight matrix, $\mathbf{H}^{(l+1)}$ and $\mathbf{H}^{(l)}$ are the node embeddings of layer $l+1$ and $l$, respectively.
The input node embeddings $\mathbf{H}^{(1)}$ are initialized using the node feature $\mathbf{X}$, i.e., $\mathbf{H}^{(1)}=\mathbf{X}$.
The final node representations are denoted by GNN$(\mathbf{A}, \mathbf{X})=\mathbf{H}^{(L_\text{GNN})}\in \mathbb{R}^{N \times h}$ with $L_\text{GNN}$ iterations.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=1\linewidth]{framework-eps-converted-to.pdf}
\end{center}
\caption{An overview of the proposed framework. (A) Given an input graph, we build graph capsules by learning disentangled node representations in order to take into account the heterogeneous factors behind each node. TGNNs are established to characterize the part-whole relationship, and a routing strategy is used to predict higher-level capsules that receive a cluster of similar votes. (B) The residual connection that combines fine, low layer information with coarse, high layer information. }
\label{fig:framework}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=0.85\linewidth]{routing-eps-converted-to.pdf}
\end{center}
\caption{Cluster by agreement.}
\label{fig:routing}
\end{figure}
\section{Methodology}
In this section, we begin by first briefing the proposed HGCN as shown in Figure~\ref{fig:framework}, then we detail each component in the following sections.
The goal of HGCN is to jointly learn node embeddings and coarsen the graph through exploiting hierarchical information.
To achieve this, we disentangle node representations to build the graph capsule by considering heterogeneous factors underlying each edge connection.
Therefore, each graph capsule is composed of multiple independent latent factors that represent different properties of the same entity.
To learn hierarchical graph representations, transformation GNNs (TGNNs) are proposed to encode the part-whole relationship between lower-level and higher-level graph capsules.
Specifically, a capsule in one layer votes for the instantiation parameters of each capsule in the layer above through TGNNs which highly depend on the structure information of lower-level capsules.
Each of these votes is then routed to a higher-level capsule that receives a cluster of similar votes by a routing-by-agreement strategy.
To encourage the graph capsules to encode the instantiation parameters of the input graph and also enhance the training stability, we further introduce the auxiliary graph reconstruction to reconstruct the input adjacency matrix.
\subsection{Disentangled Graph Capsules}
In most cases, highly complex interactions are involved in the connection between each node pair in a graph.
For example, the edges between a node and its neighbors in a social network are driven by heterogeneous factors, since a person connects with others for various reasons such as exercise, work, etc.
Therefore, it is necessary to disentangle the explanatory factors of variations underlying the node representations.
Furthermore, each node embedding is considered as multiple individual scalar features in existing GNNs, which are proved to have limited capability in preserving the graph properties \cite{verma2018graph,xinyi2018capsule}.
To address these two limitations, motivated by \cite{sabour2017dynamic}, we propose graph capsules to describe the given graph.
Specifically, we disentangle the latent factors of each node embedding and use the disentangled node representation to represent graph capsules (Figure~\ref{fig:framework}A).
In this way, each graph capsule is composed of multiple heterogeneous factors, and each factor describes a specific instantiation parameter of the entity/node.
Formally, given $G=(\mathbf{A}, \mathbf{X})$, the node $i$ is denoted by $\mathbf{x}_i \in \mathbb{R}^d$.
We project the input node features into $K$ different subspaces, assuming that there are $K$ latent factors/instantiation parameters:
\begin{equation}
\mathbf{z}_{i,k}=\sigma (\mathbf{W}_k^T \mathbf{x}_i) + \mathbf{b}_k,
\end{equation}
where $\mathbf{W}_k \in \mathbb{R}^{d \times \frac{h}{K}}$ and $\mathbf{b}_k \in \mathbb{R}^{\frac{h}{K}}$ are learnable parameters, $\sigma$ is a nonlinear activation function, and $\frac{h}{K}$ is the dimension of each factor.
Although more sophisticated implementations of node disentanglement are possible \cite{ma2019disentangled}, we use linear projection in our study attributed to its efficiency and remarkable performance.
Therefore, each graph capsule is represented by a pose matrix $\mathbf{Z}_i \in \mathbb{R}^{K \times \frac{h}{K}}$ \cite{hinton2018matrix}. For simplicity, we reshape $\mathbf{Z}_i$ to the vector format $\mathbf{z}_i \in \mathbb{R}^h$.
Recall that the existence probability of an entity represented by a capsule is measured by the capsule length \cite{sabour2017dynamic}, we thus squash $\mathbf{z_i}$ as follows:
\begin{equation}
{\mathbf{p}}_{i}= squash(\mathbf{z}_{i}) = \frac{\|\mathbf{z}_{i}\|^2}{1+\|\mathbf{z}_{i}\|^2} \frac{\mathbf{z}_{i}}{\|\mathbf{z}_{i}\|},
\end{equation}
where $\mathbf{u}_i^{(1)} = {\mathbf{p}}_{i} \in \mathbb{R}^{h}$ is the primary graph capsule representing the lowest level of entities, such as atoms in the molecular graph.
\subsection{Hierarchical Capsule Layers}
To obtain hierarchical graph representation, it is essential to capture the part-whole relationship between adjacent capsule layers.
Such relationship is measured by viewpoint-invariant transformation matrix $\mathbf{T}^{(l)}_{i,j} \in \mathbb{R}^{d_{l} \times d_{l+1}}$ for each pair of lower-level capsule $\mathbf{u}^{(l)}_i$ and higher-level capsule $\mathbf{u}^{(l+1)}_j$ in previous studies \cite{sabour2017dynamic,hinton2018matrix}, where $d_l$ and $d_{l+1}$ are the capsule dimensions of $\mathbf{u}^{(l)}_i$ and $\mathbf{u}^{(l+1)}_j$, respectively.
However, $\mathbf{T}^{(l)}_{i,j}$ totally ignores the structure information within $\mathbf{u}^{(l)}$, which is especially problematic for graph structured data.
Furthermore, $\mathbf{T}^{(l)} \in \mathbb{R}^{N_{l} \times N_{l+1} \times d_{l} \times d_{l+1}}$ is extremely memory-consuming for the scenario where a large number of high-dimensional capsules are required.
To overcome these difficulties, we propose the transformation GNNs (TGNNs) to vote for the instantiation parameters of higher-level graph capsules (\textbf{voting}).
When multiple votes agree, a higher-level capsule that receives a cluster of similar pose votes becomes active (\textbf{routing}).
More concretely, we denote the graph capsules at layer $l$ as $\mathbf{u}^{(l)} \in \mathbb{R}^{N_l \times d_l}$, the capsule number as $N_l$, and the adjacency matrix as $\mathbf{A}^{(l)}$.
Our goal is to decide which capsules to activate in $\mathbf{u}^{(l+1)} \in \mathbb{R}^{N_{l+1} \times d_{l+1}}$ and how to assign each active lower-level capsule $\mathbf{u}^{(l)}_i$ to one active higher-level capsule $\mathbf{u}^{(l+1)}_j$.
In practice, we set $N_{l+1} < N_l$ in order to get coarser and coarser graph representations (Figure~\ref{fig:framework}A).
\subsubsection{Voting} For all capsules in $\mathbf{u}^{(l)}$, their poses are transformed by TGNNs to cast votes for the pose of each capsule in $\mathbf{u}^{(l+1)}$ by the following equation,
\begin{equation}
\mathbf{v}^{(l)}_j = \text{TGNN}_j(\mathbf{A}^{(l)}, \mathbf{u}^{(l)}),
\end{equation}
where ${\mathbf{v}}^{(l)}_j \in \mathbb{R}^{N_l \times d_{l+1}}$. Specifically, ${\mathbf{v}}^{(l)}_{j|i} \in \mathbb{R}^{d_{l+1}}$ is the vote for the pose of $\mathbf{u}^{(l+1)}_j$ predicted by the capsule $\mathbf{u}^{(l)}_i$.
Note, TGNNs are learned discriminatively and could learn to represent part-whole relationships by considering the structure information of capsules in $\mathbf{u}^{(l)}$.
This is different from previous studies that use one transformation matrix for each pair of $(\mathbf{u}^{(l)}_i, \mathbf{u}^{(l+1)}_j)$.
Compared to transformation matrices, TGNNs also save $N_l$ orders of magnitude memory.
\subsubsection{Routing} Each of these votes is then weighted by an routing weight $c^{(l)}_{i,j}$ with which a part is assigned to a whole, where $c^{(l)}_{i,j} \geqslant 0$ and $\sum_{j=1}^{N_{l+1}} c^{(l)}_{i,j}=1$.
Here, $c^{(l)}_{i,j}$ is iteratively updated using an "routing-by-agreement" mechanism such that each vote in ${\mathbf{v}}^{(l)}$ is routed to a capsule in $\mathbf{u}^{(l+1)}$ that receives a cluster of similar votes (Figure~\ref{fig:routing}).
Formally, $c^{(l)}_{i,j}$ is defined as $c^{(l)}_{i,j} = {\text{exp}(b^{(l)}_{i,j})}/{\sum_k \text{exp}(b^{(l)}_{i,k})}$, where $b^{(l)}_{i,j}$ is initialized as $b^{(l)}_{i,j}=0$.
To iteratively search for the vote cluster, in each iteration we have,
\begin{equation}
\mathbf{u}^{(l+1)}_j = squash(\sum_i c^{(l)}_{i,j}\mathbf{v}^{(l)}_{j|i})
\end{equation}
where $\mathbf{u}^{(l+1)}_j$ is the predicted capsule $j$ in layer $l+1$, representing a tight cluster of votes from layer $l$.
We update $b^{(l)}_{i,j}$ with $b^{(l)}_{i,j} = b^{(l)}_{i,j} + a^{(l)}_{i,j}$, where $a^{(l)}_{i,j} = \mathbf{v}^{(l)}_{j|i} \cdot \mathbf{u}^{(l+1)}_j$ indicates the agreement between each vote and vote cluster.
It is worth mentioning that such top-down feedback also has a beneficial effect on the aggregation in the proposed TGNNs, such that TGNNs can more focus on aggregating information from neighbors that are likely to be in the same cluster.
After $R$ iterations, we get higher-level graph capsules $\mathbf{u}^{(l+1)}$ and the coarsened adjacency matrix defined as $\mathbf{A}^{(l+1)}={\mathbf{C}^{(l)}}^T \mathbf{A}^{(l)} \mathbf{C}^{(l)} \in \mathbb{R}^{N_{l+1} \times N_{l+1}}$.
As opposed to generating structurally independent higher-level capsules in previous work \cite{sabour2017dynamic, xinyi2018capsule}, the capsule layer we developed is able to explicitly preserve the structure information which is encoded in $\mathbf{A}^{(l+1)}$.
Drawing inspiration from \cite{long2015fully}, we add a residual connection at each pair of consecutive capsule layers, aiming to provide fine-grained information to higher-level capsules (Figure~\ref{fig:framework}B).
Formally, we have $\mathbf{u}^{(l+1)} \leftarrow \mathbf{u}^{(l+1)} + \text{GA}(\mathbf{u}^{(l)})$, where $\text{GA}$ indicates the global average operation.
By stacking multiple capsule layers, we get the class capsules $\mathbf{u}^{(L)} \in \mathbb{R}^{O \times d_L}$ which are intended to encode feature attributes corresponding to the class, where $O$ is the number of graph categories.
The classification loss is measured by a margin loss \cite{sabour2017dynamic} which is formulated as:
\begin{equation}
\begin{aligned}
\mathcal{L}_{m}(\mathbf{A}, \mathbf{X}) = {} &
\sum\limits_{o \in O} [T_o \hspace{0.1cm} \text{max}(0, m^+ - \|\mathbf{u}^L_o\|)^2 + \\ & \lambda(1-T_o) \hspace{0.1cm} \text{max}(0, \|\mathbf{u}^L_o\|-m^-)^2],
\end{aligned}
\end{equation}
where $m^+=0.9$, $m^-=0.1$, $T_o=1$ iff the input graph has label $o$, and $\lambda$ is ued to stop the initial learning from shrinking the length of class capsules.
\subsection{Auxiliary Graph Reconstruction}
To encourage the graph capsules to encode the instantiation parameters of the input graph and to improve the training stability, we introduce a reconstruction loss to constraint the capsule reconstruction to closely match the class-conditional distribution.
Specifically, we first mask out all but the winning capsule (the capsule corresponds to ground truth) and combine them with primary capsules by following the equation,
\begin{equation}
\mathbf{Z} = \mathbf{u}^{(1)} + (\mathbf{W}_r^T \Phi(\mathbf{u}^{(L)}) + \mathbf{b}_r),
\end{equation}
where $\mathbf{Z} \in \mathbb{R}^{N \times d_1}$, $\Phi$ is the mask operation, $\mathbf{W}_r \in \mathbb{R}^{(O \times d_L) \times d_1}$, and $\mathbf{b}_r \in \mathbb{R}^{d_1}$.
The reconstruction loss is then defined as,
\begin{equation}
\mathcal{L}_{r}(\mathbf{A}, \mathbf{X}) = -\frac{1}{N^2}\sum_{j=1}^N \sum_{k=1}^N\sum_{c\in\{0,1\}} \mathbf{A}^{(j,k,c)}log(\mathbf{\tilde{A}}^{(j,k,c)}),
\end{equation}
where $\mathbf{\tilde{A}}=\mathbf{Z} \mathbf{Z}^T$ is the reconstructed adjacency matrix of the input graph.
Taken together, we reach the optimization objective of our method as $\min\limits_{\theta} \sum\limits_{G \in \mathcal{D}} \mathcal{L}_m(\mathbf{A}, \mathbf{X}) + \beta \mathcal{L}_r(\mathbf{A}, \mathbf{X})$, where $\theta$ are all learnable parameters, and $\beta$ leverages the importance of $\mathcal{L}_r$.
The whole training process is detailed in Algorithm~\ref{alg:alg_1}.
\setlength{\textfloatsep}{10pt}%
\begin{algorithm}[t]
\SetAlgoLined
\textbf{Input:} $G=(\mathbf{A}, \mathbf{X}), \mathbf{A} \in \mathbb{R}^{N \times N}, \mathbf{X} \in \mathbb{R}^{N \times d}$ \\
\KwResult{class capsules $\mathbf{u}^{(L)}$}
\For{$i\leftarrow 1$ \KwTo $N$}{
\For{$k\leftarrow 1$ \KwTo $K$}{
$\mathbf{z}_{i,k}=\sigma (\mathbf{W}_k^T \mathbf{x}_i) + \mathbf{b}_k$ \hspace{13pt} \/// $K$ latent factors
}
}
$\mathbf{u}^{(1)}_i=squash(\mathbf{z}_i)$ \hspace{14pt} \/// disentangled graph capsules
\For{$l\leftarrow 1$ \KwTo $L$}{
$b^{(l)}_{i,j} = 0$ \\
\For{$j\leftarrow 1$ \KwTo $N_{l+1}$}{
$\mathbf{v}^{(l)}_j = \text{TGNN}_j(\mathbf{A}^{(l)}, \mathbf{u}^{(l)})$ \hspace{1.5cm} \/// votes \\
}
\For{$r\leftarrow 1$ \KwTo $R$}{
$c^{(l)}_{i,j} = {\text{exp}(b^{(l)}_{i,j})}/{\sum_k \text{exp}(b^{(l)}_{i,k})}$ \\
$\mathbf{u}^{(l+1)}_j = squash(\sum_i c^{(l)}_{i,j}\mathbf{v}^{(l)}_{j|i})$ \\
$b^{(l)}_{i,j} = b^{(l)}_{i,j} + \mathbf{v}^{(l)}_{j|i} \cdot \mathbf{u}^{(l+1)}_j$
}
$\mathbf{A}^{(l+1)}={\mathbf{C}^{(l)}}^T \mathbf{A}^{(l)} \mathbf{C}^{(l)} \in \mathbb{R}^{N_{l+1} \times N_{l+1}}$
}
\textbf{return} $\mathbf{u}^{(L)} \in \mathbb{R}^{O \times d_L}$ \hspace{2cm} \/// class capsules
\caption{Training process with $K$ latent factors, $L$ capsule layers, and $R$ iterations of routing.}
\label{alg:alg_1}
\end{algorithm}
\begin{table*}
\footnotesize
\setlength\tabcolsep{4.5pt}
\begin{center}
\begin{tabular}{ @{} l|c|*{7}{c} @{} }
\toprule
\bf Algorithm & & \bf MUTAG & \bf NCI1 & \bf PROTEINS & \bf D\&D & \bf ENZYMES & \bf PTC & \bf NCI109 \\
\midrule
AWE & \multirow{4}{*}{\rotatebox[origin=c]{90}{Kernel}} &
87.87$\pm$9.76 & \textemdash & \textemdash & 71.51$\pm$4.02 & 35.77$\pm$5.93 & \textemdash &\textemdash \\
GK & &
81.58$\pm$2.11 & 62.49$\pm$0.27 & 71.67$\pm$0.55 & 78.45$\pm$0.26 & 32.70$\pm$1.20 & 59.65$\pm$0.31 & 62.60$\pm$0.19 \\
WL & &
82.05$\pm$0.36 & 82.19$\pm$0.18 & 74.68$\pm$0.49 & 79.78$\pm$0.36 & 52.22$\pm$1.26 & 57.97$\pm$0.49 & 82.46$\pm$0.24 \\
DGK & &
87.44$\pm$2.72 & 80.31$\pm$0.46 & 75.68$\pm$0.54 & 73.50$\pm$1.01 & 53.43$\pm$0.91 & 60.08$\pm$2.55 & 80.32$\pm$0.33 \\
\midrule
SAGPool &
\multirow{11}{*}{\rotatebox[origin=c]{90}{GNN}} &
\textemdash & 67.45$\pm$1.11 & 71.86$\pm$0.97 & 76.45$\pm$0.97 & \textemdash & \textemdash & 67.86$\pm$1.41 \\
CLIQUEPOOL & &
\textemdash & \textemdash & 72.59 & 77.33 & 60.71 & \textemdash & \textemdash \\
ASAP & &
\textemdash & 71.48$\pm$0.42 & 74.19$\pm$0.79 & 76.87$\pm$0.7 & \textemdash &
\textemdash & 70.07$\pm$0.55 \\
HaarPool & &
77.60$\pm$8.94 & 80.17$\pm$2.29 & 73.23$\pm$2.51 & \textemdash & \textemdash & \textemdash & 69.61$\pm$1.49 \\
EigenPooling & &
79.50 & 77.00 & 76.60 & 78.60 & 64.50 & \textemdash & 74.90 \\
DGCNN & &
85.83$\pm$1.66 & 74.44$\pm$0.47 & 75.54$\pm$0.94 & 79.37$\pm$0.94 & 51.00$\pm$7.29 & 58.59$\pm$2.47 & 75.03$\pm$1.72 \\
PSCN & &
88.95$\pm$4.37 & 76.34$\pm$1.68 & 75.00$\pm$2.51 & 76.27$\pm$2.64 & \textemdash &
62.29$\pm$5.68 & \textemdash \\
GIN & &
89.40$\pm$5.60 & 82.70$\pm$1.70 & 76.20$\pm$2.80 & \textemdash & \textemdash &
64.60$\pm$7.00 & \textemdash \\
DIFFPOOL & &
\textemdash & \textemdash & 76.25 & 80.64 & 62.53 & \textemdash & 74.10 \\
GCN & &
87.20$\pm$5.11 & 83.65$\pm$1.69 & 75.65$\pm$3.24 & 79.12$\pm$3.07 & 66.50$\pm$6.91 & \textemdash & 70.70 \\
GFN & &
90.84$\pm$7.22 & 82.77$\pm$1.49 & 76.46$\pm$4.06 & 78.78$\pm$3.49 & 70.17$\pm$5.58 & \textemdash & \textemdash \\
\midrule
CapsGNN &
\multirow{3}{*}{\rotatebox[origin=c]{90}{Caps}} &
86.67$\pm$6.88 & 78.35$\pm$1.55 & 76.28$\pm$3.63 & 75.38$\pm$4.17 & 54.67$\pm$5.67 & \textemdash & \textemdash \\
GCAPS-CNN & &
\textemdash & 82.72$\pm$2.38 & 76.40$\pm$4.17 & 77.62$\pm$4.99 & 61.83$\pm$5.39 &
66.01$\pm$5.91 & 81.12$\pm$1.28 \\
\textbf{Ours} & &
\textbf{93.16$\pm$6.10} & \textbf{84.87$\pm$1.68} & \textbf{77.99$\pm$3.16} & \textbf{80.99$\pm$2.58} & \textbf{78.00$\pm$4.89} & \textbf{66.54$\pm$7.97} & \textbf{83.91$\pm$1.27} \\
\bottomrule
\end{tabular}
\end{center}
\caption{Performance comparison on biological graphs. "Caps" indicates capsule-based GNNs.}
\label{table:biological_c1}
\end{table*}
\iffalse
\begin{table}
\caption{Performance comparison on biological graphs ($C^*$).}
\label{table:biological_c2}
\footnotesize
\setlength\tabcolsep{4.5pt}
\begin{center}
\begin{tabular}{ @{} l|c|*{3}{c} @{} }
\toprule
\bf Algorithm & & \bf PROTEINS & \bf D\&D & \bf ENZYMES \\
\midrule
STRUCTPOOL &
\multirow{4}{*}{\rotatebox[origin=c]{90}{GNN}} &
80.36 & 84.19 & 63.83 \\
MemGNN & &
81.35 & 82.92 & 75.50 \\
GMN & &
\textbf{82.25} & \textbf{84.40} & 78.66 \\
\textbf{Ours} & &
81.32$\pm$2.23 & 84.04$\pm$2.38 & \textbf{84.17$\pm$4.03} \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\fi
\begin{table}
\footnotesize
\setlength\tabcolsep{1.5pt}
\begin{center}
\begin{tabular}{ @{} l|c|*{4}{c} @{} }
\toprule
\bf Algorithm & & \bf COLLAB & \bf IMDB-B & \bf IMDB-M & \bf RE-B \\
\midrule
GK & \multirow{4}{*}{\rotatebox[origin=c]{90}{Kernel}} &
72.84$\pm$0.28 & 65.87$\pm$0.98 & 43.89$\pm$0.38 & 65.87$\pm$0.98 \\
AWE & &
73.93$\pm$1.94 & 74.45$\pm$5.83 & 51.54$\pm$3.61 & 87.89$\pm$2.53 \\
WL & &
79.02$\pm$1.77 & 73.40$\pm$4.63 & 49.33$\pm$4.75 & 81.10$\pm$1.90 \\
DGK & &
73.09$\pm$0.25 & 66.96$\pm$0.56 & 44.55$\pm$0.52 & 78.04$\pm$0.39 \\
\midrule
PSCN &
\multirow{6}{*}{\rotatebox[origin=c]{90}{GNN}} &
72.60$\pm$2.15 & 71.00$\pm$2.29 & 45.23$\pm$2.84 & 86.30$\pm$1.58 \\
DGCNN & &
73.76$\pm$0.49 & 70.03$\pm$0.86 & 47.83$\pm$0.85 & 76.02$\pm$1.73 \\
DIFFPOOL & &
75.48 & \textemdash & \textemdash & \textemdash \\
GCN & &
81.72$\pm$1.64 & 73.30$\pm$5.29 & 51.20$\pm$5.13 & \textemdash \\
GFN & &
81.50$\pm$2.42 & 73.00$\pm$4.35 & 51.80$\pm$5.16 & \textemdash \\
GIN & &
80.20$\pm$1.90 & 75.10$\pm$5.10 & 52.30$\pm$2.80 & 92.40$\pm$2.50 \\
\midrule
GCAPS-CNN &
\multirow{3}{*}{\rotatebox[origin=c]{90}{Caps}} &
77.71$\pm$2.51 & 71.69$\pm$3.40 & 48.50$\pm$4.10 & 87.61$\pm$2.51 \\
CapsGNN & &
79.62$\pm$0.91 & 73.10$\pm$4.83 & 50.27$\pm$2.65 & \textemdash \\
\textbf{Ours} & &
\textbf{82.86$\pm$1.81} & \textbf{77.20$\pm$4.73} & \textbf{52.80$\pm$2.45} & \textbf{93.15$\pm$1.58} \\
\bottomrule
\end{tabular}
\end{center}
\caption{Performance comparison on social graphs.}
\label{table:social_c1}
\end{table}
\iffalse
\begin{table}
\caption{Performance comparison on social graphs ($C^*$).}
\label{table:social_c2}
\footnotesize
\setlength\tabcolsep{1.4pt}
\begin{center}
\begin{tabular}{ @{} l|c|*{4}{c} @{} }
\toprule
\bf Algorithm & & \bf COLLAB & \bf IMDB-B & \bf IMDB-M & \bf RE-B \\
\midrule
STRUCTPOOL &
\multirow{4}{*}{\rotatebox[origin=c]{90}{GNN}} &
74.22 & 74.70 & 52.47 & \textemdash \\
MemGNN & &
77.0 & \textemdash & \textemdash & 85.55 \\
GMN & &
80.18 & \textemdash & \textemdash & 95.28 \\
\textbf{Ours} & &
\textbf{84.80$\pm$1.57} & \textbf{79.80$\pm$3.39} & \textbf{55.80$\pm$2.20} &
\textbf{95.30$\pm$1.30} \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\fi
\section{Experiments}
In this section, we conduct empirical studies on 11 benchmark datasets and demonstrate HGCN's superiority over a number of state-of-the-art graph classification methods.
Extensive ablation studies are also performed to evaluate the effectiveness of each component in our model.
\subsubsection{Datasets}
Eleven commonly used benchmarks including (i) seven biological graph datasets, i.e., MUTAG, NCI1, PROTEINS, D\&D, ENZYMES, PTC, NCI109; and (ii) four social graph datasets, i.e., COLLAB, IMDB-Binary (IMDB-B), IMDB-Multi (IMDB-M), Reddit-BINARY (RE-B), are used in this study.
It is noteworthy that the social graphs have no node attributes, while the biological graphs come with categorical node attributes.
More details about the data statistics and properties can be found in Supplementary.
\subsubsection{Baseline Methods}
We compare with two capsule-based methods, i.e., CapsGNN \cite{xinyi2018capsule} and GCAPS-CNN \cite{verma2018graph}.
We also conduct a comparison with a number of state-of-the-art GNN-based methods, including PATCHY-SAN (PSCN) \cite{niepert2016learning}, PSCN \cite{niepert2016learning}, GCN \cite{kipf2016semi}, Deep Graph CNN (DGCNN) \cite{zhang2018end}, CLIQUEPOOL \cite{luzhnica2019clique}, DIFFPOOL \cite{ying2018hierarchical}, ASAP \cite{ranjan2019asap}, SAGPool \cite{lee2019self}, EigenPooling \cite{ma2019graph}, GIN \cite{xu2018powerful}, GFN \cite{chen2019powerful}, HaarPool \cite{wang2019haarpooling}, STRUCTPOOL \cite{yuanstructpool}, and MemGNN/GMN \cite{khasahmadi2020memory}.
For kernel-based methods, we consider WL \cite{shervashidze2011weisfeiler}, DGK \cite{yanardag2015deep}, AWE \cite{ivanov2018anonymous}, and GK \cite{shervashidze2009efficient}.
\subsubsection{Experimental Settings}
We set $K=4$, $R=3$, $\lambda=0.5$, $\beta=0.1$, $L=2$, and
follow the same settings in previous studies \cite{ying2018hierarchical} to perform 10-fold cross-validation for performance evaluation.
For each dataset, we select a single epoch that has the best cross-validation accuracy averaged over the 10 folds, and report the average and standard deviation of test accuracies at the selected epoch.
HaarPool \cite{wang2019haarpooling} repeats each experiment 10 times with different random seeds, for a fair comparison, we run HaarPool with 10-fold cross-validation and report the result.
STRUCTPOOL \cite{yuanstructpool} and MemGNN/GMN \cite{khasahmadi2020memory} use a different assessment criterion that selects the epoch with the best test accuracy on each fold and then take the average.
We name such assessment criterion as $C^*$ and run our method by following the same paradigm for comparison.
Unless otherwise indicated, we use the result reported in the original paper for other baseline methods.
For TGNNs, we adopt the GCN \cite{kipf2016semi} with $L_\text{GNN}=1$.
\iffalse
\begin{figure}
\begin{center}
\includegraphics[width=0.9\linewidth]{Figure/visualize.eps}
\end{center}
\caption{Visualization of routing weights.}
\label{fig:visualization}
\end{figure}
\fi
\begin{table*}
\footnotesize
\setlength\tabcolsep{4.5pt}
\begin{center}
\begin{tabular}{ @{} l|c|*{7}{c} @{} }
\toprule
& & \bf MUTAG & \bf NCI1 & \bf PROTEINS & \bf D\&D & \bf ENZYMES & \bf PTC & \bf NCI109 \\
\midrule
A1 &
\multirow{3}{*}{\rotatebox[origin=c]{90}{ablation}} &
91.46$\pm$5.77 & 80.24$\pm$1.78 & 76.37$\pm$3.11 & 78.35$\pm$2.73 & 70.00$\pm$4.51 & 63.61$\pm$12.47 & 80.64$\pm$2.09 \\
A2 &
& 92.08$\pm$5.10 & \textbf{85.28$\pm$1.37} & 77.63$\pm$3.03 & 80.64$\pm$3.65 & 77.00$\pm$5.82 & 64.24$\pm$8.45 & 83.84$\pm$1.41 \\
A3 &
& 92.11$\pm$7.13 & 84.87$\pm$1.07 & 77.54$\pm$3.44 & 79.96$\pm$3.26 & 77.67$\pm$3.70 & 65.10$\pm$8.81 & 83.84$\pm$1.48 \\
\midrule
\textbf{Ours} &
& \textbf{93.16$\pm$6.10} & 84.87$\pm$1.68 & \textbf{77.99$\pm$3.16} & \textbf{80.99$\pm$2.58} & \textbf{78.00$\pm$4.89} & \textbf{66.54$\pm$7.97} & \textbf{83.91$\pm$1.27} \\
\midrule
\midrule
2 &
\multirow{2}{*}{\rotatebox[origin=c]{90}{$K$}} &
91.55$\pm$5.64 & 84.50$\pm$1.87 & 77.90$\pm$3.31 & 80.04$\pm$2.77 & 77.50$\pm$5.05 & 64.81$\pm$9.45 & \textbf{84.15$\pm$1.93} \\
8 &
& 93.16$\pm$6.59 & 84.65$\pm$1.17 & 77.18$\pm$2.74 & 78.69$\pm$2.99 & 77.83$\pm$4.97 & 65.71$\pm$7.95 & 84.03$\pm$2.31 \\
\midrule
1 &
\multirow{4}{*}{\rotatebox[origin=c]{90}{$R$}} &
91.05$\pm$6.59 & 83.36$\pm$1.62 & 76.64$\pm$2.46 & 79.45$\pm$2.47 & 77.33$\pm$5.34 & 64.23$\pm$8.60 & 82.26$\pm$1.75 \\
2 &
& 92.11$\pm$6.68 & 84.01$\pm$1.68 & 77.45$\pm$2.92 & 79.97$\pm$3.60 & 77.83$\pm$3.77 & 65.13$\pm$8.07 & 83.35$\pm$1.36 \\
4 &
& 92.63$\pm$5.66 & 84.26$\pm$1.20 & 76.92$\pm$2.66 & \textbf{81.15$\pm$3.79} & 77.00$\pm$4.29 & 64.21$\pm$8.31 & 83.60$\pm$2.05 \\
5 &
& 92.08$\pm$6.19 & 84.55$\pm$1.23 & 77.18$\pm$1.99 & 80.64$\pm$3.11 & 77.33$\pm$3.87 & 65.38$\pm$14.31 & 83.11$\pm$1.82 \\
\midrule
\textbf{Ours} &
& \textbf{93.16$\pm$6.10} & \textbf{84.87$\pm$1.68} & \textbf{77.99$\pm$3.16} & 80.99$\pm$2.58 & \textbf{78.00$\pm$4.89} & \textbf{66.54$\pm$7.97} & 83.91$\pm$1.27 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Ablation studies (upper part) and sensitivity analyses (lower part) on biological graphs.}
\label{table:ablation_bio_c1}
\end{table*}
\begin{table}
\footnotesize
\setlength\tabcolsep{3.5pt}
\begin{center}
\begin{tabular}{ @{} l|c|*{4}{c} @{} }
\toprule
& & \bf COLLAB & \bf IMDB-B & \bf IMDB-M & \bf RE-B \\
\midrule
A1 &
\multirow{3}{*}{\rotatebox[origin=c]{90}{ablation}} &
81.44$\pm$1.57 & 64.70$\pm$11.65 & 51.00$\pm$3.10 & 91.45$\pm$2.13 \\
A2 &
& 82.20$\pm$1.41 & 74.80$\pm$4.18 & 49.80$\pm$6.39 & 92.95$\pm$1.94 \\
A3 &
& \textbf{83.08$\pm$1.69} & 75.90$\pm$4.36 & 52.00$\pm$1.54 & 92.70$\pm$1.53 \\
\midrule
\textbf{Ours} &
& 82.86$\pm$1.81 & \textbf{77.20$\pm$4.73} & \textbf{52.80$\pm$2.45} & \textbf{93.15$\pm$1.58} \\
\midrule
\midrule
2 &
\multirow{2}{*}{\rotatebox[origin=c]{90}{$K$}} &
82.94$\pm$1.66 & 75.60$\pm$6.69 & 52.20$\pm$2.46 & 92.85$\pm$2.12 \\
8 &
& \textbf{83.10$\pm$1.80} & 74.90$\pm$5.82 & 51.67$\pm$3.99 & \textbf{93.35$\pm$1.73} \\
\midrule
1 &
\multirow{4}{*}{\rotatebox[origin=c]{90}{$R$}} &
82.52$\pm$1.87 & 74.50$\pm$4.58 & 51.87$\pm$3.23 & 92.45$\pm$1.82 \\
2 &
& 82.36$\pm$2.03 & 74.80$\pm$4.73 & 51.40$\pm$3.68 & 92.65$\pm$1.73 \\
4 &
& 83.04$\pm$1.92 & 74.70$\pm$2.87 & 51.67$\pm$3.07 & 93.10$\pm$0.91 \\
5 &
& 82.64$\pm$1.44 & 74.50$\pm$6.26 & 51.13$\pm$2.93 & 92.85$\pm$1.67 \\
\midrule
\textbf{Ours} &
& 82.86$\pm$1.81 & \textbf{77.20$\pm$4.73} & \textbf{52.80$\pm$2.45} & 93.15$\pm$1.58 \\
\bottomrule
\end{tabular}
\end{center}
\caption{Ablation studies (upper part) and sensitivity analyses (lower part) on social graphs.}
\label{table:ablation_social_c1}
\end{table}
\subsubsection{Experimental Results}
We first compare HGCN with existing state-of-the-art graph classification methods on seven biological datasets.
Our results demonstrate that HGCN achieves the best performance based on the widely used criterion (Table \ref{table:biological_c1}) and competes favorably against three $C^*$- based baselines (Supplementary Table 5).
In particular, compared with two capsule-based GNNs (i.e., GCAPS-CNN and CapsGNN), we boost of 16.17\%, 6.49\%, 3.37\%, and 2.79\% improvement on ENZYMES, MUTAG, D\&D, and NCI109, respectively.
The reason is that higher-order statistical moments of local neighbors are used in GCAPS-CNN to build graph capsules, this strategy, however, fails to identify the underlying latent factors which are important in preserving node/graph property and extracting the hierarchical representations.
Although this limitation can be partially alleviated by using multiple graph channels as reported in CapsGNN, the transformation matrices used in CapsGNN ignore the structure information involved in lower-level capsules.
In contrast, we disentangle node representations to explicitly consider the entanglement of heterogeneous factors, and propose transformation GNNs to measure structure-aware part-whole relationships.
Furthermore, it should be noted that we achieve the largest performance gains on ENZYMES which has six classes, compared to the rest of datasets with only two classes.
This observation implies that HGCN is able to capture more complicated and accurate hierarchy for the multiclass classification problem than other methods.
For instance, although assignment matrix-based methods (e.g., DIFFPOOL) are capable of mapping nodes to a set of clusters for structurally simple graphs, its representation power is limited to complex and crowded graphs such as ENZYMES.
By contrast, our method performs iterative routing to obtain the cluster of agreement, which jointly learns the hierarchical graph representation and provides the necessary deprecation of assignment ambiguity \cite{sabour2017dynamic}.
Table \ref{table:social_c1} and Table 6 (Supplementary) show the performance comparison on four social graph datasets, where the key challenge is to identify strongly connected communities.
Similarly, we achieve the significant performance improvement over baseline models, indicating that HGCN can better reason about the part-whole relationships in social networks.
This is also consistent with the fact that highly complex interactions are involved in social graphs, which can be modeled by identifying heterogeneous factors underlying each node.
Most importantly, considering the entanglement of the latent factors enables more accurate hierarchy learning.
\subsubsection{Ablation Studies}
Comprehensive ablation studies are carried out in this section to understand the contribution of each component (i.e., disentangled graph capsules, capsule layers, and auxiliary graph reconstruction) in our method.
Specifically, we (i) directly use the input node representation to serve as graph capsules, without considering the entanglement of heterogeneous factors (A1); (ii) remove the residual connection between adjacent capsule layers (A2); and (iii) remove the auxiliary graph reconstruction (A3).
The results illustrated in Table \ref{table:ablation_bio_c1} and Table \ref{table:ablation_social_c1} (upper part) reveal that (i) disentangling node representation allows us to characterize the latent factors underlying each node and in turn more accurately preserve the node/graph properties and capture the part-whole relationship; (ii) combining fine, low layer information with coarse, high layer information provides us an ability to enhance the final graph-level representation; and (iii) graph reconstruction plays an important role in encoding the instantiation parameters of the input graph and enhancing the training stability.
Thus, we reach the conclusion that each component in our method is necessary and contributes to the performance improvement.
One exception is NCI1 (A2) and COLLAB (A3), where residual connection and auxiliary graph reconstruction bring inferior performance which may be caused by overfitting.
\subsubsection{Sensitivity Analyses}
In this section, we analyze the sensitivity of HGCN to the number of latent factors $K=\{2,8\}$ and the number of routing iterations $R=\{1,2,4,5\}$, where our method with setting: $K=4$ and $R=3$.
As shown in Table \ref{table:ablation_bio_c1} and Table \ref{table:ablation_social_c1} (lower part), the results demonstrate that HGCN is not very sensitive to these two hyper-parameters.
Although $K=8$ brings limited performance improvement on COLLAB and RE-B than $K=4$ (ours), the computational complexity is doubled in calculating the disentangled graph capsules.
Similarly, $R=4$ requires more routing iterations, albeit with 0.16\% accuracy boost on D\&D.
\section{Conclusion}
In this paper, we introduce a novel HGCN framework for graph classification, which is able to explicitly extract hierarchical graph representations.
Built upon disentangled graph capsules by identifying heterogeneous factors behind each node, HGCN encodes part-whole relationships by considering the structure information of lower-level parts and iteratively infer the pose of higher-level objects.
Empirical studies demonstrate the superiority of our framework over existing graph classification methods on 11 commonly used benchmarks.
\section{Acknowledgments}
This work was partially supported by US National Science Foundation IIS-1718853, the CAREER grant IIS-1553687 and Cancer Prevention and Research Institute of Texas (CPRIT) award (RP190107).
| fd7c8bee8e9eea07a0a2eb3f78ad1381e2e65180 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Conclusions}
\label{sec:conclusions}
We explored the
To achieve this, a future research direction would be to not just model but also predict cross-device behaviour.
\section{Introduction}
\label{sec:introduction}
Information retrieval (\acs{IR}\acused{IR}) technology is at the heart of today's e-commerce platforms, in the form of search engines, recommenders, and conversational assistants that connect users to the products they may be interested in~\citep{rowley2000product}.
To help improve the effectiveness of \ac{IR} technology in an e-commerce context, the problem of analyzing, modeling, and, ultimately, predicting customers' purchase intent has been studied extensively in academia and industry~\citep{bellman1999predictors, agichtein2006learning, lo2016understanding}
\header{Purchase intent prediction} Here, purchase intent is defined as a predictive measure of subsequent purchasing behavior~\citep{morwitz1992using}.
\begin{figure}[h]
\centering
\includegraphics[clip,trim=0mm -5mm 0mm 0mm,width=0.8\columnwidth]{user_intents_5_devices.pdf}
\caption{Customer journeys across sessions, with multiple interests and devices; the colors indicate different devices.}
\label{fig:teaser}
\end{figure}
Figure~\ref{fig:teaser} illustrates the complexities of customer behavior during a sequence of sessions, when multiple tasks, interests, and devices may play a role.
Areas in the back of the figure are meant to signify different user journeys across time, purple for one that is focused on fridges, yellow for one that is focused on a birthday present. Colored rectangular blocks in the front indicate different devices used by the user. Initial exploration of a relatively expensive item (a fridge) starts on a smartphone and continues on a tablet, while the journey ends with a purchase of a fridge on a PC.
The purchase of a fridge is interleaved with the purchase of a (lower-priced) birthday present, with initial exploration on a PC, followed by further exploration on a TV and PC, and, ultimately, a purchase on a PC.
Online search behavior that targets transactions has been analyzed at scale at least since the work by~\citet{broder-2002-taxonomy}, who identified a class of so-called \emph{transactional} queries, where the user is seeking to reach a page with more interaction opportunities, e.g., to conduct a purchase, download or sign-up. In particular, factors influencing online purchases have been described as early as in 2002~\citep{george2002influences}, and work on predicting purchases goes back to at least to the work of~\citep{ben-shimon-2015-recsys}, where the task was to predict whether a given customer is going to purchase within a given session.
\header{Challenges} Despite the many advances, purchase intent prediction still has many challenges~\citep{tsagkias-2020-challenges}. In particular, previous work on purchase intent prediction has focused mostly on customers of an e-commerce platform who are identified or recognized by the platform. A diverse range of models has been considered, from traditional feature-based models such as boosted decision trees to sequence-based neural models such as RNNs.
However, based on the analysis of de-identified data from an e-commerce website available to us, more than 50\% of traffic comes from anonymous users. Purchase intent detection for anonymous users is particularly challenging because it cannot rely on historical information about the user on which many of the existing models rely.
\header{Features for purchase intent prediction} In this paper, we focus on identifying signals that suggest purchase intent in an anonymous and identified setting. We do this by analyzing purchase vs. non-purchase sessions sampled from a large European e-commerce website and testing the features based on our observations on a production-ready model. We further test the obtained feature sets on five other classifiers to explore the generalizability of our findings.
In particular, we include features derived from session-based data such as page dwell time and customer-specific data such as the number of days since the last purchase. Session-based features have the advantage that they are available both during sessions when a user is identified (i.e., the customer has logged-in or is recognized through cookies) and anonymous sessions (when the customer is not known). Customer-related features are only available during identified sessions.
Interestingly, many of the features proposed previously~\citep{seippel-2018-customer} apply only to identified sessions: \emph{purchase intent prediction for anonymous sessions has been studied very little}.
To fill this gap, we analyze a dataset of more than 95 million sessions, sampled from four weeks of anonymized user interaction data in a European e-commerce platform. We answer the following research questions:
{\em RQ1: How do purchase sessions differ from non-purchase sessions?} In Section~\ref{section:characterizingpurchaseintent} we compare purchase vs. \ non-purchase sessions in such aspects as session length, temporal variations, device and channel type, queries. Among others, we find out that purchase sessions tend to be longer than non-purchase ones, customers are more likely to purchase in the evening and during a weekday, and more likely to own more than 1 device.
{\em RQ2: What are the important session-based features that allow us to tell purchase sessions apart from non-purchase sessions? What are the important historical features that should inform predictors for identified sessions? How does the importance of features change across the session?} Based on the experiments described in Section~\ref{section:predictingpurchaseintent}, we conclude that historical features related to previous purchasing behavior are highly important for detecting purchases in the identified setting. For the anonymous setting, however, dynamic features related to page dwell time and sequence of pages are most important. Besides, the importance of dynamic features increases as the session continues, while the importance of static features decreases.
{\em RQ3: How effective are models used for purchase intent prediction for anonymous vs.\ identified sessions? Furthermore, to which degree do the proposed features help improve performance for anonymous sessions?} In Section~\ref{section:predictingpurchaseintent}, we show that in the anonymous setting, tree-based and neural classifiers demonstrate the best performance, and adding extra features to models improves $F_{1}$ by about 17\%. In contrast, for identified setting all models demonstrate high performance and adding extra features do not provide a significant gain.
\noindent%
The principal contributions of our research are the following:
\begin{itemize}[leftmargin=*,nosep]
\item We conduct an in-depth analysis of a real-world customer interaction dataset with more than 95 million sessions, sampled from a large European e-commerce platform. We identify session features such as device type and conversion rate, weekday, channel type, and features based on historic customer data such as number of previous orders and number of devices to distinguish between purchase and non-purchase sessions (see Section~\ref{section:characterizingpurchaseintent}).
\item We define two feature sets for purchase prediction, tailored towards anonymous sessions and identified sessions (see~Section~\ref{section:predictingpurchaseintent}).
\item We evaluate our proposed features by extending an existing production-ready model and run additional experiments with classifiers generally used for this task. We find $F_{1}$ improvements of up to 17\% in purchase intent prediction for anonymous sessions and reach an $F_{1}$ of 96\% for identified sessions on held-out data collected from a real-world retail platform (see ~Section~\ref{section:predictingpurchaseintent}).
\end{itemize}
\section{Dataset Description}
\label{sec:dataset}
In this section, we describe how we extract a dataset consisting of anonymized user interaction data from the search logs of an e-commerce platform, and summarize dataset statistics.
\header{Data Collection} Our dataset comprises four weeks (28 days) of anonymized visits sampled from a European e-commerce platform in October 2019.
The original sample of the log entries includes a unique non-personal customer identifier (for identified users), the type of browsing device used during the session, as well as a timestamp for every query, and a URL of each clicked page.
We convert all the timestamps to the Central European Time Zone (CET).
We additionally recorded the price of every product the customers have seen and the prices of the items they ended up buying.
In cases where a customer starts a session without logging in and ends up logging in at a later point in the session, we assign the session to the customer.
To filter out bot traffic, we apply several measures related to location and device type~\citep{bomhardt-2005-web}.
First, we filter out sessions based on location, to only include entries from the European countries from which the majority of the customers come; bots come mostly from non-European IPs, especially North-America.
Second, we specify the set of device types we are interested in and remove all the entries from other devices, leaving us with \emph{PC}, \emph{Smartphone}, \emph{Tablet}, \emph{Game Console}, and \emph{TV}; bots often do not specify a device type.
\header{Dataset Statistics} Table~\ref{tab:ds_stats} provides descriptive statistics of the resulting dataset.
Overall, the dataset contains \numprint{95757177} sessions, out of which \numprint{54144152} (about 56.5\%) are anonymous.
In total, the dataset contains \numprint{9663509} identified users.
We additionally keep track of the device types used for browsing and distinguish between five such device types: PC, smartphone, tablet, game console, and TV.
The table also lists the number of search queries; these are the queries submitted during the sessions captured in the log.
\begin{table}[!htb]
\caption{Dataset statistics.}
\label{tab:ds_stats}
\begin{tabular}{lr}
\toprule
\bf Description & \bf Total \\
\midrule
Sessions & \numprint{94402590} \\
\quad Anonymous & \numprint{55305709} \\
\quad Logged-in or recognized & \numprint{39096881} \\
Logged-in or recognized customers & \numprint{6125781} \\
Queries & \numprint{31185176} \\
\multirow{2}{*}{Device types} & \multicolumn{1}{l}{PC, Smartphone, Tablet,} \\
& \multicolumn{1}{l}{Game Console, TV} \\
\bottomrule
\end{tabular}
\end{table}
\section{Predicting Purchase Intent}
\label{section:predictingpurchaseintent}
Next, we turn to predict purchase intent when a user is anonymous (``anonymous setting'') and when a user is logged-in or recognized (``identified setting''). The goal of our experiments is to evaluate how the features which we discovered during dataset exploration influence purchase predictor performance in both settings. To accomplish this, we derive a feature set for each setting, and evaluate the features by adding them to an existing production-ready model, based on a Random Forest. To showcase the generalizability of our findings, we additionally test the impact of our features on five additional popular classifiers.
To investigate how the models' ability to predict purchase evolves throughout a session, we evaluate all models on 11 session steps (corresponding to the visits of 10 pages). We are interested in longer sessions because the outcome of such sessions is more difficult to predict. As we do not want to evaluate the model's performance on the very last step, (where the outcome is clear), we set up a buffer of 2 pages. Therefore, we filter out all the sessions which are shorter than 12 pages.
We conclude the section by analyzing the features which contributed most to the model performance in both the anonymous and the identified setting, and explore how dynamic and static feature importance change as the session continues.
\subsection{Experimental Setup}
\label{subsec:exper_setup}
In this section, we discuss the feature sets which we use in the experiments for the anonymous and identified setting, the models on which we test the features, and the evaluation setup.
\header{Feature sets}
We start by designing a set of features for purchase prediction in identified and anonymous user settings. Since our initial analysis demonstrated that about 56\% of all sessions are anonymous (see Table~\ref{tab:ds_stats}), it is worth to pay special attention to this category.
Based on the findings obtained thus far and on an analysis of best-performing features available in the literature \citep{hop2013web, lee2015online, niu2017predictive, seippel-2018-customer}, we compile a feature set presented in the Table~\ref{tab:feature_set}.
\begin{table}[!h]
\caption{Complete feature set. ``Dynamic'' indicates that a feature may change during a session.}
\label{tab:feature_set}
\begin{tabular}{llcc}
\toprule
\multicolumn{2}{l}{\bf Feature} & \bf Dynamic & \bf Baseline \\
\midrule
\multirow{8}{*}[-0.1cm]{\rotatebox{90}{Session}}
& current page dwell time, mean & \checkmark & \checkmark\\
& current page dwell time, $\sigma$ & \checkmark & \checkmark \\
& page sequence score & \checkmark & \checkmark \\
& number of pages & \checkmark & \checkmark\\
& channel type \\
& start hour \\
& week day \\
& device type \\
& device conversion rate \\
\midrule
\multirow{9}{*}[0.5cm]{\rotatebox{90}{History}}
& number of orders & & \checkmark \\
& days since last purchase & & \checkmark \\
& number of sessions \\
& number of devices \\
& device sequence score \\
& switch probability \\
\bottomrule
\end{tabular}
\end{table}
We categorize features into two classes: \emph{session features} and \emph{customer history features}. We derive session features from the information of the given session and base customer history features on the information from previous sessions of the given customer.
Since we run experiments in the anonymous and the identified setting, we use different feature sets for each setting. In the anonymous setting, the information about the customer is not available and, therefore, we can only use session features. On the other hand, when a customer is identified, we can use both session and customer history features.
The feature set contains both static and dynamic features. Dynamic features can change throughout the session, whereas static features remain constant.
\header{Models} Next, we select models on which we evaluate the features discovered during the dataset analysis. As our primary model, we use a production-ready classifier. This is a random forest (RF) with a baseline feature set as described in Table~\ref{tab:feature_set}.
Additionally, to showcase the general utility of our feature set, we experiment on additional models. After reviewing previous work in the domain of purchase prediction (see Section~\ref{sec:relatedwork}), we choose the following models for our experiments: logistic regression (LR), K-nearest neighbors (KNN), support vector machines (SVM), neural classifier, and gradient boosted decision tree (GBDT). Each model is trained on the baseline and extended feature set in both settings.
\header{Prediction setup} Since we want to explore how models' performances change across sessions, we select points of a session for which we predict the probability of purchase.
We define a point by the number of pages opened in the session up until the point of prediction. Overall, we select 11 points of measurement. The first point is at the very beginning of the session when the user did not open any pages yet. At this point, the classifier makes a prediction based solely on static features. The following point of measurement is right after the user opened the first page. The subsequent nine points happen after the next nine pages. To make the evaluation possible and to ensure that we do not predict for the very last session page, we filter out sessions with fewer than 12 pages, with 2 pages as a buffer. The buffer is there to avoid the situation when the model predicts at the very end of a session when the outcome is clear. Therefore, we filter out all the sessions which are less than 12 pages long. For example, in step 2 we only have a session with at least 12 actions, which is a hard setting.
\header{Evaluation setup} For both settings, we evaluate model performance with 10-fold cross-validation. To account for class imbalance, we set class weights to be inversely proportional to class frequencies and use $F_{1}$ score as a primary evaluation metric.
\subsection{Prediction for Anonymous Users}
First, we evaluate how the added features influence model performance in the anonymous setting, where the user is not known.
\header{Setup} For the anonymous setting, we sampled \numprint{22982} sessions. We use the data to create a feature set for the baseline model and our model. In the anonymous setting, there is no available information about customer history, therefore, we only use session features (see Table~\ref{tab:feature_set}.
As can be seen from the table, the baseline feature set comprises four dynamic features, whereas the extended feature set offers five extra features.
Since we predict purchase for different points in the session, we compute all dynamic features for a particular session point on which we evaluate. For each session point, we train the baseline and extended model on the obtained feature sets.
\header{Results} The results in Figure~\ref{fig:results} show that the additional features boost model performance across all session steps. The performance boost is especially significant at step 0 when a customer has not opened any pages yet. In general, tree-based models (RF and GBDT) and the neural classifier demonstrate the highest scores across all steps. The models are followed by SVM, LR, and KNN classifiers.
The performance of all the models with the baseline feature set improved on step 1. The gain can be explained by the introduction of dynamic session features (step 0 means that the user did not open any pages yet, hence, no dynamic session features). Conversely, for models with the extended feature set, the introduction of dynamic features on step 1 does not significantly increase the performance. After step 1, models' performances reach a plateau.
\subsection{Prediction of Identified Users}
Second, we test models' performances with baseline and extended feature sets in identified setting, when the user is known.
\header{Setup} For the identified setting, we sampled \numprint{6319} sessions.
The feature set for this setting includes session and customer history features (see Table~\ref{tab:feature_set}).
During our experiments, we found out that information about the previous session device (such as device type and conversion rate) decrease model performance, so we excluded those features from the training and evaluation sets. This can be explained by the fact that the information about what kind of device users previously used and what was the probability of purchase on that device is not relevant for predicting purchase on the current device.
In analogy with the anonymous setting, we prepare feature sets for each of the eleven session points and train and evaluate the models with the baseline and extended feature sets.
\header{Results} Figure~\ref{fig:results} shows the performance of the models. Overall, the performance of all models for both baseline and extended feature sets and across all steps stays around 96\%. The only exception is the k-nearest neighbors classifier where adding extra features on step 0 increases the model's performance by 6.74\%. On step 1, however, the gain from the extended feature set is not present. This can be explained by the introduction of the dynamic session features.
\subsection{Feature Importance Analysis}
The experimental results raise a natural question that is 'Which features contribute most to model performance in both settings?' To answer this question, we look at the feature importance scores of a production-ready classifier which shows one of the best performances in both settings, random forest.
Figure~\ref{fig:rf_anon_feat_imp} (top) demonstrates that in anonymous setting, day of the week is the feature with the highest importance. It is followed by three dynamic features (standard deviation and mean of page dwell time, and Markov page sequence score), and four static features (starting hour, channel type, device type and conversion rate).
\begin{figure}
\centering
\includegraphics[height=4.2cm]{rf_feature_importance_anon}
\includegraphics[height=4.2cm]{img/rf_feature_importance_iden}
\mbox{}\hspace*{16.6mm}
\includegraphics[clip,trim=0mm 0mm 0mm 10mm,height=4.18cm]{img/static_feature_importances_per_step_anonymous_setting_11_steps}
\caption{Feature importance for the Random Forest in the anonymous setting (top) and identified setting (center), as well as summed for static features in the anonymous setting (bottom).}
\label{fig:evaluation}
\label{fig:rf_anon_feat_imp}
\label{fig:rf_iden_feat_imp}
\label{fig:static_feature_importances}
\end{figure}
Figure~\ref{fig:rf_iden_feat_imp} (center) shows that in the identified user setting, number of previous orders, and number of days since last order are the features with the highest relative importance. Both features describe user historical purchasing behavior what can explain their high relative importance.
The features are followed by three dynamic features (standard deviation and mean of page dwelling time, and Markov page sequence score), which also have relatively high importance in the anonymous setting.
The high relative importance of the dynamic session features (standard deviation and mean of page dwelling time, and Markov page sequence score) in both settings explain the gain all models with baseline feature set got on step 1 in the anonymous setting (see Figure~\ref{fig:results}).
Next, we determine how static feature importance changes across sessions. We consider the importance in the anonymous setting because the introduction of dynamic features in this setting showed an improvement.
Figure~\ref{fig:static_feature_importances} (bottom) shows that static session feature importance decrease as the session evolves, which entails that the importance of dynamic features increases. On step 0 the cumulative importance of static features is 100\% because there are no dynamic features introduced. However, from step 1 the relative importance starts to drop. The figure supports the hypothesis that as the session progresses dynamic features become more important.
\begin{figure*}[!h]
\includegraphics[clip,trim=2mm 0mm 2mm 0mm,width=\linewidth]{models_performance}
\caption{Experimental results in anonymous (anon) and identified (iden) setting, across different session steps, $F_{1}$.}
\label{fig:results}
\end{figure*}
\section{Predicting Purchase}
\label{sec:5}
In this section, we formally describe our prediction task and outline the features used for prediction.
For a given session, we predict whether a user is going to buy or not. We do prediction at several session steps. We define step as number of pages customer seen during session.
We conduct experiments in two settings. In the first setting, we assume that users are anonymous.
In the second setting, we assume that users are logged in.
For a baseline, we use a model currently being deployed in the e-commerce website.
After reviewing models available for the task, we choose random forest as model for our experiments. We choose this model because of the transparency feature importance of tree-based models.
We analyzed features used in the literature and selected best performing features. The complete feature set is presented in the Table~\ref{tab:feature_set}.
\begin{table}[!htb]
\caption{Complete feature set. An asterisk * indicates that feature changes throughout the session.}
\label{tab:feature_set}
\begin{tabular}{ll}
\hline
\bf Type & \bf Feature \\ \hline
\multirow{8}{*}{Session Data} & current page score* \\
& current page dwell time, mean* \\
& current page dwell time, $\sigma$* \\
& number of pages* \\
& channel type \\
& start hour \\
& week day \\
& device type \\
& device conversion rate \\
\addlinespace
\multirow{8}{*}{Customer History} & number of orders \\
& days since last purchase \\
& number of sessions \\
& number of devices \\
& device sequence score \\
& switch probability \\
& previous device type \\
& previous device conversion rate \\ \hline
\end{tabular}
\end{table}
\begin{itemize}
\item Task definition: given a visit or a customer journey, predict whether the user will end up buying or not.
\item data preparation process
\end{itemize}
\begin{figure}[]
\includegraphics[width=0.8\linewidth]{device_feature_importances_per_step_anonymous_setting}
\caption{Relative feature importances across session. End is the last page of the session minus buffer.}
\label{fig:device_feature_importances}
\end{figure}
\subsection{Experimental Setup}
\begin{itemize}
\item model description: Random Forest with several features
\item anonymous setting description
\item logged-in users description
\end{itemize}
\subsection{Results}
\begin{itemize}
\item results table
\item results interpretation
\item Error pattern analysis
\end{itemize}
\section{Definitions}
\label{sec:definitions}
In the study, we operate with the following definitions:
\begin{description}
\item[Session] is a sequence of requests made by a single-end user during a visit to a particular site. Session ends if user is idle for more than 30 minutes.
We define two types of sessions: purchase and non-purchase session. Purchase session is a session during which customer buys something. Non-purchase session is a session during which customer does not buy anything.
\item[Session length] is a number of actions taken during given session.
\item[Step] is a number of pages customer seen during given session.
\item[Actions] include opening a new web page, submitting a search query.
\item[Device switch] is the act of changing device type in-between sessions.
\item[Moving average]
\item[Moving standard deviation]
\item[Markov Score]
\item[Conversion rate] is a fraction of the number of visits during which a purchase was made~\cite{moe2004dynamic}. We use this metric to compare device popularity in a purchasing context. The conversion rate can be calculated as follows:
\begin{equation}
\text{\em Conversion rate} = \frac{\text{\em\#purchase\ sessions}}{\text{\em\#all\ sessions}}
\end{equation}
\item[Standardized conversion rate] is used to protect sensitive information. We present standardized conversion rates for each device; since we are interested in differences across devices types, this suffices for our purposes. The \emph{standardized conversation rate} is computed as follows:
\begin{equation}
\begin{split}
\text{\em Standardized}&\text{\em\ conversion rate} = \\
& \frac{\text{\em Conversion rate} - \overline{\text{\em Conversion rate}}}{\sigma}
\end{split}
\end{equation}
where:
\begin{itemize}
\item $\overline{\text{\em Conversion rate}}$ is the mean conversion rate per device type
\item $\sigma$ is standard deviation of device-specific conversion rates
\end{itemize}
\end{description}
\section{Characterizing Purchase Intent}
\label{section:characterizingpurchaseintent}
We explore customer behavior and, in particular, the difference in the behavior of purchasing and non-purchasing users. These explorations aim to identify characteristics that may help us improve the effectiveness of purchase intent predictors. We analyze several aspects of sessions, such as the length of purchase and non-purchase sessions, the temporal characteristics of sessions, and device information. Furthermore, we investigate the channels from which customers start sessions and issue queries during purchase sessions and non-purchase sessions.
\subsection{Session Length}
\label{subsec:sess_len}
First, we examine the overall session length for purchase sessions and non-purchase sessions.
Figure~\ref{fig:purchase_non_purchase_length} plots the complementary cumulative distribution function (CCDF) of the session lengths of purchase sessions and non-purchase sessions per device type.
\begin{figure}[!htb]
\includegraphics[width=0.8\linewidth]{cdd_visit_length_5_devices}
\caption{CCDF of the session length per device type for purchase sessions (p.s.) and non-purchase sessions (non-p.s.).}
\label{fig:purchase_non_purchase_length}
\end{figure}
As can be seen in the area between the P50 and P90 percentiles in Figure~\ref{fig:purchase_non_purchase_length}, purchase sessions are in general longer than non-purchase sessions.
Moreover, the purchase session length per device varies less than the non-purchase session length per device. It can be explained by the fact that non-purchase sessions can be both very short or rather long, depending on the underlying user intents. For instance, a user could quickly look something up or spend some time exploring the catalog. On the other hand, in the case of purchase sessions, user intentions are less ambiguous.
Usually, users look for a specific product that they have in mind and, upon finding it, proceed to purchase.
From a device perspective, the shortest sessions take place on smartphones, whereas sessions on tablets are generally longer. The longest sessions occur on the PC, TV, and game console. This finding holds for both purchase sessions and non-purchase sessions.
However, in the tail of the distributions, the distinction between purchase session length and non-purchase session length is not as clear as between the P50 and P90 percentiles. The non-purchase session length distribution on the PC has an exceptionally long tail.
Overall, we can attribute these findings to the fact that smartphones have a smaller screen and are therefore less convenient for longer sessions. Tablet screens are bigger than smartphone screens; hence, the sessions can last longer. The PC screen is the biggest one, and therefore PC users exhibit event longer sessions.
\subsection{Temporal Variations}
Next, we look into the temporal characteristics of purchase sessions and non-purchase sessions, such as their distribution across days of the week and the sessions' starting hours.
\begin{figure}[!htb]
\includegraphics[width=0.8\linewidth]{day_of_week_purchase_vs_non_purchase_sessions}
\caption{The fraction of purchase sessions and non-purchase sessions across days of the week w.r.t. total amount of purchase and non-purchase sessions. Most activity occurs on weekdays.}
\label{fig:day_of_week_purchase_vs_non_purchase_sessions}
\end{figure}
First, we want to understand customer activity during the days of the week. Figure~\ref{fig:day_of_week_purchase_vs_non_purchase_sessions} shows how the number of purchase and non-purchase sessions varies across days of the week.
The three most popular days for purchase sessions are Thursday, Tuesday, and Wednesday. In total, the purchase sessions of these three days amount to $48.55\%$ of all purchase sessions.
On the other hand, the least popular purchase days are Sunday, Saturday, Monday, and Friday. They contribute to $51.45\%$ of purchase sessions.
The observed pattern of purchase behavior hints at the fact that customers prefer to buy during weekdays, which aligns with their workweek. Besides, we conclude that the lower purchase activity on Monday and Friday attributes to their proximity to weekends.
In the case of non-purchase sessions, the most active session days are Wednesday, Monday, and Thursday. Altogether, these days contribute to $58.55\%$ of non-purchase sessions. The least active days are Tuesday, Sunday, Saturday, and Friday. All the sessions of these days amount to $41.44\%$.
Just like for purchase sessions, the activity for non-purchase sessions also centers around weekdays.
However, the difference between the three most active days and the four least active days for non-purchase sessions is bigger than the corresponding difference for purchase sessions. For purchase sessions, the difference is only $2.9\%$, whereas, for non-purchase sessions, the difference is $17.11\%$. Moreover, Tuesday, the
2-nd most popular day for purchase sessions is the least popular day for non-purchase sessions. On the other hand, Monday, the 2-nd most popular day for non-purchase sessions is the 3-rd least popular day for purchases.
The observation indicates that people need time to consider a purchase before making the buying decision. Hence, they spend Monday, the first day of the new week on considering the purchase, and the purchase itself happens on Tuesday or later in the week.
In general, the most active day of the week is Wednesday, whereas the least active day is Sunday.
These findings strongly suggest that user behavior depends on the day of the week. In general, people are most active on weekdays, during their workweek, their activity peaks in the middle of the week. On the other hand, at the beginning and end of the workweek, user activity is generally lower.
Next, we look at user behavior on the level of the hour during which a session starts. As mentioned in Section~\ref{sec:dataset}, all the hours are represented in CET. Figure~\ref{fig:hour_of_day_purchase_vs_non_purchase_sessions} shows how purchase sessions and non-purchase sessions spread across the hours of the day.
\begin{figure}[!h]
\includegraphics[width=0.8\linewidth]{hour_of_day_purchase_vs_non_purchase_sessions}
\caption{The fraction of purchase session and non-purchase sessions across the hours of the day w.r.t. total amount of purchase and non-purchase sessions. Most purchase sessions start in the evening.}
\label{fig:hour_of_day_purchase_vs_non_purchase_sessions}
\end{figure}
As expected, the least active hours are in the early morning, in the period from 1 am to 3 am. That can be explained by the fact that most people sleep during the night. (Note that the majority of the e-commerce platform customers come from Europe; hence, time does not vary that much.)
Moreover, the activity on the platform during the period from 10 am till 5 pm is stable both for purchase and non-purchase sessions,
whereas the most active hours are in the evening, i.e., from 6 pm till 8 pm. In general, our observations correspond to the established rhythm of the daily life of the majority of people, who sleep during the night, browse e-commerce platforms both during work hours and in the evening after work.
\subsection{Channel Types}
Next, we look at whether channel types distributions change across purchase and non-purchase sessions. We define the following channel types: \textit{direct} where a user enters the platform directly; \textit{paid} where a user enters the platform through search engine advertisement, and \textit{organic} where a user enters the platform through a web search engine and unpaid results. Table~\ref{tab:channel_type} displays the channel distribution across purchase and non-purchase sessions.
\begin{table}[!h]
\caption{Channel types for purchase and non-purchase sessions.}
\label{tab:channel_type}
\begin{tabular}{l c c r}
\toprule
& \multicolumn{2}{c}{\bf Sessions} & \bf Stand\\
\bf Channel & \bf Purchase (\%) & \bf Non-purchase (\%) & \bf conv. rate \\
\midrule
Direct & \textbf{\numprint{71.07}} &
\textbf{\numprint{77.30}} & -0.56 \\
Paid & \numprint{16.74}
& \numprint{12.92} & 0.54 \\
Organic &
\numprint{11.78} & \phantom{0}\numprint{7.83} & \textbf{0.94} \\
Other &
\phantom{0}\numprint{0.31} &
\phantom{0}\numprint{1.05} & -1.33\\
\bottomrule
\end{tabular}
\end{table}
\noindent%
Both for purchase and non-purchase sessions, the direct channel is the most used channel to enter the platform. However, for purchase sessions, the percentage of sessions which start with the direct channel is $8.06\%$ less than the fraction of non-purchase sessions, which started with the direct channel.
The second most popular channel for purchase and non-purchase sessions is a paid channel. However, in the case of this channel, the fraction of purchase sessions is $12.92\%$ bigger than the corresponding channel type fraction for non-purchase sessions.
The organic channel is the third channel in terms of popularity for both session groups. The organic channel fraction for purchase sessions is $50.40\%$ bigger for purchase sessions when compared with non-purchase sessions.
Overall, during purchase sessions, users are more likely to enter the platform through paid or organic channels, whereas for non-purchase sessions the direct channel is more common. It can be explained by the fact that purchasers decide to converge after being offered an advertisement or a search result that matches their interest, whereas non-purchasers may enter the platform directly to explore the catalog.
\subsection{Devices}
In this subsection, we investigate purchase intent from the perspective of device types. In particular, we look at the device types used by purchasers and non-purchasers and analyze device switches.
\subsubsection{Device type}
First, we want to understand how many users are using multiple devices and which devices customers use for purchase and non-purchase sessions.
\begin{table}[!htb]
\caption{User device statistics per session.}
\label{tab:ds_user_device}
\begin{tabular}{lrr}
\toprule
\bf Device(s) & \bf Purchasers (\%) & \bf Non-purchasers (\%) \\
\midrule
> 1 device & 24.05 & 16.22
\\
\midrule
1 device & 75.95
& 83.78 \\
2 devices & 22.23 &
15.39 \\
3 devices & 1.82 & 0.82 \\
4 devices & $\approx 0$ & $\approx 0$ \\
5 devices & 0 & 0 \\
\bottomrule
\end{tabular}
\end{table}
\noindent%
Table~\ref{tab:ds_user_device} shows how many devices purchasers and non-pur\-chas\-ers own.
The majority of users from both groups are single-device users. However, the fraction of single-device purchasers is $9.35\%$ smaller than the corresponding fraction of non-purchasers.
On the other hand, the fraction of multi-device users for purchasers is $45.28\%$ bigger than the corresponding fraction for non-purchasers. In general, multi-device users represent almost a quarter of the purchasers.
As the number of devices increases, the difference between purchasers and non-purchasers grows.
Our observations support the statement that multi-device users tend to be more engaged~\citep{montanez2014cross}.
\begin{table}[!h]
\caption{Purchase and non-purchase sessions per device type and standardized conversion rates.}
\label{tab:ds_device_purchase}
\begin{tabular}{l @{} r r r}
\toprule
\multirow{2}{*}{\bf Device} & \bf Purchase & \bf Non-purchase & \bf Stand.\ \\
& \bf sessions (\%) &
\bf sessions (\%) & \bf conv.\ rate \\
\midrule
Smartphone & 47.00\phantom{0} & 58.09\phantom{0} & $-$0.56 \\
PC & 44.97\phantom{0} & 34.40\phantom{0} & 1.61 \\
Tablet
& 8.03\phantom{0} & 7.50\phantom{0}
& 0.61 \\
Game Console
& 0.004 &
0.004 & $-$0.40 \\
TV & 0.001
& 0.002 & $-$1.25 \\
\bottomrule
\end{tabular}
\end{table}
Next, we examine the distributions of purchase and non-purchase sessions across device types and device-specific standardized conversion rates; see Table~\ref{tab:ds_device_purchase}.
The PC is the device with the highest conversion rate. Indeed, the fraction of purchase sessions is $30.70\%$ bigger than the fraction of non-purchase sessions.
The device with the second-highest conversion rate is a tablet. For this device, purchase sessions are $7.12\%$ more frequent than non-purchase sessions.
The Smartphone is the device with the second-lowest conversion rate. For this device, the number of purchase sessions is $19.10\%$ less frequent than the number of non-purchase sessions.
Game consoles and TVs are relatively new devices in e-commerce; hence, sessions with these devices are relatively less frequent.
Nevertheless, based on our observations, we find that the game console is a device with the third-highest conversion rate. Interestingly, its conversion rate is close to that of the smartphone. It can be explained by the fact that device functionalities of smartphones and tablets in e-commerce context blur due to the similarity of their interfaces and screen sized. The number of purchase sessions on a game console is $15.47\%$ less than the number of non-purchase sessions. The TV is the least common device, with the lowest conversion rate. The number of purchase sessions on this device is $34.01$ less than the number of non-purchase sessions.
We can explain our findings by the fact that customers use different devices for different purposes. For example, PCs and tablets seem to be used for the purchase, whereas smartphones, game consoles, and TVs for exploration.
\subsubsection{Device switches}
Next, we analyze how users switch between devices before a purchase session.
\begin{figure}[!h]
\includegraphics[width=0.9\linewidth]{5_devices_transitions_incl_self_transitions_1}
\caption{Device transition probability before a purchase session, including self-transitions. The thickness of an arrow indicates the connection strength; the dashed line is the weakest connection.}
\label{fig:device_transitions_incl_self_transitions}
\end{figure}
Figure~\ref{fig:device_transitions_incl_self_transitions} shows device transition probability, including self-transi\-tions.
Generally, the situation when a user remains on the same device is the most likely outcome for all devices, except TV. There, the self-transition probability is lower than the probability of switching from TV to PC, a device with the highest self-transition probability. A probability of remaining on a smartphone is $5.03\%$ lower than a self-transition probability for PC, whereas a probability to remain on a tablet is $17.28\%$ lower than the probability to remain on PC. The game console has the second-lowest self-transition probability.
Next, we consider connections between two different devices. We characterize those interconnections based on how likely a user is to switch from one device to another one and vice versa.
\paragraph{Strong interconnections}
Some pairs of devices have high probability interconnections.
The strongest connection is between a smartphone and a PC, the two most popular devices.
The second strongest connection is between PC and tablet. There is a bigger discrepancy between probability rates, with the probability of switching to PC being $656.86\%$ higher than of switching to tablet.
The third strongest interconnection is between a smartphone and a tablet with a stronger connection switch to a smartphone, a more popular device. The probability of switching to a smartphone is $399.17\%$ higher than switching to a tablet.
Overall, the three interconnections form a triangle that includes the three most popular devices: PC, smartphone, and tablet.
\paragraph{One-sided interconnections}
For a one-sided interconnection there is a high probability of switching from one device to another, but a close to zero probability of switching back. There are six cases of this type in Figure~\ref{fig:device_transitions_incl_self_transitions}.
TV is the device with the largest number of one-sided interconnections, with PC, smartphone, tablet, and game console. In all cases, the transition probability is low when TV is a target device, which can be explained by relative difficulty to purchase on TV.
The strongest one-sided interconnection is between TV and PC. The probability of switching from TV to PC is $43.75\%$, the highest transition probability for TV, and the highest probability to transition to another device. We explain this by the fact that PC is one of the most popular devices for purchase.
The second most likely device people switch to from TV is a smartphone, whereas a probability to switch to a tablet or game console is $6.25\%$.
Another device with a significant number of one-sided interconnections is the game console. Apart from the connection with TV discussed above, the device also has this connection type with smartphone and PC. The transition probability is close to zero when a game console is a target device. Unlike the situation with TV, came console has a higher probability of switching to a smartphone, whereas the probability of switching to PC is $1.75\%$ less.
In general, all one-sided interconnection cases include switching from a less common device type such as game console or TV to a more conventional device, such as PC, smartphone, or tablet.
\paragraph{Weak interconnections}
In some cases, the switch between two devices rarely happens, i.e., the transition probability is close to zero. As can be seen in Figure~\ref{fig:device_transitions_incl_self_transitions}, there is only one case of this type. It is a connection between a game console and a tablet.
\medskip\noindent%
Overall, the analysis of device switches before a purchase session supports the conclusion that users tend to switch from less popular devices such as TV and game console to more popular ones such as PC, smartphone, and tablet.
\subsection{Queries}
The next aspect of purchase intent that we examine is queries. We look at the number of queries in purchase and non-purchase sessions and per device type. In total, the dataset contains \numprint{31185176} queries, \numprint{1302195} or 4.17\% of which are unique. Given the number of sessions in the dataset, we can conclude that queries are infrequent.
\begin{table}[!h]
\caption{Queries per session for purchase and non-purchase sessions per device type, percentages are computed w.r.t. total number of queries per purchase or non-purchase session.}
\label{tab:ds_device_query}
\begin{tabular}{l @{} rr rr}
\toprule
\multirow{2}{*}{\bf Device} & \multicolumn{2}{c}{\bf Purchase sessions} & \multicolumn{2}{c}{\bf Non-purchase sessions}\\
\cmidrule(r){2-3}\cmidrule{4-5}
& \bf query/session &
\bf \% & \bf query/session &
\bf \%
\\
\midrule
Smartphone & $ 4 $ & \textbf{52.92} & $ 0.05$ & 42.40 \\
PC & $ 2 $ & 36.38 & $0.09$ & \textbf{57.54} \\
Tablet & $ 4 $ & 10.67 & $0.0003$ & 0.049 \\
Game Console & $\approx 0$ & $\approx 0$ & $\approx 0$ & $\approx 0$ \\
TV & $\approx 0$ & $\approx 0$ & $\approx 0$ & $\approx 0$ \\
\midrule
Avg & $3.16$ & 100 & $0.06$ & 100 \\
\bottomrule
\end{tabular}
\end{table}
Table~\ref{tab:ds_device_query} shows the query per session frequencies across five devices for both purchase and non-purchase sessions. Besides, it also demonstrates which devices are most popular for querying during purchase and non-purchase sessions. Overall, queries are more common in purchase sessions. This can be explained by the fact that querying is more likely to happen when customers are determined to buy something.
Naturally, queries are most common for smartphones, PCs and tablets, and uncommon for game consoles and TVs. Indeed, the current interface of game console and TV makes it difficult to type queries, especially when compared to a PC or a smartphone.
The PC has the highest query per session frequency for non-purchase sessions and second-highest frequency for purchase sessions. A smartphone has the second-highest query per session frequency for non-purchase sessions and the highest query frequency per non-purchase session. Tablet, on the contrary, has the third-highest frequency for non-purchase sessions and the highest frequency for purchase sessions.
When it comes to query distributions per device for purchase and non-purchase sessions, the ranking is somewhat consistent for both groups. During purchase sessions, most queries are issued on a smartphone, whereas during non-purchase sessions PC prevails. On the other hand, PC is the second most popular device for purchase sessions, whereas for non-purchase sessions smartphone takes the second place. Tablet is third for both groups. In general, the query distribution across devices correlates with the session distribution across devices (see Table~\ref{tab:ds_device_purchase}).
\begin{table}[!h]
\caption{Unique query counts for purchase and non-purchase sessions per device type. The percentage is computed w.r.t. total number of queries per purchase or non-purchase session.}
\label{tab:ds_device_unique_query}
\begin{tabular}{l rr rr}
\toprule
\multirow{2}{*}{\bf Device} & \multicolumn{2}{c}{\bf Purchase sessions} & \multicolumn{2}{c}{\bf Non-purchase sessions}\\
\cmidrule(r){2-3}\cmidrule{4-5}
& \bf count &
\bf\% & \bf count &
\bf \% \\
\midrule
Smartphone
& \numprint{180542} & 36.89 & \numprint{321640} & 39.57 \\
PC
& \numprint{224763}
& 45.92 & \numprint{312855} & 38.49 \\
Tablet & \numprint{83881} & 17.14 & \numprint{175909} & 21.64 \\
Game Console
& \numprint{136} &
0.03 & \numprint{1841} & 0.23
\\
TV & \numprint{46} & 0.01
& \numprint{582} &
0.07 \\
\midrule
Total & \numprint{489368} &
1.91 & \numprint{812827} & 14.52 \\
\bottomrule
\end{tabular}
\end{table}
\noindent%
Next, we look at the number of unique queries for purchase and non-purchase sessions and per device. Table~\ref{tab:ds_device_unique_query} shows unique queries count and their corresponding fractions. The fractions are computed w.r.t. the total number of queries per session type and device. Overall, during purchase sessions users issue less unique queries, it holds for every device class but a PC. This can be explained by the fact that during purchase sessions users may retype a previous query to revisit the results they have seen earlier, whereas non-purchasers want to explore and hence use more unique queries.
\subsection{Purchase Intent Characteristics}
What have we learned from the log analysis conducted in this section that might help us to devise better models for purchase intent prediction?
We found out that purchase sessions tend to be longer what suggests that session length is an essential indicator of purchase intent. Besides, the difference in session length depends on the type of device customer use.
Moreover, we discovered how the day of week and hour of the day influence purchase behavior. In particular, customers are more likely to buy during the weekdays and in the evening.
From the perspective of channels, there is a difference, too. In particular, for non-purchase sessions, the direct channel is more common, whereas purchase sessions are more likely to start with paid or organic channels.
From the device perspective, we found out that multi-device users are more common among purchasers.
Besides, we figured the probability of purchase for every device and characterized transitions between devices.
After looking into queries in the dataset, we discovered that during purchase sessions, users issue more queries per session. Besides, during purchase session, there are less unique queries.
\section{Modeling Purchase Intent}
\label{sec:modelingpurchaseintent}
Next, we use our insights from the previous section to boost purchase intent prediction performance. Since our initial analysis demonstrated that about 56\% of all sessions are anonymous (see Table~\ref{tab:ds_stats}), it is worth to pay special attention to this category.
Thus, we analyze model performance in two settings: when a user is anonymous (the anonymous setting) and when a user is logged in (the identified setting).
Based on the findings obtained thus far and on an analysis of best-performing features available in the literature \citep{hop2013web, lee2015online, niu2017predictive, seippel-2018-customer}, we compile a feature set presented in the Table~\ref{tab:feature_set}. Initially, we added information about previous device type and conversion rate, but we subsequently excluded it from the set because the features decreased classifier performance.
We categorize features into two classes: session features and customer history features. We derive session features from the information of the given session and base customer history features on the information from previous sessions of the given customer.
Since we run experiments in anonymous and identified setting, we use different feature set for each setting. For anonymous setting, when a customer is not known, the information about the customer is not available and, therefore, we can only use session features. On the other hand, when a customer is identified, we can use both session and customer history features.
The feature set contains both static and dynamic features. Dynamic features can change throughout the session, whereas static features remain constant.
After reviewing previous work in the domain of purchase prediction (see Section~\ref{sec:relatedwork}), we choose several models for our experiments: logistic regression, K-nearest neighbors, support vector machines, neural classifier, gradient boosting decision tree, and random forest.
\begin{table}[!h]
\caption{Complete feature set. ``Dynamic'' indicates that a feature may change during a session.}
\label{tab:feature_set}
\begin{tabular}{llcc}
\toprule
\multicolumn{2}{l}{\bf Feature} & \bf Dynamic & \bf Baseline \\
\midrule
\multirow{8}{*}[-0.1cm]{\rotatebox{90}{Session}}
& current page dwell time, mean & \checkmark & \checkmark\\
& current page dwell time, $\sigma$ & \checkmark & \checkmark \\
& page sequence score & \checkmark & \checkmark \\
& number of pages & \checkmark & \checkmark\\
& channel type \\
& start hour \\
& week day \\
& device type \\
& device conversion rate \\
\midrule
\multirow{9}{*}[0.5cm]{\rotatebox{90}{History}}
& number of orders & & \checkmark \\
& days since last purchase & & \checkmark \\
& number of sessions \\
& number of devices \\
& device sequence score \\
& switch probability \\
\bottomrule
\end{tabular}
\end{table}
\section{Discussion \& Conclusion}
\label{sec:discussionandimplications}
In this paper, we have carried out an analysis of user purchase intent in e-commerce. We have analyzed four weeks of session logs from a European e-commerce platform to identify signals in user behavior that can imply purchase intent. We have considered aspects such as session length, day of the week, and session start hour, as well as information about device, channel, and queries.
In the second part of our study, we have analyzed the relevance of the discovered signals by running a series of experiments aimed at purchase intent prediction in the anonymous and identified settings. We tested the features on random forest, the model which fits production requirements. Additionally, we tested the features on five other models. The experiments demonstrated the value of the features that we engineered based on our insights into the data. We explored which features contribute to performance improvement.
One of the implications of our study is enhanced understanding of purchasing user behavior in e-commerce. Understanding the behavior is the first step towards modeling it, as we demonstrated in the second part of the paper.
Modeling user behavior can contribute towards reducing friction in the customer journey and, therefore, to better customer experience.
Besides, we explored the topic of detecting the purchase intent of anonymous users. We showed that, while anonymous users contribute to more than half of the traffic, their user intent is harder to detect because all the predictions have to be made without knowledge about the prior behavior.
Our research has several limitations; one of them is limited generalizability. Even though the data we use in our study comes from a dominant e-commerce platform, it is still only one platform. Hence, it would be interesting to verify the findings against other e-commerce platforms and explore the differences.
Moreover, we sampled four weeks of data, thereby introducing a sample bias that could make our findings sensitive to unknown temporal or seasonal patterns. Therefore, it would be interesting to explore if expanding our dataset will lead to new insights. For example, if we had several months of data, we could explore how user purchase intent changes across different months or seasons.
Furthermore, we evaluated our purchase intent prediction models in an offline setting. The next logical step is to evaluate them in an online setting.
Future research on the topic includes several directions.
First, there is an opportunity to continue research into general purchase behavior analysis and modeling in e-commerce. It would be interesting to explore more aspects of purchasing behavior and try out more models.
Another direction for further research concerns predicting purchase intent for anonymous users.
Another exciting direction for further research includes modeling device-specific purchase behavior. It can include both relatively common devices such as PC, smartphone, and tablet, and relatively less popular and studied devices such as TV or game console.
\section{Reproducibility}
All plots for our paper, as well as the code to regenerate them, can be found in our Git repository:\\ \url{https://github.com/mariyahendriksen/purchase_intent}.
\section{Related Work}
\label{sec:relatedwork}
\header{E-commerce user purchase behavior analysis} Research on understanding online users' purchasing behavior has been ongoing since the very beginning of e-commerce~\citep{bellman1999predictors}. Studies have investigated user motivation~\citep{bellman1999predictors}, factors that influence e-commerce adoption~\citep{o2003web}, as well as purchasing behavior~\citep{brown2003buying, hsu2006longitudinal}, with a focus on perceived security~\citep{salisbury2001perceived, george2002influences}, the decision-making process~\citep{senecal2005consumers}, and purchaser profiles~\citep{swinyard2004activities, hernandez2011age}. Besides, there has also been work on user behavior on content discovery platforms and its relationship to subsequent purchases~\citep{lo2016understanding} as well as work dedicated to the identification of a taxonomy of product search intents and the prediction of user satisfaction~\citep{su2018user}.
Unlike previous work, our study focuses on the exploration of user purchasing behavior by comparing purchase vs. non-purchase sessions. Besides, we analyze the data from the perspective of device types, and explore aspects such as session length, price of the seen products from the perspective of different devices. On top of that, we also look into the way customer switches between devices.
\header{Purchase prediction in e-commerce} The problem of e-commerce user behavior modeling has been studied from various angles, such as building multiple classifiers based on genetic algorithms~\cite{kim2003combination}, mining purchase patterns with association rules and using those patterns for purchase prediction~\cite{suh2004prediction}. Research has been focused on creating models robust to noise in session data~\cite{agichtein2006learning}, and using a recurrent neural network to predict customer behaviour~\cite{lang2017understanding}.
\citet{sismeiro2004modeling} predict purchasing task completion for a given user who completed at least one task earlier, whereas \citet{cheng2017predicting} explore user behavior on a content discovery platform to determine intent specificity and time in the future when a purchase is estimated to take place. Some work in the field focuses on using queries for purchasing behavior modeling. For instance, ~\citet{dai2006detecting} predict purchase based on input query. Besides using general session data, there has been work that incorporates demographic data and perceived attributes~\cite{young2004predicting}, scrolling and mouse movements~\cite{guo2010ready}, payment data~\cite{wen2018customer}, log-trace data~\cite{tao2019log2intent}, and phone touch actions~\cite{guo2019buying}. There has been work on analyzing behavioral patterns and the exploration of different model architectures. In particular, support vector machines, K-nearest neighbor approach, random forest, and logistic regression were used~\citep{lee2015online, suchacka2015k, niu2017predictive}.
Unlike previous work in this domain, our study focuses on purchase prediction with two types of users, identified and anonymous. Therefore, we develop two models, run them in two settings, and evaluate their results. The possibility to experiment with identified users also allows us to leverage information from previous user sessions, such as user purchasing history and the number of devices a user owns. In contrast, anonymous users contribute to a higher share of traffic, which makes it important to understand their behavior too. Additionally, we explore how the relevance of dynamic and static features changes as a session progresses.
\section{Background and Definitions}
\label{sec:backgroundanddefinitions}
In our study, we operate with the following definitions.
A \textit{session} is a sequence of requests made by a single end-user during a visit to a particular site. A session ends if the user is idle for more than 30 minutes.
We define two types of sessions: \textit{purchase sessions}, during which the customer buys an item, and \textit{non-purchase sessions}, during which the customer does not buy anything.
In connection to this, we define \emph{purchasers} as customers who had at least one purchase session, whereas \emph{non-purchasers} are customers who were identified but have never purchased anything.
We furthermore distinguish between \textit{identified sessions}, where a customer is logged in or recognized with a browser cookie, and \textit{anonymous} sessions where this is not the case.
Additionally, we denote the number of actions taken during a given session as the \textit{session length}, where an \textit{action} corresponds to opening a new web page, submitting a search query, or adding/removing an item to/from the shopping basket.
\textit{Device switch} is the act of changing the type of browsing device between two consecutive sessions that belong to the same journey.
For instance, if a customer first explores the platform on a smartphone and afterward accesses the platform on a PC, she switches from a smartphone to a PC.
A \textit{channel} indicates the way through which a customer enters the platform.
For example, if the customer comes to the platform via an advertisement, she uses a paid channel.
The \textit{conversion rate} denotes the fraction of visits during which a purchase was made~\citep{moe2004dynamic}.
We use this metric to compare device popularity in a purchasing context.
We calculate the conversion rate by dividing the number of purchasing sessions by the overall number of sessions.
In order to protect sensitive information, we only report \textit{standardized conversion rates} for each device; since we are interested in differences across devices types, this suffices for our purposes.
The standardized conversion rate is computed by subtracting the mean conversion rate per device type from the desired conversion rate and dividing the result by the standard deviation of the device-specific conversion rate.
For instance, if our device specific conversion rates are
$\text{\em Conversion Rates} = \{0.5, 0.2, 0.3\}$, the mean of device-specific conversion rates is $\overline{\text{\em Conversion Rate}} = 0.33$ and the standard deviation of conversion rates is $\sigma = 0.12$.
Therefore, the resulting standardized conversion rates are $\text{\em Standardized Conversion}$ $\text{\em Rates} = \{1.34, -1.07, -0.27\}$
\section{Dataset Description}
\label{sec:3}
We start by distinguishing two search schemes and explaining how we define each of the two. Afterward, we describe how exploratory and purchasing sessions were mined from the search logs and summarize dataset statistics.
\subsection{Definitions}
\label{subsec:definitions}
The data set we use comprises 4 weeks of sessions sampled from a European e-commerce website. In the study, we operate with the following definitions:
\begin{description}
\item[Session] is a sequence of requests made by a single-end user during a visit to a particular site. Session ends if user is idle for more than 30 minutes.
We define two types of sessions: purchase and non-purchase session. Purchase session is a session during which customer buys something. Non-purchase session is a session during which customer does not buy anything.
\item[Session length] is a number of actions taken during given session.
\item[Step] is a number of pages customer seen during given session.
\item[Actions] include opening a new web page, submitting a search query.
\item[Device switch] is the act of changing device type in-between sessions.
\item[Moving average]
\item[Moving standard deviation]
\item[Markov Score]
\item[Conversion rate] is a fraction of the number of visits during which a purchase was made~\cite{moe2004dynamic}. We use this metric to compare device popularity in a purchasing context. The conversion rate can be calculated as follows:
\begin{equation}
\text{\em Conversion rate} = \frac{\text{\em\#purchase\ sessions}}{\text{\em\#all\ sessions}}
\end{equation}
\item[Standardized conversion rate] is used to protect sensitive information. We present standardized conversion rates for each device; since we are interested in differences across devices types, this suffices for our purposes. The \emph{standardized conversation rate} is computed as follows:
\begin{equation}
\begin{split}
\text{\em Standardized}&\text{\em\ conversion rate} = \\
& \frac{\text{\em Conversion rate} - \overline{\text{\em Conversion rate}}}{\sigma}
\end{split}
\end{equation}
where:
\begin{itemize}
\item $\overline{\text{\em Conversion rate}}$ is the mean conversion rate per device type
\item $\sigma$ is standard deviation of device-specific conversion rates
\end{itemize}
\end{description}
\subsection{Dataset Sampling}
The data set we use comprises 4 weeks (28 days) of anonymized visits sampled from a European e-commerce website for October 2019.
Log entries include a unique customer identifier, the class of device the customer used during the session, a timestamp for every query and a URL of each clicked page.
Besides, we also keep track of the price of every product the customer has seen and the price of the item(s) he or she ended up buying.
In cases where a customer starts a session without logging in and ends up logging in at a later point in the session, we assign the session to the customer.
To filter out bots traffic, we apply several measures related to device type, location, and visit length. First, we specify a set of particular device types we are looking for and remove all the entries from different devices. Moreover, we filter out based on location, including only entries from the European countries from which the majority of the customers come from.
Besides, we specify visit length, which allows us to get rid of both very short, uninformative visits and abnormally long visits of crawlers~\citep{bomhardt-2005-web}.
\subsection{Dataset Statistics}
TBA
\begin{table}[!htb]
\caption{Dataset statistics.}
\label{tab:ds_stats}
\begin{tabular}{ll}
\toprule
\bf Description & \bf Total \\
\midrule
Sessions (total) & \numprint{95757177} \\
Sessions (logged-in customers) & \numprint{41705176} \\
Sessions (anonymous customers) & \numprint{54144152} \\
Identified customers & \numprint{9663509} \\
Queries & \numprint{42927} \\
Devices & PC, Smartphone, Tablet, Game Console, TV \\
\bottomrule
\end{tabular}
\end{table}
\section{Dataset Description}
\label{sec:3}
\todo{Due: Dec 6}
We start off by distinguishing two search schemes and explaining how we define each of the two. Afterwards, we describe how exploratory and purchasing sessions were mined from the search logs and summarize dataset statistics.
\subsection{Definitions}
This sections describes all the definitions related to the dataset.
Visit is a
Conversion rate is a fraction of visits during which purchase was made (CITE). It can be calculated as follows: TBA
\begin{equation}
Conversion\ rate = \frac{purchase\ visits}{all\ visits}
\end{equation}
Customer journey
We define \textit{distance} between two consecutive sessions $s_t$ and $s_{t+1}$ as the time difference between the last action in the session $s_t$ and the first action of session $s_{t+1}$.
\subsection{Dataset Sampling}
This section explains how dataset was sampled and preprocessed.
\subsection{Dataset Statistics}
This section describes general dataset statistics and
The dataset statistics is present in the Table~\ref{tab:ds_stats}.
From the Table~\ref{tab:ds_stats}, it is clear that an average length of the visits in terms of actions is about six times longer for the sessions where users completed purchase as opposed to the visits where users did not purchase anything.
\begin{table}[!htb]
\caption{Dataset statistics.}
\label{tab:ds_stats}
\begin{tabular}{lrrr}
\toprule
Description & Purchase & Non-purchase & Total \\
\midrule
Data size, GB & \numprint{74.12} & \numprint{80.8} & \numprint{0} \\
Sessions & \numprint{3531267} & \numprint{27719626} & \numprint{0} \\
Visits & \numprint{3649993} & \numprint{38055183} & \numprint{0} \\
\addlinespace
Rows & \numprint{175725615} & \numprint{234326049} & \\
Avg. visit length & \numprint{48.14} & \numprint{6.16} & \\
Customers & \numprint{0} & \numprint{0} & \\
TBA & \numprint{0} & \numprint{0} & \\
\bottomrule
\end{tabular}
\end{table}
As can be seen form the Table~\ref{tab:ds_device}, the most popular device types are phone and desktop. Interestingly, most purchase visits are from desktop, whereas for most non-purchase visits, phone is prevailed over desktop by $\approx 1\%$.
\begin{figure}
\includegraphics[width=\linewidth]{purchase_visit_frequency_per_user}
\caption{How many purchasing visits one customer had. The plot demonstrates that about 72\% of customers had only 1 purchasing visit, 28\% - more than one.}
\label{fig:purchase_visit_freq}
\end{figure}
\begin{table}[!htb]
\caption{Dataset characteristics from the device perspective.}
\label{tab:ds_device}
\begin{tabular}{lrrrr}
\toprule
\multirow{2}{*}{Device} & \multicolumn{2}{c}{Purchase visits} & \multicolumn{2}{c}{Non-purchase visits} \\
\cmidrule(r){2-3}\cmidrule{4-5}
& size, gb & percentage & size, gb & percentage \\
\midrule
Phone & \numprint{29.85} & 40.28 & 37.55 & 46.48 \\
Desktop & 38.40 & 51.81 & 36.56 & 45.25 \\
Tablet & 5.86 & 7.91 & 6.68 & 8.27 \\
Game Console & $4.49 \times 10^{-3}$ & $\approx 0$ & $6.43 \times 10^{-3}$ & $\approx 0$ \\
TV & $9.11 \times 10^{-4}$ & $\approx 0$ & $1.86 \times 10^{-3}$ & $\approx 0$ \\
\addlinespace
Total & 74.11 & 100 & 80.79 & 100 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[!htb]
\caption{User device statistics. \todo{Should I make it bar chart with two variables? It leaves out some data, but shows a clear pattern}}
\label{tab:ds_user_device}
\begin{tabular}{lrrrr}
\toprule
\multirow{2}{*}{Device(s)} & \multicolumn{2}{c}{Purchasers} & \multicolumn{2}{c}{Non-purchasers} \\
\cmidrule(r){2-3}\cmidrule{4-5}
& users & percentage & users & percentage \\
\midrule
Any Device & \numprint{2575877} & 100 & \numprint{5795495} & 100 \\
> 1 device & \numprint{173965} & 6.75 & \numprint{1330202} & 22.95 \\
\addlinespace
1 device & \numprint{2401912} & 93.25 & \numprint{4465293} & 77.05 \\
2 devices & \numprint{169611} & 6.58 & \numprint{1232975} & 21.27 \\
3 devices & \numprint{4353} & 0.17 & \numprint{97172} & 1.68 \\
4 devices & \numprint{1} & $\approx 0$ & \numprint{55} & $\approx 0$ \\
5 devices & \numprint{0} & 0 & 0 & 0 \\
\bottomrule
\end{tabular}
\end{table}
\begin{comment}
\begin{figure}
\includegraphics[width=\linewidth]{user_device_chart}
\caption{Fraction of users who own 1, 2, 3, 4, 5 devices. Temporary chart.}
\label{fig:user_device}
\end{figure}
\end{comment}
Table~\ref{tab:ds_user_device} shows that majority of users from both groups use only one device. What is more a bigger fraction of purchasers stick to one device when compared to non-purchasers. In case with two- and three-device users, the situation is opposite: non-purchasers fraction if bigger.
We would like to know what kind of devices users end up buying with. We count purchasing visits per device and compare counts with non-purchasing visits. The answer is present in the Table~\ref{tab:ds_device_purchase}. As can be seen, the most popular device for purchases is mobile phone, however, desktop has highest conversion rate.
\begin{table}[!htb]
\caption{Purchase visit counts across devices.}
\label{tab:ds_device_purchase}
\begin{tabular}{l rr rr r}
\toprule
\multirow{2}{*}{Device} & \multicolumn{2}{c}{Purchase visits} & \multicolumn{2}{c}{Non-purchase visits} & \multirow{2}{*}{Conv. rate} \\
\cmidrule(r){2-3}\cmidrule{4-5}
& visits & percentage & visits & percentage \\
\midrule
Phone & \numprint{1716976} & 47.00 & \numprint{22121333} & 58.09 & 7.20 \\
Desktop & \numprint{1642820} & 44.97 & \numprint{13102228} & 34.40 & 11.14 \\
Tablet & \numprint{293463} & 8.03 & \numprint{2855728} & 7.50 & 9.32 \\
Game Console & \numprint{145} & $\approx 0$ & \numprint{1788} & $\approx 0$ & 7.50 \\
TV & \numprint{44} & $\approx 0$ & \numprint{695} & $\approx 0$ & 5.95 \\
\addlinespace
Total & \numprint{3653448} & 100 & \numprint{38081772} & 100 & 8.75 \\
\bottomrule
\end{tabular}
\end{table}
\section{Tmp data}
\todo{Due: Dec 26}
\begin{table}[!htb]
\caption{Dataset characteristic from device perspective.}
\label{tab:ds_device}
\begin{tabular}{lrrrr}
\toprule
\multirow{2}{*}{Device} & \multicolumn{2}{r}{Purchases visits} & \multicolumn{2}{r}{}Non-purchase visits \\
& size, gb & percentage & size, gb & percentage \\
\midrule
Phone & \numprint{29.85} & 40.28 & 37.55 & 46.48 \\
Desktop & 38.40 & 51.81 & 36.56 & 45.25 \\
Tablet & 5.86 & 7.91 & 6.68 & 8.27 \\
Game Console & $4.49 \times 10^{-3}$ & $\approx 0$ & $6.43 \times 10^{-3}$ & $\approx 0$ \\
TV & $9.11 \times 10^{-4}$ & $\approx 0$ & $1.86 \times 10^{-3}$ & $\approx 0$ \\
\addlinespace
Total & 74.11 & 100 & 80.79 & 100 \\
\bottomrule
\end{tabular}
\end{table}
| d63252bc8cd412325678b3e39c457913d170882d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let us consider a vortex sheet $\omega$: a weak solution of the 2D Euler equation concentrated on a simple closed curve $\Gamma:=\left\{ z(\theta,t) \in \R^2 : \theta \in \mathbb{T} \right\}$ with vortex-sheet strength $\gamma(\theta,t)$, that is, for all test functions $\varphi \in C^{\infty}_c(\R^2)$, it holds that
\begin{align*}
\int_{\R^2} \varphi(x)d\omega(x,t) = \int_{-\pi}^{\pi} \varphi(z(\theta,t)) \gamma(\theta,t)d\theta.
\end{align*}
Note that the evolution of $z$ and $\gamma$ is described by
\begin{align}
\partial_t z(\theta,t) = \mathbf{v}(z,t) + c(\theta,t)\partial_\theta z(\theta,t) \label{evolv_curve}\\
\partial_t \gamma(\theta,t) = \partial_\theta\left( c(\theta,t)\gamma(\theta,t)\right), \label{evolv_strength}
\end{align}
where $\text{curl}(\mathbf{v}) = \omega$ and $c(\theta,t)$ represents the reparametrization freedom of the curve \cite{Castro-Cordoba-Gancedo:naive-vortex-sheet,Lopes-Nussenzveig-Schochet:vortex-sheets-BR-formulation,Majda-Bertozzi:vorticity-incompressible-flow,Sulem-Sulem:finite-time-analyticity-rt}. Recall that the (discontinuous) velocity $\mathbf{v}$ on the curve, generated by the vortex sheet on $\Gamma$ is given by the Birkhoff-Rott integral:
\begin{align}\label{def_BR}
\mathbf{v}(z(\theta,t),t) : =BR(z,\gamma)(z(\theta,t)) := \frac{1}{2\pi} PV\int_{-\pi}^{\pi} \frac{(z(\theta,t)-z(\eta,t))^{\perp}}{|z(\theta,t) - z(\eta,t)|^2} \gamma(\eta,t) d\eta.
\end{align}
For simplicity, we will omit the word $PV$ in the notation for principal value integral from now on. We will also denote by $\mathbf{v^{\pm}}$ the respective limits of the velocity on the two sides of $\Gamma$ (with $\mathbf{v}^+$ being the limit on the side that $\mathbf{n}$ points into).
The main goal of this paper is to find a class of uniformly-rotating vortex sheets, concentrated on a closed curve which is not a circle. This is the first proof of existence of a family of solutions of such kind. We will say that a vortex sheet is uniformly-rotating with angular velocity $\Omega$ if it is stationary in the rotating frame with angular velocity $\Omega$. See Lemma~\ref{lemma_br_eq} for the exact equations satisfied by a uniformly-rotating vortex sheet. The existence of such solutions is not evident a priori since there are rigidity results by G\'omez-Serrano--Park--Shi--Yao \cite{GomezSerrano-Park-Shi-Yao:rotating-solutions-vortex-sheet-rigidity} ruling out their existence in the case $\Omega \leq 0$ and $\gamma > 0$.
These solutions are also important since they show that one cannot expect any asymptotic stability of the radial vortex sheet with constant strength. See the recent work by Ionescu--Jia \cite{Ionescu-Jia:axisymmetrization-vortex-point} for an asymptotic stability result when the vorticity is made out of a Dirac delta part and a Gevrey smooth part. Previously, similar solutions (uniformly rotating, non-radial) had been found for the 2D Euler equations in the context of vortex patches or smooth, compactly-supported functions \cite{Castro-Cordoba-GomezSerrano:uniformly-rotating-smooth-euler,Castro-Cordoba-GomezSerrano:analytic-vstates-ellipses,Burbea:motions-vortex-patches,Hmidi-Mateu-Verdera:rotating-vortex-patch,Garcia-Hmidi-Soler:non-uniform-vstates-euler,Hassainia-Masmoudi-Wheeler:global-bifurcation-vortex-patches}.
Despite the complexity of the solutions shown in numerical/actual experiments \cite{Krasny:vortex-sheet-icm,VanDyke:fluid-book,Majda:vortex-sheet-iciam}, there have been significant efforts to prove the existence of solutions to \eqref{evolv_curve}--\eqref{evolv_strength} in various settings. For a $L^2_{loc}$ initial velocity whose vorticity has a definite sign, it turns out that there exists a global weak solution \cite{Delort:vortex-sheet,Majda:vortex-sheet} by the works of Delort, and Majda. In case that the vorticity does not have a definite sign, the existence was proved by Lopes Filho--Nussenzveig Lopes--Xin \cite{Lopes-Nussenzveig-Xin:vortex-sheets-reflection} under the assumption that the initial vorticity satisfies a reflection symmetry. For analytic initial data, local-in-time existence of analytic solutions was proved by Sulem--Sulem--Bardos--Frisch \cite{Sulem-Sulem:finite-time-analyticity-rt}.
The singularity formation was conjectured by Birkhoff--Fisher and Birkhoff \cite{Birkhoff-Fisher:vortex-sheets-roll,Birkhoff:helmholtz-taylor-instability}. For analytic initial data, the possibility that the curvature may blow up in finite time was supported by asymptotic analysis of Moore \cite{Moore:singularity-vortex-sheet} and also verified by numerical simulations by Krasny and Meiron--Baker--Orszag in \cite{Krasny:singularity-vortex-sheet-point-vortex,Meiron-Baker-Orszag:vortex-sheet}. Note that the system \eqref{evolv_curve} and \eqref{evolv_strength} is known to be ill-posed in $H^{s}$ for $s>\frac{3}{2}$ \cite{Caflisch-Orellana:singular-solutions-ill-posedness-vortex-sheet}. For more comprehensive discussion on the well-posedness theory, we refer to \cite{Majda-Bertozzi:vorticity-incompressible-flow,Saffman:book-vortex-dynamics,Wu:vortex-sheet}.
\subsection{Steady solutions of the vortex sheet}
There are very few known examples of nontrivial steady solutions, and in fact, other than the circle or the line, the list only comprises the segment of length $2a$ and density
\begin{equation}\label{example_rotate}
\gamma(x) = \Omega \sqrt{a^2 - x^2}, \qquad x \in [-a,a],
\end{equation}
which is a rotating solution with angular velocity $\Omega$ \cite{Batchelor:book-fluid-dynamics} and the family found by Protas--Sakajo \cite{Protas-Sakajo:rotating-equilibria-vortex-sheet}, made out of segments rotating about a common center of rotation with endpoints at the vertices of a regular polygon. We remark that none of these are supported on a closed curve.
Numerically, O'Neil \cite{ONeil:relative-equilibria-vortex-sheets,ONeil:collapse-vortex-sheets} used point vortices to approximate the vortex sheet and compute uniformly rotating solutions and Elling \cite{Elling:vortex-sheet-cusps} constructed numerically self-similar vortex sheets forming cusps. O'Neil \cite{ONeil:point-vortices-vortex-sheets,ONeil:point-vortices-vortex-sheets-pof} also found numerically steady solutions which are combinations of point vortices and vortex sheets.
\subsection{Main strategy}
The main strategy to prove Theorem \ref{rotatingsolution}, the Main Theorem of this paper, is to employ bifurcation theory and try to bifurcate from the simple eigenvalue $b:= \frac{1}{2\pi}\int_{-\pi}^{\pi} \gamma(x,t)dx =2$. However, the standard methods (Crandall-Rabinowitz \cite{Crandall-Rabinowitz:bifurcation-simple-eigenvalues}) fail since the linearized operator around the circle does not satisfy the transversality condition: in other words, the nontrivial zero set is not transversal to the trivial one (disks with constant vorticity amplitude). This phenomenon is usually known in the literature as a \textit{degenerate bifurcation} \cite{Kielhofer:bifurcation-book,Kielhofer:degenerate-bifurcation}. Graphically, this can be seen in Figure \ref{fig_bifurcation}. The problem is that we no longer have a single branch emanating from the disk, but two, and therefore the linearized operator fails to describe the local behaviour at the bifurcation point. To overcome this issue, we first reduce the nonlinear problem to a suitable finite dimensional space by means of a Lyapunov-Schmidt reduction since the restriction of $D\mathcal{F}$ is an isomorphism between Ker$(D\mathcal{F})^{\perp}$ and Im$(D\mathcal{F})$. After having done so, we are left with a finite dimensional system and it is there where we perform a higher order expansion around the bifurcation point, since, as expected by the failure of the transversality condition, the first order approximation is identically zero. We obtain that in suitable coordinates, the zero sets of $\mathcal{F}$ behave as $x^2 - y^2 = 0$ and thus two bifurcation branches emanate from the bifurcation point. The last part of the proof is devoted to handle the higher order terms, which can be controlled if we restrict the bifurcation domain to a suitable small enough neighbourhood. We mention here that this technique had been successfully employed by Hmidi--Mateu \cite{Hmidi-Mateu:degenerate-bifurcation-vstates-doubly-connected-euler} (in the hyperbolic case) and Hmidi--Renault \cite{Hmidi-Renault:existence-small-loops-doubly-connected-euler} (in the elliptic case).
\subsection{Organization of the paper}
The paper is organized as follows. In Section \ref{sec_eqns} we will write down the equations and describe the main spaces used in the proof. Section \ref{sec_proof} will be devoted to the bulk of the proof of Theorem \ref{rotatingsolution}, with the explicit calculations detailed in Appendix \ref{appendix}. Numerical calculations of the bifurcation branch and a brief discussion of the numerical methods employed are presented in Section \ref{sec_numerics}.
\section{The equations and the functional spaces}\label{sec_eqns}
Let $\omega(\cdot,t)=\omega_0(R_{\Omega t})$ be a stationary/rotating vortex sheet solution to the incompressible 2D Euler equation, where $ \omega_0\in \mathcal{M}(\mathbb{R}^2) \cap H^{-1}(\mathbb{R}^2)$. Here $\Omega=0$ corresponds to a stationary solution, and $\Omega\neq 0$ corresponds to a rotating solution. Assume that $\omega_0$ is concentrated on $\Gamma$. Throughout this paper we will assume that $\Gamma$ is a simple closed curve and $\Omega > 0$. Following \cite[Lemma 2.1]{GomezSerrano-Park-Shi-Yao:rotating-solutions-vortex-sheet-rigidity}, we have that:
\begin{lemma}\label{lemma_br_eq}
Assume $\omega(\cdot,t)=\omega_0(R_{\Omega t} x)$ is a stationary/uniformly-rotating vortex sheet with angular velocity $\Omega\in\mathbb{R}$, and $\omega_0$ is concentrated on $\Gamma$, with $z$ and $\gamma$ defined as above. Then the Birkhoff-Rott integral $BR$ \eqref{def_BR} and the strength $\gamma$ satisfy the following two equations:
\begin{equation}\label{BR1}
(BR-\Omega x^\perp)\cdot \mathbf{n} =\mathbf{v}^+ \cdot \mathbf{n} = \mathbf{v}^- \cdot \mathbf{n} = 0 \quad\text{ on } \Gamma,
\end{equation}
and
\begin{equation}\label{BR2}
(BR(z(\alpha))-\Omega z^\perp(\alpha)) \cdot \mathbf{s}(z(\alpha)) \,\frac{\gamma(\alpha)}{|z'(\alpha)|}= C.
\end{equation}
\end{lemma}
Note that \eqref{BR2} can be written as
\begin{align}
\left( I - P_0 \right) \left[ \left( BR(z,\Gamma)(z(\theta)) - \Omega z(\theta)^{\perp} \right) \cdot \frac{z'(\theta)\gamma(\theta)}{|z'(\theta)|^2} \right]= 0, \label{noevol_strength}
\end{align}
where $P_0$ is a projection to the mean, that is, $P_0f := \frac{1}{2\pi} \int_{-\pi}^{\pi} f(\theta)d\theta$. For simplicity, we also denote $\fint f(\theta)d\theta := \frac{1}{2\pi}\int_{-\pi}^{\pi} f(\theta)d\theta$. Now plugging $z(\theta) = (1+r(\theta))(\cos(\theta),\sin(\theta))$ and $\gamma(\theta) := b + g(\theta)$ into \eqref{BR1}, \eqref{BR2} and \eqref{noevol_strength} yields that
\begin{align}\label{mainfunctional}
\mathcal{F}(b,g,r) := (\mathcal{F}_1,\mathcal{F}_2) = (0,0),
\end{align}
where
\begin{align*}
&\mathcal{F}_1(b,g,r) := \fint_{-\pi}^{\pi}(b+g(\eta))\frac{\left(r'(\theta)\cos(\theta-\eta)-(1+r(\theta))\sin(\theta-\eta)\right)(1+r(\eta))-(1+r(\theta))r'(\theta)}{(1+r(\theta))^2+(1+r(\eta))^2-2(1+r(\theta))(1+r(\eta))\cos(\theta-\eta)}d\eta\\
& \qquad\qquad\qquad + \Omega r'(\theta)(1+r(\theta)), \\
&\mathcal{F}_2 (b,g,r) := \left(I-P_0 \right) \tilde{\mathcal{F}}_2(b,g,r), \\
&\tilde{\mathcal{F}}_2(b,g,r) := \fint_{-\pi}^{\pi}(b+g(\eta))\frac{(1+r(\theta))^2-(r'(\theta)\sin(\theta-\eta)+(1+r(\theta))\cos(\theta-\eta))(1+r(\eta))}{(1+r(\theta))^2+(1+r(\eta))^2-2(1+r(\theta))(1+r(\eta))\cos(\theta-\eta)}d\eta \\
&\qquad\qquad\quad\quad \times \frac{(b+g(\theta))}{r'(\theta)^2+(1+r(\theta))^2} -\Omega (1+r(\theta))^2 \frac{b+g(\theta)}{r'(\theta)^2+(1+r(\theta))^2}.
\end{align*}
Throughout the paper we will work with the following analytic function spaces. Let $c>0$ be a sufficiently small parameter and let $\mathcal{C}_{w}(c)$ be the space of analytic functions in the strip $|\Im(z)| \leq c$. For $k \in \mathbb{N}$, denote
\begin{align*}
X^{k}_{c} := \left\{ f(\theta) \in \mathcal{C}_{w}(c), \quad f(\theta) = \sum_{n=1}^{\infty}a_n\cos(2n\theta), \quad \sum_{\pm} \int_{-\pi}^{\pi} |f(\theta \pm ic)|^2 + |\partial^k f(\theta \pm ic)|^2 d\theta< \infty \right\} \\
\quad Y^{k}_{c} := \left\{f(\theta) \in \mathcal{C}_{w}(c), \quad f(\theta) = \sum_{n=1}^{\infty}a_n\sin(2n\theta),\quad \sum_{\pm} \int_{-\pi}^{\pi} |f(\theta \pm ic)|^2 + |\partial^k f(\theta \pm ic)|^2 d\theta< \infty \right\},
\end{align*}
From now on, due to scaling considerations, we will fix $\Omega = 1$ and $b$ will play the role as bifurcation parameter. It is clear that $\mathcal{F}(b,0,0) = (0,0)$ for all $b\in \R$ since $\mathcal{F}_1(b,0,0) = 0$ and $\tilde{\mathcal{F}}_2(b,0,0)$ is constant. Our main theorem in this paper is the following:
\begin{theorem}\label{rotatingsolution}
Let $k \geq 3$, and let $c>0$ be sufficiently small. Then, there exists a curve of solutions $(b,g,r)$ of $\mathcal{F} = (0,0)$, belonging to $\R\times X^{k}_{c}\times X^{k+1}_{c}$ and a neighbourhood of $(b,g,r) = (2,0,0)$, bifurcating from $(b,g,r) = (2,0,0)$ such that $(g,r) \neq (0,0)$.
\end{theorem}
\section{Proof of the Main Theorem}\label{sec_proof}
The goal of this section is to prove the existence of non-radial uniformly-rotating vortex sheets. To do so, we will split the proof into the following steps: first we will prove that the functional $\mathcal{F}$ is $C^3$, next we will study $D\mathcal{F}$ to show that, as mentioned in the introduction, it is a Fredholm operator of index 0, with dim(Ker$(D\mathcal{F})) = 1$. The next step is to apply Lyapunov-Schmidt theory and reduce the problem to a finite (2) dimensional one. In those coordinates, linear expansions fail to be conclusive (all the linear terms vanish) since 2 nontrivial branches emanate from the bifurcation point (as opposed to 1). Instead, we perform a quadratic expansion to determine that locally the bifurcation branches look like two pairs of straight lines (specifically as $x^2-y^2 = 0$ in some well-chosen coordinates) and hence the bifurcation does not trivialize (as if it had been of the type $x^2 + y^2 = 0$). We conclude the proof by handling the higher order terms and showing that they don't alter the quadratic behaviour in a sufficiently small neighbourhood of the bifurcation point.
\subsection{Continuity of the functional}
In this subsection, we will check the regularity of $\mathcal{F}$. As explained above, we will reduce the infinite dimensional problem to a finite dimensional problem and investigate its Taylor expansion up to quadratic order. Hence, we need to check if the functional is regular enough to do so. To this end, we have the following proposition:
\begin{proposition}\label{regularity1}
Let $k\ge 3$. Then there exists a neighborhood $U$ of $(2,0,0) \in \R\times X^{k}_{c}\times X^{k+1}_{c} $ such that $\mathcal{F} \in C^{3}\left( U ; Y^{k}_{c}\times X^{k}_{c} \right)$.
\end{proposition}
\begin{proof}
Since the stream function, $\omega * \mathcal{N}$, is invariant under rotations, it follows immediately that $\mathcal{F}$ is also invariant under rotation by $\pi$-radians, hence $\mathcal{F}$ has only even Fourier modes. Also the oddness of $\mathcal{F}_1$ and evenness of $\mathcal{F}_2$ follow from the invariance under reflection.
To prove the regularity, we briefly sketch the idea. We impose $k\ge 3$ to ensure that $H^{k}$ is a Banach algebra. It is clear that $\mathcal{F}$ is smooth in $b$. It is also straightforward that, for example, for all $(g,r)$ near $(0,0)\in X^{k}_{c}\times X^{k+1}_{c}$,
\begin{align*}
\hspace*{-0.5cm} \frac{d}{dt}\mathcal{F}_1&(b,g+th_1,r+th_2)\bigg|_{t=0} = PV \fint h_1(\eta)\frac{\left(r'(\theta)\cos(\theta-\eta)-(1+r(\theta))\sin(\theta-\eta)\right)(1+r(\eta))-(1+r(\theta))r'(\theta)}{(1+r(\theta))^2+(1+r(\eta))^2-2(1+r(\theta))(1+r(\eta))\cos(\theta-\eta)} \\
&\quad + (b+g(\eta))\left[ \frac{(h_2'(\theta)\cos(\theta-\eta)-h_2(\theta)\sin(\theta-\eta))(1+r(\eta))-h_2(\theta)r'(\theta)}{(1+r(\theta))^2+(1+r(\eta))^2-2(1+r(\theta))(1+r(\eta))\cos(\theta-\eta)} \right.\\
&\quad \left. + \frac{\left(r'(\theta)\cos(\theta-\eta)-(1+r(\theta))\sin(\theta-\eta)\right)h_2(\eta)-(1+r(\theta))h_2'(\theta)}{(1+r(\theta))^2+(1+r(\eta))^2-2(1+r(\theta))(1+r(\eta))\cos(\theta-\eta)} \right. \\
&\quad \left. -\left[\left( r'(\theta)\cos(\theta-\eta)-(1+r(\theta))\sin(\theta-\eta)\right)(1+r(\eta))-(1+r(\theta))r'(\theta) \right] \right. \\
&\quad \left. \times \frac{\left[ 2(1+r(\theta)h_2(\theta)+2(1+r(\eta)h_2(\eta)-2\cos(\theta-\eta)(h_2(\theta)(1+r(\eta))+h_2(\eta)(1+r(\theta)))\right]}{((1+r(\theta))^2+(1+r(\eta))^2-2(1+r(\theta))(1+r(\eta))\cos(\theta-\eta))^2} \right] d\eta + \Omega h_2'(\theta)\\
& =: D\mathcal{F}_1(b,g,r)[h_1,h_2],
\end{align*}
and $D\mathcal{F}_1:\R\times X^{k}_{c}\times X^{k+1}_{c} \mapsto \mathcal{L}(X^{k}_{c}\times X^{k+1}_{c} ; Y^{k}_{c}\times X^{k}_{c})$ is continuous. A similar derivation can be performed for $D\mathcal{F}_2$. For the higher derivatives, we refer to \cite{Castro-Cordoba-GomezSerrano:existence-regularity-vstates-gsqg,Castro-Cordoba-GomezSerrano:analytic-vstates-ellipses,GomezSerrano:stationary-patches,Hmidi-Mateu-Verdera:rotating-vortex-patch,Renault:relative-equlibria-holes-sqg} for the method to deal with the singular integrals arising throughout the calculations.
\end{proof}
\subsection{Fredholm index of the linearized operator $D\mathcal{F}$}
This subsection is devoted to show that $D\mathcal{F}$ is Fredholm of index zero. We can make all the calculations explicit, moreover the operator diagonalizes in Fourier modes. We have the following lemmas:
\begin{lemma}\label{linearized}
Let $g(\theta)=\sum_{n=1}^{\infty}a_n \cos(2n\theta)$ and $r(\theta) = \sum_{n=1}^{\infty}b_n\cos(2n\theta)$. Then we have that
\[
D\mathcal{F}(b,0,0)\left[ g, r\right] = \left(\begin{array}{c}\hat{g}(\theta) \\ \hat{r}(\theta) \end{array}\right),
\]
where
\[ \hat{g}(\theta) = \sum_{n=1}^{\infty} \hat{a}_n \sin(2n\theta), \quad \hat{r}(\theta) = \sum_{n=1}^{\infty} \hat{b}_n \cos(2n\theta),
\]
and the coefficients satisfy, for any $n \ge 1$:
\[
M_n
\begin{pmatrix}
a_n \\
b_n
\end{pmatrix}
:=
\begin{pmatrix}
-\frac{1}{2} & -2n\left( \Omega - \frac{b}{2}\right) \\
\frac{b}{2}-\Omega & b^2(n-1)
\end{pmatrix}
\begin{pmatrix}
a_n \\
b_n
\end{pmatrix}
=
\begin{pmatrix}
\hat{a}_n \\
\hat{b}_n
\end{pmatrix}.
\]
\end{lemma}
\begin{proof}
We use \eqref{linearf1} in Lemma~\ref{derivative} and obtain
\begin{align*}
\hat{g}(\theta) = \frac{d}{dt}\mathcal{F}_1(b,tg,tr) &= - \fint \frac{g(\eta)\sin(\theta-\eta)}{2-2\cos(\theta-\eta)}d\eta + \left( \Omega- \frac{b}{2}\right) r'(\theta)\\
& = -\sum_{n=1}^{\infty}a_n\fint \frac{\cos(2n\eta)\sin(\theta-\eta)}{2-2\cos(\theta-\eta)}d\eta + \sum_{n=1}^{\infty}(-2n)\left( \Omega - \frac{b}{2} \right)b_n \sin(2n\theta) \\
& = \sum_{n=1}^{\infty}\left ( -\frac{a_n}{2} + (-2n)\left( \Omega - \frac{b}{2}\right) b_n \right)\sin(2n\theta),
\end{align*}
where the last equality follows from ~\eqref{lemma2}. Similarly, we apply \eqref{linearf2} in Lemma~\ref{derivative} and \eqref{lemma1} to obtain
\begin{align*}
\hat{r}(\theta) & = \left( \frac{b}{2} - \Omega \right) g(\theta) + b^2\left( \fint \frac{r(\theta)-r(\eta)}{2-2\cos(\theta-\eta)}d\eta - r(\theta)\right)\\
& = \sum_{n=1}^{\infty} \left(\frac{b}{2} - \Omega \right) a_n\cos(2n\theta) + b^2 \sum_{n=1}^{\infty} b_n \left( \fint \frac{\cos(2n\theta)-\cos(2n\eta)}{2-2\cos(\theta-\eta)}d\eta - \cos(2n\theta) \right)\\
& = \sum_{n=1}^{\infty} \left( \left(\frac{b}{2} - \Omega \right) a_n+ b^2\left( n - 1\right)b_n \right)\cos(2n\theta).
\end{align*}
This proves the lemma.
\end{proof}
\begin{lemma}\label{onedimensionality}
Let us fix $b=2$ and $\Omega = 1$. We also denote $v:=(0,\cos(2\theta)) \in X^k_{c}\times X^{k+1}_{c}$ and $w := (0,\cos(2\theta)) \in Y^k_{c} \times X^k_{c}$. Then it holds that
\begin{align*}
&\text{Ker}\left(D\mathcal{F}(2,0,0)\right) = \text{span}\left\{ v \right\} \subset{X^k_{c}\times X^{k+1}_{c}},\\
& \text{Im}\left(D\mathcal{F}(2,0,0)\right)^{\perp} = \text{span}\left\{ w \right\} \subset Y^{k}_{c}\times X^{k}_{c}.
\end{align*}
\end{lemma}
\begin{proof}
From Lemma~\ref{linearized}, we have
\begin{align*}
M_n =
\begin{pmatrix}
-\frac{1}{2} & 0 \\
0 & 4(n-1)
\end{pmatrix},
\end{align*}
for all $n \ge 1$.
For all $n\ge 2$, $M_n$ is clearly an isomorphism, while $\text{Ker}(M_1) = \text{Im}(M_1)^{\perp} = \left(\begin{array}{c} 0 \\ 1 \end{array}\right)$. By orthogonality of Fourier modes, this proves the lemma.
\end{proof}
\subsection{Lyapunov-Schmidt reduction}\label{lsreduction}
In this subsection, we will aim to derive a finite dimensional system which is equivalent to \eqref{mainfunctional}. From Lemma~\ref{onedimensionality}, we have the following orthogonal decompositions of the function spaces:
\begin{align*}
X := X^k_{c}\times X^{k+1}_{c} = \text{span}\left\{ v \right\} \oplus \text{Ker}\left(D\mathcal{F}(2,0,0)\right)^{\perp} =: \text{span}\left\{ v \right\} \oplus \mathcal{X}, \quad v\in \text{Ker}\left(D\mathcal{F}(2,0,0)\right),\\
Y := Y^{k}_{c} \times X^{k}_{c}= \text{span}\left\{ w \right\} \oplus \text{Im}\left(D\mathcal{F}(2,0,0)\right) =: \text{span}\left\{ w \right\} \oplus \mathcal{Y}, \quad w\in \text{Im}\left(D\mathcal{F}(2,0,0)\right)^{\perp},
\end{align*}
where $v$ and $w$ are as defined in Lemma~\ref{onedimensionality}. Let us consider the orthogonal projections
\[
P : X \rightarrow \text{span}\left\{ v \right\},\quad
Q: Y \rightarrow \text{span}\left\{ w \right\}.
\]
More precisely, we have
\begin{align}
& P(g(\theta),r(\theta)) = \left( 0 , \left(\frac{1}{\pi}\int r(\eta)\cos(2\eta)d\eta\right) \cos(2\theta) \right) \quad \text{ for all }(g,r)\in X^{k}_{c}\times X^{k+1}_{c}, \label{defofpq}\\
&Q(G(\theta),R(\theta)) = \left( 0 , \left(\frac{1}{\pi}\int R(\eta)\cos(2\eta)d\eta\right) \cos(2\theta) \right) \quad \text{ for all } (G,R) \in Y^{k}_{c}\times X^{k}_{c}. \label{defofpq1}
\end{align}
We remark that we will sometimes abuse notation and identify $\mathcal{F}(b,g,r)$ with $\mathcal{F}(b,(g,r))$, where $(g,r) \in X$. Let us define $G:\R \times \text{span}\left\{ v \right\} \times \mathcal{X} \mapsto {Y}$ as follows:
\begin{align*}
G(b,f,x):=\mathcal{F}(b,f+x), \quad\text{ for }\quad b\in \R,\quad f\in \text{span}\left\{ v \right\},\quad x\in \mathcal{X}.
\end{align*}
Then \eqref{mainfunctional} is equivalent to (for $(g,r) = f+x$)
\begin{align}\label{Reduction1}
QG(b,f,x)=0 \quad \text{ and }\quad \left( I-Q\right)G(b,f,x)=0.
\end{align}
However, it follows from Lemma~\ref{onedimensionality} that
\begin{align}\label{isomorphism}
D_x \left((I-Q)G\right)(b,0,0)=\left(I-Q\right)D\mathcal{F}(b,0)P : \mathcal{X}\mapsto \mathcal{Y}
\end{align}
is an isomorphism, consequently, the implicit function theorem yields that there is an open set $U\subset \R\times \text{span}\left\{ v \right\}$ near $(b,0)$ and a function $\varphi:U\mapsto \mathcal{X}$ such that
\begin{align*}
\left( I-Q\right)G(b,f,\varphi(b,f)) = \left(I-Q\right)\mathcal{F}(b,f+\varphi(b,f))=0.
\end{align*}
Note that from $\mathcal{F}(b,0)=(0,0)$ for any $b\in \R$, we have
\begin{align}
\label{trivial_solution}
\varphi(b,0)=0,
\end{align}
and thus \eqref{Reduction1} is equivalent to
\begin{align}\label{Reduction2}
0 = QG(b,f,\varphi(b,f)) = Q\mathcal{F}(b,f+\varphi(b,f)), \quad (b,f)\in U.
\end{align}
Since $\text{span}\left\{ v \right\}$ is one dimensional, we have $f=tv$ for some $t\in \R$, therefore the system \eqref{Reduction2} can be written in terms of the variables $b$ and $t$ as
\begin{align*}
0 = Q\mathcal{F}(b,tv+\varphi(b,tv)) &= \int_{0}^{1}\frac{d}{d s}\left( Q\mathcal{F}(b,stv+\varphi(b,stv)) \right)ds\\
&= \int_0^1 QD\mathcal{F}(b,stv+\varphi(b,stv))(tv+t\partial_f\varphi(b,stv)v) ds,
\end{align*}
where we used \eqref{trivial_solution} to obtain the second equality.
Dividing the right-hand side by $t$ to get rid of the trivial solutions, we are led to solve the following two dimensional problem:
\begin{align}\label{Reduction4}
0 = F_{red}(b,t):=\int_0^1 Q D\mathcal{F}(b,stv+\varphi(b,stv))(v+\partial_f\varphi(b,stv)v) ds, \quad (b,tv)\in U.
\end{align}
\subsection{Quadratic expansion of the reduced functional}
The main idea is to expand the reduced functional $F_{red}$ up to quadratic terms. To this end, we recall the following proposition for the derivatives of $F_{red}$.
\begin{proposition}(\cite[Proposition 3]{Hmidi-Mateu:degenerate-bifurcation-vstates-doubly-connected-euler}, \cite[Proposition 3.1]{Hmidi-Renault:existence-small-loops-doubly-connected-euler})
Let $F_{red}$ be defined as in \eqref{Reduction4}. Then the following hold:
\begin{itemize}
\item[(a)] First derivatives:
\begin{align*}
&\partial_b F_{red}(2,0) = Q\partial_b D\mathcal{F}(2,0)v, \\
&\partial_t F_{red}(2,0) = \frac{1}{2}\frac{d^2}{dt^2}Q\mathcal{F}(2,tv)\bigg|_{t=0}.
\end{align*}
\item[(b)] Second derivatives:
\begin{align*}
&\partial_{bb}F_{red}(2,0) = 2Q\partial_bD\mathcal{F}(2,0)\tilde{v},\\
&\partial_{tt} F_{red}(2,0) = \frac{1}{3}\frac{d^3}{dt^3}\left[ Q\mathcal{F}(2,tv)\right] \bigg|_{t=0} + Q\frac{d^2}{dtds}\mathcal{F}(2,tv+s\hat{v})\bigg|_{t=s=0},\\
&\partial_{tb}F_{red}(2,0) = \frac{1}{2}\partial_b Q\frac{d^2}{dt^2} \mathcal{F}(b,tv) \bigg|_{b=2,t=0} + \frac{1}{2}Q\partial_bD\mathcal{F}(2,0)\hat{v} + Q\frac{d^2}{dtds}\mathcal{F}(2,tv+s\tilde{v})\bigg|_{t=s=0},
\end{align*}
where
\begin{align*}
&\hat{v} := -\left[ D\mathcal{F}(2,0) \right]^{-1} \frac{d^2}{dt^2}\left[ (I - Q) \mathcal{F}(2,tv) \right] \bigg|_{t=0}, \\
&\tilde{v} := - \left[ D \mathcal{F}(2,0) \right]^{-1} (I - Q)\partial_b D \mathcal{F}(2,0)v.
\end{align*}
\end{itemize}
\end{proposition}
Now using the values found in Lemma~\ref{somevalues}, we can obtain the derivatives of $F_{red}$.
\begin{proposition}\label{explicitvalues}
Let $F_{red}$ be defined as in \eqref{Reduction4}. Then it holds that
\begin{align}
&\partial_b F_{red}(2,0) = 0, \label{derivative11}\\
&\partial_t F_{red}(2,0) = 0. \label{derivative12}\\
&\partial_{bb}F_{red}(2,0) = 2w,\label{derivative21}\\
&\partial_{tt} F_{red}(2,0) = -8w, \label{derivative22}\\
&\partial_{tb}F_{red}(2,0) = 0. \label{derivative23}
\end{align}
\end{proposition}
\begin{proof}
\eqref{derivative11} follows immediately from \eqref{defofpq1} and \eqref{value1}. For \eqref{derivative12}, we use \eqref{value2} and the orthogonality of the Fourier modes. \eqref{derivative21} follows from \eqref{value11}. \eqref{derivative22} follows from \eqref{value5} and \eqref{value6}. Lastly, \eqref{derivative23} follows from \eqref{value10}, \eqref{value8} and \eqref{value9}.
\end{proof}
\subsection{Proof of Theorem~\ref{rotatingsolution}}
Now we are ready to prove the main theorem of this section.
\begin{proof}
From \eqref{Reduction4}, it suffices to show that there exist $(b,t)$ such that $t\ne 0$ and $F_{red}(b,t) = 0$. To do so, we expand $F_{red}$ up to quadratic terms and obtain that for all $(b,t)$ near $(2,0)$,
\begin{align*}
F_{red}(b,t) &= \left[F_{red}(2,0) + \partial_{b}F_{red}(2,0)(b-2) + \partial_{t}F_{red}(2,0)t \right. \\
&\left. \ + \frac{1}{2}\partial_{bb}F_{red}(2,0)(b-2)^2 + \frac{1}{2}\partial_{tt}F_{red}(2,0)t^2 + \partial_{tb}F_{red}(2,0)(b-2)t + \left((b-2)^2 + t^2 \right)\epsilon(b,t) \right]w,
\end{align*}
where $\epsilon(b,t)$ is a continuous function such that $\lim_{(b,t)\to (2,0)}\epsilon(b,t) =\epsilon(2,0)= 0$. From Proposition~\ref{explicitvalues}, it follows that (we drop $w$ for simplicity)
\begin{align*}
F_{red}(b,t) = (b-2)^2 - 4t^2 + \left( (b-2)^2 + t^2 \right)\epsilon(b,t).
\end{align*}
Now we use the change of variables $b:= x + 2$ and $t=xy$, so that
\begin{align}\label{changeofvariables}
\hat{F}(x,y) := \frac{F_{red}(x+2,xy)}{x^2} = \left(1 - 4y^2\right) + (1+y^2)\epsilon(x+2,xy).
\end{align}
Clearly, $\hat{F}\left( 0,\frac{1}{2} \right) = 0$ and $ \partial_{y}\hat{F}\left( 0,\frac{1}{2} \right) = -4\ne 0$.
Therefore the implicit function theorem implies that there exists a continuous function $\phi$ near $0$ such that $\hat{F}(x,\phi(x)) = 0$ and $\phi(0) = \frac{1}{2}$. Therefore it follows from \eqref{changeofvariables} that there exists a pair $(b,t)$ such that $t\ne 0$ and $F_{red}(b,t) = 0$. This finishes the proof.
\end{proof}
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.2]{bifurcation_final_without_labels.eps}
\caption{\label{fig_bifurcation} Plot of the bifurcation diagram of the solutions given by Theorem \ref{rotatingsolution}. The dotted red lines correspond to the linear expansion \eqref{changeofvariables} around the bifurcation point $(2,0)$. See Figure \ref{fig_panels} for a numerical plot of the solutions at the points $A$, $B$, $C$, $D$. The branches continue beyond what is calculated.}
\end{center}
\end{figure}
\section{Numerical results}\label{sec_numerics}
In this section, we describe how to compute numerically the branches of solutions emanating from the disk, previously proved (locally) in Theorem \ref{rotatingsolution}. See Figure \ref{fig_bifurcation}. To do so, we calculate solutions of the form
\[R(\theta) = 1 + \sum_{k=1}^{N} r_k \cos(2k\theta), \quad \gamma(\theta) = \sum_{k=0}^{N} \gamma_k \cos(2k\theta)
\]
with $\gamma_0 = b$. We first employed continuation in $b$, in increments of $\Delta b = 0.001$, starting from $b=1.8$ and $b=2.1$ and using as initial guess for the starting $b$ the solution given by the linear theory and for the subsequent $b$ the solution found in the previous iteration. After discovering a fold at approximately $b \sim 1.68$, we switched variables and instead we recalculated using continuation in $r_1$, which appears to be monotonic along the branches. As before, we start at $r_1 = \pm 0.125$ and take an increment $\Delta r_1 = 0.001$.
To compute a solution for a fixed $r_1$ we use the Levenberg-Marquardt algorithm. We aim to find a zero of the system of equations $\mathcal{F}(b,g,r)(\theta_j)$, with $\theta_j = \frac{j\pi}{N_\theta}, \quad j=1,\ldots,N_\theta$ and $N_\theta = 1024$ with variables $r_k, \, k \neq 1$ (recall that $r_1$ is fixed at each iteration since it is the continuation parameter) and $\gamma_k$. We take $N = 160$. In order to perform the integration in space, we desingularize the principal value at $\eta = \theta$ by subtracting $\frac12\mathcal{H}(\gamma)$ to $\mathcal{F}_1$, where $\mathcal{H}$ denotes the Hilbert transform, computed explicitly since we have the Fourier expansion of $\gamma$, and perform a trapezoidal integration on the rest (for which the integrand is smooth), with step $h = \frac{2\pi}{N_\theta}$. We remark that the integrand of $\mathcal{F}_2$ has a removable singularity (thus no principal value integration is needed) and can be integrated using the trapezoidal integration if the limit at $\eta = \theta$ is taken properly.
\begin{figure}[h!]
\hspace*{-1.2cm}\begin{tabular}{cc}
\includegraphics[width=0.25\textwidth]{Gfold.eps}
\includegraphics[width=0.25\textwidth]{Rfold.eps}&
\includegraphics[width=0.25\textwidth]{GA.eps}
\includegraphics[width=0.25\textwidth]{RA.eps} \\
\qquad \qquad (a) $\gamma(\theta)$ and $z(\theta)$ for $r_1 = 0.362, b \sim 1.6799$ & \qquad (b) $\gamma(\theta)$ and $z(\theta)$ for $r_1 = 0.825, b \sim 1.7779$ \hspace{2.5cm} \\
\includegraphics[width=0.25\textwidth]{GC.eps}
\includegraphics[width=0.25\textwidth]{RC.eps}&
\includegraphics[width=0.25\textwidth]{GD.eps}
\includegraphics[width=0.25\textwidth]{RD.eps}\\
\qquad (c) $\gamma(\theta)$ and $z(\theta)$ for $r_1 = 0.525, b \sim 4.0954$ & \qquad (d) $\gamma(\theta)$ and $z(\theta)$ for $r_1 = 0.925, b \sim 9.3439$ \hspace{2.5cm}
\end{tabular}
\caption{Panels (a)-(d): $\gamma(\theta)$ and $z(\theta)$ at the points $A$--$D$ highlighted in Figure \ref{fig_bifurcation}. In panel (b), $\gamma$ appears to tend to be concentrated only on the horizontal parts of $z$, leading to a possible solution consisting only of two symmetric curves (cf. \cite[Figure 1]{ONeil:relative-equilibria-vortex-sheets}) and a change of topology.
} \label{fig_panels}
\end{figure}
| 3aa32cd20210a63c173901f3c3fa6f39b1079aca | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Detailed Results of Greedy Attack} \label{appendix: greedy}
\begin{figure*}
\centering
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/warning_lights_MIO-10_greedy.png}
\caption{The warning lights.}
\label{fig:MIO-10_warning_greedy}
\end{subfigure}%
\hfill
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/distance_measurement_MIO-10_greedy.png}
\caption{The manipulation on distance.}
\label{fig:MIO-10_attack_distance_greedy}
\end{subfigure}%
\hfill
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/velocity_measurement_MIO-10_greedy.png}
\caption{The manipulation on velocity.}
\label{fig:MIO-10_attack_velocity_greedy}
\end{subfigure}%
\hfill
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/phase_plot_MIO-10_greedy.png}
\caption{The state trajectory.}
\label{fig:MIO-10_state_traj_greedy}
\end{subfigure}%
\caption{Greedy attack on the MIO-10 dataset.}
\label{fig:attack_MIO-10_greedy}
\end{figure*}
\begin{figure*}[t]
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/warning_lights_MIO+1_greedy.png}
\caption{The warning lights.}
\label{fig:MIO+1_warning_greedy}
\end{subfigure}%
\hfill
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/distance_measurement_MIO+1_greedy.png}
\caption{The manipulation on distance.}
\label{fig:MIO+1_attack_distance_greedy}
\end{subfigure}%
\hfill
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/velocity_measurement_MIO+1_greedy.png}
\caption{The manipulation on velocity.}
\label{fig:MIO+1_attack_velocity_greedy}
\end{subfigure}%
\hfill
\begin{subfigure}{.24\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/phase_plot_MIO+1_greedy.png}
\caption{The state trajectory.}
\label{fig:MIO+1_state_traj_greedy}
\end{subfigure}%
\caption{Greedy attack on the MIO+1 dataset.}
\label{fig:attack_MIO+1_greedy}
\end{figure*}
In this section, we provide more detailed results of the greedy attack, including warning lights before and after attack, manipulations on measurements, and the trajectory of KF state predictions. We notice that the results are very similar for different lengths of the stealthy interval $\mathcal{T}^s$. Therefore, here we only show the results for $\mathcal{T}^s=2.5$ seconds (i.e., half of the full length) as an example.
Fig.~\ref{fig:attack_MIO-10_greedy} shows the greedy attack on MIO-10, where the stealthy interval $\mathcal{T}^s=[50, 97]$. By manipulating the distance and velocity to the maximum possible value, the attacker successfully causes the FCW to output green lights in the target interval $\mathcal{T}^\dagger$. However, the attack induces side effect in $\mathcal{T}^s$, where the original yellow lights are changed to green. In contrast, our MPC-based attack does not have any side effect during $\mathcal{T}^s$. Also note that the trajectory of the KF state prediction enters ``into" the desired green region during $\mathcal{T}^\dagger$. This is more than necessary and requires larger total manipulation ($J_1$) than forcing states just on the boundary of the desired region, as does our attack.
In Fig.~\ref{fig:attack_MIO+1_greedy}, we show the greedy attack on MIO+1. The stealthy interval is $\mathcal{T}^s=[51, 99]$. Again, the attack results in side effect during the stealthy interval $\mathcal{T}^s$. Furthermore, the side effect is much more severe (green to red) than that of our MPC-based attack (green to yellow). The KF state trajectory enters ``into" the desired red region, and requires larger total manipulation ($J_1$) than our attack.
\section{Velocity Increase in Figure~\ref{fig:MIO+1_attack_velocity}} \label{appendix:increase}
In Figure~\ref{fig:attack_MIO+1_appendix}, we show again the manipulation on the velocity measurement for the MIO+1 dataset. The attacker's goal is to cause the FCW to output red warnings in the target interval $[100,139]$. Intuitively, the attacker should decrease the distance and velocity. However, in Figure~\ref{fig:attack_MIO+1_appendix}, the attacker instead chooses to increase the velocity during interval [88,96]. We note that this is because the attacker hopes to force a very negative KF acceleration estimation. To accomplish that, the attacker first strategically increases the velocity from step 88 to 96, and then starting from step 97, the attacker suddenly decreases the velocity dramatically. This misleads the KF to believe that the MIO has a very negative acceleration. In Fig~\ref{fig:MIO+1_acc}, we show the acceleration estimation produced by KF. At step 96, the estimated acceleration is 8.1m/s$^2$. However, at step 97, the estimated acceleration suddenly drops to $-$16m/s$^2$, and then stays near $-$30m/s$^2$ until the target interval. The very negative acceleration in turn causes the KF velocity estimation to decrease quickly. The resulting velocity estimation reached around $-$10m/s during the target interval, which causes FCW to output red lights.
\begin{figure}[t]
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.6\textwidth]{figures/acceleration_estimation_MIO+1.png}
\caption{Acceleration estimation.}
\label{fig:MIO+1_acc}
\end{subfigure}%
\hfill
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.6\textwidth]{figures/velocity_measurement_MIO+1.png}
\caption{Manipulation on velocity.}
\label{fig:attack_MIO+1_appendix}
\end{subfigure}%
\caption{Acceleration reduces significantly as the velocity measurement drops after step 96. This in turn causes the KF velocity estimation to decrease fast.}
\end{figure}
\section{Human Behavior Algorithm} \label{appendix: human}
In this section, we provide an algorithmic description of the human behavior model.
\begin{algorithm}
\SetAlgoLined
\SetKwInOut{Input}{Input}
\Input{light sequence $\ell_t(1\le t\le T)$, reaction time $h^*$.}
Initialize $s=0$\;
\For{$t\leftarrow 1$ \KwTo $T$}{
\uIf{$t!=1$ and $\ell_{t}!=\ell_{t-1}$}
{
$s=0$\;
}
\uElseIf{$\ell_t=\text{red}$}
{
$s=s+1$\;
}
\Else
{
$s=s-1$\;
}
\uIf{$s\ge h^*$}
{
human applies pressure on pedal\;
}
\uElseIf{$s\le -h^*$}
{
human releases brake\;
}
\Else{human stays in the previous state\;}
}
\caption{Human Behavior Algorithm.}
\end{algorithm}
\section{Simulated Raw Data Processing} \label{appendix:carla-output-processing}
CARLA outputs a single RGB image, a depth map image, and variable number of RADAR points for each 0.05 second time step of the simulation. We analyze this data at each time step to produce object detections in the same format of MATLAB FCW \cite{matlab-fcw}:
\[
\begin{bmatrix}
d^1 & v^1 & d^2 & v^2
\end{bmatrix}
\]
where $d^1$ and $d^2$ are the distance, in meters, from the vehicle sensor in directions parallel and perpendicular to the vehicle's motion, respectively. $v^1$ and $v^2$ are the detected object velocities, in m/s, relative to the ego along these parallel and perpendicular axes.
To produce these detections from vision data, we first find bounding boxes around probable vehicles in each RGB image frame using an implementation of a YOLOv2 network in MATLAB which has been pre-trained on vehicle images \cite{matlab-yolo}. Each bounding box is used to create a distinct object detection. The $d^1$ value, or depth, of each object is taken to be the depth recorded by the depth map at the center pixel of each bounding box.
The $d^2$ value of each detection is then computed as
\begin{equation}
d^2 = u * \frac{d^1}{l_{foc}}
\end{equation}
where $u$ is the horizontal pixel coordinate of the center of a bounding box in a frame, and $l_{foc}$ is the focal length of the RGB camera in pixels \cite{penn-cam-projection}. $l_{foc}$ is not directly specified by CARLA, but can be computed using the image length, 800 pixels, and the camera field of vision, 90 degrees \cite{edmund-focal-length}.
To compute $v^1$ and $v^2$ for detections of the current time step, we also consider detections from the previous time step. First, we attempt to match each bounding box from the current time step to a single bounding box from the previous step. Box pairs are evaluated based on their Intersection-Over-Union (IoU) \cite{simple-real-time-tracking}. Valued between 0 and 1, a high IoU indicates high similarity of size and position of two boxes, and we enforce a minimum threshold of 0.4 for any two boxes to be paired. For two adjacent time steps, A and B, we take the IoU of all possible pairs of bounding boxes with one box from step A, and one from B. These IoU values form the cost matrix for the Hungarian matching algorithm \cite{real-time-mot}, which produces the best possible pairings of bounding boxes from the current time step to the previous.
This matching process results in a set of detections with paired bounding boxes, and a set with unpaired boxes. For each detection with a paired box, we calculate its velocity simply as the difference between respective $d^1$ and $d^2$ values of the current detection and its paired observation from the previous time step, multiplied by the frame rate, $fps_{cam}$. For a detection, $a$, paired with a previous detection, $b$:
\begin{equation}
<v^1_a, v^2_a> = <d^1_b - d^1_a, d^2_b - d^2_a> * fps_{cam}
\end{equation}
For each detection left unpaired after Hungarian matching, we make no conclusions about $v^1$ or $v^2$ for that detection, and treat each as zero.
Each RADAR measurement output by CARLA represents an additional object detection. RADAR measurements contain altitude ($al$) and azimuth ($az$) angle measurements, as well as depth ($d$) and velocity ($v$), all relative to the RADAR sensor. We convert these measurements into object detection parameters as follows
\begin{align*}
d^1 &= d * \cos az * \cos al & v^1 &= v * \cos az * \cos al \\
d^2 &= d * \sin az * \cos al & v^2 &= v * \sin az * \cos al\\
\end{align*}
\section{Preprocessing of CARLA Measurements} \label{appendix: preprocess}
In this section, we describe how we preprocess the measurements obtained from CARLA simulation. The measurement in each time step takes the form of $t_t=[y_t^1, y_t^2]\in \{\mathbb{R}\cup \text{NaN}\}^8$, where $y_t^1\in \{\mathbb{R}\cup \text{NaN}\}^4$ is the vision detection produced by ML-based objection detection algorithm YOLOv2, and $y_t^2$ is the detection generated by radar (details in Appendix \ref{appendix:carla-output-processing}). Both vision and radar measurements contain four components: (1) the distance to MIO along driving direction, (2) the velocity of MIO along driving direction, (3) the distance to MIO along lateral direction, and (4) the velocity of MIO along lateral direction. The radar measurements are relatively accurate, and do not have missing data or outliers. However, there are missing data (NaN) and outliers in vision measurements. The missing data problem arises because the MIO sometimes cannot be detected, e.g., in the beginning of the video sequence when the MIO is out of the detection range of the camera. Outliers occur because YOLOv2 may not generate an accurate bounding box of the MIO, causing it to correspond to a depth map reading of an object at a different physical location. As such, a small inaccuracy in the location of the bounding box could lead to dramatic change to the reported distance and velocity of the MIO.
In our experiment, we preprocess detections output from CARLA to address missing data and outlier issues. First, we identify the outliers by the Matlab ``filloutliers" method, where we choose ``movmedian" as the detector and use linear interpolation to replace the outliers. The concrete Matlab command is:
$$
\text{\small filloutliers($Y$, `linear', `movmedian', 0, `ThresholdFactor', 0.5)},
$$
where $Y\in\mathbb{R}^{T\times 8}$ is the matrix of measurements and $T$ is the total number steps. In our experiment $T=295$. We perform the above outlier detection and replace operation twice to smooth the measurements.
Then, we apply the Matlab ``impute" function to interpolate the missing vision measurements. In Fig~\ref{fig:MIO-10_raw_measurement} and~\ref{fig:MIO+1_raw_measurement}, we show the preprocessed distance and velocity measurements from vision and radar compared with the ground-truth for both MIO-10 and MIO+1 datasets. Note that after preprocessing, both radar and vision measurements match with the ground-truth well.
\begin{figure}
\centering
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.85\textwidth]{figures/distance_MIO-10.png}
\caption{Distance measurements.}
\label{fig:MIO-10_raw_distance}
\end{subfigure}
\hfill
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.85\textwidth]{figures/velocity_MIO-10.png}
\caption{Velocity measurements.}
\label{fig:MIO-10_raw_velocity}
\end{subfigure}
\caption{On the MIO-10 dataset, the preprocessed vision measurements and the radar measurements match the ground-truth reasonably well.}
\label{fig:MIO-10_raw_measurement}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.85\textwidth]{figures/distance_MIO+1.png}
\caption{Distance measurements.}
\label{fig:MIO+1_raw_distance}
\end{subfigure}
\hfill
\begin{subfigure}{.23\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.85\textwidth]{figures/velocity_MIO+1.png}
\caption{Velocity measurements.}
\label{fig:MIO+1_raw_velocity}
\end{subfigure}
\caption{On the MIO+1 dataset, the preprocessed vision measurements and the radar measurements match the ground-truth reasonably well.}
\label{fig:MIO+1_raw_measurement}
\end{figure}
\section{Derivation of Surrogate Constraints} \label{appendix: surrogate_constraint}
The original attack optimization~\eqref{attack_fake_sp2:obj}-\eqref{attack_fake_sp2:stealthy} may not be convex due to that~\eqref{attack_fake_sp2:target} and~\eqref{attack_fake_sp2:stealthy} could be nonlinear. Our goal in this section is to derive convex surrogate constraints that are good approximations to~\eqref{attack_fake_sp2:target} and~\eqref{attack_fake_sp2:stealthy}. Furthermore, we require the surrogate constraints to be tighter than the original constraints, so that solving the attack under the surrogate constraints will always give us a feasible solution to the original attack. Concretely, we want to obtain surrogate constraints to $F(x)=\ell$, where $\ell\in\{\text{green, yellow, red}\}$. We analyze each case of $\ell$ separately:
\begin{itemize}
\item $\ell=\text{green}$
In this case, $F(x)=\ell$ is equivalent to $v\ge0$ according to~\eqref{eq:FCW}. While this constraint is convex, when we actually solve the optimization, it might be violated due to numerical inaccuracy. To avoid such numerical issues, we tighten it by adding a margin parameter $\epsilon>0$, and the derived surrogate constraint is $v\ge\epsilon$.
\item $\ell=\text{red}$
In this case, $F(x)=\ell$ is equivalent to
\begin{eqnarray}
& v&< 0\label{eq:appendix_red_1}\\
& d&\le -1.2 v+\frac{1}{0.8g}v^2\label{eq:appendix_red_2}.
\end{eqnarray}
Similar to case 1, we tighten the first constraint as
\begin{equation}
v\le -\epsilon.
\end{equation}
Note that by the first constraint, we must have $v<0$. The second constraint is $d\le -1.2v+\frac{1}{0.8g}v^2$. Given $v<0$, this is equivalent to
\begin{equation}
v\le 0.48g - \sqrt{(0.48g)^2+0.8g d}.
\end{equation}
We next define the following function
\begin{equation}
U(d)=0.48g - \sqrt{(0.48g)^2+0.8g d}.
\end{equation}
The first derivative is
\begin{equation}
U^\prime(d)=-\frac{0.4g}{\sqrt{(0.48g)^2+0.8g d}},
\end{equation}
which is increasing when $d\ge 0$. Therefore, the function $U(d)$ is convex.
We now fit a linear function that lower bounds $U(d)$. Specifically, since $U(d)$ is convex, for any $d_0\ge 0$, we have
\begin{equation}
U(d)\ge U^\prime(d_0)(d-d_0)+U(d_0).
\end{equation}
Therefore, $v\le U^\prime(d_0)(d-d_0)+U(d_0)$ is a tighter constraint than $v\le U(d)$. The two constraints are equivalent at $d=d_0$. Again, we need to add a margin parameter to avoid constraint violation due to numerical inaccuracy. With this in mind, the surrogate constraint becomes
\begin{equation}\label{eq:appendix_cons_surrogate_red}
v\le U^\prime(d_0)(d-d_0)+U(d_0)-\epsilon,
\end{equation}
Or, equivalently:
\begin{eqnarray}
&v&\le -\epsilon,\\
&v&\le U^\prime(d_0)(d-d_0)+U(d_0)-\epsilon,
\end{eqnarray}
This concludes the proof of our Proposition~\ref{prop:red_surrogate}.
\begin{figure}
\centerline{\includegraphics[width=.35\textwidth]{figures/surrogate_cons.png}}
\caption{Surrogate light constraints.}
\label{fig:red_surrogate}
\end{figure}
However, we still need to pick an appropriate $d_0$. In our scenario, the distance $d$ has physical limitation $d\in[\underline d, \bar d]$ with $\underline d=0$ and $\bar d=75$. The $U(d)$ curve for $d\in [0,75]$ is shown in Fig~\ref{fig:red_surrogate}. Based on the figure, we select $d_0$ such that $U^\prime(d_0)$ is equal to the slope of the segment connecting the two end points of the curve, i.e.,
\begin{equation}
U^\prime(d_0)=\frac{U(75)-U(0)}{75}=\frac{U(75)}{75}.
\end{equation}
We now derive the concrete surrogate constraints used in our experiment section. We begin with the following equation:
\begin{equation}
0.48g+\frac{0.4g}{U^\prime(d)}=U(d).
\end{equation}
From which, we can derive $d_0$:
\begin{equation}
d_0=\frac{1}{0.8g}\left((\frac{30g}{U(75)})^2-(0.48g)^2\right)
\end{equation}
and
\begin{equation}
U(d_0)=0.48g+\frac{30g}{U(75)}.
\end{equation}
By substituting $d_0$ and $U(d_0)$ into~\eqref{eq:appendix_cons_surrogate_red}, we find that the surrogate constraint is
\begin{equation}
v\le \frac{U(75)}{75}(d-d_0)+0.48g+\frac{30g}{U(75)}+\epsilon.
\end{equation}
\item $\ell=\text{yellow}$
In this case, $F(x)=\ell$ is equivalent to
\begin{eqnarray}
& v&< 0\\
& d&\ge -1.2 v+\frac{1}{0.8g}v^2.
\end{eqnarray}
Similarly, we tighten the first constraint to
\begin{equation}
v\le -\epsilon.
\end{equation}
For the second constraint, the situation is similar to $\ell=\text{red}$. $\forall d_0>0$. We have
\begin{equation}
v\ge \frac{U(d_0)}{d_0}d, \forall d\in [0, d_0]
\end{equation}
The above inequality is derived by fitting a linear function that is always above the $U(d)$ curve. Next, we select $d_0=75$ and add a margin parameter $\epsilon$ to derive the surrogate constraint:
\begin{eqnarray}
& v&\le -\epsilon\\
& v&\ge \frac{U(75)}{75}d+\epsilon.
\end{eqnarray}
To summarize, we have derived surrogate constraints for $F(x)=\ell$, where $l\in \{\text{green, yellow, red}\}$. When we solve the attack optimization, we replace each individual constraint of~\eqref{attack_fake_sp2:target} and~\eqref{attack_fake_sp2:stealthy} by one of the above three surrogate constraints. In Fig~\ref{fig:red_surrogate}, we show the surrogate constraints for red and yellow lights with $\epsilon=10^{-3}$.
\end{itemize}
\subsection{Attack Setup}
We perform preprocessing of CARLA measurements to remove outliers and interpolate missing data (see Appendix~\ref{appendix: preprocess}). Each step of our KF corresponds to one frame of the CARLA simulated video sequence (i.e., 0.05 seconds). We assume that the KF initializes its distance and velocity prediction to the average of the first vision and RADAR measurements. The acceleration is initialized to 0 in both directions. The covariance matrix is initialized to that used by Matlab FCW \cite{matlab-fcw}. Throughout the experiments, we let the effort matrix $R=I$, the margin parameter $\epsilon=10^{-3}$, and $\lambda=10^{10}$. We assume the human reaction time is $h^*=24$ steps (i.e., 1.2 seconds in our simulation).
\subsubsection{MIO-10 dataset}
We first simulate FCW to obtain the original warning lights without attack.
The first red light appears at step 98. Before this step, the lights are all yellow.
Without attack, the human driver will notice the red warning at step 98.
After 1.2 seconds of reaction time (24 steps), the driver will start braking at step 122.
The ground-truth distance to the MIO at the first application of brakes is 14.57m.
During braking, the distance between the ego vehicle and the MIO reduces by $10^2/0.8g\approx 12.76$m before stabilizing.
Since this is less than the ground-truth distance of 14.58m before braking, the crash can be avoided. This validates the potential effectiveness of FCW.
Our attacker aims to cause a crash. To accomplish this, the attacker suppresses the first 10 red warnings, so that the first red warning is delayed to step 108. As a result, the driver starts braking at step 132. The ground-truth distance to MIO at this step is 9.58m, which is below the minimum distance needed to avoid collision (12.76m). As such, a collision will occur. Therefore, we let the target interval be $\mathcal{T}^\dagger=[98, 107]$, and the target lights be $\ell_t^\dagger=\text{green}, \forall t\in \mathcal{T}^\dagger$.
\subsubsection{MIO+1 dataset}
In this scenario, the original warning lights without attack are all green. There is a trailing vehicle 7 m behind the ego vehicle, driving at the same velocity as the ego vehicle. Our attacker aims at causing the FCW to output red lights, so that the ego vehicle suddenly brakes unnecessarily and causes a rear collision with the trailing vehicle. To this end, the attacker changes the green lights in the interval [100, 139] to red, in which case the ego vehicle driver starts braking at step 124, after 1.2 seconds of reaction time. If the warning returns to green at step 140, the driver will react after 1.2 seconds and stop braking at step 164. Therefore, the driver continuously brakes for at least $(164-124)\times0.05=2$ seconds. Assuming the driver of the trailing vehicle is distracted, then during those 2 seconds, the distance between the trailing and the ego vehicle reduces by $0.2g \times2^2=7.84m>7m$, thus causing a rear-collision. Therefore, we let the target interval be $\mathcal{T}^\dagger=[100,139]$ and the target lights be $\ell_t^\dagger=\text{red}, \forall t\in \mathcal{T}^\dagger$.
\section{Acknowledgments}
This work was supported in part by the University of Wisconsin-Madison Office of the Vice Chancellor for Research and Graduate Education with funding from the Wisconsin Alumni Research Foundation. XZ acknowledges NSF grants 1545481, 1704117, 1836978, 2041428, 2023239 and MADLab AF CoE FA9550-18-1-0166.
\section{Ethics Statement}
Our paper studies attacks on advanced driver assistance systems (ADAS) with the goal of initiating research into defenses. We do not intend for the attacks to be deployed in the real world. However, studying attacks is critical to understanding what types of defenses must be built and where defense efforts should be focused. We take a first step towards robust ADAS by studying attacks on Kalman filters that are popularly used in these systems.
\section{Attack Problem Formulation}
We assume white-box setting where the attacker can access the KF parameters (e.g., through reverse engineering). The attacker can directly manipulate measurements (i.e., false data injection), but only pertaining to the vision component, and not the RADAR data. Our attack framework is agnostic of whether the attacker manipulates camera or RADAR, but we choose to only manipulate camera because of the increasing presence of deep learning techniques in ADAS and their general vulnerability to adversarial examples~\cite{szegedy2013intriguing,roadsigns17,athalye2017synthesizing,glasses}. We envision that future work can integrate our results into adversarial examples to create physical attacks.
We further restrict the attacker to only making physically plausible changes to the vision measurements. This is because an anomaly detection system might filter out physically implausible measurements (e.g., change of $10^4$m/s over one second). Concretely, we require that the distance and velocity measurement after attack must lie in $[\underline d,\bar d]$ and $[\underline v,\bar v]$ respectively. We let $[\underline d, \bar d]=[0,75]$ and $[\underline v,\bar v]=[-30,30]$. Finally, we assume that at any time step, the attacker knows the true measurement only for that time step, but does not know future measurements. To address this difficulty of an unknown future, we propose a model predictive control (MPC)-based attack framework that consists of an outer problem and an inner problem, where the inner problem is an instantiation of the outer problem with respect to attacker-envisioned future in every step of MPC. In the following, we first introduce the outer problem formulation.
\subsection{Outer Attack Problem}
Our attacker has a pre-specified target interval $\mathcal{T}^\dagger$, and aims at changing the warning lights output by FCW in $\mathcal{T}^\dagger$. As a result, the human driver sees different lights and takes unsafe actions.
Specifically, for any time $t\in \mathcal{T}^\dagger$, the attacker hopes to cause the FCW to output a desired target light $\ell_t^\dagger$, as characterized by~\eqref{attack_fake_sp2:target}, in which $F(\cdot)$ is the FCW alert logic~\eqref{eq:FCW}.
To accomplish this, the attacker manipulates measurements in an attack interval $\mathcal{T}^a$.
In our paper, we assume $\mathcal{T}^\dagger\subset \mathcal{T}^a$.
Furthermore, we consider only the scenario where $\mathcal{T}^\dagger$ and $\mathcal{T}^a$ have the same last step, since attacking after the target interval is not needed.
Let $\delta_t$ be the manipulation at step $t$, and $\tilde y_t=y_t+\delta_t$ be measurement after attack.
We refer to the $i$-th component of $\delta_t$ as $\delta_t^i$.
We next define the attack effort as the cumulative change over measurements $J = \sum_{t\in \mathcal{T}^a}\delta_t^\top R \delta_t$.
where $R\succ 0$ is the effort matrix.
The attacker hopes to minimize the attack effort.
Meanwhile, the attacker cannot arbitrarily manipulate measurements.
We consider two constraints on the manipulation.
First, MIO distance and velocity are limited by simple natural physics, as shown in~\eqref{attack_fake_sp2:physical}.
Moreover, similar to the norm ball used in adversarial examples, we impose another constraint that restricts the attacker's manipulation $\|\delta_t\|_\infty\le \Delta$ (see~\eqref{attack_fake_sp2:delta_max}).
We refer to $\mathcal{T}^s=\mathcal{T}^a\backslash\mathcal{T}^\dagger$, the difference between $\mathcal{T}^a$ and $\mathcal{T}^\dagger$, as the stealthy (or planning) interval.
During $\mathcal{T}^s$, the attacker can induce manipulations before the target interval with advance planning, and by doing so, hopefully better achieve the desired effect in the target interval. However, for the sake of stealthiness, the planned manipulation should not change the original lights $\ell_t$ during $\mathcal{T}^s$. This is characterized by the stealthiness constraint~\eqref{attack_fake_sp2:stealthy}.
Given all above, the attack can be formulated as:
\begin{eqnarray}
&\min_{\delta_t} &J=\sum_{t\in \mathcal{T}^a}\delta_t^\top R \delta_t,\label{attack_fake_sp2:obj}\\
&\mbox{s.t.} & \tilde y_t=y_t+\delta_t, \forall t\in\mathcal{T}^a,\label{attack_fake_sp2:attack}\\
& & \tilde x_{t} = A(I-H_{t-1}C)\tilde x_{t-1} +AH_{t-1} \tilde y_t,\label{attack_fake_sp2:KF}\\
& & \delta_t^i=0, \forall i\in\mathcal{I}_{\text{radar}}, \forall t\in \mathcal{T}^a,\label{attack_fake_sp2:only_vision}\\
& & \|\delta_t\|\le \Delta, \forall t\in \mathcal{T}^a,\label{attack_fake_sp2:delta_max}\\
& & \tilde d_t^{1,\nu}\in [\underline d,\bar d], \tilde v_t^{1,\nu}\in [\underline v,\bar v], \forall t\in \mathcal{T}^a,\text{\quad\quad}\label{attack_fake_sp2:physical}\\
& & F(\tilde x_t)=\ell_t^\dagger, \forall t\in \mathcal{T}^\dagger,\label{attack_fake_sp2:target}\\
& & F(\tilde x_t)=\ell_t, \forall t\in \mathcal{T}^s. \label{attack_fake_sp2:stealthy}
\end{eqnarray}
The constraint \eqref{attack_fake_sp2:KF} specifies the evolution of the state prediction under the attacked measurements $\tilde y_t$.~\eqref{attack_fake_sp2:only_vision} enforces no change on radar measurements, where $\mathcal{I}_{\text{radar}}=\{5,6,7,8\}$ contains indexes of all radar components .
The attack optimization is hard to solve due to three reasons:
\begin{enumerate}[(1).]
\item The problem could be non-convex.
\item The problem could be be infeasible.
\item The optimization is defined on measurements $y_t$ that are not visible until after $\mathcal{T}^a$, while the attacker must design manipulations $\delta_t$ during $\mathcal{T}^a$ in an online manner.
\end{enumerate}
We now explain how to address the above three issues.
The only potential sources of non-convexity in our attack are~\eqref{attack_fake_sp2:target} and~\eqref{attack_fake_sp2:stealthy}. We now explain how to derive a surrogate convex problem using $\ell_t^\dagger=\ell_t^o=\text{red}$ as an example. The other scenarios are similar, thus we leave the details to Appendix~\ref{appendix: surrogate_constraint}. The constraint $F(\tilde x_t)=\text{red}$ is equivalent to
\begin{eqnarray}
& \tilde v_t^{1,\nu}&< 0, \label{eq:cons_red_1}\\
& \tilde d_t^{1,\nu}&\le -1.2\tilde v_t^{1,\nu}+\frac{1}{0.8g}(\tilde v_t^{1,\nu})^2.\label{eq:cons_red_2},
\end{eqnarray}
The above constraints result in non-convex optimzation mainly because~\eqref{eq:cons_red_2} is nonlinear. To formulate a convex problem, we now introduce surrogate constraints that are tighter than~\eqref{eq:cons_red_1},~\eqref{eq:cons_red_2} but guarantee convexity.
\begin{proposition}\label{prop:red_surrogate}
Let $U(d)=0.48g-\sqrt{(0.48g)^2+0.8gd}$. Let $\epsilon>0$ be any positive number. Then for any $d_0\ge0$, the surrogate constraints~\eqref{eq:cons_surro_red_1},~\eqref{eq:cons_surro_red_2} are tighter than $F(\tilde x_t)=\text{red}$, and induce convex attack optimization.
\begin{eqnarray}
&\tilde v_t^{1,\nu}&\le -\epsilon,\label{eq:cons_surro_red_1}\\
&\tilde v_t^{1,\nu}&\le U^\prime(d_0)(\tilde d_t^{1,\nu}-d_0)+U(d_0)-\epsilon.\label{eq:cons_surro_red_2}
\end{eqnarray}
\end{proposition}
We provide a proof and guidance on how to select $d_0$ in Appendix~\ref{appendix: surrogate_constraint}. With the surrogate constraints, the attack optimization becomes convex. However, the surrogate optimization might still be infeasible. To address the feasibility issue, we further introduce slack variables into~\eqref{eq:cons_surro_red_1},~\eqref{eq:cons_surro_red_2} to allow violation of stealthiness and target lights:
\begin{eqnarray}
&\tilde v_t^{1,\nu}&\le -\epsilon+\xi_t,\label{eq:cons_surro_slack_red_1}\\
&\tilde v_t^{1,\nu}&\le U^\prime(d_0)(\tilde d_t^{1,\nu}-d_0)+U(d_0)-\epsilon+\zeta_t.\quad\label{eq:cons_surro_slack_red_2}
\end{eqnarray}
We include these slack variables in the objective function:
\begin{equation}\label{eq:definitionJ1J2J3}
J=\underbrace{\sum_{t\in \mathcal{T}^a}\delta_t^\top R \delta_t}_{\text{total manipulation $J_1$}}+\lambda \underbrace{\sum_{t\in \mathcal{T}^s}(\xi_t^2+\zeta_t^2)}_{\text{stealthiness violation $J_2$}}+\lambda \underbrace{\sum_{t\in \mathcal{T}^\dagger}(\xi_t^2+\zeta_t^2)}_{\text{target violation $J_3$}}.
\end{equation}
Then, the surrogate attack optimization is
\begin{eqnarray}
&\min_{\delta_t} &J=J_1+\lambda J_2+\lambda J_3,\label{attack_surrogate_sp2:obj}\\
&\mbox{s.t.} &\text{\eqref{attack_fake_sp2:attack}-\eqref{attack_fake_sp2:physical},~\eqref{eq:cons_surro_slack_red_1},~\eqref{eq:cons_surro_slack_red_2}}.\label{attack_surrogate_sp2:cons}
\end{eqnarray}
\vspace{-0.5cm}
\begin{proposition}
The attack optimization~\eqref{attack_surrogate_sp2:obj}-\eqref{attack_surrogate_sp2:cons} with surrogate constraints and slack variables is convex and feasible.
\end{proposition}
\subsection{Inner Attack Problem: MPC-based Attack}
In the outer surrogate attack~\eqref{attack_surrogate_sp2:obj}-\eqref{attack_surrogate_sp2:cons}, we need to assume the attacker knows the measurements $y_t$ in the entire attack interval $\mathcal{T}^a$ beforehand. However, the attacker cannot know the future. Instead, he can only observe and manipulate the current measurement in an online manner. To address the unknown future issue, we adopt a control perspective and view the attacker as an adversarial controller of the KF, where the control action is the manipulation $\delta_t$. We then apply MPC, an iterative control method that progressively solves~\eqref{attack_surrogate_sp2:obj}-\eqref{attack_surrogate_sp2:cons}. By using MPC, the attacker is able to adapt the manipulation to the instantiated measurements revealed over time while accounting for unknown future measurements.
Specifically, in each step $t$, the attacker has observed all past measurements $y_{1},...y_{t-1}$ and the current measurement $y_t$. Thus, the attacker can infer the clean state $\hat x_t$ in the case of no attacker intervention. Based on $\hat x_t$, the attacker can recursively predict future measurements by simulating the environmental dynamics without noise, i.e., $\forall \tau>t$:
\begin{equation}\label{eq:attack_predict}
x_{\tau}^\prime = A x_{\tau -1}^\prime, \hat y_{\tau} = C x_{\tau}^\prime.
\end{equation}
The recursion starts from $x_t^\prime=\hat x_t$. The attacker then replaces the unknown measurements in the outer attack by its prediction $\hat y_\tau$ $(\tau>t)$ to derive the following inner attack:
\begin{eqnarray}
&\min_{\delta_{\tau:\tau\ge t}} &J=\sum_{\tau\in \mathcal{T}^a}\delta_\tau^\top R \delta_\tau+\lambda \sum_{\tau\in \mathcal{T}^a}(\xi_\tau^2+\zeta_\tau^2),\quad\quad\label{eq:MPC_red_1}\\
&\mbox{s.t.} &\tilde y_\tau=\hat y_\tau+\delta_\tau, \forall \tau\ge t,\\
& & \text{\text{\eqref{attack_fake_sp2:KF}-\eqref{attack_fake_sp2:physical},~\eqref{eq:cons_surro_slack_red_1},~\eqref{eq:cons_surro_slack_red_2}}}\text{ (defined on $\tau\ge t$)} .\label{eq:MPC_cons_red_1}
\end{eqnarray}
The attacker solves the above inner attack in every step $t$. Assume the solution is $\delta_\tau (\tau\ge t)$. Then, the attacker only implements the manipulation on the current measurement, i.e., $\tilde y_t=y_t+\delta_t$, and discards the future manipulations. After that, the attacker enters step $t+1$ and applies MPC again to manipulate the next measurement. This procedure continues until the last step of the attack interval $\mathcal{T}^a$. We briefly illustrate the MPC-based attack in algorithm~\ref{alg:MPC}.
\begin{algorithm}[t]
\SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output}
\Input{target interval $\mathcal{T}^\dagger$, target lights $\ell_t^\dagger, t\in \mathcal{T}^\dagger$, stealthy interval $\mathcal{T}^s$, original lights $\ell_t, t\in \mathcal{T}^s$.}
Initialize $\hat x_1$ and $\hat \Sigma_1$. Let $\tilde x_1=\hat x_1$, $\mathcal{T}^a=\mathcal{T}^s\cup \mathcal{T}^\dagger$\;
\For{$t\leftarrow 2$ \KwTo $T$}{
environment generates measurement $y_t$\;
\uIf{$t\in\mathcal{T}^a$}
{
attacker infers clean state $\hat x_t$ without attack\;
attacker predicts future $\hat y_t$ with $\eqref{eq:attack_predict}$\;
attacker solves~\eqref{eq:MPC_red_1}-\eqref{eq:MPC_cons_red_1} to obtain $\delta_\tau$ $(\tau\ge t)$\;
attacker manipulates $y_t$ to $\tilde y_t=y_t+\delta_t$\;
$\tilde x_t$ evolves to $\tilde x_{t+1}$ according to $\tilde y_t$
}
\lElse
{
$\tilde x_t$ evolves to $\tilde x_{t+1}$ according $y_t$
}
}
\caption{MPC-based attack.}
\label{alg:MPC}
\end{algorithm}
\section{Conclusion}
We formulate the adversarial attack of Kalman Filter as an optimal control problem. We demonstrate that our planning-based attack can manipulate the FCW to output incorrect warnings, which mislead human drivers to behave unsafely and cause crash. Our study incorporates human behaviors and applies to general machine-human hybrid systems.
\section{Experiments on CARLA Simulation}
In this section, we empirically study the performance of the MPC-based attack. We first describe the simulation setup.
\input{fcw_simplification}
\input{attack_exp_setup}
\input{results}
\subsection{Simulation Setup}
We use CARLA~\cite{carla}, a high-fidelity vehicle simulation environment, to generate measurement data that we input to the Kalman filter-based FCW. CARLA supports configurable sensors and test tracks. We configure the simulated vehicle to contain a single forward-facing RGB camera (800x600 pixels), a forward-facing depth camera of the same resolution, and a single forward-facing RADAR (15$^{\circ}$ vertical detection range, 6000 points/sec, 85 m maximum detection distance). We took this configuration from a publicly-available FCW implementation~\cite{matlab-fcw}. The simulation runs at 20 frames/sec and thus, each sensor receives data at that rate. Furthermore, this configuration is commonly available on production vehicles today~\cite{tesla-adas}, and thus, our simulation setup matches real-world FCW systems from a hardware perspective.
For each time step of the simulation, CARLA outputs a single RGB image, a depth map image, and variable number of RADAR points. We use YOLOv2~\cite{yolov2} to produce vehicle bounding boxes, the Hungarian pairwise matching algorithm \cite{hungarian} to match boxes between frames, and the first derivative of paired depth map image readings to produce vehicle detections from vision with location and velocity components. Details of processing and formatting of CARLA output can be found in Appendix~\ref{appendix:carla-output-processing}. This process produces measurements that match ground truth velocity and distance closely.
Although there are infinitely many possible physical situations where an FCW alert could occur involving two vehicles, they reside in a small set of equivalence classes. The National Highway Traffic Safety Administration (NHTSA) has outlined a set of testing conditions for assessing the efficacy of FCW alerts~\cite{nhtsa-guidelines}. It involves a two vehicles on a straight test track at varying speeds. Based on these real-world testing guidelines, we develop the following two scenarios:
\subsubsection{MIO-10: Collision between two moving vehicles}
The ego and MIO travel on a straight road, with a negative relative velocity between the two vehicles. Specifically, the ego travels at 27 m/s (\textasciitilde60 mph) and the MIO at 17 m/s (\textasciitilde38 mph). These correspond to typical freeway speed differences of adjacent vehicles. In the absence of any other action, the ego will eventually collide with the MIO. In our simulations, we let this collision occur and record camera and RADAR measurements throughout. Since the relative velocity of the MIO to the ego is -10m/s, we refer to this dataset as MIO-10.
\subsubsection{MIO+1: No collision}
The ego and MIO travel on a straight road, with a positive relative velocity between the two vehicles. Specifically, the ego travels at 27 m/s (\textasciitilde60 mph) and the MIO at 28 m/s (\textasciitilde63 mph). A trailing vehicle moving at 27 m/s follows the ego 7 m behind. In the absence of any other action, the ego and trailing vehicle will not collide. We collect measurements until the MIO moves out of sensor range of the ego. We refer to this dataset as MIO+1.
The above scenarios correspond to basic situations where the ego vehicle has an unobstructed view of the MIO and represents a best-case for the FCW system. Attacks on these two settings are the hardest to achieve and comprehensively demonstrate the efficacy of our MPC-based attack.
\section{Introduction}
Advanced Driver Assistance Systems (ADAS) are hybrid human-machine systems that are widely deployed on production passenger vehicles~\cite{nhtsa-adas}. They use sensing, traditional signal processing and machine learning to detect and raise alerts about unsafe road situations and rely on the human driver to take corrective actions. Popular ADAS examples include Forward Collision Warning (FCW), Adaptive Cruise Control and Autonomous Emergency Braking (AEB).
Although ADAS hybrid systems are designed to increase road safety when drivers are distracted, attackers can negate their benefits by strategically tampering with their behavior. For example, an attacker could convince an FCW or AEB system that there is no imminent collision until it is too late for a human driver to avoid the crash.
We study the robustness of ADAS to attacks. The core of ADAS typically involves tracking the states (e.g., distance and velocity) of road objects using Kalman filter (KF). Downstream logic uses this tracking output to detect unsafe situations before they happen. We focus our efforts on Forward Collision Warning (FCW), a popular ADAS deployed on production vehicles today. FCW uses KF state predictions to detect whether the ego vehicle (vehicle employing the ADAS system) is about to collide with the most important object in front of it and will alert the human driver in a timely manner. Thus, our concrete attack goal is to trick the KF that FCW uses and make it output incorrect state predictions that would induce false or delayed alerts depending on the specific physical situation.
Recent work has examined the robustness of road object state tracking for autonomous vehicles~\cite{iclr-mot}. Their attacks create an instantaneous manipulation to the Kalman filter inputs without considering its sequential nature, the downstream logic that depends on filter output, or the physical dynamics of involved vehicles. This leads to temporarily hijacked Kalman filter state predictions that are incapable of ensuring that downstream logic is reliably tricked into producing false alerts. By contrast, we adopt an online planning view of attacking KFs that accounts for: (1) their sequential nature where current predictions depend on past measurements; and (2) the downstream logic that uses KF output to produce warnings. Our attack technique also considers a simplified model of human reaction to manipulated FCW warning lights.
We propose a novel Model Predictive Control (MPC)-based attack that can sequentially manipulate measurement inputs to a KF with the goal of stealthily hijacking its behavior. Our attacks force FCW alerts that mask the true nature of the physical situation involving the vehicles until it is too late for a distracted human driver to take corrective actions.
We evaluate our attack framework by creating a high-fidelity driving simulation using CARLA~\cite{carla}, a popular tool for autonomous vehicle research and development. We create test scenarios based on real-world driving data~\cite{nhtsa-guidelines,euro-ncap-protocol} and demonstrate the practicality of the attack in causing crashes involving the victim vehicle. Anonymized CARLA simulation videos of our attacks are available at \url{https://sites.google.com/view/attack-kalman-filter}.
\noindent\textbf{Main Contributions:}
\begin{itemize}
\item We develop an optimal control-based attack against the popular FCW driver assistance system. Our attack targets several critical parts of the FCW pipeline -- Kalman filter tracking and prediction, FCW alert logic and human decision making in crash and near-crash scenarios.
\item We evaluate our control-based attacks in a high-fidelity simulation environment demonstrating that an attacker can compromise \emph{only} the camera-based measurement data and accomplish their goals of creating end-to-end unsafe situations for an FCW system, even under the constraint of limited manipulation to measurements.
\item We show that attack planning in advance of the targeted point is beneficial to the attack compared to without planning. Given 25 steps of planning (or 1.25 seconds based on specific physical situations in our evaluation) before the targeted time point, the attacker can cause the desired effect, while the attack fails without planning. Furthermore, via comparisons against a baseline greedy attack, we show that our attack can find near-optimal planning that achieves better overall performance.
\end{itemize}
\section{Background}
Forward Collision Warning provides audio-visual alerts to warn human drivers of imminent collisions. Fig.~\ref{fig:overview} shows the pipeline of a prototypical FCW hybrid system~\cite{matlab-fcw}: (1) It uses camera and RADAR sensors to perceive the environment; (2) It processes sensor data using a combination of traditional signal processing and machine learning algorithms to derive object velocities and distances; (3) A Kalman filter tracks the Most Important Object (MIO) state and makes predictions about its future states; (4) FCW logic uses Kalman filter predictions to determine whether a collision is about to occur and creates audio-visual warnings; (5) A human driver reacts to FCW alerts. These alerts can be either: green -- indicating no danger, yellow -- indicating potential danger of forward collision, and red -- indicating imminent danger where braking action must be taken.
We focus on attacking the core steps of FCW (shaded parts of Fig.~\ref{fig:overview}). Thus, we assume there is a single MIO in front of the ego vehicle and a single Kalman filter actively tracking its state. The steps of measurement assignment and MIO identification will not be considered in this paper.
\begin{figure}[t!]
\centering \includegraphics[width=1\columnwidth,height=0.443\textwidth]{figures/fcw.png}
\caption{Overview of Forward Collision Warning (FCW) hybrid human-machine system. We take a first step to understanding the robustness of this system to attackers who can compromise sensor measurements. Therefore, we filter the problem to its essence (shaded parts) --- the Kalman filter that tracks the most important object (MIO) and the downstream logic that decides how to warn the driver.}
\label{fig:overview}
\end{figure}
We have two attack goals that will comprehensively demonstrate the vulnerability of FCW hybrid systems --- the attacker should trick FCW into showing no red alerts when there is an imminent collision with the most important object (MIO), and vice versa --- the attacker should trick FCW into showing red alerts when there is no collision, inducing a human to react with braking that can potentially lead to a rear-end crash with a trailing vehicle.
\subsection{Kalman Filtering}
At the core of FCW is the Kalman Filter, which estimates the state of the MIO based on sensor measurements. In this paper, the state of the MIO is represented as $x_t = (d_t^1, v_t^1, a_t^1, d_t^2, v_t^2, a_t^2)$, where $d_t^1$, $v_t^1$, $a_t^1$ are the distance, velocity and acceleration of the MIO along the driving direction, and $d_t^2, v_t^2, a_t^2$ for the lateral direction (perpendicular to driving direction). Then KF models the evolution of $x_t$ as
\begin{equation}\label{eq:transition_model}
x_{t+1}=Ax_t + \omega_t, t\ge 1,
\end{equation}
where $A$ is the state-transition matrix and $\omega_t \sim N(0, \Omega)$ is Gaussian noise.
The underlying state $x_t$ is unknown, but one can obtain measurements $y_t$ of the state as
\begin{equation}\label{eq:measurement_model}
y_t = C x_t+\psi_t, t\ge 1,
\end{equation}
where $C$ is the measurement matrix and $\psi_t \sim N(0, \Psi)$ is the measurement noise. In our paper, $y_t\in\mathbb{R}^8$ contains vision and radar measurements of the MIO distance and velocity along two directions, i.e., $y_t =(d_t^{1,\nu}, v_t^{1,\nu}, d_t^{2,\nu}, v_t^{2,\nu}, d_t^{1,r}, v_t^{1,r}, d_t^{2,r}, v_t^{2,r})$, where we use superscripts $\nu$, $r$ for vision and radar, and numbers 1, 2 for driving and lateral direction, respectively.
Given the state dynamics~\eqref{eq:transition_model} and measurement model~\eqref{eq:measurement_model}, KF provides a recursive formula to estimate the state based on sequential measurements obtained over time.
Concretely, KF starts from some initial state and covariance prediction $\hat x_1$ and $\hat \Sigma_1$. Then for any $t\ge 2$, KF first applies~\eqref{eq:KF_correction} to correct the predictions based on measurements $y_t$. The corrected state and covariance matrix are denoted by $\bar x_t$ and $\bar \Sigma_t$.
\begin{equation}\label{eq:KF_correction}
\begin{aligned}
\bar x_{t} &= (I-H_{t-1}C)\hat x_{t-1} +H_{t-1} y_t,\\
\bar \Sigma_{t} &= (I - H_{t-1}C)\hat \Sigma_{t-1}.
\end{aligned}
\end{equation}
where $H_{t-1}=\hat \Sigma_{t-1}C^\top (C\hat \Sigma_{t - 1}C^\top+\Psi)^{-1}$.
Next, KF applies~\eqref{eq:KF_prediction} to predict state and covariance for the next step.
\begin{equation}\label{eq:KF_prediction}
\hat x_{t} = A\bar x_{t}, \quad \hat \Sigma_{t}= A\bar \Sigma_{t}A^\top +\Omega.
\end{equation}
The correction and prediction steps are applied recursively as $t$ grows. Note that the derivation of covariance matrix is independent of $y_t$, thus can be computed beforehand.
\subsection{Warning Alert Logic and Human Model}
In this paper, we follow the FCW alert logic used in~\cite{matlab-fcw}.
Let the state prediction be $\hat x_t=(\hat d_t^1, \hat v_t^1, \hat a_t^1, \hat d_t^2, \hat v_t^2, \hat a_t^2)$, then the warning light $\ell_t$ output by FCW at step $t$ is one of the following three cases:
\begin{itemize}
\item Safe (Green): The MIO is moving away, or the distance to MIO remains constant, i.e., $\hat v_{t}^1\ge0$.
\item Caution (Yellow): The MIO is moving closer, but still at a distance further than the minimum safe distance $d^*(\hat v_t^1)$, i.e., $\hat v_{t}^1<0$ and $\hat d_{t}^1>d^*(\hat v_t^1)$. We define the safe distance as $d^*(\hat v_t^1)=-1.2\hat v_{t}^1+(\hat v_{t}^1)^2/0.8g$, where $g$ is 9.8 $m/s^2$.
\item Warn (Red): The MIO is moving closer, and at a distance less than the minimum safe distance, i.e., $\hat v_{t}^1<0$ and $\hat d_{t}^1\le d^*(\hat v_t^1)$.
\end{itemize}
The FCW alert logic can be summarized as:
\begin{equation}\label{eq:FCW}
F(\hat x_{t}) = \left\{
\begin{array}{ll}
\text{green} & \mbox{if $\hat v_{t}^1\ge0$,} \\
\text{yellow} & \mbox{if $\hat v_{t}^1< 0, \hat d_{t}^1>d^*(\hat v_t^1)$,} \\
\text{red} & \mbox{if $\hat v_{t}^1< 0, \hat d_{t}^1\le d^*(\hat v_t^1)$.}
\end{array}
\right.
\end{equation}
Given the FCW warning light, the human driver could be in one of the following two states -- applying the brake pedal, or not applying/releasing the brake. We take into account human reaction time $h^*$; warning lights must sustain at least $h^*$ steps before the human driver switches state. That is, the driver brakes after $h^*$ steps since the first red light, and releases the brake after $h^*$ steps since the first yellow/green light. Note that the yellow and green lights are treated identically in both cases because the MIO is outside the safe distance and no brake is needed. In appendix~\ref{appendix: human}, we provide an algorithmic description of the human model.
\section{Related Work}
\noindent\textbf{Attacks on Object Tracking.} Recent work has examined the vulnerability of multi-object tracking (MOT)~\cite{iclr-mot}. Although this work does consider the downstream logic that uses the outputs of ML-based computer vision, our work goes beyond in several ways. First, we consider a hybrid system that involves both human and machine. Second, we consider the more realistic case of sensor fusion involving RADAR and camera measurements that is deployed in production vehicles today. Prior work assumed a system that only uses a single camera sensor. Third, we examine a complete FCW pipeline that uses object tracking data to predict collisions and issue warnings. Prior work only considered MOT without any further logic that is necessarily present in realistic systems. Finally, our attack algorithm accounts for the sequential nature of decision making in ADAS.
\noindent\textbf{Vision Adversarial Examples.} ML models are vulnerable to adversarial examples~\cite{szegedy2013intriguing}, with a bulk of research in the computer vision space~\cite{goodfellow2014explaining,papernot2016limitations,carlini2017towards,shafahi2018poison,chen2017targeted}. Recent work has demonstrated physical attacks in the real world~\cite{patch,athalye2017synthesizing,yolo,glasses}. For example, attackers can throw inconspicuous stickers on stop signs and cause the model to output a speed limit sign~\cite{roadsigns17}. However, all of this work studies the ML model in isolation without considering the cyber-physical system that uses model decisions. By contrast, we contribute the first study that examines the security of FCW --- a hybrid human-machine system, and we introduce a novel control-based attack that accounts for these aspects while remaining stealthy to the human driver.
\noindent\textbf{Control-based Attacks on KF.} Prior work in control theory has studied false data injection attacks on Kalman filters~\cite{bai2017kalman, kung2016performance, zhang2016stealthy, chen2016cyber, yang2016false, chen2017optimal}. Our work assumes a similar attack modality -- the attacker can manipulate measurements. However, prior work does not consider the downstream logic and human behavior that depends on the KF output. By contrast, we provide an attack framework demonstrating end-to-end effects that cause crashes in distracted driving scenarios.
\noindent\textbf{Attacks on Sequential Systems.} There are recent works that study attacks of other sequential learning systems from a control perspective~\cite{chen2020optimal,zhang2020online,zhang2020adaptive,jun2018adversarial}. Most of them focus on analyzing theoretical attack properties, while we contribute an application of control-based attacks in a practical domain.
\subsection{The MPC-based Attack Is Successful}
Our first result shows that the MPC-based attack can successfully cause the FCW to output the desired warning lights in the target interval $\mathcal{T}^\dagger$. In this experiment, we let $\Delta=\infty$ and the stealthy interval $\mathcal{T}^s$ start at step 2. In Fig.~\ref{fig:MIO-10_warning} and~\ref{fig:MIO+1_warning}, we show the warning lights in $\mathcal{T}^\dagger$ (shaded in red). For MIO-10, the attacker achieves the desired red lights in the entire $\mathcal{T}^\dagger$, while maintaining the original yellow lights in $\mathcal{T}^s$. For MIO+1, the attacker failed to achieve the red warning at step 100, but is successful in all later steps. We verified that the attack still leads to a collision. In fact, the attacker can tolerate at most two steps of failure in the beginning of $\mathcal{T}^\dagger$ while still ensuring that the collision occurs. There is an unintended side effect in $\mathcal{T}^s$ where green lights are changed to yellow. However, this side effect is minor since the driver will not brake when yellow lights are produced. In many production vehicles, green and yellow lights are not shown to the driver --- only the red warnings are shown.
In Fig.~\ref{fig:MIO-10_attack_distance},~\ref{fig:MIO-10_attack_velocity}, we note that for MIO-10, the manipulation is mostly on velocity, and there are early planned manipulations starting from step 70. A large increase in velocity happens at step 100 (the first step of $\mathcal{T}^\dagger$), which causes the KF's velocity estimation to be positive, resulting in a green light. After that, velocity measurements are further increased to maintain a positive velocity estimation. In Fig.~\ref{fig:MIO+1_attack_distance},~\ref{fig:MIO+1_attack_velocity}, we show manipulations on MIO+1. The overall trend is that the attacker reduces the perceived MIO distance and velocity. As a result, KF estimates the MIO to be close than the safe distance in $\mathcal{T}^\dagger$, thus red lights are produced. During interval [88,96], There is an exceptional increase of velocity. We provide a detailed explanation for that increase in Appendix~\ref{appendix:increase}.
In Fig.~\ref{fig:MIO-10_state_traj},~\ref{fig:MIO+1_state_traj}, we show the trajectory of KF state prediction projected onto the distance-velocity space during interval $\mathcal{T}^a$. We partition the 2D space into three regions, green (G), yellow (Y) and red (R). Each region contains the states that trigger the corresponding warning light. The trajectory without attack (blue) starts from location 1 and ends at 2. After attack, the trajectory (dark) is steered into the region of the desired warning light, ending at location 3. Note that during $\mathcal{T}^\dagger$, the state after attack lies on the boundary of the desired region. This is because our attack minimizes manipulation effort. Forcing a state deeper into the desired region would require more effort, increasing the attacker's cost.
\subsection{Attack Is Easier with More Planning Space}
Our second result shows that the attack is easier when the attacker has more time to plan, or equivalently, a longer stealthy interval $\mathcal{T}^s$. The stealthy interval is initially of full length, which starts from step 2 until the last step prior to $\mathcal{T}^\dagger$. Then, we gradually reduce the length by 1/4 of the full length until the interval is empty. This corresponds to 5, 3.75, 2.5, 1.25 and 0 seconds of planning space before the target interval $\mathcal{T}^\dagger$. We denote the number of light violations in $\mathcal{T}^\dagger$ as $V^\dagger=\sum_{t\in \mathcal{T}^\dagger}\tilde \ell_t\neq \ell_t^\dagger$, and similarly $V^s$ for $\mathcal{T}^s$. We let $\Delta=\infty$. In Table~\ref{table:MIO-10} and~\ref{table:MIO+1}, we show $V^\dagger$, $V^s$ together with $J_1, J_2, J_3$ and $J$ as defined in~\eqref{eq:definitionJ1J2J3} for MIO-10 and MIO+1 respectively. Note that on both datasets, the violation $V^\dagger$ and the total objective $J$ decrease as the length of $\mathcal{T}^s$ grows, showing that the attacker can better accomplish the attack goal given a longer interval of planning.
On MIO-10, when $\mathcal{T}^s$ is empty, the attack fails to achieve the desired warning in all target steps. However, given 1.25s of planning before $\mathcal{T}^\dagger$, the attacker forces the desired lights throughout $\mathcal{T}^\dagger$. Similarly, on MIO+1, when $\mathcal{T}^s$ is empty, the attack fails in the first three steps of $\mathcal{T}^\dagger$, and the collision will not happen. Given 1.25s of planning before $\mathcal{T}^\dagger$, the attack only fails in the first step of $\mathcal{T}^\dagger$, and the collision happens. This demonstrates that planning in $\mathcal{T}^s$ benefits the attack.
\begin{table}
\centering
\caption{\small{$V^\dagger$, $V^s$, $J_1$, $J_2$, $J_3$ and $J$ for the MIO-10 dataset.}}\label{table:MIO-10}
\small
\resizebox{0.48\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
& \multicolumn{6}{|c|}{MPC-based attack} & \multicolumn{6}{|c|}{Greedy attack} \\
\hline
$\mathcal{T}^s$ &$V^\dagger$ & $V ^s$ & $J_1$ & $J_2$ & $J_3$ & $J$ &$V^\dagger$ &$V^s$ & $J_1$ & $J_2$ & $J_3$ & $J$ \\
\hline
$0$ & 1 & 0 &7.1e3 &0 &7.4 &7.4e10 & 1 & 0 &4.6e3 &98.4 &7.4 &1.1e12 \\
$1.25$ &0 & 0&4.4e3&0 &0&4.3e3 & 0 & 23&1.3e5 &3.3e3 &0 & 3.3e13 \\
$2.5$ &0 & 0 &4.4e3&0 &0 & 4.4e3 & 0 & 47 &2.0e5 &5.4e3 &0 &5.4e13 \\
$3.75$ &0 & 0 &4.4e3&0 &0 & 4.4e3 & 0 & 71 &2.5e5 &7.6e3 &0 &7.5e13\\
$5$ & 0 & 0 &4.4e3&0 &0 & 4.4e3 & 0 & 96 & 2.9e5 &9.2e3 & 0& 9.2e13 \\
\hline
\end{tabular}}
\end{table}
\begin{table}
\centering
\caption{\small{$V^\dagger$, $V^s$, $J_1$, $J_2$, $J_3$ and $J$ for the MIO+1 dataset.}}\label{table:MIO+1}
\small
\resizebox{0.48\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
& \multicolumn{6}{|c|}{MPC-based attack} & \multicolumn{6}{|c|}{Greedy attack} \\
\hline
$\mathcal{T}^s$ &$V^\dagger$ & $V ^s$ & $J_1$ & $J_2$ & $J_3$ & $J$ &$V^\dagger$ & $V ^s$ & $J_1$ & $J_2$ & $J_3$ & $J$ \\
\hline
$0$ & 3 & 0 &3.3e4 &0 &1.2e2 &1.2e12 & 3 & 0 &1.1e5 &0 &1.2e2 &1.2e12 \\
$1.25$ &1 & 14 &7.6e4&6.8 &11.0&1.8e11 & 0 & 25 &1.7e5 &6.1e3 &0 & 6.1e13 \\
$2.5$ &1& 39 &1.1e5&4.2 &6.9 & 1.1e11 & 0 & 49 &2.3e5 &1.1e4 &0 &1.1e14 \\
$3.75$ &1 & 58 &1.5e5&3.5 &5.9 & 9.4e10 & 0 & 74 &3.0e5 &1.6e4 &0 &1.6e14\\
$5$ & 1 & 58 &1.8e5&3.3 &5.6 & 9.0e10 & 0 & 98 & 3.5e5 &2.0e4 & 0& 2.0e14 \\
\hline
\end{tabular}}
\end{table}
\subsection{Attack Is Easier as $\Delta$ Increases}
In this section, we show that the attack becomes easier as the upper bound on the manipulation $\Delta$ grows. In this experiment, we focus on the MIO-10 dataset and let $\mathcal{T}^\dagger$ start from step 2. In Fig~\ref{fig:manipulation_MIO-10_delta}, we show the manipulation on measurements for $\Delta=14,16,18$ and $\infty$. The number of green lights achieved by the attacker in the target interval is 0, 4, 10 and 10 respectively. This shows the attack is easier for larger $\Delta$. Note that for smaller $\Delta$, the attacker's manipulation becomes flatter due to the constraint $\|\delta_t\|\le \Delta$. But, more interestingly, the attacker needs to start the attack earlier to compensate for the decreasing bound. We also note that the minimum $\Delta$ to achieve the desired green lights over the entire target interval (to integer precision) is 18.
\begin{figure}[H]
\begin{subfigure}{.235\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/distance_measurement_Delta_MIO-10.png}
\caption{Manipulation on distance.}
\label{fig:MIO-10_distance_Delta}
\end{subfigure}%
\hfill
\begin{subfigure}{.235\textwidth}
\centering
\includegraphics[width=1\textwidth, height=0.7\textwidth]{figures/velocity_measurement_Delta_MIO-10.png}
\caption{Manipulation on velocity.}
\label{fig:MIO-10_velocity_Delta}
\end{subfigure}%
\caption{Manipulation on measurements with different upper bound $\Delta$. As $\Delta$ grows, the attack becomes easier.}
\label{fig:manipulation_MIO-10_delta}
\end{figure}
\subsection{Comparison Against Greedy Attacker}
In this section, we introduce a greedy baseline attacker. For MIO-10, since the attack goal is to achieve green lights in $\mathcal{T}^\dagger$, the greedy attacker always increases the distance and velocity to the maximum possible value, i.e.,
$$\tilde d_t^{1,\nu}=\min\{d_t^{1, \nu}+\Delta, \bar d\}, \tilde v_t^{1,\nu}=\min\{v_t^{1, \nu}+\Delta, \bar v\}, \forall t\in \mathcal{T}^a.$$
Similarly, for MIO+1, the attacker always decreases the distance and velocity to the minimum possible value.
In table~\ref{table:MIO-10} and~\ref{table:MIO+1}, we compare the performance of greedy and our MPC-based attack. On both datasets, each attack strategy achieves a small number of violations $V^\dagger$ in $\mathcal{T}^\dagger$. However, the greedy attack suffers significantly more violations $V^s$ in $\mathcal{T}^s$ than does MPC. Furthermore, these violations are more severe, reflected by the much larger $J_2$ of the greedy attack. As an example, on MIO+1, the greedy attack changes the original green lights in $\mathcal{T}^s$ to red, while our attack only changes green to yellow. The greedy attack also results in larger total effort $J_1$ and objective value $J$. Therefore, we conclude that our attack outperforms the baseline greedy attack overall. In appendix~\ref{appendix: greedy}, we provide more detailed results of the greedy attack.
| 18fc48617d6cbcf7987efb63e98d1f57e5f01032 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The diagnostic process of dementia is challenging and takes a substantial period of time after the first clinical symptoms arise: on average 2.8 years in late-onset and 4.4 years in young-onset dementia \citep{vanvliet2013}. The window of opportunity for advancing the diagnostic process is however much larger than these few years. For Alzheimer's disease (AD), the most common form of dementia, there is increasing evidence that disease processes start 20 years or more ahead of clinical symptoms \citep{Gordon2018}. Advancing the diagnosis is essential to support the development of new disease modifying treatments, since late treatment is expected to be a major factor in the failure of clinical trials \citep{Mehta2017}. In addition, early and accurate diagnosis have great potential to reduce healthcare costs as they give patients access to supportive therapies that help to delay institutionalization \citep{prince2011alzheimer}.
Machine learning offers an approach for automatic classification by learning complex and subtle patterns from high-dimensional data. In AD research, such algorithms have been frequently developed to perform automatic diagnosis and predict the future clinical status at an individual level based on biomarkers. These algorithms aim to facilitate medical decision support by providing a potentially more objective diagnosis than that obtained by conventional clinical criteria \citep{Kloppel2012, Rathore2017}. A large body of research has been published on classification of AD and its prodromal stage, mild cognitive impairment (MCI) \citep{Ansart2021, Wen2020, Rathore2017, Arbabshirani2017, Falahati2014, Bron2015}. Overall, classification methods show high performance for classification of AD patients and control participants with an area under the receiver-operating characteristic curve (AUC) of 85-98\%. Reported performances are somewhat lower for prediction of conversion to AD in patients with MCI (AUC: 62-82\%). Structural T1-weighted (T1w) MRI to quantify neuronal loss is the most commonly used biomarker, whereas the support vector machine (SVM) is the most commonly used classifier. Following the trends and successes in medical image analysis and machine learning, neural network classifiers - convolutional neural networks (CNN) in particular - have increasingly been used since few years \citep{Wen2020, Cui2019, Basaia2019}, but have not been shown to significantly outperform conventional classifiers. Most CNN studies perform no to minimal pre-processing of the structural MRI scans as input for their classifier \citep{Wen2020, Basaia2019, Hosseini-Asl2018, Vieira2017}, while others use more extensive pre-processing strategies proven successful for conventional classifiers, such as gray matter (GM) density maps \citep{Cui2019, Suk2017}. Although CNNs are designed to extract high-level features from raw imaging data, it is imaginable that the learning process for complex tasks is improved by dedicated pre-processing that enhances disease-related features, which reduces model complexity and enables a more stable learning process. It is unclear yet whether CNNs would improve AD classification over conventional classifiers and whether they benefit from extensive MRI pre-processing.
Despite high performance of machine learning diagnosis and prediction methods for AD, it is largely unknown how these algorithms would perform in clinical practice. A next step would be to assess the generalizability of classification methods from a specific research population to another study population. \textcolor{black}{There are however only very few studies assessing classification performance on an external data set \citep{Wen2020,Bouts2019,Archetti2019,Hall2015}. Results varied from only a minor reduction in performance for some experiments \citep{Wen2020, Hall2015} to a severe drop for others \citep{Bouts2019,Archetti2019,Wen2020}.} While generalizability seemed related to how well the training data represented the testing data (e.g. an external data set with similar inclusion criteria showed a smaller performance drop than a data set with very different criteria \citep{Wen2020}, a better understanding is crucial before applying such methods in routine clinical practice.
Therefore, this work aims to assess the generalizability of MRI-based classification performance to an external data set representing a tertiary memory clinic population for both diagnosis of AD and prediction of AD in individuals with MCI. To evaluate the value of neural networks and to determine their optimal MRI pre-processing approach, we compare a CNN with a conventional SVM classifier using two pre-processing approaches: minimal pre-processing using only rough spatial alignment and more extensive pre-processing into modulated GM maps. First, we optimize the methods using a large research cohort and assess classification performance using cross-validation. Subsequently, we validate AD prediction performance in MCI patients of the same cohort as well as AD diagnosis and prediction performance in the external data set.
\begin{table}[bt]
\caption{Demographics for the ADNI data set. \label{tab:adni-demographics}}
\begin{center}
\begin{footnotesize}
\begin{tabular}{ c c c c c }
\toprule
& AD & CN & MCIc & MCInc \\
\cmidrule(lr){2-3}
\cmidrule(lr){4-5}
\emph{\# participants} & 336 & 520 & 231 & 628 \\
\emph{male / female} & 186 / 150 & 252 / 268 & 141 / 90 & 367 / 261 \\
\emph{age (y; mean$\pm$std)} & $74.9\pm7.8$ & $74.2\pm5.8$ & $73.7\pm7.0$ & $72.9\pm7.8$ \\
\bottomrule
\end{tabular}
\end{footnotesize}
\end{center}
\end{table}
\begin{table}[bt]
\caption{Demographics for the PND data set. FU: follow-up time \label{tab:pnd-demographics}}
\begin{center}
\begin{footnotesize}
\begin{tabular}{ c c c c c }
\toprule
& AD & SCD & MCIc & MCInc \\
\cmidrule(lr){2-3}
\cmidrule(lr){4-5}
\emph{\# participants} & 199 & 138 & 48 & 91 \\
\emph{male / female} & 94 / 105 & 92 / 46 & 33 / 15 & 56 / 35 \\
\emph{age (y; mean$\pm$std)} & $73.1\pm9.6$ & $63.2\pm10.3$ & $70.4\pm7.9$ & $68.8\pm12.6$ \\
\emph{FU (y; mean$\pm$std)} & N.A. & N.A. & $2.7 \pm 1.2$ & $2.2 \pm 0.8$\\
\bottomrule
\end{tabular}
\end{footnotesize}
\end{center}
\end{table}
\begin{table*}[!htb]
\caption{An overview of T1-weighted imaging protocols in the PND data set from eight centers. All sequences were 3D and used gradient recalled echo (GRE). TFE: turbo field echo, FSPGR: fast spoiled GRE, TFL: turboflash, MPRAGE: magnetization prepared rapid gradient echo, MP: magnetization prepared, SS: steady state, SP: spoiled, IR: inversion recovery, Sag: sagittal, Cor: Coronal, Ax: axial, TE: echo time, TR: repetition time, TI: inversion time. \label{tab:pnd-imaging}}
\begin{center}
\begin{scriptsize}
\begin{tabular}{ ccccccccccccccc}
\toprule
\emph{Field strength} & \multicolumn{6}{c}{3T} & \multicolumn{5}{c
}{1.5T} \\
\cmidrule(lr){1-1}
\cmidrule(lr){2-7}
\cmidrule(lr){8-12}
\emph{Vendor} & \multicolumn{4}{c}{Philips} & GE & Siemens & Philips & \multicolumn{2}{c}{GE} & \multicolumn{2}{c}{Siemens}\\
\cmidrule(lr){1-1}
\cmidrule(lr){2-5}
\cmidrule(lr){6-6}
\cmidrule(lr){7-7}
\cmidrule(lr){8-8}
\cmidrule(lr){9-10}
\cmidrule(lr){11-12}
\emph{Number of scans} & 195 & 60 & 70 & 1 & 122 & 40 & 1 & 6 & 1 & 32 & 28\\
\emph{Sequence name} & TFE & GRE & GRE & GRE & FSPGR & MPRAGE & GRE & FSPGR & FSPGR & TFL MPRAGE & MPRAGE \\
\emph{Sequence variant} & MP & MP & MP & MP & SS/SP & IR/SP/MP & MP & SS/SP & SS/SP & IR/SP/MP & IR/SP/MP \\
\emph{Plane} & Sag & Sag & Sag & Cor & Sag & Sag & Sag & Sag & Ax & Sag & Sag
\\\
\emph{Slice thickness (mm)} & 1 & 1 & 1 & 1.4 & 1 & 1 & 2 & 1.6 & 1.6 & 1 & 1\\
\emph{In-plane (mm*mm)} & 1*1 & 0.78*0.78 & 0.5*0.5 & 0.88*0.88 & 0.94*0.94 & 1*1 & 0.47*0.47 & 0.47*0.47 & 1*1 & 1*1 & 0.5*0.5\\
\emph{TE (ms)} & 4.2 & 4.6 & 3.5 & 4.6 & 3 & 4.7 & 4.6 & 2.1 & 4.2 & 3 & 3.7 \\
\emph{TR (ms)} & 8.1 & 9.9 & 9 & 9.6 & 7.8 & 2300 & 15 & 7.1 & 8.3 & 2000 & 2700 \\
\emph{TI (ms)} & N.A. & N.A. & N.A. & N.A. & 450 & 1100 & N.A. & 450 & 450 & 1100 & 950 \\
\bottomrule
\end{tabular}
\end{scriptsize}
\end{center}
\end{table*}
\section{Methods}
\subsection{Study population}
We used data from two cohorts. The first group of 1715 participants was included from the Alzheimer's Disease Neuroimaging Initiative (ADNI; \url{adni.loni.usc.edu}). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether clinical and neuropsychological assessment, serial magnetic resonance imaging (MRI), positron emission tomography (PET), and other biological markers can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD). For up-to-date information, see \url{www.adni-info.org}. We included all participants with a T1w MRI scan at baseline from the ADNI1/GO/2 cohorts: 336 AD patients, 520 control participants (CN), 231 mild cognitive impaired (MCI) patients who converted to AD within 3 years (MCIc) and 628 MCI patients who did not convert (MCInc). The CN group consisted of 414 cognitively normal participants and 106 participants with subjective cognitive decline (SCD). Demographics are shown in Table \ref{tab:adni-demographics}. A list of included participants is made available at \url{https://gitlab.com/radiology/neuro/bron-cross-cohort}.
The second group of participants was included from the Health-RI Parelsnoer Neurodegenerative Diseases Biobank (PND; \url{www.health-ri.nl/parelsnoer}), a collaborative biobanking initiative of the eight university medical centers in the Netherlands \citep{Mannien2017}. The Parelsnoer Neurodegenerative Diseases Biobank focuses on the role of biomarkers on diagnosis and the course of neurodegenerative diseases, in particular of Alzheimer’s disease \citep{Aalten2014}. It is a prospective, multi-center cohort study, focusing on tertiary memory clinic patients with cognitive problems including dementia. Patients are enrolled from March 2009 and followed annually for two to five years. In the PND biobank, a total of 1026 participants have been included. Inclusion criteria for the current research were: a high resolution T1w MRI at baseline, clinical consult at baseline, 90 days or less between MRI and clinical consult, and a baseline diagnosis of SCD, MCI, or dementia due to AD. A flow diagram of the inclusion can be found in the supplementary files (Fig. \ref{fig:pnd-inclusion}). A total of 557 participants met inclusion criteria. One person was excluded because image analysis failed. This led to inclusion of 199 AD patients and 138 participants with SCD. Of the MCI group, we included the 139 participants that had a follow-up period of at least 6 months. Of this group, 48 MCI patients converted towards dementia within the available follow-up time and 91 MCI patients remained stable. Demographics are shown in Table \ref{tab:pnd-demographics}.
\begin{figure*}[!htb]
\centering
\includegraphics[width=1\linewidth]{CNN.png}
\caption[cnn]{CNN architecture\footnotemark}
\label{fig:cnn}
\end{figure*}
\subsection{Imaging data}
We used baseline T1w structural MRI acquired at 1.5T or 3T. Acquisition protocols are previously described (ADNI: \citet{Jack2008, Jack2015}, PND: \citet{Aalten2014}). Variation in acquisition protocols used in PND is detailed in Table \ref{tab:pnd-imaging}. For the majority of scans, a 8-channel head coil was used (N=423; 76\%); other scans used a 16-channel (N=27), 24-channel (N=1), 40-channel (N=1), or unknown coil (N=104).
\subsection{Image pre-processing}
We evaluated two pre-processing approaches based on T1w images: minimal pre-preprocessing and a more extensive pre-processing into modulated GM maps.
To prepare T1w images with minimal pre-processing, scans were non-uniformity corrected using the N4 algorithm \citep{tustison2010n4itk} and subsequently transformed to MNI-space using registration of brain masks with a similarity transformation. A similarity transformation is a rigid transformation including isotropic scaling. Registrations were performed with Elastix registration software \citep{klein2010elastix, Shamonin2014}. To account for variations in signal intensity, images were normalized within the brain mask to have zero mean and unit variance.
To obtain modulated GM maps encoding gray matter density, the Iris pipeline was used \citep{Bron2014HBM}. To compute these maps a group template space was defined using a procedure that avoids bias towards any of the individual T1w images using pairwise registration \citep{seghers2004construction}. The pairwise registrations were performed using a similarity, affine, and nonrigid B-spline transformation model consecutively. We selected a subset of images for the definition of the template space. This template set consisted of the images of 50 ADNI participants that were randomly selected preserving the ratio between diagnostic groups (subject list available at \url{https://gitlab.com/radiology/neuro/bron-cross-cohort}). The other images of both ADNI and PND data sets were registered to the template space following the same registration procedure. For the current work, some changes to the template space construction procedure as used in \cite{Bron2014HBM} were made: non-uniformity correction was performed, skull-stripping was performed, and the template space corresponded to MNI-space. Using similarity registration based on brain masks, we computed the coordinate transformations of MNI space to each of the template set's images, which were subsequently concatenated with the pairwise transformations before averaging. After template space construction, probabilistic GM maps were obtained with the unified tissue segmentation method of SPM8 (Statistical Parametric Mapping) \citep{ashburner2005tissueseg}. To obtain the final feature maps, probabilistic GM maps were transformed to the template space and modulated, i.e. multiplied by the Jacobian determinant of the deformation field, to take compression and expansion into account \citep{ashburner2000voxel}. To correct for head size, modulated GM maps were divided by intracranial volume.
\subsection{Classification approaches}
Two machine learning approaches were used for classification: a support vector machine (SVM) and a convolutional neural network (CNN).
\subsubsection{Support vector machine (SVM)}
An SVM with a linear kernel was used as this approach previously showed good performance using voxel-based features for AD classification. \citep{kloppel2008automatic, cuingnet2011automatic, Bron2014HBM, Bron2015}. The c-parameter was optimized with 5-fold cross-validation on the training set. Input features, i.e. voxel values of the pre-processed images within a brain mask, were normalized to zero mean and unit variance based on the training set. The classifier was implemented using Scikit-Learn.
To gain insight into the classifications, we calculated statistical significance maps (p-maps) that show which features contributed to the SVM decision. These maps were computed using an analytical expression that approximates permutation testing \citep{Gaonkar2015}. Clusters of significant voxels were obtained using a p-value threshold of $\alpha\leq 0.05$. P-maps were not corrected for multiple comparisons, as permutation testing has a low false-positive rate \citep{gaonkar2013analytic}.
\subsubsection{Convolutional neural network (CNN)}
\footnotetext{Figure created with \url{https://alexlenail.me/NN-SVG}} \textcolor{black}{An all convolutional neural network was used \citep{Springenberg2015}, which is a fully convolutional network (FCN) architecture that uses standard convolutional layers with stride two instead of the pooling layers used in most CNNs.} This approach was chosen as it has previously shown good classification performance for AD based on structural MRI \citep{Cui2019, Basaia2019}. The used architecture is shown in Fig. \ref{fig:cnn}. Specifically, the network was built of 7 blocks consisting of a 3D convolutional layer (filter size 3; stride 1), followed by dropout, batch normalization (BN), and a rectified linear unit (ReLU) activation function, succeeded by a second 3D convolutional layer (filter size 3; stride 2), dropout, BN, and ReLU activation \citep{Cui2019, Basaia2019}\textcolor{black}. The number of filters changed over blocks: 16 filters in block 1, 32 in block 2 and 3, 64 in block 4 and 5, 32 in block 6, and 16 in block 7. The final output layer of the network was a softmax activation function, providing 2 prediction values (1 per class). The total network consisted of 577,498 parameters.
\textcolor{black}{For artificially increasing the training data set and for removing the class imbalance, data augmentation was used. The training set was augmented to 1000 samples per class based on the `mixup' approach \citep{Eatonrosen2018, Zhang2018}. Mixup is a data-agnostic augmentation approach that is not based on spatial transformations, and therefore does not degrade the spatial normalization. Augmented samples were constructed by linearly combining two randomly selected images of the same class: a fraction of $80\%$ of the first image was added to a fraction of $20\%$ of the second image.}
The network was compiled with a binary cross-entropy loss function and Adam optimizer (learning rate=0.001, epsilon=1e-8, decay=0.0). To facilitate a stable convergence, learning rate followed a step decay schedule, i.e. after each ten epochs the learning rate was divided by two. The dropout rate was set to $20\%$. Data was propagated through the network with a batch size of 4. Input images were normalized to zero mean and unit variance based on the augmented training set. \textcolor{black}{A validation set was created by randomly splitting $10\%$ of the training data which was not used for training but only for regularization by early stopping, i.e. training was stopped when the validation AUC had not increased for 20 epochs.} The model of the epoch with the highest validation AUC was selected as final model. Implementation was based on Keras and Tensorflow.
To gain insight into the classifications, we made saliency maps that show which parts of the brain contributed the most to the prediction of the CNN, i.e. which voxels lead to increase/decrease of prediction score when changed. Saliency maps were made using guided backpropagation, changing the activation function of the output layers from softmax to linear activations \citep{Springenberg2015}. Maps were averaged over correctly classified AD patients \citep{Rieke2018}.
\begin{figure*}[!htb]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1\linewidth]{Plot1_ADCN_CNNSVM_CI_03-Nov-2020_08-55.png}
\caption{}
\label{fig:adni-adcn:a}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1\linewidth]{Plot1_ADCN_CNNSVM_Acc_CI_27-Oct-2020_15-10.png}
\caption{}
\label{fig:adni-adcn:b}
\end{subfigure}
\caption[adni-adcn]{Cross-validation performance for classification of the Alzheimer's disease patients (AD) and controls (CN) of the ADNI data set expressed by (a) area under-the-ROC-curve (AUC) and (b) accuracy. Performance is shown for SVM and CNN classifiers using two inputs: minimally processed T1w scans and modulated GM images. Error bars indicate $95\%$CIs.}
\label{fig:adni-adcn}
\end{figure*}
\begin{figure*}[!htb]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1\linewidth]{Plot2_MCI_CNNSVM_27-Nov-2020_07-38.png}
\caption{}
\label{fig:adni-mci:a}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1\linewidth]{Plot2_MCI_CNNSVM_Acc_03-Nov-2020_14-38.png}
\caption{}
\label{fig:adni-mci:b}
\end{subfigure}
\caption[adni-mci]{Classification performance of patients with mild cognitive impairment (MCI) that do or do not convert to Alzheimer's disease (MCIc vs MCInc) in the ADNI data set expressed by (a) area under-the-ROC-curve (AUC) and (b) accuracy. Performance is shown for SVM and CNN classifiers using two inputs: minimally processed T1w scans and modulated GM images. Error bars indicate $95\%$CIs. P-values for significant differences are shown in (b).}
\label{fig:adni-mci}
\end{figure*}
\begin{figure*}[!htb]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1\linewidth]{Plot3_PSI_CNNSVM_27-Nov-2020_16-25.png}
\caption{}
\label{fig:pnd:a}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1\linewidth]{Plot3_PSI_CNNSVM_Acc_27-Nov-2020_16-26.png}
\caption{}
\label{fig:pnd:b}
\end{subfigure}
\caption[pnd]{Classification performance on the PND data set: (a) area under-the-ROC-curve (AUC) and (b) accuracy. Classifiers were trained on ADNI AD-CN and applied to PND AD-SCD (left figures) and PND MCIc-MCInc (right figures). Performance is shown for SVM and CNN classifiers using two inputs: minimally processed T1w scans and modulated GM images. Error bars indicate $95\%$CIs. P-values for significant differences are shown in (b).}
\label{fig:pnd}
\end{figure*}
\subsection{Analysis and statistics}
Classification performance was quantified by the area under the curve (AUC) and accuracy. For AD-CN classification, the data of the ADNI AD and CN groups were randomly split for $20$ iterations preserving relative class sizes in each training and testing sample, using $90\%$ for training and $10\%$ for testing. Random splits were the same for both SVM and CNN. In each iteration, classification model parameters were optimized on the training set as explained above. The models were optimized solely on the training set; the test set was used only for evaluation of the final model. Ninety-five percent confidence intervals ($95\%$CI) for the mean performance measures were constructed using the corrected resampled t-test based on the 20 cross-validation iterations, thereby taking into account that the samples in the cross-validation splits were not statistically independent \citep{Nadeau}
Subsequently, we retrained classifiers using all AD and CN participants from the ADNI as training set. These retrained classifiers were used for visualization and their performance was evaluated on three independent test sets: ADNI MCIc-MCInc, PND AD-SCD, and PND MCIc-MCInc. $95\%$CIs were obtained based on 500 bootstrap samples of the test set. Significant differences between classifiers were assessed using the non-parametric McNemar Chi-square test \citep{Dietterich1998} ($\alpha<0.013$ after Bonferroni correction for 4 comparisons in each test set).
Trained models, lists of included subjects and all code used in preparation of this article are available from \url{https://gitlab.com/radiology/neuro/bron-cross-cohort}\footnote{When this manuscript is accepted for publication, we will make a release of the repository and add a permanent DOI here}.
\section{Results}
Cross-validation performance for the ADNI AD-CN classification is shown in Fig. \ref{fig:adni-adcn}. For SVM, the AUC using modulated GM maps ($0.940$, $95\%$CI: $0.924-0.955$) was higher than the AUC using T1w images ($0.801$, $95\%$CI: $0.765-0.837$). For CNN, the same effect was observed, with modulated GM maps yielding a higher AUC ($0.933$, $95\%$CI: $0.918-0.948$) than T1w images ($0.898$, $95\%$CI: $0.875-0.920$), albeit here with overlapping confidence intervals. For classification based on modulated GM maps, the AUC for SVM ($0.940$; $95\%$CI: $0.924-0.955$) was similar to that of CNN ($0.933$; $95\%$CI: $0.918-0.948$). Accuracy measures showed the same patterns.
The performance of the classifiers trained on all ADNI AD and CN data to predict MCI conversion is shown in Fig. \ref{fig:adni-mci}. While AUCs with both SVM and CNN were slightly higher for modulated GM maps than for T1w images, the accuracy measures showed similar performance for both inputs. Using modulated GM maps, performance for SVM (AUC=0.756, $95\%$CI: $0.720-0.788$; accuracy=0.695, $95\%$CI: $0.665-0.723$) was higher than for CNN (AUC=0.742, $95\%$CI: $0.709-0.776$); accuracy=0.658, $95\%$CI: $0.628-0.690$). This difference was significant according to McNemar’s test ($p<0.01$).
The performance of external validation, i.e. the application of the classifiers in the PND data set, is shown in Fig. \ref{fig:pnd}. For AD-SCD diagnosis, the AUC for SVM was $0.896$ ($95\%$CI: $0.855-0.932$) and that for CNN was $0.876$ ($95\%$CI: $0.836-0.913$). Both AUC and accuracy followed the same patterns as in ADNI: SVM and CNN showed similar performance and modulated GM maps yielded higher classification performance than minimally processed T1w images (McNemar's test; $p<0.01$ for SVM, $p=0.01$ for CNN). Performances were however slightly lower; PND confidence intervals for AUC (but not for accuracy) overlapped with those of ADNI.
For prediction of MCI conversion in PND, classification performance was also lower than that in ADNI. For the GM modulated maps, the AUC for CNN was 0.702 ($95\%$CI: $0.624-0.786$) and that for SVM was 0.665 ($95\%$CI: $0.576-0.760$). Confidence intervals were relatively large and overlapped with those in the ADNI data. No significant differences between classifiers and between pre-processing approaches were seen.
Brains regions that contributed to the classifications are visualized using SVM p-maps in Fig. \ref{fig:pmaps} and using CNN saliency maps in Fig. \ref{fig:saliency}. The SVM p-map for the minimally processed T1w images showed small clusters of significant voxels, mainly located in the medial temporal lobe (hippocampus), around the ventricles and at larger sulci at the outside of the brain. For modulated GM maps, clusters of significant voxels in the p-map were larger and predominantly visible in the hippocampus. In addition, smaller clusters were located in the rest of the temporal lobe and the cerebellum. CNN saliency maps showed a very limited contribution of the temporal lobe. Instead, the saliency map for the T1w images mainly showed contribution of voxels at the edge of the brain, in white matter regions around the ventricles and in the cerebellum. For modulated GM maps, clusters of contributing voxels were located in the subcortical structures, the white matter around the ventricles and the cerebellum.
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.8\linewidth]{pmap.pdf}
\caption[pmap]{Visualization of the SVM classifiers using analytic significance maps (p-maps) based on two inputs: (a) minimally processed T1w images and (b) modulated GM maps.}
\label{fig:pmaps}
\end{figure*}
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.8\linewidth]{saliency.pdf}
\caption[saliency]{Visualization of the CNN classifiers using guided back-propagation saliency maps based on two inputs: (a) minimally processed T1w images and (b) modulated GM maps. Relevance maps were averaged over all correctly classified AD participants and thresholded at $\frac{1}{3}$ of the maximum intensity.}
\label{fig:saliency}
\end{figure*}
\section{Discussion}
We performed a comparative study focusing on the generalizability of diagnostic and predictive performance of machine learning based on MRI data of the ADNI research cohort, to the PND multi-center data set representing a tertiary memory clinic population. Both cross-validation and external validation results for AD-CN diagnosis showed similar performance using the used deep learning classifier and conventional classifier. Both approaches significantly benefited from the use of modulated GM maps instead of raw T1w images. Application to MCI conversion prediction yielded higher performance for SVM than for CNN in ADNI, but this was not seen in PND. Performances were in line with the state-of-the-art \citep{Rathore2017, Wen2020,Ansart2021}. \textcolor{black}{For MCI conversion prediction, \cite{Ansart2021} showed that performance of current methods converges to an AUC of about 75\% as the number of subjects increases, which aligns with our results.}
While in many medical imaging applications CNNs convincingly outperformed conventional classifiers \citep{Litjens2017}, our results showed similar performance for CNN and SVM, which confirms the findings by \cite{Wen2020}. \textcolor{black}{Other CNN designs could possibly improve on this, but we made an effort to follow the state-of-the-art for CNN design. Promising developments to further improve performance could come from changes in network architecture (e.g., successful standard architectures like InceptionNet or ResNet, adversarial training, discriminative auto-encoders) and improvements in data collection and handling (e.g., larger datasets to learn more complex models, or pretraining on other collections of brain imaging data).} \textcolor{red}{In addition, data augmentation could play a role in further improvement. While a strength of the mix-up approach is that it is data-agnostic, an augmentation approach using for example prior knowledge may have added value.}
\textcolor{black}{This work shows that the need for dedicated pre-processing is lower for CNN than for SVM, but nevertheless has an added value for the performance. While we evaluated only one implementation of the pre-processing procedure \citep{Bron2014HBM}, we expect that alternative implementations (e.g. SPM12, FSL-VBM) could have slightly changed results but would have led to the same conclusions.} \textcolor{red}{With sufficiently large datasets the need for dedicated pre-processing including spatial normalization may reduce.}
Although SVM and CNN classifiers yielded similar performance, their visualizations showed different brain regions to be involved in the classification. SVM significance maps showed a clear contribution of the hippocampus and medial temporal lobe as previously shown and expected based on prior knowledge \citep{Bron2017eurrad}. CNN saliency maps showed involvement of subcortical structures, regions prone to white matter hyperintensities and the cerebellum. For both classifiers, classification based on minimally processed T1w images showed voxels at the edge of the brain to be involved, which is expected as only similarity transformation to template space had been performed. \textcolor{black}{In addition to the brain edges, the CNN classifier, which outperformed the SVM for these minimally processed input images, also highlights regions similar to those shown by the saliency map for the modulated GM images. This may implicate that the CNNs non-linear operations, in contrast to the linear kernel of the SVM, could extract feature maps that partly resemble GM modulated maps.} The regions highlighted by the CNN saliency maps could possibly be related to AD using prior knowledge, but we will refrain from over-interpretation here. It is however unexpected that the medial temporal lobe is not covered as previously shown with CNN saliency maps on ADNI data \citep{Dyrba2020, Rieke2018}. Differences between the SVM and CNN classifiers in involved brain regions could be contributed to both the differences in the classification approaches as well as to the differences in the used visualization techniques. \textcolor{black}{If the first reason dominates, hence if the classifiers actually use different brain regions, combining classifiers into a hybrid approach would be an interesting future direction. However, for} full understanding of brain regions involved in CNN-based classification of AD, further research is required.
This work is one of the few to address how AD classification performance of MRI-based machine learning generalizes to an independent cohort \citep{Wen2020, Hall2015, Bouts2019, Archetti2019}. On the PND data, the resulting AUC values (0.896 for SVM, 0.876 for CNN) were competitive with values reported for AD-CN in the literature, but still they were 0.04-0.07 lower than those in the ADNI cross-validation experiment. The main patterns in the results corresponded between ADNI and PND data, i.e. similar performance for SVM and CNN and added value of dedicated MRI processing. For prediction in MCI, AUC values in the PND data set were 0.04-0.10 lower than those in ADNI. Overall, similar to experiments by \citet{Wen2020} and \citet{Hall2015}, we observed only a minor performance drop. This largely preserved performance could be related to the similarities between the ADNI and PND studies that include a multi-center set-up, within-study standardization of cognitive protocols, and diagnostic criteria for AD \citep{McKhann1984, McKhann2011} and MCI \citep{Petersen2004}. The performance reduction could be contributed to differences between the studies, such as the MRI protocols (all high resolution T1w, but more homogeneous within ADNI than within PND), country of origin (United States vs. the Netherlands), control population (a combination of cognitively normal and SCD vs. SCD only), MCI population (amnestic MCI only vs. a broad MCI group) and patient inclusion criteria (ADNI used hard cut-offs on cognitive scores and clinical dementia rating whereas PND did not) \citep{Petersen2010, Aalten2014}. Studies that found much worse generalizability in their experiments described larger differences in inclusion and diagnostic criteria between training data and validation data than we did \citep{Bouts2019, Wen2020}.
A limitation of this study is that the diagnosis was based on clinical criteria rather than post-mortem histopathological examination. Although diagnosis was typically confirmed by follow-up, it is possible that some of the patients were misdiagnosed. \textcolor{black}{An alternative could be to use amyloid data from PET imaging or cerebrospinal fluid to classify AD pathology instead of relying on the clinical diagnosis (e.g., \cite{Son2020})}. In addition, because of the limited availability of diagnostic information at follow-up in the PND data set, its MCI data is relatively small. This is reflected by the large confidence intervals for the performance metrics in the prediction task. To maximize the number of PND MCI participants, we chose to use the last available time point for final diagnosis. As a result the time-to-prediction ranged between 1-5 years, whereas for ADNI a fixed time interval of three years is chosen. As time-to-prediction is related to predictive performance \citep{Ansart2021}, a fixed time interval would be preferred for inter-cohort performance comparison.
While the external validation performance was quite high, as expected some performance drop was observed. Therefore, research focusing on approaches to mitigate such performance drops, such as transfer learning, is highly relevant \citep{Wachinger2016}. In addition, whereas this work only exploited structural MRI, other works have shown that performance can be increased with the use of multi-modal inputs, i.e. cognitive test scores, fluid-based biomarker measurements, genetic information and other imaging modalities such as PET, diffusion MRI or perfusion MRI \citep{Bron2017eurrad, Ansart2021, Venkatraghavan2019}. While multi-modal classification would therefore be a logical and important extension, this may also lead to a decrease of generalizability as chances of differences between studies increase with multiple modalities.
In conclusion, classification performance of ADNI data generalized well to the multi-center PND biobank cohort representing tertiary memory clinic patients, with only a minor drop in performance. Conventional SVM classifiers and deep learning approaches using CNN showed comparable results, and both methods benefited from dedicated MRI processing using GM modulated maps. We hope that external validation results like those presented here will contribute to setting next steps towards the implementation of machine learning in clinical practice for aiding diagnosis and prediction.
\section*{Acknowledgements}
The authors would like to thank Judith Manniën, Ilya de Groot, and Nienke Aaftink for their effort in data preparation.
The authors are grateful to SURFsara for the processing time on the Dutch national supercomputer (\url{www.surfsara.nl/systems/cartesius}). We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan V GPU used for this research.
E.E. Bron acknowledges support from Dutch Heart Foundation (PPP Allowance, 2018B011) and the Netherlands CardioVascular Research Initiative (Heart-Brain Connection: CVON2012-06, CVON2018-28). E.E. Bron and W.J. Niessen are supported by Medical Delta Diagnostics 3.0: Dementia and Stroke. V. Venkatraghavan and W.J. Niessen acknowledge funding from the Health~Holland LSH-TKI project Beyond (LSHM18049). This work is part of the EuroPOND initiative, which is funded by the European Union's Horizon 2020 research and innovation programme under grant agreement No. 666992.
The work described in this study was carried out in the context of the Health-RI Parelsnoer Neurodegenerative Diseases Biobank. Parelsnoer biobanks are part of and funded by the Dutch Federation of University Medical Centers and has received initial funding from the Dutch Government (2007-2011).
Data collection and sharing was funded by the Alzheimer's Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen
Alzheimer Immunotherapy Research \& Development, LLC.; Johnson \& Johnson
Pharmaceutical Research \& Development LLC.; Lumosity; Lundbeck; Merck \& Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.
\bibliographystyle{model2-names}\biboptions{authoryear}
| 2795dfcb909611f9a6f9313103a640dc2162baf0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Acknowledgment}
We thank our shepherd Lorenzo Cavallaro and the anonymous reviewers for their constructive and valuable feedback.
This work is sponsored in part by NSF grants CCF-18-45893, CCF-18-22965, CCF-16-19123, CNS-18-42456, CNS-18-01426, CNS-16-18771, CNS-16-17670, CNS-15-64055, and CNS-15-63843; ONR grants N00014-17-1-2010, N00014-16-1-2263, and N00014-17-1-2788; an NSF CAREER award; an ARL Young Investigator (YIP) award; a Google Faculty Fellowship; a JP Morgan Faculty Research Award; a DiDi Faculty Research Award; a Google Cloud grant; a Capital One Research Grant; and an Amazon Web Services grant.
Any opinions, findings, conclusions, or recommendations expressed herein are those of the authors, and do not necessarily reflect those of the US Government, ONR, ARL, NSF, Captital One, Google, JP Morgan, DiDi, or Amazon.
\subsection{Detailed Model Architecture}
\label{subsec:detailed_arch}
\vspace{.1cm}\noindent\textbf{Multi-head self-attention.}
Given the embeddings of all input token (see Section~\ref{subsec:pretrain_method}) in $l$-th layer, $(E_{l,1},...,E_{l,n})$, the self-attention layer updates each embedding with the following steps.
It first maps $E_{l,i}$ to three embeddings, \emph{i.e., } query embedding $q_i$, key embedding $k_i$, and value embedding $v_i$:
\begin{equation*}
\label{eq:kqv}
\centering
q_i=f_{q}(W_{q};E_{l,i}); \ k_i=f_{k}(W_{k};E_{l,i}); \ v_i=f_{v}(W_{v};E_{l,i})
\end{equation*}
Here $f_{q}$, $f_{k}$, and $f_{v}$ are affine transformation functions (\emph{i.e., } a fully-connected layer) parameterized by $W_{q}$, $W_{k}$, and $W_{v}$, respectively.
It then computes attention $a_{ij}$ between $E_{l,i}$ and all other embeddings $E_{l,\{j|j\neq i\}}$ by taking the dot product between the $E_{l,i}$'s query embedding $q_i$ and $E_{l,j}$'s key embedding $k_j$: $a_{ij}=q_i\cdot k_j$.
Intuitively, attention $a$ is a square matrix, where each cell $a_{ij}$ indicates how much attention $E_{l,i}$ should pay to $E_{l,j}$ when updating itself.
It then divides every row of $a$ by $\sqrt{d_{emb}}$ (the dimension of the embedding vectors) and scale it by softmax to ensure them sum up to 1:
\begin{equation*}
\label{eq:scale}
\centering
a'_{ij} = \frac{\exp(a_{ij})}{\sum^n_{j=1}\exp(a_{ij})}
\end{equation*}
The scaled attention $a'_{ij}$ will be multiplied with value embedding $v_j$ and summed up:
\begin{equation*}
\label{eq:single_head}
\centering
E^h_{k+1,i} = \sum^n_{j=1} a'_{ij} v_j
\end{equation*}
Here $h$ in $E^h_{k+1,i}$ denote the updated embedding belong to attention head $h$. Assume we have total $H$ attention heads, the updated embeddings will finally go through an 2-layer feedforward network $f_{out}$ parameterized by $W_{out}$ with skip connections~\cite{he2016deep} to update embeddings from all heads:
\begin{equation}
\label{eq:multi_head}
\centering
E_{l+1,i} = f_{out}(concat(E^0_{l+1,i},...,E^H_{l+1,i});W_{out})
\end{equation}
\vspace{.1cm}\noindent\textbf{How to use the embedding.}
The learned embeddings $E_{l,i}$ after the last self-attention layer encodes the execution semantics of each instruction and the overall function.
Consider predicting masked input in pretraining.
Let layer $l$ be the $g_p$'s last self-attention layer.
The model will stack a 2-layer multi-layer perceptron (MLP) for predicting the masked codes and values:
\begin{equation*}
MLP(E_{l,i}) = softmax(tanh(E_{l,i}\cdot W_1)\cdot W_2), i\in\mathbb{MP}
\end{equation*}
$W_1\in \mathbb{R}^{d_{emb}\times d_{emb}}$ and $W_2\in \mathbb{R}^{d_{emb}\times |V|}$ where $|V|$ is the vocabulary size of the code token or bytes.
As shown in Equation~\ref{eq:pretrain}, each embedding will have 9 stacked MLPs (\emph{i.e., } one for predicting code and the rest for predicting bytes).
\subsection{\textsc{Trex}\xspace Hyperparameters}
\label{subsec:hyperparm}
\vspace{.1cm}\noindent\textbf{Network architecture.} We use 12 self-attention layers with each having 8 self-attention heads.
The embedding dimension is $d_{emb}=d_{func}=768$, which is also the embedding dimension used in bi-LSTM.
We set 3072 as the hidden layer size of MLP in the self-attention layer.
We adopt GeLU~\cite{hendrycks2016gaussian}, known for addressing the problem of vanishing gradient, as the activation function for \textsc{Trex}\xspace's self-attention module. We use the hyperbolic tangent (tanh) as the activation function in finetuning MLP.
We set the dropout rate 0.1 for pretraining do not use dropout in finetuning.
\vspace{.1cm}\noindent\textbf{Pretraining.}
We fix the largest input length to 512 and choose the effective batch size for both pretraining and finetuning as 512.
As 512 batch size with each sample of length 512 is too large to fit in our GPU memory (11 GB), we aggregate the gradient every 64 batches and then updates the weight parameter.
This setting results in the actual batch size as 8 ($64\times8=512$).
We pick learning rate $5\times10^{-4}$ for pretraining.
Instead of starting with the chosen learning rate at first epoch, we follow the common practice of using small warmup learning rate at first epoch.
We use $10^{-7}$ as the initial warmup learning rate and gradually increases it until reaching the actual learning rate ($5\times10^{-4}$) after first epoch.
We use use Adam optimizer, with $\beta_1=0.9$, $\beta_2=0.98$, $\epsilon=10^{-6}$, and weight decay $10^{-2}$.
\subsection{More Experiments}
\label{subsec:more_exp}
\vspace{.1cm}\noindent\textbf{Cross-project generalizability.}
While we strictly separate the functions for pretraining, finetuning, and testing, the functions of finetuning and testing can come from the same project (but strictly different functions).
Therefore, in this section, we further separate the functions in finetuning and testing by extracting them from \emph{different projects}.
For example, we can finetune the model on Coreutils while testing on OpenSSL.
Specifically, we select 3 projects, \emph{i.e., } Binutils, Coreutils, and OpenSSL, which have the largest number of functions, and evaluate how \textsc{Trex}\xspace performs.
We allow the functions to come from different architectures, optimizations, and obfuscations, and follow the same setup as described in Section~\ref{sec:impl}.
\begin{table}[!t]
\footnotesize
\setlength{\tabcolsep}{7pt}
\centering
\renewcommand{\arraystretch}{1.1}
\setlength\aboverulesep{0.4pt}
\setlength\belowrulesep{0.4pt}
\caption{\textsc{Trex}\xspace results (in AUC score) on function pairs with training and testing functions extracted from different projects.
}
\label{tab:cross-dataset}
\begin{tabular}{r|c|c|c}
\toprule[1.1pt]
\backslashbox{Train}{Test} & \textbf{Coreutils} & \textbf{Binutils} & \textbf{OpenSSL} \\ \midrule[.9pt]
\rowcolor{lightblue} \textbf{Coreutils} & 0.947 & 0.945 & 0.940 \\ \hline
\textbf{Binutils} & 0.945 & 0.945 & 0.944 \\ \hline
\rowcolor{lightblue} \textbf{OpenSSL} & 0.936 & 0.939 & 0.956 \\ \bottomrule[1.1pt]
\end{tabular}
\end{table}
Table~\ref{tab:cross-dataset} shows that \textsc{Trex}\xspace's AUC score does not drop dramatically ($<2$\%) when the functions are coming from different projects when compared to coming from same projects (the diagonal).
This observation indicates \textsc{Trex}\xspace generalizes to unseen function pairs in an entirely different dataset.
Note that the functions in each function pair can come from arbitrary architectures, optimizations, and obfuscations (last column in Table~\ref{tab:result}).
As we have shown in Section~\ref{subsec:rq2}, the numbers achieved by \textsc{Trex}\xspace in Table~\ref{tab:cross-dataset} even outperforms the existing baselines when their functions can only come from different architectures and within only the same projects.
\vspace{.1cm}\noindent\textbf{Effectiveness of bi-LSTM encoding.}
As described in Section~\ref{subsec:input_repr}, we treat the numeric values as an 8-byte sequence and use bi-LSTM to combine them into a single representation (embedding).
The structure of bi-LSTM is known to capture the potential dependencies between different bytes (with different significance) in the byte-sequence.
Indeed, besides bi-LSTM, there are other possible differentiable modules such as multi-layer perceptron (MLP) or simple summation can also combine the input bytes.
Therefore, we study the performance of \textsc{Trex}\xspace in predicting masked tokens when we vary the modules for byte-sequence combination.
Specifically, we follow the same setting described above by selecting a random 10,000 function binaries as the testing set and evaluate the testing PPL achieved by pretraining \textsc{Trex}\xspace.
Figure~\ref{fig:pretrain-loss} shows the validation PPL in 10 epochs of pretraining \textsc{Trex}\xspace based on (1) bi-LSTM, (2) 2-layer (with 1024 hidden size) MLP, and (3) simple summation (SUM), to combine the byte-sequence.
We can observe that bi-LSTM is obviously better than other two, achieving the lowest PPL.
This indicates that bi-LSTM helps \textsc{Trex}\xspace the most in terms of learning approximate execution semantics.
\vspace{.1cm}\noindent\textbf{Vulnerability search performance.}
We quantify the accuracy of \textsc{Trex}\xspace in searching vulnerable functions in the firmware images and compare it to that of SAFE.
As SAFE does not work for MIPS, we study how it performs on NETGEAR R7000 model, the only model that runs on ARM architecture from Table~\ref{tab:cve_case}.
Specifically, we compile OpenSSL to ARM and x64 with \texttt{O3}, and feed both our compiled and firmware's function binaries to \textsc{Trex}\xspace and SAFE to compute embeddings.
Based on the function embeddings, we search the compiled OpenSSL functions in the NETGEAR R7000's embedded OpenSSL libraries, and test their top-1/3/5/10 errors.
For example, the top-10 error measures when the query function does not appear in the top-10 most similar functions in the firmware.
\begin{figure}[!t]
\centering
\subfloat[Same architecture]{
\includegraphics[width=0.47\linewidth]{./figs/case/vuln_error_same.pdf}
\label{subfig:same}}
\subfloat[Cross architecture]{
\includegraphics[width=0.47\linewidth]{./figs/case/vuln_error_cross.pdf}
\label{subfig:cross}}
\caption{Top-1/3/5/10 error of \textsc{Trex}\xspace and SAFE in searching functions in firmware. (a) The query functions and the firmware are from same architecture (ARM). (b) The query functions are from x64 but the firmware are from ARM.}
\label{fig:vuln_error}
\end{figure}
Figure~\ref{fig:vuln_error} shows that \textsc{Trex}\xspace consistently outperforms SAFE, achieving 24.3\% lower error rate on average.
Moreover, as shown in Figure~\ref{subfig:cross}, when the query functions are from x64 (the firmware binaries are from ARM), \textsc{Trex}\xspace outperforms SAFE by a greater margin, achieving 25.3\% lower error rate in cross-architecture search.
\begin{table*}[!t]
\footnotesize
\setlength{\tabcolsep}{1.5pt}
\centering
\renewcommand{\arraystretch}{1}
\setlength\aboverulesep{0.4pt}
\setlength\belowrulesep{0.4pt}
\caption{Details of the real-world 180 firmware images we collected from DD-WRT.
}
\label{tab:firmware}
\begin{tabular}{l|l|l|l?{1.1pt}l|l|l|l?{1.1pt}l|l|l|l}
\toprule[1.1pt]
\textbf{Vendor} & \textbf{Model} & \textbf{ISA} & \textbf{Year} & \textbf{Vendor} & \textbf{Model} & \textbf{ISA} & \textbf{Year} & \textbf{Vendor} & \textbf{Model} & \textbf{ISA} & \textbf{Year} \\ \midrule[.9pt]
\rowcolor{lightblue} 8devices &Carambola 2 &MIPS &2020 &8devices &Carambola 2 8MB &MIPS &2018 &8devices &Lima &MIPS &2020 \\ \hline
Actiontec &MI424WR &ARM &2019 &Alfa &AIP-W502U &MIPS &2020 &Alfa &SOLO48 &MIPS &2020 \\ \hline
\rowcolor{lightblue} Belkin &F5D8235-4 &MIPS &2020 &Buffalo &BHR-4GRV &MIPS &2020 &Buffalo &WBMR-HP-G300H &MIPS &2020 \\ \hline
Buffalo &WHR-1166D &MIPS &2020 &Buffalo &WHR-300HP2 &MIPS &2018 &Buffalo &WHR-600D &MIPS &2018 \\ \hline
\rowcolor{lightblue} Buffalo &WXR-1900DHP &ARM &2020 &Buffalo &WZR-1166DHP &ARM &2020 &Buffalo &WZR-600DHP2 &ARM &2020 \\ \hline
Buffalo &WZR-900DHP &ARM &2020 &Buffalo &WZR-HP-G450H &MIPS &2020 &Comfast &CF-E325N &MIPS &2020 \\ \hline
\rowcolor{lightblue} Comfast &CF-E355AC &MIPS &2020 &Comfast &CF-E380AC &MIPS &2020 &Comfast &CF-WR650AC &MIPS &2020 \\ \hline
Compex &WP546 &MIPS &2020 &Compex &WPE72 &MIPS &2020 &D-Link &DAP-2230 &MIPS &2020 \\ \hline
\rowcolor{lightblue} D-Link &DAP-2330 &MIPS &2020 &D-Link &DAP-2660 &MIPS &2020 &D-Link &DAP-3320 &MIPS &2020 \\ \hline
D-Link &DAP-3662 &MIPS &2020 &D-Link &DHP-1565 A1 &MIPS &2020 &D-Link &DIR-825 C1 &MIPS &2020 \\ \hline
\rowcolor{lightblue} D-Link &DIR-825 Rev.B &MIPS &2020 &D-Link &DIR-835 A1 &MIPS &2020 &D-Link &DIR-859 &MIPS &2020 \\ \hline
D-Link &DIR-860L A1 &ARM &2020 &D-Link &DIR-860L B1 &MIPS &2020 &D-Link &DIR-862 &MIPS &2020 \\ \hline
\rowcolor{lightblue} D-Link &DIR-866 &MIPS &2020 &D-Link &DIR-868l Rev.A &ARM &2020 &D-Link &DIR-868l Rev.B &ARM &2020 \\ \hline
D-Link &DIR-868l Rev.C &ARM &2020 &D-Link &DIR-869 &MIPS &2020 &D-Link &DIR-878 A1 &MIPS &2020 \\ \hline
\rowcolor{lightblue} D-Link &DIR-880l &ARM &2020 &D-Link &DIR-882 A1 &MIPS &2020 &D-Link &DIR-885l &ARM &2020 \\ \hline
D-Link &DIR-890l &ARM &2020 &D-Link &DIR632A &MIPS &2020 &GL.iNet &AR150 &MIPS &2020 \\ \hline
\rowcolor{lightblue} Gateworks &GW2382 &ARM &2018 &Gateworks &GW2388 16M &ARM &2018 &Gateworks &GW2391 &ARM &2018 \\ \hline
Gateworks &Laguna GW2382 &ARM &2020 &Gateworks &Laguna GW2388 32M &ARM &2020 &Gateworks &Laguna GW2391 &ARM &2020 \\ \hline
\rowcolor{lightblue} Gigaset &SX763 &MIPS &2020 &JJPlus &JA76PF &MIPS &2020 &Linksys &E1700 &MIPS &2020 \\ \hline
Linksys &E2100L &MIPS &2020 &Linksys &EA6350 &ARM &2020 &Linksys &EA6400 &ARM &2020 \\ \hline
\rowcolor{lightblue} Linksys &EA6500 v2 &ARM &2020 &Linksys &EA6700 &ARM &2020 &Linksys &EA6900 &ARM &2020 \\ \hline
Linksys &EA8500 &ARM &2020 &Linksys &RE7000 &MIPS &2020 &Linksys &WRT610N v1.0 &MIPS &2020 \\ \hline
\rowcolor{lightblue} NETGEAR &AC1450 &ARM &2020 &NETGEAR &EX6200 &ARM &2018 &NETGEAR &R6250 &MIPS &2020 \\ \hline
NETGEAR &R6300v2 &ARM &2020 &NETGEAR &R6400 &ARM &2020 &NETGEAR &R6700 &ARM &2020 \\ \hline
\rowcolor{lightblue} NETGEAR &R6700v3 &ARM &2020 &NETGEAR &R7000 &ARM &2020 &NETGEAR &R7000P &ARM &2020 \\ \hline
NETGEAR &R7500v1 &ARM &2020 &NETGEAR &R7500v2 &ARM &2020 &NETGEAR &R7800 &ARM &2020 \\ \hline
\rowcolor{lightblue} NETGEAR &R8500 &ARM &2020 &NETGEAR &R8900 &ARM &2020 &NETGEAR &R9000 &ARM &2020 \\ \hline
NETGEAR &WG302v2 &ARM &2020 &NETGEAR &WNDR3700v4 &MIPS &2020 &NETGEAR &WNDR4300 &MIPS &2020 \\ \hline
\rowcolor{lightblue} NETGEAR &WNDR4500 &MIPS &2020 &NETGEAR &WNDR4500v2 &MIPS &2020 &NETGEAR &XR450 &ARM &2020 \\ \hline
NETGEAR &XR500 &ARM &2020 &NETGEAR &XR700 &ARM &2020 &Pronghorn &SBC &ARM &2020 \\ \hline
\rowcolor{lightblue} Senao &ECB3500 &MIPS &2020 &Senao &ECB9750 &MIPS &2019 &Senao &EOC1650 &MIPS &2020 \\ \hline
Senao &EOC5510 &MIPS &2020 &Senao &EOC5610 &MIPS &2020 &Senao &EOC5611 &MIPS &2020 \\ \hline
\rowcolor{lightblue} Senao &NOP8670 &ARM &2020 &TP-Link &Archer A7 V5 &MIPS &2020 &TP-Link &Archer C1900 &ARM &2020 \\ \hline
TP-Link &Archer C25 V1 &MIPS &2020 &TP-Link &Archer C5 V1 &MIPS &2020 &TP-Link &Archer C7 V1 &MIPS &2020 \\ \hline
\rowcolor{lightblue} TP-Link &Archer C7 V2 &MIPS &2020 &TP-Link &Archer C7 V3 &MIPS &2020 &TP-Link &Archer C7 V4 &MIPS &2020 \\ \hline
TP-Link &Archer C8 V1 &ARM &2020 &TP-Link &Archer C9 V1 &ARM &2020 &TP-Link &Archer C9 V2 &ARM &2020 \\ \hline
\rowcolor{lightblue} TP-Link &Archer C9 V3 &ARM &2020 &TP-Link &Deco-M4 &MIPS &2020 &TP-Link &TL-WDR3600 V1 &MIPS &2020 \\ \hline
TP-Link &TL-WDR4300 V1 &MIPS &2020 &TP-Link &TL-WDR4310 V1 &MIPS &2020 &TP-Link &TL-WDR4900 V2 &MIPS &2020 \\ \hline
\rowcolor{lightblue} TP-Link &TL-WR1043N V5 &MIPS &2020 &TP-Link &TL-WR1043ND &MIPS &2020 &TP-Link &TL-WR1043ND V2 &MIPS &2020 \\ \hline
TP-Link &TL-WR1043ND V4 &MIPS &2020 &TP-Link &TL-WR2543ND &MIPS &2020 &TP-Link &TL-WR710N V1 &MIPS &2020 \\ \hline
\rowcolor{lightblue} TP-Link &TL-WR710N V2.1.0 &MIPS &2020 &TP-Link &TL-WR810N V1 &MIPS &2020 &TP-Link &TL-WR810N V2 &MIPS &2020 \\ \hline
TP-Link &TL-WR842N V1 &MIPS &2020 &TP-Link &TL-WR842N V2 &MIPS &2020 &TRENDnet &TEW-811DRU &ARM &2018 \\ \hline
\rowcolor{lightblue} TRENDnet &TEW-812DRU V2 &ARM &2018 &TRENDnet &TEW-818DRU &ARM &2020 &TRENDnet &TEW-828DRU &ARM &2020 \\ \hline
Ubiquiti &AirGrid M2 &MIPS &2020 &Ubiquiti &AirGrid M5 &MIPS &2020 &Ubiquiti &AirGrid-M5-XW &MIPS &2020 \\ \hline
\rowcolor{lightblue} Ubiquiti &AirRouter &MIPS &2020 &Ubiquiti &AirRouter-HP &MIPS &2020 &Ubiquiti &AirWire &MIPS &2020 \\ \hline
Ubiquiti &BulletM2 HP &MIPS &2020 &Ubiquiti &BulletM5 HP &MIPS &2020 &Ubiquiti &LS SR71A &MIPS &2020 \\ \hline
\rowcolor{lightblue} Ubiquiti &NanoBeam AC &MIPS &2020 &Ubiquiti &NanoBeam M2 XW &MIPS &2020 &Ubiquiti &NanoBeam M5 XW &MIPS &2020 \\ \hline
Ubiquiti &NanoBridge M2 &MIPS &2020 &Ubiquiti &NanoBridge M2 XW &MIPS &2020 &Ubiquiti &NanoBridge M3 &MIPS &2020 \\ \hline
\rowcolor{lightblue} Ubiquiti &NanoBridge M365 &MIPS &2020 &Ubiquiti &NanoBridge M5 XW &MIPS &2020 &Ubiquiti &NanoBridge M900 &MIPS &2020 \\ \hline
Ubiquiti &NanoStation M2 &MIPS &2020 &Ubiquiti &NanoStation M3 &MIPS &2020 &Ubiquiti &NanoStation M365 &MIPS &2020 \\ \hline
\rowcolor{lightblue} Ubiquiti &Pico M5 &MIPS &2020 &Ubiquiti &Power AP N &MIPS &2020 &Ubiquiti &PowerBeam M5 M400 XW &MIPS &2020 \\ \hline
Ubiquiti &PowerBridge M10 &MIPS &2020 &Ubiquiti &PowerBridge M5 &MIPS &2020 &Ubiquiti &Rocket M2 Titanium XW &MIPS &2020 \\ \hline
\rowcolor{lightblue} Ubiquiti &Rocket M2 XW &MIPS &2020 &Ubiquiti &Rocket M5 Titanium XW &MIPS &2020 &Ubiquiti &Rocket M5 X3 XW &MIPS &2020 \\ \hline
Ubiquiti &Rocket M5 XW &MIPS &2020 &Ubiquiti &RocketM2 &MIPS &2020 &Ubiquiti &RocketM3 &MIPS &2020 \\ \hline
\rowcolor{lightblue} Ubiquiti &RocketM365 &MIPS &2020 &Ubiquiti &RocketM5 &MIPS &2020 &Ubiquiti &RocketM900 &MIPS &2020 \\ \hline
Ubiquiti &RouterStation &MIPS &2020 &Ubiquiti &RouterStation Pro &MIPS &2020 &Ubiquiti &sunMax &MIPS &2020 \\ \hline
\rowcolor{lightblue} Ubiquiti &UAP-AC-MESH &MIPS &2020 &Ubiquiti &UAP-AC-PRO &MIPS &2020 &Ubiquiti &UAP-LR &MIPS &2020 \\ \hline
Ubiquiti &UAP-LR-v2 &MIPS &2020 &Ubiquiti &UAP-v2 &MIPS &2020 &Ubiquiti &locoM2 &MIPS &2020 \\ \hline
\rowcolor{lightblue} Ubiquiti &locoM2 XW &MIPS &2020 &Ubiquiti &locoM5 &MIPS &2020 &Ubiquiti &locoM5 XW &MIPS &2020 \\ \hline
Ubiquiti &locoM900 &MIPS &2020 &WiliGear &WBD-500 &MIPS &2020 &YunCore &XD3200 &MIPS &2020 \\ \bottomrule[1.1pt]
\end{tabular}
\end{table*}
\subsection{Probing Learned Execution Semantics}
\label{subsec:probe_case}
As discussed in Section~\ref{sec:overview}, training the model to predict masked micro-trace codes and values compels the model to learn execution semantics by concrete examples.
It thus automates the extraction of the code's dynamic features without manual effort.
In this section, we study concrete code examples showing the potential hint in the code that the model likely leverages to predict the masked part.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{./figs/case/arithmetic.pdf}
\caption{The partial micro-trace code and value sequence from function \texttt{BinarySource\_get\_rsa\_ssh1\_pub} in PuTTy-0.74 compiled to x64 with \texttt{O0}. We mask the register \colorbox{lightgray}{\texttt{rax}} at line 3 and the opcode \colorbox{lightgray}{\texttt{add}} at line 4. We also mask their corresponding micro-trace values (where opcode has only dummy values (8 \texttt{\#\#}s) as described in Section~\ref{subsec:input_repr}).
We highlight the contextual hint in \colorbox{myblue}{blue} that \textsc{Trex}\xspace leverages to predict the masked \colorbox{lightgray}{\texttt{rax}} and \colorbox{lightred}{red} that \textsc{Trex}\xspace uses to predict the masked \colorbox{lightgray}{\texttt{add}}.}
\label{fig:arithmetic}
\end{figure}
\vspace{.1cm}\noindent\textbf{Predicting arithmetic instructions.}
Consider the example in Figure~\ref{fig:arithmetic}, which shows the function's micro-trace (\emph{i.e., } the dynamic value and assembly code).
For ease of exposition, the format of micro-trace values (\emph{e.g., } the value of opcode and constants) are not exactly the same as the actual format we handle, as described in Section~\ref{subsec:input_repr}.
We mask the register \texttt{rax} and opcode \texttt{add} and their corresponding values in a sequence of arithmetic instructions.
To correctly predict the masked value \texttt{0x4c0469a} of \texttt{rax}, the model should understand the approximate execution semantics of \texttt{sub} in line 2, which subtracts \texttt{rax} with \texttt{rcx}.
It may also observe the value of \texttt{rax} at line 4, which exactly equals the result of adding \texttt{rax} with 7 at line 3.
Therefore, we see the model predicts \texttt{rax} with 96\% confidence and \texttt{rcx} with only 3\%.
It also predicts the value of \texttt{rax} correctly (right table).
To correctly predict \texttt{add} at line 4, the model should observe that the \texttt{rax} at line 5 has the same exact value with the result of \texttt{rax+rcx} at line 4, so it predicts the opcode at line 4 to be \texttt{add}.
As shown in Figure~\ref{fig:arithmetic}, our pretrained model predicts \texttt{add} with 98\% confidence and \texttt{mov} with only 1\%, which implies that the model approximately understands the execution semantics of \texttt{add} and \texttt{mov} so it predicts that \texttt{add} is much more likely.
\vspace{.1cm}\noindent\textbf{Predicting stack operations.}
Consider the example in Figure~\ref{fig:stack}.
We mask the register and constant of the instruction in function epilogue -- it increments the stack pointer \texttt{rsp} by \texttt{0x20} to deallocate the local variable stored on stack.
To correctly predict the masked \texttt{rsp} and its value \texttt{0x8b4a3f}, the model should observe the \texttt{rsp} is decremented (due to opcode \texttt{sub}) by \texttt{0x20} from \texttt{0x8b4a3f} at line 3, which is part of the function prologue.
Therefore, the model should understand the execution semantics of \texttt{sub} and basic syntax of function prologue and epilogue.
To predict the masked constant \texttt{0x20}, the model should notice that line 3 decrements the stack pointer by \texttt{0x20}, which is the size of local variables.
As the model predicts \texttt{0x20} with 99\% confidence, this implies the model likely learns patterns of function prologue and epilogue, including the context such as \texttt{push ebp}, \texttt{pop ebp}, etc.
This observation indicates that our pretraining task assists in learning common function idiom (function calling convention) beyond execution semantics of individual instructions.
Learning such knowledge can potentially help other tasks beyond function similarity, such as identifying function boundaries~\cite{pei2021xda}.
\begin{figure}[!t]
\centering
\includegraphics[width=.95\linewidth]{./figs/case/stack.pdf}
\caption{The partial code and micro-trace from function \texttt{closeUnixFile} in SQLite-3.34.0 compiled to x64 with \texttt{O0}.
We mask the register \colorbox{lightgray}{\texttt{rsp}} and constant \colorbox{lightgray}{\texttt{0x20}} at line 4. We also mask their corresponding micro-trace values.
We highlight the contextual hint in \colorbox{lightred}{red} that \textsc{Trex}\xspace leverages to predict the masked \colorbox{lightgray}{\texttt{rsp}} and \colorbox{lightgray}{\texttt{0x20}}.}
\label{fig:stack}
\end{figure}
\end{appendix}
\section{Case Studies}
\label{sec:case}
In this section, we study how \textsc{Trex}\xspace can help discover new vulnerabilities in a large volume of latest firmware images.
\begin{table}[!t]
\footnotesize
\setlength{\tabcolsep}{4pt}
\centering
\renewcommand{\arraystretch}{1}
\setlength\aboverulesep{0.4pt}
\setlength\belowrulesep{0.4pt}
\caption{Vulnerabilities we have confirmed (\cmark) in firmware images (latest version) from 4 well-known vendors and products.}
\label{tab:cve_case}
\begin{tabular}{r?{1.1pt}c|c|c|c}
\toprule[1.1pt]
\multicolumn{1}{c|}{\multirow{2}{*}{\textbf{CVE}}} &
\multicolumn{1}{c|}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Ubiquiti\\ sunMax\end{tabular}}}} &
\multicolumn{1}{c|}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}TP-Link\\ Deco-M4\end{tabular}}}} &
\multicolumn{1}{c|}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}NETGEAR\\ R7000\end{tabular}}}} &
\multicolumn{1}{c}{\multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}Linksys\\ RE7000\end{tabular}}}} \\
\multicolumn{1}{c|}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c}{} \\ \midrule[.9pt]
CVE-2019-1563 & \cmark & \cmark & \cmark & \xmark \\ \hline
\rowcolor{lightblue} CVE-2017-16544 & \xmark & \cmark & \xmark & \xmark \\ \hline
CVE-2016-6303 & \cmark & \cmark & \cmark & \cmark \\ \hline
\rowcolor{lightblue} CVE-2016-6302 & \cmark & \cmark & \cmark & \cmark \\ \hline
CVE-2016-2842 & \cmark & \cmark & \cmark & \cmark \\ \hline
\rowcolor{lightblue} CVE-2016-2182 & \cmark & \cmark & \cmark & \cmark \\ \hline
CVE-2016-2180 & \cmark & \cmark & \cmark & \cmark \\ \hline
\rowcolor{lightblue} CVE-2016-2178 & \cmark & \cmark & \cmark & \cmark \\ \hline
CVE-2016-2176 & \cmark & \cmark & \xmark & \cmark \\ \hline
\rowcolor{lightblue} CVE-2016-2109 & \cmark & \cmark & \xmark & \cmark \\ \hline
CVE-2016-2106 & \cmark & \cmark & \xmark & \cmark \\ \hline
\rowcolor{lightblue} CVE-2016-2105 & \cmark & \cmark & \xmark & \cmark \\ \hline
CVE-2016-0799 & \cmark & \cmark & \xmark & \cmark \\ \hline
\rowcolor{lightblue} CVE-2016-0798 & \cmark & \cmark & \xmark & \cmark \\ \hline
CVE-2016-0797 & \cmark & \cmark & \xmark & \cmark \\ \hline
\rowcolor{lightblue} CVE-2016-0705 & \cmark & \cmark & \xmark & \cmark \\ \bottomrule[1.1pt]
\end{tabular}
\end{table}
Firmware images often include third-party libraries (\emph{i.e., } BusyBox, OpenSSL).
However, these libraries are frequently patched, but the manufacturers often fall behind to update them accordingly~\cite{OWASP}.
Therefore, we study whether our tool can uncover function binaries in firmware images similar to known vulnerable functions.
We find existing state-of-the-art binary similarity tools (\emph{i.e., } SAFE~\cite{massarelli2019safe}, Asm2Vec~\cite{ding2019asm2vec}, Gemini~\cite{xu2017neural}) all perform their case studies on the firmware images and vulnerabilities that have already been studied before them~\cite{costin2014large, feng2016scalable, david2016statistical, chen2016towards}.
Therefore, we decide to collect our own dataset with more updated firmware images and the latest vulnerabilities, instead of reusing the existing benchmarks.
This facilitates finding new (1-day) vulnerabilities in most recent firmware images that are not disclosed before.
We crawl firmware images (in their latest version) in 180 products including WLAN routers, smart cameras, and solar panels, from well-known manufacturers' official releases and third-party providers such as DD-WRT~\cite{ddwrt}, as shown in Table~\ref{tab:firmware}.
We collect firmware images with only the latest version, which are much more updated than those studied in the state-of-the-art (dated before 2015, which are all patched before their studies).
For each function in the firmware images, we construct function embedding and build a firmware image database using Open Distro for Elasticsearch~\cite{elastic}, which supports vector-based indexing with efficient search support based on NMSLIB~\cite{boytsov2013engineering}.
We extract the firmware images using binwalk~\cite{binwalk}.
In total, we collect 180 number of firmware images from 22 vendors.
These firmware images are developed by the original vendors, or from third-party providers such as DD-WRT~\cite{ddwrt}.
Most of them are for WLAN routers, while some are deployed in other embedded systems, such as solar panels and smart cameras.
Among the firmware images, 127 of them are compiled for MIPS 32-bit and 53 are compiled for ARM 32-bit.
169 of them are 2020 models, and the rest are 2019 and 2018 models.
\begin{table}[!t]
\footnotesize
\setlength{\tabcolsep}{7pt}
\centering
\renewcommand{\arraystretch}{1}
\setlength\aboverulesep{0.4pt}
\setlength\belowrulesep{0.4pt}
\caption{We study 16 vulnerabilities from OpenSSL and BusyBox, two widely-used libraries in firmware images.}
\label{tab:cve}
\begin{tabular}{r|l|l}
\toprule[1.1pt]
\textbf{CVE} & \textbf{Library} & \textbf{Description} \\ \midrule[.9pt]
CVE-2019-1563 & OpenSSL & Decrypt encrypted message \\ \hline
\rowcolor{lightblue} CVE-2017-16544 & BusyBox & Allow executing arbitrary code \\ \hline
CVE-2016-6303 & OpenSSL & Integer overflow \\ \hline
\rowcolor{lightblue} CVE-2016-6302 & OpenSSL & Allows denial-of-service \\ \hline
CVE-2016-2842 & OpenSSL & Allows denial-of-service \\ \hline
\rowcolor{lightblue} CVE-2016-2182 & OpenSSL & Allows denial-of-service \\ \hline
CVE-2016-2180 & OpenSSL & Out-of-bounds read \\ \hline
\rowcolor{lightblue} CVE-2016-2178 & OpenSSL & Leak DSA private key \\ \hline
CVE-2016-2176 & OpenSSL & Buffer over-read \\ \hline
\rowcolor{lightblue} CVE-2016-2109 & OpenSSL & Allows denial-of-service \\ \hline
CVE-2016-2106 & OpenSSL & Integer overflow \\ \hline
\rowcolor{lightblue} CVE-2016-2105 & OpenSSL & Integer overflow \\ \hline
CVE-2016-0799 & OpenSSL & Out-of-bounds read \\ \hline
\rowcolor{lightblue} CVE-2016-0798 & OpenSSL & Allows denial-of-service \\ \hline
CVE-2016-0797 & OpenSSL & NULL pointer dereference \\ \hline
\rowcolor{lightblue} CVE-2016-0705 & OpenSSL & Memory corruption \\ \bottomrule[1.1pt]
\end{tabular}
\end{table}
Table~\ref{tab:cve_case} shows 16 vulnerabilities (CVEs) we use to search in the firmware images.
We focus on the CVEs of OpenSSL and BusyBox, as they are widely included in the firmware.
For each CVE, we compile the corresponding vulnerable functions in the specified library version and computes the vulnerable function embeddings via \textsc{Trex}\xspace.
As the firmware images are stripped so that we do not know with which optimizations they are compiled, we compile the vulnerable functions to both MIPS and ARM with \texttt{-O3} and rely on \textsc{Trex}\xspace's capability in cross-architecture and optimization function matching to match functions that are potentially compiled in different architectures and with different optimizations.
We then obtain the firmware functions that rank top-10 similar to the vulnerable function and manually verify if they are vulnerable.
We leverage \texttt{strings} command to identify the OpenSSL and BusyBox versions indicative of the corresponding vulnerabilities.
Note that such information can be stripped for other libraries so it is not a reliable approach in general.
As shown in Table~\ref{tab:cve_case}, we have confirmed all 16 CVEs in 4 firmware models developed by well-known vendors, \emph{i.e., } Ubiquiti, TP-Link, NETGEAR, and Linksys.
These cases demonstrate the practicality of \textsc{Trex}\xspace, which helps discover real-world vulnerabilities in large-scale firmware databases.
Table~\ref{tab:cve} shows the details of the 16 vulnerabilities that \textsc{Trex}\xspace uncover in the 4 firmware images shown in Table~\ref{tab:cve_case}.
The description of ``allow denial-of-service'' usually refers to the segmentation fault that crashes the program.
The cause of such error can be diverse, which is not due to the other typical causes shown in the table (\emph{e.g., } integer overflow, buffer over-read, etc.).
\section{Conclusion}
\label{sec:conclusion}
We introduced \textsc{Trex}\xspace to match semantically similar functions based on the function execution semantics.
Our key insight is to first pretrain the ML model to explicitly learn approximate execution semantics based on the functions' mico-traces and then transfer the learned knowledge to match semantically similar functions.
Our evaluation showed that the learned approximate execution semantics drastically improves the accuracy of matching semantically similar functions -- \textsc{Trex}\xspace excels in matching functions across different architectures, optimizations, and obfuscations.
We plan to explore in our future work how the learned execution semantics of the code can further boost the performance of broader (binary) program analysis tasks such as decompilation.
\section{Discussion}
\label{sec:discussion}
\vspace{.1cm}
\noindent\textbf{Definition of pretraining and finetuning.}
A typical difference between pretraining and finetuning in NLP is pretraining is often self-/un-supervised while finetuning is supervised.
Besides, pretraining is often believed responsible for heavy-lifting, with huge amount of training corpus and parameter updates. While in finetuning, the training process often does not significantly update the ``knowledge'' of the pretrained model. Instead, it is often light-weight with small extra layers and number of training epochs.
In general, pretraining learns task-agnostic general knowledge while finetuning learns for task-specific downstream tasks.
In the setting of this project, such distinction becomes not so obvious, as our finetuning task (function similarity) does not necessarily requires any labeling effort. Particularly, we can cheaply collect the same source code and compile into different architectures, optimizations, and obfuscations. Therefore, we believe our finetuning task can also be leveraged to learn general semantic representation of the code. By encouraging the model to produce similar embeddings for same functions with different transformations, we teach the model invariances of underlying code semantics.
We plan to explore this direction in the future work.
\section{Evaluation}
\label{sec:eval}
Our evaluation aims to answer the following questions.
\begin{itemize}
\item RQ1: How accurate is \textsc{Trex}\xspace in matching semantically similar function binaries across different architectures, optimizations, and obfuscations?
\item RQ2: How does \textsc{Trex}\xspace compare to the state-of-the-art?
\item RQ3: How fast is \textsc{Trex}\xspace compared to other tools?
\item RQ4: How much does pretraining on micro-traces help improve the accuracy of matching functions?
\end{itemize}
\begin{table}[!t]
\footnotesize
\setlength{\tabcolsep}{5.5pt}
\centering
\renewcommand{\arraystretch}{1}
\setlength\aboverulesep{0.4pt}
\setlength\belowrulesep{0.4pt}
\caption{\textsc{Trex}\xspace results (in AUC score) on function pairs across architectures, optimizations, and obfuscations.
}
\label{tab:result}
\begin{tabular}{r?{1.1pt}ccccc}
\toprule[1.1pt]
\multirow{2}{*}{} & \multicolumn{5}{c}{\textbf{Cross-}} \\ \cline{2-6}
& \multicolumn{1}{c|}{\textbf{ARCH}} & \multicolumn{1}{c|}{\textbf{OPT}} & \multicolumn{1}{c|}{\textbf{OBF}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}l@{}}ARCH+\\ OPT\end{tabular}}} & \multicolumn{1}{c}{\textbf{\begin{tabular}[c]{@{}l@{}}ARCH+\\ OPT+\\ OBF\end{tabular}}} \\ \midrule[.9pt]
\rowcolor{lightblue} \textbf{Binutils} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.959} & 0.947 \\ \hline
\textbf{Coreutils} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.956} & 0.945 \\ \hline
\rowcolor{lightblue} \textbf{Curl} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.958} & 0.956 \\ \hline
\textbf{Diffutils} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.970} & 0.961 \\ \hline
\rowcolor{lightblue} \textbf{Findutils} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.965} & 0.963 \\ \hline
\textbf{GMP} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.968} & 0.966 \\ \hline
\rowcolor{lightblue} \textbf{ImageMagick} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.989} & \multicolumn{1}{c|}{0.960} & 0.951 \\ \hline
\textbf{Libmicrohttpd} & \multicolumn{1}{c|}{0.994} & \multicolumn{1}{c|}{0.994} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.972} & 0.969 \\ \hline
\rowcolor{lightblue} \textbf{LibTomCrypt} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.994} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.981} & 0.970 \\ \hline
\textbf{OpenSSL} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.989} & \multicolumn{1}{c|}{0.964} & 0.956 \\ \hline
\rowcolor{lightblue} \textbf{PuTTy} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.995} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.961} & 0.952 \\ \hline
\textbf{SQLite} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.994} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.980} & 0.959 \\ \hline
\rowcolor{lightblue} \textbf{Zlib} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.991} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.979} & 0.965 \\ \midrule[.9pt]
\textbf{Average} & \multicolumn{1}{c|}{0.992} & \multicolumn{1}{c|}{0.993} & \multicolumn{1}{c|}{0.990} & \multicolumn{1}{c|}{0.967} & 0.958 \\ \bottomrule[1.1pt]
\end{tabular}
\end{table}
\subsection{RQ1: Accuracy}
\label{subsec:rq1}
We evaluate how accurate \textsc{Trex}\xspace is in matching similar functions across different architectures, optimizations, and obfuscations.
As shown in Table~\ref{tab:result}, we prepare function pairs for each project (first column) with 5 types of partitions.
(1) ARCH: the function pairs have \emph{different architectures} but same optimizations without obfuscations (2nd column).
(2) OPT: the function pairs have \emph{different optimizations} but same architectures without obfuscations (3rd column).
(3) OBF: the function pairs have \emph{different obfuscations} with same architectures (x64) and no optimization (4th column).
(4) ARCH+OPT: the function pairs have \emph{both different architectures and optimizations} without obfuscations (5th column).
(5) ARCH+OPT+OBF: the function pairs can come from arbitrary architectures, optimizations, and obfuscations (6th column).
Table~\ref{tab:result} reports the mean testing AUC scores of \textsc{Trex}\xspace on each project with 3 runs.
On average, \textsc{Trex}\xspace achieves $>0.958$ (and up to 0.995) AUC scores, even in the most challenging setting where the functions can come from different architectures, optimizations, and obfuscations at the same time.
We note that \textsc{Trex}\xspace performs the best on cross-optimization matching.
This is intuitive as the syntax of two functions from different optimizations are not changed significantly (\emph{e.g., } the name of opcode, operands remain the same).
Nevertheless, we find the AUC scores for matching functions from different architectures is only 0.001 lower, which indicates the model is robust to entirely different syntax between two architectures.
On matching functions with different obfuscations, \textsc{Trex}\xspace's results are 0.026, on average, lower than that of cross-optimizations, which indicates the obfuscation changes the code more drastically.
Section~\ref{subsec:rq2} will show the specific results of \textsc{Trex}\xspace on each obfuscations.
\subsection{RQ2: Baseline Comparison}
\label{subsec:rq2}
\vspace{.1cm}\noindent\textbf{Cross-architecture search.}
As described in Section~\ref{sec:impl}, we first compare \textsc{Trex}\xspace with SAFE and Gemini on OpenSSL-1.0.1f and OpenSSL-1.0.1u with their reported numbers (as they only evaluate on these two projects).
We then run SAFE's released model on our dataset and compare to \textsc{Trex}\xspace.
We use our testing setup (see Section~\ref{sec:impl}) to evaluate SAFE's trained model, where 90\% of the total functions of those in Table~\ref{tab:dataset} are used to construct the testing set.
These testing sets are much larger than that in SAFE, where they only evaluate on 20\% of the OpenSSL functions.
Note that the dataset used in SAFE are all compiled by \texttt{GCC-5.4} at the time when it is publicized (November 2018), while ours are compiled by \texttt{GCC-7.5} (April 2020).
We find these factors (the more diverse dataset and different compilers) can all lead to the possible dataset distribution shift, which often results in the decaying performance of ML models when applied in the security applications~\cite{jordaney2017transcend}.
To study the distribution shift, we measure the KL-divergence~\cite{kullback1951information} between SAFE's dataset (OpenSSL compiled by \texttt{GCC-5.4}) and our dataset.
Specifically, we use the probability distribution of the raw bytes of the compiled projects' binaries, and compute their KL-divergence between SAFE and ours.
As OpenSSL is also a subset of our complete dataset, we compute the KL-divergence between our compiled OpenSSL and that of SAFE as the baseline.
We find the KL-divergence is 0.02 between SAFE's dataset and ours, while it decreases to 0.0066 between our compiled OpenSSL and that of SAFE.
This indicates that the underlying distribution of our test set shifts from that of SAFE's.
Moreover, the KL-divergence of 0.0066 between the same dataset (OpenSSL) but only compiled by different GCC versions implies that the compiler version has much smaller effect to the distribution shift than the different software projects.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\linewidth]{./figs/compare/roc-drawio-safe-gemini.pdf}
\caption{ROC curves of matching functions in OpenSSL across different architectures. \textsc{Trex}\xspace outperforms the reported results of SAFE and Gemini and the results of running SAFE's trained model on our testing set.}
\label{fig:cross-arch-roc}
\end{figure}
As shown in Figure~\ref{fig:cross-arch-roc}, our ROC curve is higher than those reported in SAFE and Gemini.
While SAFE's reported AUC score, 0.992, is close to ours, when we run their trained model on our testing set, its AUC score drops to 0.976 -- possibly because our testing set is much larger than theirs.
This observation demonstrates the generalizability of \textsc{Trex}\xspace -- when pretrained to approximately learn execution semantics explicitly, it can quickly generalize to match unseen (semantically similar) functions with only a minimal training set.
\begin{figure}[!t]
\centering
\includegraphics[scale=.4]{./figs/compare/cross-arch-auc-legend.pdf}
\vspace{.1cm}
\includegraphics[width=\linewidth]{./figs/compare/cross-arch-auc.pdf}
\caption{Comparison between \textsc{Trex}\xspace and SAFE on matching functions in each project compiled to different architectures (see Table~\ref{tab:dataset}).}
\label{fig:cross-arch-auc-safe}
\end{figure}
Figure~\ref{fig:cross-arch-auc-safe} shows that \textsc{Trex}\xspace consistently outperforms SAFE on all projects, \emph{i.e., } by 7.3\% on average.
As the SAFE's model is only trained on OpenSSL, we also follow the same setting by training \textsc{Trex}\xspace on only OpenSSL, similar to the cross-project setting described in Section~\ref{subsec:rq1}.
\vspace{.1cm}\noindent\textbf{Cross-optimization search.}
We compare \textsc{Trex}\xspace with Asm2Vec and BLEX on matching functions compiled by different optimizations.
As both Asm2vec and Blex run on single architecture, we restrict the comparison on x64.
Besides, since Asm2Vec uses Precision@1 and Blex uses accuracy as the metric (discussed in Section~\ref{sec:impl}), we compare with each tool separately using their metrics and on their evaluated dataset.
\begin{table}[!t]
\footnotesize
\setlength{\tabcolsep}{8pt}
\centering
\renewcommand{\arraystretch}{1}
\setlength\aboverulesep{0.4pt}
\setlength\belowrulesep{0.4pt}
\caption{Comparison between \textsc{Trex}\xspace and Asm2Vec (in Precision@1) on function pairs across optimizations.
}
\label{tab:result-asm2vec}
\begin{tabular}{r?{1.1pt}cccc}
\toprule[1.1pt]
\multicolumn{1}{c?{1.1pt}}{} & \multicolumn{4}{c}{Cross Compiler Optimization} \\ \cline{2-5}
\multicolumn{1}{l?{1.1pt}}{} & \multicolumn{2}{c?{1.1pt}}{\texttt{O2} and \texttt{O3}} & \multicolumn{2}{c}{\texttt{O0} and \texttt{O3}} \\ \cline{2-5}
\multicolumn{1}{c?{1.1pt}}{\textbf{}} & \multicolumn{1}{c|}{\textsc{Trex}\xspace} & \multicolumn{1}{c?{1.1pt}}{Asm2Vec} & \multicolumn{1}{c|}{\textsc{Trex}\xspace} & Asm2Vec \\ \midrule[.9pt]
\rowcolor{lightblue} Coreutils & \multicolumn{1}{c|}{\textbf{0.955}} & \multicolumn{1}{c?{1.1pt}}{0.929} & \multicolumn{1}{c|}{\textbf{0.913}} & 0.781 \\ \hline
Curl & \multicolumn{1}{c|}{\textbf{0.961}} & \multicolumn{1}{c?{1.1pt}}{0.951} & \multicolumn{1}{c|}{\textbf{0.894}} & 0.850 \\ \hline
\rowcolor{lightblue} GMP & \multicolumn{1}{c|}{\textbf{0.974}} & \multicolumn{1}{c?{1.1pt}}{0.973} & \multicolumn{1}{c|}{\textbf{0.886}} & 0.763 \\ \hline
ImageMagick & \multicolumn{1}{c|}{\textbf{0.971}} & \multicolumn{1}{c?{1.1pt}}{0.971} & \multicolumn{1}{c|}{\textbf{0.891}} & 0.837 \\ \hline
\rowcolor{lightblue} LibTomCrypt & \multicolumn{1}{c|}{\textbf{0.991}} & \multicolumn{1}{c?{1.1pt}}{0.991} & \multicolumn{1}{c|}{\textbf{0.923}} & 0.921 \\ \hline
OpenSSL & \multicolumn{1}{c|}{\textbf{0.982}} & \multicolumn{1}{c?{1.1pt}}{0.931} & \multicolumn{1}{c|}{\textbf{0.914}} & 0.792 \\ \hline
\rowcolor{lightblue} PuTTy & \multicolumn{1}{c|}{\textbf{0.956}} & \multicolumn{1}{c?{1.1pt}}{0.891} & \multicolumn{1}{c|}{\textbf{0.926}} & 0.788 \\ \hline
SQLite & \multicolumn{1}{c|}{\textbf{0.931}} & \multicolumn{1}{c?{1.1pt}}{0.926} & \multicolumn{1}{c|}{\textbf{0.911}} & 0.776 \\ \hline
\rowcolor{lightblue} Zlib & \multicolumn{1}{c|}{\textbf{0.890}} & \multicolumn{1}{c?{1.1pt}}{0.885} & \multicolumn{1}{c|}{\textbf{0.902}} & 0.722 \\ \midrule[.9pt]
Average & \multicolumn{1}{c|}{\textbf{0.957}} & \multicolumn{1}{c?{1.1pt}}{0.939} & \multicolumn{1}{c|}{\textbf{0.907}} & 0.803 \\ \bottomrule[1.1pt]
\end{tabular}
\end{table}
Table~\ref{tab:result-asm2vec} shows \textsc{Trex}\xspace outperforms Asm2Vec in Precision@1 (by 7.2\% on average) on functions compiled by different optimizations (\emph{i.e., } between \texttt{O2} and \texttt{O3} and between \texttt{O0} and \texttt{O3}).
As the syntactic difference introduced by optimizations between \texttt{O0} and \texttt{O3} is more significant than that between \texttt{O2} and \texttt{O3}, both tools have certain level of decrease in AUC scores (5\% drop for \textsc{Trex}\xspace and 14\% for Asm2Vec), but \textsc{Trex}\xspace's AUC score drops much less than that of Asm2Vec.
To compare to Blex, we evaluate \textsc{Trex}\xspace on Coreutils between optimizations \texttt{O0} and \texttt{O3}, where they report to achieve better performance than BinDiff~\cite{bindiff}.
As Blex show the matched functions of each individual utility in Coreutils in a barchart without including the concrete numbers of matched functions, we estimate their matched functions using their reported average percentage (75\%) on all utilities.
\begin{figure}[!t]
\centering
\includegraphics[width=.95\linewidth]{./figs/compare/cross-opt-blex.pdf}
\caption{Cross-optimization function matching between \texttt{O0} and \texttt{O3} on Coreutils by \textsc{Trex}\xspace and Blex. We sort the 109 utility binaries in Coreutils by their number of functions, and aggregate the matched functions every 10 utilities.}
\label{fig:cross-opt-blex}
\end{figure}
Figure~\ref{fig:cross-opt-blex} shows that \textsc{Trex}\xspace consistently outperforms Blex in number of matched functions in all utility programs of Coreutils.
Note that Blex also executes the function and uses the dynamic features to match binaries.
The observation here thus implies that the learned execution semantics from \textsc{Trex}\xspace is more effective than the hand-coded features in Blex for matching similar binaries.
\vspace{.1cm}\noindent\textbf{Cross-obfuscation search.}
We compare \textsc{Trex}\xspace to Asm2Vec on matching obfuscated function binaries with different obfuscation methods.
Notably, Asm2Vec is evaluated on obfuscations including bogus control flow (\texttt{bcf}), control flow flattening (\texttt{cff}), and instruction substitution (\texttt{sub}), which are subset of our evaluated obfuscations (Table~\ref{tab:dataset}).
As Asm2Vec only evaluates on 4 projects, \emph{i.e., } GMP, ImageMagic, LibTomCrypt, and OpenSSL, we focus on these 4 projects and \textsc{Trex}\xspace's results for other projects are included in Table~\ref{tab:result}.
\begin{table}[!t]
\footnotesize
\setlength{\tabcolsep}{2.5pt}
\centering
\renewcommand{\arraystretch}{1}
\setlength\aboverulesep{0.4pt}
\setlength\belowrulesep{0.4pt}
\caption{Comparison between \textsc{Trex}\xspace and Asm2Vec (in Precision@1) on function pairs across differet obfuscations.
}
\label{tab:cross-obf-asm2vec}
\begin{tabular}{r|r?{1.1pt}l|l|l|l?{1.1pt}l}
\toprule[1.1pt]
& & GMP & LibTomCrypt & ImageMagic & OpenSSL & Average \\ \midrule[.9pt]
\rowcolor{lightblue} & \textsc{Trex}\xspace & \textbf{0.926} & \textbf{0.938} & \textbf{0.934} & \textbf{0.898} & \textbf{0.924} \\ \cline{2-7}
\rowcolor{lightblue} \multirow{-2}{*}{\texttt{bcf}} & Asm2Vec & 0.802 & 0.920 & 0.933 & 0.883 & 0.885 \\ \midrule[.9pt]
\multirow{2}{*}{\texttt{ccf}} & \textsc{Trex}\xspace & \textbf{0.943} & \textbf{0.931} & \textbf{0.936} & \textbf{0.940} & \textbf{0.930} \\ \cline{2-7}
& Asm2Vec & 0.772 & 0.920 & 0.890 & 0.795 & 0.844 \\ \midrule[.9pt]
\rowcolor{lightblue} & \textsc{Trex}\xspace & \textbf{0.949} & \textbf{0.962} & \textbf{0.981} & \textbf{0.980} & \textbf{0.968} \\ \cline{2-7}
\rowcolor{lightblue} \multirow{-2}{*}{\texttt{sub}} & Asm2Vec & 0.940 & 0.960 & 0.981 & 0.961 & 0.961 \\ \midrule[.9pt]
\multirow{2}{*}{All} & \textsc{Trex}\xspace & \textbf{0.911} & \textbf{0.938} & \textbf{0.960} & \textbf{0.912} & \textbf{0.930} \\ \cline{2-7}
& Asm2Vec & 0.854 & 0.880 & 0.830 & 0.690 & 0.814 \\ \bottomrule[1.1pt]
\end{tabular}
\end{table}
Table~\ref{tab:cross-obf-asm2vec} shows \textsc{Trex}\xspace achieves better Precision@1 score (by 14.3\% on average) throughout all different obfuscations.
Importantly, the last two rows show when multiple obfuscations are combined, \textsc{Trex}\xspace performance is not dropping as significant as Asm2Vec.
It also shows \textsc{Trex}\xspace remains robust under varying obfuscations with different difficulties.
For example, instruction substitution simply replaces very limited instructions (\emph{i.e., } arithmetic operations as shown in Section~\ref{sec:overview}) while control flow flattening dramatically changes the function code.
Asm2Vec has 12.2\% decreased score when the obfuscation is changed from \texttt{sub} to \texttt{ccf}, while \textsc{Trex}\xspace only decreases by 4\%.
\subsection{RQ3: Execution Time}
\label{subsec:rq3}
We evaluate the runtime performance of generating function embeddings for computing similarity.
We compare \textsc{Trex}\xspace with SAFE and Gemini on generating functions in 4 projects on x64 compiled by \texttt{O3}, \emph{i.e., } Binutils, Putty, Findutils, and Diffutils, which have disparate total number of functions (see Table~\ref{tab:dataset}.
This tests how \textsc{Trex}\xspace scales to different number of functions compared to other baselines.
Since the offline training (\emph{i.e., } pretraining \textsc{Trex}\xspace) of all the learning-based tools is a one-time cost, it can be amortized in the function matching process so we do not explicitly measure the training time.
Moreover, the output of all tools are function embeddings, which can be indexed and efficiently searched using locality sensitive hashing (LSH)~\cite{gionis1999similarity, rajaraman2011mining}.
Therefore, we do not compare the matching time of function embeddings as it simply depends on the performance of underlying LSH implementation.
Particularly, we compare the runtime of two procedures in matching functions.
(1) Function parsing, which transforms the function binaries into the format that the model needs.
(2) Embedding generation, which takes the parsed function binary as input and computes function embedding. We test the embedding generation using our GPU (see Section~\ref{sec:impl}).
\begin{figure}[!t]
\centering
\includegraphics[scale=.4]{./figs/runtime/runtime_bar_legend.pdf}
\subfloat[Function parsing]{
\includegraphics[width=0.47\linewidth]{./figs/runtime/runtime_bar_parse.pdf}
\label{subfig:runtime_bar_parse}}
\subfloat[Embedding generation]{
\includegraphics[width=0.47\linewidth]{./figs/runtime/runtime_bar_emb.pdf}
\label{subfig:runtime_bar_emb}}
\caption{Runtime performance (lower is better) of \textsc{Trex}\xspace, SAFE, and Gemini on (a) function parsing and (b) embedding generation. The time is log-scaled.}
\label{fig:runtime}
\end{figure}
Figure~\ref{fig:runtime} shows that \textsc{Trex}\xspace is more efficient than the other tools in both function parsing and embedding generation for functions from 4 different projects with different number of functions (Table~\ref{tab:dataset}).
Gemini requires manually constructing control flow graph and extracting inter-/intra-basic-block feature engineering. It thus incurs the largest overhead.
For generating function embeddings, our underlying network architectures leverage Transformer self-attention layers, which is more amenable to parallezation with GPU than the recurrent (used by SAFE) and graph neural network (used by Gemini)~\cite{vaswani2017attention}.
As a result, \textsc{Trex}\xspace runs up to 8$\times$ faster than SAFE and Gemini.
\subsection{RQ4: Ablation Study}
\label{subsec:rq4}
In this section, we aim to quantify how much each key component in \textsc{Trex}\xspace's design helps the end results.
We first study how much does pretraining, argued to assist learning approximate execution semantics, help match function binaries.
We then study how does pretraining \emph{without micro-traces} affect the end results.
We also test how much does incorporating the micro-traces in the pretraining tasks improve the accuracy.
\vspace{.1cm}\noindent\textbf{Pretraining effectiveness.}
We compare the testing between AUC scores achieved by \textsc{Trex}\xspace (1) with pretraining (except the target project that will be finetuned), (2) with 66\% of pretraining functions in (1), (3) with 33\% of pretraining functions in (1), and (4) without pretraining (the function embedding is computed by randomly-initialized model weights that are not pretrained).
The function pairs can come from arbitrary architectures, optimizations, and obfuscations.
\begin{figure}[!t]
\centering
\includegraphics[scale=.4]{./figs/ablation/ablation-pretrain-effective-legend.pdf}
\includegraphics[width=.95\linewidth]{./figs/ablation/ablation-pretrain-effective.pdf}
\caption{Comparison of testing AUC scores between models pretrained with different fraction of the pretraining set.}
\label{fig:ablation-pretrain}
\end{figure}
Figure~\ref{fig:ablation-pretrain} shows that the model's AUC score drops significantly (on average 15.7\%) when the model is not pretrained.
Interestingly, we observe that the finetuned models achieve similar AUC scores, \emph{i.e., } with only 1\% decrease when pretrained with 33\% of the functions compared to pretraining with all functions.
This indicates that we can potentially decrease our pretraining data size to achieve roughly the same performance.
However, since our pretraining task does not require collect any label, we can still collect unlimited binary code found in the wild to enrich the semantics that can be observed.
\vspace{.1cm}\noindent\textbf{Pretraining w/o micro-traces.}
We try to understand whether including the micro-traces in pretraining can really help the model to learn better execution semantics than learning from only static assembly code, which in turn results in better function matching accuracy.
Specifically, we pretrain the model on the data that contains only dummy value sequence (see Section~\ref{sec:method}), and follow the same experiment setting as described above.
Besides replacing the input value sequence as dummy value, we accordingly remove the prediction of dynamic values in the pretraining objective (Equation~\ref{eq:pretrain}).
We also compare to SAFE in this case.
Figure~\ref{fig:ablation-pretrain-finetune-trace} shows that the AUC scores decrease by 7.2\% when the model is pretrained without micro-trace.
The AUC score of pretrained \textsc{Trex}\xspace without micro-traces is even 0.035 lower than that of SAFE.
However, the model still performs reasonably well, achieving 0.88 AUC scores even when the functions can come from arbitrary architectures, optimizations, and obfuscations.
Moreover, we observe that pretraining without micro-traces has less performance drop than the model simply not pretrained (7.2\% vs. 15.7\%).
This demonstrates that even pretraining with only static assembly code is indeed helpful to improve matching functions.
One possible interpretation is that similar functions are statically similar in syntax, while understanding their inherently similar execution semantics just further increases the similarity score.
\begin{figure}[!t]
\centering
\includegraphics[scale=.3]{./figs/ablation/ablation-pretrain-finetune-trace-legend.pdf}
\includegraphics[width=\linewidth]{./figs/ablation/ablation-pretrain-finetune-trace.pdf}
\caption{Comparison of testing AUC scores between models pretrained \emph{with} micro-trace and pretrained \emph{without} micro-trace.}
\label{fig:ablation-pretrain-finetune-trace}
\end{figure}
\vspace{.1cm}\noindent\textbf{Pretraining accuracy.}
We study the testing perplexity (PPL defined in Section~\ref{sec:impl}) of pretraining \textsc{Trex}\xspace to directly validate whether it indeed helps \textsc{Trex}\xspace to learn the approximate execution semantics.
The rationale is the good performance, \emph{i.e., } the low PPL, on unseen testing function binaries indicates that \textsc{Trex}\xspace highly likely learns to generalize based on its learned approximate execution semantics.
\begin{figure}[!t]
\centering
\includegraphics[width=0.85\linewidth]{./figs/pretrain/input-combine.pdf}
\caption{Testing PPL of pretraining \textsc{Trex}\xspace in 10 epochs. We compare different designs of combining byte-sequence in the micro-trace (see Section~\ref{subsec:input_repr}).}
\label{fig:pretrain-loss}
\end{figure}
The green line in Figure~\ref{fig:pretrain-loss} shows the validation PPL in 10 epochs of pretraining \textsc{Trex}\xspace.
The testing set is constructed by sampling 10,000 random functions from the projects used in pretraining (as described in Section~\ref{sec:impl}).
We observe that the PPL of \textsc{Trex}\xspace drops to close to 2.1, which is far below that of random guessing (\emph{e.g., } random guessing PPL is $2^{-\log(1/256)}=256$), indicating that \textsc{Trex}\xspace has around 0.48 confidence on average on the masked tokens being the correct value.
Note that a random guessing has only 1/256=0.004 confidence on the masked tokens being the correct value.
\section{Implementation and Experimental Setup}
\label{sec:impl}
We implement \textsc{Trex}\xspace using fairseq, a sequence modeling toolkit~\cite{ott2019fairseq}, based on PyTorch 1.6.0 with CUDA 10.2 and CUDNN 7.6.5.
We run all experiments on a Linux server running Ubuntu 18.04, with an Intel Xeon 6230 at 2.10GHz with 80 virtual cores including hyperthreading, 385GB RAM, and 8 Nvidia RTX 2080-Ti GPUs.
\vspace{.1cm}\noindent\textbf{Datasets.}
To train and evaluate \textsc{Trex}\xspace, we collect 13 popular open-source software projects.
These projects include Binutils-2.34, Coreutils-8.32, Curl-7.71.1, Diffutils-3.7, Findutils-4.7.0, GMP-6.2.0, ImageMagick-7.0.10, Libmicrohttpd-0.9.71, LibTomCrypt-1.18.2, OpenSSL-1.0.1f and OpenSSL-1.0.1u, PuTTy-0.74, SQLite-3.34.0, and Zlib-1.2.11.
We compile these projects into 4 architectures \emph{i.e., } x86, x64, ARM (32-bit), and MIPS (32-bit), with 4 optimization levels (OPT), \emph{i.e., } \texttt{O0}, \texttt{O1}, \texttt{O2}, and \texttt{O3}, using \texttt{GCC-7.5}.
Specifically, we compile the software projects based on its makefile, by specifying \texttt{CFLAGS} (to set optimization flag), \texttt{CC} (to set cross-compiler), and \texttt{--host} (to set the cross-compilation target architecture).
We always compile to dynamic shared objects, but resort to static linking when we encounter build errors.
We are able to compile all projects with these treatments.
We also obfuscate all projects using 5 types of obfuscations (OBF) by Hikari~\cite{hikari} on x64 -- an obfuscator based on \texttt{clang-8}.
The obfuscations include bogus control flow (\texttt{bcf}), control flow flattening (\texttt{cff}), register-based indirect branching (\texttt{ibr}), basic block splitting (\texttt{spl}), and instruction substitution (\texttt{sub}).
As we encounter several errors in cross-compilation using Hikari (based on Clang)~\cite{hikari}, and the baseline system (\emph{i.e., } Asm2Vec~\cite{ding2019asm2vec}) to which we compare only evaluates on x64, we restrict the obfuscated binaries for x64 only.
As a result, we have 1,472,066 functions, as shown in Table~\ref{tab:dataset}.
\begin{table*}[!t]
\footnotesize
\setlength{\tabcolsep}{2.2pt}
\centering
\renewcommand{\arraystretch}{1}
\setlength\aboverulesep{0.2pt}
\setlength\belowrulesep{0.2pt}
\caption{Number of functions for each project across 4 architectures with 4 optimization levels and 5 obfuscations.
}
\label{tab:dataset}
\rowcolors{1}{}{lightblue}
\begin{tabular}{c|c|l|l|l|l|l|l|l|l|l|l|l|l|l?{1.1pt}r}
\toprule[1.1pt]
\multirow{2}{*}{ARCH} & \multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}OPT\\ OBF\end{tabular}} & \multicolumn{14}{c}{\bf \# Functions} \\ \cline{3-16}
&
&
\multicolumn{1}{c|}{Binutils} &
\multicolumn{1}{c|}{Coreutils} &
\multicolumn{1}{c|}{Curl} &
\multicolumn{1}{c|}{Diffutils} &
\multicolumn{1}{c|}{Findutils} &
\multicolumn{1}{c|}{GMP} &
\multicolumn{1}{c|}{ImageMagick} &
\multicolumn{1}{c|}{Libmicrohttpd} &
\multicolumn{1}{c|}{LibTomCrypt} &
\multicolumn{1}{c|}{OpenSSL} &
\multicolumn{1}{c|}{PuTTy} &
\multicolumn{1}{c|}{SQLite} &
\multicolumn{1}{c?{1.1pt}}{Zlib} &
\multicolumn{1}{c}{\textbf{Total}} \\ \midrule[.9pt]
& \texttt{O0} & 25,492 & 19,992 & 1,067 & 944 & 1,529 & 766 & 2,938 & 200 & 779 & 11,918 & 7,087 & 2,283 & 157 & 75,152 \\ \cline{2-16}
& \texttt{O1} & 20,043 & 14,918 & 771 & 694 & 1,128 & 704 & 2,341 & 176 & 745 & 10,991 & 5,765 & 1,614 & 143 & 60,033 \\ \cline{2-16}
& \texttt{O2} & 19,493 & 14,778 & 765 & 693 & 1,108 & 701 & 2,358 & 171 & 745 & 11,001 & 5,756 & 1,473 & 138 & 59,180 \\ \cline{2-16}
& \texttt{O3} & 17,814 & 13,931 & 697 & 627 & 983 & 680 & 2,294 & 160 & 726 & 10,633 & 5,458 & 1,278 & 125 & 55,406 \\ \cline{2-16}
\multirow{-5}{*}{ARM} & \multicolumn{14}{c?{1.1pt}}{Total \# Functions of ARM} & \textbf{249,771} \\ \midrule[.5pt]
\multirow{5}{*}{MIPS} & \texttt{O0} & 28,460 & 18,843 & 1,042 & 906 & 1,463 & 734 & 2,840 & 200 & 779 & 11,866 & 7,003 & 2,199 & 153 & 76,488 \\ \cline{2-16}
& \texttt{O1} & 22,530 & 13,771 & 746 & 653 & 1,059 & 670 & 2,243 & 176 & 745 & 10,940 & 5,685 & 1,530 & 139 & 60,887 \\ \cline{2-16}
& \texttt{O2} & 22,004 & 13,647 & 741 & 653 & 1,039 & 667 & 2,260 & 171 & 743 & 10,952 & 5,677 & 1,392 & 135 & 60,081 \\ \cline{2-16}
& \texttt{O3} & 20,289 & 12,720 & 673 & 584 & 917 & 646 & 2,198 & 161 & 724 & 10,581 & 5,376 & 1,197 & 121 & 56,187 \\ \cline{2-16}
& \multicolumn{14}{c?{1.1pt}}{Total \# Functions of MIPS} & \textbf{253,643} \\ \midrule[.5pt]
& \texttt{O0} & 37,783 & 24,383 & 1,335 & 1,189 & 1,884 & 809 & 3,876 & 326 & 818 & 12,552 & 7,548 & 2,923 & 204 & 95,630 \\ \cline{2-16}
& \texttt{O1} & 32,263 & 20,079 & 1,013 & 967 & 1,516 & 741 & 3,482 & 280 & 782 & 11,578 & 6,171 & 2,248 & 196 & 81,316 \\ \cline{2-16}
& \texttt{O2} & 32,797 & 21,082 & 1,054 & 1,006 & 1,524 & 728 & 3,560 & 265 & 784 & 11,721 & 6,171 & 2,113 & 183 & 82,988 \\ \cline{2-16}
& \texttt{O3} & 34,055 & 22,482 & 1,020 & 1,052 & 1,445 & 707 & 3,597 & 284 & 760 & 11,771 & 5,892 & 1,930 & 197 & 85,192 \\ \cline{2-16}
\multirow{-5}{*}{x86} & \multicolumn{14}{c?{1.1pt}}{Total \# Functions of x86} & \textbf{358,261} \\ \midrule[.5pt]
\multirow{5}{*}{x64} & \texttt{O0} & 26,757 & 17,238 & 1,034 & 845 & 1,386 & 751 & 2,970 & 200 & 782 & 12,047 & 7,061 & 2,190 & 151 & 73,412 \\ \cline{2-16}
& \texttt{O1} & 21,447 & 12,532 & 739 & 600 & 1,000 & 691 & 2,358 & 176 & 745 & 11,120 & 5,728 & 1,523 & 137 & 58,796 \\ \cline{2-16}
& \texttt{O2} & 20,992 & 12,206 & 734 & 596 & 976 & 689 & 2,374 & 171 & 742 & 11,136 & 5,703 & 1,380 & 132 & 57,831 \\ \cline{2-16}
& \texttt{O3} & 19,491 & 11,488 & 662 & 536 & 857 & 667 & 2,308 & 160 & 725 & 10,768 & 5,390 & 1,183 & 119 & 54,354 \\ \cline{2-16}
& \multicolumn{14}{c?{1.1pt}}{Total \# Functions of x64} & \textbf{244,393} \\ \midrule[1.1pt]
& \texttt{bcf} & 27,734 & 17,093 & 998 & 840 & 1,388 & 746 & 2,833 & 200 & 782 & 10,768 & 7,069 & 2,183 & 151 & 72,785 \\ \cline{2-16}
& \texttt{cff} & 27,734 & 17,093 & 998 & 840 & 1,388 & 746 & 2,833 & 200 & 782 & 10,903 & 7,069 & 2,183 & 151 & 72,920 \\ \cline{2-16}
& \texttt{ibr} & 27,734 & 17,105 & 998 & 842 & 1,392 & 746 & 2,833 & 204 & 782 & 12,045 & 7,069 & 2,183 & 151 & 74,084 \\ \cline{2-16}
& \texttt{spl} & 27,734 & 17,093 & 998 & 840 & 1,388 & 746 & 2,833 & 200 & 782 & 10,772 & 7,069 & 2,183 & 151 & 72,789 \\ \cline{2-16}
& \texttt{sub} & 27,734 & 17,093 & 998 & 840 & 1,388 & 746 & 2,833 & 200 & 782 & 11,403 & 7,069 & 2,183 & 151 & 73,420 \\ \cline{2-16}
\multirow{-5}{*}{x64} & \multicolumn{14}{c?{1.1pt}}{Total \# Obfuscated Functions} & \textbf{365,998} \\ \midrule[1.1pt]
\multicolumn{15}{c?{1.1pt}}{Total \# Functions} & \textbf{1,472,066} \\ \bottomrule[1.1pt]
\end{tabular}
\end{table*}
\vspace{.1cm}\noindent\textbf{Micro-tracing.}
We implement micro-tracing by Unicorn~\cite{quynh2015unicorn}, a cross-platform CPU emulator based on QEMU~\cite{bellard2005qemu}.
We micro-execute each function 3 times with different initialized registers and memory, generating 3 micro-traces (including both static code and dynamic values) for pretraining (masked LM).
We leverage multi-processing to parallelize micro-executing each function and set 30 seconds as the timeout for each run in case any instruction gets stuck (\emph{i.e., } infinite loops).
For each function (with 3 micro-traces), we append one additional dummy trace, which consists of only dummy values (\texttt{\#\#}).
This setting encourages the model to leverage its learned execution semantics (from other traces with concrete dynamic values) to predict the masked code with \emph{only} code context, which helps the finetuning task as we only feed the functions' static code as input (discussed in Section~\ref{sec:intro}).
\vspace{.1cm}\noindent\textbf{Baselines.}
For comparing cross-architecture performance, we consider 2 state-of-the-art systems.
The first one is SAFE~\cite{massarelli2019safe} that achieves state-of-the-art function matching accuracy.
As SAFE's trained model is publicly available, we compare \textsc{Trex}\xspace with both SAFE's reported results on their dataset, \emph{i.e., } OpenSSL-1.0.1f and OpenSSL-1.0.1u, and running their trained models on all of our collected binaries.
The second baseline is Gemini~\cite{xu2017neural}, which is shown outperformed by SAFE.
As Gemini does not release their trained models, we compare our results to their reported numbers directly on their evaluated dataset, \emph{i.e., } OpenSSL-1.0.1f and OpenSSL-1.0.1u.
For cross-optimization/obfuscation comparison, we consider Asm2Vec~\cite{ding2019asm2vec} and Blex~\cite{egele2014blanket} as the baselines.
Asm2Vec achieves the state-of-the-art cross-optimization/obfuscation results, based on learned embeddings from static assembly code.
Blex, on the other hand, leverages functions' dynamic behavior to match function binaries.
As we only find a third-party implementation of Asm2Vec that achieves extremely low Precision@1 (the metric used in Asm2Vec) from our testing (\emph{e.g., } 0.02 vs. their reported 0.814), and we have contacted the authors and do not get replies, we directly compare to their reported numbers.
Blex is not publicly available either, so we also compare to their reported numbers directly.
\vspace{.1cm}\noindent\textbf{Metrics.}
As the cosine similarity between two function embeddings can be an arbitrary real value between -1 and 1, there has to be a mechanism to threshold the similarity score to determine whether the function pairs are similar or dissimilar.
The chosen thresholds largely determine the model's predictive performance~\cite{pagani2018beyond}.
To avoid the potential bias introduced by a specific threshold value, we consider the receiver operating characteristic (ROC) curve, which measures the model's false positives/true positives under different thresholds.
Notably, we use the area under curve (AUC) of the ROC curve to quantify the accuracy of \textsc{Trex}\xspace to facilitate benchmarking -- the higher the AUC score, the better the model's accuracy.
Certain baselines do not use AUC score to evaluate their system. For example, Asm2Vec uses Precision at Position 1 (Precision@1), and Blex uses the number of matched functions as the metric.
Therefore, we also include these metrics to evaluate \textsc{Trex}\xspace when needed.
Specifically, given a sequence of query functions and the target functions to be matched, Precision@1 measures the percentage of matched query functions in the target functions.
Here ``match'' means the query function should have the \emph{top-1 highest} similarity score with the ground truth function among the target functions.
To evaluate pretraining performance, we use the standard metric for evaluating the language model -- perplexity (PPL). The lower the PPL the better the pretrained model in predicting masked code.
The lowest PPL is 1.
\vspace{.1cm}\noindent\textbf{Pretraining setup.}
To strictly separate the functions in pretraining, finetuning, and testing, we pretrain \textsc{Trex}\xspace on all functions in the dataset, \emph{except the functions in the project going to be finetuned and evaluated for similarity.}
For example, the function matching results of Binutils (Table~\ref{tab:result}) are finetuned on the model pretrained on all other projects without Binutils.
Note that pretraining \emph{is agnostic to any labels} (\emph{e.g., } ground-truth indicating similar functions).
Therefore, we can always pretrain on large-scale codebases, which can potentially include the functions for finetuning (this is the common practice in transfer learning~\cite{devlin2018bert}).
It is thus worth noting that our setup of separating functions for pretraining and finetuning makes our testing significantly more challenging.
We keep the pretrained model weights that achieve the best validation PPL for finetuning.
The validation set for pretraining consists of 10,000 random functions selected from Table~\ref{tab:dataset}.
\vspace{.1cm}\noindent\textbf{Finetuning setup.}
We choose 50,000 random function pairs for each project and randomly select \emph{only 10\%} for training, and the remaining is used as the testing set.
We keep the training and testing functions \emph{strictly non-duplicated} by ensuring the functions that appear in training function pairs not appear in the testing.
As opposed to the typical train-test split (\emph{e.g., } 80\% training and 20\% testing~\cite{massarelli2019safe}), our setting requires the model to generalize from few training samples to a large number of unseen testing samples, which alleviates the possibility of overfitting.
Moreover, we keep the ratio between similar and dissimilar function pairs in the finetuning set as roughly 1:5.
This setting follows the practice of contrastive learning~\cite{saunshi2019theoretical,chen2020simple}, respecting the actual distribution of similar/dissimilar functions as the number of dissimilar functions is often larger than that of similar functions in practice.
\vspace{.1cm}\noindent\textbf{Hyperparameters.}
We pretrain and finetune the models for 10 epochs and 30 epochs, respectively. We choose $\alpha=0.125$ in Equation~\ref{eq:pretrain} such that the cross-entropy loss of code prediction and value prediction have the same weight.
We pick $\xi=0.1$ in Equation~\ref{eq:cosembloss} to make the model slightly inclined to treat functions as dissimilar because functions in practice are mostly dissimilar.
We fix the largest input length to be 512 and split the functions longer than this length into subsequences for pretraining.
We average the subsequences' embeddings during finetuning if the function is split to more than one subsequences.
In this paper, we keep most of the hyperparameters fixed throughout the evaluation if not mentioned explicitly (complete hyperparameters are defined in Appendix~\ref{subsec:hyperparm}).
While we can always search for better hyperparameters, there is no principled method to date~\cite{bergstra2012random}.
We thus leave as future work a more thorough study of \textsc{Trex}\xspace's hyperparameters.
\section{Introduction}
\label{sec:intro}
Semantic function similarity, which quantifies the behavioral similarity between two functions, is a fundamental program analysis capability with a broad spectrum of real-world security usages, such as vulnerability detection~\cite{brumley2008automatic}, exploit generation~\cite{avgerinos2014automatic}, tracing malware lineage~\cite{jang2013towards, bayer2009scalable}, and forensics~\cite{luo2014semantics}.
For example, OWASP lists ``using components with known vulnerabilities'' as one of the top-10 application security risks in 2020~\cite{OWASP}.
Therefore, identifying similar vulnerable functions in massive software projects can save significant manual effort.
When matching semantically similar functions for security-critical applications (\emph{e.g., } vulnerability discovery), we often have to deal with software at \emph{binary level}, such as commercial off-the-shelf products (\emph{i.e., } firmware images) and legacy programs.
However, this task is challenging as the functions' high-level information (\emph{i.e., } data structure definitions) are removed during the compilation process.
The problem aggravates when the functions are compiled to run on different instruction set architectures with various compiler optimizations or even simple obfuscations.
There has been a growing interest in using Machine Learning (ML) to tackle these challenges~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec}.
Notably, ML models learn function representations (\emph{i.e., } embeddings) from instructions in function binaries and match the functions based on the learned representation.
These approaches have obtained the state-of-the-art results~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec}, outperforming the traditional signature-based approaches~\cite{bindiff} that rely on the hand-coded features (\emph{e.g., } number of basic blocks) extracted by domain experts.
The ML-based approaches are also amenable to large-scale function matching as computing similarity score as embedding distance is exceptionally efficient (\emph{e.g., } taking around 0.1 seconds searching over one million functions)~\cite{feng2016scalable}.
\vspace{0.01cm}\noindent\textbf{Execution semantics.}
Despite the impressive progress, it remains challenging for these approaches to match semantically similar functions with disparate syntax and structure~\cite{mckee2019software}.
An inherent cause is that the code semantics is characterized by \emph{its execution effects}.
However, all existing learning-based approaches are \emph{agnostic to program execution semantics}, taking the static code as the only input.
Such a setting can easily lead a model into learning simple pattern matching, limiting their accuracy when such patterns are absent or changed~\cite{payer2014similarity, aghakhani2020malware}.
Take a pair of x86 instructions as an example. \texttt{mov eax,2;lea ecx,[eax+4]} are semantically equivalent to \texttt{mov eax,2;lea ecx,[eax+eax*2]}.
To determine their similarity statically, the most salient pattern is their common substring (both sequences share the tokens \texttt{mov}, \texttt{eax}, \texttt{lea}, \texttt{ecx}), which does not encode the key reason of equivalence.
Without grasping the execution semantics, an ML model can easily learn such spurious patterns without understanding the inherent cause of the equivalence: \texttt{[eax+eax*2]} computes the same exact address as \texttt{[eax+4]} when \texttt{eax} is 2.
\vspace{0.01cm}\noindent\textbf{Limitations of existing dynamic approaches.}
Conceptually, dynamic analysis captures the perfect execution semantics.
However, traditional dynamic analysis is inappropriate here as it is hard to find a concrete program input exercising a certain function of interest.
While symbolic execution can help, it is known for poor scalability to large programs.
Therefore, existing dynamic approaches~\cite{egele2014blanket} \emph{relax} the function initial states with random values that are often under-constrained -- they may not be feasible from executing the program regularly.
For example, a function \texttt{bar()} can only be executed when called by function \texttt{foo()}.
When directly executing \texttt{bar()}, the constraints of \texttt{bar()}'s initial states imposed in \texttt{foo()} is ignored.
While these approaches can directly execute individual functions and compare their execution behavior, such a relaxation introduces \emph{noisy and incomplete traces}.
For example, for functions taking only values from specified range, \emph{e.g., } $\sqrt{x}$ and $arcsin(x)$, the randomized function inputs (\emph{e.g., } negative numbers for $\sqrt{x}$ and numbers outside $[-1,1]$ for $arcsin(x)$) often trigger the similar behavior such as the exception handling.
Therefore, existing dynamic approaches suffer from false positives~\cite{ding2019asm2vec}.
\vspace{0.01cm}\noindent\textbf{Our approach.}
This paper presents \textsc{Trex}\xspace (\emph{TRansfer-learning EXecution semantics}) that matches semantically similar functions based on their static instructions but augmented with execution semantics.
Instead of using noisy and incomplete dynamic trace to directly determine similarity, \textsc{Trex}\xspace first pretrains the model to observe and understand how each instruction executes in a function.
It then \emph{transfers} and \emph{composes} the learned knowledge from pretrained model to match semantically similar functions.
Our extensive experiments suggest that the learned knowledge of executions in pretraining significantly boosts the performance of matching semantically similar function binaries -- \textsc{Trex}\xspace excels in matching functions from different architectures, optimizations, and obfuscations.
\begin{figure*}[!t]
\centering
\includegraphics[width=\linewidth]{./figs/general.pdf}
\caption{
The workflow of \textsc{Trex}\xspace. We first pretrain the model on the function code and its micro-traces (\emph{i.e., } dynamic register values) using the masked LM task.
We then finetune the pretrained model on the function pairs (\emph{e.g., } same source function compiled to different architectures with different optimizations/obfuscations) by stacking new neural network layers for function similarity tasks.
Finetuning updates both the pretrained model and the stacked layers.
During inference, the finetuned model computes the function embedding, whose distance encodes the function similarity.}
\label{fig:general}
\end{figure*}
Our key insight is that the function's behavior can be \emph{composed} by each instruction's execution semantics, which can be learned by observing and understanding how each instruction executes in the dynamic traces.
For example, given instructions \texttt{mov eax,2;sub eax,1;add eax,3} and their dynamic traces, we can train the model to predict the value of \texttt{eax} after executing these instructions.
This training objective is designed to force the model to understand the execution semantics of \texttt{mov}, \texttt{sub}, and \texttt{add}, \emph{i.e., } how they manipulate the operands and updates the results, and learn to \emph{compose their execution effects} to understand their overall behavior.
We can then \emph{transfer} the model's learned knowledge to compose instructions' execution semantics to match semantically similar functions.
As a result, our approach is not constrained by noisy and incomplete traces -- it matches functions based \emph{only} on their \emph{complete} static instructions but \emph{augmented} with rich knowledge of execution semantics.
Such a design also saves significant runtime overhead as we do not need to execute any function on-the-fly when matching them~\cite{kapravelos2013revolver}.
Given a sequence of assembly instructions and their micro-traces (\emph{i.e., } register values), we pretrain the model with the masked language modeling (masked LM) task.
Notably, it masks random parts in the sequence and asks the model to predict masked parts based on their context.
This design forces the model to learn how a function executes to correctly infer the missing values.
It thus \emph{automates} learning execution semantics without any manual feature engineering.
It also enables learning more \emph{comprehensive} execution semantics of diverse instructions, as we can mask various parts of the function, imposing the model to learn different aspects of execution semantics to make correct prediction.
Moreover, this task is fully \emph{unsupervised}.
Therefore, \textsc{Trex}\xspace can be trained and further improved using arbitrary binaries found in the wild.
To encode both the static and dynamic information as the model inputs, we design a hierarchical Transformer~\cite{vaswani2017attention} that supports modeling concrete micro-trace values.
By contrast, existing approaches often treat the numerical values as a dummy token~\cite{massarelli2019safe, ding2019asm2vec} to avoid prohibitively large vocabulary size, which cannot effectively learn the rich dependencies between concrete values that likely encode key function semantics.
Moreover, our architecture's self-attention layer is designed to model long-range dependencies in a sequence~\cite{vaswani2017attention} efficiently.
Therefore, \textsc{Trex}\xspace can support roughly 170$\times$ longer sequence and runs 8$\times$ faster than existing neural architectures, essential to learning embeddings of long function execution traces.
We evaluate \textsc{Trex}\xspace on 1,472,066 functions collected from 13 popular open-source software projects across 4 architectures (x86, x64, ARM, and MIPS) and compiled with 4 optimizations (\texttt{O0}-\texttt{O3}), and 5 obfuscation strategies~\cite{hikari}.
\textsc{Trex}\xspace outperforms the state-of-the-art systems by 7.8\%, 7.2\%, and 14.3\% in matching functions across different architectures, optimizations, and obfuscations, respectively.
Our ablation studies show that the pretraining task significantly improves the accuracy of matching semantically similar functions (by 15.7\%).
We also apply \textsc{Trex}\xspace in searching vulnerable functions in 180 real-world firmware images developed by well-known vendors and deployed in diverse embedded systems, including WLAN routers, smart cameras, and solar panels.
Our case study shows that \textsc{Trex}\xspace assists in finding 16 real-world CVEs in these firmware images, which have not been disclosed in any previous studies.
We make the following contributions.
\begin{itemize}
\item We propose a new approach to matching semantically similar functions: we first train the model to learn program execution semantics and then transfer the learned knowledge to match semantically similar functions.
\item We implement micro-execution to collect dynamic traces on different architectures. We then develop a novel neural architecture -- hierarchical Transformer -- that can efficiently learn execution semantics from micro-traces.
\item We implement \textsc{Trex}\xspace and evaluate it on 1,472,066 functions from 13 popular software projects and libraries. \textsc{Trex}\xspace outperforms the state-of-the-art tools by 7.8\%, 7\%, and 14.3\%, in cross-architecture, optimization, and obfuscation function matching, respectively, while running up to 8$\times$ faster.
Moreover, \textsc{Trex}\xspace helps uncover 16 vulnerabilities in 180 real-world firmware images with latest version that are not disclosed by previous studies.
We release \textsc{Trex}\xspace at \url{anonymous.4open.science/r/644f4ceb-ab22-4f04-a91c-0565e091ca31}.
\end{itemize}
\section{Introduction}
\label{sec:intro}
Semantic function similarity, which quantifies the behavioral similarity between two functions, is a fundamental program analysis capability with a broad spectrum of real-world security usages, such as vulnerability detection~\cite{brumley2008automatic}, exploit generation~\cite{avgerinos2014automatic}, tracing malware lineage~\cite{jang2013towards, bayer2009scalable}, and forensics~\cite{luo2014semantics}.
For example, OWASP lists ``using components with known vulnerabilities'' as one of the top-10 application security risks in 2020~\cite{OWASP}.
Therefore, identifying similar vulnerable functions in massive software projects can save significant manual effort.
When matching semantically similar functions for security-critical applications (\emph{e.g., } vulnerability discovery), we often have to deal with software at \emph{binary level}, such as commercial off-the-shelf products (\emph{i.e., } firmware images) and legacy programs.
However, this task is challenging, as the functions' high-level information (\emph{e.g., } data structure definitions) are removed during the compilation process.
Establishing semantic similarity gets even harder when the functions are compiled to run on different instruction set architectures with various compiler optimizations or obfuscated with simple transformations.
Recently, Machine Learning (ML) based approaches have shown promise in tackling these challenges~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec} by learning robust features that can identify similar function binaries across different architectures, compiler optimizations, or even some types of obfuscation. Specifically, ML models learn function representations (\emph{i.e., } embeddings) from function binaries and use the distance between the embeddings of two functions to compute their similarity. The smaller the distance, the more similar the functions are to each other. Such approaches have achieved state-of-the-art results~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec}, outperforming the traditional signature-based methods~\cite{bindiff} using hand-crafted features (\emph{e.g., } number of basic blocks). Such embedding distance-based strategy is particularly appealing for large-scale function matching---taking only around 0.1 seconds searching over one million functions~\cite{feng2016scalable}.
\vspace{.1cm}\noindent\textbf{Execution semantics.}
Despite the impressive progress, it remains challenging for these approaches to match semantically similar functions with disparate syntax and structure~\cite{mckee2019software}.
An inherent cause is that the code semantics is characterized by \emph{its execution effects}.
However, all existing learning-based approaches are \emph{agnostic to program execution semantics}, training only on the static code.
Such a setting can easily lead a model into matching simple patterns, limiting their accuracy when such spurious patterns are absent or changed~\cite{payer2014similarity, aghakhani2020malware}.
For instance, consider the following pair of x86 instructions: \texttt{mov eax,2;lea ecx,[eax+4]} are semantically equivalent to \texttt{mov eax,2;lea ecx,[eax+eax*2]}.
An ML model focusing on syntactic features might pick common substrings (both sequences share the tokens \texttt{mov}, \texttt{eax}, \texttt{lea}, \texttt{ecx}) to establish their similarity, which does not encode the key reason of the semantic equivalence.
Without grasping the approximate execution semantics, an ML model can easily learn such spurious patterns without understanding the inherent cause of the equivalence: \texttt{[eax+eax*2]} computes the same exact address as \texttt{[eax+4]} when \texttt{eax} is 2.
\vspace{.1cm}\noindent\textbf{Limitations of existing dynamic approaches.}
Existing dynamic approaches try to avoid the issues described above by directly comparing the dynamic behaviors of functions to determine similarity.
As finding program inputs reaching the target functions is an extremely challenging and time-consuming task, the prior works perform under-constrained dynamic execution by initializing the function input states (\emph{e.g., } registers, memory) with random values and executing the target functions directly~\cite{egele2014blanket}.
Unfortunately, using such under-constrained execution traces directly to compute function similarities often result in many false positives~\cite{ding2019asm2vec}. For example, providing random inputs to two different functions with strict input checks might always trigger similar shallow exception handling codes and might look spuriously similar.
\vspace{.1cm}\noindent\textbf{Our approach.}
This paper presents \textsc{Trex}\xspace (\emph{TRansfer-learning EXecution semantics}) that trains ML models to learn the approximate execution semantics from under-constrained dynamic traces.
Unlike prior works, which use such traces to directly measure similarity, \textsc{Trex}\xspace pretrains the model on diverse traces to learn each instruction's execution effect in its context.
\textsc{Trex}\xspace then finetunes the model by \emph{transferring} the learned knowledge from pretraining to match semantically similar functions (see Figure~\ref{fig:general}).
Our extensive experiments suggest that the approximately learned knowledge of execution semantics in pretraining significantly boosts the accuracy of matching semantically similar function binaries -- \textsc{Trex}\xspace excels in matching functions from different architectures, optimizations, and obfuscations.
\begin{figure*}[!t]
\centering
\includegraphics[width=\linewidth]{./figs/general.pdf}
\caption{
The workflow of \textsc{Trex}\xspace. We first pretrain the model on the functions' micro-traces, consisting of both instructions and dynamic values, using the masked LM task.
We then finetune the pretrained model on the semantically similar function (only static instruction) pairs by stacking new neural network layers for function similarity tasks.
Finetuning updates both the pretrained model and the stacked layers.
During inference, the finetuned model computes the function embedding, whose distance encodes the function similarity.}
\label{fig:general}
\end{figure*}
Our key observation is that while under-constrained dynamic execution traces tend to contain many infeasible states, they still encode precise execution effects of many individual instructions.
Thus, we can train an ML model to observe and learn the effect of different instructions present across a large number of under-constrained dynamic traces collected from diverse functions.
Once the model has gained an approximate understanding of execution semantics of various instructions, we can train it to match semantically similar functions by leveraging its learned knowledge.
As a result, during inference, we do not need to execute any functions on-the-fly while matching them~\cite{kapravelos2013revolver}, which saves significant runtime overhead.
Moreover, our trained model does not need the under-constrained dynamic traces to match functions, it only uses the function instructions, but they are \emph{augmented} with rich knowledge of execution semantics.
In this paper, we extend micro-execution~\cite{godefroid2014micro}, a form of under-constrained dynamic execution, to generate \emph{micro-traces} of a function across multiple instruction set architectures.
A micro-trace consists of a sequence of aligned instructions and their corresponding program state values.
We pretrain the model on a large number of micro-traces gathered from diverse functions as part of training data using the masked language modeling (masked LM) task.
Notably, masked LM masks random parts in the sequence and asks the model to predict masked parts based on their context.
This design forces the model to learn approximately how a function executes to correctly infer the missing values, which automates learning execution semantics without manual feature engineering.
Masked LM is also fully \emph{self-supervised}~\cite{devlin2018bert} -- \textsc{Trex}\xspace can thus be trained and further improved with arbitrary functions found in the wild.
To this end, we design a hierarchical Transformer~\cite{vaswani2017attention} that supports learning approximate execution semantics.
Specifically, our architecture models micro-trace values explicitly.
By contrast, existing approaches often treat the numerical values as a dummy token~\cite{massarelli2019safe, ding2019asm2vec} to avoid prohibitively large vocabulary size, which cannot effectively learn the rich dependencies between concrete values that likely encode key function semantics.
Moreover, our architecture's self-attention layer is designed to model long-range dependencies in a sequence~\cite{vaswani2017attention} efficiently.
Therefore, \textsc{Trex}\xspace can support roughly 170$\times$ longer sequence and runs 8$\times$ faster than existing neural architectures, essential to learning embeddings of long function execution traces.
We evaluate \textsc{Trex}\xspace on 1,472,066 functions collected from 13 popular open-source software projects across 4 architectures (x86, x64, ARM, and MIPS) and compiled with 4 optimizations (\texttt{O0}-\texttt{O3}), and 5 obfuscation strategies~\cite{hikari}.
\textsc{Trex}\xspace outperforms the state-of-the-art systems by 7.8\%, 7.2\%, and 14.3\% in matching functions across different architectures, optimizations, and obfuscations, respectively.
Our ablation studies show that the pretraining task significantly improves the accuracy of matching semantically similar functions (by 15.7\%).
We also apply \textsc{Trex}\xspace in searching vulnerable functions in 180 real-world firmware images developed by well-known vendors and deployed in diverse embedded systems, including WLAN routers, smart cameras, and solar panels.
Our case study shows that \textsc{Trex}\xspace helps find 16 CVEs in these firmware images, which have not been disclosed in previous studies.
We make the following contributions.
\begin{itemize}
\item We propose a new approach to matching semantically similar functions: we first train the model to learn approximate program execution semantics from micro-traces, a form of under-constrained dynamic traces, and then transfer the learned knowledge to identify semantically similar functions.
\item We extend micro-execution to support different architectures to collect micro-traces for training. We then develop a novel neural architecture -- hierarchical Transformer -- to learn approximate execution semantics from micro-traces.
\item We implement \textsc{Trex}\xspace and evaluate it on 1,472,066 functions from 13 popular software projects and libraries. \textsc{Trex}\xspace outperforms the state-of-the-art tools by 7.8\%, 7\%, and 14.3\%, in cross-architecture, optimization, and obfuscation function matching, respectively, while running up to 8$\times$ faster.
Moreover, \textsc{Trex}\xspace helps uncover 16 vulnerabilities in 180 real-world firmware images with the latest version that are not disclosed by previous studies.
We release the code and dataset of \textsc{Trex}\xspace at \url{https://github.com/CUMLSec/trex}.
\end{itemize}
\section{Methodology}
\label{sec:method}
This section describes \textsc{Trex}\xspace's design specifics, including our micro-tracing semantics, our learning architecture's details, and pretraining and finetuning workflow.
\subsection{Micro-tracing Semantics}
\label{subsec:microx-semantics}
We implement micro-execution by Godefroid~\cite{godefroid2014micro} to handle x64, ARM, and MIPS, where the original paper only describes x86 as the use case.
In the following, we briefly explain how we micro-execute an individual function binary, highlighting the key algorithms in handling different types of instructions.
\vspace{.1cm}\noindent\textbf{IR Language.}
To abstract away the complexity of different architectures' assembly syntax, we introduce a low-level intermediate representation (IR) that models function assembly code.
We only include a subset of the language specifics to illustrate the implementation algorithm.
Figure~\ref{fig:microx-syntax} shows the grammar of the IR.
Note that the IR here only serves to facilitate the discussion of our micro-tracing implementation.
In our implementation, we use real assembly instructions and tokenize them as model’s input (Section~\ref{subsec:input_repr}).
Notably, we denote memory reads and writes by \texttt{load($e$)} and \texttt{store($e_{v},e_{a}$)} (\emph{i.e., } store the value expression $e_v$ to address expression $e_a$), which generalize from both the load-store architecture (\emph{i.e., } ARM, MIPS) and register-memory architecture (\emph{i.e., } x86).
Both operations can take as input $e$ -- an expression that can be an explicit hexadecimal number (denoting the address or a constant), a register, or a result of an operation on two registers.
We use \texttt{jmp} to denote the general jump instruction, which can be both direct or indirect jump (\emph{i.e., } the expression $e_a$ can be a constant $c$ or a register $r$).
The jump instruction can also be unconditional or conditional. Therefore, the first parameter in \texttt{jmp} is the conditional expression $e_c$ and unconditional jump will set $e_c$ to \texttt{true}.
We represent function invocations and returns by \texttt{call} and \texttt{ret}, where \texttt{call} is parameterized by an expression, which can be an address (direct call) or a register (indirect call).
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{./figs/microx-semantics/language.pdf}
\caption{Low-level IR for representing assembly code. The IR abstracts away the actual assembly syntax that are disparate across different architectures.}
\label{fig:microx-syntax}
\end{figure}
\vspace{.1cm}\noindent\textbf{Micro-tracing algorithm.}
Algorithm~\ref{alg:microx} outlines the basic steps of micro-tracing a given function $f$.
First, it initializes the memory to load the code and the corresponding stack.
It then initializes all registers except the special-purpose register, such as the stack pointer or the program counter.
Then it starts linearly executing instructions of $f$.
We map the memory address \emph{on-demand} if the instruction access the memory (\emph{i.e., } read/write).
If the instruction reads from memory, we further initialize a random value in the specific memory addresses.
For call/jump instructions, we first examine the target address and skip the invalid jump/call, known as ``forced execution''~\cite{peng2014x}.
By skipping unreachable jumps and calls, it can keep executing the function till the end of the function and exposes more behaviors, \emph{e.g., } skipping potential input check exceptions.
Since the \texttt{nop} instructions can serve as padding between instructions within a function, we simply skip \texttt{nop}.
We terminate the micro-tracing when it finishes executing all instructions, reaches \texttt{ret}, or times out.
Figure~\ref{fig:arithmetic} and~\ref{fig:stack} demonstrate sample micro-traces of real-world functions.
\subsection{Input Representation}
\label{subsec:input_repr}
Formally, given a function $f$ (\emph{i.e., } assembly code) and its micro-trace $t$ (by micro-executing $f$), we prepare the model input $x$, consisting of 5 types of token sequence with the same size $n$.
Figure~\ref{fig:arch} shows the model input example and how they are masked and processed by the hierarchical Transformer to predict the corresponding output as a pretraining task.
\input{microx}
\vspace{.1cm}\noindent\textbf{Micro-trace code sequence.}
The first sequence $x_f$ is the assembly code sequence: $x_f=\{mov, eax, +, ...\}^n$, generated by tokenizing the assembly instructions in the micro-trace.
We treat all symbols appear in the assembly instructions as tokens.
Such a tokenization aims to preserve the critical hint of the syntax and semantics of the assembly instructions.
For example, we consider even punctuation to be one of the tokens, \emph{e.g., } ``,'', ``['', ``]'', as ``,'' implies the token before and after it as destination and source of \texttt{mov} (in Intel syntax), respectively, and ``['' and ``]'' denote taking the address of the operands reside in between them.
We take special treatment of numerical values appear in the assembly code.
Treating numerical values as regular text tokens can incur prohibitively large vocabulary size, \emph{e.g., } $2^{32}$ number of possibilities on 32-bit architectures.
To avoid this problem, we move all numeric values to the micro-trace value sequence (that will be learned by an additional neural network as detailed in the following) and replace them with a special token \texttt{num} (\emph{e.g., } last token of input in Figure~\ref{fig:arch}).
With all these preprocessing steps, the vocabulary size of $x_f$ across all architectures is 3,300.
\vspace{.1cm}\noindent\textbf{Micro-trace value sequence.}
The second sequence $x_t$ is the micro-trace value sequence, where each token in $x_t$ is the dynamic value from micro-tracing the corresponding code.
As discussed in Section~\ref{sec:overview}, we keep \emph{explicit} values (instead of a dummy value used by existing approaches) in $x_t$.
Notably, we use the dynamic value for each token (\emph{e.g., } register) in an instruction before it is executed.
For example, in \texttt{mov eax,0x8; mov eax,0x3}, the dynamic value of the second \texttt{eax} is \texttt{0x8}, as we take the value of \texttt{eax} before \texttt{mov eax,0x3} is executed.
For code token without dynamic value, \emph{e.g., } \texttt{mov}, we use dummy values (see below).
\vspace{.1cm}\noindent\textbf{Position sequences.}
The position of each code and value token is critical for inferring binary semantics. Unlike natural language, where swapping two words can roughly preserve the same semantic meaning, swapping two operands can significantly change the instructions.
To encode the inductive bias of position into our model, we introduce \emph{instruction position sequence} $x_c$ and \emph{opcode/operand position sequence} $x_o$ to represent the relative positions between the instructions and within each instruction.
As shown in Figure~\ref{fig:arch}, $x_c$ is a sequence of integers encoding the position of each instruction. All opcodes/operands within a single instruction share the same value.
$x_o$ is a sequence of integers encoding the position of each opcode and operands within a single instruction.
\vspace{.1cm}\noindent\textbf{Architecture sequence.}
Finally, we feed the model with an extra sequence $x_a$, describing the input binary's instruction set architecture. The vocabulary of $x_a$ consists of 4 architectures: $x_a=\{$x86, x64, ARM, MIPS$\}^n$.
This setting helps the model to distinguish between the syntax of different architecture.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{./figs/arch-pretrain.pdf}
\caption{Model architecture with input-output example. The masked input is marked in gray. In pretraining, the loss function consists of the cross-entropy losses of both (1) code prediction and (2) value prediction, where the value prediction consists of predicting each of the 8 bytes of the micro-trace value.}
\label{fig:arch}
\end{figure}
\vspace{.1cm}\noindent\textbf{Encoding numeric values.}
As mentioned above, treating concrete values as independent tokens can lead to prohibitively large vocabulary size.
We design a \emph{hierarchical input encoding scheme} to address this challenge.
Specifically, let $x_{t_i}$ denote the $i$-th value in $x_t$.
We represent $x_{t_i}$ as an (padded) 8-byte fixed-length byte sequence $x_{t_i}=$\{\texttt{0x00}, ..., \texttt{0xff}\}$^8$ ordered in Big-Endian.
We then feed $x_{t_i}$ to a 2-layer bidirectional LSTM (bi-LSTM) and take its last hidden cell's embedding as the value representation $t_i=bi$-$LSTM(x_{t_i})$.
Here $t_i$ denotes the output of applying the embedding to $x_{t_i}$.
To make the micro-trace code tokens without dynamic values (\emph{e.g., } opcode) align with the byte sequence, we use a dummy sequence (\texttt{\#\#}) with the same length.
Figure~\ref{fig:arch} (right-hand side) illustrates how bi-LSTM takes the byte sequence and computes the embedding.
Such a design significantly reduces the vocabulary size as now we have only 256 possible byte values to encode.
Moreover, bi-LSTM encodes the potential dependencies between high and low bytes within a single value.
This setting thus supports learning better relationships between different dynamic values (\emph{e.g., } memory region and offset) as opposed to treating all values as dummy tokens~\cite{shi2019learning}.
\subsection{Pretraining with Micro-traces}
\label{subsec:pretrain_method}
This section describes the pretraining task on the function binaries and their micro-traces, focusing on how the model masks its input $x$ (5 sequences) and predicts the masked part to learn the approximate execution semantics.
\vspace{.1cm}\noindent\textbf{Learned embeddings.}
We embed each token in the 5 sequences with the same embedding dimension $d_{emb}$ so that they can be summed as a single sequence as input to the Transformer.
Specifically, let $E_f(x_f)$, $E_t(x_t)$, $E_c(x_c)$, $E_o(x_o)$, $E_a(x_a)$ denote applying the embedding to the tokens in each sequence, respectively.
We have $E_{i}$, the embedding of $x_i$:
\begin{equation*}
E_{i} = E_f(x_{f_i})+E_t(x_{t_i})+E_c(x_{c_i})+E_o(x_{o_i})+E_a(x_{a_i})
\label{eq:pos}
\end{equation*}
Here $x_{f_i}$ denote the $i$-th token in $x_f$, where other sequence (\emph{i.e., } $x_t$, $x_c$, $x_o$, $x_a$) follow the similar denotation.
Note that $E_t(x_{t_i})$ leverages the bi-LSTM to encode the byte sequences (see Section~\ref{subsec:input_repr}), while the others simply multiplies the token (\emph{i.e., } one-hot encoded~\cite{harris2010digital}) with an embedding matrix.
Figure~\ref{fig:arch} (right-hand side) illustrates the two embedding strategies.
\vspace{.1cm}\noindent\textbf{Masked LM.}
We pretrain the model using the masked LM objective.
Formally, for each sequence $(x_1,x_2,...,x_n)$ in a given training set, we randomly mask out a selected percentage of tokens in each sequence.
Specifically, we mask the code token and value token in $x_f$ and $x_t$, respectively, and replace them with a special token \texttt{<MASK>} (marked gray in Figure~\ref{fig:arch}).
As masked LM trains on micro-traces without requiring additional labels, it is fully unsupervised.
Let $m(E_i)$ denote the embedding of the masked $x_i$ and $\mathbb{MP}$ a set of positions on which the masks are applied.
The model $g_p$ (to be pretrained) takes as input a sequence of embeddings with random tokens masked: $(E_1,..., m(E_i),...,E_n), i\in \mathbb{MP}$, and predicts the code and the values of the masked tokens: $\{\hat{x}_{f_i}, \hat{v}_i|i\in \mathbb{MP}\}=g_p(E_1,..., m(E_i),...,E_n)$.
Let $g_p$ be parameterized by $\theta$, the objective of training $g_p$ is thus to search for $\theta$ that minimizes the cross-entropy losses between (1) the predicted masked code tokens and the actual code tokens, and (2) predicted masked values (8 bytes) and the actual values. For ease of exposition, we omit summation over all samples in the training set.
\begin{equation}
\label{eq:pretrain}
\centering
\argmin_{\theta} \sum\limits_{i=1}^{|\mathbb{MP}|} (-x_{f_i}\log(\hat{x}_{f_i})+ \alpha\sum\limits_{j=1}^{8}-x_{t_{ij}}\log(\hat{x}_{t_{ij}}))
\end{equation}
$\hat{x}_{t_{ij}}$ denotes the predicted $j$-th byte of $x_{t_i}$ (the $i$-th token in $x_t$).
$\alpha$ is a hyperparameter that weighs the cross-entropy losses between predicting code tokens and predicting values.
\vspace{.1cm}\noindent\textbf{Masking strategy.}
For each chosen token to mask, we randomly choose from the masking window size from $\{1,3,5\}$, which determines how many consecutive neighboring tokens of the chosen tokens are also masked~\cite{joshi2020spanbert}.
For example, if we select $x_5$ (the 5-th token) in Figure~\ref{fig:arch} to mask and the masking window size is 3, $x_4$ and $x_6$ will be masked too.
We then adjust accordingly to ensure the overall masked tokens still account for the initially-chosen masking percentage.
\vspace{.1cm}\noindent\textbf{Contextualized embeddings.}
We employ the self-attention layers~\cite{vaswani2017attention} to endow contextual information to each embedding $E_i$ of the input token.
Notably, let $E_{l}=(E_{l,1},...,E_{l,n})$ denote the embeddings produced by the $l$-th self-attention layer.
We denote the embeddings before the model ($E_i$ as defined in Section~\ref{subsec:input_repr}) as $E_{0,i}$.
Each token's embedding at each layer will attend to all other embeddings, aggregate them, and update its embedding in the next layer.
The embeddings after each self-attention layer are known as \emph{contextualized embeddings}, which encodes the context-sensitive meaning of each token (\emph{e.g., } \texttt{eax} in \texttt{mov eax,ebx} has different embedding with that in \texttt{jmp eax}).
This is in contrast with static embeddings, \emph{e.g., } word2vec commonly used in previous works~\cite{duandeepbindiff, ding2019asm2vec}, where a code token is assigned to a fixed embedding regardless of the changed context.
The learned embeddings $E_{l,i}$ after the last self-attention layer encodes the approximate execution semantics of each instruction and the overall function.
In pretraining, $E_{l,i}$ is used to predict the masked code. While in finetuning, it will be leveraged to match similar functions (Section~\ref{subsec:finetuning_method}).
\subsection{Finetuning for Function Similarity}
\label{subsec:finetuning_method}
Given a function pair, we feed each function's \emph{static code} (instead of micro-trace as discussed in Section~\ref{sec:intro}) to the pretrained model $g_p$ and obtain the pair of embedding sequences produced by the last self-attention layer of $g_p$: $E^{(1)}_{k}=(E^{(1)}_{k,1},...,E^{(1)}_{k,n})$ and $E^{(2)}_{k}=(E^{(2)}_{k,1},...,E^{(2)}_{k,n})$ where $E^{(1)}_{k}$ corresponds to the first function and $E^{(2)}_{k}$ corresponds to the second.
Let $y=\{-1,1\}$ be the ground-truth indicating the similarity (1 -- similar, -1 -- dissimilar) between two functions.
We stack a 2-layer Multi-layer Perceptrons $g_t$, taking as input the average of embeddings for each function, and producing a function embedding:
\begin{equation*}
g_t(E_{k})=tanh((\sum\limits_{i=1}^n E_{k,i})/n)\cdot W_1)\cdot W_2
\end{equation*}
Here $W_1\in \mathbb{R}^{d_{emb}\times d_{emb}}$ and $W_2\in \mathbb{R}^{d_{emb}\times d_{func}}$ transforms the average of last self-attention layers embeddings $E_k$ with dimension $d_{emb}$ into the function embedding with the dimension $d_{func}$.
$d_{func}$ is often chosen smaller than $d_{emb}$ to support efficient large-scale function searching~\cite{xu2017neural}.
Now let $g_t$ be parameterized by $\theta$, the finetuning objective is to minimize the cosine embedding loss ($l_{ce}$) between the ground-truth and the cosine distance between two function embeddings:
\begin{equation*}
\argmin_{\theta}\ l_{ce}(g_t(E^{(1)}_{k}),g_t(E^{(2)}_{k}), y)
\end{equation*}
where
\begin{equation}
\label{eq:cosembloss}
l_{ce}(x_1,x_2,y) = \left\{\begin{matrix}
1-cos(x_1,x_2) & y=1\\
max(0,cos(x_1,x_2)-\xi) & y=-1
\end{matrix}\right.
\end{equation}
$\xi$ is the margin usually chosen between 0 and 0.5~\cite{paszke2019pytorch}.
As both $g_p$ and $g_t$ are neural nets, optimizing Equation~\ref{eq:pretrain} and Equation~\ref{eq:cosembloss} can be guided by gradient descent via backpropagation.
After finetuning $g_t$, the 2-layer multilayer perceptrons, and $g_p$, the pre-trained model, we compute the function embedding $f_{emb}=g_t(g_p(f))$ and the similarity between two functions is measured by the cosine similarity between two function embedding vectors: $cos(f^{(1)}_{emb},f^{(2)}_{emb})$.
\section{Overview}
\label{sec:overview}
In this section, we use the real-world functions as motivating examples to describe the challenges of matching semantically similar functions.
We then overview our approach, focusing on how our pretraining task (masked LM) addresses the challenges.
\subsection{Challenging Cases}
\begin{figure*}[!t]
\centering
\subfloat[Cross-architecture]{
\includegraphics[width=0.3\linewidth]{./figs/challenges/x86_arm.pdf}
\label{subfig:x86_arm}}
\subfloat[Cross-optimization]{
\includegraphics[width=0.33\linewidth]{./figs/challenges/O0_O3.pdf}
\label{subfig:O0_O3}}
\subfloat[Cross-obfuscation]{
\includegraphics[width=0.33\linewidth]{./figs/challenges/subobf.pdf}
\label{subfig:subobf}}
\caption{Challenging cases of matching semantically similar functions across different instruction architectures, optimizations, and obfuscations. (\textbf{Left}) the function \texttt{priv\_encode\_gost} is from \texttt{libcrypto.a} in \texttt{openssl-1.0.1f}. The upper function is compiled to x86 while the lower is compiled to ARM.
(\textbf{Middle}) the function \texttt{<wd\_comparator>} is from \texttt{basenc} in \texttt{coreutils-8.32}. The upper and lower function is compiled by GCC-7.5 with \texttt{-O0} and \texttt{-O3}, respectively.
(\textbf{Right}) the function \texttt{<CMS\_add0\_cert>} is from \texttt{libcrypto.a} in \texttt{openssl-1.0.1u}. The upper function is compiled using \texttt{clang} with default options. The lower function is compiled by turning on the instruction substitution using Hikari~\cite{hikari}, \emph{e.g., } \texttt{-mllvm -enable-subobf}.}
\label{fig:challenge}
\end{figure*}
We use three semantically equivalent but syntactically different function pairs to demonstrate some challenges of learning from only static code.
Figure~\ref{fig:challenge} shows the (partial) assembly code of each function.
\vspace{.1cm}\noindent\textbf{Cross-architecture example.}
Consider the functions in Figure~\ref{subfig:x86_arm}. Two functions have the same execution semantics as both functions take the lower 12-bit of a register and compare it to \texttt{0x80}.
Detecting this similarity requires understanding the approximate execution semantics of \texttt{and} in x86 and \texttt{lsl}/\texttt{lsr} in ARM.
Moreover, it also requires understanding how the values (\emph{i.e., } \texttt{0xfff} and \texttt{0x14}) in the code are manipulated.
However, all existing ML-based approaches~\cite{massarelli2019safe} only learn on static code without observing each instruction's real execution effect.
Furthermore, to mitigate the potentially prohibitive vocabulary size (\emph{i.e., } all possible memory addresses), existing approaches replace all register values and memory addresses with an abstract dummy symbol~\cite{massarelli2019safe,duandeepbindiff}. They thus cannot access the specific byte values to determine inherent similarity.
\vspace{.1cm}\noindent\textbf{Cross-optimization example.}
Now consider two functions in Figure~\ref{subfig:O0_O3}.
They are semantically equivalent as \texttt{[ebp+8]} and \texttt{[esp+4]} access the same memory location, \emph{i.e., } the function's first argument pushed on the stack by the caller.
To detect such similarity, the model should understand \texttt{push} decreases the stack pointer \texttt{esp} by 4.
The model should also notice that \texttt{mov} at line 2 assigns the decremented \texttt{esp} to \texttt{ebp} such that \texttt{ebp+8} in the upper function equals \texttt{esp+4} in the lower function.
However, such dynamic information is not reflected in the static code.
\vspace{.1cm}\noindent\textbf{Cross-obfuscation example.}
Figure~\ref{subfig:subobf} demonstrates a simple obfuscation by instruction substitution, which essentially replaces \texttt{eax+1} with \texttt{eax-(-1)}.
Detecting the equivalence requires understanding approximately how arithmetic operations such as \texttt{xor}, \texttt{sub}, and \texttt{add}, executes.
However, static information is not enough to expose such knowledge.
\subsection{Pretraining Masked LM on Micro-traces}
\label{subsec:overview_pretrain}
This section describes how the pretraining task, masked LM, on functions' micro-traces encourages the model to learn execution semantics.
Although it remains an open research question to explicitly prove certain knowledge is encoded by such language modeling task~\cite{saunshi2019theoretical}, we focus on describing the intuition behind the masked LM -- why predicting masked codes and values in micro-traces can help address the challenging cases in Figure~\ref{fig:challenge}.
\vspace{.1cm}\noindent\textbf{Masked LM.} Recall the operation of masked LM: given a function's micro-trace (\emph{i.e., } values and instructions), we mask some random parts and train the model to predict the masked parts using those not masked.
Note that pretraining with masked LM does not need any manual labeling effort, as it only predicts the masked part in the input micro-traces without any additional labeling effort.
Therefore, \textsc{Trex}\xspace can be trained and further improved with a substantial number of functions found in the wild.
The benefit of this is that a certain instruction not micro-executed in one function is highly likely to appear in at least one of the other functions’ micro-traces, supporting \textsc{Trex}\xspace to approximate diverse instructions' execution semantics.
\vspace{.1cm}\noindent\textbf{Masking register.} Consider the functions in Figure~\ref{subfig:subobf}, where they essentially increment the value at stack location \texttt{[rbp-0x2c]} by 1.
The upper function directly loads the value to \texttt{eax}, increments by 1, and stores the value in \texttt{eax} back to stack.
The lower function, by contrast, takes a convoluted way by first letting \texttt{ecx} to hold the value -1, and decrements \texttt{eax} by \texttt{ecx}, and stores the value in \texttt{eax} back to stack.
We mask the \texttt{eax} at line 3 in the upper function. We find that our pretrained model can correctly predict its name and dynamic value.
This implies the model understands the semantics of \texttt{add} and can deduce the value of \texttt{eax} in line 3 after observing the value of \texttt{eax} in line 2 (before the addition takes the effect).
We also find the model can recover the values of masked \texttt{ecx} in line 4 and \texttt{eax} in line 5, implying the model understands the execution effect of \texttt{xor} and \texttt{sub}.
The understanding of such semantics can significantly improve the robustness in matching similar functions -- when finetuned to match similar functions, the model is more likely to learn to attribute the similarity to their \emph{similar execution effects}, instead of their syntactic similarity.
\vspace{.1cm}\noindent\textbf{Masking opcode.}
Besides masking the register and its value, we can also mask the opcode of an instruction. Predicting the opcode requires the model to understand the execution effect of each opcode.
Consider Figure~\ref{subfig:O0_O3}, where we mask \texttt{mov} in line 2 of upper function.
We find our pretrained model predicts \texttt{mov} with the largest probability (larger than the other potential candidates such as \texttt{add}, \texttt{inc}, etc.).
To correctly predict the opcode, the model should have learned several key aspects of the function semantics.
First, according to its context, \emph{i.e., } the value of \texttt{ebp} at line 3 and \texttt{esp} at line 2, it learns \texttt{mov} is most probable as it assigns the value of \texttt{esp} to \texttt{ebp}. Other opcodes are less likely as their execution effect conflicts with the observed resulting register values.
This also implicitly implies the model learns the approximate execution semantics of \texttt{mov}.
Second, the model also learns the common calling convention and basic syntax of x86 instructions, \emph{e.g., } only a subset of opcodes accept two operands (\texttt{ebp,esp}). It can thus exclude many syntactically impossible opcodes such as \texttt{push}, \texttt{jmp}, etc.
The model can thus infer \texttt{ebp} (line 3 of upper function) equals to \texttt{esp}. The model may have also learned \texttt{push} decrements stack pointer \texttt{esp} by 4 bytes, from other masked samples.
Therefore, when the pretrained model is finetuned to match the two functions, the model is more likely to learn that the semantic equivalence is due to that \texttt{[ebp+8]} in the upper function and \texttt{[esp+4]} in the lower function refer to the same address, instead of their similar syntax.
\vspace{.1cm}\noindent\textbf{Other masking strategies.}
Note that we are not constrained by the number or the type of items (\emph{i.e., } register, opcode, etc.) in the instructions to mask, \emph{i.e., } we can mask complete instructions or even a consecutive sequence of instructions, and we can mask dynamic values of random instructions’ input-output.
Moreover, the masking operation dynamically selects random subsets of code blocks and program states at each training iteration and on different training samples.
As a result, it enables the model to learn the diverse and composite effect of the instruction sequence, essential to detecting similarity between functions with various instructions.
In this paper, we adopt a completely randomized strategy to choose what part of the micro-trace to mask with a fixed masking percentage (see Section~\ref{subsec:pretrain_method} for details).
However, we envision a quite interesting future work to study a better (but still cheap) strategy to dynamically choose where and how much to mask.
\section{.5pt}{5pt plus 2pt minus 2pt}{5pt plus 2pt minus 2pt}
\begin{document}
\title{\textsc{Trex}\xspace: Learning Execution Semantics from\\ Micro-Traces for Binary Similarity}
\makeatletter
\newcommand{\linebreakand}{%
\end{@IEEEauthorhalign}
\hfill\mbox{}\par
\mbox{}\hfill\begin{@IEEEauthorhalign}
}
\makeatother
\author{\IEEEauthorblockN{Kexin Pei}
\IEEEauthorblockA{Columbia University\\
[email protected]}
\and
\IEEEauthorblockN{Zhou Xuan}
\IEEEauthorblockA{University of California, Riverside\\
[email protected]}
\linebreakand
\IEEEauthorblockN{Junfeng Yang}
\IEEEauthorblockA{Columbia University\\
[email protected]}
\and
\IEEEauthorblockN{Suman Jana}
\IEEEauthorblockA{Columbia University\\
[email protected]}
\and
\IEEEauthorblockN{Baishakhi Ray}
\IEEEauthorblockA{Columbia University\\
[email protected]}}
\date{}
\maketitle
\input{abst}
\input{intro}
\input{overview}
\input{threat}
\input{methodology}
\input{impl}
\input{eval}
\input{case}
\input{related}
\input{conclusion}
\input{ack}
\bibliographystyle{plain}
\section{Related Work}
\label{sec:related}
\subsection{Binary Similarity}
\vspace{.1cm}\noindent\textbf{Traditional approaches.}
Existing static approaches often extract hand-crafted features by domain experts to match similar functions.
The features often encode the functions' syntactic characteristics.
For example, BinDiff~\cite{bindiff} extracts the number of basic blocks and the number of function calls to determine the similarity.
Other works~\cite{myles2005k, khoo2013rendezvous, crussell2013andarwin, farhadi2014binclone} introduce more carefully-selected static features such as n-gram of instruction sequences.
Another popular approach is to compute the structural distance between functions to determine the similarity~\cite{bindiff, bourquin2013binslayer, david2014tracelet, pewny2014leveraging, eschweiler2016discovre, david2017similarity, huang2017binsequence}.
For example, BinDiff~\cite{bindiff} performs graph matching between functions' call graphs.
TEDEM~\cite{pewny2014leveraging} matches the basic block expression trees.
BinSequence~\cite{huang2017binsequence} and Tracelet~\cite{david2014tracelet} uses the edit distance between functions' instruction sequence.
As discussed in Section~\ref{sec:intro}, both static features and structures are susceptible to obfuscations and optimizations and incur high overhead.
\textsc{Trex}\xspace automates learning approximate execution semantics without any manual effort, and the execution semantics is more robust to match semantically similar functions.
In addition to the static approaches, dynamic approaches such as iLine~\cite{jang2013towards}, BinHunt~\cite{ming2012ibinhunt}, iBinHunt~\cite{ming2012ibinhunt}, Revolver~\cite{kapravelos2013revolver}, Blex~\cite{egele2014blanket}, Rieck \textit{et al.~}\cite{rieck2011automatic}, Multi-MH~\cite{pewny2015cross}, BinGo~\cite{chandramohan2016bingo}, ESH~\cite{david2016statistical}, BinSim~\cite{ming2017binsim}, CACompare~\cite{hu2017binary}, Vulseeker-pro~\cite{gao2018vulseeker}, and Tinbergen~\cite{mckee2019software} construct hand-coded dynamic features, such as values written to stack/heap~\cite{egele2014blanket} or system calls~\cite{ming2017binsim} by executing the function to match similar functions.
These approaches can detect semantically similar (but syntactically different) functions by observing their similar execution behavior.
However, as discussed in Section~\ref{sec:intro}, these approaches can be expensive and can suffer from false positives due to the under-constrained dynamic execution traces~\cite{jiang2009automatic, ding2019asm2vec}.
By contrast, we only use these traces to learn approximate execution semantics of individual instructions and transfer the learned knowledge to match similar functions without directly comparing their dynamic traces.
Therefore, we are much more efficient and less susceptible to the imprecision introduced by these under-constrained dynamic traces.
\vspace{.1cm}\noindent\textbf{Learning-based approaches.}
Most recent learning-based works such as Genius~\cite{feng2016scalable}, Gemini~\cite{xu2017neural}, Asm2Vec~\cite{ding2019asm2vec}, SAFE~\cite{massarelli2019safe}, DeepBinDiff~\cite{duandeepbindiff} learn a function representation that is supposed to encode the function syntax and semantics in low dimensional vectors, known as function embeddings.
The embeddings are constructed by learning a neural network that takes the functions' structures (control flow graph)~\cite{feng2016scalable,xu2017neural,duandeepbindiff} or instruction sequences~\cite{ding2019asm2vec, massarelli2019safe} as input and train the model to align the function embedding distances to the similarity scores.
All existing approaches are based only on static code, which lacks the knowledge of function execution semantics.
Moreover, the learning architectures adopted in these approaches require constructing expensive graph features (attributed CFG~\cite{feng2016scalable, xu2017neural}) or limited in modeling long-range dependencies (based on Word2Vec~\cite{ding2019asm2vec, duandeepbindiff}).
By contrast, \textsc{Trex}\xspace learns approximate execution semantics to match functions.
Its underlying architecture is amenable to learning long-range dependencies in sequences without heavyweight feature engineering or graph construction.
\subsection{Learning Program Representations}
There has been a growing interest in learning neural program representation as embeddings from ``Big Code''~\cite{allamanis2018survey}.
The learned embedding of the code encodes the program's key properties (\emph{i.e., } semantics, syntax), which can be leveraged to perform many applications beyond function similarity, such as program repair~\cite{parihar2017automatic, wang2017semantics}, recovering symbol names and types~\cite{chua2017neural, patrick2020probabilistic}, code completion~\cite{raychev2014code}, decompilation~\cite{fu2019coda, katz2018using}, prefetching~\cite{shi2019learning}, and many others that we refer to Allamanis \textit{et al.~}\cite{allamanis2018survey} for a more thorough list.
Among these works, the most closest work to us is XDA~\cite{pei2021xda}, which also leverages the transfer learning to learn general program representations for recovering function and instruction boundaries.
However, XDA only learns from static code at the raw byte level, which lacks the understanding of execution semantics.
The core technique proposed in this paper -- learning approximate execution semantics from micro-traces -- is by no means limited to only function similarity task but can be applied to any of the above tasks.
Indeed, we plan to explore how the learned semantics in our model can transfer to other (binary) program analysis tasks in our future work.
\section{Threat Model}
\label{sec:threat}
We assume no access to the debug symbols or source while comparing binaries.
Indeed, there exist many approaches to reconstruct functions from stripped binaries~\cite{bao2014byteweight, shin2015recognizing, andriesse2017compiler, di2017rev, pei2021xda}.
Moreover, we assume the binary can be readily disassembled, \emph{i.e., } it is not packed nor transformed by virtualization-based obfuscator~\cite{vmprotect, ugarte2015sok}.
\vspace{.1cm}\noindent\textbf{Semantic similarity.}
We consider two semantically similar functions as having the same input-output behavior (\emph{i.e., } given the same input, two functions produce the same output).
Similar to previous works~\cite{xu2017neural, massarelli2019safe, ding2019asm2vec}, we treat functions compiled from the same source as similar, regardless of architectures, compilers, optimizations, and obfuscation transforms.
| 77b7872015b7a62413514156fc8cd7ab150afdba | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Quantum key distribution (QKD) enables two remote network users to exchange provably-secure keys when it is implemented faithfully~\cite{lo2014secure,Diamanti2016,XuRMP2020}. To ensure implementation security, the research community has been focusing on the security of practical systems in recent years,
developing methods to narrow the gap between the theory and practice of QKD.
On the theoretical side, robust QKD protocols have been proposed, which not only make practical systems more secure against device imperfections but also easier to calibrate and validate in practice (since fewer assumptions are required).
On the experimental side, efforts have been focused on tackling quantum side-channels and a wide variety of countermeasures have been proposed and developed~\cite{Nitin2016,XuRMP2020}.
Trojan-horse attacks (THAs)~\cite{Vakhitov2001, gisin_trojan-horse_2006} represent one of the biggest threats to QKD security. These attacks aim to steal the secret key information via the injection of unauthorized light pulses, seeking to carry critical modulation information out of the transmitters. More specifically, in these attacks, the adversary (henceforth called Eve) injects bright light pulses into the transmitter and collects the reflected light pulses. Consequently, this allows Eve to learn some information about the secret key. It has been shown that these kind of attacks can be readily implemented using standard optical methods~\cite{Vakhitov2001, gisin_trojan-horse_2006,jain2014trojan, Sajeed2017}. To mitigate this issue, one can use specialized security analyses to include security against specific types of THAs; for instance, by modeling the unauthorized input light pulses as coherent states. Then, under the assumption that the energy of the reflected light pulses is bounded, one can compute the secret key rate, as was done in~Refs.~\cite{lucamarini_practical_2015,tamaki_decoy-state_2016}.
The bright illumination attacks are another particularly powerful class of side-channel attacks. These include laser damage attacks~\cite{bugge2014laser,makarov_creation_2016,huang_laser-damage_2020} and blinding attacks~\cite{makarov2009controlling,lydersen2010hacking,yuan_resilience_2011}. In these attacks, bright light pulses are used to control QKD devices by exploiting their implementation knowledge. Consequently, these allow Eve to avoid eavesdropping detection and hence security is no longer guaranteed. Fortunately, there exist countermeasures which are pretty effective against such attacks~\cite{yuan_avoiding_2010,yuan_resilience_2011} and innovative QKD protocols which are completely immune against detection side-channel attacks are known as well, e.g., see measurement-device-independent QKD (MDI QKD)~\cite{Lo2012, Braunstein2012}.
Based on the above, it can therefore be said that the injection of (unauthorized) bright light pulses into quantum communication systems is a catalyst for side-channel attacks. This is not so surprising since the presence of bright light pulses essentially breaks one of the most important assumptions of quantum cryptography---that the energy of the underlying quantum signals is at the single-photon level (or sufficiently small). To overcome these potential loopholes, one promising solution is to limit the energy of incoming light. Indeed, if this is achieved, one can be sure that the QKD system is operating at the single-photon level and the energy of any outgoing light pulse is bounded as well. Consequently, this will allow the system to operate faithfully in the quantum regime.
In practice, this solution would mean introducing a kind of \emph{quantum power limiting} device into the QKD system. Based on current research on side-channel attacks, we believe an ideal quantum power limiter should possess the following properties: (a) able to provide a reliable and adjustable photon energy limiting down to the order of a few photons to hundreds of photons for each quantum state, (b) have a minimum insertion loss if the input power is below the threshold and stop the transmission or maintain at the threshold power once the input power exceeds the threshold, and (c) the power limiting effects are independent of other physical degrees of freedoms, e.g., frequency, polarization, etc. In terms of practical considerations, the power limiter device should also be cost-effective, passive, and easily replaceable (if it cannot recover to its normal state after being exposed to strong light).
Here, we propose and demonstrate a novel and practical quantum power limiter that can secure a broad class of QKD setups~\cite{patent}. The device is based on a form of thermo-optical defocusing effect, which effectively bounds the output optical power by some predetermined threshold. By modeling the system using a set of physically relevant assumptions, we show that the output-input optical power relation of the proposed device can be precisely controlled by changing the system parameters, e.g. the length of the prism and the diaphragm width. Consequently, this allows us to tailor the device to different quantum cryptographic applications. The feasibility and performance of our proposed power limiter device are confirmed using COMSOL (a multi-physics simulation software) and experimental data.
The paper is organized as follows. In Section.~\ref{sect2}, we first present the design details and the modeling of our power limiter. Thereafter, simulation and experimental results are illustrated. Section.~\ref{sect3} discusses the potential implementation loopholes and the robustness of the proposed power limiter.
In Section.~\ref{sect3.5}, we experimentally verify that the power limiter is essentially transparent to standard quantum encoding choices such as intensity, phase, and polarization degrees of freedom.
In Section.~\ref{sect4}, we illustrate the broad utility of the proposed power limiter over three different QKD systems. In the first application, we provide a general security analysis of MDI QKD that allows for Eve to inject in any kind of state in a given Trojan-horse optical mode. Thereafter, a detailed study on the application of our power limiter in MDI QKD is presented, followed by the simulation results. In the second and third applications, we discuss how the proposed power limiter could be utilized to deter bright illumination attacks and to enhance the implementation security of plug-and-play QKD~\cite{muller_plug_1997, xu_measurement-device-independent_2015, tang2016experimental}. In Section.~\ref{sect5} we end with a conclusion.
\section{Optical power limiter design}\label{sect2}
\begin{figure}[tbp]
\centering
\includegraphics[width=1\columnwidth]{figure1-1.pdf}
\caption{Schematic of the power limiter design. An acrylic prism is used as the active medium. The input collimated Gaussian beam diverges due to the thermo-optical defocusing effect, when the absorbed energy introduces temperature gradients inside the prism. A diaphragm is placed after the prism to control the collectable optical power. The optical filter restricts the working wavelength range for security analysis. The inset is the top view of the acrylic prism and the diverged Gaussian beam. Owing to the isotropic nature of acrylic, both the optical and thermal responses are assumed to be axially symmetric along the optical axis.}
\label{fig1}
\end{figure}
Our power limiter design is shown in Fig.~\ref{fig1}. The input light and output light are collimated using a pair of fiber collimators. An acrylic prism is placed along the optical path as the core part of our proposal, whose negative thermo-optical coefficient (TOC) $ dn/dT $ is exploited, where $n$ is the refractive index and $T$ is the temperature. Noted here any material with negative TOC could be used with similar analysis. The absorption of input light generates a heat gradient inside the prism, which is then converted to a refractive index gradient accordingly. The negative TOC leads to a relatively smaller reflective index at the center of the prism, resulting in the whole optical architecture works as a concave lens and diverges the transmitting light, as shown in the inset of Fig.~\ref{fig1}. By adding a diaphragm with customizable width, the amount of output power can be suitably controlled. An optical filter is then introduced to restrict the working wavelength range of the device for security analysis, which will be discussed in detail in Section.~\ref{sect3}. We remark that all the components used here are cost-effective and commercially available.
The mechanism of thermal optical defocusing and related power limiting phenomenon have been widely studied in both theory and experiments~\cite{smith_high-power_1977,leite_thermal_1967,derosa_fiber-optic_2003}. In our case, We first simulate the temperature and electric field distribution inside a 10~cm acrylic prism with 7.9~mW input power using COMSOL, whose results are shown in~\ref{fig2} (a) and (b). The simulation results indicate a distinct temperature distribution inside the medium, and a clear divergence of the light field. Towards a better quantitative understanding, we model our power limiter design by balancing the optical absorption and the heat transfer inside the prism under steady-state condition~\cite{smith1969thermal}:
\begin{equation}\label{eq1-1}
\alpha I = -\frac{k}{r}\frac{\partial}{\partial r} \left(r \frac{\partial T}{\partial r} \right),
\end{equation}
where $\alpha$ is the absorption coefficient of the material, $I$ represents the input light power density, $T$ is the temperature, and $k$ is the thermal conductivity. If we assume that the light propagates along the $z$-direction and follows a Gaussian profile, temperature gradient in the $z$-direction is negligible, and the radiative and convective heat transfer is minimal, the steady-state laser radiation intensity at position $(r,z)$ can be solved as~\cite{smith1969thermal}
\begin{equation}\label{eq2}
\begin{split}
I(r,z) =& I(r,0) \\
&\cdot \exp\Bigg[-\alpha z + \frac{\displaystyle\frac{\partial n}{\partial T} P_0 e^\frac{-r^2}{a^2} \left(z - \frac{1 - e^{-\alpha z}}{\alpha} \right)}{\pi kna^2}\Bigg],
\end{split}
\end{equation}
where the input intensity $ I(r,0)=\frac{P_0}{\pi a^2} e^{-r^2/a^2}$, $a$ is the radius where the light intensity drops to $1/e$ of its axial value, and $P_0$ is the incident laser power.
The output optical power can be obtained by integrating the light intensity over a certain area which depends on the position (prism length) and the width of diaphragm.
\begin{figure}[tbp]
\centering
\includegraphics[width=1\columnwidth]{figure2-5.pdf}
\caption{Simulated (a) temperature and (b) E-field distribution in the region marked with the dashed box in Fig.~\ref{fig1} using a 2-dimensional model in COMSOL. The results clearly indicate the high-temperature gradient in the $r$-direction and the gradual divergence of Gaussian profile in the $r$-direction. Simulated (c) maximum output power and (d) insertion loss at different diaphragm width and prism length using Eq.~\ref{eq2}, where $\alpha = 25.95~{\rm m}^{-1}$ (measured~value),~TOC = $1.3\times10^{-4}~{\rm K}^{-1}$~\cite{zhang2006thermo}, $n = 1.47$~\cite{zhang2020complex}, $k = 0.19~ {\rm Wm^{-1}K^{-1}}$ \cite{rudtsch2004intercomparison},~$a = 0.14~{\rm mm}$.
(e) Experimental output-input power relationship at different diaphragm width with the same prism length of 101.6~mm. The results indicate that smaller the diaphragm width, lower the output power threshold. Also with a diaphragm width larger than the beam width, the insertion loss will remain minimum. (f) Experimental output-input power relationship at different prism length but with the same diaphragm width of 25~\textmu m. Longer prism length could provide lower output power threshold but the insertion loss will be higher. Both the simulation and experimental results confirmed the power limiting effect of our design and an adjustable power limiting threshold is feasible.}
\label{fig2}
\end{figure}
The maximum output power (defined as power limiting threshold) and the insertion loss at different prism length and diaphragm width are shown in Fig.~\ref{fig2} (c) and (d). Since a larger prism length will lead to a greater photon absorption as well as a larger light divergence, a higher insertion loss and a smaller power limiting threshold can thus be expected. Likewise, a smaller diaphragm collects less photon energy, which also results in a higher insertion loss and a smaller power limiting threshold. Therefore, depending on the application, it is possible to choose a set of parameters that balance the insertion loss and power limiting threshold that meet system requirements.
Note here that the Gaussian profile assumption only holds when the beam divergence is relatively small, thus the analytical model may only be able to provide a quick guidance for parameter selection. Hence, experiments are conducted to verify the feasibility of our proposal.
A proof-of-concept experiment is performed using a simplified version of Fig.~\ref{fig1}. A collimator is used for light coupling from single mode optical fiber to free-space. Here a transmissive collimator based on graded-index (GRIN) lens is used in the setup for feasibility demonstration, and it can be conveniently replaced by reflective collimators to ensure the proper functioning over a wide range of wavelengths for security reasons (See Section.\ref{sect3} for details). Then the Gaussian beam with a beam width of 0.4~mm is directed into the Acrylic Prism. Three acrylic prisms with lengths of 25.4, 50.8, and 101.6~mm are tested. The output light will then be collected after the diaphragm. Diaphragm width of 25, 50, 380, 750 and 1300~\textmu m are used in our experiment. Fig.~\ref{fig2} (e) shows the measured output-input relationship at different diaphragm width and the same prism length of 101.6~mm, while Fig.~\ref{fig2} (f) shows the result at different prism length with the same diaphragm width of 25~\textmu m. The results clearly show the power limiting effect in various conditions. The output power linearly increases with the input power at low power region. As the input power further increases, the output power will increase slowly, and finally be limited to a certain threshold.
Besides, the experimental results verified that the power limiting feature of our proposal can be readily adjusted by modifying the prism length and diaphragm width. Among all of our system configurations, the lowest power limiting threshold of -27.9~dBm is measured, with a insertion loss of -34.0~dB, when a 101.6~mm prism and 25~\textmu m diaphragm are chosen. Similarly, a lower insertion loss of -5.1~dB can be obtained, together with a 10.3~dBm output power limiting threshold, when a 50.8~mm prism and 750~\textmu m diaphragm are used. For different applications, one can expect different requirements for power limiting device. For example, for protecting transmitters against THA, the insertion loss of the power limiter is less concerned since we can always adjust the optical attenuators to generate expected quantum states. While in order to protect receivers from bright illumination attacks, the insertion loss of the device can be a critical factor to system performance. Thus, we would imagine customised power limiter configurations for different application scenarios.
\section{Robustness against potential implementation loopholes}\label{sect3}
The above analyses so far only show the feasibility of the proposed power limiter under a steady-state condition. Below, we analyze the robustness of the proposed device against potential implementation loopholes that could happen via the variation of standard optical properties.
One important consideration is the finite response time of the proposed device. To investigate this property, we install an electronic variable optical attenuator (EVOA) after a continuous-wave (CW) laser source to create a laser pulse with relatively long pulse width and measure the output response of the power limiter. The experimental results are shown in Fig.~\ref{fig3} (a), where a 101.6~mm prism is used with a 750~\textmu m diaphragm. The settling time of our power limiter is measured to be 300~ms. We observed that the peak output power close to the starting time can be a few times higher than the steady-state output power (which happens after about 300~ms). Crucially, this suggests that one could exploit the finite response time of the power limiter to breach the desired energy threshold.
\begin{figure}[tbp
\centering
\includegraphics[width=1\columnwidth]{figure3-4.pdf}
\caption{(a) Experimental output response of the power limiter at different input optical power. An 101.6~mm prism is used with a 750~\textmu m diaphragm. The power limiting effect has a long settling time of around 300~ms. (b) Experimental output response of the power limiter with constant-energy pulse input. An 101.6~mm prism is used with a 25~\textmu m diaphragm. The peak power and duty cycle are selected to maintain the same energy per pulse or the same average input power of 10.5~dBm and 13~dBm. (c) The corresponding average output power at different duty cycle which shows that the maximum output power occurs at duty cycle of 1, i.e. CW input. (d) The COMSOL simulated maximum temperature inside the prism with constant average input power of 20~mW. The result shows that higher peak power heats the prism faster and reaches a higher temperature. Thus, a higher thermo-optical effect can be expected.}
\label{fig3}
\end{figure}
However, as we will show in Section. \ref{sect4}. A, the information leakage due to THAs can, in fact, be bounded using only the average energy constraint (integrated over the finite response time); it is not necessary to bound the maximum (peak) energy for security. Thus, in the experiment, we study the average output optical power at constant-energy pulse input but with different duty cycles. The time domain results are shown in Fig.~\ref{fig3} (b), where a 101.6~mm prism is used with a 25~\textmu m diaphragm. The input laser pulse is modulated at 1~Hz frequency with average input power of 10.5~dBm and 13~dBm. The corresponding average output power is shown in Fig.~\ref{fig3} (c). The results indicate that the average output power is higher at a larger duty cycle. The maximum appears at duty cycle equals 1, i.e. CW light input. In other words, given fixed average input power, CW input will give the largest averaged output power, where Eve is getting the most amount of information about the transmitter. As such, we will be using the power limiting threshold obtained under the CW Trojan horse input assumption for THA analysis; see Section. \ref{sect4}. A.
To explain this effect, we study the temperature response inside the medium under constant-energy pulse input with different peak power and different duty cycles using COMSOL. The results are shown in Fig.~\ref{fig3} (d). The simulation results indicate that a higher input peak power will lead to a higher maximum temperature, even with the same amount of average power. Therefore, a higher refractive index gradient and larger divergence of input laser are expected with a higher instantaneous power of the input light, leading to a larger thermo-optical defocusing effect and consequently a lower output power.
Another possible attack is to try to change the power limiting threshold by varying the wavelength of the incoming light. This could allow Eve to send in brighter light pulses with a different wavelength. To investigate the possibility of such an attack, we analyze how different input wavelength could affect the TOC and heat generation of the power limiter device.
Generally, the TOC can be modeled by~\cite{soave2009refractive,qiu_synthesis_2009}
\begin{equation}\label{eq3}
{\rm TOC} = \frac{\mathrm{d} n}{\mathrm{d} T} = f(n(\lambda)) \left(\Phi-\beta\right),
\end{equation}
where $f(n(\lambda))$ is defined as $(n^2-1)(n^2+2)/(6n)$, $n$ is the reflective index, $\lambda$ is the wavelength of input light, $\Phi$ is the electronic polarizability and $\beta$ is the volumetric expansion coefficient. In most polymers, the volumetric expansion coefficient is more dominant, i.e. $\Phi\ll\beta$, and hence the overall TOC is typically negative~\cite{soave2009refractive}. More importantly, notice that the volumetric expansion coefficient is physically independent of the wavelength. As such, the wavelength dependency of TOC is only related to $f(n(\lambda))$. The $f(n(\lambda))$ for acrylic as the function of wavelength is shown in Fig.~\ref{fig4} (a). The corresponding TOC change will introduce a small difference in the output power threshold calculation, as referenced to the power at 1550~nm, which is shown as the red curve in Fig.~\ref{fig4} (a).
\begin{figure}[tbp
\centering
\includegraphics[width=1\columnwidth]{figure5-5.pdf}
\caption{(a) Calculated $f(n(\lambda))$ from Eq.~\ref{eq3} and the output power threshold difference caused by the corresponding TOC change, as referenced to the value at 1550~nm. (b) Absorption loss spectrum of acrylic, taken from Ref.~\cite{zhang2020complex,zhang2020complex2}. The loss at 1310~nm and 1550~nm communication bands are marked. The minimum loss is about 0.15~dB/cm at visible wavelength. (c) The maximum output power of the power limiter and the corresponding input power with silicon absorber (visible light absorbing filter) of different lengths. The silicon absorber significantly reduces the power threshold at the visible wavelength. (d) The total maximum output photon number per second calculated from (c) with the silicon absorber thickness of 1~mm.}
\label{fig4}
\end{figure}
As for heat generation, it is related to the absorption loss of the material. A lower loss indicates less energy converted from the optical energy to heat energy, thereby resulting in a lower temperature gradient and a higher power limiting threshold. Based on this, a spectral filter with a large power handling capability can be applied to limit the transmission spectrum of the device; in which case the peak power (over the transmitted spectrum) is considered for the security analysis.
Considering optical fiber-based applications at 1550 nm, the optical fiber itself is, in fact, a bandpass filter for about 300-2100 nm wavelength, beyond which the transmission loss is higher than 100~dB/km~\cite{agrawal2000nonlinear}. Thus, by applying a secure fiber with adequate length, light beyond this wavelength range can be suppressed to a negligible level. In this way, it is effective to only consider the wavelength dependency feature within this band.
For the material of our power limiter, acrylic, its absorption loss spectrum is shown in
Fig.~\ref{fig4} (b) with some low loss bands marked~\cite{zhang2020complex,zhang2020complex2}. The loss is about 1.29~dB/cm at 1550~nm and 0.82~dB/cm at 1310~nm, which are standard communication bands. The loss below 1100~nm is even lower. The minimum occurs at about 800~nm with 0.15~dB/cm loss. Based on the absorption spectrum and considering a prism length of 10~cm with 750~\textmu m diaphragm, the maximum output power spectrum and the corresponding input power is calculated based on Eq.~\ref{eq2}, as shown in Fig.~\ref{fig4} (c). The power threshold below 1100~nm is about 11~dB higher than the 1310~nm band and more than 17~dB higher than the 1550~nm band. Although a pessimistic power bound of about 8~dBm can be set and used as the system power bound (marked as a red triangle), it is better to use an optical filter to block the light below 1100~nm wavelength. The silicon absorber can be a good candidate, which provides a stable and robust filtering performance. By adding a thin layer of silicon sheet after the power limiter, the output power can be significantly suppressed. As shown in Fig.~\ref{fig4} (c), with only about 1~mm thick silicon, the maximum output power shifts back to the communication bands. The maximum output photon number per second can be further calculated, as shown in Fig.~\ref{fig4} (d). The maximum output photon number per second appears at 1260~nm wavelength with a photon energy of $1.58 \times 10^{-19}~J$; as such, this wavelength is considered in the security analysis using the worst-case approach.
Similarly, for other degrees of freedom, e.g. the state of polarization, the polymer acrylic used in our design inherently possesses isotropic behavior. Thus, by nature, it will not introduce any birefringence related changes and is independent of the thermo-optical effect, which avoids introducing related loopholes to the system.
Another consideration is laser damage attacks~\cite{bugge2014laser,makarov_creation_2016,huang_laser-damage_2020}.
Preliminary simulations indicate that the acrylic prism could be damaged with only about 400~mW of input power~\cite{m2014hole,berrie1980drilling}.
If the acrylic is damaged or burnt, the thermal defocusing effect is not applicable anymore and the light might be collected by the output collimator directly. Consequently, the power limiting effect may not hold. However, this issue can be resolved by replacing the crossing-through prism with a total internal reflection structure, where the input beam is non-coaxial with the output. In this way, any damage to the material will not weaken the robustness of our proposal; instead the device works as an optical fuse to permanently block the optical path.
\section{Quantum signal integrity}\label{sect3.5}
To determine if the proposed power limiter is useful for practical use, it is also important to study if the quantum signals will be disturbed when passing through the device. To this end, experiments based on time-bin (intensity), phase, and polarization encoding are implemented to see whether the quantum signal integrity will be affected.
Here, we study the QBER of the system, which is defined as the number of errors ($N_{error}$) over the total number of detection counts ($N_{correct} + N_{error}$),
\begin{equation}\label{eq_qber}
{\rm QBER} = \frac{N_{error}}{N_{correct} + N_{error}}.
\end{equation} In addition, given that the detector is well characterised (e.g., its background noise and single-photon efficiency are known), we may further write ${\rm QBER}= {\rm QBER}_{opt} + {\rm QBER}_{det}$,
where QBER$_{opt}$ comes from quantum optical imperfections (e.g., imperfect state preparation, optical misalignment, etc) and QBER$_{det}$ comes from the detector dark counts. Here, as mentioned above, our main focus is the QBER$_{opt}$ for intensity, phase and polarization encoding schemes, which represent three of the most popular choices for QKD in practice.
The QBER of intensity or time-bin encoding scheme is measured first. As shown in Fig.~\ref{fig_qber_i}~(a), the intensity extinction ratio of a pulsed laser is measured to infer the QBER. The pulsed laser is attenuated to about 0.1 photon per pulse and measured by an avalanche photodiode (APD) operating in the Geiger mode (gated). The laser pulse has a repetition frequency of 100~MHz and a pulse width of 400~ps. The APD has a gate width of 1~ns. The delay on the APD gate signal is scanned to cover both the laser pulse (bit 1) and dark region (bit 0). The dark counts here are subtracted after the data acquisition for an accurate extinction ratio measurement of the optical pulse. The power limiter used here has a length of 101.6~mm and a diaphragm width of 750~\textmu m. An average input power of 14.49 dBm and -19.72 dBm are tested to demonstrate the cases when the input power is close to and far below the power limiting threshold, respectively. The schematic of the signal controls are shown in Fig.~\ref{fig_qber_i}~(b). The resulting counts as a function of delay is shown in Fig.~\ref{fig_qber_i}~(c) and (d). For the input power and the cases with and without the power limiter, the resulting extinction ratios are all above 35 dB, indicating a QBER of less than 0.032\%. Therefore, we conclude that the introduction of the proposed power limiter will not introduce any significant noise to QKD systems based on time-bin encoding.
\begin{figure}[tbp]
\centering
\includegraphics[width=1\columnwidth]{figure_qber_i.pdf}
\caption{(a) Experimental setup for QBER measurement of a time-bin encoding scheme (which uses on-off keying). An APD is used to measure the extinction ratio of a pulsed laser with and without the power limiter. (b) Schematic of the signal control for the pulsed laser and the APD gate signal. (c,d) The count value as a function of gate signal delay with and without power limiter (PL). (c) The case for 14.49 dBm input power, which is close to the power limiting threshold. (d) The case for -19.72 dBm input power. The discontinuity of the curve is because of the negative count values obtained due to the statistical noise. }
\label{fig_qber_i}
\end{figure}
For the phase encoding scheme, the experimental setup is shown in Fig.~\ref{fig_qber_p}~(a). The input CW laser is modulated using a phase modulator switching between 0 and $\pi$ phase with 50~MHz frequency. The laser output power is 10.28 dBm. Schematics of the signal controls are shown in Fig.~\ref{fig_qber_p}~(b). The modulated signal is then decoded using an asymmetric Mach-Zehender interferometer (AMZI) with a path delay of around 10~ns. Moreover, a phase shifter is added in one of the paths of the AMZI to lock the relative phase. As such, the interference visibility as well as the QBER can be obtained. Finally the output is attenuated to 0.1 photon per gate and measured by an APD. The counts as a function of delay are shown in Fig.~\ref{fig_qber_p}~(c) and (d), which corresponds to the case with and without power limiter installed, respectively. The interference visibility $V$ is shown in Fig.~\ref{fig_qber_p}~(e) as $(1-V)$ for a clear view. The maximum visibility with and without power limiter are 0.9844 and 0.9836, corresponding to a QBER of 0.78\% and 0.82\%, respectively. Thus like in the case of time-bin encoding, we conclude that the proposed power limiter device is also suitable for phase-encoding QKD systems.
\begin{figure}[tbp]
\centering
\includegraphics[width=1\columnwidth]{figure_qber_p.pdf}
\caption{(a) Experimental setup for the QBER measurement of a typical phase encoding scheme. The input CW light is modulated with a phase modulator in 0 and $\pi$ phase and decoded with an AMZI. The output light is attenuated to single-photon level and measured with an APD. (b) Schematics of the waveform for the modulated phase in both the path of AMZI and the outputs from AMZI with different phase shifter setting resulting in minimum power and maximum power. Similar to the time-bin (intensity) encoding scheme, the APD gate signal delay is scanned over the whole period. (c,d) The count value with minimum and maximum phase shifter setting as a function of gate signal delay with and without the power limiter. (e) The calculated interference visibility as a function of delay with and without the power limiter.}
\label{fig_qber_p}
\end{figure}
Finally, we study the impact of the device on polarization encoding. The experimental setup is shown in Fig.~\ref{fig_qber_pol}, where a CW laser with an output power of 11.41 dBm is used and polarization is manually tuned with a polarization controller. The attenuated output goes through a polarization beam splitter (PBS) and the outputs are measured by two APDs. The polarization extinction ratio is calculated from the ratio between the two APD counts. The result shows a polarization extinction ratio of 30.1 dB and 32.6 dB for the case with and without the power limiter, corresponding to QBERs of 0.098\% and 0.055\%, respectively. This clearly shows that the power limiter will not significantly disturb the state of polarisation of the photon.
\begin{figure}[tbp]
\centering
\includegraphics[width=1\columnwidth]{figure_qber_pol.pdf}
\caption{(a) Experimental setup for the QBER measurement of the polarization encoding scheme. The polarization of a attenuated CW laser is manually tuned to match the polarization of a polarization beam splitter (PBS). The output from the PBS is measured by two APDs.}
\label{fig_qber_pol}
\end{figure}
All in all, we experimentally confirmed that our power limiter device does not introduce any significant noise (in terms of the QBER) to standard QKD systems based on time, phase, and polarization encoding schemes. However, it should be noted that the power limiter does introduce extra losses (insertion loss) to the signal so the photon collection efficiency would decrease when it is deployed on the receiver side. In our experiment, a minimum insertion loss of -5.1~dB is measured, which is equivalent to an transmission efficiency of around 31\%, or a transmission distance of 25.5~km(assuming single mode fiber with a transmission loss of 0.2~dB/km). We note that this issue could be mitigated by using materials with higher TOC values so that smaller amount of light absorption is required to trigger the power limiting effect.
\section{Applications and countermeasures}\label{sect4}
\subsection{Security against THAs}\label{subsect4}
As an application of our proposed power limiter, we consider a phase-encoding MDI QKD protocol~\cite{Lo2012, Braunstein2012,tamaki2012phase} with energy constrained THAs. A schematic of our system is shown in Fig.~\ref{fig5}, where Alice and Bob are distant quantum transmitters and supposed to prepare the required phase-encoding coherent states, then send them to Charlie for Bell-state measurement. The protocol is outlined below:
\begin{figure}[tbp
\centering
\includegraphics[width=1\columnwidth]{figure6-4.pdf}
\caption{Schematic of the phase-encoding MDI QKD system with the power limiter installed. Alice and Bob are the users preparing the phase-encoding coherent states using their lasers, modulators, and attenuators. The prepared states are sent to Charlie for Bell state measurement. The distance between Alice and Bob is contributed by the two fiber spools combined. The Trojan horse attack from Eve could provide her with a maximum $\nu$ Trojan horse photon, which is taken into consideration for secure key rate calculation.}
\label{fig5}
\end{figure}
Alice and Bob randomly prepare one of the four coherent states $\{\ket{e^{ix\frac{\pi}{2}}\alpha}\}$ and $\{\ket{e^{iy\frac{\pi}{2}}\beta}\}$, where $x,y\in\{0,1,2,3\}$ are the classical information of Alice and Bob, respectively. Then Alice and Bob send the quantum states to Charlie via the quantum channel for Bell-state measurement. The distance between Alice and Bob is contributed by the two fiber spools combined.
Charlie interferes the incoming states from Alice and Bob using a 50-50 beam-splitter and measures the outputs using single-photon detectors. Thereafter, he announces the measurement result $z\in\{L,R,\varnothing\}$ through the authenticated classical channel, which corresponds to the left detector clicks, the right detector clicks, and none of the detector clicks or both detectors click. Alice and Bob repeat the state preparation and measurement for $N$ rounds.
Upon receiving the Bell-state measurement results from Charlie, Alice and Bob only keep data of those rounds give $z = L, R$. Besides, Bob flips the value of $y$ if $z=R$. Alice and Bob then obtain the statistics of all the state combinations of Alice and Bob, conditioned on $z=L,R$. Particularly, for rounds with $x,y=0,2$, Alice and Bob keep the data for extracting the secret keys.
Alice and Bob then implement parameter estimation and apply error correction and privacy amplification thereafter to extract a pair of identical and secure keys.
To take THAs into consideration, different models for Trojan horse states have been proposed. For example, in Ref. \cite{lucamarini_practical_2015}, the Trojan horse state is modeled as a pure coherent state with a fixed phase and intensity. However, this model might be too restrictive as Eve can send other states. In practice, she could send a mixture of coherent states with different intensities or other states that could potentially leak more information. Another model that can address potential THAs is presented in Refs. \cite{pereira2019leaky, pereira2020correlated}. There, the non-vacuum component of the Trojan horse state is modeled by an arbitrary state that lives outside the qubit space in which the legitimate parties encode the information. While this model is very general and could take into accounts of any source side-channels, the resulting bound can be overly pessimistic, since in the worst-case scenario the leakages might correspond to orthogonal quantum states and hence would leak full information about the modulation (key information).
In our analysis, we take the intermediate step and allow Eve to send any Trojan horse state in a given optical mode. However, because the modulators in Alice's and Bob's labs are trusted, the resultant Trojan horse states will not be orthogonal after the modulation. As such, the THA will not leak complete information about Alice's and Bob's key information. Without loss of generality, the Trojan horse state can be written as
\begin{equation}\label{eq4}
\ket{\xi}=\sum_{n,m}{c_{nm}\ket{n}\ket{m}}\ket{\mathcal{E}_{nm}},
\end{equation}
where $\ket{n},\ket{m}$ are the Fock states injected into Alice's and Bob's apparatus, respectively. $\ket{\mathcal{E}_{nm}}$ is an ancilla that is kept in Eve's lab. The coefficients $c_{nm}$ are the quantum amplitudes of the Fock states. Note that the state of the form \eqref{eq4} includes Trojan horses that are mixed (after tracing out Eve's ancilla) and may even be entangled.
The states $\ket{n}$ and $\ket{m}$ will accumulate some phases introduced by Alice's and Bob's modulators and hence they would leak some information about $x$ and $y$. On the other hand, the states $\ket{\mathcal{E}_{nm}}$ will not accumulate any phase since it is kept in Eve's lab. After gathering the modulation information from the modulators, the output THA state thus with the form
\begin{equation}\label{eq5}
\ket{\xi'_{xy}}=\sum_{n,m}{c_{nm}e^{i(nx+my)\frac{\pi}{2}}\ket{n}\ket{m}}\ket{\mathcal{E}_{nm}},
\end{equation}
Both the quantum states prepared by Alice and Bob and the THA state will be sent to Charlie via the quantum channel. Thus, the untrusted measurement can be modeled by a quantum-to-classical map, which can be described by an isometry $\mathcal{U}$ (with an appropriate purification):
\begin{align}\label{eq6}
\ket{\phi_{xy}}&=\ket{e^{ix\frac{\pi}{2}}\alpha}\ket{e^{iy\frac{\pi}{2}}\beta}\ket{\xi'_{xy}}\nonumber\\
&\stackrel{\mathcal{U}}{\longrightarrow}\sum_{z}{\ket{\mathrm{e}^z_{xy}}\ket{z}}.
\end{align}
Therefore, given the fact that the Gram matrix $G$ based on Eq. (\ref{eq6}) is positive semi-definite and linearly constrained, a tight upper bound of the phase error rate can be obtained by solving the dual problem of a semi-definite program (SDP), similar to the security analysis presented in Ref.~\cite{wang_characterising_2019,primaatmaja2019versatile}.
The asymptotic secret key rate can thus be obtained using the so-called Shor-Preskill key rate formula~\cite{Shor2000}:
\begin{equation}\label{eq8}
R \geq \max\{0,P_{\rm pass} [1-h_2(e_{\rm ph})-h_2(e_{\rm bit})] \},
\end{equation}
where $e_{\rm bit}(e_{\rm ph})$ is the bit (phase) error rate of the system, $P_{\rm pass}$ represents the probability of successful Bell state measurement when Alice and Bob choose the key generation basis, and $h_2(\cdot)$ is the binary entropy function. A detailed security analysis is given in the supplemental material. To restrict information leakage, the mean photon-number, $\nu$, of the THA state should be much less than one. This requirement can be achieved using the proposed power limiter together with an optical attenuator. These devices can be readily implemented in standard quantum transmitters as shown in Fig.~\ref{fig5}.
To be more specific, based on the power limiting threshold obtained in Section.~\ref{sect3},
a maximum photon number of injected eavesdropping light can be strictly constrained by the proposed optical power limiter. Then, the injected light will go through the attenuator twice before being collected by Eve, while the quantum state for QKD has just been attenuated once. Consider a QKD system working at a frequency of 1~GHz, with a power limiting threshold of 1~mW and an ideal phase modulator that does not introduce any extra insertion loss. In this case, an attenuation of 69~dB is sufficient to guarantee an average energy output of $\nu=10^{-7}$. At the same time, the laser output can be adjusted to optimize the intensity $\mu$ for QKD, where an averaged optical power of 23~\textmu W can be used to generate quantum states with $\mu=0.0183$. This is similar to the optimized intensity for MDI QKD with a detector efficiency of $\eta_\text{det} = 85\%$, dark count rate $p_\text{dc} = 10^{-7}$, and a 100~km transmission distance. Comparing to Ref.~\cite{lucamarini_practical_2015} where the 12.8~W optical fiber damage threshold was used as the upper bound, the proposed power limiter could limit the power by 4 to 5 orders of magnitude lower. As a result, the requirement for attenuator and optical isolator is significantly reduced. Also, removing the need to have isolators could benefit future chip-based integration of such MDI QKD systems.
As mentioned, due to the finite response time of the power limiter, only the average power of the THA (instead of the maximum power of the THA pulses) can be bounded. To this end, we develop a general proof technique (see the supplemental material) that uses only the average photon number information of the THA. In particular, the security proof takes into account attacks where Eve employs a mixture of bright Trojan horse pulses with the vacuum (where the probability of sending a bright light is small enough such that the energy constraint is satisfied). As such, the proposed optical power limiter can be used to ensure that the assumptions of the security proof are enforced during the protocol.
\begin{figure}[tbp
\centering
\includegraphics[width=1\columnwidth]{figure7-4.pdf}
\caption{Simulation for asymptotic key rate for phase-encoding MDI QKD under two set of parameters: (a) detector's efficiency $\eta_\text{det} = 10\%$, dark count rate $p_\text{dc} = 10^{-5}$, (b) detector's efficiency $\eta_\text{det} = 85\%$, dark count rate $p_\text{dc} = 10^{-7}$. Trojan horse photon number $\nu$ of $10^{-5}$, $10^{-6}$, $10^{-7}$ and 0 are shown. The output intensity $\mu$ of each transmitter is optimized for each distance to maximize the key rate.}
\label{fig6}
\end{figure}
To benchmark the performance of the protocol, we simulate the achievable asymptotic key rate with two sets of parameters: (1) detector's efficiency $\eta_\text{det} = 10\%$, dark count rate $p_\text{dc} = 10^{-5}$, (2) detector's efficiency $\eta_\text{det} = 85\%$, dark count rate $p_\text{dc} = 10^{-7}$. For both sets of parameters, misalignment error $e_{\text{ali}}$ is set to be $2\%$, and the transmission loss of optical fiber is set to be 0.2 dB/km. We also assume that the central node is equidistant to Alice and Bob and $\mu_A = \mu_B = \mu$, which has been optimized over the simulation. As for the THA intensity, we set $\nu_A=\nu_B=\nu$. The results of the simulation are shown in Fig.~\ref{fig6}. The results indicate Alice and Bob can get a promising key rate without being affected much by the Trojan horse attack if the energy of the THA is properly upper bounded.
\begin{figure}[tbp]
\centering
\includegraphics[width=1\columnwidth]{figure_blind.pdf}
\caption{(a) A schematic of the detector blinding attack proposed in Ref.~\cite{lydersen_thermal_2010}. (b) The current-voltage relationship of a typical APD. In normal working conditions, the breakdown voltage is $V_{Br0}$. A fixed bias voltage $V_{Bias}>V_{Br0}$ enables Geiger mode operation with single-photon sensitivity. The resulting output current above a threshold $I_{Th}$ is registered as a successful count. However, the breakdown voltage increases with the temperature $T$ of the device. Thus, it is possible to blind the detector by driving it into the linear mode ($V_{Bias}<V_{Br1}$), thus nullifying the single-photon sensitivity. (c) The current-input power relationship of an APD operating in linear mode. By controlling the input power below or above the power threshold $P_{Th}$, the detector could be controlled to register fake-states. (d) The input power on Bob's detector with and without a power limiter.}
\label{fig_blind}
\end{figure}
\subsection{Potential countermeasures}
\subsubsection{Bright illumination attacks}
Laser damage attacks are a particularly powerful class of bright illumination attacks. In Ref.~\cite{bugge2014laser,makarov_creation_2016,huang_laser-damage_2020}, it is shown that the detectors and optical components are prone to permanent changes and damages when Eve sends in a bright damaging laser with power in the order of Watts. This is crucial because the security of most QKD systems depend on the integrity of their devices---that they behave according to design specifications.
Another class is detector blinding attacks~\cite{makarov2009controlling,lydersen2010hacking,yuan_resilience_2011}. By exploiting the implementation knowledge of single-photon detectors and the imperfect detector performances, Eve can send in a relatively strong eavesdropping light to change the working condition of the detector and get partial (or even full) control over the outcomes~\cite{makarov2009controlling,lydersen2010hacking,yuan_resilience_2011,lydersen_controlling_2011,qin_homodyne-detector-blinding_2018}.
For illumination-related attacks, a common feature is that Eve must send in relatively bright light pulses. Hence, by restricting the input optical power using the proposed power limiter, it is expected that some of these attacks could be thwarted. To illustrate this possibility, we sketch out a method that could prevent the bright illumination attack presented in Ref.~\cite{lydersen_thermal_2010}; see Fig.~\ref{fig_blind} (a). To start with, we note that standard single-photon detectors based on APD typically require low-temperature operation to minimize the detectors' background noise, i.e., to limit the dark count rate. To cool the detectors, thermoelectric coolers (TECs) are used but these have limited cooling capacity. In Ref.~\cite{lydersen_thermal_2010}, it is shown that injection of bright light pulses can create a situation in which the generated heat fails to dissipate completely. This leads to the breakdown voltage of the APD going above the predetermined value, which consequently puts the detector into the \emph{linear} mode (instead of \emph{Geiger} mode); see Fig.~\ref{fig_blind} (b). In this case, the detector is no longer sensitive to single-photon input (i.e., \emph{blinded}) and Eve can manipulate its outcome by sending in a control light pulse superimposed on the bright light pulse, as depicted in Fig.~\ref{fig_blind} (c).
According to Ref.~\cite{lydersen_thermal_2010}, a bright CW light with an optical power of around 10~mW is required to blind the commercial QKD detectors, and a control light pulse with a peak power of around 1 mW is sufficient to fully control the detector's outcome. If the power limiter is in place, as shown in Fig. 7 (a), the input light power can be limited below than this blinding threshold, which would prevent the temperature of detector from raising and hence the detector from being blinded (see Fig.~\ref{fig_blind} (d)). For example, we can use an acrylic prism with length of 50.8~mm and a diaphragm width of 380~\textmu m to provide a power limiting threshold of 6.03~dBm (with -6.02~dB insertion loss) to prevent such attack. It's important to note that in normal working conditions, quantum signals (i.e., optical signal with small energy levels) in principle will experience a small amount of loss while passing through the power limiter device, since the power limiting effect hasn't been triggered. As such, the introduction of our power limiter are not expected to greatly reduce the overall performance of practical QKD systems. As with our current design, we can expect a smaller insertion loss as well as a stronger power limiting effect, for example, by using material with a higher TOC.
\subsubsection{Plug-and-play QKD with untrusted light sources}\label{sect_twowaycomm}
Plug-and-play QKD is a two-way communication configuration~\cite{muller_plug_1997} that aims to simplify implementation requirements such as polarization compensation and reference frame calibration. This approach is especially useful for practical MDI QKD systems since it naturally guarantees near-perfect mode matching for the required two-photon interference~\cite{xu_measurement-device-independent_2015,tang2016experimental,wang_phase-reference-free_2015}.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.7\columnwidth]{figure_pnp-2.pdf}
\caption{Schematics of the power limiter used in (a) a plug-and-play MDI QKD system and (b) a plug-and-play BB84 QKD system.}
\label{fig_pnp}
\end{figure}
However, in using external (untrusted) light sources instead of trusted light sources, plug-and-play systems are prone to transmitter-based attacks~\cite{gisin_trojan-horse_2006,xu_experimental_2010,sun_passive_2011}. Again, the central issue here is that Eve can inject bright light pulses to break the working assumptions of QKD. To overcome this issue, one popular approach is to monitor the energy of the incoming light with a classical detector~\cite{zhao_quantum_2008,zhao_security_2010,xu_measurement-device-independent_2015}. However, it has been shown that such active monitoring methods are not entirely robust and the classical detectors can still be hacked by exploiting their electrical circuitry, e.g., see Refs.~\cite{makarov_creation_2016,sajeed_attacks_2015}.
In light of the above, it is thus interesting to explore alternative countermeasures that are based on passive devices instead of active devices such as detectors. To this end, we propose to replace (or augment) the active power monitoring device with a passive power limiter as shown in Fig.~\ref{fig_pnp}. Similar to the arguments provided in Section.~\ref{subsect4}, the power limiter would limit the energy of the outgoing light and hence Eve's knowledge about the key information as well; we leave a more careful security analysis to future work.
In addition, it is worthwhile to add that existing methods to limit incoming light energy are typically based on isolators/circulators and laser damage threshold of devices~\cite{lucamarini_practical_2015,makarov_creation_2016,huang_laser-damage_2020}. These are however one-directional and add additional attenuation on the propagating direction of the eavesdropping light. In the case where quantum signal has the same propagation as Eve's light, i.e., plug-and-play QKD or quantum receivers to resist against bright-illumination attacks, they may either pessimistically estimate Eve's information---since the actual input power significantly deviates from the device damage threshold, which will be used for security analysis, or introduce large insertion loss so the system performance will be greatly affected.
As a comparison, the proposed power limiter is shown to be able to provide an adjustable power limiting threshold on the output optical power, and capable of protecting the system where the eavesdropping light and quantum signal have the same propagation direction.
\section{CONCLUSION}\label{sect5}
In this report, we have proposed and demonstrated a passive power limiter design based on the thermo-optical defocusing effect of an acrylic prism. By numerical simulations and the experimental demonstration, we rigorously studied the feasibility and performance of our power limiter design. In our experiment, the lowest optical power limiting threshold of -27.9~dBm with an insertion loss of -34.0~dB is measured. With a different setting, a low insertion loss of -5.1~dB is achieved with a 10.3~dBm power limiting threshold. The values are adjustable according to different system requirements. It is possible to further reduce the insertion loss at a certain power threshold by switching to a material with higher TOC values or/and reduce the beam width. Besides, our design possess desirable features like compactness, robustness, plus polarization and spectrum-dimension independence.
To illustrate the applicability of our proposed power limiter, we have quantitatively developed a general security analysis that allows for arbitrary of Trojan-horse states. By properly limiting the THA energy leakage in a MDI QKD system, a desirable secure key rate and transmission distance can be achieved. Moreover, based on the previous evidences, we remarked that the power limiter can be useful for deterring bright illumination attacks in a quantum cryptography system. We took the thermal CW-blinding attack on the APD detectors as an example, and show how the power limiter can be designed to prevent such an attack. We further discussed the possibility of using a power limiter to secure the plug-and-play QKD systems without active elements.
As demonstrated in our paper, by simply limiting the incoming/outgoing optical energy, a broad class of QKD protocols can be practically protected without introducing cumbersome device modification. Beyond these, one can also expect such a power-limiting device to find applications in securing semi-device-independent quantum protocols based on energy constraints~\cite{van2017semi,van_himbeeck_correlations_2019,avesani2020semi,rusca2019self_testing}, line-topology or ring-topology multiparty quantum communication systems~\cite{grice_two-party_2015,schmid_experimental_2005}. As such, we believe it will attract much interest and possess the potential to become a standard tool for quantum cryptography applications.
\begin{acknowledgments}
This research is supported by the National Research Foundation (NRF) Singapore, under its NRF Fellowship programme (NRFF11-2019-0001) and Quantum Engineering Programme 1.0 projects (QEP-P2, QEP-P3, and QEP-P8)
\end{acknowledgments}
| 4bd6e2ffe7ba35433228f83c950dc7b8f2f74898 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{I. Introduction}
Strongly correlated systems with orbital degrees of freedom exhibit interesting phenomena, with the Kondo effect \cite{Kondo1964}, which is the many-body phenomenon arising from an antiferromagnetic interaction between a conduction electron and a localized magnetic moment, as a prominent example. It was originally studied in the context of the enhancement of the resistivity in magnetic alloys at low temperature, and it is now a ubiquitous problem in condensed matter physics. Also, the Kondo lattice model, where the localized spins are aligned periodically, is a paradigmatic model of a heavy fermion system. Its phase diagram, called the Doniach phase diagram \cite{Doniach1977}, contains the paramagnetic phase due to the Kondo screening in the strong coupling regime and the Ruderman-Kittel-Kasuya-Yoshida (RKKY) ordered phase in the weak coupling regime.
Although the Kondo effect has been intensively studied in solid-state and mesoscopic systems, by exploiting its novel possibilities in the control of system parameters and the detection, ultracold atomic gases in an optical lattice allow one to study the Kondo system in a unique manner, which is challenging to investigate in other systems \cite{Bloch2012}. In particular, alkaline-earth-like atoms have received much attention due to the presence of the metastable states \state{3}{P}{0} and \state{3}{P}{2} as well as the ground state \state{1}{S}{0}. Taking advantage of the long-lived states, the quantum simulator with orbital degrees of freedom using the \state{1}{S}{0} and \state{3}{P}{0} or \state{3}{P}{2} atoms has been proposed, and the Kondo system using the two-orbital system has been studied theoretically \cite{Gorshkov2010,Foss2010,Nakagawa2015,Zhang2016,Kanasz2018,Nakagawa2018,Kuzmenko2018,Goto2019}. In order to implement the Kondo system with cold atoms, an antiferromagnetic spin-exchange interaction between mobile and immobile atoms is required. The clock transition spectroscopy in the state-independent optical lattice, called the magic-wavelength optical lattice, reveals that, in contrast to a ferromagnetic coupling of \Yb{173} and {$^{87}\mathrm{Sr}$ \cite{Cappellini2014,Scazza2014,Zhang2014}, the fermionic isotope of \Yb{171} has an antiferromagnetic coupling between the \state{1}{S}{0} atom and the \state{3}{P}{0} atom \cite{Ono2019}. This suggests that the two-orbital system using \Yb{171} is a promising natural candidate for the quantum simulator of the Kondo effect in contrast to the tuning of the spin-exchange coupling via confinement-induced resonances \cite{Riegger2018}. Another feature of \Yb{171} is the weak interatomic interaction of \state{1}{S}{0} atoms, suggesting that the \state{1}{S}{0} atoms in an optical lattice can be described as a non-interacting metallic state, which is suitable for the exploration of the Kondo physics. Motivated by these unique properties of the two-orbital system using \Yb{171}, the numerical simulation of the dipole oscillation of the \state{1}{S}{0} atoms in the presence of the localized \state{3}{P}{0} atom is performed \cite{Goto2019}, showing that the Kondo effect manifests itself in such a way that the center-of-mass motion of \state{1}{S}{0} atoms is suppressed as the temperature is lowered due to the antiferromagnetic spin-exchange interaction.
In this letter, we report on the observation of the spin-exchange dynamics between \Yb{171} atoms in the ground state $\ket{g}=\ket{^1\mathrm{S}_0}$ and in the metastable state $\ket{e} = \ket{^3\mathrm{P}_0}$. Using a two-orbital lattice system consisting of a two-dimensional (2D) magic-wavelength optical lattice and a 1D near-resonant optical lattice giving strong confinement to the $\ket{e}$ atom alone and no net effect to the $\ket{g}$ atom, the quasi (0+1)D system is implemented, where the $\ket{g}$ atom behaves as the quasi 1D free fermion interacting with the $\ket{e}$ atom mimicking a localized magnetic moment. By exploiting the optical Stern-Gerlach method, we observe the relaxation of the spin polarization caused by the interorbital spin-exchange process and the suppression of the spin depolarization in a high magnetic field. The rate of spin-exchange dynamics is also controlled by the excited-state population. These observations are an important first step towards the quantum simulation of the Kondo effect.
\begin{figure*}
\includegraphics[width=1\linewidth]{fig1ab_ver3.pdf}
\caption{Schematic diagram of experiment. (a) Relevant energy diagram of Yb atom. (b) Schematic illustration of beam configuration. The polarization of the clock excitation light is perpendicular to the quantization axis defined by the magnetic
field, and it amounts to the equal mixture of $\sigma_+$ and $\sigma_-$ polarization. The circularly polarized OSG light propagates along the quantization axis. (c) Schematic illustration of experimental procedure. (1) Initially, some fraction of the atoms in the $\ket{g\downarrow}$ state (a green ball) are excited to the $\ket{e\uparrow}$ state (a yellow ball) in a magnetic field of 30 G. The upper figure shows the schematic representation of the optical lattice potentials for the $\ket{g}$ atom (a green curve) and the $\ket{e}$ atom (a yellow curve). (2) After the excitation, a magnetic field is lowered to 0.5 G, and the spin-exchange dynamics is started. (3) After the hold time, the population of the atoms in the $\ket{g\uparrow}$ and in the $\ket{g\downarrow}$ is detected with an OSG technique. The upper figure shows the spin-dependent optical gradient potential to spatially separate the $\ket{g\uparrow}$ atom and the $\ket{g\downarrow}$ atom. The lower figure shows a typical example of the simultaneous observation of both spin states in the false color ToF image of the \Yb{171} gas in the $\ket{g}$ state subjected to the OSG light.}
\label{fig1}
\end{figure*}
\section{II. Methods}
We first explain how we implement the quasi-(0+1)D system using a near-resonant optical lattice. An optical dipole potential $V(\vb*{r})$ is proportional to the laser intensity $I(\vb*{r})$:
\begin{equation}
V(\vb*{r})=-\frac{1}{4}\alpha I(\vb*{r}).
\label{eq1}
\end{equation}
The coefficient $\alpha$ is called polarizability:
\begin{equation}
\alpha = \sum_i \frac{6\pi c^2}{\omega_{i}^3}\left(\frac{\Gamma_i}{\omega_{i}-\omega}+\frac{\Gamma_i}{\omega_{i}+\omega}\right),
\label{eq2}
\end{equation}
where $\omega$ is the laser angular frequency and $c$ is the speed of light. Here $\omega_i$ and $\Gamma_i$ correspond to the resonant angular frequency and the natural linewidth of the $i$th-state, respectively. The wavelength of the near-resonant optical lattice is chosen to be 650.7 nm, which is close to the \state{3}{P}{0}--\state{3}{S}{1} transition wavelength of 649.1 nm, resulting in the large polarizability for the $\ket{e}$ atom $\alpha_e$ (see Fig.~1(a) for relevant energy levels). Using the Eq.~(\ref{eq2}), the polarizability is obtained as $\alpha_e/h=$1.4 kHz/(mW/cm$^2$), $h$ being the Planck constant. In this calculation, we assume that the \state{3}{P}{0}--\state{3}{S}{1} transition makes the dominant contribution and the other transitions are negligible. Similarly, the polarizability for the $\ket{g}$ atom $\alpha_g$ due to the 650.7 nm light is also calculated as $\alpha_g/h=39$ Hz/(mW/cm$^2$). As a result, the near-resonant optical lattice has the large polarizability ratio $\alpha_e/\alpha_g=36$, while a similar experiment in Ref. \cite{Riegger2018} uses a state-dependent lattice with $\alpha_e/\alpha_g=3.3$. In our experiment, the $\ket{e}$ atom is deeply confined by the near-resonant lattice while the lattice potential is regarded as the continuum system for the $\ket{g}$ atom. A natural concern of using near-resonant light is the possibly non-negligible photon scattering loss rate $\gamma_\mathrm{sc}$ of the $\ket{e}$ atoms. By using narrow-linewidth band-pass filters with less than 0.1~nm to suppress the 649.1 nm resonant frequency component in the single-mode narrow-linewidth 650.7 nm laser, we obtain the loss rate of the $\ket{e}$ atom in the absence of the $\ket{g}$ atom $\gamma_\mathrm{sc}$ = 5.0 Hz. Although this is larger than the theoretically estimated value of 0.59 Hz obtained by assuming only the off-resonant excitation, the corresponding lifetime is long enough to clearly observe the spin-exchange dynamics (see III. RESULTS). A state-independent optical lattice is obtained with the magic wavelength of 759.4 nm.
Figure \ref{fig1}(b) illustrates the schematic diagram of the beam configuration. The 2D array of the tube traps is produced using the 2D magic-wavelength lattice ($x$ and $z$), and the 1D near-resonant optical lattice is superimposed along the axis of the tubes ($y$). As a result, the $\ket{e}$ atom is localized by the 3D confinement while the $\ket{g}$ atom is mobile along the $y$ direction in the tube potential. The maximum potential depth for the $\ket{e}$ atom due to the near-resonant lattice amounts to 27$E_\mathrm{R}$, with $E_\mathrm{R}=k_\mathrm{B}\times96$~nK being the recoil energy for the magic wavelength. Here $k_\mathrm{B}$ is the Boltzmann constant. The corresponding trap frequency of the lattice site at the trap center is 24 kHz, and the residual harmonic trap frequency due to the Gaussian beam shape is estimated as 24 Hz from the beam radius.
The near-resonant lattice depth is calibrated using diffraction of the $\ket{e}$ atoms by a pulsed optical lattice technique with the near-resonant lattice \cite{Denschlag2002}. We use the \Yb{171} atom for the calibration of the lattice depth since the bosonic isotopes would suffer from the severe inelastic loss in the \state{3}{P}{0} states, in addition to another technical merit that the Rabi frequency of the clock transition for the fermionic isotopes is larger than that for bosonic isotopes. After the excitation to the $\ket{e}$ state in the 3D magic-wavelength lattice, the remaining atoms in the $\ket{g}$ state are blasted with the resonant light with the \state{1}{S}{0}--\state{1}{P}{1} transition. Then the magic-wavelength lattice potential along the $y$ axis is ramped down in 1 ms, and the pulsed lattice is irradiated along the $y$ axis immediately after switching off the remaining magic-wavelength lattice potentials along the $x$ and $z$ axes. During the time-of-flight (ToF), the atoms are repumped into the $\ket{g}$ state using the resonant light with the \state{3}{P}{0}--\state{3}{D}{1} transition, and the diffraction pattern is probed by absorption imaging with the \state{1}{S}{0}--\state{1}{P}{1} transition. From the oscillatory behavior of the diffraction pattern, we can calibrate the near-resonant lattice depth.
Our experiments start with the preparation of the quantum degenerate gas of \Yb{171} using the sympathetic evaporative cooling with \Yb{173} \cite{Ono2019}. During the evaporative cooling, the optical pumping into the $\ket{g\downarrow}$ state is performed with the \state{1}{S}{0}--\state{3}{P}{1}($F'=1/2$) transition, where $\ket{\uparrow}=\ket{m_F=+1/2}$, $\ket{\downarrow}=\ket{m_F=-1/2}$ denote the projections of the nuclear spin $F$ onto the quantization axis defined by a magnetic field. The number of atoms $N$ and the temperature scaled by the Fermi temperature $T/T_{\mathrm{F}}$ are $N \simeq 2\times10^4$ and $T/T_{\mathrm{F}}\simeq0.3$, respectively. After the removal of \Yb{173} atoms using the resonant light associated with the \state{1}{S}{0}--\state{3}{P}{1}$(F'=7/2$) transition, the atoms are loaded into the optical lattices, where the initial depths of the magic-wavelength optical lattice and the near-resonant optical lattice for the $\ket{e}$ atom are set to $30E_{\mathrm{R}}$ and $6.8E_{\mathrm{R}}$, respectively. Figure \ref{fig1}(c) shows the experimental procedure after loading atoms into an optical lattice. (1) Some fraction of the atoms are coherently transferred to the $\ket{e\uparrow}$ state in a magnetic field of 30 G by a stabilized clock laser \cite{Takata2019} with a typical linewidth of a few Hz. To localize the $\ket{e}$ atom, the near-resonant optical lattice is then ramped up to 27$E_\mathrm{R}$ in 1 ms, which is longer than the inverse of the lattice-site trap frequency, regarded as an adiabatic ramp. To reduce the spatial inhomogeneity of the clock transition frequency due to the residual harmonic trap created by the near-resonant optical lattice, the $\ket{g}$ atoms are coherently transferred to the $\ket{e}$ state in the shallower optical lattice. A moderate lattice depth is required for the sideband-resolved excitation, on the other hand, which allows one to prepare $\ket{e}$ atoms in the lowest band. We thus perform the clock excitation with the lattice depth of 6.8$E_\mathrm{R}$. The hopping energies between the nearest neighbor tubes $J_\perp$ and between the adjacent sites of the near-resonant optical lattice $J_y$ are estimated as $J_\perp=h\times1.0$ Hz and $J_y=h\times7.0$ Hz for the $\ket{e}$ atom, indicating that the hopping energy is negligible within the experimentally relevant timescales. (2) After the clock excitation, the magnetic field is rapidly lowered to 0.5 G with about 3 ms settling time, and the spin-exchange dynamics is initiated. (3) After the hold time, the spin polarization of the $\ket{g}$ atoms is detected with the optical Stern-Gerlach technique (OSG) \cite{Taie2010}, which enables one to separately observe the atoms in the $\ket{g\uparrow}$ and $\ket{g\downarrow}$ states using a spin-dependent optical potential gradient. The OSG light is blue-detuned by 875 MHz from the \state{1}{S}{0}--\state{3}{P}{1}($F'=1/2$) transition.
\section{III. Results}
\begin{figure}
\includegraphics[width=0.95\linewidth]{fig2a.pdf}
\includegraphics[width=0.8\linewidth]{spin_exchange_fig2b.pdf}
\caption{Observation of spin-exchange dynamics. (a) Time evolution of the spin polarization of the $\ket{g}$ atoms $\Delta N_g/N_g$ with the different excitation rates to the $\ket{e}$ state: $r=0.56$, $0.49$, $0.41$, and $0.31$. Error bars show the standard deviations of the mean values obtained by averaging ten measurements. Solid lines represent fits to the data with the Eqs~(\ref{eq3})-(\ref{eq6}). False color time-of-flight images of the $\ket{g}$ atoms after the spin-exchange dynamics with $r=0.56$ are shown. The left and right figures correspond to the hold time of 3 ms and 68 ms, respectively. (b) Spin-exchange rate $\Omega_\mathrm{ex}$ as a function of the excitation rate. Error bars are 1$\sigma$ confidence intervals of the data fits. The solid line represents linear fits to the data.}
\label{fig2}
\end{figure}
Figure \ref{fig2}(a) shows the time evolution of the spin polarization of the $\ket{g}$ atoms, defined as $\Delta N_g/N_{g}$. Here $\Delta N_g$ denotes the atom number difference between the $\ket{g\uparrow}$ and $\ket{g\downarrow}$ states, and $N_g$ is the total number of the $\ket{g}$ atoms. The result clearly shows the spin depolarization due to the spin-exchange interaction with $\ket{e}$ atoms. We note that we did not observe the depolarization in the case of no $\ket{e}$ atoms. Also, the relaxation rate of the spin polarization is controlled by the clock excitation rate $r$, which is associated with the number of $\ket{e}$ atoms. The spin polarization less than unity at the initial time could be ascribed to the imperfect optical pumping and the photon-scattering of the OSG light. It is noted that the remaining unwanted spin component is not removed after the optical pumping.
The observed relaxation dynamics is quantitatively analyzed with the following two-body rate equations \cite{Scazza2014}:
\begin{align}
\dot{p}_{g\uparrow}(t) =& \Omega_\mathrm{ex}(p_{g\downarrow}(t)p_{e\uparrow}(t)-p_{e\downarrow}(t)p_{g\uparrow}(t)) \nonumber \\
&-\Gamma_{eg}p_{g\uparrow}(t)(p_{e\uparrow}(t)+p_{e\downarrow}(t)) \nonumber \\
&+ \frac{\gamma_\mathrm{sc}}{2}(p_{e\uparrow}(t)+p_{e\downarrow}(t)),
\label{eq3}
\end{align}
\begin{align}
\dot{p}_{g\downarrow}(t) =& \Omega_\mathrm{ex}(p_{e\downarrow}(t)p_{g\uparrow}(t)-p_{g\downarrow}(t)p_{e\uparrow}(t)) \nonumber \\
&-\Gamma_{eg}p_{g\downarrow}(t)(p_{e\uparrow}(t)+p_{e\downarrow}(t)) \nonumber \\
&+ \frac{\gamma_\mathrm{sc}}{2}(p_{e\uparrow}(t)+p_{e\downarrow}(t)),
\label{eq4}
\end{align}
\begin{align}
\dot{p}_{e\uparrow}(t) =& \Omega_\mathrm{ex}(p_{e\downarrow}(t)p_{g\uparrow}(t)-p_{g\downarrow}(t)p_{e\uparrow}(t)) \nonumber \\
&-\Gamma_{eg}p_{e\uparrow}(t)(p_{g\uparrow}(t)+p_{g\downarrow}(t)) \nonumber \\
&- \gamma_\mathrm{sc}p_{e\uparrow}(t),
\label{eq5}
\end{align}
\begin{align}
\dot{p}_{e\downarrow}(t) =& \Omega_\mathrm{ex}(p_{g\downarrow}(t)p_{e\uparrow}(t)-p_{e\downarrow}(t)p_{g\uparrow}(t)) \nonumber \\
&-\Gamma_{eg}p_{e\downarrow}(t)(p_{g\uparrow}(t)+p_{g\downarrow}(t)) \nonumber \\
&- \gamma_\mathrm{sc}p_{e\downarrow}(t).
\label{eq6}
\end{align}
Here $p_{\alpha\sigma}(t)=\bar{n}_{\alpha\sigma}(t)/\bar{n}_0$ denotes the relative population of the atom in the $\ket{\alpha\sigma}$ state ($\alpha=g,e$, $\sigma = \uparrow, \downarrow$), where $\bar{n}_{\alpha\sigma}$ and $\bar{n}_0$ denote the mean density of the atom in the $\ket{\alpha\sigma}$ state and the mean density of the total atoms in the initial state, respectively. Also, $\Omega_\mathrm{ex}$ and $\Gamma_{eg}$ correspond to the spin-exchange rate and the two-body loss rate between the $\ket{g}$ atom and the $\ket{e}$ atom, respectively, and they are propotional to $\bar{n}_0$. In addition, $\gamma_\mathrm{sc}$ is the one-body loss rate of the $\ket{e}$ atom.
We assume that the inelastic collision between the $\ket{e}$ atoms is ignored since the hopping rates $J_\perp$ and $J_y$ are much smaller than the spin depolarization rate.
In addition, $\Gamma_{eg}$ is assumed to be independent of the spin state and is estimated from the measurement of the lifetime of the $\ket{e}$ atom during the spin-exchange dynamics, resulting in $\Gamma_{eg}=10$ Hz.
On the other hand, using the inelastic loss-rate coefficient $\beta_{eg\pm} \leq 2.6(3)\times 10^{-16} $ cm$^3$/s obtained by the measurement of the lifetimes of the $\ket{eg^+}$ and $\ket{eg^-}$ states in the 3D magic-wavelength optical lattice \cite{Bettermann2020}, the two-body loss rate is calculated as $2.3\times10^{-2}$ Hz, where $\ket{eg^+}$ and $\ket{eg^-}$ correspond to the spin-singlet state and the spin-triplet state, respectively.
Although the origin of the discrepancy is not known, here we note that the analysis using Eqs.~(\ref{eq3})-(\ref{eq6}) does not depend sensitively on the value of $\Gamma_{eg}$, due to the existence of photon scattering loss $\gamma_\mathrm{sc} = 5$ Hz, and in fact $\Gamma_{eg} =10$ Hz and $\Gamma_{eg} = 0$ give almost the same results. In the following analysis, we adopt $\Gamma_\mathrm{eg} =10$ Hz.
It should be noted that the two-body loss between the $\ket{e}$ atoms via a tunneling process is suppressed owing to the on-site repulsive interaction and the two-body dissipation via a quantum Zeno effect \cite{Syassen2008,Tomita2019}. In our experiment, this effective loss rate is estimated as 0.28 Hz, suggesting that the inelastic loss between the $\ket{e}$ atoms would not occur during the spin-exchange dynamics.
Solid lines in Fig.~\ref{fig2}(a) represent the fits to the data using the two-body rate equations (\ref{eq3})-(\ref{eq6}) by treating $\Omega_\mathrm{ex}$ as a free parameter. Figure \ref{fig2}(b) shows the spin-exchange rate obtained from the data fits in Fig.~\ref{fig2}(a) as a function of the excitation rate to the $\ket{e}$ state, exhibiting an enhancement of the spin-exchange rate with the increase of the number of atoms in the $\ket{e}$ state.
The linear dependence of the spin-exchange rate $\Omega_\mathrm{ex}$ on the excitation rate $r$ is expected when no correlation between the $\ket{e}$ atoms is considered. The validity of this assumption is related to the characteristic energy of the RKKY interaction $V_\mathrm{ex}^2/\epsilon_\mathrm{F}$ \cite{Tsunetsugu1997}, where $V_\mathrm{ex}$ and $\epsilon_\mathrm{F}$ are the spin-exchange energy and the Fermi energy, respectively. In our experiment, this is estimated as $k_\mathrm{B}\times1.4$~nK, which is much smaller than the atomic temperature. Thus, the RKKY correlation is negligible. In addition, this linearity is expected for a small excitation rate, and the investigation of the spin-exchange dynamics with a higher excitation rate will be an interesting future study.
\begin{figure}
\includegraphics[width=0.85\linewidth]{spin_exchange_fig3_mag_dep_150.pdf}
\caption{Magnetic-field dependence of spin-exchange dynamics with the \state{3}{P}{0} excitation rate $r=0.56$. Error bars show the standard deviations of the mean values obtained by averaging ten measurements. In this experiment, the initial state of the exchange dynamics is prepared by the excitation $\ket{g\uparrow}\to\ket{e\uparrow}$ of the spin-balanced \Yb{171} atoms using $\pi$-polarized light since this scheme results in the reduced uncertainty.}
\label{fig3}
\end{figure}
Also, we investigate the magnetic-field dependence of the spin-exchange dynamics. The interorbital spin-exchange interaction energy $V_\mathrm{ex}$ can be estimated as
\begin{equation}
V_\mathrm{ex} = \frac{4\pi\hbar^2}{m}\frac{a_{eg}^+-a_{eg}^-}{2}\int d^3\vb*{r}n_g(\vb*{r})\abs{\psi_e(\vb*{r})}^2,
\label{eq8}
\end{equation}
where $m$ denotes the mass of the atom and $\psi_e(\vb*{r})$ is the single-particle wavefunction of the $\ket{e}$ atom. The spin-exchange is characterized by the difference between the spin-singlet scattering length $a_{eg}^+=240(4)a_0$ and the spin-triplet scattering length $a_{eg}^- = 389(4)a_0$ \cite{Bettermann2020}, resulting in an antiferromagnetic coupling $V_\mathrm{ex}<0$. Here $a_0$ denotes the Bohr radius. In the central tube, $V_\mathrm{ex}$ is estimated to be $V_\mathrm{ex}/h = -0.25$ kHz.
On the other hand, the differential Zeeman shift between the $\ket{g\uparrow}$ state and the $\ket{e\uparrow}$ state amounts to $-200.0(6)$ Hz/G \cite{Ono2019}. Figure \ref{fig3} shows a comparison between the spin relaxation dynamics in a magnetic field of 0.5 G and that in a magnetic field of 150 G, where the Zeeman energy is two orders of magnitude larger than $V_\mathrm{ex}$. The result shows that the spin depolarization dynamics is frozen in a higher magnetic field, which is consistent with the fact that the spin-exchange process is energetically costly.
\section{IV. Conclusions}
In conclusion, we successfully realized the quasi-1D fermion system in the presence of the immobile spin using the 2D magic-wavelength optical lattice and the 1D near-resonant optical lattice. Using this system, the spin-exchange dynamics between the itinerant $\ket{g}$ atom and the localized $\ket{e}$ atom is observed. Our works can pave the way to the quantum simulation of the Kondo effect. Although the near-resonant lattice causes the one-body loss of the $\ket{e}$ atom, which is detrimental to the exploration of many-body physics, the scattering rate $\gamma_\mathrm{sc}$ will be reduced by using a far-detuned laser source with, for example, a wavelength of 652 nm. It will be an interesting in future work to compare the observed spin-exchange rates with theoretical calculations \cite{Cheng2017,Zhang2018,GotoPrivCom}. The Kondo effect manifests itself as a many-body singlet state, resulting in the screening of a localized spin by itinerant fermions, which is called Kondo screening. The screening cloud of itinerant fermions could be observed in the two-orbital system using a Yb quantum gas microscope \cite{Yamamoto2016,Miranda2015}, as in a quantum dot system \cite{Borzenets2020}. In addition, it is interesting to detect the spin correlation between the excited-state atoms as a signature of the RKKY interaction \cite{Gorshkov2010}.
\section{acknowledgments}
\begin{acknowledgments}
We thank Ippei Danshita, Shimpei Goto for fruitful discussions. KO acknowledges support from the JSPS (KAKENHI grant number 19J11413). The experimental work was supported by the Grant-in-Aid for Scientific Research of JSPS(Nos.~JP17H06138, JP18H05405, and JP18H05228), the Impulsing Paradigm Change through Disruptive Technologies (ImPACT) program, JST CREST (No.~JP-MJCR1673), and MEXT Quantum Leap Flagship Program (MEXT Q-LEAP) Grant No.~JPMXS0118069021.
\end{acknowledgments}
| 89c2a47c1de65b1ef6ef6752a154beec23b5952d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{INTRODUCTION}
Although interest in topological data analysis has grown sharply in recent years, most of the important developments have appeared outwith the statistical literature. We suggest that a consequence is that the potential benefits of adapting some established statistical methods to topological data may not have been fully exploited. In this paper we argue that event history methods are highly suitable for the analysis of data on certain topological features, that the use of such methods can increase the power and effectiveness of topological data analyses, and in turn that exploitation of topological features can bring added value to statistical analyses.
We use Fig. \ref{fig:intro1} and Fig. \ref{fig:intro2} to illustrate, at the cost of preempting Section 3, where they will be explained more fully. Figure \ref{fig:intro1} shows four simulated stationary and isotropic random fields, each on a $60\times 60$ discrete lattice. They are simulated to have correlation functions that are indistinguishable for practical purposes. All four have standard N(0,1) marginal distributions.
Panel (a) is a simulation from a Gaussian random field, panels (b) and (c) are independent realisations of one particular non-Gaussian random field, and panel (d) is a realisation of a second non-Gaussian random field. Panel (d) is visually quite different from the others, but we see no obvious differences between (b) or (c) and the Gaussian random field (a). Figure \ref{fig:intro2} however shows simple Nelson-Aalen plots based on topological features of these data, to be described in Section 3. Now the differences are stark. Given that all four have the same marginal and correlation properties, we know of no other method that can draw out higher order differences so readily. Further, panel (a) is seen to be consistent with a Gaussian random field while the other panels are not.
\begin{figure}
\centerline{\includegraphics[height=5in,width=5in]{fig1.pdf}}
\caption{Four simulated random fields, each on a 60$\times$60 lattice.}
\label{fig:intro1}
\end{figure}
\begin{figure}
\centerline{\includegraphics[height=3in,width=3in]{fig2.pdf}}
\caption{Nelson-Aalen plots for connected components for the random fields of Fig. \ref{fig:intro1}. Short dashed lines correspond to panel (a), dotted lines to panels (b) and (c), and long dashed lines to panel (d). The shaded regions indicate $\pm$ one standard deviation. The solid line shows the expected value for a Gaussian random field with the same correlation structure as the data in Fig. \ref{fig:intro1}.}
\label{fig:intro2}
\end{figure}
Topological data analysis is in part founded on persistent homology. Important early contributions include \cite{frosini1992measuring}, \cite{robins1999towards} and, particularly, \cite{Edelsbrunner2002}, who introduced the idea of persistence of features along filtrations. However, the earliest appearance of the expression ``topological data analysis'' seems not to have been until work by \cite{de2004topological} and \cite{bremer2004topological}. Later, \cite{carlsson2009topology} was key in the popularisation of topological data analysis, demonstrating how ideas from topology could solve many of the issues faced when applying geometric methods to complex data. Other developments are described by \cite{perea2019}.
In Section 2 we give an introduction to the key constructs and concepts underpinning persistent homology and topological data analysis, aimed at a statistical reader. In Section 3 we develop the idea of Nelson-Aalen plots for topological features of random fields, as already seen in our introductory example. We apply the methods in two application areas, one being properties of climate model residuals on a sphere, and the other the distribution of neutral atomic hydrogen \textup{H}\,{\sc i} in the interstellar medium. In Section 4 we consider topological methods for embedded metric trees and propose Cox regression methods as a simple way to deal with potentially censored tree data. This method is applied to vascular structures in fundus images of the human eye, where it has potential to be an effective screening tool for the early detection of diseases.
\section{TOPOLOGICAL DATA ANALYSIS}
Topology is the study of features in a space that are invariant under continuous deformation. We will concentrate on topological data analysis methods that are based on persistent homology, which is the most popular variant of topological data analysis. Detailed descriptions are provided by, inter alia, \cite{Edelsbrunner2002}, \cite{edelsbrunner2008persistent}, \cite{carlsson2009topology} and \cite{Edelsbrunner2012PersistentHT}, with an excellent computational description given by \cite{Otter2017}. Other variants of topological data analysis are summarised by \cite{wass18}.
We assume we have a metric space and inherited distance function. Dealing with arbitrary features in a space is complicated, and so a starting point is to give a representation in a common form. One such is a simplicial complex, which is a collection of vertices, edges, triangles, tetrahedra and higher dimensional equivalents. Thus the letters T, D, and A when written on a page could be represented by four vertices connected by three edges (T), a triangle (D) and a triangle with two extra external edges to two extra vertices (A).
Simplicial complexes are not unique but they facilitate the use of linear algebra for the computation of topological invariants, which are features that do not change when the underyling space is bent, stretched or otherwise deformed, but without gluing or tearing.
The most important topological invariants are the homology groups, which are formally algebraic constructs, one per natural number, that collect together invariant features at each dimension. The ranks of these groups are called Betti numbers. The zero-dimension homology group $H_0(\mathcal{X})$ of a simplicial complex $\mathcal{X}$ consists of the disjoint elements, often referred to as connected components. For example, if the letters T, D, and A are written on a page then there are three connected components and the zero-order Betti number is $\beta_0=3$. In three dimensions, if we take for example letter balloons of T, D and A, there are still three connected components.
The next homology group $H_1(\mathcal{X})$ consists of holes constructed from one-dimensional edges. These are loops or cycles that cannot be shrunk to a single point. For T, D, and A on a page there are two, so $\beta_1=2$. For T, D, and A as letter balloons we have $\beta_1=4$. This is because the D and the A are each homologous to a torus. The hole through the centre of the torus is obvious: the other hole arises because a loop around the circle of revolution that forms the torus cannot be shrunk to a point without leaving the surface.
The second-dimension homology group $H_2(\mathcal{X})$ consists of two-dimensional cycles which enclose an empty three-dimensional space, often called a void. For T, D, and A in balloon form there are three, with one per connected component. For T, D, and A on a page there are none.
We now turn to persistent homology, which begins with a filtration of a space or simplicial complex $\mathcal{X}$. This is a nested sequence of subsets
\begin{equation}
\emptyset = \mathcal{X}_0 \subseteq \mathcal{X}_1 \subseteq \cdots \subseteq \mathcal{X}_d \cdots \subseteq \mathcal{X}_n = \mathcal{X},
\label{eqn:filtration}
\end{equation}
indexed by a parameter $d$ that can be discrete or continuous. Each subset
is itself a valid simplicial complex, with associated homology groups. Panel (a) of Fig. \ref{fig:trees} illustrates. It shows a rooted metric tree embedded in two dimensions, and is a simplified version of the metric trees we will consider in Section 4. There are no cycles and so no holes. In this example the feature of interest is the connectivity of the tree as we move towards the root, filtering inward by radial distance. So $\mathcal{X}_d$ consists of the subset of $\mathcal{X}$ that is outside a circle of radius $r=1/d$ from the root. The plot shows two examples. The complex outside the outer circle consists of just one connected component. The complex outside the inner circle consists of nine connected components.
\begin{figure}
\centerline{\includegraphics[height=5in,width=5in]{treefig.pdf}}
\caption{Illustrations of persistent homology: (a) embedded metric tree filtered by radius from root; (b) corresponding persistence diagram for connected components; (c) point cloud filtered by ball diameter; (d) corresponding persistence diagram for holes.
}
\label{fig:trees}
\end{figure}
In persistent homology the appearance and disappearance of features is tracked as we move through the filtration. A feature is said to be born when it appears and to die when it disappears or merges with another feature. Thus if we start with $r=\infty$ in Fig. \ref{fig:trees} there are no features. This continues as $r$ is decreased until it reaches the point on the tree that is furthest from the root, at radius $r_{\rm max}$ say. The first component is born at this level. As $r$ is decreased further, additional components are born at vertices that have only inward facing edges. Existing components merge at vertices with two outward facing edges. In these cases, by convention the component that was born first is assumed to continue, and the component that was born second is said to die at this radius. The process continues until the root is reached and the tree consists of a single connected component.
The birth and death levels of components are often represented by a persistence diagram, as in panel (b) of Fig. \ref{fig:trees}. This is a plot of birth times and death times of features as we move through the filtration. In the example we have plotted against $r_{\rm max}-r$ so that the first component to be born is on the left of the horizontal axis. Points near the diagonal are short-lived, points further from the diagonal are longer-lived. In the example we have two subtrees with common root, which merge only at $r=0$ and so there are two points with the same maximum death time.
A common way to represent a topological space is through a point cloud, as in
panel (c) of Fig. \ref{fig:trees}. In these cases a filtration is usually defined by growing balls of increasing diameter, centred on the points. There are different ways of forming simplicial complexes in these cases, the most common of which is the Vietoris-Rips complex, which defines a subset of points to be a simplex in $\mathcal{X}_d$ when all pairwise distances between points are less than $2d$. The persistence diagram in panel (d) relates to the birth and death of holes for this example. The grey broken lines indicate the ball radius shown in (c). We see that a medium sized hole is about to be born towards the upper right and this will have modest persistence. A large hole in the centre will be born when the ball radius is increased a little more. This will be the last hole to die and is represented by the highest point in panel (d).
Topological data analysis is often used when there is a real-valued function $z_x$ associated with each point $x$ in a space $\mathcal{X}$. A lower level set $\mathcal{X}_t$ is defined as
\begin{equation}
\mathcal{X}_t= \{x: z_x\leq t, x \in \mathcal{X} \},
\label{eqn:levelset}
\end{equation}
and upper level sets can defined similarly. A filtration is defined by increasing or decreasing the threshold $t$ and persistence of topological features in the level sets is tracked. Figure \ref{fig:levsets} illustrates. In panel (a) we show a two-dimensional image on a discrete $20 \times 20$ grid. We assume that pixels that share an edge are connected but those that share only a vertex are not. Panels (b) and (c) show two level sets: pixels with image values less than $t=-1$ and less than $t=0$ respectively, marked as dark squares. In panel (b) there are $\beta_0=6$ connected components. There are also two regions in that panel that are not in the level set and are completely separated by connected components. Hence $\beta_1=2$. The level set has expanded by $t=0$ in (c) and the six previous components have merged, along with others, to form one large component. One new component has appeared to the lower left, and a small single-pixel component is close by, so $\beta_0=3$. Five holes are evident.
\begin{figure}
\centerline{\includegraphics[height=5in,width=5in]{revlevfig.pdf}}
\caption{A simulated Gaussian random field together with level sets $\mathcal{X}_{-1}$ and $\mathcal{X}_{0}$, and barcode for connected components.}
\label{fig:levsets}
\end{figure}
The final panel of Fig. \ref{fig:levsets} shows a so-called barcode for connected components for the image in panel (a). This is an alternative to a persistence diagram for the presentation of lifespans of topological features as we filter through the level sets. There is one line per feature, starting at the birth level and ending at the death level. In the example the lowest field value is in the dark region to the top right. This corresponds to the first pixel to appear in the level set, which by convention is the last to die, giving the long line at the bottom of the barcode in (d). The second component to be born is at a local minimum toward the top-left of the image. This component also has the second-longest lifetime until it merges with the first at around $t=-0.25$. The third-longest lifetime corresponds to the cluster at the lower left of the figure, which is relatively late to be born but quite isolated from other components and hence persists for some time.
A variety of approaches for the analysis of persistence diagrams and barcodes have been proposed. Wasserstein or bottleneck distances can be used to measure the difference between two persistence diagrams, but these measures are not calibrated to any interpretable scale. There remains no useful equivalents of, say, means or variances for persistence diagrams, which would be needed for inference. \cite{mileyko11}, \cite{turner14} and \cite{munch15} showed how to obtain a {F}r\'{e}chet mean for a group of persistence diagrams, but unfortunately this is not unique. An alternative is to base inference on summaries such as the landscape functions of
\cite{bubenik2015}, which essentially consider nested profiles of the point patterns, or on counts of points over subregions, possibly with smoothing to produce the persistent images of, for example, \cite{adams2015}. Other proposals include a form of parametric bootstrap suggested by
\cite{adler17} and the accumulated persistence suggestion of \cite{bisc19}.
\cite{wass18} provides a recent review. Our approach is quite different.
\section{NELSON-AALEN FOR RANDOM FIELDS}
\subsection{Method}
Let $\mathcal{Z}$ be a zero-mean real-valued random field defined on a finite discrete space $\mathcal{X}$. We assume a discrete space because in the applications we have in mind the field will often be an image, and invariably pixellated. Let
$z_x$ be the field value at location $x \in \mathcal{X}$. We assume $z_x$ is drawn from a continuous distribution with no jump discontinuities and that there are no ties. We will require cor$(z_x,z_{x^\prime})<1$ for all $x \neq x^\prime$ in $\mathcal{X}$.
Our filtration corresponding to \eqref{eqn:filtration} will be the nested sequence
of sub-level sets (\ref{eqn:levelset}). But instead of tracking the persistence of features, which underpins most published applications of topological data analysis, we will take a counting process approach based on the birth levels of new features in the filtration. As well as being computationally much more simple, the approach leads naturally to the use of methods that are familiar in event history analysis, with level playing the role of time and with negative values allowed. As features of interest we will concentrate here on connected components within $\mathcal{X}_t$, namely the $0$-dimensional features.
A component is born at location $x$ at level $t$ if $z_x=t$ and $x$ corresponds to a local minimum of $\mathcal{Z}$. To define a local minimum in the discrete space $\mathcal{X}$ we need to define the neighbours $(x)$ of each location $x$. We are free to do this as we wish, though shared edges or nearest neighbours seem sensible for pixellated or lattice-based fields.
Let $N_x(t)$ count the number of components that are born at location $x$ up to and including level $t$, and let $Y_x(t)$ be the associated predictable at-risk indicator. This is an indicator of whether, given the evolution of level sets to just before level $t$, it is possible for location $x$ to hold a local minimum at $t$. So
\begin{equation*}
N_x(t) = \left\{ \begin{array}{ll}
1\;\;\;\; & z_x\leq t,\;\; z_{(x)}> z_x 1_x,\\
\\
0& {\rm otherwise}, \end{array} \right.
\;\;\;\;\;\;\
Y_x(t) = \left\{ \begin{array}{ll}
1\;\;\;\; & z_x \geq t,\;\; z_{(x)}> t 1_x,\\
\\
0 & {\rm otherwise}, \end{array} \right.
\label{eqn:counts}
\end{equation*}
where $1_x$ is a unit vector of the same length as $z_{(x)}$. Define $N(t)=\sum_xN_x(t)$ and $Y(t)=\sum_xY_x(t)$. Now let
\begin{equation}
\hat{A}(t) = \int_{-\infty}^t \frac{dN(u)}{Y(u)}.
\label{eqn:nelaal}
\end{equation}
This is an analogue of the Nelson-Aalen estimator of the cumulative hazard for a survival time variable \citep{aal08}. If preferred, we can redefine the level scale $t$ to $t^\ast=\exp(t)$ so as to be a more familiar non-negative argument. Like the Nelson-Aalen estimator, $\hat{A}(t)$ is defined
for all $t$ and has a finite number of discontinuities at the observed birth times. It is non-negative and uniformly bounded for fixed $\mathcal{X}$. It is a cadlag process, meaning continuous to the right with left limits. However, the
standard Nelson-Aalen variance estimator
\begin{equation}
{\hat{\rm{var}}\{\hat{A}(t)}\} = \int_{-\infty}^t \frac{dN(u)}{Y^2(u)}
\label{eqn:var}
\end{equation}
does not in general apply, for a variety of reasons. One is possible non-independence between counting processes at different locations given non-independence of the underlying field values. Another, related, reason is informative censoring: knowledge that location $x$ falls out of the risk set at level $u$ provides information about other locations. And a third reason is that we have not constructed a martingale with respect to the observed data filtration.
We propose such Nelson-Aalen estimators for topological features as tools for comparing properties of fields beyond marginal and correlation structures, for comparing observed field data or model residuals with an assumed parametric model, and as a diagnostic for assumed or fitted correlation structures. We allow either a single field $\mathcal{Z}$ or independent and identically distributed replicates $\mathcal{Z}_1, \mathcal{Z}_2, \ldots,\mathcal{Z}_N$ defined on a common space $\mathcal{X}$.
\subsection{Asymptotics}
If $\mathcal{X}$ is fixed and there are $N$ independent replicates $\hat{A}_1(t), \ldots, \hat{A}_N(t)$, then as each $\hat{A}_i(t)$ is non-negative and bounded above, for each level $t$ the variance of $\hat{A}_i(t)$ is finite and the classic central limit theorem applies. Hence pointwise inference for large $N$ is straightforward, based on the sample mean $\bar{A}(t)$ and associated sample standard deviation.
For simultaneous confidence bounds we need to be a little careful because of the discontinuities in $\hat{A}_i(t)$. \cite{hahn77} for instance points out that there can be uniformly bounded cadlag processes for which no central limit theorem applies, depending upon properties of their jumps. \cite{bloz94} provide sufficient conditions under which $\surd N\bar{A}(t)$ converges weakly to a Gaussian process with finite variance and smooth sample paths in $(0,\tau)$, for any arbitrarily large $\tau$ for which pr$\{Y(\tau)>0\}>0$. These conditions are verified in Supplementary Material. Hence a functional central limit theorem applies.
Further, for fixed $\mathcal{X}$ we can show that $\bar{A}(t)$ is consistent in $N$ for
\begin{equation}
A(t) = \displaystyle\int_{-\infty}^t J(u)\frac{ \sum_x {\rm pr}(z_{(x)}> u 1_x \mid z_x=u) f_x(u)}{\sum_x {\rm pr}(z_x>u, \;z_{(x)}> u 1_x)} du,
\label{eqn:limit}
\end{equation}
where $f_x(.)$ is the marginal density of $z_x$ and $J(u)=1$ if $E\{Y(u)\}>0$, zero otherwise. If $N=1$ and there is just a single realisation, the estimator $\hat{A}(t)$ is also consistent for (\ref{eqn:limit}), but this time as the cardinality of $\mathcal{X}$ increases, provided cor$(z_x, z_{x^\prime})$ approaches zero as the distance between $x$ and $x^\prime$ increases.
\subsection{Inference}
For inference, we need to distinguish between the single and multiple replication situations. We will begin by assuming multiple replications: $\hat{A}_1(t), \ldots, \hat{A}_N(t)$.
We can consider the replicates $\hat{A}_i(t)$ as functional data. It is helpful to discretise the time/level scale of the estimator (\ref{eqn:nelaal}) to a grid of $M$ distinct levels $\tau_1=t_1,\ldots,t_M=\tau_2$. This is a common assumption in functional data analysis and deals effectively with the discontinuities in our estimators, which with probability one will not occur at measurement points.
There are then a huge array of analysis techniques immediately available: see for example the excellent review by \cite{wang16} for general approaches, and
\cite{chiou09} for a specific example of treating discretised hazards as functional data. With our applications in mind, we will mention only the construction of simultaneous confidence bands. This problem has had attention for functional data that are either sparse \citep{yao05, ma12} or dense \citep{degras11, cao12}. Given that our discretisation is by design, a dense functional data approach is appropriate.
Let $\hat{A}_{ij}$ be the Nelson-Aalen estimator for field $\mathcal{Z}_i$ at level $t_j$. Let $\bar{A}_j$ and $\hat{\sigma}_j^2$ be the sample mean and sample variance of the estimators at $t_j$, and let $\hat{\rho}_{jk}$ be the estimated correlation between $\hat{A}_{.j}$ and $\hat{A}_{.k}$. \cite{degras11}
establishes that for fixed $\tau_1$ and $\tau_2$, as both $N$ and $M$ increase
\begin{equation*}
\bar{A}_j\pm c_\gamma\hat{\sigma}_j/\surd N
\label{eqn:ci}
\end{equation*}
has asymptotic coverage $1-\alpha$, where $c_\alpha$ is the upper $\alpha$-quantile
of the maximum absolute value over $(\tau_2-\tau_1)/\tau_2$ of a Gaussian process with standard margins and with correlation function equal to an appropriately scaled limit of $\hat{\rho}_{jk}$. The threshold $c_\gamma$ needs to be obtained numerically: software is available in the R package {\tt SCBmeanfd} \citep{scb16}.
Turning to the case of a single replication only, we assume a parametric model $F_\theta$ is available for the random field $\mathcal{Z}$ and we propose a parametric bootstrap approach. Parametric and more general bootstrap methods for functional data have been studied empirically and theoretically by, for example, \cite{cuevas06}. We require
a consistent estimator $\hat{\theta}$ of the parameter vector $\theta$. We then generate $B$ simulated realisations of $\mathcal{Z}$ on $\mathcal{X}$ from $F_{\hat{\theta}}$, and obtain the bootstrap Nelson-Aalen estimators $\hat{A}_b(t)$ for $b=1,\ldots,B$. Pointwise intervals that have correct coverage asymptotically in $B$ and $\mid \mathcal{X} \mid$ can be obtained from the bootstrap quantiles or a Normal approximation justified by the central limit theorem.
For simultaneous confidence bands we suggest a Monte Carlo method used by \cite{crain11} that avoids explicit estimation of the correlation function. We again discretise the original and bootstrap estimators $\hat{A}_b(t)$ to a dense grid of $M$ points over a fixed range of levels $(\tau_1, \tau_2)$. Let $\hat{A}_j$ be the Nelson-Aalen estimate (\ref{eqn:nelaal}) at level $t_j$ obtained from the original data, and let $\hat{A}_{bj}$ be corresponding value from bootstrap replicate $b$. Let $\bar{A}_j$ and $\hat{\sigma}_j$ be the bootstrap mean and estimated standard deviation at $t_j$. Define
\[ G_b= \max_j\{ \mid \hat{A}_{bj}- \bar{A}_j \mid/\hat{\sigma}_j \},\]
and take $\hat{d}_\alpha$ to be the upper $\alpha$-quantile of the empirical distribution of $G_1, \ldots, G_B$. Then for large $M$ an approximate simultaneous confidence band for $A(t)$ over $(\tau_1, \tau_2)$ is
\[ \hat{A}_j\pm \hat{d}_\alpha \hat{\sigma}_j \hspace*{1cm} j=1,\ldots, M.\]
\subsection{Simulation studies}
We consider random fields on $d\times d$ lattices and investigate three models. The first,
M1, is a stationary and isotropic Gaussian random field with standard N(0,1) marginals and Mat\'{e}rn correlation function. If locations $x$ and $x'$ are separated by Euclidean distance $u$ then
\[ \mbox{cor}\left(z_x,z_{x'}\right)= \frac{2^{1-\nu}}{\Gamma(\nu)}\left(\sqrt{2\nu}u/\eta\right)^\nu K_\nu\left(\sqrt{2\nu} u/\eta\right), \]
where $K_\nu(.)$ is a modified Bessel function of the third kind. Let $\theta_1=(\eta, \nu)$.
The second model, M2, is a marginally transformed $\chi^2_1$ field. We simulate from the Gaussian random field model M1, with parameters $\theta_2$. Values are then squared to give $\chi^2_1$ margins. Let $y_x$ be the value at location $x$. We next marginally transform to N(0,1) using $z_x=\Phi^{-1}\{F_1(y_x)\}$, where $\Phi(.)$ and $F_1(.)$ are the N(0,1) and $\chi^2_1$ cumulative distribution functions, respectively.
The third model, M3, is a marginally transformed $F_{3,3}$ field. We begin by constructing two independent $\chi^2_3$ fields, each from the sum of three squared independent
Gaussian random fields M1 with Mat\'{e}rn parameters $\theta_3$. The ratio of these $\chi^2_3$ fields has $F_{3,3}$ distribution and we again marginally transform back to N(0,1).
All simulated fields are mean-corrected to help with the comparisons. The purpose of the marginal transformations is to generate fields that are not Gaussian random fields but which have Gaussian N(0,1) marginal distributions. This is so that we can assess the added value of topological methods over and above simple comparisons of marginal distributions. Similarly, in our simulations, for each choice of correlation $\theta_1$ for
model M1 we choose $\theta_2$ and $\theta_3$ so as to match as closely as we can the correlation of the final M2 and M3 fields to Mat\'{e}rn with parameters $\theta_1$. Details of this are presented in Supplementary Material. Again, the point is to show that differences between fields are not explained by differences in correlation functions. We have no particular interest in M2 and M3 other than as easily generated non-Gaussian random fields to be used as comparators.
The previously-presented Fig. \ref{fig:intro1} gives in panel (a) an example simulation from M1 on a $60 \times 60$ lattice with $\theta=(5,1).$ Panels (b) and (c) provide two examples of M3, and panel (d) an example of M2, with matched correlations. The topological Nelson-Aalen plots for these fields are in Fig. \ref{fig:intro2}. The solid line is the limiting value \eqref{eqn:limit}, which we obtained using the multivariate Gaussian distribution function routine in the R package {\tt mvtnorm}.
Table \ref{tab:cov1} shows coverage of nominal 95\% confidence intervals and bands based on Nelson-Aalen plots for connected components from Gaussian random fields M1 simulated on a 60$\times$60 lattice. Pointwise coverage is evaluated at selected percentiles of the at-risk distributions, ie levels $t$ satisfying $E\{Y_x(t)\}=p$ for five different values of $p$. The simultaneous confidence bands cover the central 90\% of the at-risk distributions, discretised to a grid of 200 equally spaced points. For each parameter combination we give results for three methods. The first, for reference, is a naive approach that assumes standard methods based on the variance estimator (\ref{eqn:var}) can be applied and using \cite{borgan87} for simultaneous confidence bands. The second method assumes we have
a single replication and uses a parametric bootstrap with $B=200$ simulated Gaussian random fields, generated with Mat\'{e}rn correlation parameters taken as the maximum likelihood estimates from the sole replicate.
The third method assumes we have replicates. The standard and parametric bootstrap results are based on 1000 single simulations, the results with replications are from 1000 batches of $N=40$ replicates, which matches an application in the next subsection.
There is severe undercoverage at times
for the standard confidence intervals, but the other methods both produce good results.
Simulation results for M2 and M3 are provided in Supplementary Material, along with further simulations to assess size and power when using our method to test for a Gaussian random field.
\begin{table}
\begin{center}
\caption{Coverage of nominal 95\% pointwise confidence intervals and 95\% simultaneous confidence band (SCB) around Nelson-Aalen component plots for Gaussian random fields. Results from 1000 simulations on $60 \times 60$ lattices, with Mat\'{e}rn correlation function with parameters $\eta$ and $\nu$ }
\label{tab:cov1}
\begin{tabular}{clrrrrrr}
& & \multicolumn{5}{c}{Percentile}\\
$(\eta, \nu)$ & Method & 0.9 & 0.7 & 0.5 & 0.3 & 0.1 & SCB \\
\\
(5,1) & Standard & 84.0& 92.2& 93.5& 88.9& 82.2 & 73.9\\
&Par. bootstrap& 95.3 &94.4 & 95.8& 95.3& 96.2& 94.1\\
&Replications & 94.2 &94.7 &95.7 &93.9 &94.7 &94.4\\
\\
(10,1) & Standard & 67.8 & 82.7 & 83.3 & 73.7 & 65.5 & 39.1\\
&Par. bootstrap& 95.3 &95.1 & 95.6 & 95.7 & 94.7 & 93.6 \\
&Replications & 95.1 &94.0& 94.9 & 95.9 & 94.0 & 96.4\\
\\
\\
(5,2) & Standard & 93.4 & 94.6 & 95.0 & 93.2 & 90.8 & 82.9\\
&Par. bootstrap & 95.4 &95.2& 95.7& 95.5 &95.8& 95.9\\
&Replications & 94.0 & 94.6 &95.0& 94.3& 93.7 & 93.4\\
\end{tabular}
\end{center}
\end{table}
In Table \ref{tab:class} we provide a brief comparison between using the proposed Nelson-Aalen estimator and existing topological data analysis techniques when discriminating between random fields, all on 60$\times$60 lattices. The statistical properties of topological summaries such as persistence diagrams and landscape functions are not understood, which means we cannot use these methods to assess the adequacy of an assumed model such as a Gaussian random field, or for single-sample problems. However, if we have two or more groups of replicated data we can compare classification performance.
We took training data to be batches of 40 random fields simulated from models M1, M2 and M3, for each of the three parameter choices. We then used a support vector machine to develop pairwise probabilistic classifiers, using the default parameter choices for the {\tt svm} routine in the R package {\tt e1071}. We built classifiers based on the proposed Nelson-Aalen estimator $\hat{A}(t)$ defined at (\ref{eqn:nelaal}) and on the first 10 landscape functions \citep{bubenik2015} for each of components and holes, choosing the best of these to compare with $\hat{A}(t)$. Landscape functions are popular summaries of persistence diagrams and are available through the R package {\tt TDA}. We replicated the training data 20 times and assessed classifier performance on separate groups of 200 simulated random fields for each model.
We assume test data would be allocated to the group with the highest class probability. In Table \ref{tab:class} we provide the accuracy of the classifier, which is the proportion of test data that are allocated to the correct group, and as a calibration measure we give the mean class probabilities for the correct groups.
For example, a classifier that estimates the probability of the correct group to be 0.51 gives accurate allocation but is poorly calibrated, whereas one that estimates the correct class probability to be 0.99 would be both accurate and well calibrated. The table shows that classification based on the Nelson-Aalen estimator consistently outperforms classification based on the landscape functions in both accuracy and calibration. Additional simulation results provided in Supplementary Material support this conclusion.
\begin{table}
\begin{center}
\caption{Classification of random fields. Accuracy (Acc.) measures the proportion of random fields that were correctly allocated and calibration (Cal.) gives the mean estimated class probabilities for the correct classes. Values are multiplied by 100 for presentation}
\label{tab:class}
\begin{tabular}{clrrrrrr}
Mat\'{e}rn &Method & \multicolumn{2}{c}{M1 v M2}& \multicolumn{2}{c}{M1 v M3}& \multicolumn{2}{c}{M2 v M3}\\
$(\eta, \nu)$ & & Acc. & Cal. & Acc. & Cal. & Acc. & Cal. \\
\\
(5,1) & Nelson-Aalen & 100.0 &97.7 & 99.4 & 96.6 & 100.0 & 97.5\\
& Landscape &96.5& 92.5& 91.7& 85.5& 95.4& 91.6\\
\\
(10,1) & Nelson-Aalen &99.7 &97.2 &95.2 &91.0 &100.0 &97.3\\
& Landscape &99.7& 97.0& 80.7& 72.0& 95.8& 90.1\\
\\
(5,2) & Nelson-Aalen &100.0 &97.6 &99.7 &97.1 &100.0& 97.4\\
& Landscape & 92.3& 87.7& 79.4& 71.0& 87.5 &80.1\\
\end{tabular}
\end{center}
\end{table}
\subsection{Applications: climate modelling, interstellar medium}
The Community Earth System Model Large Ensemble project \citep{kay2015community} is a publicly available ensemble of 40 high resolution climate model simulations. Although the models used to generate the data are deterministic, spatial and spatio-temporal statistical modelling is often applied as a means to both explore and concisely summarise the output \citep{cast16assess, edwards19}. Such models often assume that the modelled residuals form a Gaussian random field. We can use the topological Nelson-Aalen approach to assess this assumption, at least in part, and as a quick and easy method to compare ensemble members and to examine the suitability of assumed correlation structures.
To illustrate our method we considered annual wind intensities at $10$m for the year 2020. Each ensemble member has data on a 192$\times$288 grid over the surface of a sphere and at each data point we standardised by the mean and standard deviation across members. To keep the correlation matrices manageable for this illustration we used
every third lattice point in both the latitudinal and longitudinal directions. Following
\cite{castruccio2014beyond} and \cite{castruccio2016compressing} we also trimmed the regions near the poles, leaving data on a $51 \times 96$ lattice for each ensemble member.
Topological Nelson-Aalen plots for connected components are shown as functional boxplots \citep{sun11} in the left panel of Fig. \ref{fig:naexamples} together with a simultaneous confidence band for the mean. There is reassurance that the standard assumption of a Gaussian random field for these climate residuals is reasonable: the band almost uniformly includes the expected value for a Gaussian random field
with correlation function taken to be the empirical correlation found in the data. To illustrate the potential of our method as a diagnostic for fitted correlation models, the figure also includes expected values under three modelled correlation structures. These are the three dotted lines in the plot. In order from the top, the first line corresponds to a stationary
Mat\'{e}rn model. It is clearly poor for these data, and the same is true for the second dotted line, which corresponds to a regional block model with separate Mat\'ern correlation functions fitted to land, ocean and coastal regions. In contrast, as seen by the lowest dotted line, a semiparametric non-stationary model proposed by
\cite{konzen2019} fits these data well.
\begin{figure}
\centerline{\includegraphics{naexamplesfig.pdf}}
\caption{Nelson-Aalen plots for connected components. In (a) the grey region and lines represent functional boxplots of 40 separate realisations of climate data residuals. The long dashed lines represent the mean and associated 95\% simultaneous confidence band, and the other lines are limiting values for Gaussian random fields with different correlation matrices. The solid line uses the empirical correlation matrix, the others are explained in the text. Panel (b) considers the neutral hydrogen data: the short dashed lines are Region 1, the dotted lines Region 2 and the long dashed lines Region 3. The shaded regions indicate $\pm$ two standard deviations. The solid lines
show the limiting values for Gaussian random fields with matched correlations. The upper lines, which are almost indistinguishable, are for Regions 1 and 2, the lower is for Region 3. }
\label{fig:naexamples}
\end{figure}
Our second application is into the distribution in the Milky Way of neutral atomic hydrogen, \textup{H}\,{\sc i}, using data obtained by the Galactic All-Sky Survey using the Parkes 64 m radio telescope
\citep{mclure09, kalberla10}. Neutral atomic hydrogen makes up around 90\% of the atoms in the interstellar medium, which embeds stars and galaxies. Turbulence and shocks lead to a highly heterogeneous random pattern to the local density of \textup{H}\,{\sc i}, and a common assumption in astrophysics is that after the removal of trends and large-scale features, the distribution of the log-density of \textup{H}\,{\sc i} is that of a Gaussian random field \citep{elmeg04, monin07}. This assumption was investigated by \cite{hend2020}, who considered topological features of three distinct regions of the southern sky. Based on counts of components and holes, but not event history, the authors concluded that two regions were consistent, after monotonic transformation, with Gaussian random fields but the third was not.
The right panel of Fig. \ref{fig:naexamples} shows Nelson-Aalen plots for connected components for these data. Each region consists of density measurements over a 256$\times$256 lattice and following \cite{hend2020} we considered the residuals after trend removal using a thin plate smoothing spline. The data were again marginally transformed to N(0,1) to preempt questions as to whether simple monotonic transformations could suffice to convert the data to Gaussian random fields. Region 3 is clearly distinct from the others and not consistent with the fitted Gaussian random field. Regions 1 and 2 have very similar Nelson-Aalen plots and almost indistinguishable limiting values for Gaussian random fields with Mat\'ern correlations fitted to the data. There is evidence however that at the higher levels these regions are not consistent with the fitted Gaussian random fields.
\section{COX PROPORTIONAL HAZARDS FOR EMBEDDED METRIC TREES}
\subsection{Topology and event history for tree data}
We now turn to trees $\mathcal{T}=\{\mathcal{N},\mathcal{E}\}$ made up of nodes $\mathcal{N}$ and associated edges $\mathcal{E}$ embedded in $\mathbb{R}^2$ or $\mathbb{R}^3$, and with a single common ancestor. A simple example was shown in panel (a) of Fig. \ref{fig:trees}.
A number of approaches for the comparison of tree structures have been inspired by persistent homology. \cite{kanari2018} considered persistence of one-dimensional features on contours equidistant from the root, whether path length or radial distance, as previously illustrated in Panel (a) of Fig. \ref{fig:trees}. \cite{LI2017} independently proposed a similar approach in the comparison of neuronal tree structures. \cite{BENDICH2016}
applied persistent homology methods to brain artery trees in $\mathbb{R}^3$, \cite{MAOLI2017} used a geodesic distance function to derive persistence diagrams for the branching architectures of plants, and
\cite{BRODZKI2018} used persistent homology to classify bronchial trees.
None of the methods of which we are aware can deal with the possibility of censored or incomplete trees. For example, an image may capture only part of a tree, in which case edges that cross the image boundary would not be fully observed. Edges that reach the boundary are not true first order nodes or leaves, and should
not be considered as birth or death locations when applying any of the persistent homological techniques. Equally they do contain information and these edges should not be ignored or discarded.
This type of problem is of course fundamental to event history and survival analysis methodology, where methods for censored or truncated data are very well established. Hence our proposal is straightforward: construct a filtration as above but consider as an event the occurrence of either a leaf or a branching point. More specifically we propose the following.
\begin{enumerate}
\item Filter outwards by radial distance $r$.
\item Consider a leaf event to occur at $r$ if a known leaf has that radial distance from the root. Consider a branch event similarly. If desired, we might distinguish bifurcations from trifurcations and so on. Censored points cannot be events.
\item Construct a risk set consisting of all edges in $\mathcal{E}$ that intersect a circle of radius $r$ centered at the root, as in panel (a) of Fig. \ref{fig:trees}.
\item Use standard event history methods to contrast events with risk sets.
\end{enumerate}
Let $\mathcal{C}_r$ be a disc of radius $r$ centred on the root and define $\mathcal{T}_r=\mathcal{T} \cap \mathcal{C}_r$. Covariates can be included provided they are exogenous or dynamically defined on
$\mathcal{C}_r$. Thus spatial location, branch order, edge weights, proximity to other branches in $\mathcal{C}_r$ and so on can be handled easily using a Cox or other event time regression model. Further, we can construct risk sets from multiple trees and use fixed or random effects (frailties) to include tree-to-tree comparisons.
For this method to work we need three common event history assumptions. First, that events occur independently given the risk sets and covariates. Second, that censoring is independent. Third, that whichever hazard or intensity model is assumed is appropriate.
Under these assumptions standard results apply, including asymptotics for Cox or other model estimates.
\subsection{Application: vascular patterns in the eye}
The left panel of Fig. \ref{fig:eyefig} shows the vascular pattern of a human eye, taken
from a high dimensional image provided by \cite{budai13}. The right panel of the figure shows a tree representation, with vertices joined by linear edges, a simplification of the optic nerve area to provide a root, and with the omission of vessels that appear to originate outwith the plot region. Vessels that reach the boundary of the plot region are marked as censored points. This is a two-dimensional projection of a three-dimensional tree and branches can cross. There are no cycles.
\begin{figure}
\centerline{\includegraphics{eyefig.pdf}}
\caption{Left panel: vascular patterns in a healthy human eye, with data taken from
\cite{budai13}. Right panel: tree representation. Censored edges are marked by points. }
\label{fig:eyefig}
\end{figure}
The data in Fig. \ref{fig:eyefig} are from one of the 15 fundus images of eyes of healthy patients provided by \cite{budai13}, who also give fundus images of 15 patients with diabetic retinopathy. We use our methods to compare diabetic retinopathy and healthy patients, and also to investigate the effect of a number of covariates. These are the width of the vessel, the Euclidean distance between nodes, the ratio of path length to Euclidean distance between nodes as a measure of curvature, the number of other nodes within radius 200 units of the current node (about 10\% of the image size), the generation order, and the azimuth to the root measured from the vertical. We assumed log-linear effects for all covariates except azimuth, for which we assumed a smooth effect via a $P$-spline with four degrees of freedom. We also included patient-level terms as either fixed effects or frailties. In total over the 30 patients there were 12590 edges. At this large sample size the fixed effect and frailty model fits were almost identical.
Table \ref{tab:cox} summarises some of the results. In the upper part we consider an event to be the termination of a vessel at a leaf, and in the lower part we consider branching points. Given the large sample, statistical significance is almost irrelevant and so to compare the magnitudes of effects we take the hazard ratio (hr) between
the 20\% and 80\% points of the covariate distributions.
As might be expected, large vessels are highly unlikely to terminate in a leaf but are more likely to branch. The hazard functions for both types of event fall as Euclidean distance from an earlier node increases. There is no real effect of either curvature of an edge or proximity to other nodes, but there is some evidence that branching becomes less likely as the generation of an edge increases. There is a modest azimuth effect on branching, with a plot of the smooth fit (not shown here) indicating a decrease in branching hazard as aziumth moves away from the vertical, fairly symmetrically in both directions.
Of most interest is the difference between patients. The estimated frailty terms for both leaf and branch events are shown in Fig. \ref{fig:frailfig}. There are relatively small differences between patients in vessel branching frequency, but large differences in the hazard for terminating in a leaf, given covariate effects. Interestingly, we have complete separation between healthy and diabetic retinopathy patients in these frailties.
\begin{table}[h]
\begin{center}
\caption{Cox proportional hazards results for eye trees}
\label{tab:cox}
\begin{tabular}{lrrrrr}
& \multicolumn{5}{c}{Event = leaf}\\
& Coefficient & SE & $\chi^2$ & df & hr\\
Width & -0.3849 & 0.0079 & 2369.51 & 1 & 0.02\\
Euclid & -0.0051 & 0.0002 & 797.16 & 1 & 0.41\\
Path length ratio & 1.0142 & 0.1260 & 64.76 & 1 & 1.05\\
Nodes within 200 & 0.0037 & 0.0037 & 1.02 & 1 & 1.02\\
Order & -0.0147 & 0.0049 & 8.90 & 1 & 0.86\\
Azimuth & & & 34.66 & 4 & 1.18\\
Patient effect & & & 823.98 & 29 & 2.29\\
\\
& \multicolumn{5}{c}{Event = branch}\\
& Coefficient & SE & $\chi^2$ & df & hr\\
Width & 0.0591 & 0.0016 & 1420.28 & 1 & 1.88\\
Euclid & -0.0058 & 0.0001 & 1874.37 & 1 & 0.37\\
Path length ratio & -0.9058 & 0.2406 & 14.18 & 1 & 0.95\\
Nodes within 200 & 0.0375 & 0.0033 & 128.54 & 1 & 1.21\\
Order & -0.0622 & 0.0051 & 149.35 & 1 & 0.54\\
Azimuth & & & 119.30 & 4 & 1.41\\
Patient effect & & & 103.60 & 29 & 1.37\\
\end{tabular}
\end{center}
\end{table}
\begin{figure}
\centerline{\includegraphics[height=3in,width=3in]{treefrail.pdf}}
\caption{Subject-level frailties for 15 diabetic retinopathy patients (triangles) and 15 healthy patients (circles).}
\label{fig:frailfig}
\end{figure}
\section{DISCUSSION}
There are a number of areas for further work. In Section 3 we suggested empirical variances for the Nelson-Aalen estimator when there are repeated samples of random fields, as in the climate example, and parametric bootstrap estimators when there was interest in a particular model, such as a Gaussian random field. In Supplementary Material, an adjusted parametric bootstrap assuming local misspecification as in \cite{Copas2005} is suggested and
provides a conservative approach for the case of a single sample and no model. However, it would be useful formally to develop a non-parametric variance estimator.
In both Section 3 and Section 4 we concentrated on the birth times of connected components. We might also consider death times, though the identification of risk sets is less straightforward in this case, which in turn means that limiting properties will be harder to derive. For complexes in two or higher dimensions, including the random fields of Section 3, we could consider higher order features. In two-dimensions it will be particularly simple to use our methods to investigate death times of holes, which as a result of Alexander duality \citep{edel12} can be analysed in the same manner as births of components simply by reversing the direction of the filtration.
\section{ACKNOWLEDGEMENT}
We thank Evandro Konzen and Jian Shi for providing software to fit their non-stationary correlation model to the climate data.
\clearpage
\bibliographystyle{biometrika}
| 20163f3ae1ff7822511a38fa3e239fc7137a94f6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\label{Introduction}Introduction}
Although FL is born for high efficient distributed model building, it’s not as efficient enough as we thought. Many existing methods may perform well for IID data setting, but non-ideal for non-IID data setting. In these methods,thier continuous global model sharing strategy, which makes them over focusing on global data distribution fitting but ignoring local data distribution. In FL, the methodology for this kind of problem is called personalization. To improve the personalization performance of FL, we bring the double head design in this paper. This idea is inspired from the multi-task learning, where multiple heads design \citep{DBLP:journals/corr/LiH16e, DBLP:journals/corr/abs-1905-07835} is introduced to remember the model info from history missions. Similarly, we expect our double head design will achieve a good balance for model optimization between global and local data distribution.
Communication efficiency is another concerned topic in FL. Here we create the gradual sharing design to handle it. We get this idea from the analysis of model convergence rules. As we know, clients’ models share a similar convergence direction during their initial rounds of training. So it is not necessary to share model in this stage during this stage. In addition, we notice that once a frozen model restarts to share, it could recover to the normal performance effectively. All this gives us confidence to try the gradual sharing design instead of always model sharing in FL to save communication cost.
We summarize three main contributions of our method as below.
\begin{itemize}[leftmargin=*]
\item We propose a double head design for FL, which makes our FL model benefit from the local and global data simultaneously.
\item We introduce a gradual model sharing strategy for FL to save its communication cost greatly.
\item We split the data distribution setting over train and test data, which contributes to better evaluation for FL methods over various data distribution settings.
\end{itemize}
The rest of our paper is organized as follow. In Sec.\ref{RelatedWork} we overview existing methods towards FL efficient problems. In Sec.\ref{OurDesign} we introduce our creative designs in detail. In Sec.\ref{Experiments} extensive experiments are made to evaluate our method. Finally, in Sec.\ref{Conclusion}, we summarize our method.
\section{\label{RelatedWork}Related Work}
Since the concept of FL was proposed by Google \citep{mcmahan2017communication}, personalization effectiveness and communication efficiency have always been concerned by researchers.
\subsection{Personalization Effectiveness}
As mentioned before, the classic FL method has the drawback of handling non-IID data \citep{10.1145/3286490.3286559}. Roughly speaking, the principle idea for this problem is located at heterogeneous model parameter optimization, which is attempting to adapt a global model by fine-tuning, data-sharing or model-mixture. For the fine-tuning method \citep{wang2019federated, li2020lotteryfl}, a global model is given to each client in FL, then each cliect retrains the model with a new objective function based on their local data.. In the data-sharing method, some researchers \citep{DBLP:journals/corr/abs-1806-00582} try to upload a small part of local data to the server to tackle the non-IID problem, although it’s an obvious violation of data privacy. Other researchers attempt to build a generative model \citep{DBLP:journals/corr/abs-1811-11479} to cope with clients’ sample insufficient problem under the non-IID setting. As to the model-mixture method, researchers \citep{hanzely2020federated, deng2020adaptive, huang2020personalized} try to optimize the FL’s overall performance on global and local data distributions by ensembling models on server and clients. Other researchers attempt to achieve this by freezing data sharing on the last layer \citep{yang2020heterogeneity} of clients’ model. Currently, most of methods achieve better personalization performance at the price of more complicated computation cost and a sacrifice of performance on the IID data setting.
\subsection{Communication Efficiency}
In the field of multi-task learning, some researchers aim to alleviate communication cost by compressing the communication content or reducing model update frequency \citep{tang2020communication}. Similar ideas continue in the FL area. At first, researchers \citep{DBLP:journals/corr/KonecnyMYRSB16, DBLP:journals/corr/Alistarh0TV16} propose to compress the model through a combination of quantization, subsampling, or encoding before sending it to the server. Then, other researchers \citep{DBLP:journals/corr/abs-1804-03235, DBLP:journals/corr/abs-1811-11479, yu2020salvaging} try to implement knowledge distillation \citep{hinton2015distilling} to save communication cost for the model aggregation. Recently, besides choosing sketched info to communicate over clients and server \citep{DBLP:journals/corr/abs-1903-04488}, more researchers start to optimize the communication efficient in FL by only synchronizing sub-model of the base model \citep{liang2020think, yang2020heterogeneity, li2020lotteryfl}.
Several methods bear similar ideas with us. In method \citep{yang2020heterogeneity}, the last layer of the model is locked while training. As for method \citep{liang2020think}, the local model is divided into local and global part, the local part's model sharing is forbidden while training. In our method, we take similar measures of constraining the model sharing across layers, while our specific freeze strategy is quite different from them. In addition, our method achieves better stable performance across various data distribution settings than those two methods. The method \citep{wang2020federated} also has a layer-wise model sharing strategy, which seems like our gradual sharing design. However, it is optimized for the probabilistic model and can be seen as an extension for the bayesian nonparametric FL \citep{yurochkin2019bayesian} in CNN and LSTM setting. As a contrast, our method aims for a general model setting, and owns a completely different communication saving strategy.
\section{\label{OurDesign}Our Design}
\begin{figure*}[htbp]
\subfigure{
\begin{minipage}[t]{0.47\linewidth}
\centering
\includegraphics[width=1\linewidth]{pic/DoubleHeadDesignNew.png}
\setcounter{figure}{0}\caption{\label{fig:DoubleHeadDesign}Double Head Design}
\end{minipage}
}
\quad
\subfigure{
\begin{minipage}[t]{0.45\linewidth}
\centering
\includegraphics[width=1\linewidth]{pic/GradualSharingDesignNew.png}
\caption{\label{fig:GradualSharingDesign}Gradual Sharing Design}\setcounter{figure}{3}
\end{minipage}
}
\end{figure*}
\subsection{Double Head}
As discussed before, our double head design is originated from the multiple head design in multi-task learning. These two heads are responsible for fitting the global and local data distribution respectively. For the sake of convenience, we take a sketch map of a classification task's network as an example to illustrate our theory in the follow-up, as shown in Figure \ref{fig:DoubleHeadDesign}. In our design, the network of the model could be divided into two main parts: the base part and the specific part.In the specific part, there're two tracks of network structure, each of which is consisted with several full connected layers and a softmax layer. These tracks are called the global and the local head respectively. Only the local head is forbidden model sharing while training, other parts of the model will always join model sharing. We formula the global and local heads’ output as
\begin{equation}
f_g(\overrightarrow{u})_i = \frac{e^{u_i}}{\sum_{j=1}^Ce^{u_j}}, \mbox{ for } i = 1, ..., C \mbox{ and } j = 1, ..., C
\end{equation}
\begin{equation}
f_l(\overrightarrow{v})_i = \frac{e^{v_i}}{\sum_{j=1}^Ce^{v_j}}, \mbox{ for } i = 1, ..., C \mbox{ and } j = 1, ..., C
\end{equation}
where the \(f_g\) and \(f_l\) are the output of a client model’s global and local head. \(\overrightarrow{u}\) and \(\overrightarrow{v}\) represent the penultimate output from the two heads. \(u_i\) and \(v_i\) mean the \(i\)-th element of vector \(\overrightarrow{u}\) and \(\overrightarrow{v}\). Parameter \(i\) and \(j\) are index of classes range from \(1\) to \(C\).
The final prediction result of the model is extracted from the output of these two heads’ softmax layers. We take the index of max value in the two output vectors as result, which is formulated as
\begin{equation}
{pred} = \mathop{\arg\max}_i(f_g(\overrightarrow{u})_i \oplus f_l(\overrightarrow{v})_i)
\end{equation}
where \(f_g\), \(f_l\), \(\overrightarrow{u}\) and \(\overrightarrow{v}\) bear the same meaning as in previous formulas; \(\oplus\) represents for the concatenation operation.
The complete pseudo-code of our double head design is given in Algorithm \ref{algorithmDHFL}, where \(K\) is clients' quantity in out FL system, \(T\) is the round number of model training, \(E\) is the number of local epochs, \(\eta\) represent the learning rate, \(w\) is the weight of server's model, \(wb, wg\) represent for the model weights of the base part and the global part, and \(w'\) is the weight of clients' model.
\begin{figure}
\begin{minipage}{0.5\textwidth}
\begin{algorithm}[H]
\caption{\label{algorithmDHFL}Double Head Design}
\SetKwProg{ServerExecutes}{Server executes}{:}{end}
\SetKwProg{ClientUpdate}{ClientUpt}{:}{end}
\SetKwProg{ClientInit}{ClientInit}{:}{end}
\SetKwFunction{ClientUpdateFn}{ClientUpt}
\SetKwFunction{ClientInitFn}{ClientInit}
\SetKwFunction{concat}{concat}
\SetKwFunction{split}{split}
\SetKwFunction{valueFn}{value}
\ServerExecutes{}{
initialize \(w_0 \rightarrow\) \concat{\(wb_0, wg_0\)} \\
\ForEach{client \(k = 1, ..., K\) in parallel}{
\ClientInitFn{\(k, w_0\)} \\
}
\For{\(t = 1, 2, ..., T\)}{
\(s_t \leftarrow (\mbox{random set of } m \mbox{ clients})\) \\
\ForEach{client \(k \in s_t\)}{
\(w_{t+1}^k \leftarrow\) \ClientUpdateFn{\(k, w_t\)} \\
}
\(w_{t+1} \leftarrow \sum_{k=1}^K\frac{n_k}{n}w_{t+1}^k\) \\
}
}
\ClientUpdate{\((k, w_t)\)}{
\(wb_t, wg_t \leftarrow\) \split{\(w_t\)} \\
\({wb'}_{t-1}^k, {wg'}_{t-1}^k, {wl'}_{t-1}^k \newline \leftarrow\) \split{\({w'}_{t-1}^k\)} \\
\({w'}_t^k \leftarrow\) \concat{\(wb_t, wg_t, {wl'}_{t-1}^k\)} \\
\For{local epoch \(i = 1, 2, ..., E\)}{
\({w'}_t^k \leftarrow {w'}_{t}^k - \eta \nabla l({w'}_t^k; D_k)\) \\
}
\({wb'}_{t}^k, {wg'}_{t}^k, {wl'}_{t}^k \leftarrow\) \split{\({w'}_{t}^k\)} \\
\({w'}_t^{t+1} \leftarrow\) \concat{\({wb'}_{t}^k, {wg'}_{t}^k\)} \\
\Return \(w_{t+1}^k\) to server \\
}
\ClientInit{\((k, w_0)\)}{
\(wb_0, wg_0 \leftarrow\) \split{\(w_0\)} \\
\({wb'}_0^k, {wg'}_0^k, {wl'}_0^k \leftarrow\) \valueFn{\(wb_0, wg_0, wl_0\)} \\
\({w'}_0 \leftarrow\) \concat{\({wb'}_0^k, {wg'}_0^k, {wl'}_0^k\)} \\
}
\end{algorithm}
\end{minipage}
\quad
\begin{minipage}{0.5\textwidth}
\begin{algorithm}[H]
\caption{\label{algorithmGSFL}Gradual Sharing Design}
\SetKwProg{ServerExecutes}{Server executes}{:}{end}
\SetKwProg{ClientUpdate}{ClientUpt}{:}{end}
\SetKwProg{GetMask}{GetMask}{:}{end}
\SetKwFunction{ClientUpdateFn}{ClientUpt}
\SetKwFunction{GetMaskFn}{GetMask}
\SetKwFunction{concat}{concat}
\ServerExecutes{}{
initialize \(w_0\) \\%\tcp*[l]{initialize server model}
\ForEach{phase \(p = 1, 2, ..., P\)}{
\(m_p \leftarrow\) \GetMaskFn{\(p\)} \\
\For{\(t = 1, 2, ..., T\)}{
\(s_t \leftarrow (\mbox{random set of } m \mbox{ clients})\) \\
\ForEach{client \(k \in s_t\)}{
\(ws_t \leftarrow m_p \otimes w_t\) \\
\(ws_{t+1}^k \leftarrow\) \ClientUpdateFn{\(k, p, ws_t\)} \\
\(w_{t+1}^k \leftarrow\) \concat{\newline\(ws_{t+1}^k, (1 - m_p) \otimes w_t\)} \\
}
\(w_{t+1} \leftarrow \sum_{k=1}^K\frac{n_k}{n}w_{t+1}^k\) \\
}
}
}
\ClientUpdate{\((k, p, ws_t)\)}{
\(m_p \leftarrow\) \GetMaskFn{\(p\)} \\
\({w'}_t^k \leftarrow\) \concat{\(ws_t, (1-m_p) \otimes {w'}_{t-1}^k\)} \\
\For{local epoch \(i = 1, 2, ..., E\)}{
\({w'}_t^k \leftarrow {w'}_t^k - \eta\nabla l({w'}_t^k; D_k)\) \\
}
\(ws_{t+1}^k \leftarrow m_p \otimes {w'}_t^k\) \\
\Return \(ws_{t+1}^k\) to server \\
}
\GetMask{\((p)\)}{
\(m \leftarrow (\mbox{initialize model size mask with } 0)\) \\
\ForEach{layer \(l = 1, 2, ..., p\)}{
\(m \leftarrow (\mbox{turn } l \mbox{ part of mask into } 1)\)
}
\Return \(m\)
}
\end{algorithm}
\end{minipage}
\end{figure}
\subsection{Gradual Sharing}
Unlike other FL methods' full size model sharing, our method implements a new gradual sharing design, which is actually a model gradual sharing strategy for FL. This design helps us achieve more communication saving without hurting FL’s accuracy performance. As shown in Figure \ref{fig:GradualSharingDesign}, we will freeze clients sharing model with the server in initial training stages. Then, along with the training, clients could be allowed to share their model in layer-wise manner from shallower to deeper at certain frequency. Similarly, to better illustrate our idea, we give its pseudo-code in Algorithm \ref{algorithmGSFL}, in which \(K, E, \eta, w, w'\) are same to Algorithm \ref{algorithmDHFL}, and \(P\) indicates the current model sharing phase, which is calculated from the gradual sharing frequency, with value ranges from 1 to number of layers.
\section{\label{Experiments}Experiments}
\subsection{Experimental Detail}
\paragraph{Datasets}
Two challenge datasets are used here to evaluate our method: the FEMINIST dataset \citep{DBLP:journals/corr/abs-1812-01097} and the TCP Traffic Classification (hereinafter called the ‘TTC’) dataset. FEMINIST is a public dataset of image classification, containing over 800k samples of 62 classes. Each sample (\(28 \times 28 \times 1\) pic) has its own writers and there are more than 3k writers in the dataset. Its data characteristics like: with plentiful enough classes; samples of the same class with various label types, are quite suitable for federated learning simulation. As for TTC dataset, it is a proprietary dataset from Huawei for traffic classification task. There are about 900k samples (\(10 \times 1\) vector) of 34 classes in it. It offers TCP traffic data which is captured and desensitized from two various data scenarios. This makes the evaluation on it more believable for industrial application.
According to industrial scenarios might met, we set three kinds of data distribution:
\begin{itemize}[leftmargin=*]
\item The \textbf{IID} setting: dataset distribution across clients is same. For FEMINIST, each client owns a dataset with the same amount of every writer’s samples. For TTC, each client is equally allocated data from two scenarios.
\item The \textbf{non-IID} setting: dataset distribution over clients is different, but each client’s data could cover the whole labels. For FEMINIST, clients are given whole label included data from different writers. For TTC, clients’ dataset are distributed from different data scenarios.
\item The \textbf{dispatch} setting: different clients own samples of uncrossed classes. This is a relative extreme non-IID setting for FL, but it is common for industrial scenario and discriminative enough for FL methods evaluation.
\end{itemize}
It's worth noting that, for most of current FL methods, the test data distribution will follow as the train data. But in real industrial scenarios, future data distribution is unknown, thus makes us to extend our test data into two basic modes, as follows.
\begin{itemize}[leftmargin=*]
\item The global mode: a new added test mode, where the distribution of test data will be close to the combination of all training data, and all clients model share the same test data.
\item The local mode: the classic test mode, where the distribution of test data on each client will inherit their local train data, and therefore each client’s test data is different.
\end{itemize}
\paragraph{Implementations}
To simplify compuation, we take a 5 layer network as the basic network, consisted of 2 convolution (conv) layers, 2 fully connected (fc) layers and a softmax layer in sequential. In our method, the first two conv + pooling layers are set as the base part and the rest layers are duplicated twice as the specific part. As a contrast, we take the Google’s FedAvg method \citep{mcmahan2017communication} as baseline, and also make a comparison with the HDAFL method \citep{yang2020heterogeneity}, who also declared itself better than FedAvg over non-IID data setting. All methods are implemented by Tensorflow v1.14 on Ubuntu 16.04. All models will be trained for over 400 rounds, and communication frequency is set at every 2 model iterations. In FEMINIST dataset, we simulate 10 clients, each has 10k samples. In TCC dataset, we have 2 clients, each owns around 200k samples.
\subsection{Double Head Effect}
\begin{figure}[htbp]
\centering
\subfigure[IID + global test]{
\begin{minipage}[t]{0.24\linewidth}
\label{fig:DBHeadGIid}
\centering
\includegraphics[width=1.1\linewidth]{pic/DBHeadGIid.png}
\end{minipage}
}
\subfigure[non-IID + global test]{
\begin{minipage}[t]{0.24\linewidth}
\label{fig:DBHeadGNonIid}
\centering
\includegraphics[width=1.1\linewidth]{pic/DBHeadGNonIid.png}
\end{minipage}
}
\subfigure[dispatch + global test]{
\begin{minipage}[t]{0.24\linewidth}
\label{fig:DBHeadGDispatch}
\centering
\includegraphics[width=1.1\linewidth]{pic/DBHeadGDispatch.png}
\end{minipage}
}
\subfigure[IID + local test]{
\begin{minipage}[t]{0.24\linewidth}
\label{fig:DBHeadLIid}
\centering
\includegraphics[width=1.1\linewidth]{pic/DBHeadLIid.png}
\end{minipage}
}
\subfigure[non-IID + local test]{
\begin{minipage}[t]{0.24\linewidth}
\label{fig:DBHeadLNoniid}
\centering
\includegraphics[width=1.1\linewidth]{pic/DBHeadLNoniid.png}
\end{minipage}
}
\subfigure[dispatch + local test]{
\begin{minipage}[t]{0.24\linewidth}
\label{fig:DBHeadLDispatch}
\centering
\includegraphics[width=1.1\linewidth]{pic/DBHeadLDispatch.png}
\end{minipage}
}
\caption{\label{fig:DoubleHeadEffect}Compare of model accuracy performance over FEMNIST dataset}
\end{figure}
We firstly explore the advantages of double head design. As shown in Figure \ref{fig:DBHeadGIid} and \ref{fig:DBHeadGNonIid}, under the global test mode, with IID and non-IID data setting, our method has similar accuracy performance with FedAvg and performs better than HDAFL. This result can be explained by the global test data includes all clients' data distribution. Methods with more global model sharing power will gain more effective info while training. Likewise, method with less global sharing ability will achieve less in this setting.The result for IID and non IID data setting under local test mode is shown in Figure \ref{fig:DBHeadLIid} and \ref{fig:DBHeadLNoniid}. Our method performs a bit worse than HDAFL, but better than FedAvg. This might due to local test data mode setting, where methods with stronger fitting for local data distribution could perform better. In Figure \ref{fig:DBHeadGDispatch} and \ref{fig:DBHeadLDispatch}, the result of the dispatch data setting further supports above conclusions. In global test mode, FedAvg gets much better performance than others. While on local test mode, our method and HDAFL have similar performance and outperform FedAvg greatly.
All this turns out our double head design enhance FL's fitting power for global and local data distribution simultaneously. Although it didn't get best accuracy result in all data settings, it performs more stable than other methods, which is more favorable in industry. After all, in industrial application, a better generalization performance is foremost, especially when future data distribution is unknown.
\subsection{\label{GradualSharingEffect}Gradual Sharing Effect}
We further explore the effectiveness of our gradual sharing design for FL. First of all, we want to evaluate the impact of gradual sharing frequency. For FEMNIST dataset, according to our preliminary estimate, client’s model may reach convergence status after around 400 rounds of training. Accordingly, we set three types of sharing frequencies: release one layer every 10 rounds, every 20 rounds, and every 80 rounds respectively.
\begin{figure}[htbp]
\centering
\subfigure[IID + global test]{
\begin{minipage}[t]{0.23\linewidth}
\label{fig:GSGIid}
\centering
\includegraphics[width=1.1\linewidth]{pic/GSGIid.png}
\end{minipage}
}
\subfigure[non-IID + global test]{
\begin{minipage}[t]{0.23\linewidth}
\label{fig:GSGNonIid}
\centering
\includegraphics[width=1.1\linewidth]{pic/GSGNonIid.png}
\end{minipage}
}
\subfigure[dispatch + global test]{
\begin{minipage}[t]{0.23\linewidth}
\label{fig:GSGDispatch}
\centering
\includegraphics[width=1.1\linewidth]{pic/GSGDispatch.png}
\end{minipage}
}
\subfigure[IID + local test]{
\begin{minipage}[t]{0.23\linewidth}
\label{fig:GSLIid}
\centering
\includegraphics[width=1.1\linewidth]{pic/GSLIid.png}
\end{minipage}
}
\subfigure[non-IID + local test]{
\begin{minipage}[t]{0.23\linewidth}
\label{fig:GSLNoniid}
\centering
\includegraphics[width=1.1\linewidth]{pic/GSLNoniid.png}
\end{minipage}
}
\subfigure[dispatch + local test]{
\begin{minipage}[t]{0.23\linewidth}
\label{fig:GSLDispatch}
\centering
\includegraphics[width=1.1\linewidth]{pic/GSLDispatch.png}
\end{minipage}
}
\subfigure[FEMINST]{
\begin{minipage}[t]{0.23\linewidth}
\label{fig:GSEffectTransmissionRate}
\centering
\includegraphics[width=1.1\linewidth]{pic/GSEffectTransmissionRate.png}
\end{minipage}
}
\subfigure[TTC]{
\begin{minipage}[t]{0.23\linewidth}
\label{fig:TTCGSEffectTransmissionRate}
\centering
\includegraphics[width=1.1\linewidth]{pic/TTCGSEffectTransmissionRate.png}
\end{minipage}
}
\caption{\label{fig:GradualSharingEffect}Evaluation of Gradual Sharing Effectiveness: (a)-(f) are compare of convergence over various sharing frequencies; (g) and (h) are communication saving compare over two datasets}
\end{figure}
Results in Figure \ref{fig:GradualSharingEffect} has shown that, no matter in what kind of data setting or test mode, the gradual sharing applied model would resume to the normal accuracy performance quickly, although there would be a temporary decline once a layer sharing was relaxed. It’s worth noting that the frequency is an experimental parameter ranges from 0 to N (various according to dataset scenario). Once it's set too big, the model would be risky to resume to the normal accuracy within limited training rounds.
We also provide the comparison results for communication saving evaluation. In Figure \ref{fig:GSEffectTransmissionRate}, our method could achieve over 50\% communication saving than other two methods when gradual sharing frequency is set at 80 rounds/layer. This proves that our gradual sharing design would help FL system to save communication cost greatly without obvious accuracy performance decline.
\subsection{\label{ComprehensiveCompareInTTCDataset}Evalutaion on TTC dataset}
The TTC dataset is extracted from the telecom scenario, so an evaluation on this dataset could give a better reflect of our method’s performance in real industrial scenarios. The quantitative results in Table \ref{tab:QCOfModelAccPerfOverTTCDataset} show that our method's good accuracy performance across different data settings. And the result in Figure \ref{fig:TTCGSEffectTransmissionRate} further proves that our method have an obvious advantage in communication saving. Besides, our method even performs better in TTC dataset than it is in FEMNIST dataset. This might due to its bigger model complexity induced by the double head design, which improves model’s ability to tackle a complex dataset like TTC.
\begin{table}
\centering
\caption{\label{tab:QCOfModelAccPerfOverTTCDataset}Quantitative compare of model accuracy performance over TTC dataset}
\begin{tabular}{l|lll|lll}
\toprule
& \multicolumn{3}{c}{Global Test} & \multicolumn{3}{c}{Local Test} \\
\cmidrule(r){2-7}
Name & IID & non-IID & dispatch & IID & non-IID & dispatch \\
\midrule
FedAvg & 0.8375 & 0.8341 & \textcolor{red}{0.7994} & 0.8503 & 0.8385 & 0.7374 \\
HDAFL & 0.8370 & 0.8021 & 0.4420 & 0.8498 & 0.8583 & 0.8792 \\
Our\_DH & \textcolor{red}{0.8544} & \textcolor{red}{0.8512} & 0.7495 & \textcolor{red}{0.8680} & \textcolor{red}{0.8722} & 0.8723 \\
Our\_DH+GS & 0.8430 & 0.8444 & 0.7472 & 0.8562 & 0.8704 & \textcolor{red}{0.8818} \\
\bottomrule
\end{tabular}
\end{table}
\section{\label{Conclusion}Conclusion}
Improving the personalization effectiveness and the communication efficient are always the research focus in FL. In this paper, we propose the double head design and the gradual sharing design to tackle these challenges. Our experimental results show that the double head design effectively enhance FL's accuracy performance under the non-IID data setting. And the gradual sharing design could save communication cost hugely without impeding model convergence. Although our method didn't get the best accuracy result over all data settings, it achieves a more stable performance across various test data settings compared to other SOTAs. This helps it to be more industry attractive.
\bibliographystyle{plainnat}
| 900508ae46ee30a54ff7c33fbc70ec4a99fa5634 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The classical Kida's formula relates the $\la$-invariants of CM-extensions of cyclotomic $\Zp$-extensions of number field (see \cite{Ki}; also see \cite{Iw}). An analogue of this formula for Selmer groups of elliptic curves was later discovered by Hachimori-Matsuno \cite{HM}. Subsequently, this has been generalized to Selmer groups of ordinary modular forms and more general Galois representations (for instances, see \cite{CDLSS,PW06}). A common phenomenon in deriving such a formula is that one has to assume an appropriate $\mu=0$ assumption. It is then natural to ask if such a formula still holds without the $\mu=0$ assumption. Indeed if the Galois representation possess a Galois-invariant lattice whose Selmer group satisfies $\mu=0$ hypothesis, one does have an analogous formula (see \cite[Remark 3.2.3]{Gr11}). However, this observation cannot be carried out for more general situation. Indeed, the work of Drinen \cite{Dr} (also see \cite{DD15}) has exhibited examples of Galois representations whose Galois-invariant lattices do not satisfy the $\mu=0$ assumption.
The aim of this paper is to at least provide a conceptual (but conjectural) explanation to the validity of Kida's formula when $\mu\neq 0$.
We briefly explain the idea behind our approach here, leaving the details to the body of the paper. Let $F_{\infty}$ be a strongly admissible pro-$p$ extension of $F$, where we write $H=\Gal(F_{\infty}/F^{\cyc})$ and $\Ga=\Gal(F^{\cyc}/F)$. Suppose for now we are in the $\mu=0$ situation. In other words, our dual Selmer group $X(A/F^{\cyc})$ is finitely generated over $\Zp$ (for simplicity of exposition here, we assume our Galois module has coefficients in $\Zp$; see the body of the paper for a more general context). A by-now standard argument then shows that the dual Selmer group $X(A/F_{\infty})$ is finitely generated over $\Zp\ps{H}$ (for instances, see \cite[Theorem 6.4(ii)]{CH} or \cite[Theorem 2.1]{CS12}). Furthermore, there is a well-known formula relating the $\Zp\ps{H}$-rank of $X(A/F_{\infty})$ with $\Zp$-rank of $X(A/F^{\cyc})$ (see \cite[Theorem 5.4]{HS} or \cite[Corollary 2.12]{Ho}). There are two approaches towards proving this formula. The first approach is via a descent argument comparing the cyclotomic Selmer groups under a finite $p$-base change. This approach made heavy usage of Kida's formula (see \cite[Theorem 3.1]{HV} or \cite[Theorem 5.4]{HS}). A second approach is a homological descent argument due to Howson (cf. \cite[Corollary 2.12]{Ho}). We emphasize that Howson's approach avoids use of Kida's formula. An important observation, which does not seem to be written down in literature (although the author believes this might be known) is that the above said rank formula can be used to recover Kida's formula. We briefly say a few words on this observation. For a subgroup $H'$ of $H$ with finite index, one has an equality $|H:H'|\rank_{\Zp\ps{H}}\big(X(A/F_{\infty})\big) = \rank_{\Zp\ps{H'}}\big(X(A/F_{\infty})\big)$. As seen above, the quantity $\rank_{\Zp\ps{H}}\big(X(A/F_{\infty})\big)$ is related to $\Zp$-rank of $X(A/F^{\cyc})$, and on the other hand, $\rank_{\Zp\ps{H'}}\big(X(A/F_{\infty})\big)$ can be related to $\Zp$-rank of $X(A/L^{\cyc})$, where $L$ is a finite $p$-extension of $F$ contained in $F_{\infty}$ with $L^{\cyc} = (F_{\infty})^{H'}$. Combining these three relations, we obtain Kida's formula.
Now we turn to the $\mu\neq 0$ situation. Inspired by the above discussion, the natural first step towards establishing a Kida formula is then to note that under the validity of the $\M_H(G)$-conjecture, Howson's proof can be adapted to establish an analogous formula relating the $\Zp\ps{H}$-rank of $X(A/F_{\infty})/X(A/F_{\infty})(p)$ over $F_{\infty}$ with $\Zp$-rank of $X(A/F^{\cyc})/X(A/F^{\cyc})(p)$ (see Proposition \ref{rank formula}). Consequently, we may combine this formula with the above strategy to obtain Kida's formula (see Theorem \ref{Kida formula theorem}).
Although this gives a conceptual explanation to Kida's formula, we have to confess that we are not able to establish the validity of such a formula (much less the $\M_H(G)$-conjecture) unconditionally.
Nevertheless, we at least content ourselves with recovering the original Kida formula for the $\mu=0$ case via the approach given here (see Theorem \ref{Kida formula theorem mu=0} and also discussion in Section \ref{elliptic review}).
We now give a brief description of the contents of each section of the paper. Section \ref{algebra} introduces some algebraic notations and results required for the paper. In Section \ref{Selmer}, we introduce the (strict) Selmer group following Greenberg \cite{G89} and prove a formula relating the ranks for the dual Selmer groups of the admissible extension and cyclotomic extension. Section \ref{Kida section} is where we prove our main result. Finally, in Section \ref{elliptic review}, we revisit the elliptic curve situation as illustration of our result.
\begin{ack}
The author like to thank John Coates for his interest and comments on the paper.
He would also like to thank the anonymous referee for many useful comments and suggestions on the article. Some part of the research of this article was conducted when the author was visiting the National University of Singapore and the National Center for Theoretical Sciences in Taiwan, and he would like to acknowledge the hospitality and conducive working conditions provided by these institutes.
The author's research is supported by the National Natural Science Foundation of China under Grant No. 11550110172 and Grant No. 11771164. \end{ack}
\section{Algebraic Preliminaries} \label{algebra}
Throughout the paper, $p$ will denote a fixed prime. Let $\Op$ be the ring of integers of a finite extension of $\Qp$. Fix a local parameter
$\pi$ for $\Op$ and denote by $k$ the residue field of $\Op$. Let $G$ be a compact pro-$p$ $p$-adic
Lie group without $p$-torsion. The completed group algebra of $G$ over $\Op$ is given by
\[ \Op\ps{G} = \plim_U \Op[G/U], \]
where $U$ runs over the open normal subgroups of $G$ and the inverse
limit is taken with respect to the canonical projection maps. It is well known that $\Op\ps{G}$ is
an Auslander regular ring (cf. \cite[Theorem 3.26]{V02} or \cite[Theorem A.1]{LimFine}) with no zero divisors (cf.\
\cite{Neu}). Therefore, it admits a skew field $Q(G)$ which is flat
over $\Zp\ps{G}$ (see \cite[Chapters 6 and 10]{GW} or \cite[Chapter
4, \S 9 and \S 10]{Lam}). Thanks to this, we can define the notion of $\Op\ps{G}$-rank of a finitely generated $\Op\ps{G}$-module $M$, which is in turn given by
$$ \rank_{\Op\ps{G}}(M) = \dim_{Q(G)} (Q(G)\ot_{\Op\ps{G}}M). $$
The module $M$ is then said to be a
\textit{torsion} $\Op\ps{G}$-module if $\rank_{\Op\ps{G}} (M) = 0$.
The completed group algebra $k\ps{G}$ is defined similarly as above.
By \cite[Theorem 3.30(ii)]{V02}, the ring $k\ps{G}$ is Auslander
regular. Furthermore, it has no zero divisors (cf. \cite[Theorem
C]{AB}). Therefore, one has a notion of $k\ps{G}$-rank defined as
above. Similarly, we shall say that
the module $N$ is a \textit{torsion} $k\ps{G}$-module if
$\rank_{k\ps{G}}(N) = 0$.
For a given finitely generated $\Op\ps{G}$-module $M$, denote by
$M(\pi)$ the $\Op\ps{G}$-submodule of $M$ consisting of elements
of $M$ which are annihilated by some power of $\pi$. Since the ring
$\Op\ps{G}$ is Noetherian, the module $M(\pi)$ is also finitely generated
over $\Op\ps{G}$. Thus, there exists a sufficiently large integer $r$ so
that $\pi^r$ annihilates $M(\pi)$. Following \cite[Formula (33)]{Ho}, we
define
\[\mu_{\Op\ps{G}}(M) = \sum_{i\geq 0}\rank_{k\ps{G}}\big(\pi^i
M(\pi)/\pi^{i+1}\big). \]
(For another alternative, but equivalent, definition, see
\cite[Definition 3.32]{V02}.) Note that the sum on
the right is a finite one by the above discussion.
Now if $M$ is a finitely generated $\Op\ps{G}$-module, then its homology groups $H_i(G,M)$ are finitely generated over $\Op$ (see \cite[Proof of Theorem 1.1]{Ho} or \cite[Lemma 3.2.3]{LS}). Hence the quantity $\rank_{\Op}\big(H_i(G,M)\big)$ is well-defined. In view of this observation, we can now state the following result of Howson (see \cite[Theorem 1.1]{Ho} or \cite[Lemma 4.3]{LimFine}).
\bp[Howson] \label{Howson}
Let $M$ be a finitely generated $\Op\ps{G}$-module. Then we have
\[ \rank_{\Op\ps{G}}(M) = \sum_{i=0}^d(-1)^i\rank_{\Op}\big(H_i(G,M)\big),\]
where here $d$ denotes the dimension of the $p$-adic group $G$.
\ep
In this paper, we are mostly interested in the class of $p$-adic Lie groups $G$ which contains a closed normal subgroup $H$ such that $G/H\cong\Zp$. The following relative formula for the $\mu$-invariant will be of use later.
\bl \label{mu compare Mf}
Let $G$ be a pro-$p$ compact $p$-adic group without $p$-torsion which contains a closed normal subgroup $H$ with the property that $\Ga:=G/H\cong \Zp$. Let $M$ be a finitely generated $\Op\ps{G}$-module. Then we have
\[ \mu_{\Op\ps{G}}(M) = \sum_{i=0}^d(-1)^i\mu_{\Op\ps{\Ga}}\big(H_i(H,M(\pi))\big).\]
\el
\bpf
See \cite[Lemma 2.2]{LimMHG}.
\epf
For a group $G$ defined as above, a finitely generated $\Op\ps{G}$-module $M$ is said to
\textit{satisfy the $\M_H(G)$-property} if $M_f:=M/M(\pi)$ is
finitely generated over $\Op\ps{H}$. It has been conjectured for
certain Galois representations coming from abelian varieties with
good ordinary reduction at $p$ or cuspidal eigenforms with good
ordinary reduction at $p$, the dual Selmer group associated to such
a Galois representation satisfies the $\M_H(G)$-property (see
\cite{BV, CFKSV, CS12, LimMHG}).
For the subsequent discussion, we shall write $M_H$ for the largest quotient of $M$ on which $H$ acts trivially. Note that $M_H = H_0(H,M)$. We now record another useful lemma.
\bl \label{rank compare Mf}
Let $G$ be a pro-$p$ compact $p$-adic group without $p$-torsion which contains a closed normal subgroup $H$ such that $\Ga:=G/H\cong \Zp$. Let $M$ be a finitely generated $\Op\ps{G}$-module which satisfies the $\M_H(G)$-property and has the property that $H_i(H,M)$ is finitely generated over $\Op$ for all $i\geq 1$. Then we have
\[ \mu_{\Op\ps{G}}(M) = \mu_{\Op\ps{\Ga}}\big(M_H\big)\]
and \[ \rank_{\Op}\big(H_i(H, M)\big) =\rank_{\Op}\big(H_i(H, M_f)\big)\]
for every $i\geq 1$.
\el
\bpf
Taking $H$-homology
of the following short exact sequence
\[ 0\lra M(\pi)\lra M\lra M_f\lra 0,\]
yields an exact sequence
\[ H_{i+1}(H, M_f)\lra H_{i}(H, M(\pi))\lra H_{i}(H,
M) \stackrel{f_i}\lra H_i(H, M_f)\lra H_{i-1}(H, M(\pi))\]
for $i\geq 1$. As $M_f$ is finitely generated over $\Op\ps{H}$, the group $H_i(H, M_f)$ is therefore finitely generated over $\Op$ for every $i$ (cf. \cite[Proof of Theorem 1.1]{Ho} or \cite[Lemma 3.2.3]{LS}). Combining this with the hypothesis that $H_i(H,M)$ is finitely generated over $\Op$ for all $i\geq 1$, we see that $H_i(H,M(\pi))$ is finitely generated over $\Op$ for $i\geq 1$. As noted before, there exists a sufficiently large $r$ such that $\pi^r$ annihilates $M(\pi)$, and hence all the $H$-homology groups of $M(\pi)$. Therefore, $H_i(H,M(\pi))$ must be finite for $i\geq 1$. From this and the above exact sequence, we have the second equality of the lemma. Furthermore, the finiteness implies $\mu_{\Op\ps{\Ga}}\big(H_i(H,M(\pi))\big) =0$ for $i\geq 1$. Putting these into Lemma \ref{mu compare Mf}, we obtain
\[ \mu_{\Op\ps{G}}(M) = \mu_{\Op\ps{\Ga}}\big(M(\pi)_H\big).\]
On the other hand, from the exact sequence
\[ H_{1}(H, M_f)\lra M(\pi)_H\lra M_H \lra H_0(H, M_f)\lra 0,\]
and the observation that each $H_i(H,M_f)$, being finitely generated over $\Op$, has trivial $\mu_{\Op\ps{\Ga}}$-invariants, we see that
\[ \mu_{\Op\ps{\Ga}}\big(M(\pi)_H\big) = \mu_{\Op\ps{\Ga}}\big(M_H\big).\]
Combining this with the above, we obtain the first equality of the lemma.
Finally, the finiteness of $H_i(H,M(\pi))$ (for $i\geq 1$) immediately gives the second equality for $i\geq 2$. For $i=1$, it remains to verify that $\mathrm{coker} f_1$ to be finite. But this follows from the facts that $\mathrm{coker} f_1$ is finitely generated over $\Op$ (being a quotient of $H_1(H, M_f)$) and is annihilated by $\pi^r$ (being a submodule of $M(\pi)_H$). The proof of the lemma is now completed.
\epf
\section{Selmer groups} \label{Selmer}
Throughout this section, we let $F$ be a number
field which is further assumed to have no real primes if $p=2$. As before, denote by $\Op$ the ring of integers of a fixed finite extension of $\Qp$. In this section, we define the strict Selmer groups associated to certain datum in the sense of Greenberg \cite{G89}. As a start, we introduce the axiomatic conditions on our datum which is denoted by $\big(A, \{A_v\}_{v|p}\big)$ and satisfies all of the following four conditions \textbf{(C1)-(C4)}.
\begin{enumerate}
\item[(\textbf{C1})] $A$ is a
cofinitely generated cofree $\Op$-module of $\Op$-corank $d$ with a
continuous, $\Op$-linear $\Gal(\bar{F}/F)$-action which is
unramified outside a finite set of primes of $F$.
\item[(\textbf{C2})] For each prime $v$ of $F$ above $p$, there is a
distinguished $\Gal(\bar{F}_v/F_v)$-submodule $A_v$ of $A$ which is cofree of
$\Op$-corank $d_v$.
\item[(\textbf{C3})] For each real prime $v$ of $F$, $A_v^+:=
A^{\Gal(\bar{F}_v/F_v)}$ is cofree of
$\Op$-corank $d^+_v$.
\item[(\textbf{C4})] The following equality
\[\sum_{v|p} (d-d_v)[F_v:\Qp] = dr_2(F) +
\sum_{v~\mathrm{real}}(d-d^+_v)\]
holds. Here $r_2(F)$ denotes the number of complex primes of $F$.
\end{enumerate}
As we need to work with Selmer groups defined over a tower of number fields, we need to consider the base change of our datum which we now do. For a finite extension $L$ of $F$, the base change of our datum $\big(A,
\{A_w\}_{w|p} \big)$ over $L$ is given as follows:
(1) $A$ can be viewed as a $\Gal(\bar{F}/L)$-module via restriction of the Galois action.
(2) For each prime $w$
of $L$ above $p$, we set $A_w =A_v$, where $v$ is the prime of $F$
below $w$, and view it as a $\Gal(\bar{F}_v/L_w)$-module via the appropriate restriction.
Note that we then have $d_w= d_v$.
(3) For each real prime $w$ of $L$ which lies above a real prime $v$ of $F$, we set
$A_w^+= A^{\Gal(\bar{F}_v/F_v)}$ and write $d^+_w
= d^+_v$.
We now record the following lemma which gives some
sufficient conditions for equality in \textbf{(C4)} to hold for
the datum $\big(A, \{A_w\}_{w|p}\big)$ over $L$.
\bl \label{data base change} Suppose that $\big(A, \{A_v\}_{v|p}\big)$ is a datum defined over $F$.
Suppose further that at least one of the following statements holds.
\begin{enumerate}
\item[$(i)$] $[L:F]$ is odd.
\item[$(ii)$] $F$ has no real primes.
\end{enumerate}
Then the datum $\big(A, \{A_w\}_{w|p}\big)$ obtained by base change satisfies \textbf{(C1)-(C4)}. In particular, we have the equality
\[ \sum_{w|p} (d-d_w)[L_w:\Qp] = dr_2(L) +
\sum_{w~\mathrm{real}}(d-d^+_w).\]
\el
\bpf See \cite[Lemma 3.0.1]{LimCMu}. \epf
We shall impose one more condition on our datum.
\medskip\textbf{(Fin)} $H^0(L^{\cyc}, A^*)$ is finite for every $L$ contained in $F_{\infty}$. Here $A^* =
\Hom_{\cts}(T_{\pi}(A),\mu_{p^{\infty}})$, where $T_{\pi}(A) =
\plim_i A[\pi^i]$.
\medskip
We mention two basic examples of our datum.
(i) $A = \mathcal{A}[p^{\infty}]$, where $\mathcal{A}$ is an abelian
variety defined over an arbitrary finite extension $F$ of $\Q$ with
good ordinary reduction at all places $v$ of $F$ dividing $p$. For
each $v|p$, it follows from \cite[P. 150-151]{CG} that we have a
$\Gal(\bar{F}_v/F_v)$-submodule $A_v$ characterized by
the property that $A/A_v$ is the maximal
$\Gal(\bar{F}_v/F_v)$-quotient of $\mathcal{A}[p^{\infty}]$ on which
some subgroup of finite index in the inertia group $I_v$ acts
trivially. It is not difficult to verify that \textbf{(C1)-(C4)} are
satisfied. The condition \textbf{(Fin)} is a well-known
consequence of a theorem of Imai \cite{Imai}.
(ii) Let $V$ be the Galois representation attached to a primitive
Hecke eigenform $f$ for $GL_2 /\Q$, which is ordinary at $p$,
relative to some fixed embedding of the algebraic closure of $\Q$
into $\Qp$. By the work of Mazur-Wiles \cite{MW}, $V$ contains a
one-dimensional $\Qp$-subspace $V_v$ invariant under
$\Gal(\bar{\Q}_p/\Qp)$ with the property that the inertial subgroup
$I_p$ acts via a power of the cyclotomic character on $V_v$ and
trivially on $V/V_v$. By compactness, $V$ will always contain a free
$\Op$-submodule $T$, which is stable under the action of
$\Gal(\bar{F} /F)$. For such an $\Op$-lattice $T$, we write $A =
V/T$ and $A_v = V_v/ (T\cap V_v)$. The condition \textbf{(Fin)} is shown in
the proof of \cite[Lemma 2.2]{Su}.
We return to general discussion.
A Galois extension $F_{\infty}$ of $F$ is said to be a strongly admissible pro-$p$ $p$-adic Lie
extension of $F$ if (i) $\Gal(F_{\infty}/F)$ is a compact pro-$p$ $p$-adic Lie group with no $p$-torsion, (ii) $F_{\infty}$ contains the cyclotomic $\Zp$-extension
$F^{\cyc}$ of $F$ and (iii) $F_{\infty}$ is unramified outside a finite set of primes. We shall always write $G = \Gal(F_{\infty}/F)$, $H = \Gal(F_{\infty}/F^{\cyc})$ and $\Ga =\Gal(F^{\cyc}/F)$. By Lemma \ref{data base change}, the datum $\big(A, \{A_w\}_{w|p}, \{A^+_w\}_{w|\R} \big)$ obtained by base changing to any intermediate subextension of $F_{\infty}/F$ satisfies \textbf{(C1)-(C4)}. (Recall that when $p=2$, our standing assumption is that $F$ has no real primes.)
We can now define the strict Selmer group associated to our datum following Greenberg \cite{G89}. Let $S$ be a finite set of
primes of $F$ which contains all the primes above $p$, the ramified
primes of $A$, the ramified
primes of $F_{\infty}/F$ and all the infinite primes of $F$. Denote by $F_S$ the
maximal algebraic extension of $F$ unramified outside $S$ and write
$G_S(\mathcal{L}) = \Gal(F_S/\mathcal{L})$ for every algebraic
extension $\mathcal{L}$ of $F$ which is contained in $F_S$. Let $L$
be a finite extension of $F$ contained in $F_{\infty}$. For a prime $w$ of $L$ lying over $S$, set
\[ H^1_{str}(L_w, A)=
\begin{cases} \ker\big(H^1(L_w, A)\lra H^1(L_w, A/A_w)\big), & \text{\mbox{if} $w$
divides $p$},\\
\ker\big(H^1(L_w, A)\lra H^1(L^{ur}_w, A)\big), & \text{\mbox{if} $w$ does not divide $p$,}
\end{cases} \]
where $L_w^{ur}$ is the maximal unramified extension of $L_w$.
We then define
\[ J_v(A/L) = \bigoplus_{w|v}H^1_s(L_w, A),\]
where $w$ runs over the (finite) set of primes of $L$ above $v$ and $H^1_s(L_w, A) = H^1(L_w, A)/H^1_{str}(L_w, A)$. If
$\mathcal{L}$ is an infinite extension of $F$, we define
\[ J_v(A/\mathcal{L}) = \ilim_L J_v(A/L),\]
where the direct limit is taken over all finite extensions $L$ of
$F$ contained in $\mathcal{L}$. For a (possibly infinite) extension $K$ of $F$ contained in $F_{\infty}$, the (strict) Selmer group is defined by
\[ S^{str}(A/K) := \Sel^{str}(A/K) := \ker\left( H^1(G_S(K),A)\lra
\bigoplus_{v\in S}J_v(A/K)\right).\]
It follows from the definition that $S^{str}(A/K) = \ilim_L S^{str}(A/L)$, where the direct limit is taken
with respect to the natural maps, as $L$ varies over all finite subextensions
of the base field $F$ in the larger extension $K$. Write $X(A/K)$ for its Pontryagin dual and $X_f(A/K) =X(A/K)/X(A/K)(\pi)$.
Suppose that $X_f(A/F_{\infty})$ satisfies the $\M_H(G)$-property. It then makes sense to speak of the quantity $\rank_{\Op\ps{H}}\big(X_f(A/F_{\infty})\big)$. Also, it follows from \cite[Proposition 2.5]{CS12} that $X(A/F^{\cyc})$ is torsion over $\Op\ps{\Ga}$ and so the quantity $\rank_{\Op}(X_f(A/F^{\cyc}))$ is well-defined. We can therefore state the following.
\bp \label{rank formula}
Let $F_{\infty}$ be a strongly admissible pro-$p$ Lie extension of $F$. Suppose that the data $\big(A, \{A_v\}_{v|p}, \{A^+_v\}_{v|\R} \big)$ satisfies $\textbf{(C1)-(C4)}$ and $\textbf{(Fin)}$. Assume that $X(A/F_{\infty})$ satisfies the $\M_H(G)$-property. Then we have
\[ \mu_{\Op\ps{G}}\big(X(A/F_{\infty})\big) = \mu_{\Op\ps{\Ga}}\big(X(A/F^{\cyc})\big)\]
and
\[ \rank_{\Op\ps{H}}\big(X_f(A/F_{\infty})\big) = \rank_{\Op}\big(X_f(A/F^{\cyc})\big) - \corank_{\Op}\big(A(F^{\cyc})\big) +
\sum_{\substack{w\in S(F^{\cyc}),
\\ \dim H_w\geq 1}}\corank_{\Op}\big(D_w(F^{\cyc}_w)\big).\]
Here $D_w$ denotes $A/A_w$ or $A$ accordingly as $w$ divides $p$ or not.
\ep
Proposition \ref{rank formula} has been proved for an elliptic curve $E$ under the stronger assumption that $X(E/F_{\infty})$ is finitely generated over $\Zp\ps{H}$ (see
\cite[Corollary 6.10]{CH}, \cite[Theorem 5.4]{HS}, \cite[Theorem
3.1]{HV} and \cite[Theorem 2.8]{Ho}). The approach for the proof in \cite{CH, HS, HV} builds on Kida's formula which we want to avoid for our purposes. On the other hand, the proof given in \cite[Theorem 2.8]{Ho} does not assume the validity of Kida's formula. The goal of Proposition \ref{rank formula} is to show that Howson's approach works under the $\M_H(G)$-property which the remainder of the section is devoted to.
As a start, we have the following lemma.
\bl \label{short exact sequences}
Retaining the assumptions of Proposition \ref{rank formula}, we have short exact sequences
\[ 0 \lra S^{str}(A/F^{\cyc})\lra H^1(G_S(F^{\cyc}), A) \lra \bigoplus_{v\in S} J_v(A/F^{\cyc})\lra 0\]
and
\[ 0 \lra S^{str}(A/F_{\infty})\lra H^1(G_S(F_{\infty}), A) \lra \bigoplus_{v\in S} J_v(A/F_{\infty})\lra 0.\]
\el
\bpf
Since $X(A/F_{\infty})$ satisfies the $\M_H(G)$-property, it follows from \cite[Proposition 2.5]{CS12} that for every finite extension $L$ of $F$ contained in $F_{\infty}$, $X(A/L^{\cyc})$ is torsion over $\Op\ps{\Ga_L}$, where $\Ga_L=\Gal(L^{\cyc}/L)$. In view of $\textbf{(Fin)}$, we may apply a similar argument to that in \cite[Proposition 3.3]{LimMHG} to obtain a short exact sequence
\[ 0 \lra S^{str}(A/L^{\cyc})\lra H^1(G_S(L^{\cyc}), A) \lra \bigoplus_{v\in S} J_v(A/L^{\cyc})\lra 0.\]
In particular, this yields the first short exact sequence by taking $L=F$.
On the other hand, by taking direct limit over $L$, we obtain the second short exact sequence.
\epf
The next two lemmas are concerned with the $H$-homology of global cohomology groups and local cohomology groups.
\bl \label{global calculation}
Retain the assumptions of Proposition \ref{rank formula}. Then $H^i\big(H, H^1\big(G_S(F_{\infty}),A)\big)$ is cofinitely generated over
$\Op$ for every $i\geq 1$. Moreover, we have an exact sequence
\[ 0 \lra H^1(H,
A(F_{\infty}))\lra H^1(G_S(F^{\cyc}), A) \lra H^1(G_S(F_{\infty}), A)^H\lra H^2(H,
A(F_{\infty}))\lra 0\]
and isomorphisms \[
H^i\big(H, H^1(G_S(F_{\infty}), A)\big)\cong H^{i+2}\big(H,
A(F_{\infty})\big) \mbox{ for } i\geq 1.\]
\el
\bpf
Since $X(A/F_{\infty})$ is assumed to satisfy the $\M_H(G)$-property, it follows from \cite[Proposition 2.5]{CS12} that for every finite extension $L$ of $F$ contained in $F_{\infty}$, $X(A/L^{\cyc})$ is torsion over $\Op\ps{\Ga_L}$, where $\Ga_L=\Gal(L^{\cyc}/L)$. Via similar arguments to those in \cite[Proposition 3.3 and Corollary 3.4]{LimMHG}, we have that
$H^2(G_S(F^{\cyc}),A)=0$ and $H^2(G_S(F_{\infty}),A)=0$. Hence the spectral sequence
\[ H^i\big(H, H^j(G_S(F_{\infty}), A)\big)\Longrightarrow
H^{i+j}(G_S(F^{\cyc}), A)\] degenerates to yield an exact sequence
\[ 0 \lra H^1(H,
A(F_{\infty}))\lra H^1(G_S(F^{\cyc}), A) \lra H^1(G_S(F_{\infty}), A)^H\lra H^2(H,
A(F_{\infty}))\lra 0\]
and isomorphisms \[
H^i\big(H, H^1(G_S(F_{\infty}), A)\big)\cong H^{i+2}\big(H,
A(F_{\infty})\big) \mbox{ for } i\geq 1.\]
Finally, the $\Op$-cofinitely generation of the latter
groups follows from the fact that the cohomology groups $H^i(H,W)$ are cofinitely generated
$\Op$-modules for any $p$-adic Lie group $H$ and any $\Op$-cofinitely generated
$H$-module $W$. Consequently, $H^i\big(H, H^1\big(G_S(F_{\infty}),A)\big)$ is cofinitely generated over
$\Op$ for every $i\geq 1$.
\epf
\bl \label{local calculation}
Retain the assumption of Proposition \ref{rank formula}. Then $H^i(H, \bigoplus_{v\in S}J_v(A/F_{\infty}))$ is cofinitely generated over
$\Op$ for every $i\geq 1$. Moreover, we have an exact sequence
\[ 0 \lra \bigoplus_{w\in S(F^{\cyc})}H^1\big(H_w,
D_v(F_{\infty,w})\big)\lra \bigoplus_{v\in S}J_v(A/F^{\cyc}) \lra \left(
\bigoplus_{v\in S}J_v(A/F_{\infty})\right)^H \] \[\lra \bigoplus_{w\in S(F^{\cyc})}H^2\big(H_w,
D_v(F_{\infty,w})\big)\lra 0\]
and isomorphisms
\[
H^i\left(H, \bigoplus_{v\in S}J_v(A/F_{\infty})\right)\cong\bigoplus_{w\in S(F^{\cyc})}H^{i+2}\big(H_w,
D_v(F_{\infty,w})\big) \mbox{ for } i\geq 1.\]
Here $D_v$ denotes $A/A_v$ or $A$ accordingly as $v$ divides $p$ or not.
\el
\bpf
This is a local version of Lemma \ref{global calculation} with a similar proof noting that
$H^2(F^{\cyc}_w,A)=0$ and $H^2(F_{\infty,w},A)=0$ by \cite[Theorem 7.1.8(i)]{NSW}.
\epf
We can now give the proof of Proposition \ref{rank formula}.
\bpf[Proof of Proposition \ref{rank formula}]
By Lemma \ref{short exact sequences}, we have the following commutative diagram
\[ \entrymodifiers={!! <0pt, .8ex>+} \SelectTips{eu}{}\xymatrix{
0 \ar[r]^{} & S^{str}(A/F^{\cyc}) \ar[d]_{\al} \ar[r] & H^1(G_S(F^{\cyc}), A)
\ar[d]_{\be}
\ar[r] & \displaystyle\bigoplus_{v\in S}J_v(A/F^{\cyc}) \ar[d]_{\ga} \ar[r] & 0 &\\
0 \ar[r]^{} & S^{str}(A/F_{\infty})^H \ar[r]^{} & H^1(G_S(F_{\infty}), A)^H \ar[r] & \
\displaystyle\bigoplus_{v\in S}J_v(A/F_{\infty})^H \ar[r] & H^1\big(H, S^{str}(A/F_{\infty})\big) \ar[r] & \cdots } \]
with exact rows. To simplify notation, we write $W_{\infty}=
H^1(G_S(F_{\infty}), A)$ and
$J_{\infty} = \displaystyle\bigoplus_{v\in S}J_v(A/F_{\infty})$. From the commutative diagram, we have a long exact
sequence
\[ \ba{c} 0\lra \ker\al \lra \ker \be
\lra \ker \ga
\lra \coker \al \lra \coker \be \\
\lra \coker\ga \lra H^1\big(H, S^{str}(A/F_{\infty})\big)
\lra H^1(H, W_{\infty})
\lra H^1(H, J_{\infty})\lra \cdots \\
\cdots\lra H^{i-1}(H, J_{\infty}) \lra H^i\big(H, S^{str}(A/F_{\infty})\big)
\lra H^i(H, W_{\infty})
\lra H^i(H, J_{\infty})\lra \cdots .\ea \]
By Lemmas \ref{global calculation} and \ref{local calculation}, the groups $\ker \be$, $\ker \ga$, $\coker \be$, $\coker\ga$, $ H^i(H, W_{\infty})$ (for $i\geq 1$) and $H^{i}(H, J_{\infty})$ (for $i\geq 1$) are cofinitely generated over $\Op$. Thus, combining this observation with the above exact sequence, we have that $\ker \al$, $\coker \al$ and $H^1\big(H, S(A/F_{\infty})\big)$ (for $i\geq 1)$ are cofinitely generated over $\Op$. Therefore, we have
\[ \mu_{\Op\ps{G}}\big(X(A/F_{\infty})\big) = \mu_{\Op\ps{\Ga}}\big(X(A/F_{\infty})_H\big) = \mu_{\Op\ps{\Ga}}\big(X(A/F^{\cyc})\big),\]
where the first equality follows from Lemma \ref{rank compare Mf} and the second follows from the facts that $\ker \al$ and $\coker \al$ are cofinitely generated over $\Op$. This establishes the first equality of the proposition. Moreover, we have
\[ \ba{c} \corank_{\Op}(\ker\al) - \corank_{\Zp}(\coker\al)= -\displaystyle\sum_{i\geq 1}(-1)^i\corank_{\Op}H^i(H,S^{str}(A/F_{\infty}))\hspace{2in} \\
\hspace{1in}+\displaystyle\sum_{i\geq 1}(-1)^{i}\corank_{\Op}H^i(H,
A(F_{\infty}))-\displaystyle\sum_{\substack{w\in S(F^{\cyc}),
\\ \dim H_w\geq 1}}\left(\sum_{i\geq 1}(-1)^{i}\corank_{\Op}H^i(H_w,
D_v(F_{\infty,w}))\right), \ea \]
where here $D_v$ denotes $A/A_v$ or $A$ accordingly as $v$ divides $p$ or not.
Applying Proposition \ref{Howson} and Lemma \ref{rank compare Mf}, the right hand side is just
\[ -\displaystyle\sum_{i\geq 1}(-1)^i\rank_{\Op}H_i(H,X_f(A/F_{\infty}))-\corank_{\Op}H^0(H,
A(F_{\infty}))
+\displaystyle\sum_{\substack{w\in S(F^{\cyc}),
\\ \dim H_w\geq 1}}\corank_{\Op}H^0(H_w,
D_v(F_{\infty,w})).
\]
Now consider the following commutative diagram
\[ \entrymodifiers={!! <0pt, .8ex>+} \SelectTips{eu}{}\xymatrix{
& X(A/F_{\infty})(p)_H \ar[d]_{h'} \ar[r] & X(A/F_{\infty})_H
\ar[d]_{\al^{\vee}}
\ar[r] & X_f(A/F_{\infty})_H \ar[d]_{h''} \ar[r]& 0 \\
0 \ar[r]^{} & X(A/F^{\cyc})(p) \ar[r]^{} & X(A/F^{\cyc}) \ar[r] & \
X_f(A/F^{\cyc}) \ar[r] & 0 } \]
with exact rows. From this, we have a long exact
sequence
\[ \ker h' \lra \ker (\al^{\vee})
\stackrel{f}{\lra} \ker h''
\lra \coker h' \lra \coker (\al^{\vee}) \lra \ker h''\lra 0. \]
Since $X(A/F_{\infty})$ satisfies $\M_H(G)$-property, it follows from Nakayama's lemma that $X_f(A/F_{\infty})_H$ is finitely generated over $\Op$. But $X_f(A/F^{\cyc})$ is also finitely generated over $\Op$ by \cite[Proposition 2.5]{CS12}. Hence $\ker h''$ and $\coker h''$ are finitely generated over $\Op$, and we have
\[ \rank_{\Op} (\ker h'') - \rank_{\Op}(\coker h'')= \rank_{\Op}\big(X_f(A/F_{\infty})_H\big) -\rank_{\Op}(X_f(A/F^{\cyc}))\]
On the other hand, as already seen above, $\ker (\al^{\vee})$ and $\coker (\al^{\vee})$ are finitely generated over $\Op$. Hence so are $\ker f$ and $\coker h'$. But since these latter groups are $\pi$-primary, they must be finite. Thus, we have
\[ \rank_{\Op} (\ker (\al^{\vee})) - \rank_{\Op}(\coker (\al^{\vee}))= \rank_{\Op}\big(X_f(A/F_{\infty})_H\big) -\rank_{\Op}(X_f(A/F^{\cyc})).\]
Combining this with the above calculations and applying Proposition \ref{Howson} for $X_f(A/F_{\infty})$, we obtain the required formula.
\epf
\section{Kida's formula} \label{Kida section}
We are in position to prove the main theorem of the paper.
\bt \label{Kida formula theorem}
Let $F_{\infty}$ be a strongly admissible pro-$p$ Lie extension of $F$ and $L$ a finite extension of $F$ contained in $F_{\infty}$ with $F^{\cyc}\cap L= F$. Let $\big(A, \{A_v\}_{v|p}, \{A^+_v\}_{v|\R} \big)$ be a datum satisfying $\textbf{(C1)-(C4)}$ and $\textbf{(Fin)}$. Assume that $X(A/F_{\infty})$ satisfies the $\M_H(G)$-property. Then we have
\[ \mu_{\Op\ps{\Ga_L}}\big(X(A/L^{\cyc})\big) = [L:F]\mu_{\Op\ps{\Ga}}\big(X(A/F^{\cyc})\big)\] and
\[ \rank_{\Op}(X_f(A/L^{\cyc})) = [L:F]\rank_{\Op}(X_f(A/F^{\cyc})) + \corank_{\Op}(A(L^{\cyc})) -[L:F] \corank_{\Op}(A(F^{\cyc})) \] \[
+ \sum_{w\in R(L^{\cyc}/F^{\cyc})}\left([L:F]\corank_{\Op}\big(D_w(F^{\cyc}_w)\big) -\sum_{u|w}\corank_{\Op}\big(D_w(L^{\cyc}_{u})\big) \right).\]
Here $D_w$ denotes $A/A_w$ or $A$ accordingly as $w$ divides $p$ or not, and $R(L^{\cyc}/F^{\cyc})$ is set of primes of $F^{\cyc}$ consisting of all the primes above $p$ and all the ramified primes in $L^{\cyc}/F^{\cyc}$.
\et
\bpf
For a $\Op\ps{G}$-module $M$, one has the observation that $\mu_{\Op\ps{G}}(M) = [G:G']\mu_{\Op\ps{G'}}(M)$. Combining this with the first equality in Proposition \ref{rank formula}, we obtain the first equality of the theorem. On the other hand, one also has
$\rank_{\Op\ps{H}}(M) = [H:H_L]\rank_{\Op\ps{H_L}}(M)$. Applying this to $X(A/F_{\infty})$ and combining with Proposition \ref{rank formula}, we have
\[ \rank_{\Op}(X_f(A/L^{\cyc})) - \corank_{\Op}(A(L^{\cyc})) + \sum_{\substack{u\in S(L^{\cyc}), \\ \dim H_{L,u}\geq 1}}\corank_{\Op}\big(D_u(L^{\cyc}_u)\big) = \]
\[ [L:F]\left(\rank_{\Op}(X_f(A/F^{\cyc})) - \corank_{\Op}(A(F^{\cyc})) + \sum_{\substack{w\in S(F^{\cyc}), \\ \dim H_w\geq 1}}\corank_{\Op}\big(D_w(F^{\cyc}_w)\big)\right).\]
Now since $L^{\cyc}/F^{\cyc}$ is a finite extension, it follows that for each $w\in S(F^{\cyc})$ and $u\in S(L^{\cyc})$ which lies above $w$, we have $\dim H_{L,u} \geq 1$ if and only $\dim H_{w} \geq 1$. Hence the summation $\displaystyle\sum_{\substack{u\in S(L^{\cyc}) \\ \dim H_{L,u}\geq 1}}$ can be rewritten as $\displaystyle\sum_{\substack{w\in S(F^{\cyc})\\ \dim H_{w}\geq 1}}\sum_{u|w}$. Now, upon rearranging the terms, we obtain
\[ \rank_{\Op}(X(A/L^{\cyc})) = [L:F]\rank_{\Op}(X(A/F^{\cyc})) + \corank_{\Op}(A(L^{\cyc})) -[L:F] \corank_{\Op}(A(F^{\cyc})) \] \[
+ \sum_{w\in S(F^{\cyc})}\left([L:F]\corank_{\Op}\big(D_w(F^{\cyc}_w)\big) -\sum_{u|w}\corank_{\Op}\big(D_w(L^{\cyc}_{u})\big) \right).\]
Hence the required equality of this theorem will follow once we show that the summation $\displaystyle\sum_{w\in S(F^{\cyc})}$ can be rewritten as $\displaystyle\sum_{w\in R(L^{\cyc}/F^{\cyc})}$.
Clearly $R(L^{\cyc}/F^{\cyc})\subseteq S(F^{\cyc})$. Thus, it suffices to show that the term in the summand is zero when $w\in S(F^{\cyc})\setminus R(L^{\cyc}/F^{\cyc})$. But since such a prime does not divide $p$ and is not ramified in $L^{\cyc}/F^{\cyc}$, it must therefore split completely in $L^{\cyc}/F^{\cyc}$. Hence for each such prime $w$, we have $D_w(L^{\cyc}_u) = D_w(F^{\cyc}_w)$ for every $u|w$, where there are $[L:F]$ of these $u$'s. Consequently, the term in the sum for this prime $w$ is zero. Thus, we have proven our result.
\epf
For the remainder of this section, we discuss how the above theorem can be used to recover Kida's formula for the $\mu=0$ case. Before doing so, we record the following preliminary observation.
\bl \label{burn-ven}
Let $F$ be a number field which contains a primitive $p$-th root of unity and $L$ a finite $p$-extension of $F$. Then there exists a strongly admissible pro-$p$ Lie extension $F_{\infty}$ of $F$ containing $L$.
\el
\bpf
This can be proven similarly to that in \cite[Lemma 6.1]{BV}.
\epf
We now apply Theorem \ref{Kida formula theorem} to recover Kida's formula for the $\mu=0$ case.
\bt \label{Kida formula theorem mu=0}
Let $F$ be a number field which contains a $p$-th root of unity and $L$ a finite $p$-extension of $F$. Let $\big(A, \{A_v\}_{v|p}, \{A^+_v\}_{v|\R} \big)$ be a datum satisfying $\textbf{(C1)-(C4)}$ and $\textbf{(Fin)}$. Suppose that $X(A/F^{\cyc})$ is finitely generated over $\Op$. Then $X(A/L^{\cyc})$ is finitely generated over $\Op$ and
\[ \rank_{\Op}(X(A/L^{\cyc})) = [L:F]\rank_{\Op}(X(A/F^{\cyc})) + \corank_{\Op}(A(L^{\cyc})) -[L:F] \corank_{\Op}(A(F^{\cyc})) \] \[
+ \sum_{w\in R(L^{\cyc}/F^{\cyc})}\left([L:F]\corank_{\Op}\big(D_w(F^{\cyc}_w)\big) -\sum_{u|w}\corank_{\Op}\big(D_u(L^{\cyc}_{u})\big) \right),\]
where $R(L^{\cyc}/F^{\cyc})$ is set of primes of $F^{\cyc}$ consisting of primes above $p$ and the primes that are ramified in $L^{\cyc}/F^{\cyc}$.
\et
\bpf
Let $F_{\infty}$ be a strongly admissible pro-$p$ Lie extension $F_{\infty}$ of $F$ containing $L$ which is obtained via Lemma \ref{burn-ven}. By a similar argument to that in \cite[Corollary 2.5]{PW06} or \cite[Corollary 3.4]{HM}, $X(A/L^{\cyc})$ is finitely generated over $\Op$. Also, it follows from \cite[Theorem 2.1]{CS12} that $X(A/F_{\infty})$ is finitely generated over $\Op\ps{H}$, where $H=\Gal(F_{\infty}/F^{\cyc})$. In particular, $X(A/F_{\infty})$ satisfies the $\M_H(G)$-property. Thus, Theorem \ref{Kida formula theorem} applies to yield the required conclusion.
\epf
\section{Review of elliptic curves} \label{elliptic review}
In this section, we revisit Kida's formula for the elliptic curve of Hachimori-Matsuno \cite{HM} and show that our theorem specializes to recover their theorem. For simplicity, we shall assume that $p\geq 5$. Let $E$ be an elliptic curve defined over a number field $F$ such that for every prime $v$ of $F$ above $p$, $E$ has either good ordinary reduction or split multiplicative reduction at $v$. We shall also assume that $F$ contains all the $p$th roots of unity. Recall that by \cite[p.\ 150-151]{CG}, for each prime $v$ of $F$ above $p$, we have a short exact sequence
\[ 0\lra C_v\lra E(p) \lra D_v\lra 0 \]
of discrete $\Gal(\bar{F}_v/F_v)$-modules which is characterized by the facts that $C_v$ is divisible and $D_v$ is the maximal quotient of $E(p)$ by a divisible subgroup such that the inertia group acts on $D_v$ via a finite quotient. In fact, by our assumptions (also see \cite[pp. 69-70]{GR99}), $D_v$ has the following description
\[ D_v =\begin{cases} \widetilde{E}_v(p),& \mbox{if $E$ has good ordinary reduction at $v$}, \\
\Qp/\Zp, & \mbox{if $E$ has split multiplicative reduction at $v$}.\end{cases}\]
Here $\widetilde{E}_v$ is the reduced curve $E$ at $v$.
The following finiteness result of $\widetilde{E}_v(F^{\cyc}_w)(p)$ for a good ordinary prime $v$ is well-known but for a lack of reference, we include a proof here.
\bl \label{finiteness}
Let $E$ be an elliptic curve over $F$ with good ordinary reduction at the prime $v$ which lies above $p$. Then for every prime $w$ of $F^{\cyc}$ above $v$, we have that $\widetilde{E}_v(F^{\cyc}_w)(p)$ is finite.
\el
\bpf
The long cohomology exact sequence of
\[ 0\lra C_v\lra E(p) \lra \widetilde{E}_v(p)\lra 0 \]
gives rise to an exact sequence
\[ E(F^{\cyc}_w)(p)\lra \widetilde{E}_v(F^{\cyc}_w)(p)\lra H^1(F^{\cyc}_w, C_v)\lra H^1(F^{\cyc}_w, E(p)) \lra H^1(F^{\cyc}_w, \widetilde{E}_v(p)) \lra 0,\]
where the final zero follows from that fact that $H^2(F^{\cyc}_w,C_v)=0$ (cf.\ \cite[Theorem 7.1.8(i)]{NSW}). Since our elliptic curve has good ordinary reduction at the prime $v$, Imai's theorem \cite{Imai} asserts that $E(F^{\cyc}_w)(p)$ is finite. On the other hand, a local Euler characteristics argument (cf.\ \cite[\S 3]{G89}) shows that $\corank_{\Zp}(H^1(F^{\cyc}_w, E(p))) =
\corank_{\Zp}(H^1(F^{\cyc}_w, C_v)) + \corank_{\Zp}(H^1(F^{\cyc}_w, \widetilde{E}_v(p))$. Putting this information into the exact sequence, we see that the $\widetilde{E}_v(F^{\cyc}_w)(p)$ has trivial $\Zp$-corank.
\epf
Now for the local terms, we have
\[\corank_{\Zp}\big(D_w(F^{\cyc}_w)\big) = \begin{cases} 1,& \mbox{if $E$ has split multiplicative reduction at $w$}, \\
2, & \mbox{if $w\nmid p$ and $E$ has good reduction at $w$ with $E(F^{\cyc}_w)[p]\neq 0$},
\\
0, & \mbox{otherwise,}\end{cases}
\]
where we have made use of \cite[Proposition 5.1]{HM} for the primes not dividing $p$.
Let $L$ be a finite $p$-extension of $F$. By Lemma \ref{burn-ven}, one can find a strongly admissible pro-$p$ $p$-adic Lie extension $F_{\infty}$ of $F$ which contains $L$.
Denote by $P_0(F^{\cyc})$ the set of primes of $S(F^{\cyc})$, whose decomposition group of $H$ at $w$ has dimension $\geq 1$. Let $M(F^{\cyc})$ (resp., $P_1(F^{\cyc})$) be the set of split multiplicative primes of $E$ in $P_0(F^{\cyc})$ which lie above $p$ (resp., do not lie above $p$). Let $P_2(F^{\cyc})$ be the primes $w$ of $P_0(F^{\cyc})$ at which $w\nmid p$ and $E$ has good reduction at $w$ with $E(F^{\cyc}_w)[p]\neq 0$. Assuming that $X(E/F_{\infty})$ satisfies the $\M_H(G)$-conjecture, Theorem \ref{Kida formula theorem} then reads as
\[ \rank_{\Zp}(X_f(E/L^{\cyc})) = [L:F]\rank_{\Zp}(X_f(E/F^{\cyc}))
+ \sum_{w\in M(F^{\cyc})\cup P_1(F^{\cyc})}\left([L:F] -\sum_{u|w}1\right) \]
\[+ 2\sum_{w\in P_2(F^{\cyc})}\left([L:F] -\sum_{u|w}1 \right), \]
where the $u$'s denote the primes in $S(L^{cyc})$, and noting that $E(F^{\cyc})(p)$ and $E(L^{\cyc})(p)$ are finite by a result of Ribet \cite{Ri}. One can see that the above agrees with that in \cite[Theorems 3.1 and 8.1]{HM} and \cite[Lemma 3.6]{HV}.
\br
One can also check that our Theorem \ref{Kida formula theorem} (or Theorem \ref{Kida formula theorem mu=0}) recovers \cite[Theorem 4.2]{CDLSS} and \cite[Theorem 2.8]{PW06}.
\er
We end by discussing an example taken from \cite{DD15}. Let $E$ be the elliptic curve defined by
\[ y^2 = x^3 -24z^7\sqrt{z+3} x^2 +zx,\]
where $z= \frac{\sqrt{5}-1}{2}$. This elliptic curve is defined over $\Q(\sqrt{z+3})$ and has good ordinary reduction at the prime above $5$. Let $F$ be the field $\Q(\zeta_{20})$ and $F_{\infty}$ the $\Z_5^5$-extension of $F$. It is shown in \cite{DD15} that the $\mu$-invariant of $X(E/F^{\cyc})$ is positive. Thus, if we assume that $\M_H(G)$-conjecture holds for $X(E/F_{\infty})$, then it follows from the above discussion that for every finite extension $L$ of $F$ contained in $F_{\infty}$, we have
\[ \rank_{\Zp}(X_f(E/L^{\cyc})) = [L:F]\rank_{\Zp}(X_f(E/F^{\cyc})).\]
Unfortunately, we do not have a way of establishing $\M_H(G)$-conjecture for $X(E/F_{\infty})$ as yet (but see \cite[Section 3]{CS12} and \cite[pp 1985]{DD15} for some partial results in this direction). It could be of interest to at least verify the above rank equality numerically (for certain $L$). We however do not know how to approach this problem at this point of writing.
\section{Remarks on the classical case}
We give a sketch how the discussion of Section \ref{Kida section} specializes to the classical case. From now on, $F$ will denote a totally real number field. Write $\Delta= \Gal(F(\mu_p)/F)$. Let
\[ \kappa: \Gal(F(\mu_{p^\infty})/F)\lra \mathrm{Aut}(\mu_{p^\infty})\cong \Zp^{\times} \]
be the cyclotomic character. Writing $d=|\Delta|$, we define
\[ e_{i} = \frac{1}{d}\sum_{\sigma\in \Delta} \kappa^{-i}(\sigma)\sigma^i \in \Zp[\Delta].\]
Let $i$ be an even integer such that $0\leq i\leq d$. Set $A=\Qp/\Zp(\kappa^{-i})$ and $A_v=\Qp/\Zp(\kappa^{-i})$ for all $v|p$, where
$\Qp/\Zp(\kappa^{-i})$ is $\Qp/\Zp$ as $\Zp$-module with a Galois action by $\sigma\cdot x = \kappa^{-1}(\sigma) x$. Then one can check that
\[S^{str}(A/F(\mu_{p^{\infty}})) = H^1\big(G_S(F(\mu_{p^{\infty}})),A\big) \cong \big(e_iG_S(F(\mu_{p^{\infty}}))^{ab}(p)\big)^{\vee}. \]
Here $G_S(F(\mu_{p^{\infty}}))^{ab}$ is the abelianization of the group $G_S(F(\mu_{p^{\infty}}))$. Note that $e_iG_S(F(\mu_{p^{\infty}}))^{ab}(p)$ is a torsion $\Zp\ps{\Ga}$-module by \cite[Proposition 11.4.5]{NSW}, where $\Ga=\Gal(F^\cyc/F)$.
We may proceed as in the previous section to obtain a Kida formula for $S^{str}(A/F(\mu_{p^{\infty}}))$ for every even $i$ under the assumption of the $\M_H(G)$-conjecture. Summing over these $i$'s, we obtain \cite[Corollary 11.4.11]{NSW}. Upon combining this with \cite[Corollary 11.4.4]{NSW}, we obtain the classical Kida's formula for the minus class groups \cite{Iw, Ki}. We should mention that Hachimori-Sharifi has proven the classical Kida's formula under a slight weakening of the usual $\mu=0$ assumption (cf. \cite[Theorem 2.1]{HS}). However, it would seem that their approach cannot be extended to prove the general case (see \cite[Remark 2.2.2]{Gr11}). Of course, we should mention that in this classical situation, Iwasawa has conjectured that $\mu=0$ always holds, although this has only been verified for abelian number fields (\cite{FW}).
\footnotesize
| 3428b7f5099b9eeae91c879034f268c74e8e2a44 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{ Introduction }
In the present paper,
we shall begin with introducing one-dimensional weighted Hardy's inequalities with sharp remainders under one-sided boundary condition.
As weights we shall deal with the so-called non-doubling weights in addition to usual doubling ones.
Then we shall establish $N$-dimensional weighted Hardy's inequalities with non-doubling
weights of the distance $\delta(x):= { \rm dist}(x,\partial\Omega)$ to the boundary $\partial\Omega$, where
$\Omega$ is a $C^2$ class bounded domain of $\mathbf R^N\, ( N\ge 2)$.
A positive \textcolor{black}{ continuous} function $w(t)$ on $(0,\infty)$ is said to be a doubling weight if there exists a positive number $C$ such that we have
\begin{equation} C^{-1} w(t)\le w(2t)\le C w(t) \quad (0<t<\infty),\label{doubling} \end{equation}
where $C$ is independent of each $t \in (0,\infty)$.
When $w(t)$ does not possess this property, $w(t)$ is said to be a non-doubling weight in the present paper.
In one-dimensional case we typically treat a weight function $w(t)$ that may vanish or blow up in infinite order such as $e^{-1/t}$ or $e^{1/t}$ at $t=0$. In such cases the limit of ratio $w(t)/w(2t)$ as $t\to +0$ may become $0$ or $+\infty$, and hence they are regarded as non-doubling weights according to our notion.
\begin{df} Let $1<p<\infty$ and
set $p'= p/(p-1)$. $\Lambda_p$ denotes the one-dimensional Hardy best constant defined by
$$\Lambda_p=\frac{1}{(p')^p}=\left(1- \frac1p\right)^{p}.$$
\end{df}
Let $\eta>0$. By $ C_c^1((0,\eta])$ we denote the set of all $C^1$ functions with compact supports in $(0,\eta]$.
Then
one-dimensional Hardy's inequalities with one-sided boundary condition in this paper are typically represented by the followings.
\begin{prop}\label{Prop} Assume that $1<p<\infty$, $\mu>0$ and $\eta>0$. \par\noindent
1.
For every $u\in C_c^1((0,\eta])$ we have
\begin{equation} \int_0^\eta |u'(t)|^p e^{-(p-1)/t}\,dt + \frac{(\Lambda_{p})^{1/p'} }{\mu^{p-1} }|u(\eta)|^p\ge
\Lambda_{p} \int_0^\eta \frac{|u(t)|^p {e^{-(p-1)/t}}\,dt}{\left(e^{-1/t}\left( \int_t^\eta e^{1/s}\,ds +\mu \right)\right)^p}. \label{DD}
\end{equation}
\par \noindent 2.
For every $u\in C_c^1((0,\eta])$ we have
\begin{equation} \int_0^\eta |u'(t)|^p e^{(p-1)/t}\,dt \ge
\Lambda_{p} \int_0^\eta \frac{|u(t)|^p {e^{(p-1)/t}}\,dt}{\left( e^{1/t} \int_0^t e^{-1/s} \,ds \right)^p}+ \frac{(\Lambda_{p})^{1/p'}}{\left(\int_0^\eta e^{-1/s}\,ds\right)^{p-1} }|u(\eta)|^p
. \label{GG}
\end{equation}
\end{prop}
Proposition \ref{Prop} will be established as Corollary \ref{cor2.2} to Theorem \ref{T3.1}.
When $u(\eta)=0$, these inequalities are a variant of classical weighted Hardy's inequalities (see \cite{ KO,ahl,Ma}).
It is interesting that both coefficients $\Lambda_{p}$ and $(\Lambda_{p})^{1/p'}$ appearing in (\ref{DD}) and (\ref{GG}) are best, even though the inequalities contain two parameters $\eta$ and $\mu$.
To see the sharpness of (\ref{DD}) and (\ref{GG}), by the density argument it suffices to employ $u_\varepsilon(t)= \left( \mu+ \int_t^\eta e^{1/s}\,ds\right)^{1/p'-\varepsilon}$ and
$u_\varepsilon(t)= \left( \int_0^t e^{-1/s}\,ds\right)^{1/p'+\varepsilon}$ for
test functions ($\varepsilon \to +0$) respectively. (For the detail see Subsection 4.2 (Part 1).)
\par
Our first purpose in this paper is not only to establish a general version of Proposition \ref{Prop} but also improve it by adding sharp remainder terms.
By $W(\mathbf R_+)$ we denote a class of functions
$$ \{ w\in C^1({\bf R}_+): w>0, \lim_{t\to+0}w(t)=a\, \text{ for some }\,a\in [0,\infty] \}$$
with $\mathbf R_+=(0,\infty)$.
As weights we adopt functions $W_p(t)=w(t)^{p-1}$ with $w(t)\in P(\mathbf R_+) \cup Q(\mathbf R_+)$, where
\begin{equation} \begin{cases}&P(\mathbf R_+)= \{ w(t)\in W(\mathbf R_+) : \, w(t)^{-1} \notin L^1((0,\eta)) \, \text{ for some} \, \eta >0\},\\
&Q(\mathbf R_+) =\{ w(t)\in W(\mathbf R_+) : \, w(t)^{-1}\in L^1((0,\eta)) \, \text{ for any } \, \eta >0 \}.
\end{cases}
\end{equation}
Clearly $W(\mathbf R_+)=P(\mathbf R_+) \cup Q(\mathbf R_+)$,
$e^{-1/t} \in P(\mathbf R_+)$ and $e^{1/t} \in Q(\mathbf R_+)$ for $t>0$ ( For the precise definitions see Section 2 ).
By virtue of Proposition \ref{Prop}, it is clearly seen that
our results on this matter essentially depend on whether $w$ belongs to
$ P(\mathbf R_+) $ or $ Q(\mathbf R_+)$.
In particular when $w(t)\in P(\mathbf R_+)$, it follows from Proposition \ref{ct1} that
\begin{equation}
\inf_{u\in V_\eta}
\int_0^\eta |u'(t)|^p W_p(t)\,dt =0, \label{C}
\end{equation}
where $V_\eta= \{ u\in C^1([0,\eta]) : u(0)=0, u(\eta)=1\}. $
Nevertheless we have sharp Hardy type inequalities (\ref{nct1}) and (\ref{nct2}) in Theorem \ref{T3.1}.
\par
As an important application, in Theorem \ref{CT3} we shall establish
$N$-dimensional Hardy's inequalities with
weights being functions of the distance $\delta (x)=\rm{ dist}(x,\partial\Omega)$ to the boundary $\partial\Omega$.
In this task it is crucial to establish sharp weighted Hardy's inequalities in the tubler neighborhood $\Omega_\eta$ of $\Omega$, which are reduced to the one dimensional inequalities in Theorem \ref{T3.1}.
To this end, $\Omega $ is assumed to be a bounded domain of $\mathbf R^N$ ( $N\ge 2$ ) whose boundary $\partial\Omega$ is a $C^2$ compact manifolds in the present paper.
We prepare more notations to describe our results.
For $ W_p(t) = w(t)^{p-1}$ with $ w(t)\in W(\mathbf R_+)$, we define a weight function $W_p(\delta(x))$ on $\Omega$ by $$W_p(\delta(x))=(W_p\circ\delta)(x).$$
By
$ L^p(\Omega; W_p(\delta))$ we denote the space of Lebesgue measurable functions with weight $ W_p(\delta(x))$,
for which
\begin{equation} \| u \|_{ L^p(\Omega; W_p(\delta))} = \bigg( \int_{\Omega}|u(x)|^p W_p(\delta(x))\,dx\bigg ) ^{1/p} < +\infty.\label{2.1} \end{equation}
By $ C_c^\infty(\Omega)$ we denote the set of all $C^\infty$ functions with compact supports in $\Omega$.
$W_{0}^{1,p}(\Omega,W_p(\delta))$ is given by the completion of $C_c^\infty(\Omega)$ with respect to the norm
defined by
\begin{equation}
\| u\|_{W^{1,p}_{0}(\Omega; W_p(\delta)) } =
\| |\nabla u| \|_{ L^p(\Omega; W_p(\delta))} + \| u\|_{ L^p(\Omega; W_p(\delta))}. \label{2.2}
\end{equation}
Then
$W^{1,p}_{0}(\Omega; W_p(\delta)) $ becomes a Banach space with the norm $\| \cdot\|_{W^{1,p}_{0}(\Omega; W_p(\delta)) }
$.
Under these preparation
we will establish $N$-dimensional weighted Hardy's inequality as
Theorem \ref{CT3}, which is the counter-part to Theorem \ref{T3.1}.
In particular for $w(t)\in Q(\mathbf R_+)$,
as in Corollary \ref{NCC1} we have a simple inequality which is a generalization of classical Hardy's inequality:
\begin{equation}\label{HI}
\int_\Omega |\nabla u|^p W_p(\delta(x)) \,dx\ge \gamma \int_\Omega \frac{|u|^pW_p(\delta(x))}{F_\eta(\delta(x))^p}\,dx, \quad \forall u\in W^{1,p}_{0}(\Omega; W_p(\delta)) ,
\end{equation}
where $\eta$ is a sufficiently small positive number, $\gamma$ is some positive constant and $F_\eta(\delta(x))=(F_\eta\circ \delta)(x)$ is a nonnegative function defined in Definition \ref{D2}.
If $w=1$, then $F_{\eta}(t)=t \, (t\le \eta)$; $\eta\,(t\ge \eta)$ and
(\ref{HI}) is a well-known Hardy's inequality \textcolor{black}{having $\left(\min(\delta(x),\eta)\right)^{-p}$ as the Hardy potential,} which is valid for a bounded domain $\Omega $ of $\mathbf R^N$ with Lipschitz boundary (cf. \cite{BM,D, MMP}).
Further
if $\Omega$ is convex, then $\gamma = \Lambda_{p}$ holds for arbitrary $1<p<\infty$ (cf. \cite{MS,YZ}). \par
It is worthy to remark that (\ref{HI}) is never valid in the case that $w(t)\in P(\mathbf R_+)$
by (\ref{C}) (see also Proposition \ref{ct1} and Proposition \ref{CT1}). Nevertheless, in this case
we shall establish weighted Hardy's inequalities with a switching function in Theorem \ref{CT3} and Corollary \ref{NCC1}, which correspond to Theorem \ref{T3.1} and its corollaries.
We remark that these Hardy's inequalities with a compact perturbation are closely relating to
the so-called weak Hardy property of $\Omega$. In fact, if $w(t) \in P(\mathbf R_+)$, then
a constant $\gamma^{-1}$ in (\ref{2.7}) concerns the weak Hardy constant, but in this case the strong Hardy constant is $+\infty$ ( see \cite{D} for the detail). In \cite{ahl}, we have improved Hardy's inequalities adopting $\delta(x)^{\alpha p}$ ( powers of the distance $\delta(x)$ to the boundary $\partial\Omega$ ) as weight functions instead of $W_p(\delta)$. In the present paper, some inequalities
of Hardy type in \cite{ah0} and \cite{ahl} will be employed with minor modifications, especially when $1<p<2$
(see also \cite{anm}).\par
We remark that our results will be applicable to variational problems with critical Hardy potentials in a coming paper \cite{ah4} (c.f. \cite{ah3})
and also applicable to the Caffarelli-Kohn-Nirenberg type inequalities with non-doubling weights in the coming paper \cite{ah5}.
\par
This paper is organized in the following way: In Section 2 we introduce a class of weight functions $W(\mathbf R_+)$ and two subclasses $P(\mathbf R_+)$ and $ Q(\mathbf R_+)$ together with
functions such as $F_\eta(t)$ and $G_\eta(t)$,
which are crucial in this paper. Further a notion of admissibilities for $P(\mathbf R_+)$ and $ Q(\mathbf R_+)$ is introduced.
In Section 3, the main results are described.
The results are divided into two cases (one-dimensional case and $N$-dimensional case), which are described in Subsection 3.1 and Subsection 3.2, respectively.
Theorem \ref{T3.1} and Theorem \ref{T3.2} are established in Section 4. Theorem \ref{CT3} together with Corollary \ref{NCC1} are proved in Section 5.
The proof of Theorem \ref{NCT2} is given in Section 6 and the proofs of Proposition \ref{ct1} and Proposition \ref{CT1} are given in Section 7.
In Appendix the proof of Lemma \ref{lemma4.4} is provided.
Some auxiliary inequalities are also given as Lemma \ref{l2}.
\par\medskip
\section{Preliminaries }
First we introduce a class of weight functions which is crucial in this paper.
\begin{df}\label{D0}
Let us set $\mathbf R_+= (0,\infty)$ and
\begin{equation}W(\mathbf R_+)=
\{ w(t)\in C^1({\bf R}_+): w(t)>0, \lim_{t\to+0}w(t)=a\, \text{ for some }\, a\in [0,\infty] \}.
\end{equation}
\end{df}
In the next we define two subclasses of this rather large space.
\begin{df}\label{D1} Let us set
\begin{equation} P(\mathbf R_+)= \{ w(t)\in W(\mathbf R_+) : \, w(t)^{-1} \notin L^1((0,\eta)) \, \text{ for some} \, \eta >0\}.\end{equation}
\begin{equation}
Q(\mathbf R_+) =\{ w(t)\in W(\mathbf R_+) : \, w(t)^{-1}\in L^1((0,\eta)) \, \text{ for any } \, \eta >0 \}.\end{equation}
\end{df}
\begin{rem} \begin{enumerate}
\item From Definition \ref{D0} and Definition \ref{D1} it follows that $W(\mathbf R_+)= P(\mathbf R_+)\cup Q(\mathbf R_+)$ and $P(\mathbf R_+)\cap Q(\mathbf R_+)=\phi.$
\item If $w^{-1} \notin L^1((0,\eta)) \, \text{ for some} \, \eta >0$, then $w^{-1} \notin L^1((0,\eta)) \, \text{ for any} \, \eta >0$. Similarly
if $w^{-1} \in L^1((0,\eta)) \, \text{ for some} \, \eta >0$, then $w^{-1} \in L^1((0,\eta)) \, \text{ for any} \, \eta >0$.
\item
If $w\in P(\mathbf R_+)$, then $\lim_{t\to +0}w(t)=0$. Hence
by setting $w(0)=0$, $w$ is uniquely extended to a continuous function on $[0,\infty)$.
On the other hand if $w\in Q(\mathbf R_+)$, then possibly $\lim_{t\to +0}w(t)=+\infty$.
\end{enumerate}
\end{rem}
Here we give some fundamental examples:
\begin{exam}
\begin{enumerate}\item $ e^{-1/t}\in P(\mathbf R_+)$ and $ e^{1/t}\in Q(\mathbf R_+)$. (See Corollary \ref{cor2.2}.)
\item For $p'=p/(p-1)$,
$t^{\alpha p'} \in P(\mathbf R_+)$ if $\alpha \ge 1/p'$ and
$t^{\alpha p'} \in Q(\mathbf R_+)$ if $\alpha < 1/p'$. (See Corollary \ref{cor2.3} with $W_p(t)=w(t)^{p-1}= t^{\alpha p}$.)
\item For $ \alpha\in \mathbf R$, $ t^\alpha e^{-1/t}\in P(\mathbf R_+)$ and $t^\alpha e^{1/t}\in Q(\mathbf R_+)$.
\end{enumerate}
\end{exam}
In the next we define functions such as $F_\eta(t)$ and $G_\eta(t)$,
which are crucial in considering variants of the Hardy potential like $F_\eta(\delta(x))^{-p}$ in (\ref{HI}).
\begin{df}\label{D2} Let $\mu>0$ and $\eta>0$. For
$w \in W(\mathbf R_+)$, we define the followings:
\begin{enumerate}
\item When $w\in P(\mathbf R_+)$,
\begin{align}
F_{\eta}(t; w,\mu) & = \begin{cases} &w(t)\left( \mu+ \int_t^\eta {w(s)}^{-1}\, {ds}\right), \qquad \text{ if } \, t\in (0,\eta], \\
&w(\eta) \mu, \qquad \qquad \qquad \qquad\quad \text{ if } \, t\ge \eta.
\end{cases}\\
G_\eta(t;w,\mu) &= \begin{cases} &\mu+ \int_t^\eta {F_\eta(s;w,\mu)}^{-1}\, {ds}, \qquad \text{ if } \, t\in (0,\eta],\\
&\mu, \qquad \qquad \qquad \qquad\qquad \quad \text{ if } \, t\ge \eta.
\end{cases} \label{PG}
\end{align}
\item When $w\in Q(\mathbf R_+)$,
\begin{align}
F_{\eta}(t; w) &= \begin{cases} &w(t)\int_0^t {w(s)}^{-1} \,{ds}, \qquad \qquad \text{ if } \, t\in (0,\eta], \\
&w(\eta) \int_0^\eta w(s)^{-1} \,ds, \qquad \quad \text{ if } \, t\ge \eta.
\end{cases}\\
G_\eta(t;w,\mu)&= \begin{cases} &\mu+ \int_t^\eta {F_\eta(s;w)}^{-1}\, {ds}, \qquad \text{ if } \, t\in (0,\eta], \\
&\mu, \qquad \qquad \qquad \qquad \qquad \text{ if } \, t\ge \eta.
\end{cases}
\end{align}
\item $F_{\eta}(t;w,\mu)$ and $F_{\eta}(t;w)$ are abbreviated as $F_{\eta}(t)$.
$ G_{\eta}(t; w,\mu)$ is abbreviated as $ G_{\eta}(t)$.
\item For $w\in W(\mathbf R_+)$,
we define
\begin{equation} W_p(t)= w(t)^{p-1}. \label{1.4}\end{equation}
\end{enumerate}
\end{df}
\begin{rem}\label{defofG} In the definition (\ref{PG}), one can replace $G_\eta(t; w,\mu)$ with the more general
$G_\eta(t; w,\mu,\mu')$ $= \mu'+ \int_t^\eta {F_\eta(s;w,\mu)}^{-1}\, {ds}, \, t\in (0,\eta]; {\, } \mu', \, t\ge \eta \,$
for $\mu'>0 $.
However, for simplicity this paper uses (\ref{PG}).
\end{rem}
\begin{exam} Let $w(t)=t^{\alpha p'} $ for $\alpha\in \bf R$, $1<p<\infty$ and $p'=p/(p-1)$.
\begin{enumerate}
\item When $\alpha>1/p'$, $F_\eta(t)=t/(\alpha p'-1) \,$ and
$G_\eta(t)=\mu +(\alpha p'-1)\log(\eta/t)$ provided that $ \mu= \eta^{1-\alpha p'}/(\alpha p'-1)$.
\item When $\alpha=1/p'$, $F_\eta(t)=t(\mu+ \log(\eta/t)) \,$ and
$G_\eta(t)=\mu -\log \mu +\log\left( \mu+ \log(\eta/t)\right)$.
\item When $\alpha<1/p'$, $F_\eta(t)=t/(1-\alpha p') \,$ and
$G_\eta(t)=\mu +(1-\alpha p')\log(\eta/t)$.
\end{enumerate}
\end{exam}
By using integration by parts we see the followings:
\begin{exam}
\begin{enumerate}\item If either $w(t)=e^{-1/t} \in P(\bf R_+)$ or $w(t)=e^{1/t}\in Q(\bf R_+)$, then $F_\eta(t)= O(t^2)$ as $t\to +0$.
\item Moreover,
if $w(t)= \exp({\pm t^{-\alpha}}) $ with $\alpha> 0$, then $F_\eta(t)= O(t^{\alpha+1})$ as $t\to +0$.
In fact, it holds that $\lim_{t\to +0} F_\eta(t)/t^{\alpha+1}=1/\alpha$.
\end{enumerate}
\end{exam}
In a similar way we define the following:
\begin{df}\label{fg} Let $p'= p/(p-1)$, $\mu>0$ and $\eta>0$. For
$w \in W(\mathbf R_+)$ and $t\in (0,\eta)$, we define the followings:
\begin{enumerate}
\item When $w\in P(\mathbf R_+)$,
\begin{align}
f_{\eta}(t; w,\mu) &= \begin{cases} & \mu+ \int_t^\eta {w(s)}^{-1}\, {ds}, \qquad \text{ if } \, t\in (0,\eta],\\
&\mu, \qquad \qquad \qquad \qquad \text{ if } \, t\ge \eta.
\end{cases}\\
g_\eta(t;w,\mu)&= (p' f_\eta(t;w,\mu))^{{1}/{p'}}, \qquad \qquad \text{ if } \, t>0.
\end{align}
\item When $w\in Q(\mathbf R_+)$,
\begin{align}
f_{\eta}(t; w) &= \begin{cases}&
\int_0^t {w(s)}^{-1} \,{ds}, \qquad \qquad \text{ if } \, t\in (0,\eta],\\
&\int_0^\eta w(s)^{-1}\,ds,\qquad \qquad \text{ if } \, t\ge \eta.
\end{cases}\\
g_\eta(t;w) &= (p' f_\eta(t;w))^{{1}/{p'}}, \qquad \qquad \quad \text{ if } \, t>0.
\end{align}
\item $f_{\eta}(t;w,\mu)$ and $f_{\eta}(t;w)$ are abbreviated as $f_{\eta}(t)$.
$g_{\eta}(t;w,\mu)$ and $g_{\eta}(t;w)$ are abbreviated as $ g_{\eta}(t)$.
\end{enumerate}
\end{df}
\begin{rem}\label{Remark2.2}\begin{enumerate}\item
We note that for $0<t<\eta$
\begin{equation}\label{2.13}\begin{cases}& \frac {d}{dt}\log f_{\eta}(t)= -{F_\eta(t) }^{-1}, \quad \text{ if } w\in P({\mathbf R}_+),\\
&\frac {d}{dt}\log f_{\eta}(t)= {F_\eta(t) }^{-1},\quad \mbox{ if } w\in Q({\mathbf R}_+),\\
&\frac {d}{dt}\log G_{\eta}(t)= -({F_\eta(t) G_{\eta}(t)})^{-1},\quad \mbox{ if } w\in W({\mathbf R}_+),\\
&\frac {d}{dt}G_{\eta}(t)^{-1}= ({F_\eta(t) G_{\eta}(t)^2})^{-1},\quad \mbox{ if } w\in W({\mathbf R}_+).
\end{cases}
\end{equation}
By Definition \ref{D1} and (\ref{2.13}), we see that $F_\eta(t)^{-1}\notin L^1((0,\eta))$, $( F_\eta (t)G_\eta(t))^{-1}\notin L^1((0,\eta))$ but $ (F_\eta(t) G_\eta(t)^2 )^{-1}\in L^1((0,\eta))$.
\item
If $w\in W(\bf R_+)$, then immediately we have $\liminf_{t\to +0}F_\eta(t)=0$ from 1.
\end{enumerate}
\end{rem}
\begin{exam}
If either
$w(t)=t^2e^{-1/t}\in P(\mathbf R_+)$ or $w(t)=t^2e^{1/t}\in Q(\mathbf R_+)$, then
$F_{\eta}(t)=O(t^2)$ and $G_{\eta}(t)=O(1/t)$ as $t\to 0$.
\end{exam}
Now we introduce two admissibilities for $P(\mathbf R_+)$ and $Q(\mathbf R_+)$.
\begin{df} \begin{enumerate}\item
A function $w\in P(\mathbf R_+)$ is said to be admissible if there exist positive numbers $\eta$ and $K$ such that we have
\begin{equation}\label{cond1}
\int_t^\eta w(s)^{-1} \,ds \le e^{K/\sqrt t} \qquad (0<t<\eta).
\end{equation}
\item
A function $w\in Q(\mathbf R_+)$ is said to be admissible if there exist positive numbers $\eta$ and $K$ such that we have
\begin{equation}\label{cond2}
\int_0^t w(s)^{-1} \,ds \ge e^{-K/\sqrt t} \qquad (0<t<\eta).
\end{equation}
\end{enumerate}
\end{df}
\begin{df} \label{admissibility}
By $P_A (\mathbf R_+)$ and $ Q_A(\mathbf R_+)$ we denote the set of all admissible functions in $P(\mathbf R_+)$ and $ Q(\mathbf R_+)$ respectively. We set \begin{equation}W_A(\mathbf R_+)= P_A (\mathbf R_+)\cup Q_A(\mathbf R_+).\end{equation}
\end{df}
Here we give typical examples:
\begin{exam} $e^{-1/t}\notin P_A(\mathbf R_+)$, $e^{1/t}\notin Q_A(\mathbf R_+)$ but $e^{-1/\sqrt{t}}\in P_A(\mathbf R_+)$, $e^{1/\sqrt{t}}\in Q_A(\mathbf R_+)$.\par\medskip
\par\noindent
{\bf Verifications: } \par
\noindent $e^{-1/t} \notin P_A(\bf R_+):$ For small $t>0$, we have $\int_t^\eta e^{1/s} \,ds \ge \int_t^{2t} e^{1/s} \,ds\ge te^{1/(2t)}$.
But this contradicts to (\ref{cond1}) for any $K>0$. \par\noindent
{ $e^{-1/\sqrt t} \in P_A(\bf R_+) :$}
Since $e^{1/\sqrt s}\le e^{1/\sqrt t}\, (t<s<\eta)$, we have $\int_t^\eta e^{1/\sqrt s}\,ds \le \eta e^{1/{\sqrt t}}\le e^{K/\sqrt t}$ for some $K>1$.
\par\noindent $e^{-1/t} \notin Q_A(\bf R_+):$
For $0<s\le t$, we have $\int_0^t e^{-1/s}\,ds\le t e^{-1/t}$. But this contradicts to (\ref{cond2}) for any $K>0$.
\par\noindent
{$e^{-1/\sqrt t} \in Q_A(\bf R_+):$} For $t/2<s<t$, we have
$\int_0^t e^{-1/\sqrt s}\,ds \ge \int_{t/2}^t e^{-1/\sqrt s}\,ds$ $\ge (t/2) e^{-\sqrt 2/ \sqrt s}$ $\ge e^{-K/\sqrt t}$ for some $ K>\sqrt 2$.
\end{exam}
\begin{prop}\label{prop2.1}
\begin{enumerate}\item
Assume that either $w\in P_A (\mathbf R_+)$ or $w\in Q_A (\mathbf R_+)$. Then
there exist positive numbers $\eta$ and $K$ such that we have
\begin{equation}
\sqrt t \,G_\eta (t) \le K\qquad t\in (0,\eta).\label{2.10}
\end{equation}
\item
Assume that $w\in W(\mathbf R_+)$ and $w$ satisfies (\ref{2.10}) for some positive numbers $\eta$ and $K$.
Then
$w$ is admissible in the sense of Definition \ref{admissibility}, that is, $w$ belongs to $W_A (\mathbf R_+)$.
\end{enumerate}
\end{prop}
\noindent {\bf Proof:}
By integrating
(\ref{2.13}), we have for $0<t<\eta$
\begin{equation}\begin{cases}&G_\eta (t)= \log \left( \int_t^\eta { w(s)}^{-1}\,ds +\mu\right)-\log \mu + \mu , \quad \text{ if } w\in P(\mathbf R_+),\\ \\
&G_\eta (t)= \log \left( \int_0^\eta { w(s)}^{-1}\,ds \right)- \log\left(\int_0^t { w(s)}^{-1}\,ds \right) + \mu ,\quad \mbox{ if } w\in Q(\mathbf R_+).
\end{cases}
\end{equation}
Hence the inequality (\ref{2.10}) with positive numbers $\eta$ and $K$ is equivalent to
\begin{equation}\begin{cases}& \int_t^\eta { w(s)}^{-1}\,ds\le \mu( e^{ {K}/{\sqrt t} -\mu}-1)\quad (0<t<\eta), \quad \text{ if } w\in P_A(\mathbf R_+),\\ \\
&\int_0^t {w(s)}^{-1}\,ds \ge e^\mu \int_0^\mu {w(s)}^{-1}\,ds \,e^{-{K}/{\sqrt t}}\quad (0<t<\eta),\quad \mbox{ if } w\in Q_A(\mathbf R_+).
\end{cases} \label{2.10'}
\end{equation}
Here we
note that for each $\mu>0$, $\eta>0$ and $w\in Q(\mathbf R_+)$ there exist some positive numbers $K_1$, $K_2$ and $K_3$ such that $K_1\le K_2\le K_3 $ and we have
\begin{equation} \begin{cases}&
e^{ {K_1}/{\sqrt t} } \le \mu( e^{ {K_2}/{\sqrt t} -\mu}-1) \le e^{ {K_3}/{\sqrt t}} \quad (0<t<\eta)\\ &\\
&
e^{- {K_3}/{\sqrt t}}\le e^\mu \int_0^\mu {w(s)}^{-1}\,ds \,\,e^{-{K_2}/{\sqrt t}} \le
e^{- {K_1}/{\sqrt t}} \quad (0<t<\eta).\end{cases}
\end{equation}
\noindent Therefore the assertion is now clear. \hfill$ \Box$
\section{Main results}
\subsection{Results in the one dimensional case}
We introduce function spaces to state the results in one dimensional case:\par\noindent
By
$ L^p((0,\eta]; W_p)$ we denote the space of Lebesgue measurable functions with weight $ W_p(t)$,
for which
\begin{equation} \| u \|_{ L^p((0,\eta]; W_p)} = \bigg( \int_0^\eta|u(t)|^p W_p(t)\,dt\bigg ) ^{1/p} < +\infty.\label{2.1'} \end{equation}
By $ C_c^\infty((0,\eta])$ we denote the set of all $C^\infty$ functions with compact supports in $(0,\eta]$.
$W_{0}^{1,p}((0,\eta]; W_p)$ is given by the completion of $C_c^\infty((0,\eta])$ with respect to the norm
defined by
\begin{equation}
\| u\|_{ W_{0}^{1,p}((0,\eta]; W_p)} =
\|u' \|_{ L^p((0,\eta]; W_p)} + \| u\|_{ L^p((0,\eta]; W_p)}. \label{2.2'}
\end{equation}
Then
$W_{0}^{1,p}((0,\eta]; W_p) $ becomes a Banach space with the norm $\| \cdot\|_{ W_{0}^{1,p}((0,\eta]; W_p)}$.
We also define a switching function:
\begin{df} $($ \rm{Switching function} $)$
For $w\in W(\mathbf R_+)= P(\mathbf R_+) \cup Q(\mathbf R_+)$ we set
\begin{equation} s(w)= \begin{cases} -1 \quad & \text{if } \quad w\in P(\mathbf R_+),\\
\,\, 1\quad & \text{if } \quad w\in Q(\mathbf R_+).\end{cases}
\end{equation}
\end{df}
We state one-dimensional Hardy's inequalities with one-sided boundary condition.
\par\medskip
\begin{thm}\mbox{$( w\in W(\mathbf R_+) )$ } \label{T3.1}\par \noindent
Assume that $1<p<\infty$, $\eta>0$, $\mu>0$ and $w\in W(\mathbf R_+)= P(\mathbf R_+) \cup Q(\mathbf R_+)$. Then we have the followings:
\begin{enumerate}\item For every $u\in W_{0}^{1,p}((0,\eta]; W_p)\cap C((0,\eta])$, we have
\begin{align} \int_0^\eta |u'(t)|^p W_p(t)\,dt \ge \Lambda_{p}
\int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p}+s(w)\frac{(\Lambda_p)^{1/p'}}{f_\eta(\eta)^{p-1}} |u(\eta)|^p
.\label{nct1}
\end{align}
Moreover the coefficients $\Lambda_p$ and $s(w)(\Lambda_p)^{1/p'}$ are optimal.
\item There exist positive numbers $C=C(w,p,\eta,\mu )$ and $L=L(w,p,\eta,\mu )$ such that for every
$u\in W_{0}^{1,p}((0,\eta]; W_p)\cap C((0,\eta])$, we have
\begin{equation}
\begin{split} \int_0^\eta |u'(t)|^p W_p(t)\,dt
\ge & \Lambda_{p}
\int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p} + C\int_0^\eta \frac{ |u(t)|^p W_p(t) \,dt}{ F_\eta(t)^p G_\eta(t)^2}
+s(w)L|u(\eta)|^p,\label{nct2}
\end{split}
\end{equation}
where $C$ and $L$ can be taken independent of each $u$.
\end{enumerate}
\end{thm}
The following is a direct consequence from this theorem.
We remark that $C_c^1((0,\eta])$ is densely contained in $W_{0}^{1,p}((0,\eta]; W_p) $.
\begin{cor}\label{cor2.2}
\begin{enumerate}\item Let $1<p<\infty$, $\eta>0$ and $\mu>0$.
Then, for every $u\in C_c^1((0,\eta])$ we have
\begin{equation} \int_0^\eta |u'(t)|^p e^{-(p-1)/t}\,dt + \frac{(\Lambda_{p})^{1/p'}}{\mu^{p-1} }|u(\eta)|^p\ge
\Lambda_{p} \int_0^\eta \frac{|u(t)|^p {e^{-(p-1)/t}}\,dt}{\left(e^{-1/t}\left( \int_t^\eta e^{1/s}\,ds +\mu \right)\right)^p}. \label{D}
\end{equation}
\item
Let $1<p<\infty$ and $\eta>0$.
Then, for every $u\in C_c^1((0,\eta])$ we have
\begin{equation} \int_0^\eta |u'(t)|^p e^{(p-1)/t}\,dt \ge
\Lambda_{p} \int_0^\eta \frac{|u(t)|^p {e^{(p-1)/t}}\,dt}{\left( e^{1/t} \int_0^t e^{-1/s} \,ds \right)^p}+ \frac{(\Lambda_{p})^{1/p'}}{\left(\int_0^\eta e^{-1/s}\,ds\right)^{p-1} }|u(\eta)|^p
. \label{G}
\end{equation}
\end{enumerate}
\end{cor}
\begin{df}For $1<p<+\infty$ and $\alpha \in \mathbf R$ we set
\begin{equation}\Lambda_{\alpha,p}=\left |\frac{1}{p'}-\alpha\right |^p = \left |\frac{p-1-\alpha p}{p}\right |^p.\label{2.5}
\end{equation}
\end{df}
\begin{cor}\label{cor2.3}
\begin{enumerate}\item If $\alpha>1/p'$, $1<p<\infty$ and $\eta>0$, then for every $u\in C_c^1((0,\eta])$ we have
\begin{equation} \int_0^\eta |u'(t)|^p t^{\alpha p}\,dt + \frac{(\Lambda_{\alpha,p})^{1/p'}}{\eta^{p-1-\alpha p}} |u(\eta)|^p\ge\Lambda_{\alpha,p} \int_0^\eta {|u(t)|^p}{t^{(\alpha -1)p}}\,dt. \label{F}
\end{equation}
\item
If $\alpha=1/p'$, $1<p<\infty$, $\eta>0$ and $R>e$, then for every $u\in C_c^1((0,\eta])$ we have
\begin{equation} \int_0^\eta |u'(t)|^p t^{p-1}\,dt
+ \frac{(\Lambda_{p})^{1/p'} } { (\log R)^{p-1} } |u(\eta)|^p \ge\Lambda_{p} \int_0^\eta \frac{|u(t)|^p}{t A_1(t/\eta)^p}\,dt,\label{E}
\end{equation}
where $ A_1(t)= \log (R/t)$ with $R>e$.
\item
If $ \alpha<1/p'$, $1<p<\infty$ and $\eta>0$, then for every $u\in C_c^1((0,\eta])$ we have
\begin{equation} \int_0^\eta |u'(t)|^p t^{\alpha p}\,dt \ge
\Lambda_{\alpha,p}\int_0^\eta {|u(t)|^p}{t^{(\alpha-1) p}}\,dt +
\frac{({\Lambda_{\alpha,p}})^{1/p'}}{\eta^{p-1-\alpha p}} |u(\eta)|^p.\label{B}
\end{equation}
\end{enumerate}
\end{cor}
\par\noindent{\bf Proof of Corollary \ref{cor2.3}}:
In Theorem \ref{T3.1} we set $W_p(t)=w(t)^{p-1}= t^{\alpha p}$. Then $ w(t)\in P(\mathbf R_+)$ for $\alpha \ge1/p'$,
and $ w(t)\in \textcolor{black}{ Q}(\mathbf R_+)$ for
$\alpha <1/p'$.
For $R>e$ we set
\begin{equation} \mu =
\begin{cases}& (\Lambda_{p}/ \Lambda_{\alpha,p})^{1/p} \eta^{(p-1-\alpha p)/(p-1)}, \quad \text{ if }\alpha>1/p',
\\ & \log R, \qquad\quad \qquad \quad\quad \quad \qquad\quad \text{ if }\alpha=1/p'.
\end{cases}
\end{equation}
When $\alpha \neq 1/p'$ we have
\begin{equation}
F_\eta(t)= (\Lambda_{p}/ \Lambda_{\alpha,p})^{1/p} \, t \quad\mbox{and}\quad
f_\eta(\eta)= \mu.
\end{equation}
When $\alpha =1/p'$ we have
\begin{equation}
F_\eta(t)=t \log (R\eta/t)\quad\mbox{and}\quad f_\eta(\mu)= \mu.
\end{equation}
Then the assertions (\ref{F}), (\ref{E}) and (\ref{B}) follow from (\ref{nct1}).\qed
\begin{rem}\begin{enumerate}\item
The sharpness of coefficients in theses corollaries \ref{cor2.2}, \ref{cor2.3} will be seen in Section 4 (4.2, Part 1).
\item If $u\in C^1_c((0,\infty))$, the inequalities (\ref{G}), (\ref{F}) and (\ref{B}) remain valid for $\eta=+\infty$.
\end{enumerate}
\end{rem}
\par\medskip
In order to establish Hardy's inequalities in a bounded domain of $\mathbf R^N$ as an application, we need a further refinement of
the previous results.
\begin{thm}\mbox{$( w\in W_A(\mathbf R_+)= P_A(\mathbf R_+) \cup Q_A(\mathbf R_+))$ } \label{T3.2}\par \noindent
Assume that $1<p<\infty$, $\mu>0$ and $ w\in W_A(\mathbf R_+)$. If $\eta>0$ is sufficiently small, then
there exist positive numbers $C_0=C_0(w,p,\eta,\mu)$, $C_1=C_1(w,p,\eta,\mu)$ and $L=L(w,p,\eta,\mu)$ such that for every $u\in W_{0}^{1,p}((0,\eta]; W_p)\cap C((0,\eta])$, we have
\begin{equation}
\begin{split} \int_0^\eta & \left( |u'(t)|^p -\frac{|u(t)|^p}{{F_\eta(t)}^p} \left(\Lambda_{p} + \frac{C_0 }{G_\eta(t)^2}\right)\right) W_p(t)\,dt
\label{c1} \\ &\ge C_1
\int_0^\eta \left( |u'(t)|^p + \frac{|u(t)|^p}{{F_\eta(t)}^p} \left(\Lambda_{p} + \frac{C_0 }{G_\eta(t)^2}\right) \right) W_p(t)\, t\,dt +s(w)L |u(\eta)|^p,
\end{split}
\end{equation}
where $C_0$, $C_1$ and $L$ can be taken independent of each $u$.
\end{thm}
Lastly we
state a fundamental result which will be useful in the subsequent.
\begin{prop}\label{ct1} $( w\in P(\mathbf R_+) )$\par\noindent
Assume that $ w\in P(\mathbf R_+)$ and $1<p<\infty$. Then we have
\begin{equation}\label{}
\inf_{u\in V_\eta}
\int_0^\eta |u'(t)|^p W_p(t)\,dt =0,
\end{equation}
where $V_\eta= \{ u\in C^1([0,\eta]) : u(0)=0, u(\eta)=1\}. $
\end{prop}
\begin{rem}
The proof will be given in Section 6.
\end{rem}
\medskip
\subsection{Results in a domain of $\mathbf R^N \,(N\ge 2)$ }
As an important application of one dimensional Hardy's inequalities in the previous section, we describe Hardy's inequalities in a bounded domain
$\Omega$ of $\mathbf R^N \,(N\ge 2)$. Let $\delta(x):= { \rm dist}(x,\partial\Omega)$.
For each small $\eta >0$, $\Omega _\eta $ denotes a tubular neighborhood of $\partial \Omega$ and
$\Sigma_\eta$ denotes the boundary of $ \Omega\setminus \Omega_\eta$, namely
\begin{equation}
\Omega_\eta = \{ x\in \Omega: \delta(x)<\eta \}\quad\mbox{and} \quad \Sigma_\eta = \{ x\in \Omega: \delta(x)=\eta \} .\label{NBD}
\end{equation}
For the sake of simplicity,
by $W_p(\delta)$, $F_\eta(\delta)$ and $G_\eta(\delta)$ we denote
$W_p(\delta(x))$, $F_\eta(\delta(x))$ and $G_\eta(\delta(x))$ respectively.\par
The proofs of Theorem \ref{CT3} and Corollary \ref{NCC1} will be given in Section 5. Theorem \ref{NCT2} will be proved in Section 6.
\begin{thm}\label{CT3} \mbox{\rm ( $ w\in W_A(\mathbf R_+)$ )} Assume that $\Omega$ is a bounded domain of class $C^2$ in
$\mathbf R^N$.
Assume that $1<p<\infty$ and $ w\in W_A(\mathbf R_+)$.
Assume that $\mu>0$ and $\eta $ is a sufficiently small positive number.
Then, there exist positive numbers $C=C(w,p,\eta,\mu)$ and $L'=L'(w,p,\eta,\mu)$ such that
for every $ u\in W_0^{1,p}(\Omega;W_p(\delta))\cap C(\Omega)$, we have
\begin{equation}\label{2.11}
\begin{split}
\int_{\Omega_\eta} \left( |\nabla u|^p -\Lambda_{p} \frac{|u|^p}{F_\eta(\delta)^p} \right)W_p(\delta)\,dx
&\ge C \int_{\Omega_\eta} \frac{|u|^p {W_p(\delta)}}{F_\eta(\delta)^p G_\eta(\delta)^2}\,dx \\
&+ s(w)L' \int_{\Sigma_\eta} |u|^p W_p(\delta)\,d\sigma_\eta,
\end{split}
\end{equation}
where $d\sigma_\eta$ denotes surface elements on $\Sigma_\eta$, and $C$, $L'$ are independent of each $u$.
\end{thm}
\begin{rem} We remark that the assumpsion $w\in W_A(\bf R_+)$ is needed
even if we do not have the first term involving $G_\eta(t)$ in the right-hand side.
(See also Corollary \ref{NCC1}).
\end{rem}
\begin{cor}\label{NCC1}
Assume that $\Omega$ is a bounded domain of class $C^2$ in $\mathbf R^N$.
Assume that $1<p<\infty$ and $ w\in W_A(\mathbf R_+)$. Assume that $\mu>0$ and $\eta $ is a sufficiently small positive number.
Then, there exist positive numbers $\gamma=\gamma(w,p,\eta,\mu)$ and $ L'=L'(w,p,\eta,\mu)$ such that for every
$ u\in W^{1,p}_{0}(\Omega;W_p(\delta))\cap C(\Omega)$, we have
\begin{equation}\label{2.7}
\int_{\Omega} \left( |\nabla u|^p -\gamma\frac{|u|^p}{F_\eta(\delta)^p} \right)W_p(\delta)\,dx
\ge s(w) L' \int_{\Sigma_\eta} |u|^pW_p(\delta)\,d\sigma_{\eta},
\end{equation}
where $d\sigma_\eta$ denotes surface elements on $\Sigma_\eta$, and $\gamma$, $L'$ are independent of each $u$.
\end{cor}
Moreover
we have the followings:
\begin{thm}\label{NCT2}\mbox{ \rm( $ w\in W_A(\mathbf R_+) $ )} Assume that $\Omega$ is a bounded domain of class $C^2$ in $\mathbf R^N$.
Assume that $1<p<\infty$, $\mu>0$ and $ w\in W_A(\mathbf R_+)$. Then, the followings are equivalent to each other.
\begin{enumerate}
\item There exist positive numbers $\gamma$, $\eta$ and $L'$ such that the inequality (\ref{2.7}) is valid for every $ u\in W_0^{1,p}(\Omega;W_p(\delta))\cap C(\Omega)$.
\item For a sufficiently small $\eta>0$,
there exist positive numbers $\kappa$, $C$ and $L'$ such that the inequality (\ref{2.11}) with $\Lambda_{p}$ replaced by $\kappa$ is valid for every $ u\in W_0^{1,p}(\Omega;W_p(\delta))\cap C(\Omega)$.
\end{enumerate}
\end{thm}
\par\medskip
\begin{prop}\label{CT1} \mbox{\rm ( $ w\in P(\mathbf R_+) $)} Assume that $\Omega$ is a bounded domain of class $C^2$ in $\mathbf R^N$.
Assume that $1<p<\infty$ and $ w\in P(\mathbf R_+)$.
Then, for an arbitrary $\eta \in (0, \sup_{x\in \Omega} \delta(x) )$ we have
\begin{equation}\label{2.8}
\inf
\left\{
\int_{\Omega} |\nabla u|^p W_p(\delta) \,dx : u\in C^1_c (\Omega), u=1 \mbox{ on } \{ \delta(x)=\eta\}\right \}=0.
\end{equation}
\end{prop}
\begin{rem}\begin{enumerate}\item
The proof will be given in Section 6 together with Proposition \ref{ct1}
\item
From this we see that constant functions belong to $W^{1,p}_{0}(\Omega;W_p(\delta))$, and hence
Hardy's inequality (\ref{HI}) never holds when $w\in P(\mathbf R_+)$.
\end{enumerate}
\end{rem}
\section{Proofs of Theorem \ref{T3.1} and Theorem \ref{T3.2} }
\subsection{ Lemmas }
First we prepare the following fundamental inequalities which are established in \cite{anm} as Lemma 2.1 for $X>-1$.
\begin{lem}\label{lem3.1}
\begin{enumerate}\item
For $p\ge 2$ we have
\begin{equation}
|1+X|^p-1-pX\ge c(p)|X|^q, \quad \mbox{ for any } q\in [2,p] \mbox{ and } X\in \mathbf {R}.
\label{3.0}\end{equation}
\item
For $1<p\le2$ and $M\ge1$, we have
\begin{equation} |1+X|^p-1-pX\ge c(p)\begin{cases} &M^{p-2}X^2,\qquad \,\, |X|\le M,\\
& |X|^p,\qquad\qquad |X|\ge M.\end{cases}\label{3.1}\end{equation}
\end{enumerate}
Here $c(p)$ is a positive number independent of each $X$, $M\ge1$ and $q\in [2,p]$.
\end{lem}
{\noindent\bf Proof.} By Taylar expansion we have (\ref{3.0}) with $q=2$.
For $p>1$, we note that
\begin{equation} \lim_{X\to 0} \frac{|1+X|^p-1-pX}{X^2}=\frac{p(p-1)}{2}, \quad \lim_{|X|\to \infty} \frac{|1+X|^p-1-pX}{|X|^p} =1. \label{Add}
\end{equation}
Therefore (\ref{3.0}) is valid for any $q\in [2,p]$ for a small $c(p)>0$.
If $X>-1$, then (\ref{3.1}) also follows from Taylar expansion and (\ref{Add}).
If we choose $c(p)$ sufficiently small, then it
remains valid for $X\le -1$.
\hfill$\Box$
\begin{rem}\label{density}
$C_c^1((0,\eta]) $ is densely contained in $W_{0}^{1,p}((0,\eta]; W_p)$.
If $u\in W_{0}^{1,p}((0,\eta]; W_p)$, then $|u|\in W_{0}^{1,p}((0,\eta]; W_p)$ and
$\| u\|_{ W_{0}^{1,p}((0,\eta]; W_p)}$ $ =$ $ \| |u| \|_{ W_{0}^{1,p}((0,\eta]; W_p)}$.
Therefore,
in the proofs of Theorem \ref{T3.1} and Theorem \ref{T3.2}, without the loss of generality
we may assume that $u\in C_c^1((0,\eta]) $ and
$ u\ge 0$ in $(0,\eta)$.
\end{rem}
For $u\in C_c^1((0,\eta])$ let us set
\begin{equation}
u(t)= g_\eta(t) v(t),
\end{equation}
where $g_\eta(t)$ is given by Definition \ref{fg}.
Clearly $v(t)\in C_c^1((0,\eta])$ and $u(t)'= g'_\eta(t) v(t)+g_\eta(t) v'(t)$. We define
\begin{equation} X(t)= \begin{cases}&\frac{g_\eta(t)}{g'_\eta(t)} \frac{ v'(t)}{v(t)}=p'F_\eta(t) \frac{ v'(t)}{v(t)}
\textcolor{black}{s(w)} \quad \text{if } v(t)\neq 0,\\
& 0\,\,\quad\qquad\qquad\qquad\qquad\qquad\text{ if } v(t)= 0. \end{cases} \end{equation}
Then we have
\begin{equation}
u'(t)= g'_\eta(t) v(t)\left( 1+ X(t)\right), \quad \text{ if } v(t)\neq 0. \end{equation}
Then it follows immediately from Lemma \ref{lem3.1} that we have the following.
\begin{lem}\label{prop4.1}
Assume that $\eta>0$, $\mu>0$, $X=X(t)$ and $ w\in W(\mathbf R_+)$.
Then, there exists a positive number $c(p)$ such that we have the followings:
\begin{enumerate}\item If $p\ge 2$, then for any $u\in C_c^1((0,\eta])$ and $q\in [2,p]$ we have
\begin{align} \int_0^\eta |u'(t)|^p W_p(t)\,dt
\ge \int_0^\eta |v(t)|^p |g_\eta'|^p \left( 1+ pX +c(p) |X|^q \right)W_p(t)\,dt. \label{1.lemma4.2}
\end{align}
\item If $1<p<2$, then for any $u\in C_c^1((0,\eta])$ and $M\ge 1$ we have
\begin{align} \int_0^\eta |u'(t)|^p W_p(t)\,dt
&\ge \int_0^\eta |v(t)|^p |g_\eta'|^p \left( 1+ pX \right)W_p\,dt \label{2.lemma4.2} \\
+c(p) \int_0^\eta |v(t)|^p & |g_\eta'|^p \left( M^{p-2} X^2 \chi_{ \{|X|\le M\}} + |X|^p \chi_{\{|X|>M\}}\right)W_p(t)\,dt,\notag
\end{align}
where $\chi_S(t)$ denotes a characteristic function of a set $S$.
\end{enumerate}
\end{lem}
Now we prepare the following lemma,
recalling a switching function;
\begin{equation} s(w)= \begin{cases} -1 \quad & \text{if } \quad w\in P(\mathbf R_+),\\
\,\,1\quad & \text{if } \quad w\in Q(\mathbf R_+).\end{cases}
\end{equation}
\begin{lem}\label{lemma4.2}Assume that $\eta>0$, $\mu>0$ and $ w\in W(\mathbf R_+)$. Assume that $u\in C_c^1((0,\eta])$, $X=X(t)$ and $ u(t)=g_\eta(t) v(t)$. Then we have:
\begin{enumerate}
\item
\begin{equation} |g'_\eta(t)|^{p-1} g_\eta(t) W_p(t)=1\label{4.13}
\end{equation}
\item
\begin{equation} p |v(t)|^p |g_\eta'(t)|^p X(t) W_p(t) =s(w) (|v(t)|^p)' \label{4.14}
\end{equation}
\item
\begin{equation} |v(t)|^p |g_\eta'(t)|^p W_p(t)= \Lambda_p \frac{ |u(t)|^pW_p(t)}{F_\eta(t)^p}\label{4.15}
\end{equation}
\item
\begin{equation}|v(t)|^p |g_\eta'(t)|^p |X(t)|^2 W_p(t)= \frac{4p'}{p^2} | (v(t)^{p/2})'|^2 F_\eta(t) \label{4.16}
\end{equation}
\item
\begin{equation} |v(t)|^p |g_\eta'(t)|^p |X(t)|^p W_p(t)= (p')^{p-1} |v'(t)|^p F_\eta(t)^{p-1} \label{4.17}
\end{equation}
\end{enumerate}
\end{lem}
\noindent{\bf Proof of Lemma \ref{lemma4.2}:}
\par\noindent
Proof of (\ref{4.13}): By Definition \ref{fg} (2) we see
\begin{equation} g'_\eta= (p')^{-1/p } (f_\eta)^{-1/p } w^{-1} s(w). \label{4.18} \end{equation} Then we have (\ref{4.13}).
\par \noindent Proof of (\ref{4.14}): Using (\ref{4.13}) together with
\begin{equation} p |v|^p |g_\eta'|^p X = s(w) (|v|^p)' |g'_\eta|^{p-1} g_\eta, \label{4.19} \end{equation}
we have (\ref{4.14}).
\par \noindent Proof of (\ref{4.15}): Noting that $p+p'=pp'$ and $(p')^{-p}= \Lambda_p$, we have
\begin{equation*} |v|^p|g_\eta'|^p=\frac{|u|^p}{ g^{p+p'}_\eta w^p}=\frac{|u|^p}{ g^{pp'}_\eta w^p}=\Lambda_p \frac{|u|^p}{ F_\eta^p}.
\label{4.20} \end{equation*}
\par \noindent Proof of (\ref{4.16}): Using (\ref{4.18}) and $2-(p-2)/(p-1)= p'$, we have
\begin{equation*}|v|^p |g_\eta'|^p |X|^2 W_p=|v|^{p-2} (v')^2 |g'_\eta|^{p-2}g_\eta^2 W_p=
\frac{4p'}{p^2} | (v^{p/2})'|^2 F_\eta.
\end{equation*}
\par \noindent Proof of (\ref{4.17}): Using (\ref{4.18}) and $ g_\eta^p= (p' f_\eta)^{p-1}$ we have
\begin{equation*} |v|^p |g_\eta'|^p |X|^p W_p= |v'|^p g_\eta^pW_p= (p')^{p-1} |v'|^p F_\eta^{p-1}.
\end{equation*}
\hfill $\Box$\par
By the definition of $\Lambda_p$ we have the following.
\begin{lem}\label{lem4.4}Assume that $\eta>0$ and $ w\in W(\mathbf R_+)$. Then we have
\begin{equation}
\int_0^\eta (|v(t)|^p)' \,dt=\frac{ (\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{p-1}} |u(\eta)|^p,
\label{4.21}
\end{equation}
where $v= u/g_\eta$ and $f_\eta(\eta)=\mu$, if $w\in P(\mathbf R_+)$; $\int_0^\eta w(s)^{-1}\,ds$, if $w\in Q(\mathbf R_+)$.
\end{lem}
By virtue of Lemma \ref{prop4.1}, Lemma \ref{lemma4.2} and Lemma \ref{lem4.4} , we have
the following:
\begin{lem}\label{prop4.2}
Assume that $\eta>0$, $\mu>0$, $X=X(t)$ and $ w\in W(\mathbf R_+)$.
Then, there exists a positive number $c(p)$ such that we have the followings:
\begin{enumerate}\item If $p\ge 2$, then for any $u\in C_c^1((0,\eta])$ we have
\begin{align} \int_0^\eta |u'(t)|^p W_p(t)\,dt &\ge \Lambda_p \int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p} +s(w)\frac{ (\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{p-1}} |u(\eta)|^p \label{4.222}\\&
+ d(p)\int_0^\eta |(|v(t)|^{p/2})'|^2 F_\eta(t) \,dt,\notag
\end{align}
where $d(p)= c(p) 4p'/p^2$.\par
Moreover the last term can be replaced by $c(p)(p')^{p-1}\int_0^\eta |v'(t)|^p F_\eta(t)^{p-1} \,dt$.
\item If $1<p<2$, then for any $u\in C_c^1((0,\eta])$ and $M\ge 1$ we have
\begin{align} & \int_0^\eta |u'|^p W_p(t)\,dt
\ge \Lambda_p \int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p} + s(w)\frac{ (\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{p-1}} |u(\eta)|^p \label{4.23}\\ &\notag
+c(p)(p')^{p-1}\int_0^\eta |v'(t)|^p F_\eta(t)^{p-1} \chi_{\{|X|>M\}}\,dt \\&+ M^{p-2}d(p)\int_0^\eta
|(|v(t)|^{p/2})'|^2 F_\eta(t) \chi_{ \{|X|\le M\}} \,dt, \notag
\end{align}
where $d(p)= c(p) 4p'/p^2$ and $\chi_S(t)$ denotes a characteristic function of a set $S$.
\end{enumerate}
\end{lem}
\par\noindent{\bf Proof of Lemma \ref{prop4.2}:}
\par\noindent{Proof of (\ref{4.222}): } This follows from (\ref{1.lemma4.2}) of
Lemma \ref{prop4.1}, (\ref{4.13}), (\ref{4.14}), (\ref{4.15}), (\ref{4.16}) of Lemma \ref{lemma4.2} and Lemma \ref{lem4.4}. By using (\ref{4.17}) instead of (\ref{4.16}), the alternative inequality follows.
\par\noindent{Proof of (\ref{4.23}): } This follows from (\ref{2.lemma4.2}) of
Lemma \ref{prop4.1}, (\ref{4.13}), (\ref{4.14}), (\ref{4.15}), (\ref{4.16}), (\ref{4.17}) of Lemma \ref{lemma4.2} and Lemma \ref{lem4.4}. \qed
\begin{rem}\label{rem4.2}
From
(\ref{4.222}) with the last term
replaced by $c(p)(p')^{p-1}\int_0^\eta |v'|^p F_\eta^{p-1} \,dt$, we have the following particular inequality for $p\ge 2$:
\begin{align} \int_0^\eta & |u'|^p W_p\,dt
\ge \Lambda_p \int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p} +s(w)\frac{ (\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{p-1}} |u(\eta)|^p\\ &
+\frac{d(p)}{2}\int_0^\eta |(|v(t)|^{p/2})'|^2 F_\eta(t) \,dt +\frac{c(p)(p')^{p-1}}{2}\int_0^\eta |v'(t)|^p F_\eta(t)^{p-1} \,dt.\notag
\end{align}
\end{rem}
\subsection{Proof of Theorem \ref{T3.1}}
\par\noindent{\bf Part 1: Proof of the inequality (\ref{nct1}) and its sharpness }\par
Assume that $w\in W(\mathbf R_+)= P(\mathbf R_+)\cup Q(\mathbf R_+)$.
It follows from Lemma \ref{prop4.2},
we clearly have the inequality (\ref{nct1}) in Theorem \ref{T3.1}. Hence we proceed to the proof of optimality of the coefficients $\Lambda_p$ and $(\Lambda_p)^{1/p'}$ in the right-hand side of (\ref{nct1}). By the density argument,
we can adopt as test functions
\begin{align} u_\varepsilon(t) = f_\eta(t)^{1/p'+s(w)\varepsilon } \qquad (0<\varepsilon).
\end{align}
We note that
\begin{equation} u'_\varepsilon(t)= \left( 1/p' +s(w)\varepsilon\right) f_\eta(t)^{s(w)\varepsilon -1/p}
{s(w)}{w}^{-1}.
\end{equation}
Then the left-hand side becomes
\begin{align*}
\int_0^\eta|u_\varepsilon'(t)|^pW_p(t)\,dt&= \left( 1/p' +s(w)\varepsilon\right)^p\int_0^\eta
f_\eta(t)^{s(w)\varepsilon p-1}{w(t)}^{-1}\,dt \notag\\
&= \left( 1/p' +s(w)\varepsilon\right)^p {f_\eta(\eta)^{s(w) \varepsilon p}}({p\varepsilon})^{-1}.
\end{align*}
In a similar way the right-hand side becomes
\begin{align*}\Lambda_p &\int_0^\eta \frac{|u_\varepsilon(t)|^p W_p(t)}{F_\eta(t)^p }\,dt+ s(w) \Lambda_p^{1/p'} \frac{|u_\varepsilon (\eta)|^p}{f_\eta(\eta) ^{p-1}}\notag\\
&= \left( 1/{p' }\right)^{p}\int_0^\eta f_\eta(t)^{s(w)\varepsilon p-1}w(t)^{-1}\,dt+ s(w) \Lambda_p^{1/p'} \frac{|u_\varepsilon (\eta)|^p}{f_\eta(\eta) ^{p-1}}
\notag\\
&= \left( 1/p' \right)^{p} {f_\eta(\eta)^{s(w) \varepsilon p}}({p\varepsilon})^{-1} + s(w) \Lambda_p^{1/p'} \frac{|u_\varepsilon (\eta)|^p}{f_\eta(\eta) ^{p-1}}.
\end{align*}
Finally we reach to
\begin{equation} \left(\left( 1/p' +s(w)\varepsilon\right)^p-\left( 1/{p' } \right)^{p}\right)
\ge s(w) \Lambda_p^{1/p'}p\varepsilon \qquad (0<\varepsilon ),
\end{equation}
and this inequality is clear from the convexity. \hfill $\Box$
\par\medskip
\par\noindent{\bf Part 2: Proof of the inequality (\ref{nct2}) }\par
In order to establish the inequality (\ref{nct2}) in Theorem \ref{T3.1} we shall estimate the term involving
$ |(|v|^{p/2})'|^2 F_\eta^{}$ from below and
use a positive term involving $ |v'|^p F_\eta^{p-1}$ to absorb negative error terms.
To this end we need more notations:\par
For $u,v \in C^1_c((0,\eta])$, we retain
$$
u(t)= g_\eta(t) v (t)\quad \mbox{and }\quad X(t)= \frac{g_\eta(t)}{g'_\eta(t)} \frac{ v'(t)}{v(t)}=p'F_\eta(t) \frac{ v'(t)}{v(t)}
s(w)\quad \textcolor{black}{(v(t)\neq0)}.
$$
By Remark \ref{density}, it suffices to assume in the proof that $u$ and $v$ belong to the class
$ \textcolor{black}{G((0,\eta])}$ defined by the following:
\begin{df}\label{df4.2}
\begin{equation}\label{4.22} G((0,\eta])=\{ v\in C_c^1((0,\eta]): v(t)\ge 0 \,\text { in } (0,\eta]\, \} \end{equation}
\end{df}
\begin{df}\label{df4.3} For $v \in G((0,\eta])$ and $M>1$ we define two subsets of $[0,\eta]$ as
follows:
\begin{equation} \begin{cases}& A(v,M)= \{ t\in [0,\eta] : \,\, \,
p' F_\eta(t) |v'(t)| \le M {v(t)}\}\\
& B(v,M)= \{ t\in [0,\eta] :\,\, \, p' F_\eta(t) |v'(t)| > M {v(t)}\}\\
\end{cases}
\end{equation}
$A(v,M)$ and $B(v,M)$ are sometimes abbreviated as $A$ and $B$ respectively.
\end{df}
\begin{rem}
From Definition \ref{df4.2} and Definition \ref{df4.3} we see $A\cup B=[0,\eta]$.
\textcolor{black}{ If $v(t)\in G((0,\eta])$ and $v(s)=0$ for some $s\in [0,\eta)$, then $v'(s)=0$, and hence $s\in A(v,M)$ for any $M>0$. }
We note that the set
$C(v,M)=\{ t\in [0,\eta] :\,\, \, p' F_\eta(t) |v'(t)| = M {|v(t)|}\}$ coincides with the set of critical points of $\, \Psi(t):=\log (|v|f_\eta^{\pm M/p'})$. Namely, $\{ t\in[0,\eta] : \Psi'(t)=0 \}$.
\end{rem}
The following lemma will be established in Appendix.
\begin{lem}\label{lemma4.4}Assume that $\eta>0$, $\mu>0$, $M>1$ and $ w\in W(\mathbf R_+)$. For any
$v\in G((0,\eta])$ we have the followings:
\begin{enumerate}\item For $A=A(v,M)$ and $B=B(v,M)$,
\begin{equation}\label{4.32}
\begin{split} &\int_A |(v(t)^{p/2})'|^2 F_\eta(t)\,dt \ge -\frac{ (\Lambda_p)^{1/p'}} {2\mu f_{\eta}(\eta)^{p-1}} |u(\eta)|^p
+\frac{1}{4(p')^{p-1}}\int_A\frac{|u(t)|^pW_p(t)}{F_\eta(t)^pG_\eta(t)^2} \,dt \\
&+\frac{1}{2(p')^{p-1}}\int_B\frac{|u(t)|^pW_p(t)}{F_\eta(t)^pG_\eta(t)^2}\,dt- \frac{p}{2\mu}\left(\frac{p'}{M}\right)^{p-1}\int_B|v'(t)|^p F_\eta(t)^{p-1}\,dt.
\end{split}
\end{equation}
\item
\begin{equation}\label{4.32'}
\begin{split}
\int_0^\eta |(v(t)^{p/2})'|^2 F_\eta(t)\,dt &\ge -\frac{ (\Lambda_p)^{1/p'}} {2\mu f_{\eta}(\eta)^{p-1}} |u(\eta)|^p\\&
+\frac{1}{4(p')^{p-1}}\int_0^\eta \frac{|u(t)|^pW_p(t)}{F_\eta(t)^pG_\eta(t)^2} \,dt
\end{split}
\end{equation}
\end{enumerate}
\end{lem}
\begin{rem}
The assertion 2 follows from Lemma \ref{lemma8.1} with $S=[0,\eta]$.
The inequality (\ref{nct2}) will follow from this lemma together with Lemma \ref{prop4.2}.
\end{rem}
\par\noindent
{\bf End of the proof of (\ref{nct2}):}
If $p\ge 2$, then for any $u\in G((0,\eta])$ we have by using (\ref{4.32'}) of Lemma \ref{lemma4.4}
\begin{align*} \int_0^\eta |u'(t)|^p W_p(t)\,dt
&\ge s(w)\frac{ (\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{p-1}} |u(\eta)|^p+ \Lambda_p \int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p}\\ &
+d(p)\left(-\frac{ (\Lambda_p)^{1/p'}} {2\mu f_{\eta}(\eta)^{p-1}} |u(\eta)|^p
+\frac{1}{4(p')^{p-1}}\int_0^\eta\frac{|u(t)|^pW_p(t)}{F_\eta(t)^pG_\eta(t)^2}\,dt\right)
\notag\\
&= s(w)\frac{ (\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{p-1}} \left( 1- s(w)\frac{d(p)}{2\mu}\right) |u(\eta)|^p +\Lambda_p \int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p}\,dt\notag
\\ &+\frac{d(p)}{4(p')^{p-1}}\int_0^\eta \frac{|u(t)|^pW_p(t)}{F_\eta(t)^pG_\eta(t)^2}\,dt.\notag
\end{align*}
Here we assume that $d(p) $ is so small that $ d(p)/\mu<1$ if $w\in Q(\mathbf R_+)$.
Then we get (\ref{nct2}) with
\begin{equation}\begin{cases}&L= { (\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{1-p}} \left( 1-s(w) {d(p)}/({2\mu})\right),\\ &\\&
C={d(p) (\Lambda_p)^{1/p'}}/{4}.\end{cases}
\end{equation}
If $1<p<2$, then for any $u\in G((0,\eta])$ and $M\ge 1$ we have
\begin{align*} & \int_0^\eta |u'|^p W_p\,dt
\ge s(w)\frac{ (\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{p-1}} |u(\eta)|^p+ \Lambda_p \int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p}\\ &
+c(p)(p')^{p-1}\int_{B(v,M)} |v'(t)|^p F_\eta(t)^{p-1} \,dt+ M^{p-2}d(p)\int_{A(v,M)}
|(|v(t)|^{p/2})'|^2 F_\eta \,dt. \notag
\end{align*}
Here $c(p)$ is a positive number independent of each $u$ and $d(p)= c(p) 4p'/p^2$.
For the last term we use (\ref{4.32}) of Lemma \ref{lemma4.4}, then we have
\begin{align*} & \int_0^\eta |u'(t)|^p W_p(t)\,dt
\ge
\frac{ s(w)(\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{p-1}} \left( 1-s(w) \frac{d(p)M^{p-2}}{2\mu}\right) |u(\eta)|^p \\&+\Lambda_p \int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p}
+c(p)(p')^{p-1}\left( 1- \frac{2}{\mu(p-1)M} \right)
\int_{B(v,M)} |v'(t)|^p F_\eta(t)^{p-1} \,dt
\\&
+\frac{d(p)M^{p-2} (\Lambda_p)^{1/p'}}{2}\left( \frac12\int_{A(v,M)} \frac{|u(t)|^pW_p(t)}{F_\eta(t)^pG_\eta(t)^2}\,dt+
\int_{B(v,M)} \frac{|u(t)|^pW_p(t)}{F_\eta(t)^pG_\eta(t)^2}\,dt\right).
\notag
\end{align*}
Then we take a sufficiently large $M$ so that we have $1- {2}/({\mu(p-1)M}) >0$ and ${d(p)M^{p-2} }/({2\mu})<1$ if $w\in Q(\mathbf R_+)$. Then
\begin{align*}
&\int_0^\eta |u'(t)|^p W_p(t)\,dt
\ge
s(w)L |u(\eta)|^p +\Lambda_p \int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p}
+C \int_0^\eta \frac{|u(t)|^pW_p(t)}{F_\eta(t)^pG_\eta(t)^2}\,dt,
\end{align*}
where \begin{equation}\begin{cases}&L= { (\Lambda_p)^{1/p'}} {f_{\eta}(\eta)^{1-p}} \left( 1-s(w) {d(p)M^{p-2}}/({2\mu})\right),\\ &\\&
C={d(p)M^{p-2} (\Lambda_p)^{1/p'}}/{4}.\end{cases}
\end{equation}
\qed
\subsection{Proof of Theorem \ref{T3.2}}
Let $v\in G((0,\eta])$.
Recall that
$
u(t)= g_\eta(t) v(t)$, $u'(t)= g'_\eta(t) v(t)\left( 1+ X(t)\right) $ $(v\neq0)$ and
\begin{equation*} X(t)= \frac{g_\eta(t)}{g'_\eta(t)} \frac{ v'(t)}{v(t)}=p'F_\eta (t)\frac{ v'(t)}{v(t)}\textcolor{black}{s(w)}\quad (v(t)\neq 0); 0\quad (v(t)=0).
\end{equation*}
Then we have two elementary lemmas.
\begin{lem}\label{lemma4.7} For any $X\in \mathbf R$ and any $M>1$,
\begin{equation} |1+X|^p \le \begin{cases}& (1+M)^p, \qquad |X|\le M, \\
& 2^p |X|^p,\qquad \quad |X|> M.
\end{cases}
\end{equation}
\end{lem}
\begin{lem}\label{lemma4.8}For any $M>1$,
\begin{equation} |u'(t)|^p W_p(t)\le \begin{cases}& \Lambda_p(1+M)^p { u(t)^p W_p(t)}{ F_\eta(t)^{-p}}, \quad\qquad |X(t)|\le M,\\
& 2^p\Lambda_p^{-1/p'} |v'(t)|^p F_\eta(t)^{p-1},\qquad \qquad \quad \, \,\, |X(t)|> M.
\end{cases}
\end{equation}
\end{lem}
\par\noindent{\bf Proof:} Since $u'(t)= g'_\eta(t) v(t)\left( 1+ X(t)\right) $, we get
$$|u'(t)|^pW_p(t)\le |g'_\eta(t)|^pv(t)^p|\,1+X(t)|^p \textcolor{black}{W_p(t).}$$
If $|X(t)|\le M$, then from Lemma \ref{lemma4.7} and (\ref{4.15})
we get the desired estimate. If $|X(t)|\ge M$, then from Lemma \ref{lemma4.7} and (\ref{4.17}) we get the desired one.\hfill $\Box$\par
By \textcolor{black}{ Proposition \ref{prop2.1} } we have the following.
\begin{lem}\label{lemma4.9} Assume that $w$ is admissible in the sense of Definition \ref{admissibility}. Then, there are some positive numbers $K$ and $\eta$ such that we have for
any $u\in C_c^1(0,\eta])$
\begin{equation}\int_0^\eta \frac{ u(t)^pW_p(t)}{ F_\eta(t)^p}t\,dt \le K^2\int_0^\eta \frac{ |u(t)|^p W_p(t) \,dt}{ F_\eta(t)^p G_\eta(t)^2}.\end{equation}
\end{lem}
\par\noindent{\bf Proof of Theorem \ref{T3.2}:}
For a sufficiently small $\eta>0$, temporally we set
\textcolor{black}{\begin{equation} Q_{ L}(u)= \int_0^\eta |u'(t)|^p W_p(t)\,dt
-\Lambda_p \int_0^\eta \frac{ |{u(t)}|^p W_p(t)\,dt }{F_\eta(t)^p}-s(w) {L}|u(\eta)|^p,
\end{equation}}
where $L$ is a positive constant independent of each $u$, \textcolor{black}{which will be specified later}.
In order to prove Theorem \ref{T3.2}, it suffices to control three terms below by $Q_{L}(u)$ :
$$\int_0^\eta \frac{ |u(t)|^p W_p(t) \,dt}{ F_\eta(t)^p G_\eta(t)^2}, \quad \int_0^\eta \frac{ u(t)^pW_p(t)}{ F_\eta(t)^p}\, t\,dt \quad\text{and} \quad \int_0^\eta |u'(t)|^p W_p(t) \, t\,dt.$$
It follows form Theorem \ref{T3.1} (\ref{nct2}) that we have for $D=C^{-1}$
$$\int_0^\eta \frac{ |u(t)|^p W_p(t) \,dt}{ F_\eta(t)^p G_\eta(t)^2}\le D\cdot Q_L(u),$$
where $L=L(w,p,\eta,\mu)$ is the same constant in (\ref{nct2}).
From this and Lemma \ref{lemma4.9} we also have
$$\int_0^\eta \frac{ u(t)^pW_p(t)}{ F_\eta(t)^p}\, t\,dt \le K^2 D\cdot Q_L(u).$$
By using Lemma \ref{lemma4.8} we see that
\begin{align*}\int_0^\eta |u'(t)|^p W_p(t) \, t\,dt &\le \Lambda_p(1+M)^p\int_{|X|\le M} \frac{ u(t)^pW_p(t)}{ F_\eta(t)^p}\, t\,dt \\&+2^p\Lambda_p^{1/p-1}\int_{|X|> M}|v'(t)|^p F_\eta(t)^{p-1}\, t\,dt.\end{align*}
Then, the first term in the right-hand side can be controlled by the same $Q_L(u)$.
Now we claim that for some positive number $K'$ \begin{equation} \int_{|X|> M}|v'(t)|^p F_\eta(t)^{p-1}t\,dt\le K'\cdot Q_{L}(u), \label{4.43}\end{equation}
where $L= {(\Lambda_p)^{1/p'}}{f_\eta(\eta)^{1-p}} $.
\par\noindent
First we assume that $p\ge 2$.
By (\ref{4.222}) of Lemma \ref{prop4.2} with the last term replaced by $$c(p)(p')^{p-1}\int_0^\eta |v'(t)|^p F_\eta(t)^{p-1} \,dt,$$we see that (\ref{4.43}) is valid, provided that $\eta$ is sufficiently small (See also Remark \ref{rem4.2}). Secondly we assume that $1<p<2$.
Then in a similar way the assertion follows from (\ref{4.23}) of Lemma \ref{prop4.2}.
As a result we have the desired inequality (\ref{c1}) for some positive numbers $C_0, C_1$ and $L$ which are independent of each $u$.
\hfill$\Box$
\section{ Proofs of Theorem \ref{CT3} and Corollary \ref{NCC1}}
We establish Theorem \ref{CT3} and Corollary \ref{NCC1} using Theorem \ref{T3.2}.
\par \medskip
\noindent{\bf Proof of Theorem \ref{CT3}:}
Let us prepare some notations and fundamental facts.
Define $\Sigma = \partial \Omega$ and $\Sigma_t =\{ x\in \Omega : \delta(x)= t\}$.
Since $\Sigma$ is is of class $C^2$, there exists an $\eta_0>0$ such that we have
a $C^2$ diffeomorphism $ G: \Omega_\eta \mapsto (0,\eta)\times \Sigma$
for any $\eta\in (0,\eta_0)$. By $G^{-1}(t, \sigma) \, ((t,\sigma) \in (0,\eta_0) \times \Sigma)$ we denote the inverse of $G$. Let $H^{}_t$ denote the mapping
$G^{-1}(t, \cdot) $ of $\Sigma$ onto $\Sigma_t$.
This mapping is also a $C^2$ diffeomorphism and its Jacobian is close to $1$ in $(0,\eta_0)\times \Sigma$.
Therefore,
for every non-negative continuous function $u$ on $\overline{\Omega_\eta}$ with $\eta\in (0,\eta_0)$ we have
\begin{align}& \int_{\Omega_\eta}u = \int _0^\eta \,dt \int_{\Sigma_t} u \,d\sigma_t =\int _0^\eta \,dt \int_\Sigma u( H_t(\sigma)) ( \rm{Jac}\, H_t )\,d\sigma, \label{5.1}\\
& | \rm{Jac}\, H_t (\sigma)-1|\le ct, \quad \mbox{ for every } (t,\sigma)\in (0,\eta_0)\times \Sigma, \label{5.2}
\end{align}
where $c$ is a positive constant independent of each $(t,\sigma)$, $d\sigma$ and $d\sigma_t$ denote surface elements on $\Sigma$ and $\Sigma_t$ respectively.
Then we have
\begin{align}
&
\int_\Sigma u( H_\eta(\sigma)) ( 1-c\eta)\,d\sigma \le
\int_{\Sigma_\eta}u\,d\sigma_\eta\le \int_\Sigma u( H_\eta(\sigma)) ( 1+c\eta)\,d\sigma. \label{4.2}
\end{align}
Again for the sake of simplicity, we denote $W_p(\delta(x))$ and $F_\eta(\delta(x))$ by $W_p(\delta)$ and $F_\eta(\delta)$ respectively, and
the symbol $dx$ is often abbreviated.
Then we immediately have for $v=u(H_t(\sigma))$
\begin{align}
\int_{\Sigma} \,d\sigma \int_0^\eta \left |\frac{\partial v}{\partial t} \right |^p (1-ct) &
W_p(t) \,dt \le \int_{\Omega_\eta}|\nabla u|^p W_p(\delta)\,dx, \notag\\
\int_{\Sigma} \,d\sigma \int_0^\eta \frac{| v|^pW_p(t) }{F_\eta(t)^p} (1-ct) \,dt &\le \int_{\Omega_\eta}\frac{| u|^pW_p(\delta)\,dx }{F_\eta(\delta)^p}\le \int_{\Sigma} \,d\sigma \int_0^\eta\frac{| v|^pW_p(t) }{F_\eta(t)^p} (1+ct) \,dt. \notag
\end{align}
\noindent{\bf Proof of (\ref{2.11}):} Under these consideration,
(\ref{2.11}) is reduced to the following one dimensional Hardy's inequality. Setting $v(t) =u(H_t(\sigma))$ and $v'= \partial v/\partial t$ we have
\begin{equation}
\begin{split} \int_0^\eta & |v'(t)|^p W_p(t)(1-ct)\,dt \ge
\Lambda_{p} \int_0^\eta \frac{|v(t)|^p W_p(t)}{F_\eta(t)^p} (1+ct)\,dt \\
&+ C \int_0^\eta \frac{|v(t)|^p W_p(t)}{F_\eta(t)^p \textcolor{black}{G_\eta(t)^2} }(1+ct)\,dt
+ s(w) L' |v(\eta)|^p W_p(\eta)(1+s(w)c\eta).
\end{split}
\end{equation}
Equivalently we have
\begin{equation} \label{5.7}
\begin{split} \int_0^\eta & \left( |v'(t)|^p -\frac{|v(t)|^p}{F_\eta(t)^p} \left(\Lambda_{p} + \frac{C }{G_\eta(t)^2}\right)\right) W_p(t)\,dt
\\ &\ge c
\int_0^\eta \left( |v'(t)|^p + \frac{|v(t)|^p}{F_\eta(t)^p} \left(\Lambda_{p} + \frac{C }{G_\eta(t)^2}\right) \right) W_p(t)\, t \,dt \\&
+s(w) L' |v(\eta)|^p W_p(\eta)(1+s(w)c\eta).
\end{split}
\end{equation}
Assume that $0<\eta\le 1/(2c)$. Then $1-c\eta\ge 1/2$ and (\ref{5.7}) clearly follows from Theorem \ref{T3.2} with
$ C_0=C $, $C_1=c$ and $L =L'W_p(\eta)(1+s(w)c\eta)$. \hfill $\Box$
\par\medskip
\noindent{\bf Proof of Corollary \ref{NCC1}: }
First we treat the case that \textcolor{black}{$w\in P_A(\mathbf R_+)$}.
Assume that Hardy's inequality (\ref{2.7}) does not hold.
Then there exists a sequence of functions
$\{u_k\} \subset W^{1,p}_{0}(\Omega; W_p(\delta)) \cap C(\Omega)$
such that
\begin{equation} \begin{cases}&\lim_{k\to \infty}\left(\int_{\Omega} |\nabla u_k|^pW_p(\delta)\,dx + \int_{\Sigma_\eta}|u_k|^p W_p(\delta)\,d\sigma_\eta\right)=0,\\ &
\int_{\Omega} {|u_k|^pW_p(\delta)}/{F_\eta^p(\delta)}\, dx =1 \quad (k=1,2,\cdots).
\end{cases} \label{5.8}
\end{equation}
For a sufficiently small $\eta>0$,
let $W^{1,p}(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta))$ be
given by the completion of $C^\infty(\Omega\setminus \overline{\Omega_\eta} )$ with respect to the norm
defined by
\begin{equation*}
\| u\|_{W^{1,p}(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta)) } =
\| |\nabla u| \|_{ L^p(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta))} + \| u\|_{ L^p(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta))}.
\end{equation*}
Since $W_p(\delta)>0$ in $\overline{\Omega\setminus{\Omega_\eta}}$, $W^{1,p}(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta)) $ is well-defined and becomes
a Banach space with the norm $\| \cdot\|_{W^{1,p}(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta)) }
$.
We note that
$\int_{\Sigma_\eta}|u_k|^p W_p(\delta)\,d\sigma_\eta$ $(k=1,2,3\ldots)$ is bounded, because the trace operator $T: W^{1,p}(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta))$ $\mapsto $
$L^p(\Sigma_\eta; W_p(\eta) )$ is continuous.
By Theorem {\ref{CT3}} we have
\begin{align*}
&\int_{\Omega} |\nabla u_k|^pW_p(\delta)\,dx = \int_{\Omega_\eta} |\nabla u_k|^pW_p(\delta) \,dx
+ \int_{\Omega\setminus \Omega_\eta} |\nabla u_k|^pW_p(\delta)\,dx \\
&\ge \Lambda_{p}\left( 1- \int_{\Omega \setminus \Omega_\eta}\frac{ {|u_k|^p W_p(\delta)}}{F_\eta^p(\delta)}\,dx\, \right)
+ \int_{\Omega\setminus \Omega_\eta} |\nabla u_k|^pW_p(\delta)\,dx-L' \int_{\Sigma_\eta} |u_k|^pW_p(\delta)
\,d\sigma_\eta.
\end{align*}
Since $\delta \ge \eta$ in $\Omega\setminus \Omega_\eta$, by the standard argument we have $u_k\to C \,( constant)$ in $W^{1,p}(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta))$ as $k\to \infty$. By (\ref{5.8}) we have $C=0$.
Hence we see
$ 0 \ge \Lambda_{p}$, and we reach to a contradiction.
\par Secondly we treat the case that $w\in Q(\mathbf R_+)$. In this case we assume that
\begin{align} &\lim_{k\to \infty}\int_{\Omega} |\nabla u_k|^pW_p(\delta)\,dx=0, \quad
\int_{\Omega} \frac{|u_k|^pW_p(\delta)}{F_\eta^p(\delta)}\,dx =1 \quad (k=1,2,\cdots). \label{5.9}
\end{align}
Then,
by Theorem {\ref{CT3}} we have
\begin{align}
&\int_{\Omega} |\nabla u_k|^pW_p(\delta)\,dx = \int_{\Omega_\eta} |\nabla u_k|^pW_p(\delta)\,dx + \int_{\Omega\setminus \Omega_\eta} |\nabla u_k|^pW_p(\delta) \,dx \label{5.10}\\
&\ge \Lambda_{p}\left( 1- \int_{\Omega \setminus \Omega_\eta}\frac{ {|u_k|^p W_p(\delta)}}{F_\eta^p(\delta)}\,dx \right)
+ \int_{\Omega\setminus \Omega_\eta} |\nabla u_k|^pW_p(\delta)\,dx+L' \int_{\Sigma_\eta} |u_k|^pW_p(\delta)\,d\sigma_\eta. \notag
\end{align}
Again we have $u_k\to C \,( constant)$ in $W^{1,p}(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta))$ as $k\to \infty$. By (\ref{5.9}) and (\ref{5.10}) we have $C=0$.
Hence
we see
$ 0 \ge \Lambda_{p}$, and this is a contradiction.
\hfill $\Box$
\section{Proof of Theorem \ref{NCT2} }
By virtue of Corollary \ref{NCC1} and its proof, it suffices to show the implication $1\to 2$. Since $F_\eta^{-1}\notin L^1((0,\eta))$, for an arbitrary $\varepsilon>0$ we have
$G_\eta (\delta)^{-1} <\varepsilon$ in $ \Omega_\eta$ provided that $\eta$ is sufficiently small.
Therefore we assume that $C=0$ without the loss of generality.
\par\medskip
Here we introduce a Banach space ${W_0^{1,p}(\Omega_\eta; W_p(\delta)) }$ which corresponds to $W_{0}^{1,p}((0,\eta]; W_p)$ defined in Subsection 3.1.
Let $ \tilde{C}_c^\infty(\Omega_\eta)= \{ u|_{\Omega_\eta}: u\in C_c^\infty(\Omega)\}$, where
by $ u|_{\Omega_\eta}$ we denote the restriction of $u $ to $\Omega_\eta$. Let $W_0^{1,p}(\Omega_\eta; W_p(\delta))$ be
given by the completion of $\tilde{C}_c^\infty(\Omega_\eta)$ with respect to the norm
defined by
\begin{equation*}
\| u\|_{W_0^{1,p}(\Omega_\eta; W_p(\delta)) } =
\| |\nabla u| \|_{ L^p(\Omega_\eta; W_p(\delta))} + \| u\|_{ L^p( \Omega_\eta; W_p(\delta))}.
\end{equation*}
Then,
$W_0^{1,p}(\Omega_\eta; W_p(\delta)) $ is a Banach space with the norm $\| \cdot\|_{W^{1,p}_{0}(\Omega_\eta; W_p(\delta)) }
$.
Then we prepare a lemma on extension:
\begin{lem}\label{l11}\mbox{\rm ( Extension )} Assume that $\eta_0$ is a sufficiently small positive number. Then
for any $\eta \in (0,\eta_0) $ there exists an extension operator
$E=E(\eta) : W_0^{1,p}(\Omega_\eta; W_p(\delta))\mapsto W_0^{1,p}(\Omega; W_p(\delta))$ such that:
\begin{enumerate}\item $E (u)= u \quad $ a.e. in $\Omega_\eta$
\item There exists some positive number $C=C(\eta)$ such that for any
$u\in W_0^{1,p} ( {\Omega_\eta}; W_p(\delta))$
$$ \| |\nabla E( u)| \|_{L^p(\Omega; W_p(\delta))}
\le C\Big( \||\nabla u\||_{L^p (\Omega_{\eta/2}; W_p(\delta) )}+ \|u\|_{W^{1,p} (\Omega_\eta \setminus \overline{\Omega_{\eta/2}} ; W_p(\delta))}\Big).$$
\end{enumerate}
\end{lem}
Admitting this for the moment, we prove Theorem \ref{NCT2}.
First we treat the case that $w\in P(\mathbf R_+)$. Then we see that $W_p/F_\eta^p\in L^1(0,\eta)$.
In fact we have $\int_0^\eta W_p(t)/F_\eta(t)^p\,dt= \mu^{1-p}/(p-1).$
Now we assume that
Hardy's inequality (\ref{2.11}) with $C=0$ does not hold.
Then there exists a sequence of functions $\{u_k\} \subset W_0^{1,p}(\Omega,W_p(\delta))\cap C(\Omega)$ satisfying
\begin{equation} \begin{cases}&\lim_{k\to \infty}\left(\int_{\Omega_\eta} |\nabla u_k|^pW_p(\delta)\,dx + \int_{\Sigma_\eta}|u_k|^p W_p(\delta)\,d\sigma_\eta\right)=0,\\ &
\int_{\Omega_\eta} {|u_k|^pW_p(\delta)}/{F_\eta^p(\delta)}\,dx =1 \quad (k=1,2,\cdots).
\end{cases} \label{6.1}
\end{equation}
By $w_k=u_k|_{\Omega_\eta}$ we denote the restriction of $u_k$ to $\Omega_\eta$.
Then $ E(w_k) \in W_0^{1,p}(\Omega;W_p(\delta))$ for $k=1,2,\ldots$.
On the other hand, it follows from (\ref{6.1}) that
$w_k=u_k\to C$ a.e. in $\Omega_\eta$ for some constant $C$ as $k\to\infty$. From (\ref{6.1}) we have $C=0$. Then, by the assumption 1 and the continuity of the trace operator $T: W^{1,p}(\Omega\setminus \overline{\Omega_\eta}; W_p(\delta)) \mapsto
L^p(\Sigma_\eta; W_p(\delta ) )$ for a small $\eta>0$, we have
\begin{align} 1 &\le \int_\Omega \frac{|E(w_k)|^pW_p(\delta)}{F_\eta^p(\delta)}\,dx \label{6.2}\\& \le \gamma^{-1}\left(
\int_\Omega |\nabla E(w_k)|^pW_p(\delta)\,dx + L' \int_{\Sigma_\eta} |w_k|^pW_p(\delta)\,d\sigma_\eta \right) \notag\\
& \le \gamma^{-1}C'\Big( \||\nabla u_k| \| _{L^p (\Omega_{\eta/2}; W_p(\delta) )}
+ \textcolor{black}{ \|u_k\|_{{W^{1,p}}(\Omega_\eta \setminus \overline{\Omega_{\eta/2}};W_p(\delta))} }\Big)^{\textcolor{black}{p}} \to 0, \quad\text{ as } k\to\infty,\notag
\end{align}
where $C'$ is some positive number independent of each $u_k$.
But this is a contradiction.
\par Secondly we treat the case that $w\in Q(\mathbf R_+)$. Since
$\int_0^\eta W_p(t)/F_\eta(t)^p\,dt= +\infty$, we see that $W_p/F_\eta^p\notin L^1(0,\eta)$.
There is a sequence of functions $\{u_k\} \subset W_0^{1,p}(\Omega; W_p(\delta))\cap C(\Omega)$ satisfying
\begin{align} &\lim_{k\to \infty}\int_{\Omega_\eta} |\nabla u_k|^pW_p(\delta)\,dx =0, \quad
\int_{\Omega_\eta} \frac{|u_k|^pW_p(\delta)}{F_\eta^p(\delta)} \,dx =1 \quad (k=1,2,\cdots). \label{6.3}
\end{align}
As before we see that $u_k\to C$ (constant) in $\Omega_\eta$. From (\ref{6.3}) we have $C=0$.
Then, by the assumption 1 and the continuity of the trace operator $T$, we have
\begin{align} 1 &\le \int_\Omega \frac{|E(w_k)|^pW_p(\delta)}{F_\eta^p(\delta)}\,dx \label{6.4}\\ &\le \gamma^{-1}\left(
\int_\Omega |\nabla E(w_k)|^pW_p(\delta) \,dx - L' \int_{\Sigma_\eta} |w_k|^pW_p(\delta) \,d\sigma_\eta \right)\\
& \le \gamma^{-1}C'\Big( \| |\nabla u_k| \|_{L^p (\Omega_{\eta/2}; W_p(\delta) )}
+ \textcolor{black}{ \|u_k\|_{{W^{1,p}} (\Omega_\eta \setminus \overline{\Omega_{\eta/2}}; W_p(\delta))}} \Big)^{\textcolor{black}{p}}\to 0, \quad\text{ as } k\to\infty,\notag
\notag
\end{align}
where $C'$ is some positive number.
But this is a contradiction.
\hfill $\Box$
\par\medskip
\noindent{\bf Proof of Lemma \ref{l11}: } Since $\eta $ is small and $\delta$ is Lipschitz continuous,
$\partial \Omega_\eta$ and $\partial \Omega_{\eta/2}$ are Lipschitz compact manifolds. By the standard theory
( Theorem 1 in Section 5.4 of \cite{E} for example ) we have an extension operator $\tilde{E} : W^{1,p} (\Omega_\eta \setminus \overline{\Omega_{\eta/2}} ; W_p(\delta))\mapsto
W^{1,p}(\Omega\setminus \overline{\Omega_{\eta/2}}; W_p(\delta))$ such that $\tilde E(u)= u$ a.e. in $\Omega_\eta \setminus \overline{\Omega_{\eta/2}}$, and
$$ \| |\nabla \tilde{E}( u)|^p\|_{L^p(\Omega\setminus \overline{\Omega_{\eta/2}}; W_p(\delta))}
\le C(\eta) \textcolor{black}{ \|u\|_{{W^{1,p} }(\Omega_\eta \setminus \overline{\Omega_{\eta/2}};W_p(\delta))}}.$$
Define for $u\in W_0^{1,p}(\Omega_\eta; W_p(\delta))$
\begin{equation} { E}(u) = u, \, (x\in \overline{\Omega_{\eta/2}},); \quad \tilde{E}(u), \, (x \in \Omega\setminus \overline{\Omega_{\eta/2}} ).
\end{equation}
Then the assertion follows. \hfill$\Box$
\section{Proofs of Propositions \ref{ct1} and \ref{CT1}}
Proposition \ref{ct1} is known in a more general fashion. In fact
a variant is seen in Maz'ya \cite{Ma} ( Lemma 2, p144). For the sake of reader's convenience we give an elementary verification.
We note that
Proposition \ref{CT1} is a direct consequence of Proposition \ref{ct1}.
\par\medskip
\noindent{\bf Proof of Proposition \ref{ct1}:}
For $w\in P(\bf R_+)$ and $\overline\varepsilon \in (0,\eta/2)$, define
\begin{equation} \varphi_{\overline\varepsilon}= \,\,
0 \,\,\,( 0\le t\le \overline\varepsilon) ; \,\,\, \frac {f_\eta(\overline\varepsilon)-f_\eta(t) }{f_\eta(\overline\varepsilon)-f_\eta(\eta/2) }\, \,\,(\overline\varepsilon \le t\le \eta/2);\,\,\, 1\, \,\,(\eta/2 \le t \le \eta). \label{7.1}\end{equation}
Noting that $f_\eta(\overline\varepsilon)= \mu+ \int_{\overline\varepsilon}^\eta 1/w(s)\,ds\to +\infty$ as $\overline\varepsilon\to +0$, we see that
$ \int _0^\eta|\varphi_{\overline\varepsilon} '|^p W_p(t)\,dt=( f_\eta(\overline\varepsilon)-f_\eta(\eta/2))^{1-p} \to 0 \, \mbox{ as } \overline\varepsilon\to +0.$
On the other hand we have $\varphi_{\overline\varepsilon}(0)=0, \varphi_{\overline\varepsilon} (1)=1$ and hence the assertion is now clear.
Further we note that
$$\int_{\overline\varepsilon}^\eta |\varphi_{\overline\varepsilon}|^p \frac{W_p(t)}{F_\eta(t)^p}\,dt
\ge \int_{\eta/2}^\eta \frac{W_p(t)}{F_\eta(t)^p}\,dt=
\frac{f_\eta(\eta)^{1-p}- f_\eta(\eta/2)^{1-p}}{p-1} >0
\quad \mbox{ as } \overline\varepsilon\to +0.$$
\qed
\noindent{\bf Proof of Proposition \ref{CT1}:}
If a positive number $\eta_0$ is sufficiently small, then one can assume that $\delta\in C^2(\Omega_{\eta_0})$, $|\nabla \delta|=1$ in $ \Omega_{\eta_0} $ and
a manifolds $\{ x\in \Omega; \delta =\eta\}$ is of $C^2$ class for $\eta \in (0,\eta_0]$.
Let $\varphi_{\overline\varepsilon}$ be defined by (\ref{7.1}).
By virtue of (\ref{4.2}) we have
$$ \int_{\Omega_\eta}|\nabla \varphi_{\overline\varepsilon}(\delta(x))|^p {W_p(\delta(x))}\,dx\le
\int_{\Sigma} \,d\sigma \int_0^\eta \left | \varphi'_{\overline\varepsilon}(t) \right |^p {W_p(t)}
(1+ct) \,dt,$$
hence the assertion follows from Proposition \ref{ct1}.\qed
\section{Appendix}
\subsection{ Proof of Lemma \ref{lemma4.4}}
Let $v\in G((0,\eta])$.
We recall that
$$ g_\eta=( p' f_\eta)^{{1}/{p'}},\quad F_\eta= w f_\eta, \quad G_\eta= \mu+ \int_t^\eta\frac{ds}{F_\eta(s)}\quad\mbox{and}\quad u= g_\eta \nu.$$
Let us prepare more notations.
Let us set
\begin{equation} v^{{p}/{2}}=z,\quad z= a\varphi \quad\mbox{and}\quad a=(G_\eta)^{1/2},
\end{equation}
where $ \varphi $ is some function in $ C^1((0,\eta])$ with $\varphi(0)=0$.\par\noindent
Then we have
\begin{align*} z' &= \varphi' a + \varphi a'\quad\mbox{and}\quad
|z'|^2 = (\varphi')^2 a^2 + \frac 12 (\varphi^2)' (a^2)' + \varphi^2 (a')^2.
\end{align*}
Noting that $a^2= G_\eta,
\,(a^2)' = -{1}/{F_\eta}\, ${and} $\, a' = - {1}/({2 F_\eta a}),$
we have
\begin{align*} |z'|^2F_\eta= (\varphi')^2F_\eta G_\eta -\frac12 (\varphi^2)' + \frac{\varphi^2}{ 4F_\eta G_\eta} \ge -\frac12 (\varphi^2)' + \frac{\varphi^2}{ 4F_\eta G_\eta},
\end{align*}
and we also have
$$\varphi^2=\frac{v^p}{a^2}= \frac{ u^p}{g_\eta^pG_\eta}
= \textcolor{black}{ \frac{1}{(p')^{p-1}}}\frac{u^p W_p}{ F_\eta^{p-1} G_\eta}. $$
As a result, we have the following that is valid for any measurable set $S\subset[0,\eta]$.
\begin{lem}\label{lemma8.1} Let $S$ be a measurable set contained in $[0,\eta]$.
Assume that $\eta>0$, $\mu>0$ and $ w\in W(\mathbf R_+)$. For any nonnegative $v\in G([0,\eta])$
\begin{align} \int_S|(v(t)^{p/2})'|^2 F_\eta(t)\,dt
\ge -\frac12 \int_S(\varphi(t)^2)' \,dt+ \frac{1}{4(p')^{p-1}}\int_S\frac{u(t)^p W_p}{ F_\eta(t)^p G_\eta(t)^2}\,dt.
\end{align}
\end{lem}
\par\noindent{\bf Proof of Lemma \ref{lemma4.4}: }
\textcolor{black}{
By using Lemma \ref{lemma8.1} with $S= [0,\eta]$ we immediately have (\ref{4.32'}).} Then we proceed to (\ref{4.32}).
We assume that $A=A(v,M)$ and $B=B(v,M)$ are defined by Definition \ref{df4.3}.
As for the first term of the right-hand side, we have
\begin{align*}
-\frac12 \int_A(\varphi(t)^2)' \,dt &=-\frac12 \int_0^\eta(\varphi(t)^2)'\,dt +\frac12 \int_B(\varphi(t)^2)' \,dt\\&
=-\frac12 \varphi(\eta)^2 -\int_B\frac{z(t)^2 a'(t)}{a(t)^3} +\int_B\frac{z(t) z'(t)}{a(t)^2}.
\end{align*}
Moreover we see that:
\begin{enumerate}
\item
\begin{equation}\varphi(\eta)^2=\frac{z(\eta)^2}{a(\eta)^2}=\frac{1}{\mu}
\frac{|u(\eta)|^p}{g_\eta(\eta)^p} = \frac{|u(\eta)|^p}{\mu(p')^{p-1} f_\eta(\eta)^{p-1}}.
\end{equation}
\item
\begin{equation}-\int_B\frac{z(t)^2 a(t)'}{a(t)^3}\,dt=
\frac12 \int_B\frac{z(t)^2 }{a(t)^4 F_\eta(t)}\,dt=\frac{1}{2(p')^{p-1}}\int_B\frac{u(t)^p W_p(t)}{ F_\eta(t)^p G_\eta(t)^2}\,dt.
\end{equation}
\item
Noting that $ zz'= p v^{p-1}v'/2 $, $ |v|\le {p'}|v'|F_\eta /{M}$ in $B$ and $G_\eta\ge \mu$, we have
\begin{align*}\left|\int_B\frac{z(t)z'(t) }{a(t)^2 }\,dt\right|
&\le
\frac{p}{2}\int_B \frac{\left(\frac{p'}{M} |v'(t)| F_\eta(t)\right)^{p-1}|v'| }{G_\eta(t)} \,dt\\
& \le \frac{p}{2\mu} \left(\frac{p'}{M} \right)^{p-1}
\int_B |v'(t)| ^{p}F_\eta(t)^{p-1}\,dt.
\end{align*}
\end{enumerate}
Finally we get the desired inequality (\ref{4.32}) which proves Lemma \ref{lemma4.4}.
\hfill $\Box$
\subsection{Auxiliary inequalities in the noncritical case}
If we restrict ourselves to the case that $w\in Q(\mathbf R_+)$, then Hardy's inequality (\ref{8.10})
follows from the next simple lemma which is provided in \cite{BM} for $p=2$ and $W_p=w=1$.
\begin{lem}\label{l2} \mbox{\rm ( $w\in Q(\mathbf R_+)$)}
Assume that \textcolor{black}{ $ f\in C^{}([0,\eta])\cap C^1((0,\eta])$ } is a monotone nondecreasing function such that $f(\eta)\le 1$. Assume that $1<p<\infty$ and $w\in Q(\mathbf R_+)$.
Then for every
$u\in C_c^1((0,\eta])$, we have
\begin{equation} \int_0^\eta\left ( |u'(t)|^p -\Lambda_{p} \frac{|u(t)|^p}{F_\eta(t)^p} \right)W_p(t)\,dt \ge
\int_0^\eta \left( |u'(t)|^p -\Lambda_{p} \frac{|u(t)|^p}{F_\eta(t)^p} \right)W_p (t)f (t) \,dt. \label{3.2}
\end{equation}
In particular we have
\begin{equation} \int_0^\eta |u'(t)|^p W_p(t)\,dt \ge
\Lambda_{p} \int_0^\eta \frac{|u(t)|^p}{F_\eta(t)^p} W_p (t) \,dt. \label{8.10}
\end{equation}
\end{lem}
\par\medskip
\noindent{\bf Proof of Lemma \ref{l2}:} Without loss of generality
we assume that $f\ge 0$, $f(\eta)=1$, and $u\ge0$. Define $ g= 1-f$. Then $g\ge 0$ and $ g'\le 0$.
Noting that $u\in C^1_c((0,\eta])$ and $$ \frac d{dt}\left(\int_0^t\frac 1{w(s)}\,ds \right)^{1-p}= (1-p)\frac{W_p(t)}{F_\eta(t)^p},$$
by integration by parts we have
\begin{align*} & (p-1)\int_0^\eta \frac{u(t)^p W_p(t)}{F_\eta(t)^p} g\,dt \\ &=
\int_0^\eta u(t)^p \left( \int_0^t \frac1{w(s)}\,ds\right)^{1-p} g'(t)\,dt + p \int_0^\eta u(t)^{p-1} u'(t) g(t)\left( \int_0^t \frac1{w(s)}\,ds\right)^{1-p}\,dt.
\end{align*}
Since $g'=-f'\le 0$ and $g\ge 0$,
\begin{align*} \frac1{p'} \int_0^\eta \frac{u(t)^p W_p(t)}{F_\eta^p(t)} g(t)\,dt
\le \int_0^\eta u(t)^{p-1} u' (t) g(t) \left( \int_0^t \frac1{w(s)}\,ds\right)^{1-p}\,dt.
\end{align*}
By H\"older's inequality, we have
\begin{equation*}\begin{split}
\int_0^\eta u(t)^{p-1} u' (t) g(t) &\left( \int_0^t \frac1{w(s)}\,ds\right)^{1-p}\,dt\\
&\le \left( \int_0^\eta \frac{u(t)^p W_p(t)}{F_\eta(t)^p} g(t)\,dt \right)^{1/{p'}}
\left( \int_0^\eta |u'(t) |^p
W_p (t)g(t)\,dt\right)^{1/p}.
\end{split}
\end{equation*}
Hence we have
\begin{align*} \frac1{p'}\left( \int_0^\eta \frac{u(t)^p W_p(t)}{F_\eta(t)^p} g(t)\,dt \right)^{1/p}
\le \left( \textcolor{black}{\int_0^\eta} |u'(t) |^p
W_p (t)g(t)\,dt\right)^{1/p}.
\end{align*}
Using $ g=1-f$ and the definition of $\Lambda_{p}$, we have (\ref{3.2}).
\hfill $\Box$
\par\bigskip
\noindent{{\bf Acknowledgments}}\par\medskip
The author would like to thank the referee for the precise advice in completing this paper.
\medskip
| 6726c7e8ed7d96883c47fb81d347b81222e8a287 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The recent rapid progress of information technologies,
including machine learning,
has led to studies on novel computing concepts and hardware,
such as neuromorphic processing \cite{Markovic2020,Rabinovich2006,Furber2016,Merolla2014,Feldmann2019,Tait2017},
reservoir computing \cite{Versraeten2007,Jaeger2004,Tanaka2019,Appeltant2011,Brunner2013,Inubushi2020},
and deep learning \cite{LeCun2015,Xu2018,Lin2018,Shen2017}.
In particular, deep learning has become a groundbreaking
tool for data processing owing to its high-level performance \cite{LeCun2015}.
Furthermore, the energy-efficient computing for deep learning is gaining importance
with the rising need for processing large amounts of data \cite{XChen2014}.
An underlying key factor of deep learning is
its high expressive power, which is the result of
the layer-to-layer propagation
of information in the deep network.
This expressive power enables the representation of
extremely complex functions in a manner
that cannot be achieved using shallow networks with the same number of neurons
\cite{Poole2016,Montufar2014}.
Interestingly, recent studies have reported
that the information propagation in
multilayer systems can be expressed as the time evolution of
dynamical systems \cite{Liu2019,Chen2018,Benning2019,Haber2017}.
From the point of view of dynamical systems, the learning process of
networks can be regarded as the optimal control of the dynamical systems
\cite{Liu2019,Chen2018}.
This viewpoint suggests that there is a
connection between deep neural networks
and dynamical systems and indicates the possibility of using
dynamical systems as physical deep-learning machines.
In this paper, we reveal the potential of dynamical systems
with optimal control for the physical implementation.
We propose a deep neural network-like architecture
using dynamical systems with delayed
feedback and show that delayed feedback allows for
the virtual construction of a deep network structure
in a {\it physically single node} using a time-division multiplexing
method.
In the proposed approach, the virtual deep network for information propagation
comes from the time evolution of delay systems;
the systems are optimally controlled such that
information processing, including classification, is facilitated.
The significant difference between our deep network
and ordinary deep neural network architectures is
that the learning via optimal control is realized
by only a few control signals and minimal weight parameters,
whereas the learning by conventional deep neural networks requires
a large number of weight parameters \cite{Shen2017,Lin2018}.
The proposed approach using optimal control is applicable for
a wide variety of experimentally controllable systems;
it allows for simple but large-scale deep networks in physical systems
with a few control parameters.
\section{Multilayer neural networks and dynamical systems}
First, we briefly discuss the relationship
between multilayer neural networks and dynamical systems.
Let a dataset to be learned be composed of $K$ inputs,
${\bm x}_{k} \in {\mathbb R}^M$ and their corresponding target vectors, ${\bm t}_k\in{\mathbb R}^{L}$,
where $k \in \{1,2,\cdots,K\}$.
$M$ and $L$ are dimensions of the inputs and target vectors, respectively.
The goal of supervised learning
is to find a function that maps inputs
onto corresponding targets, $\mbox{\boldmath $G$}: {\bm x}_k \rightarrow {\bm t}_k$.
To this end, we consider an output function,
${\bm y} = \tilde{\mbox{\boldmath $G$}}({\bm x},{\bm w}) \in {\mathbb R}^L$,
parameterized by the $M_w$-dimensional vector, ${\bm w} \in {\mathbb R}^{M_w}$,
and the following loss function:
\begin{align}
J =
\sum_{k=1}^K
\Psi({\bm t}_k,{\bm y}_k), \label{cost-eq1}
\end{align}
where $\Psi({\bm t}_k,{\bm y}_k)$ is a function of the distance between the target
${\bm t}_k$ and the output, ${\bm y}_k = \tilde{\mbox{\boldmath $G$}}({\bm x}_k,{\bm w})$.
${\bm w}$ is determined such that loss function $J$ is minimized, i.e.,
output ${\bm y}_k$ corresponds to target ${\bm t}_k$.
It is well-known that a neural network model with an appropriate
activation function is a good candidate
for representing function $\mbox{\boldmath $G$}$ owing to its universal approximation
capability \cite{Cybenko1989,Funahashi1989,Sonoda2017}.
In multilayer neural networks,
the output, ${\bm y}_k = (y_{0,k},y_{1,k},\cdots,y_{L-1,k})^\mathrm{T}$,
is given by the layer-to-layer propagation of an input, ${\bm x}_k$
[Fig.~\ref{fig_ds}(a)].
The layer-to-layer propagation based on multilayer network structures
plays a crucial role in increasing expressivity \cite{Poole2016}
and enhancing learning performance.
In this study, instead of standard multilayer networks,
we utilize information propagation in
a continuous-time dynamical system,
\begin{align}
\dev{{\bm r}(t)}{t}{} = \mbox{\boldmath $F$}\left[{\bm r}(t),{\bm u}(t)
\right], \label{eq1}
\end{align}
where ${\bm r}(t) \in {\mathbb R}^M$ is the state vector at time $t$
and ${\bm u}(t) \in {\mathbb R}^{M_u}$ represents a control signal vector.
Based on the correspondence between a multilayer network [Fig.~\ref{fig_ds}(a)]
and a dynamical system [Fig.~\ref{fig_ds}(b)],
we suppose that
an input ${\bm x}_k$ is set as an initial state ${\bm r}(0)$.
Additionally, the corresponding output, ${\bm y}_k$,
is given by the time evolution
(feedforward propagation) of
the state vector, ${\bm r}_k(T) = {\bm r}(T,{\bm x}_k)$, up to the end time $t = T$,
i.e., ${\bm y}_k = {\bm y}[{\bm r}_k(T),\mbox{\boldmath $\omega$}]$, where
$\mbox{\boldmath $\omega$} \in {\mathbb R}^{L\times M}$ is a parameter matrix determined in
the training process.
Loss function $J$ is obtained by repeating the aforementioned feedforward
propagation for all training data instances and using their outputs.
The goal of the learning is to find
an optimal control vector, ${\bm u}^*(t)$, and a parameter vector,
$\mbox{\boldmath $\omega$}^*$, such that $J$ is minimized, i.e.,
${\bm w}^* = (\{{\bm u}^*(t)\}_{0 < t \le T}, \mbox{\boldmath $\omega$}^*) = \mbox{argmin}_{{\bm w}} J$.
It should be noted that learning using a discretized version of Eq. (\ref{eq1}) directly
corresponds to that using a residual network (ResNet) \cite{Chen2018,Benning2019}.
One strategy for finding optimal controls and parameters is to compute
the gradients of the loss function, i.e., the direction of the steepest
descent, and to update control and parameter vectors
in an iterative manner, i.e,
${\bm u}(t) \rightarrow {\bm u}(t) + \delta {\bm u}(t)$ and
$\mbox{\boldmath $\omega$} \rightarrow \mbox{\boldmath $\omega$} + \delta \mbox{\boldmath $\omega$}$, respectively.
In a simple gradient descent method,
$\delta {\bm u}(t)$ and $\delta \mbox{\boldmath $\omega$}$ are chosen such that
the largest local decrease of $J$ is obtained.
$\delta \mbox{\boldmath $\omega$}$ is simply selected
in the opposite direction of the gradient $\pdev{J}{\mbox{\boldmath $\omega$}}{}$,
e.g., $\delta \mbox{\boldmath $\omega$} = -\alpha_{\omega}\pdev{J}{\mbox{\boldmath $\omega$}}{}
=
-\alpha_{\omega}\sum_k(\pdel{\Psi}{{\bm y}_k}{}\pdel{{\bm y}_k}{\mbox{\boldmath $\omega$}}{})$,
where $\alpha_{\omega}$ is the learning rate,
which is usually a small positive number.
$\delta {\bm u}(t)$ is obtained using the adjoint method
developed in the context of optimal control problems \cite{Kirk2004,Sage1977}
as follows:
\begin{align}
\delta{\bm u}(t)
=
-\alpha_{u}
\sum_{k=1}^K
\left(
{\bm p}_{k}^\mathrm{T}(t)\del{\mbox{\boldmath $F$}_{k}}{{\bm u}}{}
\right)^\mathrm{T},
\label{eq1-j}
\end{align}
where $\alpha_u$ is a small positive number,
$\mbox{\boldmath $F$}_k = \mbox{\boldmath $F$}[{\bm r}_k(t),{\bm u}(t)]$ and ${\bm r}_k(t) = {\bm r}(t,{\bm x}_k)$.
${\bm p}_{k}(t) \in {\mathbb R}^M$ is the adjoint state vector that satisfies
the end condition at $t = T$,
${\bm p}_{k}(T) = \pdel{\Psi[{\bm t}_k,{\bm y}_k(\mbox{\boldmath $\omega$},{\bm r}_k)]}{{\bm r}_k}{}|_{t = T}$.
For $0 \le t < T$, the time evolution of ${\bm p}_{k}(t)$ is given
by
\begin{align}
\dev{{\bm p}_{k}^\mathrm{T}(t)}{t}{}
= - {\bm p}_{k}^\mathrm{T} (t)\del{\mbox{\boldmath $F$}_{k}}{{\bm r}_{k}}{}.
\label{eq1-p}
\end{align}
The derivation of Eqs. (\ref{eq1-j}) and
(\ref{eq1-p}) is shown in Appendix \ref{app1}.
We note that integrating Eq. (\ref{eq1-p}) in the backward direction (from $t=T$ to
$t=0$) corresponds to backpropagation in neural networks.
In summary, the algorithm for computing optimal ${\bm u}^*$ and $\mbox{\boldmath $\omega$}^*$ is
as follows:\\
\noindent (i) Set the input, ${\bm x}_k$, for the $k$th data instance as
an initial state, i.e., ${\bm r}(0) = {\bm x}_k$.\\
\noindent (ii) Forward propagation:
Starting from initial state ${\bm x}_k$, integrate Eq. (\ref{eq1})
and obtain the end state, ${\bm r}_k(T) = {\bm r}(T,{\bm x}_k)$.
Then, compute the output, ${\bm y}_k = {\bm y}[{\bm r}_k(T),\mbox{\boldmath $\omega$}]$. \\
\noindent (iii) Repeat the forward propagation for all data instances
and compute loss function $J$. \\
\noindent (iv) Backpropagation: Integrate adjoint Eq. (\ref{eq1-p})
for ${\bm p}_{k}(t)$ in the backward direction from $t = T$
with ${\bm p}_{k}(T) = \pdel{\Psi}{{\bm r}_{k}(T)}{}$. \\
\noindent (v) Compute $\delta\mbox{\boldmath $\omega$}$
and $\delta{\bm u}(t)|_{0 < t < T}$ using Eq. (\ref{eq1-j}) with
appropriate learning rates, $\alpha_{\omega}$ and $\alpha_{u}$.\\
\noindent (vi) Update control signal ${\bm u}(t)$ and parameter $\mbox{\boldmath $\omega$}$.
For the updates, one can use different optimization algorithms
\cite{Ruder2016}.
\begin{figure}[htbp]
\centering\includegraphics[width=7.5cm]{figure1.eps}
\caption{\label{fig_ds}
Schematics of (a) multilayer neural network and (b) dynamical system.
In (b), the output, ${\bm y}_k = (y_{0,k},y_{1,k},\cdots,y_{L-1,k})^\mathrm{T}$,
is given by the end state ${\bm r}(T)$ and weight parameter $\mbox{\boldmath $\omega$}$.
${\bm u}(t)$ and $\mbox{\boldmath $\omega$}$ can be updated
using a gradient-based optimization algorithm.
}
\end{figure}
\subsection{Binary classification problem \label{sec_ode}}
Here, we use an abstract dynamical system
for solving a typical fundamental problem, the binary classification problem.
The goal of the binary classification is to classify a given dataset
into two categories labeled as, for example, ``0'' or ``1''.
For this, we here consider a simple dynamical model,
$\dot{{\bm r}} = \tanh[{\bm a}(t){\bm r}+{\bm b}(t)]$, where
the state vector is two-dimensional, ${\bm r} =(\xi,\eta)^\mathrm{T}$.
Weight ${\bm a}(t) \in {\mathbb R}^{2\times 2}$ and
bias ${\bm b}(t) \in {\mathbb R}^2$ are used as control signals.
We apply this model to the binary classification problem for
a spiral dataset, $\{{\bm x}_k,c_k\}_{k=1}^K$, where ${\bm x}_k \in {\mathbb R}^2$
is distributed around
one of two spirals in the $\xi\eta$-plane, as shown in
Fig.~\ref{fig_ode1}(a), and ${\bm x}_k$
is labeled by $c_k$ as ``0'' or ``1'' according to the classes.
For the classification, we used one-hot encoding, i.e.,
target ${\bm t}_k$ corresponding to input ${\bm x}_k$ was set
as ${\bm t}_k = (t_{0,k},t_{1,k})^\mathrm{T} = (1,0)^\mathrm{T}$ if $c_k = 0$
and ${\bm t}_k = (0,1)^\mathrm{T}$ if $c_k = 1$.
For the output, ${\bm y}_k = (y_{0,k},y_{1,k})^\mathrm{T}$, the softmax function
was used: $y_{l,k}= \exp(z_{l,k})/\sum_l\exp(z_{l,k})$, where
$z_{l,k} = \mbox{\boldmath $\omega$}_l^\mathrm{T}{\bm r}_k(T) + \omega^{bias}_{l}$,
$\mbox{\boldmath $\omega$}_l \in {\mathbb R}^2$, and $\omega^{bias}_{l} \in {\mathbb R}$.
If $z_{0,k} \gg z_{1,k}$, ${\bm y}_k$ approaches $(1,0)^\mathrm{T}$,
whereas if $z_{0,k} \ll z_{1,k}$, ${\bm y}_k$ approaches $(0,1)^\mathrm{T}$.
$J$ was selected as a cross-entropy loss function,
$J = -1/K\sum_{k=1}^K \sum_{l=0}^1t_{l,k}\ln y_{l,k}$.
A training set of $K = 1000$ data points was used to
train ${\bm u}(t)=\{{\bm a}(t),{\bm b}(t)\}$ and
$\mbox{\boldmath $\omega$} = \{\mbox{\boldmath $\omega$}_l,\omega^{bias}_{l}\}_{l=0,1}$.
The classification accuracy was evaluated
using a test set of $1000$ data points.
For the gradient-based optimization, we used the Adam optimizer
\cite{Kingma2014} with a batch size of $K$.
Figures \ref{fig_ode1}(b) and \ref{fig_ode1}(c) show
the learning curve and classification accuracy
for the training and test datasets, respectively.
The loss function monotonically decreases, and the
classification accuracy approaches 100 $\%$.
For sufficient training over 300 training epochs,
the classification accuracy was over 99 $\%$
when end time $T$ is set as $200\Delta t$,
where $\Delta t \approx 0.01$ is the time step used in the simulation.
Figure \ref{fig_ode2} shows the time evolution of the two distributions
constituting the spiral dataset.
During the evolution,
the distributions of the initial states are disentangled
[Figs.~\ref{fig_ode2}(a-d)] and become
linearly separable at end time $T$ [Fig.~\ref{fig_ode2}(d)]
to aid the classification at the softmax output layer.
As a result, any input state can be classified into either of
two classes [Fig.~\ref{fig_ode2}(e)].
We note that
the classification based on disentanglement
is different from
that of other schemes utilizing dynamical systems, e.g.,
reservoir computing,
whose classification is based on the mapping of
input information onto a high-dimensional feature space \cite{Tanaka2019}.
In addition, we note that the disentanglement is facilitated as end time
$T$ increases, i.e., the number of layers increases,
and high classification accuracy is achieved for $T \le 400\Delta t$,
as shown in Fig. \ref{fig_ode2}(f).
A slight decrease in classification accuracy at $T = 600 \Delta t$
is attributed to slowdown of the training due to
a local plateau of loss function $J$,
which is occasionally caused in a non-convex optimization problem
\cite{Dauphin2014}.
At $T =600\Delta t$,
we confirmed that
a classification accuracy of over 99 $\%$ was obtained
when the number of training epochs is extended to 400.
\begin{figure}[htbp]
\centering\includegraphics[width=8.8cm]{figure2.eps}
\caption{\label{fig_ode1}
(a) Spiral dataset for binary classification.
The dataset consists of the two data groups labeled as ``0'' or ``1'',
colored purple or green, respectively.
(b) Loss function $J$ and (c) classification accuracy as a function of
the training (test) epoch.
}
\end{figure}
\begin{figure}[htbp]
\centering\includegraphics[width=8.8cm]{figure3.eps}
\caption{\label{fig_ode2}
(a--d) Configuration of the states constituting the two spirals
over the time evolution in the trained system
up to the end time $T=200\Delta t$.
$t/\Delta t$ effectively represents the number of layers from the viewpoint of
a neural network.
The initial spiral distribution is disentangled according to
the time evolution (layer-to-layer propagation) and
becomes linearly separable at end time $t = T$.
In (d), the dotted line represents the decision boundary for separating
the two distributions.
(e)
Result of binary classification.
The inputs can be classified into two regions
indicated by pink and blue colors.
(f)
Classification accuracy for a test dataset
as a function of end time $T$.
}
\end{figure}
\section{Physical implementation in delay systems}
In the previous section, we showed binary classification
based on optimal control in a two-dimensional dynamical model,
where weight ${\bm a}(t) \in {\mathbb R}^{2\times 2}$ and bias ${\bm b}(t) \in {\mathbb R}^2$
were used as control signals.
The excellent performance of this demonstration suggests
the realization of deep learning in various physical systems,
such as coupled oscillators, fluids, and elastic bodies.
However, for processing high-dimensional data,
a number of signals must be used to control
the high-dimensional degrees of freedom in the systems;
this may be difficult
in terms of physical implementation in an actual system.
To overcome the difficulty, we propose the use of delay systems
to achieve feasible optimal control of numerous degrees of freedom
with limited control signals.
It is known that delay systems can be regarded as
infinite-dimensional dynamical systems, as visualized
in a time--space representation \cite{Arecchi1992}.
Furthermore, delay systems can support numerous virtual neurons
using a time-division multiplexing method \cite{Appeltant2011}.
In addition, they can exhibit
various dynamical phenomena, including stable motion,
periodic motion, and high dimensional chaos, with experimentally controllable
parameters, e.g., delay time and feedback strength
\cite{Uchida,Soriano2013}.
Thus, their high expressivity as well as controllability
are promising.
\subsection{Learning by optimal control in a delay system \label{sec_delay}}
We introduce a training method based on
the optimal control of a delay system,
the time evolution of which
is governed by the following equation:
\begin{align}
\dev{{\bm r}(t)}{t}{}
= \mbox{\boldmath $F$}\left[{\bm r}(t),{\bm r}({t-\tau}),{\bm u}(t)
\right], \label{delay-eq}
\end{align}
where ${\bm r}(t) \in {\mathbb R}^{M_r}$ and ${\bm u}(t) \in {\mathbb R}^{M_u}$
represent the state vector and control signal
vector at time $t$, respectively, and $\tau$ is the delay time.
The aforementioned equation can be integrated by setting
${\bm r}(t)$ for $-\tau \le t \le 0$ as an initial condition.
The information dynamics can intuitively be interpreted by
a space--time representation \cite{Arecchi1992}
based on the time discretization of Eq. (\ref{delay-eq}),
${\bm r}_{n}^{j+1} = {\bm r}_n^j +\Delta t\mbox{\boldmath $F$}({\bm r}_n^j,{\bm r}_{n-1}^j,{\bm u}_n^j)$,
where $t = n\tau + j\Delta t$, ${\bm r}_n^j = {\bm r}(n\tau + j\Delta t)$,
$n \in \{-1,0,1,\cdots, N-1\}$, $j \in \{0,1,\cdots, M_{\tau}-1\}$,
and $M_{\tau} = \tau/\Delta t$.
In this representation, ${\bm r}_n^j$ can be regarded as
the $j$th network node in the $n$th layer,
which is affected by an adjacent node, ${\bm r}_n^{j-1}$,
and node ${\bm r}_{n-1}^j$ in the $(n-1)$th layer, as shown in
Fig. \ref{fig_dn}.
Feedforward propagation is carried out as follows:
First, an input, ${\bm x}_{k} = (x_{1,k},\cdots,x_{M,k})^\mathrm{T}$,
is encoded as ${\bm r}_{-1}^{j} = {\bm r}_{-1}^j({\bm x}_{k})$
for $j \in \{0,1,\cdots, M_{\tau}\}$
in the initial condition.
Then, Eq. (\ref{delay-eq}) is numerically solved to obtain
${\bm r}_{N-1}^j$ in the $(N-1)$th layer (corresponding to
$\{{\bm r}_k(t)\}_{T-\tau \le t < T}$ in continuous time).
The output, ${\bm y}_k \in {\mathbb R}^L$, is computed using the nodes
in the $(N-1)$th layer, $\{{\bm r}_{N-1}^j\}_{j=0}^{M_\tau-1}$,
as shown in Fig. \ref{fig_dn}.
In the continuous time representation, the output is defined as
${\bm y}_k =
{\bm y}\left(
{\bm z}_{k}
\right)
$,
where
${\bm z}_{k} = \int_{T-\tau}^T\mbox{\boldmath $\omega$}(t)
{\bm r}_{k}(t)dt + {\bm b}$.
$\mbox{\boldmath $\omega$}(t) \in {\mathbb R}^{L\times M_{r}}$ and ${\bm b} \in {\mathbb R}^L$
are the weight and bias parameters to be trained, respectively,
and ${\bm r}_k(t)$
is the state vector starting from the initial state $\{{\bm r}(t,{\bm x}_k)\}_{-\tau \le t \le 0}$.
Finally, loss function $J$ [Eq. (\ref{cost-eq1})]
is computed.
To minimize loss function $J$, a gradient-based optimization is used,
where ${\bm u}(t)$, $\mbox{\boldmath $\omega$}(t)$, ${\bm b}$ are updated in an iterative manner.
In a gradient descent method, the
update variations, $\delta{\bm u}(t)$, $\delta\mbox{\boldmath $\omega$}(t)$, and $\delta{\bm b}$,
can be chosen as follows:
\begin{align}
\delta{\bm u}(t)
= -\alpha_{u}\sum_{k=1}^K
\left(
{\bm p}_{k}^\mathrm{T}(t)\del{\mbox{\boldmath $F$}_{k}}{{\bm u}}{}
\right)^T, \label{update-delayeq1}
\end{align}
\begin{align}
\delta\mbox{\boldmath $\omega$}
= -\alpha_{\omega}
\sum_{k=1}^K
\left(
{\bm r}_k
\del{\Psi}{{\bm z}_{k}}{}
\right)^\mathrm{T},
\hspace{1mm}
\delta{\bm b}_l
= -\alpha_{b}
\sum_{k=1}^K
\left(
\del{\Psi}{{\bm z}_{k}}{}
\right)^\mathrm{T}.
\end{align}
The details of these derivations are provided in Appendix~\ref{app2}.
In the aforementioned equations,
$\mbox{\boldmath $F$}_{k} = \mbox{\boldmath $F$}[{\bm r}_{k}(t),{\bm r}_{k}(t-\tau), {\bm u}(t)]$,
$\Psi = \Psi({\bm t}_k,{\bm y}({\bm z}_k))$ is a function of ${\bm z}_k$,
and
$\alpha_{i}$ for $i \in \{u,\omega,b\}$ is the learning rate
which is a small positive number.
In Eq. (\ref{update-delayeq1}),
${\bm p}_{k}(t)$ is the adjoint state vector,
which satisfies ${\bm p}_k(T)=0$.
${\bm p}_k(t)$ can be obtained by solving the
following adjoint equations in
the backward direction,
\begin{align}
\del{{\bm p}_{k}^\mathrm{T} (t)}{t}{}
=
-\del{\Psi}{{\bm z}_{k}}{}\mbox{\boldmath $\omega$}(t)
-
{\bm p}_{k}^{\mathrm{T}}(t)
\dfrac{\partial \mbox{\boldmath $F$}_{k}}{\partial {\bm r}_{k}}, \label{delay-eqp1}
\end{align}
for $T-\tau \le t < T$, and
\begin{align}
\dev{{\bm p}_{k}^\mathrm{T} (t)}{t}{}
=
-{\bm p}^\mathrm{T} (t)
\dfrac{\partial \mbox{\boldmath $F$}_{k}}{\partial {\bm r}_{k}}
-{\bm p}_{k}^\mathrm{T} (t+\tau)
\dfrac{\partial \mbox{\boldmath $F$}_{k}(t+\tau)}{\partial {\bm r}_{k}}, \label{delay-eqp2}
\end{align}
for $0 \le t < T-\tau$.
\begin{figure}[htbp]
\centering\includegraphics[width=8cm]{figure4.eps}
\caption{\label{fig_dn}
Schematic of virtual network of a delay system.}
\end{figure}
\subsection{Optoelectronic delay system}
As an effective and feasible example,
we consider the use of an optoelectronic delay system,
as shown in Fig.~\ref{fig_delaysys}.
The delay system is
composed of a laser, optoelectronic intensity modulator, photodetector,
and electrical filter to construct a time-delay feedback loop.
The time evolution of the system state,
${\bm r}(t) = (\xi(t),\eta(t))^\mathrm{T}$, is given
by the following equations \cite{Murphy2010}:
\begin{align}
\tau_L\dfrac{d\xi}{dt} =
-\left(
1 + \dfrac{\tau_L}{\tau_H}
\right)\xi - \eta
+ \beta \cos^2
\left[
u_1(t)\xi(t-\tau)
+u_2(t)
\right], \label{eq-optdelay1}
\end{align}
\begin{align}
\tau_H\dfrac{d\eta}{dt} = \xi, \label{eq-optdelay2}
\end{align}
where
$\xi(t)$ is the normalized voltage, and
$\tau_H$ and $\tau_L$ are the time constants of the low-pass and
high-pass filters, respectively.
$\beta$ represents the feedback strength.
$u_1(t)$ and $u_2(t)$ are electronic signals added to the feedback loop,
which are used as control signals in the system.
\begin{figure}[htbp]
\centering\includegraphics[width=8cm]{figure5.eps}
\caption{\label{fig_delaysys}
Schematic of optoelectronic delay system. MZM, Mach--Zehnder modulator;
Delay line, optical fiber delay line; PD, photodetector; Amp, electric
amplifier; Filter, a two-pole band-pass filter (consisting of low-pass and high-pass filters).
}
\end{figure}
\subsection{Results}
\subsubsection{Binary classification}
We demonstrate binary classification for a spiral dataset, as shown in
Fig.~\ref{fig_ode1}(a),
with the aforementioned optoelectronic delay system.
The goal of binary classification is to classify two categories labeled
as ``0'' or ``1'' for the spiral dataset, $\{{\bm x}_k, c_k\}_{k=1}^K$.
In the same manner as that described in Sec. \ref{sec_ode},
target ${\bm t}_k$ is set
as ${\bm t}_k = (1,0)^\mathrm{T}$ if $c_k = 0$,
whereas ${\bm t}_k = (0,1)^\mathrm{T}$ if $c_k =1$.
Output $y_{l,k}$ is set as the softmax function, i.e.,
$y_{l,k} = \exp{z_{l,k}}/\sum_{l^{'}}\exp{z_{l^{'},k}}$,
where
$z_{l,k} = \int_{T-\tau}^T \omega_l(t)\xi_k(t) dt + b_l$,
$\omega_l(t) \in {\mathbb R}$ and $b_l \in {\mathbb R}$.
Then, $J$ is defined as
the cross-entropy loss function, $-1/K\sum_{k=1}^K\sum_{l=0}^{L-1}t_{l,k}\log{y_{l,k}}$.
In this simulation, the following parameters settings were applied:
$\tau_H = 1.59$ ms, $\tau_L = 15.9$ $\mu$s, and $\tau = 230$ $\mu$s.
The input, ${\bm x}_k = (x_{1,k},x_{2,k})^\mathrm{T}$,
is encoded as the initial state of $\xi$, i.e.,
$\xi_k(t) = x_{1,k}$ for $-\tau \le t < -\tau/2$ and
$\xi_k(t) = x_{2,k}$ for $-\tau/2 \le t \le 0$.
We set $u_1(t) = 1.0$ and $u_2(t) = -\pi/4$ as the initial control signals
and $\omega_l(t) = 0$ and $b_l = 0$ ($l \in \{0, 1\}$)
as the initial weight and bias parameters.
We used the Adam optimizer \cite{Kingma2014} with a batch size of $K$
for the gradient-based optimization.
The update equations for $u_1(t)$, $u_2(t)$, $\omega_l$,
and $b_l$ are shown in Appendix~\ref{sec_app2-2}.
In the aforementioned conditions,
the classification accuracy at training epoch 100 was 99.1$\%$
when the feedback strength was $\beta = 3.0$ and the end time was
$T = 5\tau$.
To gain insight into the classification mechanism, we
investigated the effect of the control signals, $u_1(t)$ and $u_2(t)$,
and weights, $\omega_l(t)$, $l \in \{0,1\}$, on the delay dynamics.
Figures~\ref{fig_dbr1}(a--e) show the trained control signals, $u_1(t)$ and $u_2(t)$,
weights, $\omega_0(t)$ and $\omega_1(t)$,
and four instances of $\xi_k(t)$ at training epoch 100.
$\xi_k(t)$ in a range of $T-\tau \le t \le T$
is used for computing the
(softmax) outputs, $y_{l,k}$, which
represents the probability that the input ${\bm x}_k$ is
classified as class $l \in \{0,1\}$.
Considering that $y_{l,k}$ is a function of
$z_{l,k} = \int_{T-\tau}^T\omega_l\xi_k(t)dt + b_l$,
we computed $\tilde{z}_{l,k_{l'}}
= \int_{T-\tau}^T\omega_l(t)\xi_{k_{l'}}(t) dt$.
$\tilde{z}_{l,k_{l'}}$ corresponds to the correlation between
(softmax) weights $\omega_l(t)$ used for the classification
as class $l$ and the $k_{l'}$th instance, $\xi_{k_{l'}}(t)$, which starts
from the initial states labeled as $l' \in \{0,1\}$.
Figures~\ref{fig_dbr1}(f) and \ref{fig_dbr1}(g) show
the histograms of the correlation values, $\tilde{z}_{l,k_{l'}}$,
for 500 instances.
The correlation values are positive for $l = l'$ in most cases,
whereas they are negative for $l \ne l'$.
In other words, $\tilde{z}_{l,k_{l}} > \tilde{z}_{l,k_{l'}} (l \ne l')$
in most cases.
Thus, the softmax output $y_{l,k_{l}}$ for classification as $l$ is
activated by $\xi_{k_l}(t)$ starting from
the initial states with the same class $l$.
These results reveal that
the trained control signals control
each trajectory such that
it is positively correlated to the weight $\omega_l(t)$
and $y_{l,k_l}$ is maximized.
Figure \ref{fig_dbr2} shows the classification accuracy at training epoch 100
as a function of feedback strength $\beta$ and end time $T/\tau$.
As seen in this figure,
classification performance is low for $\beta < 2.0$.
In this regime, the system exhibits transient behavior
to stable limit cycle motion, which
is insensitive to external perturbations
[Figs.~\ref{fig_dbr3}(a) and \ref{fig_dbr3}(b)].
This means that it is difficult to control the system.
When $\beta$ increases ($\beta > 2.5$),
the system starts to exhibit complex behavior and
becomes sensitive to the control signals for a large end time, $T$,
as shown in Figs. \ref{fig_dbr3}(c) - \ref{fig_dbr3}(f).
Sensitivity plays a role in aiding the classification.
However,
extremely high sensitivity makes it difficult to control the system
and decreases classification performance,
as observed for $\beta > 4.0$ and $T > 7.0\tau$ in Fig.~\ref{fig_dbr2}.
Consequently, a high classification performance of over 99$\%$ is achieved
with moderate values of $T$ and $\beta$, suggesting
that the transient behavior around the edge of chaos plays a crucial role in
classification.
\begin{figure}[htbp]
\centering\includegraphics[width=8.8cm]{figure6.eps}
\caption{\label{fig_dbr1}
(a,b) Control signals, $u_1(t)$ and $u_2(t)$
at training epoch 100.
(c,d) Weight parameters, $\omega_0(t)$ and $\omega_1(t)$,
at training epoch 100.
(e) Four instances of $\xi_k(t)$ starting from different initial states
labeled as ``0'' or ``1'', which are
within a distance $|{\bm x}_k-{\bm x}_{k^{'}}| < 0.1$.
The end time is set as $T = 5\tau$.
$\xi_k(t)$ for $4\tau \le t < 5\tau$ (indicated by light yellow color)
is used to obtain $z_{l,k} = \int_{T-\tau}^T\omega_l(t)\xi_k(t)dt + b_l$.
(f,g) Histograms of correlation values, $\tilde{z}_{l,k_{l'}}$,
between softmax weights, $\omega_l(t)$, and the $k_{l'}$-th
instance, $\xi_{k_{l'}}$.
}
\end{figure}
\begin{figure}[htbp]
\centering\includegraphics[width=8.8cm]{figure7.eps}
\caption{\label{fig_dbr2}
Classification accuracy as a function of feedback strength $\beta$
and end time $T$.
}
\end{figure}
\begin{figure}[htbp]
\centering\includegraphics[width=8.8cm]{figure8.eps}
\caption{\label{fig_dbr3}
Typical examples of $\xi_k(t)$ starting from an initial state
at training epochs 0 and 100 for various values of $\beta$ and $T$.
}
\end{figure}
\subsubsection{MNIST handwritten digit classification}
To investigate the classification performance for a higher dimensional dataset,
we use the MNIST handwritten digit dataset, commonly
used as a standard benchmark for learning \cite{MNIST,LeCun1998}.
The dataset has a training set of 60,000 28$\times$28 pixel grayscale images of ten handwritten digits, along with a test set of 10,000 images.
To set an initial state of the system state,
an input image of $28\times 28$ pixels is enlarged to double its length
and width; it is transformed to a $m$-dimensional vector, where
$m = (28\times 2)^2$.
The vector components are sequentially set
as $\xi_k(t_j)$ at time $t_j = -\tau + j\Delta t$
with a time interval of $\Delta t = \tau/M_{\tau} \approx 0.07$ $\mu$s.
The input process is repeated $M_{\tau}/m$-times
to encode the information of the input image as $\xi_k(t_j)$
for all $j \in \{0,1,\cdots,M_{\tau}\}$.
The training of the delay system is based on
the gradient-based optimization using
the Adam optimizer with a batch size of 100.
The maximum number of epochs to train was set as 50 to ensure the convergence
of the training process.
As a demonstration of the classifications at epoch 50, we show four examples of the softmax outputs, which represents the probability that the input image belongs to one of the 10-classes, in Figs. \ref{fig_dmr}(a)--(d).
Figures \ref{fig_dmr}(e) and \ref{fig_dmr}(f) show
the classification accuracy for various values of
feedback strength $\beta$ and delay time $\tau$, where $T/\tau = 3$
is fixed.
For this image dataset, the delay system with $\beta = 3.0$
exhibits relatively
high classification performance.
The best classification accuracy for this system is 97 $\%$.
We emphasize that accurate classification is achieved
with two training signals, $u_1(t), u_2(t)$, and minimal weight parameters,
$\omega_l(t)$, and $b_l$, $l \in \{0,1,\cdots,9\}$, owing to the
time-division multiplexing encoding method based on the delay structure,
as discussed in Sec. \ref{sec_delay}.
This is in contrast to conventional neural networks,
where more than hundreds or thousands of weight parameters need to be trained.
We can see that classification accuracy improves
as delay time $\tau$ increases [Fig.\ref{fig_dmr}(f)].
As discussed in Sec. \ref{sec_delay},
the effective number of the network nodes, $M_{\tau}$,
depends on $\tau$ in the delay system, i.e., $M_{\tau} \approx \tau/\Delta t$.
This suggests that
larger-scale networks, i.e., systems with a longer delay,
play an important role in achieving
better classification for this large dataset.
\begin{figure}[htbp]
\centering\includegraphics[width=8.8cm]{figure9.eps}
\caption{\label{fig_dmr}
(a-d) Examples of softmax outputs, $\{y_{l,k}\}_{l=0}^9$,
for the MNIST handwritten images.
Each inset shows the input handwritten image corresponding to
(a) ``$2$'', (b) ``$3$'', (c) ``$7$'', and (d) ``$9$''.
The softmax output, $y_{l,k}$, represents the probability that
the input image $k$ belongs to class $l$.
In (a--d), $\beta = 3.0$, $\tau = 1610$ $\mu$s, and $T = 3\tau$.
(e--f) Classification accuracy for the MNIST dataset (10,000 test images)
as a function of
(e) feedback strength $\beta$ and (f) delay time $\tau$
in the delay system.
In (e), $\tau = 3220$ $\mu$s and $T = 3\tau$.
In (f), $\beta = 3.0$ and $T = 3\tau$.
}
\end{figure}
\section{Conclusion}
We discussed the applicability of optimally-controlled dynamical systems
to information processing.
The dynamics-based processing provides insight into
the mechanism of information processing based on deep network structures,
and it can be easily implemented in physical systems.
As a particular example, we introduced an optoelectronic delay system.
The delay system can be trained to perform nonlinear classification and image recognition with only a few control signals and classification weights
based on the time-division multiplexing method.
This feature of delay systems is an advantage to
hardware implementation of the systems and
is distinctively different from
conventional neural networks, which require a large number of
training parameters.
The dynamics-based processing based on optimal control
can be applied to various physical systems to construct not only
feedforward networks but also recurrent neural networks, including
reservoir computing.
This provides a novel direction for physics-based computing.
\begin{acknowledgements}
This work was supported, in part, by JSPS KAKENHI (Grant
No. 20H042655) and JST PRESTO (Grant No. JPMJPR19M4).
The authors thank Profs. Kazutaka Kanno and Atsushi Uchida for
valuable discussions on optoelectronic delay systems.
\end{acknowledgements}
| 3483a453f8db98a9ddda233fc7c21573d70390ec | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Considerable efforts are being made to detect the Warm-Hot Intergalactic Medium (WHIM) in various bands of the electromagnetic spectrum. Detecting radio synchrotron emission from intercluster filaments or bridges could provide a new probe of the WHIM and could also shed light on mechanisms of particle acceleration and magnetic fields in a poorly studied environment \citep{2019A&A...627A...5V}.
The magnetic field in the WHIM is less likely to be affected by a small-scale dynamo or by outflows from galaxies and active galactic nuclei (AGN). This field could therefore be a tracer of magnetic fields from an earlier epoch in the Universe that has subsequently been merely compressed by structure formation processes \citep[e.g.][]{2005ApJ...631L..21B, 2017CQGra..34w4001V}.
Moreover, interacting galaxy clusters produce diffuse radio sources by accelerating electrons via shocks or turbulence \citep[e.g.][]{2014IJMPD..2330007B}. Giant radio halos are diffuse radio synchrotron sources found in galaxy clusters that have recently suffered a merger, as indicated by a disturbed intracluster medium (ICM) or other indicators of the cluster’s dynamical state \citep[e.g.][]{2001A&A...378..408S, 2010A&A...517A..10C}. Cluster radio shocks or radio relics are arc-shaped, diffuse sources that appear to be related to shock waves in the ICM \citep{2020MNRAS.493.2306B}. Moreover, there is a class of sources that trace old radio plasma from AGN that has been re-energised through processes in the ICM. Low-frequency radio observations are starting to show more of these types of sources. The common properties of these sources are the AGN origin of the plasma and their ultra-steep radio spectra. Radio phoenices and gently reenergised tails (GReETs) are examples of such sources (see e.g. \cite{2019SSRv..215...16V} for a recent review).\\
The cluster Abell 3395 (hereafter A3395) is double-peaked, both in its galaxy distribution \citep{1997ApJ...482...41G} as well as in its X-ray emission, first observed with the Einstein satellite \citep{1981ApJ...243L.133F}. It is accompanied to the north by A3391 at a separation of about 3 Mpc. The mean redshift of A3395 ($z=0.0518$) and A3391 ($z=0.0555$) is $\langle z\rangle =0.053$. In between the clusters lies a galaxy group called ESO 161-IG 006 ($z=0.0520$, Alvarez et al. 2018). Both clusters have masses of around $M_{200}\sim 2\times 10^{14} M_\odot$ \citep{2011A&A...534A.109P}, and X-ray temperatures of around $kT\sim5$\,keV \citep{2009ApJ...692.1033V}. ASCA, ROSAT, Planck, and Suzaku observations have confirmed that A3395 and A3391 are connected by a gas bridge \citep{2001ApJ...563..673T, 2013A&A...550A.134P, 2017PASJ...69...93S}.
Diffuse radio emission from pairs of galaxy clusters is rare \citep{2018MNRAS.478..885B, 2019A&A...630A..77B, 2019Sci...364..981G, 2020MNRAS.tmpL.159B}.
The first known bridge of low surface brightness connecting a radio halo and radio relic was found in the Coma cluster, first reported by \citet{1989Natur.341..720K} and further studied afterwards (e.g. \citealt{1990AJ.....99.1381V}, \citealt{2011MNRAS.412....2B}). A moderate correlation between radio and X-ray brightness has also been found in the Coma bridge. This correlation suggests that the radio and X-ray emission originate from the same volume and that the correlation is not produced by projection effects.
The bridges detected so far have different sizes and are observed in systems that are believed to be in a state preceding the merger. Currently, the best examples are the merging galaxy clusters Abell 399 and Abell 401, which are separated by about 5 Mpc. Diffuse synchrotron emission was detected in the region connecting the two clusters using the Low Frequency Array (LOFAR) at $140$
MHz \citep{2019Sci...364..981G}. X-ray observations of A399-A401 have revealed a hot (6 - 7 keV) filament of plasma in the region between the two clusters \citep[e.g.][]{2008PASJ...60S.343F}. Their masses are $5.7\times 10^{14}M_\odot$ and $9.3\times 10^{14}M_\odot$, respectively, and so the system is more than three times as massive as the A3391/5 system. The redshift of Abell 399 is 0.0718, while that of Abell 401 is 0.0736.
Here, both clusters host a radio halo. The presence of a bridge was
confirmed through the Sunyaev-Zeldovich effect \citep{2013A&A...550A.134P} but it is not seen in the GaLactic and Extragalactic All-Sky MWA (GLEAM) low-frequency radio survey \citep{ 2015PASA...32...25W}.
Integrating the average surface brightness over the Abell 399/401 bridge region, namely $\langle I \rangle_{140\,{\rm MHz}} = 2.75 \pm 0.08$ mJy beam$^{-1}$ (or 0.38 $\upmu$Jy arcsec$^{-2}$), over an area of $3 \times 1.3$ Mpc$^2$ and excluding the emission from the two halos, \cite{2019Sci...364..981G} obtain a total flux density of $S_{140\,{\rm MHz}} = 822 \pm 24$ mJy.
This flux density corresponds to a radio power of $L_{140\,{\rm MHz}} = 1.0 \times 10^{25}$ W Hz$^{-1}$ and a mean radio emissivity of $\langle J \rangle_{140\,{\rm MHz}} = 8.6 \times 10^{-44}$ W Hz$^{-1}$ m$^{-3}$.
\cite{2020MNRAS.tmpL.159B} discovered a radio bridge in the pre-merging galaxy clusters A1758N and A1758S, which are 2 Mpc apart. The bridge is clearly detected in the LOFAR image at 144 MHz and tentatively detected at 53 MHz.
The clusters A1758N and A1758S also host radio halos but the mean radio emissivity in the bridge is more than one order of magnitude lower than that of the halos. Interestingly, radio and X-ray surface brightness
are found to be correlated in the bridge.
The short lifetime for relativistic electrons emitting at 140 MHz limits the maximum distance that these relativistic electrons can travel in their lifetime to $<$0.1 Mpc \citep{2019Sci...364..981G}. Hence, the sizes of radio bridges combined with the short cooling time of the synchrotron-emitting electrons suggest that relativistic particles are accelerated in situ. The large area filling factor, that is, a very smooth radio surface brightness, observed in the radio bridge of A399-A401 clearly disfavours shock acceleration as the main source of the observed radio emission.
The origin of radio synchrotron emission in radio bridges is not understood. Using cosmological simulations, \cite{2019Sci...364..981G} proposed that the synchrotron emission could come from electrons re-accelerated by weak shock waves provided there exists a volume-filling population
of pre-accelerated electrons in the bridge. However, this would not produce a radio-X correlation because shocks cover at most 10 percent of the volume of the bridge. Also, \cite{2019Sci...364..981G} concluded that the steep radio spectrum challenges this scenario for the origin of the radio synchrotron bridge.
Recently, \citet{2020PhRvL.124e1101B} proposed that stochastic acceleration of relativistic electrons by turbulence could explain the radio emission observed in the system A399-A401. According to this model, the emission is expected to be volume-filling and with a steep spectrum\footnote{We define the spectral index $\alpha$ as $S \propto \nu^\alpha$.} ($\alpha <$ -1.3). Their model also predicts a natural correlation between radio and X-ray surface brightness. Moreover, it is expected that the radio emission gets clumpier at higher frequencies.
\begin{figure*}
\center{\includegraphics[width=1.9\columnwidth]{A3391_95_eRosita_EMU_DDF3.pdf}}
\caption{ASKAP/EMU radio and eROSITA X-ray overlay of the A3391-A3395 system. The eROSITA adaptively smoothed image covers the energy range from
0.3 to 2.0 keV. The inlays show notable individual radio galaxies, some of which show signs of interactions with the ambient medium. All sources highlighted and magnified in this image fall within the mean redshift of the cluster system except for F3 which is at $z=0.193$. Details of these sources are given in Tables~\ref{tab:RGs_EMU} and \ref{tab:flux}.}
\label{fig:overlay}
\end{figure*}
In this paper, we present the results of our search for diffuse radio synchrotron emission in the bridge of thermal gas between A3395 and A3391 that was recently observed with eROSITA. Moreover, we present an inventory of interesting radio sources in this field around this cluster merger. The outline of this paper is as follows: In Sect.~\ref{sec:xray} we give a short overview of the eROSITA observations that are published in an accompanying paper (Reiprich et al. 2020). In Sect.~\ref{sec:radio} we describe ASKAP and MWA observations. Results are presented in Sect.~\ref{sec:results} and we present our conclusions in Sect.~\ref{sec:conclusions}.
Throughout this paper, we use the fiducial cosmology $H_0 = 70$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_M = 0.3$, and $\Omega_\Lambda = 0.7$. At the mean redshift of A3395 and A3391, 1 arcsec on the sky corresponds to $\sim 1.04$ kpc.
\section{X-ray observations}
\label{sec:xray}
On July 13, 2019, {eROSITA} (extended ROentgen Survey with an Imaging Telescope Array) was successfully launched and made its way into an L2 halo orbit \citep{Predehl2020}. During its Performance Verification (PV) phase, {eROSITA} surveyed the A3391-A3395 system as part of its first light observations. Reiprich et al. (2020) present the first results from four eROSITA PV phase observations of A3395-A3391. These cover a total area of about 15 deg$^2$, or about half the size of the accompanying Australian Square Kilometre Array Pathfinder \citep[ASKAP:][]{askap} field, with at least 30 s exposure ($\sim$10 deg$^2$ with at least 1000 s). The resulting X-ray image is shallower than the XMM-Newton observation of the A3395-A3391 system but covers a much wider area.
Also, owing to its better response to soft X-rays, in between A3395 and A3391, Reiprich et al. (2020) observe warm gas in emission in the bridge, which may constitute a WHIM detection. Emission along this bridge includes the galaxy group ESO 161-IG006, which produces only a small fraction of the total emission. While most gas in the bridge appears hot, these latter authors discover hints for cool primordial gas between the clusters. Moreover, several clumps of matter around the merging system were discovered. This is enabled by the X-ray hot gas morphologies and radio jet and lobe structures of their central AGN. Finally, Reiprich et al. (2020) discover a filamentary emission region north of and well beyond the virial radius of A3391 connecting to an apparently infalling cluster and a connected emission filament towards the south of A3395, suggesting a total length of the filament of $\sim$15 Mpc ($\sim$4 degrees). The Planck SZ Y-maps, as well as galaxy density maps from DECam observations \citep[DECam; see, e.g.][]{Flaugher_2015}, appear to confirm this new filament. For more information we refer the reader to Reiprich et al. (2020).
\section{Radio observations}
\label{sec:radio}
\subsection{ASKAP observations}
The A3391-A3395 system was observed as part of the Evolutionary Map of the Universe \citep[EMU:][]{emu} survey Early Science observations which use the new ASKAP telescope in Australia to make a census of radio sources in the sky south of +30$^\circ$ declination; see Fig.~\ref{fig:overlay} for an X-ray radio overlay. Our ASKAP observations reach a depth of 25--35 $\mu$Jy/beam root mean square (rms) at a spatial resolution of $\sim$10 arcsec in the frequency range 846.5--1134.5 MHz. These latter are therefore 15 times deeper than the NRAO VLA Sky Survey \citep[NVSS:][]{NVSS}, and have a five times better angular resolution (we note that the A3391-A3395 field is not covered by NVSS).
The ASKAP early science data covering A3395-A3391 are from Scheduling Block (SB) 8275 observed on March 22, 2019, and are publicly available at the \href{https://data.csiro.au/collections/#domain/casdaObservation/search/}{CSIRO ASKAP Science Data Archive} (CASDA; \citealp{2017ASPC..512...73C}). SB8275 contains 36 separate measurement sets corresponding to 36 separate beams, each with a unique phase-tracking centre. Each beam covers $\sim$1 deg$^{2}$ on the sky and all beams were observed simultaneously, giving an instantaneous field of view (FOV; after accounting for beam overlap) of $\sim$30 deg$^{2}$. Of the 36 ASKAP antennas, 35 were used for the observations, with a minimum baseline of $\sim$ 22 m and maximum baseline of $\sim$ 6.4 km. Each measurement set covers a bandwidth of 288~MHz with a central frequency of 990.5~MHz. These data were processed through the official ASKAP processing pipeline (ASKAPsoft; \citealp{2019ascl.soft12003G}). The steps implemented in the ASKAPsoft pipeline will be described in an upcoming system description paper, but we refer the reader to the \href{https://www.atnf.csiro.au/computing/software/askapsoft/sdp/docs/current/pipelines/introduction.html}{CSIRO ASKAPsoft documentation} for further details.
ASKAP forms beams electronically through a chequerboard Phased Array Feed at the prime focus of each antenna. Individual beams are then correlated between antennas. The ASKAP correlator generates 15552 spectral-line channels, which are then averaged to 288 1-MHz channels.
Beam measurement sets are bandpass calibrated using a standard calibrator (PKS B1934--638)\footnote{Reynolds J. E., 1994, Technical report, A Revised Flux Scale for the AT Compact Array. ATNF, Epping}. The calibration observations occurred in a preceding scheduling block wherein PKS B1934--638 was placed at the centre of each beam for 200 seconds. The derived bandpass solutions also account for the other fundamental interferometric calibration terms, such as the flux-density, delay, and phase-referencing solutions. One cycle of frequency-independent, direction-independent, phase-only self-calibration is also applied for each beam as part of the imaging pipeline.
Individual calibrated measurement sets are imaged and deconvolved separately with imaging weights determined by Wiener pre-conditioning. These weights have a setting equivalent to a Brigg’s robustness value of 0.0, which was determined to provide an optimal combination of resolution, sensitivity, and point spread function (PSF) quality. The $w$-projection technique is used to account for the non-coplanarity of the sky, using a total of 557 w-planes. Multi-term deconvolution (with two terms) is used to model variation in source flux over the synthesised bandwidth and multi-scale deconvolution (with six scales) is used to improve the modelling of extended sources. Beam images are then averaged with the same weight to form a mosaic in the image plane. Primary beam correction is applied during the linear mosaic process using circular Gaussian models whose size is determined by holography observations. Although the beams differ in shape and are coma-distorted at the edge of the FOV, an average value is used for all 36 beams with an estimated error of 10\%.
As part of the routine ASKAP/EMU validation procedures, a preliminary source catalogue was created from the mosaic image and compared with other large-area radio survey catalogues. For these data, compact radio sources were crossmatched with the Sydney University Molonglo Sky Survey \citep[SUMSS:][]{SUMSS} catalogue \citep{2003MNRAS.342.1117M}, resulting in 111 matches. The median astrometric offset of $\sim 1.3$ arcsec between these matches
is consistent with zero offset given the dispersion in this measurement. The median flux ratio of the ASKAP sources versus their SUMSS counterparts is 0.81 overall but is approximately 1.0 for the brightest matches. Given that this flux comparison is complicated by the large difference in spatial resolution and a modest difference in frequency (for which we assume a spectral index of -0.8) we do not interpret this result as evidence for a systematic flux error. However, there is one additional concern that contributes to the flux uncertainty in these data and that is the lack of a common restoring beam among the images. Because the individual beam images are processed independently, the beams can have different UV coverage (e.g. because of flagging or declination) and therefore different PSFs. Gaussian fits to each of the individual 36 PSFs yielded beam sizes with geometric means (i.e. $\sqrt{b_{\rm maj}\,b_{\rm min}}$) ranging from 9.8 to 11.1 arcsec. As we were unable to propagate these individual beam sizes to the final linear mosaic image, we assumed a constant beam size of 11.2 $\times$ 9.5 arcsec at a position angle of 87 degrees (a geometric mean of 10.3 arcsec). The difference between the true and assumed PSF size will lead to a systematic error in integrated flux measurements, which is typically only $\sim$3\% but could be as large as 10\% in some places in the mosaic.
The resulting images of the beams covering A3395-A3391, as well as the full ASKAPsoft mosaic image of SB8275, show artefacts that mainly emanate from two bright and extended radio galaxies at the centres of the clusters: sources N1 and S1 (see Fig.~\ref{fig:overlay} and \ref{fig:ASKAPradiomap}). Negative bowl-like artefacts are present around the radio galaxies, with larger-scale positive and negative rings extending up to $\sim 1$~deg from their centres. The cause of these artefacts is currently being investigated and will be addressed in an upcoming technical paper (Wilber et al. in prep).
In an attempt to improve the image quality, \citet{2020arXiv200601833W} recently tested direction-dependent calibration and imaging on ASKAP early science and pilot survey observations. These latter authors report a significant reduction in artefacts around bright, compact sources after implementing further data processing via the third-generation software packages KillMS (kMS) \citep{2014A&A...566A.127T, 2015MNRAS.449.2668S} and DDFacet (DDF) \citep{2018A&A...611A..87T}, which are officially being used for data products of the LOFAR Two-Metre Sky Survey (LoTSS; \citealp{2017A&A...598A.104S}). As part of their testing, \citet{2020arXiv200601833W} developed a direction-dependent ASKAP pipeline, which was used to create a mosaic image of SB8275. This direction-dependent mosaic image is referred to as the ASKAP-DD image hereafter and is used to measure flux densities of radio sources.
We refer the reader to \citet{2020arXiv200601833W} and an upcoming paper, Wilber et al. in prep, for details of the ASKAP direction-dependent pipeline and specific processing steps.
Finally, the convention for naming sources discovered with EMU is EMU NA JHHMMSS.s+DDMMSS where `N' will be replaced with `E' for `early science' or `P' for `pilot survey', and `A' is `S' for `source', `C' for component, or `D' for `diffuse emission'.
\subsection{Murchison Widefield Array observations}
\begin{table}
\centering
\caption{\label{tab:obs:mwa}Details of MWA-2 images used.}
\begin{tabular}{c c c c c}
\hline
Band & $\nu_\text{c}$ & $\sigma_\text{rms}$ & $\Delta S_\nu$ & Beam \\\hline
& (MHz) & (mJy\,beam$^{-1}$) & \% & (${}^{\prime\prime}\times{}^{\prime\prime}$) \\\hline
\multicolumn{5}{c}{MWA-2 robust $+1.0$ tapered}\\\hline
72--103 & 88 & 12 & 9 &$192 \times 120$ \\
103--134 & 118 & 6.2 & 9 & $167 \times 105$ \\
139--170 & 154 & 6.8 & 9 &$181 \times 139$\\
170--200 & 185 & 7.6 & 9 & $165 \times 130$ \\
200-231 & 216 & 9.2 & 9 & $157 \times 125$ \\\hline
\multicolumn{5}{c}{MWA-2 robust $0.0$}\\\hline
139--170 & 154 & 2.6 & 9 &$85 \times 75$\\
170--200 & 185 & 2.8 & 9 & $70 \times 61$ \\
200-231 & 216 & 2.9 & 9 & $61 \times 53$ \\\hline
\multicolumn{5}{c}{MWA-2 robust $+1.0$}\\\hline
170--231$^\text{a}$ & 200 & 2.1 & - & $91 \times 79$ \\\hline
\end{tabular}\\
{\footnotesize \textit{Notes.} $^\text{a}$ Only used to create lower right panel in Fig.~\ref{fig:mwa2}.}
\end{table}
The A3391-95 system has been observed with the Murchison Widefield Array Phase 2 \citep[hereafter MWA-2;][]{tgb+13,wtt+18} in its `extended' configuration. The observations cover five frequencies: 88, 118, 154, 185, and 216 MHz, each with 30~MHz bandwidth. Full processing details for this form of MWA-2 data are described by \citet{Duchesne2020}, but we briefly describe this process here. Observations for the MWA-2 are taken in two-minute snapshot observing mode, with each snapshot calibrated and imaged independently; imaged snapshots are linearly stacked to create deep mosaic images. Calibration is performed using a sky model generated from the GLEAM extra-galactic catalogue \citep{2017MNRAS.464.1146H} using the full-Jones \texttt{Mitchcal} algorithm as described by \citet{oth+16} and full-embedded element primary beam model \citep{scs+17}. For this work, imaging is performed with `Briggs' weighting with robust parameters between $+0.5$ and $+1.0$ with additional Gaussian tapering applied to the 154-, 185-, and 216-MHz datasets to ensure a consistent PSF and $u$--$v$ sampling across the full MWA band. We use the multi-scale CLEAN algorithm within the widefield imager \texttt{WSClean} \citep[][]{wsclean1,wsclean2} for deconvolution. We make a wideband ($\Delta\nu = 60$~MHz) image centred on 200~MHz by stacking the 185- and 216-MHz mosaics made with robust $0.0$ weighting with no tapering as a higher-resolution reference image but note that at this weighting a significant fraction of flux is lost for extended sources. A separate CLEAN component model and residual mosaics are also created and are used for flux-density measurements, with a factor of $\sim0.6$ (dependent on the source size) applied to integrated residual flux density to account for a difference in CLEAN and dirty flux. Image details are presented in Table~\ref{tab:obs:mwa}. GLEAM and MWA-2 images are shown in Fig.~\ref{fig:mwa2}. \par
\section{Results}
\begin{figure}
\center{\includegraphics[width=\columnwidth]{A3391_95_eRosita_Filament_Region.pdf}}
\caption{Region of the bridge from the eROSITA X-ray map. The green box denotes the volume we assume for calculating the limits on the emissivity in the bridge region; see Sec. 4.1.}
\label{fig:region}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{DD_A3391-95_paper.jpg}
\caption{Our ASKAP mosaic image of SB8275 covering the Abell 3391-95 system and surrounding sources after direction-dependent calibration. Magenta contours are at $4\sigma$ and cyan contours are $6\sigma \times l$ where $l$ is [1, 2, 8, 32, 128, 512, 1024, 2048] and $\sigma = 36\,\mu$Jy beam$^{-1}$. White dashed contours are $-4\sigma$. All sources highlighted and magnified in this image fall within the mean redshift of the cluster system except for F3 which is at $z=0.193$. Details on these sources are summarised in Tables 2 and 3.}
\label{fig:ASKAPradiomap}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=1\linewidth]{A3391-95_MWA2-GLEAM_multiplot.pdf}
\caption{\label{fig:mwa2} GLEAM (top row) and MWA-2 (bottom row) images of the A3391-95 system. The solid white circle indicates the location of A3391 and the dashed white circle indicates A3395. Both circles have a 1~Mpc radius at the mean cluster redshift. We note that the colour scale on each map is linear, ranging from $-1\sigma_\text{rms}$ to $20\sigma_\text{rms}$.}
\end{figure*}
\label{sec:results}
In the ASKAP image (Fig.~\ref{fig:ASKAPradiomap}), a radio bridge is not detected, although the short spacings allow us to retain sensitivity up to scales of $\sim$1 degree, significantly larger than the X-ray bridge. In our image after direction-dependent calibration, artefacts on the largest scales as well as negative bowl artefacts are reduced. The noise in the image ranges from $\sim20~\mu$Jy to $\sim50~\mu$Jy per beam where the strongest artifacts emanate from the brightest cluster galaxies at the centre of Abell 3391 and Abell 3395. There does not appear to be any radio emission coincident with the bridge detected by eROSITA.
No radio bridge is seen in the MWA-2 or GLEAM data (Fig. \ref{fig:mwa2}). We note that the GLEAM data may show faint large-scale emission around bright, complex sources, but this is due to residual un-deconvolved
sidelobes of the PSF characteristic of MWA data. As with the ASKAP data, the bright, complex radio galaxies in each cluster limit the sensitivity that can be achieved in this region. We note that MWA-2 and GLEAM data, despite the expected steep spectra of the bridge synchrotron emission, may not achieve better surface brightness sensitivity than ASKAP \citep[][though we note that for their analysis, MWA-2 are imaged at robust $0.0$, reducing the surface brightness sensitivity for larger scales compared to robust $+1.0$ used here]{Hodgson2020}.
At the same time, the merging cluster system contains several nearby radio galaxies with extended features that may be interacting with the intracluster environments. In the following sections we report on the upper limit of the non-detection of the radio bridge and report flux densities for radio galaxies within the cluster field.
\subsection{Upper limit on the radio bridge}
There seems to be a consensus that the following physical conditions need to be met in order for a radio-synchrotron emission bridge to exist: (1) seed electrons (either from earlier structure formation shocks or from AGN), (2) a magnetic field, (3) an ongoing acceleration mechanism as described in Sect. 1, and finally (4) sufficient time for the acceleration to have produced a sufficient number of electrons of the required Lorentz factors.
In order to compute the flux values in the intercluster region, we placed a square box of 1120 arcsec per side on the bridge region and then compared it to 36 other such boxes outside of the bridge region.
We then blanked areas with flux densities above 0.4 mJy and calculated the mean and rms in the boxes. Subsequently, we subtracted the mean of the 36 boxes outside of the bridge from the mean of the box on the bridge, and calculated the scatter in the means between the boxes (an appropriate measure of the uncertainty of the flux in each box). The mean (residual) flux in the bridge box is then 39 mJy, and the scatter between boxes is 25 mJy, meaning a non-detection.
The 25 mJy scatter between boxes is approximately 7.5 times larger than the fluctuations on a smaller scale caused by large-scale ripples in the image.
The resulting upper limit on the emissivity is given by $\langle J \rangle_{1\,{\rm GHz}} =4\pi d_L^2 S/V$, where $V$ is the volume of the emitting region and $d_L$ the luminosity distance, which is $d_L\sim 237$ Mpc. For $S$ we take the $3\sigma$, that is, 75 mJy. The boundary of the bridge is difficult to define. If we assume that the bridge roughly follows the emission detected by eROSITA, the radius of the bridge is about 600 kpc or $10'$, and its length is about 1.3 Mpc or $21'$ (see Fig.~\ref{fig:region}). If we assume a cylindrical volume of dimensions 1.3 Mpc $\times \pi\, (0.6$ Mpc)$^2$, we obtain
\begin{eqnarray}
\langle J \rangle_{1\,{\rm GHz}} &<& \frac{ 7.5\times 10^{-28} {\rm W}\,{\rm Hz}^{-1}{\rm m}^{-2}\times 4\pi\times 5.4\times 10^{49}{\rm m}^{2}}{4.3\times 10^{67} {\rm m}^{-3}} \nonumber \\
&<& 1.2 \times 10^{-44} {\rm W}\, {\rm Hz}^{-1} {\rm m}^{-3} .
\end{eqnarray}
If we take $\alpha\sim -1.3$ as in \cite{2019Sci...364..981G}, one would expect an emissivity at 140 MHz of $\langle J \rangle_{140\,\text{MHz}}=1.5\times10^{-43}$~W~Hz$^{-1}$m$^{-3}$, which is a factor 1.7 higher than the emissivity derived for the A399-A401 system.
The reasons for a non-detection of synchrotron emission from the bridge region at the level of A399-401, for example, may lie in the absence of any one of the four points mentioned at the beginning of this section. If we assume that the relation between thermal gas density and magnetic field strength found in galaxy clusters is also valid in the intercluster region, we can estimate the magnetic field strength using the observed scaling $B\propto n_e
^{0.5}$ \citep{2020ApJ...888..101J}. In cluster centres with electron number densities of $n_e\sim 0.1$ cm$^{-3}$, typical field values of 1-10 $\mu$G are observed. Consequently, the thermal gas density in the bridge region of A3395-A3391 is $n_e\sim 1.5 \times 10^{-4}$ cm$^{-3}$ as inferred from SZ-observations \citep{2017PASJ...69...93S}, which would correspond to a magnetic field strength of $B\sim 0.03-0.3\,\mu$G. We note that a similar value for the electron density, $n_e= 1.08 \times 10^{-4}$ cm$^{-3}$ , was inferred from X-ray observations with Chandra and XMM-Newton by \cite{2018ApJ...858...44A}.
In comparison, the electron density inferred for the intercluster filament in the A399-A401 system is $n_e\sim 2.5-3.3 \times 10^{-4}$ cm$^{-3}$, which is very close to the estimate for A3395-A3391 \citep{2017A&A...606A...1A}. Also, $B=0.1-0.2\, \mu$G was found by \cite{2019Sci...364..981G} in their simulations of A399-401 and $B=0.5\, \mu$G was found by \cite{2020PhRvL.124e1101B} including dynamo amplification.
From the limit on the surface brightness of $I_0\sim 0.3\,\mu$Jy arcsec$^{-2}$, one can derive an upper limit on the equipartition magnetic field $B_{\rm eq}=(24\pi u_{\rm min}/7)^{1/2}$, where \citep{2004IJMPD..13.1549G}
\begin{eqnarray}
u_{\rm min} & \sim & 2.7\times10^{-11}(1+k)^{4/7}(\nu/\mathrm{GHz})^{-4\alpha/7} \nonumber\\
& &(I_0/\mathrm{(mJy/arcsec}^2))^{4/7} (d/\mathrm{kpc})^{-4/7}\mathrm{erg\, cm}^{-3} .
\end{eqnarray}
For $\alpha=-1.3$, a proton-to-electron energy ratio of $k=1$, and a source depth of $d_s=800$ kpc, we get $u_{\rm min}\sim 8.1\times 10^{-14} \mathrm{erg\, cm}^{-3}$, and consequently $B_{\rm eq} = 0.9\, \mu$G. If we were to detect the bridge at the level of our upper limit then the equivalent equipartition field would be 0.9 $\mu$G. This is higher than derived from our density-scaling argument, and so it is possible that the magnetic fields are simply too low to be detected.
However, we note that the electron density for A399-401 is only slightly higher, $n_e=2.5-3 \times 10^{-4}$ cm$^{-3}$, but the observed surface brightness of its bridge requires much higher fields than the scaled values.
\cite{2020PhRvL.124e1101B} suggest that radio bridges originate from second-order Fermi acceleration of electrons interacting with turbulence.
The complex dynamics of substructures associated with massive filaments connecting cluster pairs can generate turbulence even before the two clusters
undergo their central merger.
Specifically, they explored the role of solenoidal turbulence using a mechanism proposed by \cite{2016MNRAS.458.2584B} where particles interact with magnetic field lines diffusing into super-Alfvenic incompressible turbulent flows.
Provided that the acceleration mechanism in A3391-A3395 is Fermi-II acceleration, the non-detection would imply that the average acceleration time for electrons is longer than the average cooling time; as shown by \cite{2020PhRvL.124e1101B}, this is independent of the magnetic field in the bridge region.
The implication would be that either the turbulent injection scale is larger in A3391-A3395 than in A399-A401 or the turbulent velocity is smaller. We consider the latter to be the most likely because the turbulent injection length should be similar for systems of similar sizes while the turbulent velocity can vary significantly.
A smaller turbulent velocity in A3391-A3395 could be motivated by the smaller mass of the system and by the fact that these two systems are separated by a greater distance (in terms of virial radius of the clusters) than that separating A399-A401, possibly meaning that the merger is in a less advanced stage.
Also, the A3391-A3395 clusters are less massive than either A1758 or A399-A401.
Furthermore, as also briefly mentioned in paper 1, we cannot know the true three-dimensional distance between the systems. It is possible that A3391 and 3395 are much further apart than what is implied by their projection, as is indeed suggested by the X-ray analysis in Reiprich et al. (2020). This would likely result in lower estimates of electron densities, and might also explain the absence of giant radio halos in the constituent clusters, as the occurrence of giant radio halos depends strongly on mass \citep{2010A&A...517A..10C}.
Given their similarity to A399-A401 in other aspects (both premerging, similar gas densities in the bridge region), the A3391-A3395 system is a promising target for deeper radio observations.
For a more detailed study of the implications of the non-detection, dedicated MHD simulations are needed, such as those recently presented in Locatelli et al. (2020). These latter infer upper limits in LOFAR observations of two cluster pairs, RXCJ1659.7+3236-RXCJ1702.7+3403 and RXCJ1155.3+2324-RXCJ1156.9+2415, yielding upper limits on the magnetic field in the intercluster region of $B<0.2\, \mu$G. Using cosmological simulations, these latter authors infer primordial magnetic fields of less than 10 nG.
\subsection{Other notable radio features}
In addition to the absence of diffuse emission in the intercluster bridges, the environment of this unusual cluster-pair field shows a number of very interesting sources that are worth cataloguing given the potential impact of the cluster system. By no means is this a complete sample, as it was compiled by identifying extended objects by eye. The field contains a number of candidates for diffuse radio emission that will need to be followed up with deeper X-ray and radio observations. In addition, the field contains a number of Giant Radio Galaxies (GRGs) of which all but one lie in the background of this cluster pair. The exception is the FR\,I radio galaxy J0621--5647 in the far south, at the same $z$ as the
cluster pair. Polarisation measurements of these GRGs could be interesting for rotation-measure studies of the magnetic fields in the double cluster region \citep{2013MNRAS.432..243P,2020A&A...638A..48S}. Finally, there are some other notable radio galaxies that show signs of interaction with the ICM of A3391-A3395, such as the wide-angle tail source F2 as well as the radio galaxy F1. The latter shows lobes that are transversally advected, which may be caused by the galaxy falling into the cluster (see Fig.~\ref{fig:overlay}).\\
Table~\ref{tab:flux} lists the measurements of flux density, $S$, and spectral index, $\alpha$, for radio sources in the A3391-A3395 cluster field. Flux density was measured within a region marking the 3$\sigma$ contour line using a local rms value for $\sigma$. To measure the spectral index of specific sources or source regions (except for S3), we convolved the ASKAP-DD image to the same resolution as our tapered MWA-2 observations. For S3, to ensure S2 and S3 were resolved enough for flux-density measurements, we use the native resolution ASKAP-DD image and the robust 0.0 MWA-2 images. We do not measure S2 directly with the MWA-2 data as the diffuse emission becomes resolved out in the robust 0.0 MWA-2 images. We used \href{https://gist.github.com/Sunmish/198ef88e1815d9ba66c0f3ef3b18f74c}{fluxtools.py} to measure the flux density and error on the flux density when considering the rms. The uncertainties are calculated taking into account the image noise and flux-scale errors as follows,
\begin{equation}
\Delta S = \sqrt{\Delta S^2_\text{image} + \Delta S^2_\text{fluxscale} }= \sqrt{N\times \sigma^2 + (f \times S)^2},
\end{equation}
where $N$ is the number of independent beams covering the source, $f=10\%$ is the flux-scale error for the EMU observation, and $f=9\%$ for the MWA-2 observations. The spectral index is estimated by fitting the flux densities of the sources to a power law as a function of the observed frequency.\\
\begin{table*}
\caption{Properties of radio galaxies in Fig.~\ref{fig:ASKAPradiomap} and their respective host galaxies. Columns are (1) a shortened name where `EMU ES' stands for Evolutionary Map or the Universe Early Science Source, (2) the source label according
to Figs.~\ref{fig:overlay} and \ref{fig:ASKAPradiomap}, (3) the name of the optical host galaxy, (4) its
spectroscopic redshift, (5) its largest angular size (LAS) measured
in a straight line between opposite ends of the detectable radio emission,
(6) its largest projected linear size (LLS) assuming the cosmological parameters
listed at the end of Sect. 1.}
\begin{tabular}{ccccccccc}
\hline
EMU ES & Our label & Host name & Redshift & \multicolumn{2}{c}{Extent of radio lobes} \\
name & (in Fig.~\ref{fig:ASKAPradiomap})& & $z$ & LAS & LLS & Comments \\
& & & & [arcmin] & [kpc] \\
\hline
J0626--5341 & N1 & ESO 161-IG 007 NED02 & 0.0551 & 5.35 & 346 & FR\,I in A3391 \\
J0626--5433 & S1 & WISEA J062649.57--543234.4 & 0.0520 & 6.07 & 372 & PKS\,0625--545, in A3395 \\
J0626--5432 & S2 & -- ? & -- & 5.52 & -- & relic/remnant in A3395 \\
J0625--5427 & S3 & WISEA J062557.04--542750.4 ? & 0.0603 & 3.25 & 229 & remnant/relic in A3395 \\
J0622--5334 & F1 & WISEA J062255.56--533434.5 & 0.0567 & 5.71 & 379 & asym. FR\,II?, PMN\,0622--5334 \\
J0621--5241 & F2 & 2MASX J06214330--5241333 & 0.0511 & 6.21 & 374 & WAT, PKS\,0620--52 \\
J0632--5404 & F3 & WISEA J063201.16--540457.4 & 0.193 & 5.80 & 1117 & FR\,II plume \\
\hline
\end{tabular}
\label{tab:RGs_EMU}
\end{table*}
\begin{table*}
\centering
\caption{\label{tab:flux} Integrated flux density measurements (within $3\sigma$ contours where $\sigma$ is the local rms) of sources labelled in Fig. \ref{fig:ASKAPradiomap}.}
\begin{tabular}{c | c c | c c c c c c }\hline
Source & ASKAP-DD & rms ($\sigma$) & MWA-2 & MWA-2 & MWA-2 & MWA-2 & MWA-2 & ASKAP-DD$^\text{a}$ \\\hline
& 1013 MHz & local & 88 MHz & 118 MHz & 154 MHz & 185 MHz & 216 MHz & 1013 MHz \\
& [mJy] & [$\mu$Jy/beam] & [Jy] & [Jy] & [Jy] & [Jy] & [Jy] & [Jy] \\ \hline
N1 & $8799 \pm 880$ & 50 & $89.6 \pm 9.7$ & $62.1 \pm 6.2$ & $48.6 \pm 4.3$ & $41.1 \pm 3.7$ & $35.3 \pm 3.2$ & $9.06 \pm 0.91$ \\
S1 & $4314 \pm 431$ & 50 & $33.2 \pm 4.0$ & $24.5 \pm 2.7$ & $19.8 \pm 1.9$ & $17.2 \pm 1.6$ & $15.2 \pm 1.6$ & $4.41 \pm 0.44$ \\
S2 & $72.1 \pm 7.2$ & 48 & - & - & $1.38\pm0.29$ $^{c}$ & $1.17\pm0.23$ $^{c}$ & $0.91\pm0.20$ $^{c}$ & - \\
S3 & $226 \pm 22.6$ & 48 & - & - & $1.26\pm0.12$ & $0.99\pm0.10$ & $0.83\pm0.10$ & - \\
S2$+$S3$^\text{b}$ & $298 \pm 29.8$ & 48 & $5.92 \pm 0.87$ & $3.81 \pm 0.48$ & $2.64 \pm 0.26$ & $2.15 \pm 0.21$ & $1.73 \pm 0.17$ & $0.392 \pm 0.039$ \\
F1 & $175 \pm 17.5$ & 30 & $0.786 \pm 0.104$ & $0.532 \pm 0.063$ & $0.565 \pm 0.057$ & $0.428 \pm 0.043$ & - & $0.167 \pm 0.017$ \\
F2 & $4883 \pm 488$ & 40 & $25.5 \pm 3.2$ & $18.8 \pm 2.1$ & $15.5 \pm 1.5$ & $13.8 \pm 1.3$ & $11.6 \pm 1.1$ & $4.90 \pm 0.49$ \\
F3 & $556 \pm 55.9$ & 26 & $3.14 \pm 0.36$ & $2.28 \pm 0.24$ & $1.89 \pm 0.17$ & $1.58 \pm 0.14$ & $1.39 \pm 0.13$ & $0.586 \pm 0.059$ \\\hline
\end{tabular}\\
{\footnotesize \textit{Notes.} $^\text{a}$ Convolved to a resolution of $150$~arcsec $\times$ $131$~arcsec to match the tapered MWA-2 images, which results in an additional contribution from nearby but faint point sources. $^\text{b}$ S2 and S3 are convolved together in the tapered and 88--118 MHz MWA-2 images. $^{c}$ Residual flux density after subtracting S3 from S2$+$S3.}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=8cm]{F2.RGB-image.png}
\includegraphics[width=8cm]{N1.RGB-image.png}
\includegraphics[width=8cm]{F3.RGB-image.png}
\includegraphics[width=8cm]{F1.RGB-image.png}
\includegraphics[width=8cm]{S1.RGB-image.png}
\includegraphics[width=8cm]{S2-3.RGB-image.png}
\caption{ASKAP 1 GHz radio continuum images of the radio galaxies shown in Fig. 1 overlaid on DSS2 $R$-band images. Top left: F2, Top right: N1, Middle left: F3, Middle right: F1, Bottom left: S1, Bottom right: S2 and S3. The radio contour levels starting at 3$\sigma$ are typically 0.012, 0.025, 0.05, 0.1, 0.25, 0.5, 1, 2.5, 5, 10, 25, 50, 100, 250 and 500 mJy/beam, except for F1-F3 where a 0.006 mJy/beam contour is added, and N1 where the 0.012 mJy/beam contour is omitted.}
\label{fig:mainsource}
\end{figure*}
\subsubsection{Source N1 (EMU ES J0626--5341)}
The bright radio galaxy EMU ES J0626--5341 (EMU ES: Evolutionary Map of the Universe Early Science Source) is located in the centre of the Abell~3391 cluster and extends over 4.5 arcmin (see Fig.~\ref{fig:mainsource}). It is also located at the centre of the X-ray emission (see Fig.~\ref{fig:xrayover}). Its host is the elliptical galaxy 2MASX J06262045--5341358 ($PA$ = 54\degr), which is the eastern component of ESO 161-G008, with a redshift of $z$ = 0.055. We estimate a linear projected radio size of 290 kpc. The radio lobes show numerous twists possibly due to density variations in the ICM.
\subsubsection{Source S1 (EMU ES J0626--5432)}
The bright radio galaxy EMU ES J0626--5432 (PKS 0625--545) is a Fanaroff-Riley class I (FR I) \citep{1974MNRAS.167P..31F} source located in the centre of the A3395 cluster and extends over 5.5 arcmin (see Fig.~\ref{fig:mainsource}). Its host galaxy is WISEA J062649.57--543234.4 ($z$ = 0.052). Radio fluxes from 145 MHz to 8.5 GHz as well as 148 GHz are listed in NED. The SUMSS 843 MHz flux is catalogued as $\sim$5 Jy and the Australia Telescope Compact Array (ATCA) 1.4 GHz flux as $\sim$3.1 Jy \citep{2011ApJ...743...78L}. To the south, just as it starts to exit the denser X-ray gas (see Fig.~\ref{fig:chandra_s1}), the tip of the radio jet starts to expand spherically and the brightness of the emission drops. This could be the onset of the formation of an X-ray cavity. Comparison with the eROSITA data also shows that the BCG is offset from the centroid of the X-ray emission (see Fig.~\ref{fig:xrayover}). This was studied by \cite{2014ApJ...797...82L, 2020MNRAS.tmp.3079D} for example. In X-ray groups in COSMOS, we showed that BCGs can be offset from the group centre \citep{2019MNRAS.483.3545G}. This has also been studied in HIFLUGCS clusters \citep{2011A&A...526A.105Z}.
\begin{figure*}
\includegraphics[width=\textwidth]{chandra_radio2.pdf}
\caption{Chandra X-ray image of the S1 region and radio contours. Left: Exposure-corrected Chandra image in the 0.5-5 keV band after masking point sources, smoothed by a Gaussian of $\sigma=3.9$~arcsec. The contours show the EMU data at levels of $10^{-4}$, $10^{-3}$, $0.01$, $0.05$ and $0.1$ Jy~beam$^{-1}$. Centre: Fractional difference between an exposure-corrected image smoothed by a Gaussian with $\sigma=7.9$~arcsec and a smooth model which is its average at each radius (measured from the radio nucleus). Right: Edge-filtered X-ray image showing the gradient magnitude on scales of $16$~arcsec, meaning that components on scales larger than $16$~arcsec have been filtered out.}
\label{fig:chandra_s1}
\end{figure*}
The interaction of the radio source with the surrounding ICM can be seen in more detail in a Chandra X-ray image of the region. After reprocessing the Chandra observation (OBSID 4944), images and exposure maps were created, and point sources were masked. Figure~\ref{fig:chandra_s1} shows the smoothed exposure-corrected image, the fractional difference from a model created from the radial average, and an edge-filtered X-ray image made using using the GGM algorithm \citep{2016MNRAS.457...82S, 2016MNRAS.460.1898S}. There is some indication of a depression in the X-ray surface brightness in the region of the round, southern tip, but it is of low significance \citep{2011ApJ...743...78L}. The northern tip appears to end where there is an excess of emission in the smooth model residual map. The edge-filtered map confirms that the two ends of the radio source are located where there are steep gradients in the X-ray surface brightness. There is also a clear connection between the structure within the radio source and the X-ray gradient filtered image.
\begin{figure}
\includegraphics[width=1.\columnwidth]{N1_overa.png}
\includegraphics[width=0.95\columnwidth]{S1_over2.png}
\includegraphics[width=0.95\columnwidth]{S2-3_c.png}
\caption{Overlays of radio images on optical images from the g, r, and z images from the DECam Legacy Survey. The turquoise contours are from eROSITA. Top: Multi-wavelength close-up around radio galaxy N1. The radio image from the EMU pilot survey at a resolution of 11" is shown in orange, with the high-brightness areas masked out and replaced with white contours. Middle: Multi-wavelength close-up around the radio S2/S3 region. The radio emission from the EMU ES observation at a resolution of $11''$ is shown in blue, overlaid in orange with a lower, $25''$ resolution image. The high brightness areas are colour-coded separately. Bottom: Multi-wavelength close-up around radio galaxy S1. The radio image from the EMU pilot survey at a resolution of $11''$ is shown in orange.}
\label{fig:xrayover}
\end{figure}
\subsection{Source S2+S3}
The sources S2 and S3 form a complex conglomerate of radio galaxies and interconnected diffuse emission (see Fig.~\ref{fig:mainsource}). The integrated spectral index between 88 MHz and 1013 MHz is $\sim -1$. Figure~\ref{fig:mainsource} shows an overlay of the S2-S3 complex over the optical DSS2 R-band image.
The L-shaped source S2 is fed by the radio galaxy 2MASX J06261051-5432261 that fans out into two arms towards the north and the west that are roughly perpendicular to one another. The arms extend for about 4 arcmin. There are further faint filaments of emission that stretch out to the source S3.
From the brightest spot close to the galaxy 2MASX J06255706-5427502, the source S3 extends to the northwest before it turns to form a faint, linear source that extends for about 6 arcmin towards the southwest. This could be a radio relic or phoenix that shows re-acceleration of plasma that may originate from an AGN. Interestingly, the southern part of this linear feature lies in a high-temperature region as indicated by the oxygen-to-softband ratio in the eROSITA data (see Fig. 18 in Reiprich et al. 2020). It is not clear whether this region corresponds to a shock or is merely a region of higher pressure that could boost the synchrotron emission from old radio plasma that may have come from a radio galaxy.
\subsection{Source F1 (EMU ES J0622--5334)}
EMU ES J0622--5334 (PMN\,0622--5334) is a large, rather asymmetric radio galaxy with Fanaroff-Riley class I morphology in Abell S0584 (see Fig.~\ref{fig:mainsource}). Its host galaxy is WISEA J062255.56--533434.5 at $z$ = 0.0567. The radio lobes extend over $\sim$6 arcmin, corresponding to $\sim$400 kpc. The SUMSS 843 MHz flux of PMN\,0622--5334 is catalogued as $75 \pm 6$ mJy, likely not including the low-surface brightness emission. The 4.85 GHz flux is catalogued as $47 \pm 8$ mJy \citep{1996ApJS..103..145W}.
\subsection{Source F2 (EMU ES J0621--5241)}
EMU ES J0621--5241 (PKS\,0620--52) is a large wide-angle tail (WAT) radio galaxy (see Fig.~\ref{fig:mainsource}) which is located far beyond the virial radius of A3391 and could be infalling into the system.
Its host galaxy is 2MASX J06214330--5241333 at $z$ = 0.0511. It is associated with RX J0621.6--5241. The two radio lobes have extents of 4 arcmin (NE; 240 kpc) and 2.5 arcmin (NW; 150 kpc).
\subsection{Source F3 (EMU ES J0632--5404)}
EMU ES J0632--5404 is a GRG with an FR\,II morphology extending 5.8 arcmin (see Fig.~\ref{fig:mainsource}) along a position angle of $PA \sim 60$\degr. Its host galaxy is the quasi-stellar object (QSO) WISEA J063201.16--540457.4, 6dFGS gJ063201.2--540458, SWIFT J0632.1--5404 for which \cite{1998AJ....115.1253P} measured a redshift of $z=0.193$, while \cite{2005AJ....130..896S} measured $z=0.2036$. Comparing both spectra we believe the former redshift (0.193) to me more reliable, and therefore find an LLS of 1.12 Mpc. \cite{2005AJ....130..896S} measure a 843 MHz flux of 694 mJy.
This source is also GLEAM J063153-540527 with a low-frequency
spectral index of -0.88, and so ASKAP may show more of the extended
emission as the spectrum seems to flatten above 800 MHz.
The 4.85 GHz flux from PMN (source J0631-5405) is $0.155 \pm 0.011$ Jy, resulting in a spectral index of $\alpha^{1013}_{4850}=-0.85$, in agreement with low frequencies.
The host is also listed as blazar BZQ J0632-5404 by \cite{2009A&A...495..691M}.
\subsection{Large and Giant Radio Galaxies}
Based on visual inspection of the whole 30 deg$^2$ field for radio galaxies, we identified close to 200 objects larger than $\sim$1 arcmin. Of these, 27 are GRGs with linear projected sizes greater than 1~Mpc (7 of these are presently still candidate GRGs). Only a single one was already known, namely J0632-5404, published as SGRS J0631--5405 by \cite{2005AJ....130..896S}.
A further 30 radio galaxies not listed in the present paper are larger than 700~kpc (of which 7 are candidate GRGs).
Giant Radio Galaxies are thought to reside in low-density environments where the jets face little resistance. However, about 10 percent of GRGs have now been found to reside in cluster environments. According to \cite{2020A&A...642A.153D}, about 820 GRGs larger
than 0.7 Mpc are known to date. However, owing to their large extent and low surface brightness, they are notoriously difficult to detect. Using Data Release 1 of LoTSS, \cite{2020A&A...635A...5D} found 239 GRGs (there, defined as galaxies with LLS $>0.7$ Mpc) in 424 deg$^2$ of sky, albeit in a frequency range of 120--168 MHz. For galaxies with LLS $>0.7$ Mpc, we obtain a sky density of $\approx 1.7$ deg$^{-2}$, while \cite{2020A&A...635A...5D} found 239/424 = 0.55 deg$^{-2}$, three times less than our value.
Those GRGs larger than 1 Mpc that we identify in this field are listed in Table~\ref{tab:RGs}. The GRGs in the area of the ASKAP field covered also by eROSITA are marked in Fig.~\ref{fig:large_overlay2}, and radio-optical overlays for a subset of GRGs are shown in Fig.~\ref{fig:radiogal}. Photometric redshifts were taken from \cite{2016ApJS..225....5B} and from \cite{2019ApJS..242....8Z}. We measured the flux densities, $S$, of the GRGs listed in Table~\ref{tab:RGs} within a region marking the 2$\sigma$ contour line using the local rms for $\sigma$. We used \href{https://gist.github.com/Sunmish/198ef88e1815d9ba66c0f3ef3b18f74c}{fluxtools.py} to measure the flux density and error on the flux density when considering the rms.
During our search we also identified a number of relics and cluster halos (see Table~\ref{tab:relics}). In order to identify the respective host galaxies we used optical and infrared surveys, in particular the Dark Energy Survey (DES; \cite{2020PhRvD.102b3509A}), 2MASS (\cite{2006AJ....131.1163S}), and WISE (\cite{2010AJ....140.1868W}). In the following we briefly describe some of the most remarkable objects.
\begin{figure*}
\center{\includegraphics[width=\textwidth]{A3391_95_eRosita_as20.60.5_DDF_large.png}}
\caption{eROSITA X-ray 0.3--2.0~keV adaptively smoothed image of the A3391-A3395 field overlaid with the ASKAP/EMU radio contours (smoothed with a five-pixel Gaussian). The red circles show the locations of the GRGs within the FOV of the eROSITA observations. This is only the central half of the area of the ASKAP image, and it therefore only shows 11 of the 27 GRGs in
Table 4. These GRGs in this field have higher redshifts than the A3391-A3395 cluster (see Table~\ref{tab:RGs}). Except for
perhaps two exceptions, there is no indication that these GRGs are correlated with X-ray emission peaks. The GRG with the strongest X-ray
peak is J0632-5404.}
\label{fig:large_overlay2}
\end{figure*}
\subsubsection{EMU ES J0621--5217}
EMU ES J0621--5217 is an FR\,I-type radio galaxy with LAS$\sim15$\,arcmin
with a radio core of 18.2 mJy/beam peak brightness, inner twin jets,
and evidence for precession given the shape of its outermost lobes (see Fig.~\ref{fig:radiogal}, left panel). It is highly asymmetric, likely due to projection effects, with the western lobe much brighter and closer to the core than the eastern lobe. We identify the host galaxy as WISEA J062112.81--521700.6 ($z$ = 0.042816; 6dF \cite{2009MNRAS.399..683J}). The projected linear extent of J0621--5217 is $\sim$760 kpc. This galaxy is located in the foreground of A3395-A3391.
\begin{figure*}
\centering
\includegraphics[width=8cm]{J0621-5213.RGB.png}
\includegraphics[width=8cm]{J0639-5455.RGB.png}
\includegraphics[width=0.8\textwidth]{J0621-5647.RGB.png}
\caption{ASKAP 1 GHz radio continuum images of the radio galaxies EMU ES J0621--5217 (top left), EMU ES J0639--5455 (top right), and EMU ES J0621--5647 (bottom) overlaid on DSS2 $R$-band images. In the top images, the radio contour levels are 0.006, 0.012, 0.025 0.05, 0.1, 0.25, 0.5, 1, and 2 mJy/beam and 0.1, 0.5, 2, 10, and 50 mJy/beam for the bottom image. The host of EMU ES J0621--5647 (bottom) is the elliptical galaxy 2MASX J06215057--5647566 ($z = 0.0539$). The very faint radio lobes appear to extend over more than $\sim15$\,arcmin. The convolved beam is 15 arcsec.}
\label{fig:radiogal}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=8cm]{J0624-5414.RGB.png}
\includegraphics[width=8cm]{J0630-5216.RGB.png}
\includegraphics[width=8cm]{J0627-5426.RGB.png}
\includegraphics[width=8cm]{J0617-5200.RGB.png}
\caption{ASKAP 1 GHz radio continuum images of some of the diffuse radio sources listed in Table~5: EMU ES J0624--5411 (A3395, top left), EMU ES J0630-5216 (A3397A/B, top right), ESO 161-G008 (in A3395N, $z$ = 0.0486, bottom left), and EMU ES J0617--5201 in A3385 (bottom right), overlaid on a DSS2 $R$-band image. In the lower right panel, the right half of the overlay shows the FRII radio galaxy DES J061722.22-515919.1 (Sect. 4.8.3). The radio contour levels are 0.006, 0.012, 0.025, 0.05, 0.1, 0.25, 0.5, 1, 2, 5, 10, and 20 mJy/beam. The convolved beam is 15 arcsec.}
\label{fig:halos}
\end{figure*}
\subsubsection{EMU ES J0644--5152}
EMU ES J0644--5152 has the largest LLS of all GRGs catalogued in this field with radio lobes extending nearly 5 arcmin and FR\,II morphology. Its host galaxy is SMSS J064437.48--515205.5, possibly a QSO, with a photometric redshift of $z \sim 0.75$ as estimated from the WISE colours (see Fig. 2 in \citealt{2018ApJS..235...10K}). We estimate its largest linear size (LLS) as 2.15 Mpc.
\subsubsection{EMU ES J0639--5455}
EMU ES J0639--5455 is a GRG with a rather unusual, WAT-like morphology (see Fig.~\ref{fig:radiogal}, right panel). Its host galaxy (2MASX J06391009--5455156, WISEA J063910.02--545515.2) is located at the radio-brightest spot in the far WSW and has a photometric redshift of $\sim$0.142, thus well behind A3391-A3395 \citep{2016ApJS..225....5B}.
We measure an angular extent of 7.15 arcmin, corresponding to a projected linear extent of 1.07 Mpc, a size very rarely seen for sources of this radio morphology.
\subsubsection{EMU ES J0621--5647}
EMU ES J0621--5647 is a FR\,I-type GRG, possibly extending at least $\sim$15 arcmin and possibly $\sim$32 arcmin (see Fig.~\ref{fig:radiogal}). Higher radio sensitivity is required to confirm the presence and extent of its faint radio lobes due ESE and WNW. We identify 2MASX J06215057--5647566
as the host galaxy at a redshift of $z=0.05391$ (6dF). EMU ES J0621--5647 lies at the
southern edge of the ASKAP field, but at the same redshift as A3391/3395.
If the outer lobes are confirmed, the LLS of J0621--5647 is $\sim$2\,Mpc.
\subsubsection{Summary of Giant Radio Galaxy properties}
In Table \ref{tab:RGs} we list the basic properties of the 20 GRGs (LLS$>$ 1 Mpc) and 7 candidate
GRGs found in the ASKAP field. Here we compare some of these properties
with 213 GRGs larger than 1\,Mpc as compiled from the literature
by \cite{2018ApJS..238....9K}. While the median redshift of the latter
GRGs is 0.248, our 27 GRGs have a median of 0.6, albeit with the caveat
that only two of them have spectroscopic redshifts.
We adjust the 1.4-GHz radio luminosities listed in
\cite{2018ApJS..238....9K}
to the ASKAP observing frequency of 1.0\,GHz by assuming an average radio
spectral index
of $\alpha=-0.8$, which implies a 1-GHz flux 1.3 times larger than the
1.4-GHz
flux, or $\log P_{\rm 1GHz} = \log P_{\rm 1.4GHz}+0.12$. While the 213 GRGs from the
literature have a median $\log P_{\rm 1GHz}$[W/Hz]=25.6, our 27 GRGs have a
median
$\log P_{\rm 1GHz}$[W/Hz]=25.5, and the two distributions are statistically
indistinguishable, ranging from $\sim$23.3 to $\sim$27.4. Consistent with
the literature sample we find no trend for the median linear size or the
median radio luminosity to change with redshift. Together with the fact
that our much smaller sample has a much larger median redshift,
we conclude that there is no evidence for cosmological evolution of the
population of GRGs.
If we exclude four GRGs with complex or hybrid radio morphologies from our sample of 27
and divide the rest into three classes of radio morphology, we
find median values of $\log P_{\rm 1GHz}$[W/Hz]=24.0 for the 3 FR\,Is,
24.5 for 7 FR\,IIs with remnant-type lobes,
and 26.1 for the 15 clear-cut FR\,IIs.
\subsection{Diffuse radio sources}
We discovered a number of candidate diffuse radio sources in the ASKAP field which are listed in Table~\ref{tab:relics}. The data are not sufficient for a reliable classification but we discuss the most interesting sources below. The diffuse sources in the cluster A3404 are the subject of a forthcoming paper and are not discussed here.
\subsubsection{Abell~S0592}
\cite{2020arXiv200601833W} detected a giant radio halo in the nearby cluster Abell S0592 ($z$ = 0.2216), also known as SPT CL J0553--3342 and MACS J0553.4--3342, using the same Early Science data used here.
The radio halo has a diameter of about 1 Mpc and an integrated flux density of $S_{\rm 1013 MHz} = 9.95 \pm 2.16$ mJy. The diffuse emission is also seen in ATCA data at 2.215 GHz.
\subsubsection{Abell~3395E}
We find extended radio emission associated with the bright elliptical galaxy ESO 161-G008 ($z$ = 0.0486), the central galaxy of A3395E. This could be diffuse emission from a radio halo, although image artefacts in that area make the full size of the structure difficult to assess. Three other cluster ellipticals in the area also have radio emission with significant tails: WISEA J062744.63--542644.4 (PGC~019090; $z$ = 0.044281) has a tail of 2 arcmin in length towards the east, while DES~J062719.2--542515.24 ($z$ = 0.0456) has a tail of 1 arcmin pointing north and 2MASS J06274388--5424266 ($z$ = 0.044971) has a short tail in the NW direction. The aforementioned galaxies lie within the X-ray cocoon of the cluster, which is aligned with and centred on ESO 161-G008. There are at least 40 known cluster
members within 5 arcmin of ESO 161-G008.
\subsubsection{Abell~3385}
We find diffuse emission in Abell~3385 ($z$ = 0.1245; 6df) also known as 1RXS J061747.8--520132. This emission lies right at the NW edge of the eROSITA FOV. North of the 1RXS position is a bright radio source that appears to connect to two faint strands of radio emission that run from the NW to the SE. The angular length of the filaments is 2 arcmin which corresponds to 270 kpc. The flux density of this diffuse source is $S_{\rm 1013 MHz} = 13.7 \pm 2.3$ mJy. This source is likely to be a radio relic. Finally, on the right (W) half of the bottom-right panel in Fig.~\ref{fig:halos} we see
the regular and straight FR II source hosted by DES J061722.22-515919.1,
a likely member of A3385, with $z_{\rm ph}\sim 0.128$, LAS$\sim 4.0'$, and LLS$\sim 550$ kpc.
\subsubsection{Abell~3397}
A3397 is a poorly studied galaxy cluster that is located north of A3391-A3395 with no published redshift, but the 6dF redshift survey \citep{2009MNRAS.399..683J} shows seven galaxies with mean measured redshift $\langle z\rangle =0.715$ (A3397A) and four others with $\langle z \rangle=0.1063$ (A3397A) within one Abell radius (1.7$'/z$) from the Abell centre \citep{2014MNRAS.445.4073C}. One more galaxy, 2MASX J062940.23-521418.5, with a redshift measured by \cite{2014ApJ...797...82L}, makes a total of eight spectroscopic members of A3397A
with a velocity dispersion of $750 \pm 80$ km\,s$^{-1}$ based on the BIweight
scale of ROBUST \citep{1990AJ....100...32B}. The eROSITA image shows an extension towards the west that is indicative of a merger in the east--west direction. The X-ray image also shows an extension towards the north. An extended radio source lies on the western side of the X-ray peak with hints of diffuse emission (see Fig.
~\ref{fig:halos}, top right panel) and appears to be composed of several components; its LAS is 1.4 arcmin which corresponds to an LLS of 600 kpc at $z=0.715$. The flux density of this source is $S_{\rm 1013 MHz} = 0.23 \pm 0.04$ Jy. The nature of the source is unclear.
\subsubsection{EMU ED J0624-5414}
The source sits in a patch of diffuse radio emission whose angular extent is about 4 arcmin. The redshift of WISEA J0624-5414 is $z=0.045$. It appears as two connected bars of emission that stretch from the SE to the NW. The main blob in the northern part of the structure extends for about 1.8 arcmin before it fades quite abruptly. The flux density of this source is $S_{\rm 1013 MHz} = 0.13 \pm 0.019$ Jy and there is no counterpart in the eROSITA image. This source is likely to be a radio halo (see Fig.~\ref{fig:halos}, upper left panel).
\begin{table*}
\centering
\caption{Properties of GRGs (LLS $>$ 1 Mpc) discovered in the 30 deg$^2$ ASKAP field towards Abell 3391/5 sorted in decreasing order of LLS. We list the host galaxy name, their spectroscopic (s), photometric (p), or estimated (e) redshifts, flux densities, $S$, luminosities, $\log P$(1013~MHz)/[W/Hz], and radio morphology. The superscript `C' after the source name denotes a GRG candidate for which we list the most likely host name. Flux densities, $S$, were derived by integration of the ASKAP-DD image and measured within the $3\sigma$ contour level where $\sigma$ is the local rms. Flux densities, $S$, marked with $^{\dag}$ were measured at $2\sigma$.}
\begin{tabular}{lllccccccc}
\hline
EMU ES & Host name & Redshift & \multicolumn{2}{c}{Extent of radio lobes} & $S$(1013~MHz) & $\log P$ & Radio\\
Name & $\alpha,\delta$(J2000) & $z$~~~~type & LAS & LLS & mJy & & morphology& \\
& & & [$'$] & [Mpc] & \\
\hline
J0644--5152 & SMSS J064437.48--515205.5 & 0.75~~~~p & 4.89 & 2.15 & $17.6 \pm 1.8$& 25.6 & FR\,II \\
J0621--5647 & 2MASX J06215057--5647566 & 0.0539s & 32? & 2.0? & $23.3 \pm 5.8$ $^{\dag}$& 23.3 & FR\,I/II remn., WAT? \\
J0608--5409 $^\text{C}$ & DES--J060805.04--540918.2 & 1.2~~~~~~p & 3.7 & 1.84 & $21.5 \pm 2.2$ & 26.2 & FR\,II \\
J0634--5309 & DES~J063428.15--530928.4 & 0.34~~~~p & $>$5.6 & $>$1.6 & $4.5 \pm 0.5 $ $^{\dag}$& 24.2 & FR\,II remn. \\
J0612--5157 $^\text{C}$ & DES--J061215.47--515738.1 & 0.7~~~~~~e & 3.7 & 1.59 & $54.4 \pm 5.4$& 26.0 & FR\,I/II \\
\\
J0613--5117 $^\text{C}$ & DES--J061339.02-511703.5 & 1.8~~~~~~e & 3.1 & 1.57 & $5.1 \pm 0.5$& 25.9 & FR\,II asym. \\
J0639--5136 & VHS J063919.22--513652.1 & 0.45~~~~e & 4.16 & 1.44 & $41.2 \pm 4.1$&25.5 & FR\,II \\
J0640--5257 & WISEA J064013.78--525702.2 & 0.6~~~~~~e & 3.57 & 1.43 & $9.9 \pm 1.0$& 25.1& FR\,II relic \\
J0633--5424 & 2MASX J06333235--5424190 & 0.165~~p & 7.65 & 1.30 & $20.8 \pm 2.1$& 24.2& FR\,I \\
J0613--5621 & DES~J061337.62--562114.6 & 0.35~~~~p & 4.3 & 1.27 & $7.9 \pm 0.8$ $^{\dag}$& 24.5 & FR\,I/II remn. \\
\\
J0629--5341 & DES~J062935.40--534124.4 & 1.65~~~~p & 2.48 & 1.26 & $ 102.0 \pm 10.3 $& 27.2 & FR\,II \\
J0636--5616 & DES~J063633.92--561635.6 & 0.6~~~~~~e & 3.06 & 1.23 & $10.8 \pm 1.1$& 25.2 & FR\,II \\
J0641--5136 & SMSS J064130.39--513602.7 & 0.395~~p & 3.67 & 1.18 & $6.1 \pm 0.6$& 24.5 & FR\,I/II \\
J0619--5558 & DES~J061909.04--555843.7 & 1.3~~~~~~p & 2.33 & 1.17 & $98.6 \pm 9.9$& 26.9 & FR\,II cpx \\
J0630--5218 & DES~J063043.12--521821.2 & 1.8~~~~~~p & 2.27 & 1.15 & $19.6 \pm 2.0$& 26.5 & FR\,II \\
\\
J0610--5412 $^\text{C}$ & DES--J061042.30--541225.8 & 0.8~~~~~~e & 2.5? & 1.13? & $1.2 \pm 0.1 $ $^{\dag}$& 24.5 & FR\,II? \\
J0631--5317 & DES~J063113.59--531719.6 & 0.45~~~~e & 3.27 & 1.13 & $27.5 \pm 2.8$& 25.3 & FR\,II \\
J0615--5506 $^\text{C}$ & DES--J061518.25-550647.5 & 1.015~~p & 2.33 & 1.13 & $51.2 \pm 5.1$ & 26.4 & FR\,II, no core? \\
J0632--5404 & 6dFGS gJ063201.2--540458 & 0.193~~s & 5.80 & 1.12 & $556.0 \pm 55.9$& 25.8 & FR\,II plume \\
J0649--5501 & SMSS J064947.02--550109.6 & 0.20~~~~p & 5.49 & 1.09 & $370.6 \pm 37.2$& 25.6 & FR\,II, X-shaped \\
\\
J0640--5353 & WISEA J064024.95--535347.1 & 0.8~~~~~~~e & 2.38 & 1.07 & $81.6 \pm 8.2$& 26.3 & FR\,II \\
J0621--5638 & DES~J062133.55--563822.9 & 0.321~~~e & 3.7? & 1.03 & $ 5.4 \pm 0.5$ $^{\dag}$& 24.2& FR\,II remn. \\
J0625--5137 $^\text{C}$ & DES--J062522.66--513750.1 & 0.6~~~~~~~e & 2.5 & 1.00 & $2.3 \pm 0.2$ & 24.5 & FR\,II remn. \\
J0639--5455 & 2MASX J06391009--5455156 & 0.142~~~p & 7.15 & 1.07 & $61.2 \pm 6.1$& 24.5 & FR\,I\\
J0615--5136 & 2MASX J06153812--5136193 & 0.126~~~p & 7.9 & 1.07 & $43.8 \pm 4.4$& 24.3 & FR\,II relic \\
\\
J0634--5644 & WISEA J063429.02-564436.0 & 1.2~~~~~~~e & 2.06 & 1.02 & $41.2 \pm 4.1$ & 26.4 & FR\,II \\
J0636--5646 $^\text{C}$ & DES J063630.11-564617.7 & 0.832~~~p & 2.32 & 1.06 & $95.6 \pm 9.6$ & 26.4 & FR\,II \\
\hline
\end{tabular}
\label{tab:RGs}
\end{table*}
\begin{table*}
\centering
\caption{Properties of diffuse radio sources discovered in the 30 deg$^2$ ASKAP field towards A3391/95. We list the spectroscopic (s) or photometric (p) redshifts. A question mark in the second column indicates uncertainty as to whether the cited object is the host of the radio source.}
\begin{tabular}{cccccccccc}
\hline
EMU ES& centre position & \multicolumn{2}{c}{redshift} & \multicolumn{2}{c}{extent} \\
Name & & $z$ & type & [arcmin] & [kpc] & type & comments \\
\hline
J0617--5201 & WISEA J061750.63--520113.8 & 0.1245 & s & 4? & 550 & cluster relics & A3385 \\
J0624--5411 & WISEA J062447.61--540939.6 ? & 0.0490 & s & 2.5? & 150? & relic or halo ? & A3395 substructure \\
J0627--5426 & ESO 161-G008 & 0.0486 & s & 3? & 170? & radio halo? & A3395E \\
J0628--5448 & ... & ? & & 3+ & 200? & relic & in A3395 ? \\
J0630--5216 & WISEA J063003.92--521634.0 & 0.0765 & p & 4? & ? & ? & A3397A ? \\
~~~~~" & WISEA J063002.67--521636.3 & 0.0874 & p & " & ? & ? & A3397B ? \\
J0638--5358 & PSZ2 G263.14-23.41& 0.2220 & s & 3 & 650 & radio halo & AS0592 \\
J0645--5413 & 2MASX J06452948--5413365 & 0.1670 & s & 5? & 850 & radio halo + NATs & A3404 \\
\hline
\end{tabular}
\label{tab:relics}
\end{table*}
\section{Conclusions}
\label{sec:conclusions}
We present 1 GHz radio observations of the pre-merging system of galaxy clusters A3391-A3395 with ASKAP/EMU. Accompanying eROSITA observations of this system yield the best large-scale X-ray characterisation of the emission bridge between A3391 and A3395 to date (Reiprich et al., submitted).
We do not detect diffuse radio emission in the X-ray bridge between these two clusters. Provided that the acceleration mechanism is Fermi-II acceleration, the non-detection implies that the turbulent velocity in this system is smaller than in the A399-401 system, possibly because of the smaller mass. This is something that can be verified with the future ATHENA X-ray mission \citep{2013arXiv1306.2307N}.
Moreover, we find a plethora of other interesting radio sources in a 30 deg$^2$ field around A3391-A3395. We identified around 200 objects extending over more than 1 arcmin. Of these, we present a list of 27 that are GRGs
with linear projected sizes of greater than 1 Mpc (see Table~\ref{tab:RGs}). This surface density is four times that found by \cite{2020A&A...635A...5D} who find about 0.2 GRGs (LLS $>1$ Mpc per deg$^2$ based on LoTSS DR1. For galaxies with LLS $>0.7$ Mpc, we obtain a sky density of $\approx 1.7$ deg$^{-2}$, three times higher than previously reported \citep{2020A&A...635A...5D}. The 1.4-GHz luminosities of our sample of GRGs are consistent with distributions found in the literature. Our sample shows no evidence in support of cosmological evolution of the GRG population.
Finally, we identify seven candidates for radio relics and halos, which we list in Table~\ref{tab:relics} together with previously published examples \citep{2020arXiv200601833W}. It will be important to follow these up with deeper observations and to make spectral index maps in order to classify the sources and understand their origin. This will be the subject of future work.
\begin{acknowledgement}
This work is based on data from eROSITA, the primary instrument aboard SRG, a joint Russian-German science mission supported by the Russian Space Agency (Roskosmos), in the interests of the Russian Academy of Sciences represented by its Space Research Institute (IKI), and the Deutsches Zentrum f\"ur Luft- und Raumfahrt (DLR). The SRG spacecraft was built by Lavochkin Association (NPOL) and its subcontractors, and is operated by NPOL with support from the Max Planck Institute for Extraterrestrial Physics (MPE).
The development and construction of the eROSITA X-ray instrument was led by MPE, with contributions from the Dr. Karl Remeis Observatory Bamberg \& ECAP (FAU Erlangen-N\"urnberg), the University of Hamburg Observatory, the Leibniz Institute for Astrophysics Potsdam (AIP), and the Institute for Astronomy and Astrophysics of the University of T\"ubingen, with the support of DLR and the Max Planck Society. The Argelander Institute for Astronomy of the University of Bonn and the Ludwig Maximilians Universit\"at Munich also participated in the science preparation for eROSITA.
The eROSITA data shown here were processed using the eSASS software system developed by the German eROSITA consortium.
The Australian SKA Pathfinder is part of the Australia Telescope National Facility which is managed by CSIRO. Operation of ASKAP is funded by the Australian Government with support from the National Collaborative Research Infrastructure Strategy. ASKAP uses the resources of the Pawsey Supercomputing Centre. Establishment of ASKAP, the Murchison Radio-astronomy Observatory and the Pawsey Supercomputing Centre are initiatives of the Australian Government, with support from the Government of Western Australia and the Science and Industry Endowment Fund. We acknowledge the Wajarri Yamatji people as the traditional owners of the Observatory site. The Australia Telescope Compact Array (/ Parkes radio telescope / Mopra radio telescope / Long Baseline Array) is part of the Australia Telescope National Facility which is funded by the Australian Government for operation as a National Facility managed by CSIRO. This paper includes archived data obtained through the Australia Telescope Online Archive (http://atoa.atnf.csiro.au). This work was supported by resources provided by the Pawsey Supercomputing Centre with funding from the Australian Government and the Government of Western Australia. We acknowledge and thank the builders of ASKAPsoft.
Support for the operation of the MWA is provided by the Australian Government (NCRIS), under a contract to Curtin University administered by Astronomy Australia Limited.
MB acknowledges support from the Deutsche Forschungsgemeinschaft under Germany's Excellence Strategy - EXC 2121 "Quantum Universe" - 390833306.
HA benefited from grant CIIC 90/2020 of Universidad de Guanajuato, Mexico. AB and DNH acknowledge support from the ERC StG DRANOEL 714245. AB acknowledges support from the MIUR grant FARE "SMS". SWD acknowledges an Australian Government Research Training Program scholarship administered through Curtin University. LR receives support from the U.S. National Science Foundation grant AST17-14205 to the University of Minnesota. The authors made use of the database CATS \citep{2005BSAO...58..118V} of the Special Astrophysical Observatory.
\end{acknowledgement}
\bibliographystyle{aa}
| 0d1cebc6354e706bb8f13c9327f07242a614f1df | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Asymptotic freeness in the large dimension limit of independent GUE random matrices is a fundamental phenomenon discovered by Voiculescu
in the nineties \cite{MR1094052}. It allows us to
understand the spectrum of non-commuting polynomials in independent
random matrices in the limit of large dimensions and under very general assumptions.
Subsequently, Voiculescu proved that independent Haar unitary random matrices are almost surely asymptotically free as the dimension goes to infinity
\cite{MR1601878}.
While the above result addresses unitary groups, i.e.~the group of symmetries of a Euclidian space, it is natural to consider other groups
of symmetries, and in particular, of a finite set, i.e., symmetric groups. Here, the counterpart of Voiculescu's results was solved by
Nica in \cite{MR1197059}.
Asymptotic freeness for random permutations notably provides fine spectral information for operators acting on a random Schreier
graph or random coverings of a fixed graph. It has also become folklore that some specific non-random permutations' asymptotic freeness can be achieved easily by explicitly exhibiting a sequence of finite quotients of a free group
whose kernels intersect with its trivial subgroup.
As soon as compact matrix groups are involved, the joint asymptotic behavior of independent random variables can legitimately be expected
to depend on which representation of the group is being considered, and
for example, although the above results are conclusive examples of asymptotic freeness, irrespective of the representation
in the case of the symmetric group, we are not aware of any complete result in this direction for unitary or orthogonal groups,
and the initial results deal instead with the very particular case of the fundamental representation.
In a recent development by
the second author, together with Gaudreau Lamarre and Male, addressed part of the problem
in \cite{MR3573218}, when a signature is fixed.
However, although asymptotic freeness describes the macroscopic spectrum of non-commuting polynomials in the generators in large dimension efficiently,
it is not enough to analyze the existence of eigenvalues away from the limiting spectrum.
The absence of such eigenvalues (also known as outliers) is a difficult problem whose positive answer corresponds to strong asymptotic freeness.
We refer to Section \ref{sec:main}
for a formal definition of strong asymptotic freeness.
The first breakthrough in this direction was achieved in \cite{MR2183281}
where the authors proved strong asymptotic freeness of independent GUE matrices, and it was followed by many improvements. A notable progress was
strong asymptotic freeness of Haar unitaries in \cite{MR3205602}.
Later, both authors in \cite{MR4024563} showed that strong asymptotic freeness holds
for random independent permutations (i.e., Haar unitaries on the symmetric group) when viewed as $n\times n$ matrices.
To continue the parallel between asymptotic freeness and strong asymptotic freeness,
let us point out here that all results on strong asymptotic freeness involving independent copies of Haar measure on groups are basically
only valid with respect to the
fundamental representation --- with the notable exception of \cite{MR4024563}, where we show that we can also extend the result
to the tensor product of two fundamental representations.
For particular operators, though, partial results have also been obtained in the context of quantum expanders,
see \cite{MR3226740,MR2553116,MR2453786,MR2486279}.
In other words, there is no hint at the fact that the same strong asymptotic freeness would hold if one were considering
all representations simultaneously --- or equivalently, for our purposes, the left regular representation.
This seems, however, to be a very natural question, as there exist sequences of permutations
(such as those of the Ramanujan graphs, \cite{MR0484767,MR963118})
for which a simple polynomial --- the sum of generators --- is known to behave well for all nontrivial irreducible representations.
This prompts us to state a question:
given an integer $n$, consider a nontrivial signature $\rho$, and the irreducible representation
$V_{n,\rho}$ of $\mathbb{U}_n$ associated with the signature.
We consider the quotient $\mathbb{U}_{n,\rho}$ of $\mathbb{U}_n$ viewed as a matrix subgroup of $\mathbb{U}(V_{n,\rho})$.
In particular, the Haar measure on $\mathbb{U}_{n,\rho}$ has `less randomness' than the Haar measure on the full unitary group
$\mathbb{U}(V_{n,\rho})$,
in the sense that the group is of a much smaller dimension --- say, as a manifold. Then the question is: \emph{for which sequence of pairs
$(n,\rho )$ is it true that $d$ independent Haar random variables of
$\mathbb{U}_{n,\rho}$ are strongly asymptotically free?}
As a trivial observation, $\dim (V_{n,\rho})$ must tend to infinity because freeness does not occur in finite dimension, so
either $n$ or $\rho$ must tend to infinity. Fixing $n$ and letting $\rho$ tend to infinity seems to be a fascinating
problem to which we do not have an answer.
Our main theorem is an answer to this theorem when $\rho$ is fixed, and $n$ tends to infinity. Our result is that
strong asymptotic freeness holds as soon as $\rho$ is nontrivial. It can be made uniform, provided
that $\rho$ varies slowly as a function of $n$.
The precise statement can be found in Corollary \ref{cor:main}.
The above corollary is a consequence of our main theorem, Theorem \ref{th:main}. This result establishes strong asymptotic freeness for
tensor products of the fundamental representation
and of the contragredient representation on the orthogonal of fixed points.
We believe the main theorem to be of independent interest in mathematical physics and quantum information theory,
but for the purpose of this introduction, let us
note that, as explained in Section \ref{sec:main}, thanks to elementary results in representation theory and a few classical facts of operator algebras,
Corollary \ref{cor:main} and Theorem \ref{th:main} are equivalent.
Our main result, Theorem \ref{th:main}, can be interpreted as a $0$-$1$ law over natural sequences of irreducible representations:
either it is trivial (if the sequence of representations is one-dimensional) or strongly asymptotically free (in all other cases).
In other words, this quantifies very precisely the fact that the sole known obstructions to strong asymptotic freeness
are the fixed points of a representation. Hence, there does not seem to be intermediate behavior between
strong asymptotic freeness and triviality.
Let us digress a bit to hint at the fact that strong asymptotic freeness can naturally be expected to be much harder to achieve than
plain asymptotic freeness.
To be more specific, let us try to investigate whether the operator norm of a non-commutative linear function of independent copies of
Haar-distributed variables could be bounded above uniformly (which must happen in the case of strong asymptotic freeness
thanks to the Haagerup inequality).
A naive attempt could consist of a combination of a net argument and a union bound --- and this attempt would succeed
in the case of the context of the fundamental representation.
For this purpose, for $\varepsilon, \eta >0$, let us observe that the size of an $\eta$-net of vectors in the unit ball of $V_{n,\rho}$
is of the order of $(C/\eta)^{\dim(V_{n,\rho}) }$, whereas for any $1$-Lipschitz function, the likelihood of
being $\varepsilon$-away from the median is of order $\exp (-c\varepsilon^2 n)$;
this follows from Gromov's comparison theorem, see \cite[Theorem 4.4.27]{MR2760897}.
The involved exponential speeds $n$ and $\dim(V_{n,\rho})$ are comparable as $n$ grows iff
$V_{n,\rho}$ is the fundamental representation or its contragradient.
In all other cases $\dim (V_{n,\rho})\gg n$ and it is impossible, given $\eta$, to fix $\varepsilon$ such that
$(C/\eta)^{\dim(V_{n,\rho}) }\cdot \exp (-c\varepsilon^2 n)$ remains small uniformly over the dimension.
Hence, we might expect that `soft' geometric techniques such as those exposed in the monograph \cite{MR3699754}
are not of use here and make it less obvious
that strong asymptotic must hold.
Likewise, analytic proofs extending \cite{MR3205602} and \cite{MR2183281} are not handy in this case because of the tensor structure
of the objects -- there is no prospect for an easy folding trick like in \cite{MR3205602} or
a Schwinger-Dyson equation like in \cite{MR2183281} because of the tensor structure.
Because of this lack of geometric or analytic methods,
a natural approach turns out to be based on moments, which is the core technology of this paper.
It is implemented with the help of an operator-valued non-backtracking theory.
The non-backtracking theory was observed by a few authors
to be very powerful in studying outliers in the context of graph theory, see, e.g., \cite{Friedman1996,bordenaveCAT}.
A version of the non-backtracking theory with non-commuting weights was developed further in \cite{MR4024563}
to prove strong asymptotic freeness of random permutations.
Here, to achieve our goal, we generalize this theory even further, beyond the case of permutation operators, to the case of all bounded operators.
The results are gathered in Section \ref{sec:SAFNB}.
However, moment methods require a good understanding of the moments of the random objects under consideration. While moment
estimates for random permutations boil down to combinatorial formulas, the situation is more involved for random unitary or
orthogonal matrices. In these cases, formulas for moments require Weingarten calculus, as developed in
\cite{MR2217291,MR1959915}.
One of our key results relies on the comparison of moments of unitary random matrices and Gaussian random matrices, which
extends the line of research of \cite{MR2257653,MR3769809} and others.
Our main result in this direction is Theorem \ref{theorem-with-brackets}, which quantifies how well unitary random matrices
can be estimated by Gaussian matrices.
A critical preliminary result to achieve this goal is to develop a systematic method to handle centering, for which we use a bracket notation.
This is based on another notion of lattices, and a similar moment theoretical work was initiated in \cite{MR2830615}.
Our main result here is Theorem \ref{lemma:unitary-Wg-expansion3}.
The paper is organized as follows. This introduction is followed by Section \ref{sec:main}, which states the main results and describes the steps
of the proof and its applications.
Section \ref{sec:Wg} contains the necessary new developments on uniform centered Weingarten functions.
Section \ref{sec:SAFNB} describes the non-commutative non-backtracking theory in full generality.
Section \ref{sec:trace} is devoted to trace estimates based on the previous sections and the completion of the proof of the main theorems.
We finish the paper with Section \ref{sec:appendix}, which is an appendix in which we provide a constructive proof of Pisier's unitary linearization trick, on which this paper relies heavily.
\paragraph{Acknowledgements}
This paper was initiated while the first author was a visiting JSPS scholar at KU in 2018, and he acknowledges the hospitality of
JSPS and of Kyoto University.
BC was supported by JSPS KAKENHI 17K18734 and 17H04823 and CB by ANR grant ANR-16-CE40-0024.
We are indebted to Narutaka Ozawa for drawing our attention to reference \cite{MR2553116}.
We also are very grateful to the anonymous referees for their efforts in improving the manuscript.
\section{Strong asymptotic freeness for random unitary tensors}
\label{sec:main}
\subsection{Main result}
In the sequel, for an integer $k \geq 1$, we set
$\INT{k} = \{1, \ldots, k \}.$
For any compact group, there exists a unique probability measure that is left- and right-invariant under translation; it is called the
normalized Haar measure. In this paper, we are interested in the case of the unitary group $\mathbb{U}_n$ and the orthogonal group $\mathbb{O}_n$.
We focus on the unitary case to keep the paper to a reasonable length. The orthogonal case boils down to technical modifications of the unitary case, which
we summarize in Subsection \ref{subsec:orthogonal-variations}.
Let $d \geq 2$ be an integer and let $U_1, \cdots,U_{d}$ be independent Haar-distributed random
unitary matrices in $\mathbb{U}_n$. For $d+1 \leq i \leq 2d$, we set $U_i = U_{i-d}^*$. We consider the involution on $\INT{2d}$, defined by
$i^* = i+d$ for $1 \leq i \leq d$ and $i^* = i-d$ for $d+1 \leq i \leq 2d$. We thus have $U_{i^*} = U_i^* $ for all $i \in \INT{2d}$.
If $M \in M_n(C)$ and $\varepsilon \in \{\cdot, -\}$, we set $M^{\varepsilon} = M$ for $\varepsilon = \cdot$ and $M^\varepsilon = \bar M$ for $\varepsilon = -$.
We fix a triple of integers $q = q_- + q_+ \geq 1$. For each $ i \in \INT{2d}$, we introduce the unitary matrix in $\mathbb{U}_{n^q}$,
\begin{equation}\label{eq:defVi}
V_i = \bar U_i^{\otimes q_-} \otimes U_i^{\otimes q_+} = U_i^{\varepsilon_1} \otimes \cdots \otimes U_i^{\varepsilon_{q}},
\end{equation}
where $(\varepsilon_1,\ldots, \varepsilon_{q } )\in \{.,-\}^ q$ is the symbolic sequence: $\varepsilon_p = -$ for $p \leq q_-$ and $\varepsilon_p = \cdot$ otherwise.
We recall the following fact:
\begin{proposition}
The average matrix $\mathbf {E} V $
is
the orthogonal projection,
denoted by $P_{H}$,
onto the vector subspace $H$ of elements invariant under (left) multiplication by
$\bar U ^ {q_-} \otimes U ^ {q_+}$ for all $U \in \mathbb{U}_n$.
\end{proposition}
Although this result is classical and its proof is elementary, it is not very well referenced in our specific context, so we include an
outline for the sake of being self-contained, see also \cite[Section III.A]{MR2553116}.
\begin{proof}[Outline of the proof]
The fact that $\mathbf {E} V $ is a projection
follows from a direct application of the fact that the distribution of a Haar-distributed element
$U$ is the same as
$\tilde U\cdot \hat U$ where $\tilde U$ and $\hat U$ are two iid copies of $U$.
The fact that it is self-adjoint follows from the fact that $U$ and $U^*$ have the same distribution.
The fact that every invariant vector is invariant under the mean is trivial.
Finally, let $x\notin H$. By definition, there exists $V_0$ such that $V_0x\ne x$. Since $V$ is a unitary operator and therefore preserves the Euclidean norm,
it follows from the strict convexity of the Euclidean norm that $\|\mathbf {E} V x\|_2 <\|x\|_2$ and therefore $x$ is not in the image of $\mathbf {E} V$, which
concludes the proof.
\end{proof}
We refer to
Subsection \ref{subsec:Wg} for a method to compute $\mathbf {E} V$. For example if the vector $(\varepsilon_1,\ldots, \varepsilon_{q } )$ is not balanced,
that is $q_- \ne q_+$ then $\mathbf {E} V = 0$. We set
\begin{equation}\label{eq:bracket}
[V_i] = V_i - \mathbf {E} V_i = V_i- P_{H}.
\end{equation}
For a fixed integer $r \geq 1$, let $(a_0, a_1, \ldots, a_{2d})$ be matrices in $M_r(\mathbb {C})$. We introduce the following matrix on
$\mathbb {C}^r \otimes \mathbb {C}^{n^q}$:
\begin{equation}\label{eq:defA}
A = a_0 \otimes 1 + \sum_{i=1}^{2d} a_i \otimes V_i,
\end{equation}
where $1$ is the identity. From what precedes, the vector space $H_r = \mathbb {C}^r \otimes H$ is an invariant subspace of $A$ and its adjoint $A^*$.
We denote by $H_r ^\perp$ the orthogonal of $H_r$ and denote by $A_{|H_r^\perp}$ the restriction of $A$ to this vector space. Our goal
is to describe the spectrum of $A_{|H_r^\perp}$ as $n$ goes to infinity.
To this end, the key observation is the following. Consider the unitary representation of the free group ${\mathbb{F}_d} $ on $ \mathbb {C}^{n^q}$ defined
by $\pi(g_i) = V_i$, where $(g_i)_{i \in \INT{2d}}$ are the free generators and their inverses: $g_{i^*} = g_i ^{-1}$. Note that
this representation $\pi$ is random and depends implicitly on $n$ and $(q_-,q_+)$. The matrix $A$ is the image by $\pi$ of the
following operator in $\ell^2({\mathbb{F}_d})$ in the (left)-group algebra on ${\mathbb{F}_d}$:
\begin{equation}\label{eq:defAfree}
A_\star = a_0 \otimes 1 + \sum_{i=1}^{2d} a_i \otimes \lambda(g_i),
\end{equation}
where $g \mapsto \lambda (g)$ is the left-regular representation (that is, left multiplication by group elements).
The main technical result of this paper is the following theorem. In the sequel $\| \cdot \|$ denotes the operator norm of an operator in
a Hilbert space: $\| T \| = \sup_{x \ne 0} \| Tx \|_2 / \|x \|_2$ where $\| x\|_2$ is the Euclidean norm in the Hilbert space.
\begin{theorem}\label{th:main}
There exists a universal constant $c >0$ such that the following holds. Let $r \geq 1$ be an integer and let $(a_0, a_1, \ldots, a_{2d})$
be matrices in $M_r(\mathbb {C})$. If $q = q(n)$ is a sequence such that $q \leq c \ln(n) / \ln (\ln (n))$ for all $n$ large enough, then with
probability one, we have
$$
\lim_{n \to \infty} \| A_{|H_r^\perp} \| = \| A_\star \|.
$$
\end{theorem}
Thanks to the ``linearization trick" of the theory of operator algebras (see Pisier \cite{MR1401692} and the monograph by
Mingo and Speicher \cite[p256]{MR3585560} for an accessible treatment), Theorem \ref{th:main} has
an apparently much more general corollary that we describe in the sequel of this subsection.
Although Pisier's linearization trick is enough for our paper, we include, for the convenience of the reader, an appendix in Section \ref{sec:appendix}, which provides a self-contained constructive proof of the linearization theorem.
A reader interested in linearization tricks might find this section of independent interest.
In the simplest case where $r=1$, all $a_i$'s are equal to one and $q_- = q_+$, then Theorem \ref{th:main} is contained
in \cite[Theorem 6]{MR2553116}. In this case, a more direct method of moments based on Schwinger-Dyson equations can be performed.
Let us start with some basic notation of representation theory.
For any $n$, let $\rho = \rho (n)$ be a signature (to lighten the notation, we omit the dependence in $n$).
More precisely, $\rho$ is a pair of Young diagrams: two pairs of non-negative integer sequences $\lambda =(\lambda_1,\lambda_2,\ldots)$ and
$\mu= (\mu_1,\mu_2,\ldots)$ satisfying the following properties:
$\mu_i \geq \mu_{i+1}$, $\lambda_i\geq \lambda_{i+1}$ and $\sum_i\lambda_i+\mu_i<\infty$.
We introduce $l(\lambda)=\max\{i, \lambda_i>0\}$
and $|\lambda |=\sum_i\lambda_i$ (and likewise for $\mu$). For notation, we refer, for example to \cite{MR0473098}.
If $n\ge l(\lambda)+l(\mu)$ we call
$V_{n,\rho}$ the Hilbert space such that the group homomorphism $\tilde \rho: \mathbb{U}_n \to \mathbb{U}(V_{n,\rho})$
is the
irreducible representation of $\mathbb{U}_n$ whose highest weights are
$$\lambda_1\ge\cdots\ge\lambda_{l(\lambda )}\ge 0\ge \cdots \ge 0
\ge -\mu_{l(\mu)}\ge\ldots \ge -\mu_{1}.$$
The signature formulation is a reformulation of the highest weight theory, but it is convenient as it allows
to define simultaneously representations for all unitary groups $\mathbb{U}_n$ with $n\ge l(\lambda)+l(\mu)$.
We call $\mathbb{U}_{n,\rho}=\tilde\rho (\mathbb{U}_n)$. It is the quotient of
$\mathbb{U}_n$ under the above representation map.
We view it as a matrix subgroup of $\mathbb{U}(V_{n,\rho})$.
Letting $U_i$ be a Haar-distributed random matrix in $\mathbb{U}_n$, we call $W_i:=\tilde\rho (U_i)$ its image under the irreducible representation map associated with $\rho$.
Since the map is surjective, $W_i$ is Haar-distributed according to $\mathbb{U}_{n,\rho}$.
Consider any operator on $\mathbb {C}^r \otimes \ell^2 ({\mathbb{F}_d})$ of the form
$$
P_\star = \sum_{g \in {\mathbb{F}_d}} a_g \otimes \lambda (g),
$$
where for all $g \in {\mathbb{F}_d}$, $a_g$ is a matrix in $M_r(\mathbb {C})$. We assume that $a_g$ is non-zero for a finite number of group elements.
In other words, $P_\star$ is a matrix-valued non-commutative polynomial.
The image of $P_\star$ by the representation $\pi$ is denoted
by $P$:
$$
P = \sum_{g \in {\mathbb{F}_d}} a_g \otimes W(g),
$$
where $W(g) = W_{i_1} \ldots W_{i_k}$ if $g = g_{i_1} \ldots g_{i_k}$. Obviously, $H_r$ and $H_r^\perp$ are again invariant subspaces
of $P$ and $P^*$. The operator $P_\star$ (and thus $P$) is self-adjoint if the following condition is met:
\begin{equation}
\label{eq:symP}
a_{g^{-1}} = a_g^* \quad \hbox{ for all } g \in {\mathbb{F}_d}.
\end{equation}
The following result is a corollary of Theorem \ref{th:main}.
\begin{corollary}\label{cor:main}
Let $r \geq 1$ be an integer, $P$ and $P_\star$ be as above, and $c$ be as in Theorem \ref{th:main}. If $q = q(n)=|\lambda |+|\mu |$
is a sequence such that
$q \leq c \ln (n) / \ln (\ln (n))$ for all $n$ large enough, then with probability one, we have
$$
\lim_{n \to \infty} \| P_{|H_r^\perp} \| = \| P_\star \|.
$$
Moreover, if \eqref{eq:symP} holds then, with probability one, the Hausdorff distance between the spectrum of $P_{|H_r^\perp}$ and
$P_\star$ goes to $0$ as $n$ goes to infinity.
\end{corollary}
The reason why
Theorem \ref{th:main} is sufficient to allow general polynomials in the above Corollary \ref{cor:main}
follows from \cite[Section 6]{MR4024563}.
This is a consequence of \cite[Proposition 6]{MR1401692}.
Details can also be found in the appendix of section \ref{sec:appendix}.
As for the fact that we can replace tensors with irreducible representation, this follows from the following two facts:
(i) contracting by an orthogonal projection does not increase the operator norm, and (ii) there exists an orthogonal projection $Q$
that commutes with all $V_i$'s such that on $\Im(Q)$, $W_i=QV_iQ$, when $q_+=|\lambda |, q_-=|\mu|$.
We refer, for example, to \cite{MR2522486} for details on the tools involved, such as Schur-Weyl duality.
Corollary \ref{cor:main} establishes the almost-sure strong asymptotic freeness of unitary representations of independent
Haar-distributed unitary matrices provided that the unitary representation is a sub-representation of a tensor product representation of
not too large dimension. We observe that this last point is critical. Indeed, the determinant is a representation of dimension $1$ (which is
a sub-representation of the tensor product representation with $q = n$), and there is no freeness in nontrivial finite-dimensional spaces.
\subsection{Overview of the proof}
We now explain the strategy behind the proof of Theorem \ref{th:main}. The strategy is the same as the one used in \cite{MR4024563}
for permutation matrices. The model in \cite{MR4024563} is, however, very different, and new technical achievements were necessary
for tensor products of unitary matrices.
First of all, from \cite{MR4024563}, it is enough to prove that, for any $\varepsilon > 0$, we have
\begin{equation}\label{eq:tbdA}
\| A_{|H_r^\perp} \| \leq \| A_\star \| + \varepsilon,
\end{equation}
with high probability. To achieve this, it is possible to restrict ourselves to self-adjoint operators. Indeed, the operator norm of an
operator $M$ is also the square root of the right-most point in the spectrum of the non-negative operator
$$
\begin{pmatrix}
0 & M \\
M^*& 0
\end{pmatrix} = E_{12} \otimes M + E_{21} \otimes M^*,
$$
where $E_{ij}$ is the $2 \times 2$ canonical matrix whose entry $(i,j)$ is equal to $1$ all other entries are equal to $0$. In particular,
if $A$ is of the form \eqref{eq:defA}, at the cost of changing $r$ in $2r$, we may assume without loss of generality that
\begin{equation}
\label{eq:symA}
a_{i^*} = a_{i}^* \quad \hbox{for all $i \in \INT{2d}$}.
\end{equation}
For such a given operator $A$, we will define a family of companion operators, denoted by $B_\mu$ and indexed by a real parameter
$\mu$, with the property that if, for all $\varepsilon >0$, with high probability, for all $\mu$, we have
\begin{equation}\label{eq:tbdB}
\rho( B_\mu) \leq \rho((B_\star)_\mu) + \varepsilon
\end{equation}
then \eqref{eq:tbdA} holds. In the above expression, $(B_\star)_\mu$ is the companion operator of $A_\star$ defined in
\eqref{eq:defAfree} and $\rho(M) = \sup \{ |\lambda| : \lambda \in \sigma(M)\}$ is the spectral radius. These operators $B_\mu$ are
called the non-backtracking operators. This will be explained in Section \ref{sec:SAFNB}. This part is an extension of
\cite{MR4024563} in a more general setting.
It is not a priori obvious why the claim \eqref{eq:tbdB} is easier to prove than claim \eqref{eq:tbdA}. A reason is that the powers
of $(B_\star)_\mu$ are much simpler to compute than the powers of $A_\star$.
In Section \ref{sec:trace}, we prove that \eqref{eq:tbdB} holds by using the expected high trace method popularized by F\"uredi and
Koml\'os \cite{MR637828} in random matrix theory. It can be summarized as follows: assume that we aim at an upper bound of the
form $\rho(M) \leq ( 1 + o (1)) \theta$ for some $\theta >0$ and $M \in M_n (\mathbb {C})$ random. We observe that for any $\ell$ integer,
$$
\rho ( M)^{2\ell} \leq \| M^{\ell} \|^2 = \| M^\ell (M^*)^\ell \| \leq {\rm tr} ( M^\ell (M^*)^\ell).
$$
Moreover, at the last step, we lose a factor at most $n$ (and typically of this order). If we can prove that
$$
{\mathbf{E}} {\rm tr} ( M^\ell (M^*)^\ell) \leq n \theta^{2\ell},
$$
then we will deduce from Markov's inequality that the probability that $\rho(M) \leq n^{1/(2\ell)} ( 1+ \delta) \theta$ is at least
$1 - (1+ \delta)^{-\ell}$. In particular, if $n^{1/(2\ell)} \to 1$, that is $\ell \gg \ln(n)$, then this last upper bound is sharp enough
for our purposes.
A usual strategy to evaluate $ {\mathbf{E}} {\rm tr} ( M^\ell (M^*)^\ell)$ is to expand the trace and the powers as the sum of product matrix entries
and then use the linearity of the expectation. We thus need to combine two ingredients: (i) a sharp upper bound on the expectation of
the product of matrix entries in terms of combinatorial properties of the entries and (ii) a counting machinery to estimate the number
of entries in the trace that have the given combinatorial properties.
In our setting, the matrix $M$ is the non-backtracking matrix $B_\mu$ and $\theta = \rho( (B_\star)_\mu)$ for a fixed\,$\mu$. Among
the difficulties, the value of $\theta$ is unknown exactly, and also, a probabilistic control of an event where all $\mu$ are considered is
needed in \eqref{eq:tbdB}. These issues were already present in \cite{MR4024563}. Here, the presence of the tensor
products will also create a significant complication in the counting arguments of ingredient (ii). An important step for ingredient (i) is performed
in Section \ref{sec:Wg}, where we prove a new estimate on the expectation of the product of a large number of entries of a Haar unitary
matrix. This will be done by developing recent results on Weingarten calculus.
\section{High order Gaussian approximation for random unitary\;matrices}
\label{sec:Wg}
This section aims to develop an efficient machinery to compare the expectation of products of entries of a Haar-distributed unitary
matrix with the average of the same product when we replace the unitary matrix with a complex Gaussian matrix. The main results of
these sections are Theorem \ref{theorem-warmup} and Theorem \ref{theorem-with-brackets} below.
\subsection{Wick calculus}
In this subsection, we recall the classical Wick formula. We then introduce a centered version which is new.
\subsubsection{Wick formula}
Let $V$ be a real vector space of real centered Gaussian variables.
It is called a real Gaussian space.
Similarly, let $W$ be a complex vector space of
complex centered Gaussian variables. It is called a complex Gaussian space.
In both cases, the addition is the regular addition of real (respectively complex) valued random variables, and likewise for the scaling.
Let $(x_i)_{i \in I}$ be iid real centered Gaussian variables indexed by a countable set $I$. Then $ \mathrm{span} ((x_i)_{i \in I})$ is a real vector space,
and any real vector space can be realized in this way.
Finally, a real Gaussian vector space comes with the scalar product $(x,y)\mapsto {\mathbf{E}}(xy)$ and
similarly, a complex Gaussian vector space comes with the Hilbert product $(x,y)\mapsto {\mathbf{E}}(\overline{x}y)$.
Wick's Theorem asserts that the scalar product is enough to recover the structure of the Gaussian space completely,
in other words, there is a one-to-one correspondence (in law) between Gaussian spaces and their Hilbert structure; see Janson \cite[Chapter 3]{MR1474726}.
Everything relies on a moment formula (which, in the case of Gaussian variables, determines the distribution).
In the real case,
$${\mathbf{E}}(x_1\ldots x_k)=\sum_{p\in P_k} {\mathbf{E}}_p(x_1,\ldots , x_k),$$
where $P_k$ is the collection of pair partitions of $\INT{k}$, typically denoted by
\begin{equation}\label{eq:notep}
p= \{ p_1 , \ldots, p_{k/2} \} = \{\{i_1,j_1\},\ldots , \{i_{k/2},j_{k/2}\}\}
\end{equation}
with $i_l<j_l$ and $i_{l}<i_{l+1}$ (obviously $P_k$ is empty when
$k$ is odd), and, under this notation,
$${\mathbf{E}}_p(x_1,\ldots , x_k)={\mathbf{E}}(x_{i_1}x_{j_1})\ldots {\mathbf{E}}(x_{i_{k/2}}x_{j_{k/2}}).$$
In the complex case,
\begin{equation}\label{eq:WickC}
{\mathbf{E}}(g_1\ldots g_k \overline h_1\ldots \overline h_k)=\sum_{\sigma\in S_k} {\mathbf{E}}_{\sigma} ( g_1,\ldots , g_k , h_1, \ldots , h_k).
\end{equation}
where $S_k$ is the permutation group on $\INT{k}$ and
$$
{\mathbf{E}}_{\sigma} ( g_1,\ldots , g_k , h_1,\ldots , h_k) = \prod_{l=1}^k {\mathbf{E}}(g_l \overline h_{\sigma(l)}).
$$
Note that the complex case can be deduced directly from the real case thanks to the real structure of complex Gaussian spaces
($S_k$ appears as a subset of $P_{2k}$ by identifying a permutation $\sigma \in S_{2k}$ with the partition $p = \{p_1, \ldots, p_{k} \}$
of $P_{2k}$ with $p_l = \{ l , k + \sigma(l) \}$).
As for the real case, the formula is trivial when the $x_i$'s are all the same (this is the formula for the moments of a Gaussian).
For the general case, after observing that both the right-hand side and the left-hand side are $k$-linear and symmetric, we conclude
by a polarization identity.
\subsubsection{Centered Wick formula}
If $X$ is a vector valued integrable random variable, we call $[X]$ the centered random variable
\begin{equation}\label{eq:defbracket}
[X] = X-{\mathbf{E}}(X).
\end{equation}
For integer $T \geq 1$, the Wick formula can be extended as follows: if $\pi = (\pi_t)_{t\in \INT{T}}$ is a partition of $\INT{k}$,
we have, for real Gaussian variables,
$$
{\mathbf{E}} \PAR{ \prod_{t=1}^T \big[ \prod_{i \in \pi_t} x_{i} \big] } = \sum_{p \in P_{ k}(\pi)} {\mathbf{E}}_p(x_1,\ldots , x_k),
$$
where $ P_{k}(\pi)$ is the subset of $P_k$ of pair partitions
$p=\{p_1,\ldots ,p_{k/2}\}$ with the following property: for each block $\pi_t $ of the partition,
there exists at least one $j$ such that the pair $p_j$ of $p$ has one element in $\pi_t$
and one element outside.
Although this formula is not mainstream in probability theory, it follows directly from
standard inclusion-exclusion-type formulas (see Lemma \ref{le:gaussiansieve} below for the complex case).
A similar formula holds for a complex Gaussian space:
$$
{\mathbf{E}} \PAR{ \prod_{t=1}^T \big[ \prod_{i \in \pi_t} g_{i} \bar h_i \big]}=\sum_{\sigma\in S_k(\pi)} {\mathbf{E}}_{\sigma} ( g_1,\ldots , g_k , h_1, \ldots , h_k),$$
where $S_k(\pi)$ is the subset of $S_k$ of permutations $\sigma$ such that for each block $\pi_t$, there exists at least one $i \in \pi_t$
such that $\sigma(i) \notin \pi_t$.
We will use these results in the proof of Theorem \ref{theorem-with-brackets}
when we compare Gaussian moments and unitary moments.
\subsection{Weingarten calculus with brackets}
\label{subsec:Wg}
\subsubsection{Unitary Weingarten formula}
In the sequel, if $x= (x_1, \ldots, x_k)$ and $y = (y_1, \ldots, y_k)$ are multi-indices in $\INT{n}^k$, for $\sigma\in S_k$ we set
$$\delta_{\sigma}(x,y) = \prod_{l=1}^k \delta_{x_l,y_{\sigma (l)}},$$
where $\delta_{i,j}$ is the usual Kronecker delta ($1$ if $i=j$ and $0$ otherwise).
Similarly, if $k$ is even and $p \in P_k$ is a pair partition,
$\delta_{p}(x)$ is a generalized Kronecker delta: namely, $\delta_{p}(x)$ takes the value $1$ if the map $i\mapsto x_i$ is constant on each block of $p$ and zero in all other cases. This is a product of $k/2$
such deltas.
We start by recalling the analogue of Wick calculus for the normalized Haar measure on the unitary group $\mathbb{U}_n$.
We refer, for example, to \cite{MR1959915} for early versions of this result.
\begin{theorem}\label{th:unitary-Wg}
Let $k,n \geq 1$ be integers and $x,y,x',y'$ in $\INT{n}^k$. There exists a function $\mathrm{Wg} (\cdot, \cdot , n ) : S_k\times S_k\to \mathbb{R}$ such that if $U = (U_{ij})$ is Haar-distributed on $\mathbb{U}_n$,
$${\mathbf{E}} \PAR{ \prod_{l=1}^k U_{x_l y_l}\overline U_{x_l'y_l'}} =
\sum_{p,q\in S_k}\delta_p(x,x')\delta_q(y,y')\mathrm{Wg}(p,q,n).$$
This function is uniquely defined iff $k\le n$; see Theorem
\ref{lemma:unitary-Wg-expansion1} for a formula.
\end{theorem}
The sequel of this subsection is devoted to providing a modern description of $\mathrm{Wg}$.
It follows from commutativity that
$\mathrm{Wg}(p,q,n)$ actually only depends on the conjugacy class of $pq^{-1}$ so in the unitary case we will
write $\mathrm{Wg}(pq^{-1},n)=\mathrm{Wg}(p,q,n)$, so
$\mathrm{Wg}$ becomes a central function on $S_k$.
Let $\sigma\in S_k$. We denote by $|\sigma| = k - \ell(\sigma)$ where $\ell(\sigma)$ is the number of disjoint cycles in the cycle
decomposition of $\sigma$. Classically, $|\sigma|$ is also the minimal number $m$ such that $\sigma$ can be written as a product
of $m$ transpositions. In particular $(-1)^{|\sigma|}$ is the signature of $\sigma$.
For $l \geq 0$, we call $P(\sigma,l)$ the collection of solutions of
\begin{equation}\label{hurwitz}
\sigma= (i_1,j_1)\cdots (i_{|\sigma|+l}, j_{|\sigma|+l}),
\end{equation}
where $j_p\le j_{p+1}$ and $i_p<j_p$. Note that $P(\sigma,l) = 0$ unless $l = 2g$ is even (since composing by a single
transposition negates the signature).
The following theorem is a combination of Theorem 2.7 and Lemma 2.8 in \cite{MR3680193} (beware that the definition of the map
$l \to P(\sigma,l)$ is shifted by $|\sigma|$ in this reference), see also \cite{MR3010693}. In the statement below and in the sequel, if $S$
is a finite set, we denote by $|S|$ its cardinal number (not to be confused with $|\sigma|$ for $\sigma \in S_k$).
\begin{theorem} \label{lemma:unitary-Wg-expansion1}
For $\sigma \in S_k$, we have the
expansion
\begin{equation} \label{eq:unitary-Wg-expansion1}
\mathrm{Wg}(\sigma ,n)= (-1)^{|\sigma|} n^{-k - |\sigma|} \sum_{g \ge 0} | P(\sigma, 2g) | n^{-2g}.
\end{equation}
This expansion is formal in the sense that $\mathrm{Wg}(\sigma ,n)$ is a rational fraction in $n$, and its
power series expansion in the neighborhood of infinity is as above.
\end{theorem}
Since the poles of $n\to \mathrm{Wg} (\sigma,n)$ are known to be in the set $\{-k+1,\ldots , k-1\}$, it follows
that the power series expansion is convergent as soon as $n\ge k$.
The following result can be found in \cite[Theorem 3.1]{MR3680193}; it is an estimate of the number of solutions of \eqref{hurwitz}.
It allows subsequently to give estimates on the Weingarten function.
\begin{proposition}\label{combinatorial-estimate}
Let $k$ be a positive integer.
For any permutation $\sigma \in S_k$ and
integer $g \geq 0$, we have
$$
(k-1)^g | P (\sigma , 0) | \leq
|P(\sigma , 2g)|\leq
\PAR{ 6 k^{7/2}}^g |P (\sigma , 0)|.
$$
\end{proposition}
For our purposes, we will need the following corollary:
\begin{corollary}\label{cor-upper-bound-unitary}
If $\sigma \in S_k$ and $12 k^{7/2} \leq n^{2}$,
$$
| \mathrm{Wg}(\sigma ,n) | \leq \PAR{ 1 + 24k^{7/2} n^{-2} } n^{-k - |\sigma|} 4^{|\sigma|}.
$$
\end{corollary}
\begin{proof}
By \cite[Corollary 2.11]{MR3010693}, we have $$| P(\sigma , 0)| = \prod_{i=1}^{\ell (\sigma)} C_{\mu_i - 1}, $$
where $C_n$ is the $n$-th Catalan number and $\mu_i$ is the length of the $i$-th cycle of $\sigma$. Since $C_n \leq 4^n$, we get
$| P(\sigma , 0)| \leq 4^{|\sigma|}$. From Proposition \ref{combinatorial-estimate}, we deduce that for any integer $g \geq 0$,
\begin{equation}\label{eq:upper-bound-unitary}
| P(\sigma , 2g)| \leq 4^{|\sigma |}\PAR{ 6k^{7/2}}^{g}.
\end{equation}
From Theorem \ref{lemma:unitary-Wg-expansion1}, we deduce that
$$
| \mathrm{Wg}(\sigma ,n) | \leq n^{-k - |\sigma|} 4^{|\sigma|} \sum_{g\geq 0} \PAR{ \frac{6 k^{7/2}}{n^2} }^g .
$$
The conclusion follows by using that $(1-x)^{-1} \leq 1 + 4 x$ for all $0 \leq x \leq 1/2$.
\end{proof}
\subsubsection{The centered case}
For a symbol $\varepsilon \in \{ \cdot,-\}$ and $z \in \mathbb {C}$, we take the notation that $z^{\varepsilon}=z$ if $\varepsilon=\cdot$ and
$z^{\varepsilon}=\overline z$ if
$\varepsilon = -$.
Our goal is to compute, for $U = (U_{ij})$ Haar-distributed on $\mathbb{U}_n$, expressions of the form
\begin{equation}\label{eq:prodbraex}
{\mathbf{E}} \PAR{\prod_{t=1}^T \big[ \prod_{l = 1}^{k_t} U_{x_{tl} y_{tl }}^{\varepsilon_{tl}} \big]},
\end{equation}
where we have used the bracket defined in \eqref{eq:defbracket} --
and then to estimate it in a useful way.
The polynomial to be integrated can be expanded into $2^T$ terms for which the Weingarten formula can be applied
each time. However, such an approach does not yield good estimates because the sum is signed,
and additional cancellations occur, which results in the
items to be summed not having the correct decay in the large dimension. On the other hand,
for a given pairing of indices, one can group the $2^T$ (signed) Weingarten functions into one single more general Weingarten function whose expansion turns out to be non-signed and, therefore, much more suitable for
more sophisticated asymptotics.
The following proposition addresses this issue.
\begin{proposition}\label{weingarten-centered}
Let $k$ be even, $\pi = (\pi_t)_{t \in \INT{T}}$ be a partition of $\INT{k}$, let $\varepsilon \in \{\cdot,-\}^k$ be a balanced sequence (in the sense that it has as many $\cdot$ than $-$) and $x, y$ in $\INT{n}^k$.
There exists a generalized Weingarten function
$\mathrm{Wg}[\pi] (p,q,n)$ such that
\begin{equation*}\label{eq:Wgcentered}
{\mathbf{E}} \PAR{\prod_{t=1}^T \big[ \prod_{i \in \pi_t} U_{x_{i} y_{i}}^{\varepsilon_{i}} \big]}
=\sum_{p,q\in P^\varepsilon_k}\delta_{p}(x)\delta_{q}(y)\mathrm{Wg}[\pi] (p,q,n),
\end{equation*}
where $P_k^\varepsilon \subset P_k$ are the pair partitions that match an $\varepsilon_i =\cdot$ with an $\varepsilon_{i'} = -$
(seen as bijections from the set of $i$'s such that $\varepsilon_i = \cdot$ to the set of $i$'s such that $\varepsilon_i = -$).
\end{proposition}
\begin{proof}
Let $A\subset \INT{T}$ and $X_1,\ldots ,X_T$ be random variables.
We introduce the following temporary notation:
$${\mathbf{E}}_A(X_1,\ldots , X_T)={\mathbf{E}}\PAR{\prod_{t \in A}X_t}\prod_{t \notin A}{\mathbf{E}}(X_t).$$
It follows from this definition that
\begin{equation}\label{eq:prodbracket}
{\mathbf{E}}([X_1]\ldots [X_T])=\sum_{A\subset \INT{T}}{\mathbf{E}}_A(X_1,\ldots , X_T)(-1)^{T - |A|}.
\end{equation}
We apply this equation to
$$X_t:=\prod_{i \in \pi_t} U_{x_{i} y_{i}}^{\varepsilon_{i}}.$$
We claim that for a given $A$, there exists a function $\mathrm{Wg}_A[\pi] (p,q,n)$ such that
$${\mathbf{E}}_A ( X_1, \ldots, X_T )
=\sum_{p,q\in P_k^\varepsilon}\delta_{p}(x)\delta_{q}(y)\mathrm{Wg}_A[\pi] (p,q,n).$$
To define $\mathrm{Wg}_{A}$ precisely, we call $\pi_A$ the partition of
$\INT{k}$ whose blocks are $\pi_{t}, t\notin A$,
and a last block that complements the previous ones (in other words, we merge all blocks in $A$ and leave the other unchanged).
From Theorem \ref{th:unitary-Wg}, this yields the following description of $\mathrm{Wg}_{A}$:
if either the permutation $p$ or the permutation $q$ fails to respect the partition
$\pi_A$ (that is, does not leave all blocks invariant), then its value
is zero. Or else, it is a product of Weingarten functions obtained by restricting the permutations over the block of $\pi_A$:
\begin{equation}\label{eq:WgA}
\mathrm{Wg}_A[\pi] (p,q,n) = \prod_{b \in \pi_A} \mathrm{Wg} (p_b,q_b,n),
\end{equation} where the product is over all blocks $b$ of $\pi_A$ and $p_b,q_b$ are the restrictions of $p,q$ to the block $b$.
The fact that the above holds follows from an application of the Weingarten formula, Theorem \ref{th:unitary-Wg}, for
each expectation factor appearing in the product.
In turn, the explicit formula we obtain for the generalized Weingarten formula becomes
\begin{equation}
\label{eq:WgU}\mathrm{Wg}[\pi] (p,q,n)= \sum_{ A\subset \INT{T}}(-1)^{T - |A|}\mathrm{Wg}_{A}[\pi] (p,q,n),
\end{equation}
and this concludes the proof.
\end{proof}
\begin{example}\normalfont
We take $T = 2$ and for $k_1 \in \INT{k}$, $\pi_1 = \INT{k_1}$, $\pi_2 = \INT{k} \backslash \INT{k_1}$. Then, if $p$ or $q$ do not
leave invariant $\pi$, we have $\mathrm{Wg}[\pi] (p,q ,n)=\mathrm{Wg}(p,q,n)$. Otherwise, we get
$\mathrm{Wg}[\pi] (p,q,n)=\mathrm{Wg}(p,q,n)-\mathrm{Wg}(p_{\pi_1},q_{\pi_1},n)\mathrm{Wg}(p_{\pi_2},q_{\pi_2},n)$, where $p_{\pi_t}$ is the restriction of
$p$ to the block $\pi_t$. \end{example}
We now give an analogue of Theorem \ref{lemma:unitary-Wg-expansion1}. We consider the setting of Proposition \ref{weingarten-centered}.
Let $\varepsilon \in \{\cdot,-\}^k$ be a balanced sequence and $\pi$ be a partition of $\INT{k}$. Recall that we identify pair partitions
$p,q \in P_k^\varepsilon$ with bijections from the $i$'s such that $\varepsilon_i = \cdot$ to the set of $i$'s such that $\varepsilon_i = -$. For
$p,q \in P^\varepsilon_k$, we call
$P[\pi] (p,q, l)$ the collection of solutions of
\begin{equation}\label{hurwitz-centered}
p= (i_1,j_1)\ldots (i_{|p q^{-1}|+l}, j_{|p q^{-1}| + l})q
\end{equation}
where $j_r\le j_{r+1}$, $i_r<j_r$
and that satisfies the following property:
the solution can be restricted to \emph{no single block} of $ \pi$ in the
sense
that
if there exists a block $b$ of $ \pi$ such that $p,q$ and each transposition $(i_r,j_r)$ leave $b$ invariant,
then the solution is not in $P[\pi] (p ,q , l)$, otherwise, the solution is in $P[\pi] (p ,q , l)$.
These notations allow us to reformulate combinatorially the centered Weingarten function used in
Proposition \ref{weingarten-centered}:
\begin{theorem} \label{lemma:unitary-Wg-expansion3} Let $k$ be even, $\pi = (\pi_t)_{t \in \INT{T}}$ be a partition of $\INT{k}$, let $\varepsilon \in \{\cdot,-\}^k$ be a balanced sequence. For all $p, q \in P^\varepsilon_k$, we have the expansion
\begin{equation} \label{eq:unitary-Wg-expansion3}
\mathrm{Wg} [\pi](p,q ,n)=(-1)^{|pq^{-1} |} n^{-k/2-|pq^{-1} | } \sum_{g \ge 0} | P[\pi](p,q, 2g)| n^{-2g}.
\end{equation}
\end{theorem}
\begin{proof}
We set $\sigma = p q^{-1}$, it is a bijection on the set $\dot{\INT{k}}$ of elements $\INT{k}$ such that $\varepsilon_i = \cdot$.
For $A \subset \INT{T}$, we start with the formula for $\mathrm{Wg}_A[\pi]$ in \eqref{eq:WgA}, where we recall that $\pi_A$ is the partition
obtained for $\pi = (\pi_t)$ by merging all blocks in $A$. We apply Theorem \ref{lemma:unitary-Wg-expansion1} to each term
$\mathrm{Wg} (p_b,q_b,n) = \mathrm{Wg} (p_b q^{-1}_b,n)$ in the product \eqref{eq:WgA}. One finds that
\begin{equation}\label{eq:WgAex}
\mathrm{Wg}_A[\pi](p,q,n) = (-1)^{|\sigma|} n^{-k/2 - |\sigma|} \sum_{g \ge 0} | P_A[\pi](p,q, 2g)| n^{-2g},
\end{equation}
where $P_{A}[\pi](p,q, 2g)$ is the empty set if either $p$ or $q$ do not respect the
partition $\pi_A$ (that is, do not leave all blocks of $\pi_A$ invariant)
and else, it is the collection of solutions of \eqref{hurwitz-centered},
$\sigma= (i_1,j_1)\ldots (i_{|\sigma| + 2g}, j_{|\sigma| + 2g})$
where $j_p\le j_{p+1}$ and $i_p<j_p$ in $\dot{\INT{k}}$ and every transposition $(i_p,j_p)$ respects the partition $\pi_A$.
This follows from the facts that the condition ``$j_p\le j_{p+1}$ and $i_p<j_p$'' is a total order on transpositions and that
all other quantities in the sum \eqref{eq:unitary-Wg-expansion1} are multiplicative over blocks of $\pi_A$.
We observe from the definitions that $P_{\INT{T}} [\pi] ( p, q, 2g) = P (pq^{-1},2g) $ and that
$$
P[\pi](p,q,2g) = P (pq^{-1},2g) \backslash \bigcup_{ t = 1 }^T P_{\INT{T}\backslash \{t\} } [\pi] (p,q,2g).
$$
Indeed, $P_{\INT{T}\backslash \{t\} } [\pi] (p,q,2g)$ is the set of solutions of \eqref{hurwitz-centered} such that $p$, $q$ and the
transpositions leave $\pi_t$ invariant. In addition, from the definition, we have for all sets $A,B$ in $\INT{T}$,
$$P_{A}[\pi](p,q, 2g )\cap P_{B}[\pi](p,q, 2g)
=P_{A\cap B}[\pi](p,q, 2g).$$
We recall that the inclusion-exclusion formula asserts that any sets $S_t \subset S$ we have
\begin{equation}\label{eq:sieve}
\left| S \backslash \bigcup_{ t = 1 }^T S_t\right| = |S| + \sum_{\emptyset \ne B \subset \INT{T}} (-1)^{|B|} \left| \bigcap_{ t \in B } S_t\right| = |S| + \sum_{A \subsetneq \INT{T}} (-1)^{T - |A|} \left| \bigcap_{ t \notin A } S_t\right|,
\end{equation}
We apply the inclusion-exclusion formula to the sets $S_t = P_{\INT{T}\backslash \{t\} } [\pi] (p,q,2g) $ and $S = P (pq^{-1},2g)$. From what precedes
$$
\bigcap_{t \notin A} P_{\INT{T}\backslash \{t\} }[\pi] (p,q,n) = P_{\bigcap_{t \notin A } \INT{T}\backslash \{t\} } [\pi] (p,q,n)= P_{A}[\pi] (p,q,n).
$$
We thus have proved that
$$P[\pi](p,q,2g) = \sum_{A\subset \INT{T}}(-1)^{T-|A|} |P_A[\pi](p,q,2g)|. $$
This completes the proof by plugging this last identity in Equation \eqref{eq:WgU} and Equation \eqref{eq:WgAex}. \end{proof}
Out of this, we are able to propose the key estimate for the centered Weingarten function. In the statement below, if $p$ and $q$ are
two partitions of $\INT{k}$ then $p \vee q$ is the finest partition coarser than both $p$ and $q$.
\begin{theorem}\label{thm:centered-wg-estimate}
Let $k$ even with $2 k^{7/2} \leq n^{2}$, $\pi = (\pi_t)_{t \in \INT{T}}$ be a partition of $\INT{k}$, let $\varepsilon \in \{\cdot,-\}^k$ be a
balanced sequence. For all $p, q \in P^\varepsilon_k$, we have the following estimate:
$$|\mathrm{Wg}[\pi](p,q ,n)|\le (1 + 3 k^{7/2} n^{-2} ) n^{-k/2-|pq^{-1} |} 4^{|pq^{-1} |} ( k^{7/4}n^{-1 })^{r},$$
where $r$ is the number of blocks of $\pi$ to which $p\vee q$ can be restricted.
\end{theorem}
Before we supply the proof, we give the main idea, which is quite simple:
since in Theorem \ref{lemma:unitary-Wg-expansion3}, we realize Weingarten functions as unsigned sums, it is enough to estimate
each summand separately,
In turn, our estimate is quite blunt and relies solely on the inclusion
$P[\pi](p,q , l) \subset P(p,q ,l) $. In other words, an estimate
good enough for our purposes is achieved just because a partial connectedness condition
kills the first terms of a series.
\begin{proof}[Proof of Theorem \ref{thm:centered-wg-estimate}]
Set $\sigma = pq ^{-1}$. From Theorem \ref{lemma:unitary-Wg-expansion3} we have
$$|\mathrm{Wg} [\pi](p,q,n)|= n^{-k/2-|\sigma|} \sum_{g \ge 0} | P[\pi](p,q, 2g) |n^{-2g}.$$
We observe that $P[\pi](p,q , 2g) \subset P(p,q , 2g)$. We also claim that $P[\pi](p,q , 2g)=\emptyset$ as soon as $2g < r $. Indeed, if
$b$ is a block of $p \vee q$, then any solution of \eqref{hurwitz-centered} with $l=0$ satisfies $i_p,j_p$ is in $b$. Hence, if a block $\pi_t$ is a union of blocks of $p \vee q$, then at least two extra transpositions from an element in $\pi_t$ to another block have to be added to
be
an admissible solution of \eqref{hurwitz-centered}. These two extra transpositions could be shared between two such blocks of $\pi$.
It follows that $P[\pi](p,q , 2g)\ne \emptyset$ implies $2g \geq r$ and
$$|\mathrm{Wg}[\pi](p,q,n)| \le n^{-k/2-|\sigma |} \sum_{g \ge r/2} | P(p, q ,2g) | n^{-2g}.$$
The right-hand side can be estimated
thanks to \eqref{eq:upper-bound-unitary} (applied to $k/2$). We get if $c = 3 / 2^{5/2}$ and $c k^{7/2}n^{-2} < 1$,
$$|\mathrm{Wg}[\pi](p,q,n)|\le n^{-k/2-|\sigma |} 4^{|\sigma |} \frac{ (c k^{7/2}n^{-2})^{ r/2}}{1-c k^{7/2}n^{-2}}.$$
If we assume further
$2 c k^{7/2}n^{-2}\le 1$, we obtain
$$|\mathrm{Wg}[\pi](p , q ,n)|\le \PAR{ 1 + \frac{4 c k^{7/2}}{n^2} } n^{-k/2-|\sigma |} 4^{|\sigma |} (c k^{7/2}n^{-2})^{r/2},$$
as required (since $c \simeq 0.53$). \end{proof}
\subsection{From Weingarten calculus to Wick calculus}
\subsubsection{Case without brackets}
We start with
$x, y$ in $\INT{n}^k$ and a balanced sequence $\varepsilon \in \{\cdot,-\}^k$.
If $U = (U_{ij})$ is Haar-distributed on $\mathbb{U}_n$, we want to compare
$|{\mathbf{E}}(U_{x_1y_1}^{\varepsilon_1}\ldots U_{x_ky_k}^{\varepsilon_k} )|$
with the matrix $U$ replaced by $G_{ij} / \sqrt n$, where $G_{ij}$ are independent complex standard Gaussian variables.
We need a new definition. Let $x,y \in \INT{n}^k$. If $u \in [n]$, we define the number of {\em left arms} of $u \in [n]$ in $(x,y)$ as
$\sum_i \IND (x_i = u)$ and the number of {\em right arms} as $\sum_i \IND ( y_i = u)$. The pair $(x,y)$ is called an {\em even sequence}
if for any $u \in [n]$, the number of left and right arms are even.
Our result is as follows:
\begin{theorem}\label{theorem-warmup}
Let $k$ be even, $x, y$ in $\INT{n}^k$ and let $\varepsilon \in \{\cdot,-\}^k$ be a balanced sequence.
If $n \geq 4$ and $2 k^{7/2} \leq n^2$ then
\begin{equation*}\label{eq:prodtw}
n^{k/2} \left| {\mathbf{E}} \PAR{ \prod_{i=1}^k U_{x_iy_i}^{\varepsilon_i}} \right|
\le \PAR{ 1 + 3 k^{7/2} n^{-2}} {\mathbf{E}} \PAR{ \prod_{i=1}^k \PAR{G_{x_iy_i}^{\varepsilon_i}+ k n^{-1/4}}}.
\end{equation*}
Moreover, the above expectation on the left-hand side is zero unless $(x,y)$ is an even sequence.
\end{theorem}
\begin{proof}
From Theorem \ref{th:unitary-Wg}, we have
\begin{equation}\label{eq:prodWgA}
n^{k/2} {\mathbf{E}} \PAR{ \prod_{i=1}^k U_{x_iy_i}^{\varepsilon_i}} = \sum_{p,q\in P^\varepsilon_k}\delta_{p}(x)\delta_{q}(y) n^{k/2} \mathrm{Wg} (p,q,n),
\end{equation}
where we have identified a pair partition in $P^\varepsilon_k$ with a bijection from the set $ \dot{\INT{k}} \subset \INT{k}$ of $i$'s such
that $\varepsilon_i = \cdot$ to the set of $i$'s such that $\varepsilon_i = -$. The final statement of the theorem follows directly.
Given a pair $(p,q)$ involved in the right-hand side of \eqref{eq:prodWgA},
we introduce a subset $A$ of $\INT{k}$ which is the maximal subset such that
$p_{A}, q_{A}$, the restrictions of $p$ and $q$ to $A$, are well defined and describe the \emph{same} pair partition. In other words,
$A$ is the union of all common pairs of $p$ and $q$.
Note that $A$ could be empty or $\INT{k}$, but it has an even number of elements.
Our strategy is to evaluate simultaneously all $(p,q)$ that yield the same $A$.
If we can find $\delta, \eta \geq 0$ such that for all $A \subset \INT{k}$, the sum in \eqref{eq:prodWgA} restricted to pairs
$(p,q)$ that yield $A$ is bounded above by
\begin{equation}\label{eq:wgA200}
( 1 + \delta ) {\mathbf{E}}\PAR{\prod_{i\in A} G_{x_iy_i}^{\varepsilon_i}}\eta^{|A^c|}
\end{equation}
then we would deduce that
$$n^{k/2} \left| {\mathbf{E}} \PAR{ \prod_{i=1}^k U_{x_iy_i}^{\varepsilon_i}} \right| \leq ( 1 + \delta) {\mathbf{E}} \PAR{ \prod_{i=1}^k \PAR{G_{x_iy_i}^{\varepsilon_i}+\eta }}
$$
since we have the following expansion:
$$
{\mathbf{E}} \PAR{ \prod_{i=1}^k \PAR{G_{x_iy_i}^{\varepsilon_i}+\eta }} = \sum_{A \subset \INT{k}} {\mathbf{E}} \PAR{ \prod_{i\in A} G_{x_iy_i}^{\varepsilon_i}}\prod_{i\in A^c}\eta .
$$
To that end, we set $c = 3 / \sqrt 2$ and $\sigma = pq^{-1}$ (seen as a bijection on $\dot{\INT{k}}$).
According to Corollary \ref{cor-upper-bound-unitary} (applied to $k/2$), if $c k ^{7/2} \leq n ^2$ and $\delta = 2 c k^{7/2} / n^2$, we have
$$n^{k/2 } \mathrm{Wg} (p,q,n)\le \PAR{ 1 + \delta} \PAR{ \frac 4 n }^{|\sigma |}.$$
It is standard that if $\tau \in S_m$, then $|\tau | = m - \sum_l c_l$ where $c_l$ is the number of cycles of length $l$.
Since $\sum l c_l = m \geq 2 (\sum_l c_l) - c_1$, we get $|\tau| \geq (m -c_1) /2 $. By assumption, all cycles of
$\sigma$ in $A^c \cap \dot{\INT{k}}$ have length at least $2$, so the number of cycles of length $1$ of $\sigma$ is at most
$|A \cap \dot{\INT{k}}| = |A|/2$. We deduce that
$$|\sigma | \geq (k - |A| ) /4 = |A^c|/4.$$
We thus have proved that, if $n \geq 4$, and $\eta_0 =( 4/n)^{1/4}$,
$$n^{k/2} \mathrm{Wg} (p,q,n)\le ( 1+ \delta) \eta_0^{|A^c|}.$$
Therefore, the sum in \eqref{eq:prodWgA} restricted to pair partitions
$p,q$ that yield $A$ is upper bounded by
$$(1+\delta) \eta_0^{|A^c|} \PAR{(|A^c|/2)! }^2 \sum_{\tau \in P^\varepsilon_{A}} \delta_\tau(x) \delta_\tau(y) , $$
where $P^\varepsilon_A$ is the set of bijections on $A$ from the set $i$'s such that $\varepsilon_i = \cdot$ to the set of $i$'s such that $\varepsilon_i = -$.
Moreover, the term $((|A^c|/2)! )^2$ accounts for the choices of $(p_{A^c}, q_{A^c})$. From Wick formula \eqref{eq:WickC}, we have
$$ {\mathbf{E}}\PAR{\prod_{i\in A} G_{x_iy_i}^{\varepsilon_i}} = \sum_{ \tau \in P^\varepsilon_{A}} \delta_\tau(x) \delta_\tau(y) .$$
Finally, we recall that $m! \leq (m e^{-1})^m$. We deduce that \eqref{eq:wgA200} holds with $\delta$ as above and
$\eta = \eta_0 e^{-1} (k/2)$. It concludes the proof.
\end{proof}
\subsubsection{Case with brackets}
We now move to the case with brackets. If $U = (U_{ij})$ is Haar-distributed on $\mathbb{U}_n$, we want to compare expectations as in
Equation \eqref{eq:prodbraex}
with the matrix $U$ replaced by $G_{ij} / \sqrt n$, where $G_{ij}$ are independent complex standard Gaussian variables. The main result
in this direction is:
\begin{theorem}\label{theorem-with-brackets}
Let $k$ even with $2 k^{7/2} \leq n^{2}$ and $n \geq 4$, $\pi = (\pi_t)_{t \in \INT{T}}$ be a partition of $\INT{k}$ such that each block has at most $\ell$ elements. Let $\varepsilon \in \{\cdot,-\}^k$ be a balanced sequence. For any $x,y$ in $\INT{n}^k$, we have
$$n^{k/2}\left| {\mathbf{E}} \PAR{\prod_{t=1}^T \big[ \prod_{i \in \pi_t} U_{x_{i} y_{i}}^{\varepsilon_{i}} \big]} \right|\le (1 + \delta) {\mathbf{E}} \PAR{\prod_{t=1}^T \PAR{ \big[ \prod_{i \in \pi_t} G_{x_{i} y_{i}}^{\varepsilon_{i}} \big] + \eta }},$$
with $\delta = 3 k^{7/2} n^{-2}$ and $\eta = 2 k^\ell n^{-1/4}$. Moreover, if each block $\pi_t$
contains an element with $\varepsilon_i = \cdot$ and another with $\varepsilon_i = -$, the same bound holds with $\eta = 2 k^\ell n^{-1/2}$.
Finally, the above expectation on the left-hand side is zero unless $(x,y)$ is an even sequence.
\end{theorem}
We start by evaluating the average of products of brackets of shifted Gaussian variables.
\begin{lemma}\label{le:gaussiansieve}
Let $k$ be even, let $\varepsilon \in \{\cdot,-\}^k$ be a balanced sequence and let $\pi = (\pi_t)_{t \in \INT{T}}$ be a partition of $\INT{k}$.
If $(g_i)_{i \in \INT{k}}$ is a complex Gaussian vector, then, for any complex number $\eta$, we have
$$
{\mathbf{E}} \PAR{\prod_{t=1}^T \big[ \prod_{i \in \pi_t} g_{i}^{\varepsilon_{i}}\big]} = \sum_{p \in P^\varepsilon (\pi)} \prod_{ (i,j) \in p} \mathbf {E} ( g_i \bar g_{j} ),
$$
where $P^\varepsilon (\pi)$ is the set (possibly empty) of pair partitions on $\INT{k}$ such that all pairs $(i,j)$ satisfy:
(i) $\varepsilon_i = \cdot$,
$\varepsilon_j = -$
or the other way around,
and (ii) for each block of $\pi$, there exists a pair $(i,j)$ in $p$ with one element in the block and the other outside.
\end{lemma}
The proof is elementary and closely related to the proof of Theorem \ref{lemma:unitary-Wg-expansion3}; therefore, we omit it.
\begin{proof}[Proof of Theorem \ref{theorem-with-brackets}]
We follow the same strategy as in the proof of Theorem \ref{theorem-warmup}. By Proposition\;\ref{weingarten-centered},
\begin{equation}\label{eq:prodWgU2}
n^{k/2} {\mathbf{E}} \PAR{\prod_{t=1}^T \big[ \prod_{i \in \pi_t} U_{x_{i} y_{i}}^{\varepsilon_{i}} \big]} = \sum_{p,q\in P^\varepsilon_k}\delta_{p}(x)\delta_{q}(y) n^{k/2} \mathrm{Wg}[\pi] (p,q,n),
\end{equation}
where as usual we have identified a pair partition in $P^\varepsilon_k$ with a bijection from the set $ \dot{\INT{k}} \subset \INT{k}$ of $i$'s
such that $\varepsilon_i = \cdot$ to the set of $i$'s such that $\varepsilon_i = -$. The last statement of the theorem follows directly.
On the other hand, for $B \subset \INT{T}$, we set $\INT{k}_B = \cup_{t \in B} \pi_t$. We define $P^\varepsilon_B$ as the set of pair partitions on
$\INT{k}_B$ whose pairs $(i,j)$ are such that $\varepsilon_i = \cdot$, $\varepsilon_j = -$ and we let $P^\varepsilon_B (\pi) \subset P^\varepsilon_B$ be the pair partitions
such that for each block of $\pi$, there exists a pair $(i,j)$ with one element in the block and the other outside. By Lemma \ref{le:gaussiansieve},
we have, for any numbers $\eta_1,\eta_2$,
\begin{eqnarray}
{\mathbf{E}} \PAR{\prod_{t \in T} \PAR{ \big[ \prod_{i \in \pi_t} (G_{x_iy_i}^{\varepsilon_{i}} ) \big] + \eta_1 + \eta_2 }} & =& \sum_{A \subset \INT{T}} \eta_2^{|A|^c} \sum_{B \subset A} \eta_1^{|B^c|} \sum_{\tau \in P^\varepsilon_B (\pi)} {\mathbf{E}} \PAR{\prod_{t \in B} \big[ \prod_{i \in \pi_t} (G_{x_iy_i}^{\varepsilon_{i}} ) \big] }\nonumber \\
& = & \sum_{A \subset \INT{T}} \eta_2^{|A|^c} \sum_{B \subset A} \eta_1^{|B^c|} \sum_{\tau \in P^\varepsilon_B (\pi)} \delta_\tau(x)\delta_\tau(y),\label{eq:prodbrag}
\end{eqnarray}
where $B^c = A \backslash B$ is the complement of $B$ in $A$.
Given a pair $(p,q)$ involved in the right-hand side of \eqref{eq:prodWgU2}, we introduce a subset $A$ of $\INT{T}$ which is the
maximal subset such that
$p_{A}, q_{A}$, the restrictions of $p$ and $q$ to $\INT{k}_A$, are well defined and describe the \emph{same} pair partition, say $\tau \in P^\varepsilon_A$.
According to Theorem \ref{thm:centered-wg-estimate}, if $\sigma = pq^{-1}$, we have
$$
n^{k/2} |\mathrm{Wg}[\pi](p,q ,n)|\le \PAR{ 1 + \delta } \PAR{\frac{4}{n}}^{|\sigma|} \eta^{r},
$$
with $\delta = 3 k^{7/2} n^{-2}$, $\eta=k^{7/4}n^{-1}$ and $r$ is the number of blocks of $\pi$ to which $p \vee q$ can be restricted.
The number of cycles of $\sigma$ of length $1$ is at most $k/2 - |A^c|/2$ (remark: it is at most $k/2 - |A^c|$ if each block contains
an element with $\varepsilon_i = \cdot$ and another with $\varepsilon_i = -$). As in the proof of Theorem \ref{theorem-warmup}, we deduce that
$|\sigma| \geq |A|^c / 4$ and, if $n \geq 4$,
$$
n^{k/2} |\mathrm{Wg}[\pi](p,q ,n)|\le \PAR{ 1 + \delta } \eta_0^{|A^c|} \eta^{r},
$$
with $\eta_0 = (4/n)^{1/4}$. We note also that $r$ is additive over the restrictions of $\pi$ to $\INT{k}_A$ and $\INT{k}_{A^c}$.
Since there are at most $((|\INT{k}_{A^c}|/2) !)^2$ choices for the restrictions of $p$ and $q$ to $\INT{k}_{A^c}$, the contribution in
\eqref{eq:prodWgU2} of the sum over all pairs $(p,q)$ which yields the same set $A$ is thus upper bounded by
\begin{equation}\label{eq:prodWgU2A}
\PAR{ 1 + \delta } \eta_0^{|A^c|} ((|\INT{k}_{A^c}|/2) !)^2 \sum_{\tau \in P^{\varepsilon}_{A}} \eta^{r(\tau)} \delta_{\tau}(x)\delta_{\tau}(y),
\end{equation}
where $r(\tau)$
is the number of blocks of $\pi$ which is a union of pairs of $\tau$.
By assumption $|\INT{k}_{A^c}| \leq \ell |A|^c$. Using $m! \leq (m e^{-1})^m$, we deduce that the expression \eqref{eq:prodWgU2A}
is upper bounded by
\begin{equation}\label{eq:prodWgU2A2}
\PAR{ 1 + \delta } \eta_2^{|A^c|} \sum_{\tau \in P^{\varepsilon}_{A}} \eta^{r(\tau)} \delta_{\tau}(x)\delta_{\tau}(y),
\end{equation}
with $\eta_2 = k^\ell n ^{-1/4}$. We now take a closer look at the sum in \eqref{eq:prodWgU2A2}. Let us call $B\subset A$
the complement in $A$ of the blocks of $\pi$ which are union of pairs of $\tau$. We have $r(\tau) = |B^c|$ where
$B^c = A \backslash B$. For a given set $B \subset \INT{k}_A$, let us call $P^\varepsilon_{A,B}$ the subset
$P^{\varepsilon}_{A}$ which yields the set $B$. By assumption, the restriction of $\tau$ in $P^\varepsilon_{A,B}$ to $B$ is a pair partition in
$P_B^\varepsilon(\pi)$. Conversely, recall that the blocks of $\pi$ have at most $\ell$ elements. For a given pair partition $\tau'$ in $P_B^\varepsilon(\pi)$,
there are at most $(\ell-1)^{|B^c|/2}$ partitions in
$P^\varepsilon_{A,B}$ whose restriction to $B$ is $\tau'$. If $\eta_1 = \ell^{\ell/2} \eta$, we thus have proved that
$$
\sum_{\tau \in P^{\varepsilon}_{A}} \eta^{r(\tau)} \delta_{\tau}(x)\delta_{\tau}(y) \leq \sum_{B \subset A} \eta_1^{|B^c|} \sum_{\tau \in P^\varepsilon_B (\pi)} \delta_\tau(x)\delta_\tau(y).
$$
Note also that $\eta_1 \leq \eta_2$ for our choice of $k$. This concludes the proof in view of Equation \eqref{eq:prodbrag}. \end{proof}
\begin{remark}\normalfont
Note in passing that the lower bound in Proposition \ref{combinatorial-estimate} can be used to show that there is some symmetry
in Theorem \ref{theorem-warmup} and Theorem
\ref{theorem-with-brackets}. Namely, it is possible in these results to swap the roles of the
Gaussian entries and of the unitary entries (the values of the constants $\delta,\eta$ need to be adjusted). This is just an esthetic comment about the sharpness of the comparison, as
we need the bound, as stated in the Theorem
\ref{theorem-with-brackets}.
\end{remark}
\subsection{Moment bounds for a product of unitaries}
We conclude this section with a corollary of Theorem \ref{theorem-with-brackets}. Let $x,y$ be two sequences in $\INT{n}^k$.
The {\em multiplicity} of $e = (a,b) \in \INT{n}^2$ is defined as $\sum_{i} \IND ( (x_{i},y_{i} ) = e )$. The set of pairs of multiplicity at
least one is the set of visited pairs $\cup_i \{ (x_i,y_i) \}$. Moreover if $\pi = (\pi_t)_{t \in \INT{T}}$ is a partition of $\INT{k}$, we say that
$\pi_t$ is an {\em isolated} block of $(x,y)$ if for all $i \in \pi_t$, for all $(x_{i},y_{i}) \ne (x_{j}, y_{j})$ for all
$j \in \INT{k} \backslash \pi_t$ (in other words, $(x_{i},y_{i})$ is of multiplicity $0$ in the sequence
$(x_{j},y_{j})_{j \in \INT{k} \backslash \pi_t}$).
\begin{corollary}\label{cor:WG2}
Let $k = q T$ even with $k^{q+1} \leq n^{1/4}$, $\pi = (\pi_t)_{t \in \INT{T}}$ be a partition of $\INT{k}$ such that each block has at least
$q$ elements. Let $\varepsilon \in \{\cdot,-\}^k$ be a balanced sequence. For any $x,y$ in $\INT{n}^k$, we have, for some universal
constant $c>0$,
$$\left| {\mathbf{E}} \PAR{\prod_{t=1}^T \big[ \prod_{i \in \pi_t} U_{x_{i} y_{i}}^{\varepsilon_{i}} \big]} \right| \leq c n^{-\frac{k}{2}} \eta^{ b + \frac {e_1} q } k^{\frac{m_4}{ 2}},$$
where $\eta = c k^{q/2} n^{-1/8}$, $e_1$ is the number of pairs of multiplicity $1$, $b$ is the number of isolated $(x_t,y_t)$ and
$m_{4}$ is the sum of multiplicities of pairs with multiplicity at least $4$.
\end{corollary}
\begin{proof}
Let $\eta,\delta$ be as in Theorem \ref{theorem-with-brackets} with $\ell =q$. From \eqref{eq:prodbrag}, we have
\begin{equation}\label{eq:corWG2}
I :=(1+\delta)^{-1} n^{k/2}\left| {\mathbf{E}} \PAR{\prod_{t=1}^T \big[ \prod_{i \in \pi_t} U_{x_{i} y_{i}}^{\varepsilon_{i}} \big]} \right| \leq \sum_{A \subset \INT{T}} \eta^{|A|^c} \sum_{\tau \in P^\varepsilon_A (\pi)} \delta_\tau(x)\delta_\tau(y).
\end{equation}
From the definition of $P^\varepsilon_A(\pi)$, we have $\delta_\tau(x)\delta_\tau(y) = 0$ if $A$ contains a pair of odd multiplicity or if
$A$ intersects an isolated block of $(x,y)$. We set $\INT{k}_A = \cup_{t \in A} \pi_t$. If $(G_{ij})$ and $Z$ are independent standard complex Gaussian variables,
$$
\sum_{\tau \in P^\varepsilon_A (\pi)} \delta_\tau(x)\delta_\tau(y) \leq \sum_{\tau \in P^\varepsilon_A} \delta_\tau(x)\delta_\tau(y) = {\mathbf{E}} \prod_{i \in \INT{k}_A} G^{\varepsilon_i}_{x_{i} y_{i}} = \prod_{(a,b) \in \INT{n}^2} \mathbf {E} [ Z^{m^ {\cdot}_{ab}(A)} \bar Z^{ m^{\bar{}}_{ab}(A)}],
$$
where $m^\varepsilon_{ab}(A)$ is the number of times that $(x_i,y_i,\varepsilon_i) = (a,b,\varepsilon)$ for $i \in \INT{k}_A$. If $m_{ab}$ is the multiplicity of
$(a,b)$ in $(x,y)$, we deduce that
$$
{\mathbf{E}} \prod_{i \in \INT{k}_A} G^{\varepsilon_i}_{x_{i} y_{i}} \leq \prod_{(a,b) \in \INT{n}^2} \mathbf {E} |Z|^{m_{ab}},
$$
From Wick formula, for even $m$, $\mathbf {E} |Z|^{m} = (m/2) ! \leq m^{m/2}$ and $\mathbf {E} |Z|^2 =1$. We deduce that
$$
{\mathbf{E}} \prod_{i \in \INT{k}_A} G^{\varepsilon_i}_{x_{i} y_{i}} \leq k^{m_4/2}.
$$
Let $T_0$ be set of $t \in T$ such that $\pi_t$ is isolated or contains a pair $(x_i,y_i)$ of multiplicity one. We set $t_0 = |T_0|$.
From \eqref{eq:corWG2}, we find
\begin{eqnarray*}
I & \leq & k^{m_4/2} \sum_{A \subset \INT{T}, A \cap T_0 = \emptyset} \eta^{|A^c|} \\
& \leq & k^{m_4/2} \sum_{s = 0}^{T-t_0}
{T -t_0 \choose s} \eta^{t_0+s} \\
&= & k^{m_4/2} \eta^{t_0} ( 1+ \eta)^{T - t_0},
\end{eqnarray*}
where we have upper bounded all possibilities of sets $A^c$ of size $t_0 +s$ in terms of its intersections with the set
$T_0$ and its complement. We get,
$$
I \leq k^{m_4/2} (1+ \eta)^{T} \eta^{t_0}.
$$
By assumption, $t_0 \geq \max(b , e_1/q) \geq (b + e_1/q)/2$. The statement of the corollary follows by using that, for $u,v \geq 0$,
$(1 + u)^v \leq e^{uv} \leq 1 + e^c uv$ if $uv \leq c$.
\end{proof}
\begin{remark}\normalfont Corollary \ref{cor:WG2} is tailored for our needs. This is not an optimal consequence of Theorem
\ref{theorem-with-brackets}. Along the lines of the proof of Corollary \ref{cor:WG2}, it is possible to obtain sharper bounds, for example,
by using the fact that the odd moments of a Gaussian random variable vanish.
\end{remark}
\subsection{The orthogonal case}
\label{subsec:orthogonal-variations}
This paper focuses on random unitary matrices; however, as we claim in the introduction, the obvious variant of our results
works precisely the same way for sequences of orthogonal groups. This subsection outlines how to adapt the above statements on the unitary group to the orthogonal case.
Firstly, there exists a counterpart of Proposition \ref{weingarten-centered}, which can be stated as follows:
\begin{proposition}\label{weingarten-centered-orthogonal}
Let $k$ be even, $\pi = (\pi_t)_{t \in \INT{T}}$ be a partition of $\INT{k}$, and $x, y$ in $\INT{n}^k$.
There exists a generalized Weingarten function
$\mathrm{Wg}_O[\pi] (p,q,n)$ such that
\begin{equation*}
{\mathbf{E}} \PAR{\prod_{t=1}^T \big[ \prod_{i \in \pi_t} O_{x_{i} y_{i}} \big]}
=\sum_{p,q\in P_k}\delta_{p}(x)\delta_{q}(y)\mathrm{Wg}_O[\pi] (p,q,n).
\end{equation*}
\end{proposition}
Intriguingly, this formula seems to be simpler. Indeed, there is no need to consider conjugates, specify matchings, or balancing
conditions since the entries $O_{ij}$ of a random orthogonal matrix are all real. Its proof is based on the existence of the orthogonal Weingarten
function and then repeats the argument of Proposition \ref{weingarten-centered}.
The subsequent estimate is the orthogonal counterpart of
Theorem \ref{thm:centered-wg-estimate}.
\begin{theorem}\label{thm:centered-wg-estimate-orthogonal}
Let $k$ even with $4 k^{7/2} \leq n^{2}$, $\pi = (\pi_t)_{t \in \INT{T}}$ be a partition of $\INT{k}$. For all $p, q \in P_k$, the following estimate holds true:
$$|\mathrm{Wg}_O[\pi](p,q ,n)|\le (1 + 12 k^{7/2} n^{-1} ) n^{-k/2-|pq^{-1} |} 4^{|pq^{-1} |} ( 2 k^{7/4}n^{-1 })^{r},$$
where $r$ is the number of blocks of $\pi$ to which $p\vee q$ can be restricted.
\end{theorem}
Without entering details, the proof is essentially the same as the proof of Theorem \ref{thm:centered-wg-estimate}. The main input
is an analogue of Theorem \ref{lemma:unitary-Wg-expansion1} and Proposition \ref{combinatorial-estimate} given in \cite{MR3680193} by
Theorem 4.6 and Theorem 4.9. We note that in \cite{MR3680193}, these two results are slightly more difficult to prove than their unitary counterpart.
This yields an orthogonal version of Theorem \ref{theorem-with-brackets} and
Corollary \ref{cor:WG2}.
\section{Strong asymptotic freeness through non-backtracking operators}
\label{sec:SAFNB}
This section aims to extend \cite[Section 3]{MR4024563} to general bounded operators in Hilbert spaces.
\subsection{Spectral mapping formulas}
We consider $(b_1, \ldots, b_\ell)$ elements in $\mathcal B(\mathcal H)$ where $\mathcal H$ is a Hilbert space. We assume that the set
$\INT{\ell}$ is endowed with an involution $i \mapsto i^{*}$.
The non-backtracking operator associated with the $\ell$-tuple of matrices
$(b_1, \ldots, b_\ell)$
is the operator on
$\mathcal B(\mathcal H \otimes \mathbb {C}^\ell)$ defined by
\begin{equation}
\label{eq:defBB}
B = \sum_{j \ne i^*} b_j \otimes E_{ij},
\end{equation}
where $E_{ij} \in M_\ell( \mathbb {R})$ are the canonical matrix elements. We also define the left non-backtracking operator as
\begin{equation}
\label{eq:defBBl}
\widetilde B = \sum_{j \ne i^*} b_i \otimes E_{ij}.
\end{equation}
Note that if the $b_i$'s are invertible, then $B$ and $\widetilde B$ are conjugate
$$
\widetilde B = D B D^{-1}
$$
with $D = \sum_i b_i \otimes E_{ii}$. The non-backtracking operators are used in this paper to give an alternative description of the
spectrum of an operator of the form \eqref{eq:defA}. We start with a result in the reverse direction. In the sequel for shorter notation,
the identity operator in $\mathcal H$ is denoted by $1$, and if $a$ is in $\mathcal B(\mathcal H)$ and $\lambda \in \mathbb {C}$, we write
$a- \lambda$ in place of $a - \lambda 1$.
\begin{proposition}\label{prop:nonbackgeneral}
Let $B$ be as above and let $\lambda \in \mathbb {C}$ satisfy
$\lambda^2 \notin \{ \sigma( b_i b_{i^*} ): i \in \INT{\ell} \}$. Define the operator $A^{(\lambda)}$ on $\mathcal H$ through
\begin{equation} \label{def_AM0}
A^{(\lambda)} = b_0 ( \lambda) + \sum_{i=1}^\ell b_i (\lambda) \,, \qquad b_i (\lambda) = \lambda b_{i} ( \lambda^2 - b_{i^*} b_{i} )^{-1} \quad \mathrm{ and } \quad b_0 ( \lambda) = - 1 - \sum_{i=1}^\ell b_{i}( \lambda^2 - b_{i^*} b_i )^{-1} b_{i^*} .
\end{equation}
Then $\lambda \in \sigma (B)$ if and only if $0 \in \sigma ( A^{(\lambda)})$. Similarly, $\lambda \in \sigma (\widetilde B)$ if and only if
$0 \in \sigma ( A^{(\lambda)})$.
\end{proposition}
\begin{proof} Let us assume that $\lambda$ is in the point spectrum of $B$, i.e. there is a non-zero vector
$v \in \mathcal H \otimes \mathbb {C}^\ell$ such that $B v = \lambda v$. Let $E_i, i\in \INT{\ell}$ be the canonical basis of $\mathbb {C}^\ell$.
We can write $v=\sum_{i\ }v_i\otimes E_i$.
We define a vector $u\in \mathcal H$ by
\begin{equation}\label{000}
u=\sum_{j } b_j v_j.
\end{equation}
Let us show that
$u$ is a non-zero vector $u\in \mathcal H$ such that $A^{(\lambda)} u=0$.
Component-wise, the equation $B v = \lambda v$ can be written
\begin{equation}\label{111}
\lambda v_i= \sum_{j \ne i^*} b_j v_j=
u-b_{i^*} v_{i^*}.
\end{equation}
In the above equation, if we replace $i$ by $i^*$
we get
\begin{equation}\label{222}
\lambda v_{i^*}=u-b_i v_i.
\end{equation}
Multiplying Equation \eqref{111} by $\lambda$ and substituting the last term of its right-hand side with
Equation \eqref{222} multiplied by $b_{i^*}$, we get
$$\lambda^2 v_i=
\lambda u-b_{i^*} u+b_{i^*}b_i v_i .$$
This can be rewritten as
$$(\lambda^2 -b_{i^*}b_i) v_i= (\lambda -b_{i^*}) u .$$
Since $\lambda^2-a_{i^*}a_i$ is invertible, we get, for all $i\in \INT{\ell}$,
\begin{equation}\label{333}
v_i=(\lambda^2 -b_{i^*}b_i)^{-1}(\lambda -b_{i^*}) u .
\end{equation}
If $u$ were the zero vector, then all $v_i$ would be zero. Therefore $v$ would be zero, which contradicts that $v$ is an eigenvector.
This proves that $u$ is not zero.
Let us now prove that $A^{(\lambda)} u=0$, i.e. its kernel is nontrivial.
According to the definition of $A^{(\lambda)}$, we need to prove that
$$u=\sum_{i} \PAR{ -b_{i}( \lambda^2 - b_{i^*} b_i )^{-1} b_{i^*} +
\lambda b_{i} ( \lambda^2 - b_{i^*} b_{i} )^{-1} } u.$$
The right-hand side is equal to
$$\sum_{i}
b_{i}
(\lambda^2 -b_{i^*}b_i)^{-1}(\lambda -b_{i^*}) u.$$
Substituting with the help of Equation \eqref{333} for each $i$, this is equal to
$\sum_{i}
b_{i} v_i,$
which completes the claim from the definition of $u$ in \eqref{000}.
Conversely, if $0$ is in the point spectrum of $A^{(\lambda)} $ with eigenvector $u$,
we define $v_i$ with Equation \eqref{333}. However, the above computations imply that $Bv = \lambda v$ and
$u=\sum_{j}b_jv_j$ as per the original definition of $u$, therefore $u$ would be zero,
if $v$ is zero
which contradicts the assumption. We thus have proved so far that $0$ is in the discrete spectrum of $A^{(\lambda)}$ if and only if
$\lambda$ is in the point spectrum of $B$.
The same equivalence holds for $\widetilde B$. Indeed, if $\widetilde B v = \lambda v$ then we define $u = \sum_{i } v_i$. Repeating
the above computation, we find $v_i = b_i ( \lambda^2 - b_i b_{i^*} )^{-1}( \lambda - b_i) u$ and $A^{(\lambda)} u = 0$. Conversely,
if $A^{(\lambda)} u = 0$, then we define $ v = \sum_i v_i \otimes E_i$ with $v_i = a_i ( \lambda^2 - a_i a_{i^*} )^{-1} ( \lambda - a_i) u $.
We get $Bv = \lambda v$. Arguing as above, the equivalence follows.
Next, we handle the continuous spectrum.
If $\lambda$ is in the continuous spectrum of $B$, then
there exists a sequence of unit vectors $v^{(n)}\in \mathcal H \otimes \mathbb {C}^\ell$ such that
$\|\lambda v^{(n)}-B v^{(n)}\|_2\to 0$ (see for example Kowalski \cite[p. 19]{kowalskiLN}) .
To each $v^{(n)}$ we associate a vector $u^{(n)}\in \mathcal H$
thanks to Equation \eqref{000}.
In the first part of the proof, all equations are continuous.
Specifically, Equations \eqref{111}-\eqref{333} remain correct if the right-hand side is perturbed additively
by a vector $\varepsilon_n$
whose norm goes to $0$ as $n\to\infty$.
This implies that $\|A^{(\lambda)} u^{(n)}\|_2\to 0$, i.e. $0$ is in the spectrum of $A^{(\lambda)}$.
Conversely, if $0$ is in the continuous spectrum of $A^{(\lambda)}$, let $u^{(n)}\in \mathcal H \otimes \mathbb {C}^\ell$ such that
$\|A^{(\lambda)} u^{(n)}\|_2\to 0$ as $n\to\infty$.
We define $v^{(n)}$ through equation \eqref{333}. As above, it is possible to prove that
$\|\lambda v^{(n)}-B v^{(n)}\|_2\to 0$ as $n\to\infty$, so $\lambda$ is in the spectrum of $B$.
Finally, we handle the residual spectrum. Let $v \in \mathcal H \otimes \mathbb {C}^\ell$ non-zero that is not an element in the closure of the image of
$B - \lambda 1$, and without loss of generality
let us assume that $v$ is orthogonal to the image of $B-\lambda 1$. Then, for all
$x \in \mathcal H \otimes \mathbb {C}^\ell$, $\langle Bx-\lambda x,v\rangle =0$.
This implies that $B^* v=\overline\lambda v$, that is, $\overline\lambda$ is in the point spectrum of $B^*$. Now, observe that
$$
B^* = \sum_{i \ne j ^*} b^*_{j} \otimes E_{ji} = \sum_{i \ne j ^*} b^*_{i} \otimes E_{ij} = \sum_{i \ne j ^*} \tilde b_i \otimes E_{ij},
$$
with $\tilde b_i = b_i^*$. In particular, $B^*$ is a left non-backtracking operator as defined in \eqref{eq:defBBl} with underlying operators
the $\tilde b_i$'s. From what precedes, $\overline\lambda$ is in the point spectrum of $B^*$ if and only if $0$ is in the point spectrum
of $\widetilde A^{(\lambda)} = \tilde b_0 + \sum_i \tilde b_i $ defined by, for $i \in \INT{\ell}$,
$$
\tilde b_i = \overline \lambda b^*_{i} ( \overline \lambda^2 - b^*_{i^*} b^*_{i} )^{-1} \quad \mathrm{ and } \quad \tilde b_0 = - 1 - \sum_{i=1}^\ell b^*_{i}( \overline \lambda^2 - b^*_{i^*} b^*_i )^{-1} b^*_{i^*} .
$$
It straightforward to check that $\tilde b_i = b_i (\lambda) ^*$ for each $i \in \INT{\ell}$. Thus, we have
$\widetilde A^{(\lambda)} = A^{(\lambda)*} $. From what precedes, we get that $\overline\lambda$ is in the discrete spectrum of $B^*$
if and only if $0$ is in the discrete spectrum of $A^{(\lambda)*}$. Hence, we have proved that if $\lambda$ is in the residual spectrum
of $B$, then $0$ is in the spectrum of $A^{(\lambda)}$ (since for any bounded operator $T$ on a Hilbert space, $\sigma(T^*) $ is the
complex conjugate of the set $\sigma(T)$, see Reed-Simon \cite[Theorem VI.7]{MR0493421}).
Conversely, if $0$ is in the residual spectrum of $A^{(\lambda)}$, then $0$ is in the discrete spectrum of $A^{(\lambda)*}$. From what
precedes, we get that $\overline \lambda$ is in the discrete spectrum of $B^*$. In particular, $\lambda$ is in the spectrum of $B$.
The proposition is proved.\end{proof}
We now consider the situation where the Hilbert space $\mathcal H$ is of the form $\mathbb {C}^r \otimes \mathcal K$, where $\mathcal K$
is a Hilbert space and
$$
b_i=a_i\otimes V_i
$$
with $a_i \in M_r(\mathbb {C})$ and $V_i$ unitary operator on $\mathcal K$ such that for all $i \in \INT{\ell}$,
$$
V_{i^*} = V_i^*.
$$
Moreover, we assume that $\ell = 2 d$ and the involution $i^*$ is as in Section \ref{sec:main}. In this specific case, we have
\begin{equation}\label{eq:defB0}
B = \sum_{i \ne j^*} a_j \otimes V_j \otimes E_{ij} \quad \mathrm{ and } \quad \widetilde B = \sum_{i \ne j^*} a_i \otimes V_i \otimes E_{ij}.
\end{equation}
The operator $A^{(\lambda)}$ in \eqref{def_AM0} is given by
\begin{equation} \label{def_AMU}
A^{(\lambda)} = a_0 ( \lambda) \otimes 1 + \sum_{i=1}^{2d} a_i (\lambda) \otimes V_i
\end{equation}
with
$$\qquad a_i (\lambda) = \lambda a_{i} ( \lambda^2 - a_{i^*} a_{i} )^{-1} \quad \mathrm{ and } \quad a_0 ( \lambda) = - 1 - \sum_{i=1}^{2d} a_{i}( \lambda^2 - a_{i^*} a_i )^{-1} a_{i^*}.
$$
If $\mathcal K=\ell^2(X)$ with $X$ countable and $V_i$ a permutation operator,
then Proposition \ref{prop:nonbackgeneral} recovers Proposition 9 of \cite{MR4024563} up to the minor point that
$B$ is slightly modified into $B = \sum_{j \ne i^*} a_j\otimes S_i \otimes E_{ij}$, but it is easy to check that
both $B$ are conjugate to each other, so they have the same spectrum.
It follows that all results of \cite[Section 3]{MR4024563} can be extended to this more general setting.
We now review two key results \cite[Section 3]{MR4024563} that will be used in the sequel. We take $\ell = 2d$. We start with a kind of converse of
Proposition \ref{prop:nonbackgeneral}, in the sense that for a given self-adjoint operator $A$ of the form
\begin{equation}\label{eq:defA0}
A = a_0 \otimes 1 + \sum_{i=1}^{2d} a_i \otimes V_i,
\end{equation}
satisfying the symmetry condition \eqref{eq:symA} and a real number $\mu$, we look for a non-backtracking operator $B_\mu$ which detects if $\mu \in \sigma(A)$. To perform this, we need to introduce the resolvent of the operator $A_\star$ defined in \eqref{eq:defAfree}. For
$\mu \notin \sigma( A_\star)$, we set
$$
G(\mu) = ( \mu - A_\star )^{-1}.
$$
For $x,y\in {\mathbb{F}_d}$, we define $G_{xy} (\mu) \in M_r(\mathbb {C})$ as the matrix $P_x G(\mu) P^*_y$ where $P_x$ is the projection onto the vector
space $\mathbb {C}^r \otimes \delta_x$. We denote by $o$ the neutral element of ${\mathbb{F}_d}$ for its group structure and $g_i$ the $i$-th generator of
${\mathbb{F}_d}$. Finally, if $D$ is a bounded set in $\mathbb {C}$, the convex hull of $D$ is denoted by $\mathrm{hull}( D ) $.
\begin{proposition}\label{prop:nonback2}
Let $A$ be as in \eqref{eq:defA0} satisfying \eqref{eq:symA} and $\mu \notin \mathrm{hull}( \sigma ( A_\star))$. Define the matrices for $i \in \INT{2d}$,
\begin{equation*} \label{def_AM}
\hat a_i (\mu) = G_{oo} ( \mu )^{-1} G_{o g_i} ( \mu ).
\end{equation*}
Let $B_\mu = \sum_{i \ne j^*} \hat a_j (\mu)\otimes V_j \otimes E_{ij} $ be the corresponding non-backtracking operator. Then
$\mu \notin \sigma (A)$ if and only if
$1 \notin \sigma ( B_\mu)$. The same statement holds for the corresponding left non-backtracking operator $\widetilde B_\mu$.
\end{proposition}
This statement is \cite[Proposition 10]{MR4024563} extended for a general Hilbert space and general unitaries $V_i$'s. The same proof
applies in this case thanks to Proposition \ref{prop:nonbackgeneral}.
\subsection{Spectral radius of non-backtracking operators}
We conclude this section with a sharp criterion to guarantee in terms of non-backtracking
operators that the spectrum of an operator $A$ is in a neighborhood of the spectrum of the operator $A_\star$. The following result
is \cite[Theorem 12]{MR4024563}. Again, thanks to the improvement of Proposition \ref{prop:nonbackgeneral}, we can now state it in a
more general context.
\begin{theorem}\label{prop:edgeAB}
Let $A$ be as in \eqref{eq:defA0} satisfying \eqref{eq:symA} and $A_\star$ the corresponding free operator defined by
\eqref{eq:defAfree}. The following two results hold:
\begin{enumerate}[(i)]
\item
For any $\mu \notin \mathrm{hull} ( \sigma (A_\star) ) $, we have $\rho( (B_\star)_\mu ) < 1$,
where $(B_\star)_\mu = \sum_{i \ne j^*} \hat a_j (\mu)\otimes \lambda (g_j) \otimes E_{ij} $ is the non-backtracking operator on ${\mathbb{F}_d}$
with weights as in Proposition \ref{prop:nonback2}.
\item
For any $\varepsilon >0$, there exists $\delta > 0$ such that if for all real $\mu$ at distance at least $\varepsilon$ from $\mathrm{hull}( \sigma (A_\star) )$,
$$
\rho(B_\mu) < \rho( (B_\star)_\mu ) + \delta,
$$
then $\mathrm{hull}(\sigma(A))$ is in an $\varepsilon$-neighbourhood of $\mathrm{hull}( \sigma (A_\star) )$.
\end{enumerate}
Moreover,
the same holds for the left non-backtracking operators.
\end{theorem}
The theorem is a simple consequence of the following two claims:
$$
\hbox{for any $(b_1,\ldots,b_{2d})$ in $M_r(\mathbb {C})^{2d}$, } \quad \rho(B_\star) = \sup \{ \lambda \geq 0 : \lambda \in \sigma(B_\star) \},
$$
and
\begin{equation}\label{eq:src}
\hbox{the map from $M_r(\mathbb {C})^{2d}$ to $\mathbb {R}$: $(b_1,\ldots,b_{2d}) \mapsto \rho(B_\star)$ is continuous.}
\end{equation}
We refer to \cite{MR4024563} for details.
\section{Expected high trace of non-backtracking matrices}
\label{sec:trace}
In this section, we prove Theorem \ref{th:main} by computing the spectral radius of the non-backtracking matrices associated to the
random matrix $A$ defined in \eqref{eq:defA}.
\subsection{Main technical statement}
We now come back to the setting of Theorem \ref{th:main}. Let $U_1, \ldots,U_d$ be independent Haar-distributed random unitary
matrices in $\mathbb{U}_n$. We define $V_i$ as in \eqref{eq:defVi} and its centered version $[V_i]$ as in \eqref{eq:bracket}.
The left non-backtracking operator associated to the weights $(a_1 \otimes [V_1],\ldots,a_{2d}\otimes [V_{2d}])$ is
\begin{equation}\label{eq:defBV2}
B = \sum_{i \ne j^*} a_i \otimes [V_i] \otimes E_{ij}.
\end{equation}
Note that we have omitted the tilde in \eqref{eq:defBBl} for shorter notation. The main technical result of this section is an upper bound
on the spectral radius of $B$ in terms of the spectral radius of the corresponding operator on the free group ${\mathbb{F}_d}$:
\begin{equation}\label{eq:defBstar}
B_\star = \sum_{i\ne j^*} a_i \otimes \lambda(g_i) \otimes E_{ij},
\end{equation}
where $\lambda(g)$ is the left-regular representation of $g \in {\mathbb{F}_d}$. Let $\rho (B_\star)$ be the spectral radius of the $B_\star$.
Recall that $\rho(B) \leq \| B^\ell \|^{1/\ell}$ for all integer $\ell \geq 1$ and the sequence $\| B^\ell \|$ is sub-multiplicative in $\ell$.
\begin{theorem}\label{th:FKB}
There exists a universal constant $c >0$ such that if $q \leq c \ln n / \ln \ln n $, then the following holds.
For any $\varepsilon >0$, there exists a constant $C \geq 1$ (depending on $\varepsilon$, $d$ and $r$) such that for any
$(a_1,\ldots,a_{2d}) \in M_r(\mathbb {C})$ with $\max_i \| a_i \| \leq 1$, if $\ell = \lfloor C q \rfloor$, the event
$$
\| B^\ell \|^{1 / \ell} \leq \rho (B_\star) + \varepsilon
$$
holds with probability at least $ 1 - C \exp \PAR{-(\ln n)^2}$.
\end{theorem}
In the following subsection, we prove Theorem \ref{th:FKB}. In the last subsection, we deduce Theorem \ref{th:main} from
Theorem \ref{th:FKB}.
\subsection{Proof of Theorem \ref{th:FKB}}
In the sequel, an entry of the matrix $[V_i]$ will be denoted by
$$[V_i]_{x y} = [\prod_{p=1}^q U^{\varepsilon_{q}}_i]_{x_p y_p},$$ with $ x = (x_1, \ldots, x_q) \in \INT{n}^q$ and
$y = (y_1, \ldots, y_q)\in \INT{n}^q$. We also set
$$
\vec E = \INT{n}^q \times \INT{2d}.
$$
In analogy with usual non-bactracking matrices, an element $e = (x,i)$ of $\vec E$ is thought as the directed edge attached to $x$
associated to the $i$-th unitary matrix $V_i$. If $e = (x,i), f = (y,j) \in \vec E$, the matrix-valued entry $(e,f)$ of $B$ defined
in \eqref{eq:defBV2} is
\begin{equation}\label{eq:defBV20}
B_{ef} = a_i [V_i]_{xy} \IND_{i \ne j^*} \in M_r(\mathbb {C}) .
\end{equation}
We start with the Weyl formula for the spectral radius. If $o$ is the unit of $F_{d}$, we set
$$
\rho_k = \PAR{ \max \| B_\star^k \varphi\otimes \delta_e \| }^{\frac {1 }{k }},
$$
where the maximum is over all $e = (o,i), i \in \INT{2d}$ and $\varphi \in \mathbb {C}^r$ of unit Euclidean norm. From the Weyl formula, $\rho_k$ converges
to $\rho( B_\star)$ as $k$ goes to infinity (see for example \cite[Theorem 1.3.6]{MR1074574}). In the remainder of the proof, we fix
$\varepsilon >0$, then there exists an integer $k_0$ large enough such that for all $k \geq k_0$,
\begin{equation}\label{eq:defrho}
\rho_k \leq \rho = \rho (B_\star) + \varepsilon.
\end{equation}
Let $\ell$ be an integer. We fix an integer $\theta$. To get a good probabilistic estimate, we will upper bound the
expectation of $\NRM{ B ^ {\ell}}^{2\theta}$ for $\theta$ large enough. We write
$$
\NRM{ B ^ {\ell}}^{2\theta} = \NRM{ B ^ {\ell} ( B ^ {\ell })^* }^{ \theta } = \NRM{ \PAR{ B ^ {\ell} (B ^ {\ell })^* }^{\theta}}.
$$
In particular, if ${\rm tr}_{\mathbb {C}^r}$ denotes the
partial
trace on $\mathbb {C}^r$ for operators on $\mathbb {C}^r \otimes \mathbb {C}^{\vec E}$, we get
$$
\NRM{ B ^ {\ell}}^{2\theta}\leq {\rm tr} \BRA{ {\rm tr}_{\mathbb {C}^r} \BRA{ \PAR{ B ^ {\ell} (B ^ {\ell })^* }^{\theta} }}.
$$
We expand the trace in terms of the matrix-valued entries of $B$:
\begin{eqnarray*}
{\rm tr}_{\mathbb {C}^r} \BRA{ \PAR{ B ^ {\ell} (B ^ {\ell })^* }^{\theta} } &=& \sum_{e_\alpha \in \vec E , \alpha \in \INT{2\theta}} \prod_{\alpha=1}^{\theta} (B ^ { \ell} )_{e_{2\alpha-1} e_{2\alpha}} ((B ^ { \ell} )^ *)_{e_{2\alpha}e_{2\alpha+1}},
\end{eqnarray*}
with $e_{2\theta +1} =e_1$. We expand further and use that $(B^*)_{ef} = (B_{fe})^*$, we obtain
\begin{eqnarray*}
{\rm tr}_{\mathbb {C}^r} \BRA{ \PAR{ B ^ {\ell} (B ^ {\ell })^* }^{\theta} } & =& \sum_{\gamma} \prod_{\alpha=1}^{2 \theta} \prod_{t=1}^\ell B^{\varepsilon_\alpha}_{\gamma^\alpha_{t} \gamma^\alpha_{t+1}} ,
\end{eqnarray*}
with $B_{ef}^{\varepsilon_\alpha}$ is equal to $B_{ef}$ for odd $\alpha$ and $(B_{fe})^*$ for even $\alpha$ and the sum is over all
$\gamma = (\gamma^1,\ldots,\gamma^{2\theta})$, $\gamma^\alpha = (\gamma^\alpha_1,\cdots,\gamma^\alpha_{\ell+1})$ in
$\vec E^{\ell+1}$ with the boundary conditions, for all $\alpha \in \INT{\theta}$, \begin{equation}\label{eq:bdcond}
\gamma^{2\alpha}_1 = \gamma^{2\alpha+1}_1 \quad \mathrm{ and } \quad \gamma^{2\alpha}_{\ell+1} = \gamma^{2\alpha-1}_{\ell+1},
\end{equation}
with the convention $\gamma^{2\theta +1} = \gamma^1$
Let us write
$\gamma^\alpha_t = (x^\alpha_t,i^\alpha_t)$, $x^\alpha_t \in \INT{n}^q$ and $i^\alpha_t \in \INT{2d}$. From \eqref{eq:defBV20}, we find
$$
{\rm tr}_{\mathbb {C}^r} \BRA{ B ^ { \ell} (B ^ { \ell} )^* } = \sum _{\gamma \in P_{\ell,\theta}} \prod_{\alpha=1}^{2\theta} \prod_{t=1}^\ell a_{i^\alpha_t}^{\varepsilon_\alpha} [V_{i^\alpha_t}]^{\varepsilon_\alpha}_{x^\alpha_t x^\alpha_{t+1}} ,
$$
where $(a^{\varepsilon_\alpha}_i,[V_{i}]_{xy}^{\varepsilon_\alpha})$ is equal to $(a_i, [V_i]_{xy})$ or $(a_i^* ,[\bar V_i]_{xy})$ depending on the
parity of $\alpha$ and $P_{\ell,\theta}$ is the set of all $\gamma = (\gamma^1,\ldots ,\gamma^{2\theta})$ as above which are also
non-backtracking, that is, for all $t \in \INT{\ell}$ and $\alpha \in \INT{2\theta}$,
\begin{equation}\label{eq:NBi}
i^\alpha_{t+1} \ne {i^{\alpha}_t}^*.
\end{equation}
(Note that $[\bar V_i]_{xy} = [V_{i^*}]_{yx}$ by construction).
We finally take the expectation and find:
\begin{equation}\label{eq:tr1}
\mathbf {E} \NRM{ B ^ {\ell}}^{2\theta} \leq r \sum_{\gamma \in P_{\ell,\theta}} a(\gamma) p(\gamma) ,
\end{equation}
where we have introduced the algebraic and probabilist weights of an element $\gamma \in P_\ell$ defined as
$$
a(\gamma) = \prod_{\alpha = 1}^{2\theta} \NRM{ \prod_{t=1}^\ell a^{\varepsilon_\alpha}_{i^\alpha_t} } \quad \hbox{ and } \quad p(\gamma) = \left| \mathbf {E} \BRA{ \prod_{\alpha = 1}^{2 \theta}\prod_{t=1}^\ell [V_{i^\alpha_t}]^{\varepsilon_\alpha}_{x^\alpha_t x^\alpha_{t+1}}} \right|.
$$
We organize the terms on the right-hand side of \eqref{eq:tr1} by introducing the following equivalence class on elements of
$P_{\ell,\theta}$.
We write $\gamma \sim \gamma'$ if there exist a permutation $\tau$ on $\INT{n}$ and a family of permutations $(\beta_x)_{x \in \INT{n}}$ on $\INT{2d}$ such that the image of $\gamma = (x^\alpha_{t,p},i^\alpha_{t,p})$ by these permutations is $\gamma' = ((x^\alpha_{t,p})',(i^\alpha_{t,p})')$. More precisely, if for all $\alpha \in \INT{2\theta}$, $t \in \INT{\ell}$ and $p \in \INT{q}$, $\beta_{x^\alpha_{t,p}}(i_t^\alpha) = (i_t^\alpha)' = \beta_{x^\alpha_{t+1,p}} ((i_t^\alpha )^*)^*$ and $\tau(x^\alpha_{t,p}) = (x^\alpha_{t,p})'$.
In more combinatorial language, if
$P_{\ell,\theta}$ is seen as a collection of labeled paths where the labels are the vertex entries in $\INT{n}$ and the edge colors in $\INT{2d}$; an equivalence class is the corresponding unlabeled path.
We will use colored graphs defined formally as follows.
\begin{definition}[Colored edge and graph]\label{def:colorgraph}
Let $X$ be a countable set and $C$ a countable set with an involution $* : C \to C$. A {\em colored edge} $[x,i,y] $ is an equivalence
class of $X \times C \times X$ equipped with the equivalence relation $(x,i,y) \sim (y,i^*,x)$. A {\em colored graph} $G = (V,E)$ is the
pair formed by a vertex set $V \subset X$ and a set of colored edges $[x,i,y]$ with $x,y \in V$. The {\em degree} of $v \in V$ is
$\sum_{e = [x,i,y] \in E} ( \IND ( v = x ) + \IND ( v = y) )$.
\end{definition}
We now define a colored graph naturally associated with an element $\gamma \in P_{\ell,\theta}$. We write
$x^{\alpha}_t = (x^{\alpha}_{t,1}, \ldots,x^{\alpha}_{t,q}) \in \INT{n}^q$ and define the colored graph
$G_\gamma = (V_\gamma,E_\gamma)$ with color set $\INT{2d} $,
$V_\gamma = \{ x^\alpha_{t,p} : \alpha \in \INT{2\theta}, t \in \INT{\ell+1}, p \in \INT{q}\} \subset [n]$ and
$E_\gamma = \{ [ x^\alpha_{t,p} ,i^\alpha_t, x^\alpha_{t+1,p}] : \alpha \in \INT{2\theta}, t \in \INT{\ell} , p \in \INT{q} \}$.
The {\em multiplicity} of an edge $e \in E_\gamma$ is defined as
$m(e) = \sum_{t\in \INT{\ell},\alpha \in \INT{2\theta},p\in \INT{q}} \IND ( [ x^\alpha_{t,p} ,i^\alpha_t, x^\alpha_{t+1,p}] = e )$.
If $e = |E_\gamma|$, $e_1$ is the number of edges of multiplicity one and $e_{\geq 2}$ is the number of edges of multiplicity
at least two, we have
$$
e = e_1 + e_{\geq 2} \quad \mathrm{ and } \quad e_1 + 2 e_{\geq 2} \leq 2 q \ell \theta.
$$
We deduce that
\begin{equation}\label{eq:bdee1}
e \leq q \ell\theta + e_1 /2.
\end{equation}
Moreover, for each $p \in \INT{q}$ and $\alpha \in \INT{2\theta}$, the sequence $\gamma^\alpha_p = ((x^\alpha_{t,p},i^\alpha_{t,p},x^\alpha_{t+1,p}))_{t \in \INT{\ell}}$ is a path in $G_\gamma$. Besides, from the boundary condition \eqref{eq:bdcond}, for fixed $p \in \INT{q}$, the sequence of paths $(\gamma^\alpha_p), \alpha \in \INT{2\theta},$ are connected by their ends. It follows that the graph $G_\gamma$ has at most $q$ connected components.
If $v =|V_\gamma|$, we deduce that
\begin{equation}\label{eq:genus0}
e - v + q \geq 0.
\end{equation}
In particular, combining the last two inequalities, for any $\gamma \in P_{\ell,\theta}$,
\begin{equation}\label{eq:genus}
\chi = q (\ell \theta+1) + e_1/2 - v \geq 0.
\end{equation}
For integers $v, e_1$ such that \eqref{eq:genus} holds, we denote by $P_{\ell,\theta} (v,e_1)$ the set of elements in $P_{\ell,\theta}$
with $v$ vertices and $e_1$ edges of multiplicity one. Note that if $\gamma \sim \gamma'$, the number of vertices and the number of
edges with a given multiplicity are equal. It follows that we may define $\mathcal {P}_{\ell,\theta} (v,e_1)$ as the set of equivalence classes
with $v$ vertices and $e_1$ edges of multiplicity one. Our first lemma is a rough bound on $\mathcal {P}_{\ell,\theta} (v,e_1)$.
\begin{lemma}\label{le:Plclass}
If $\chi = q (\ell \theta+1) + e_1/2 - v $, we have,
$$
|\mathcal {P}_{\ell,\theta} (v,e_1) |\leq v^{q-1} ( 2 q \ell \theta) ^{ 6 \chi }.
$$
\end{lemma}
\begin{proof}
We order the set $T = \INT{q} \times \INT{2\theta} \times \INT{\ell+1} $ with the lexicographic order on the first two coordinates and
for the last coordinate: $(p,\alpha,t) \succ (p,\alpha,t')$ if $\alpha$ is odd and $t > t'$ or if $\alpha$ is even and $t < t'$. In words,
the natural order is reversed for the last coordinate depending on the parity of $\alpha$.
We think of an element $s= (p,\alpha,t) \in T$ as a time. We also define the set $\vec E_1 = \INT{n} \times \INT{2d}$ ordered with the
lexicographic order. An element $\gamma \in P_{\ell,\theta}$ can be written as the sequence
$ (\gamma_{s})_{s \in T} \in \vec E_1^T$ with $\gamma_s = (y_s, i_s) \in \vec E_1$. For concreteness, we may define a canonical element
in each equivalence class as follows. We say that $\gamma \in P_{\ell,\theta}$ is {\em canonical} if $(\gamma_s)_{s \in T}$ is minimal
for the lexicographic order in its equivalence class. For example, if $\gamma \in P_{\ell,\theta} (v,e_1)$ is canonical then
$\gamma_{(1,1,1)} = (1,1)$, $V_\gamma = \INT{v}$ and the vertices appear for the first time in the sequence $(\gamma_s)$ in order.
Our goal is then to give an upper bound on the number of canonical elements in $P_{\ell,\theta} (v,e_1)$.
We define $T_0\subset T$ as the set of $(p,\alpha,t)$ such that if $\alpha$ is odd $t \in \INT{\ell}$ and if $\alpha$ is even $t-1 \in \INT{\ell}$.
For $s \in T$, we write $s+1$ for the successor of $T$ in the total order and $s-1$ for its predecessor, with the convention that $(1,1,1) - 1= 0$
and $(q,2\theta,1)+1 = \infty$. The set $T_0$ is the subset of $s$ in $T$ such that $s$ and $s+1$ have the same first two coordinates.
We explore iteratively the sequence $(i_s,y_{s+1})$, $s \in T_0$, and we also build a growing sequence of forests $(F_s)$ (graphs without cycles).
We define $F_0$ as the graph with no edge and vertex set $\{1 \}$. At time $s \in T_0$, if the addition to $F_{s-1}$ of the edge
$[y_s,i_s,y_{s+1}]$ does not create a cycle, then $F_s$ is the forest spanned by $F_s$ and $[y_s,i_s,y_{s+1}]$, we then say that
$[y_s,i_s,y_{s+1}]$ is a {\em tree edge}. Otherwise $F_s = F_{s-1}$.
We now build a partition of $T_0$. Let us say that a time $s \in T_0$ is a {\em first time} if the vertex
$y_{s+1}$ has not been seen before. Then, $[y_s,i_s,y_{s+1}]$ is a tree edge that is said to be associated with $y_{s+1}$. We call the
vertices $y_{(p,1,1)} \in V_{\gamma}$ with $p \in \INT{q}$, the {\em seeds}. Due to the boundary condition \eqref{eq:bdcond}, each vertex
$y \in V_\gamma$ different from a seed has an associated tree edge. Hence, the number of tree edges, say $f$, satisfies the inequality
$$
f \geq v - q.
$$
If $s$ is a first time, then the value of $y_{s+1}$ is determined by the preceding values $y_{s'+1}$, $s' < s$ and the value of the
seeds: we have $y_{s+1} = u+1$ where $u$ is the number of distinct vertices which had been seen so far. We say that a time
$s \in T_0$ is a {\em tree time} if $[y_s,i_s,y_{s+1}]$ is a tree edge that has already been visited. Finally, the other times $s \in T_0$ are
called {\em important times}. We have thus partitioned $T_0$ into first times, tree times, and important times.
Due to the non-backtracking constraint \eqref{eq:NBi}, a first time cannot be directly followed by a tree time. The sequence $\gamma$ can thus be decomposed as
the successive repetitions of: $(i)$ a sequence of first times (possibly empty), $(ii)$ an important time or a time in $T \backslash T_0$,
$(iii)$ a sequence of tree times (possibly empty).
We note that all edges are visited at least twice except $e_1$ of them. We deduce that the number of important times is at most
$$
2 q \ell \theta - 2 f + e_1 \leq 2 q (\ell\theta +1) - 2 v + e_1 = 2 \chi.
$$
We mark the important times by the vector $(i_s,y_{s+1},y_{\tau},i_\tau)$ where $\tau > s$ is the next time which is not a tree time.
We claim that the canonical sequence $\gamma$ is
uniquely determined by the value $y_{(p,1,1)}$ of the seeds, the positions of the important times, and their marks. Indeed, the sequence $(y_{s+1},i_{s+1}, \ldots, y_{\tau})$ is the unique non-backtracking path in $F_s$ from $y_{s+1}$ to $y_\tau$ (there is a unique non-backtracking path between two vertices of a tree). Moreover, if $\sigma \geq \tau$ is the next important time or time in $T \backslash T_0$, $(\tau, \ldots, \sigma-1)$ is a sequence of first times (if $\tau = \sigma$ this sequence is empty). It follows that if $u$ vertices had been seen by time $s$, we have $y_{\tau + k} = u+k$, for $k = 1, \cdots, \sigma- \tau$, by the minimality of canonical paths. Similarly, if $i_\tau \ne 1^* = d+1$, then $i_{\tau + k} = 1$ for $k = 1, \cdots, \sigma- \tau$, while if $i_\tau = d+1$, then $i_{\tau + 1} = 2$, $i_{\tau + k} = 1$ for $k = 2, \cdots, \sigma- \tau$.
There are at most $2q\ell \theta$ possibilities for the position of an important time and $(2q\ell\theta )^2$ possibilities for its mark
(there are at most $2 q \ell \theta$ edges in $G_\gamma$). In particular, the number of distinct canonical paths in $P_{\ell,\theta} (v,e_1)$
is at most
$$
v^{q-1} (2 q \ell \theta )^{6 \chi},
$$
where $v^{q-1}$ accounts for the possible values of the seeds (recall that $y_{(1,1,1)} = 1$). \end{proof}
We now estimate the probabilistic weight $p(\gamma)$. To this end, we introduce a finer partition of the set $P_{\ell,\theta}$.
Let $\gamma \in P_{\ell,\theta}$ with $\gamma^{\alpha}_t = (x^{\alpha}_t,i^{\alpha}_t), t \in \INT{\ell+1}$. We define a {\em bracket} as a
colored edge $[x, i ,y]$ on $\INT{n}^q$ with color set $\INT{2d}$ (in the sense of Definition \ref{def:colorgraph}). We associate to
$\gamma$ the sequence of brackets $([x^{\alpha}_t, i^{\alpha}_t ,x^{\alpha}_{t+1}])$, $t \in \INT{\ell}, \alpha \in \INT{2\theta}$. We say that
a bracket $[x^{\alpha}_t, i^\alpha_t ,x^{\alpha}_{t+1}]$ of $\gamma$ is {\em isolated} if for all $p \in \INT{q}$, the colored edge of
$G_\gamma$, $e = [x^{\alpha}_{t,p}, i^{\alpha}_t ,x^{\alpha}_{t+1,p}]$ is not visited at a different time: that is for all
$(t',\alpha',p') \in \INT{\ell}\times \INT{2\theta}\times \INT{q}$ with $(t',\alpha') \ne (t,\alpha)$, we have
$e \ne [x^{\alpha'}_{t',p'}, i^{\alpha'}_{t'} ,x^{\alpha'}_{t'+1,p'}]$. For integer $b$, we denote by $P_{\ell,\theta} ( v,e_1,b)$ the set of
$\gamma \in P_{\ell,\theta} (v,e_1)$ with $b$ isolated brackets.
\begin{lemma}\label{le:pgamma}
Let $\gamma \in P_{\ell,\theta} (v, e_1,b)$ and $\chi = q (\ell\theta+1) + e_1/2 - v$. If $v > q(\ell\theta+1)$ then $p(\gamma ) = 0$. Otherwise,
there exists a universal constant $c >0$ such that
$$
p(\gamma) \leq c n^{ - q \ell \theta} \eta^{b + \frac{e_1}{q}} \left( 2 q\ell \theta \right) ^{ 2\chi },
$$
with $\eta = (c q \ell\theta)^{q/2} n^{-1/8}$.
\end{lemma}
\begin{proof}
Let $T, T_0$ be equipped with their total order as in the proof of Lemma \ref{le:Plclass}. By Theorem \ref{theorem-with-brackets},
$p(\gamma) = 0$ unless the sequence $(x^{\alpha}_{t,p},x^{\alpha}_{t+1,p}), (p,\alpha,t) \in T_0$ is an even sequence. In particular, this
last condition and \eqref{eq:NBi} imply that $2 ( v - q) \leq 2 q \ell\theta $, since for all vertices $x$ of $V_\gamma$ different from the seeds
$(x_{(p,1,1)}), p \in \INT{q},$ we may associate at least two elements $(p,\alpha,t) \in T_0$ such that $ x= x^{\alpha}_{t+1,p}$ (for $\alpha$ odd)
or $ x= x^{\alpha}_{t,p}$ (for $\alpha$ even). It gives the first claim.
We now prove the second claim. By Corollary \ref{cor:WG2}, applied to $k = 2 q \ell \theta$, it suffices to prove that
\begin{equation}\label{eq:m3bd}
m_{\geq 4} \leq 4 \chi ,
\end{equation}
where $m_{\geq 4} = \sum_{e} \IND(m_e \geq 4) m_e$ is the sum of multiplicities of edges of $G_\gamma$ with multiplicity at least $4$.
Indeed, let $e_{23}$ be the number of edges of multiplicity $2$ or $3$. We have
$$
e_1 + 2 e_{23} + m_{\geq 4} \leq 2 q \ell \theta \quad \mathrm{ and } \quad e_1 + e_{23} + \frac{m_{\geq 4}}{ 4} \geq e \geq v - q,
$$
where $e = |E_\gamma|$ is the number of edges and where we have used \eqref{eq:genus0}.
We cancel $e_{23}$, and find
$$
-e_1 + \frac{m_{\geq 4}} { 2} \leq 2 q \ell \theta - 2 v + 2 q.
$$
We obtain \eqref{eq:m3bd}. \end{proof}
Our final lemma is the estimation of $\sum a(\gamma)$ where the sum is over an equivalence class. This is our main combinatorial ingredient.
\begin{lemma}\label{le:agamma}
Let $\gamma \in P_{\ell,\theta} (v, e_1,b)$, $\chi = q (\ell\theta +1) + e_1/2 - v$ and let $\rho$ and $k_0$ be as in \eqref{eq:defrho}. There exists a constant
$c >1$ (depending on $k_0,r,d$ and $\max_i \| a_i \|$), such that
$$
A(\gamma) = \sum_{\gamma' : \gamma'\sim \gamma} a(\gamma') \leq n^v c^{b + q\theta + \chi} \rho^{2 \ell\theta}.
$$
\end{lemma}
We start with a preliminary lemma.
\begin{lemma}\label{le:agamma0}
Let $\rho$ and $k_0$ be as in \eqref{eq:defrho}. There exists a constant $c > 1$ (depending on $k_0,d$ and $\max_i \| a_i \|$) such that for any integer
$k \geq 1$,
$$
\sum \NRM{\prod_{t=1}^k a_{i_t} }^2 \leq c^2 \rho^{2k}
\quad \mathrm{ and } \quad
\sum \NRM{\prod_{t=1}^k a_{i_t} } \leq c^k \rho^{k},
$$
where the sums are over all $(i_1, \cdots, i_k)$ such that $i_{t+1} \ne i_t^*$. In particular,
$$\max \NRM{\prod_{t=1}^k a_{i_t} } \leq c\rho^k$$
where the maximum is over all $(i_1, \cdots, i_k)$ such that $i_{t+1} \ne i_t^*$.
\end{lemma}
\begin{proof}
Let $k \geq 1$ be an integer. As above Proposition \ref{prop:nonback2}, for
$e,f \in {\mathbb{F}_d} \times \INT{2d}$, we denote by $(B^k_\star)_{ef} \in M_r(\mathbb {C})$ the matrix $P_e B^k _\star P_f$ with $P_f$ the orthogonal projection onto
$\mathbb {C}^r \otimes \delta_f$.
Let $e = (o,i)$ and $f = (g,j) $ be in ${\mathbb{F}_d} \times \INT{2d}$ where $g = g_{i_1}\cdots g_{i_m}$ is written in reduced form (that is, for all $t\in \INT{m-1}$, $i_{t+1} \ne i_t^*$). If $m = k$, $i_1 = i$ and $j \ne i^*_{m}$, from the definition of $B_\star$, we have
$$
(B_\star^k)_{e f} = \prod_{t=1}^k a_{i_t}.
$$
Otherwise, $m \ne k$, $i_1 \ne i$ or $j = i^*_{m}$ and we find $(B_\star^k)_{e f} = 0$. For $\varphi \in \mathbb {C}^r$, we deduce that
$$
\| B_\star ^k \varphi \otimes \delta_e \|^2 = (2 d - 1) \sum \| \prod_{t=1}^k a_{i_t} \varphi \|^2,
$$
where the sum is over all $(i_1, \cdots, i_k)$ such that $i_1 = i$ and $i_{t+1} \ne i_t^*$. Recall that $\rho_{k}$ was defined above \eqref{eq:defrho}. Taking the supremum over all $\varphi$ of unit norm, we get
$$
\rho_{k}^{2k} = (2d-1) \cdot \max_{i \in \INT{2d}, \varphi \in S^{r-1}} \sum \NRM{\prod_{t=1}^k a_{i_t} \varphi }^2 ,
$$
where the sum is over all $(i_1, \cdots, i_k)$ such that $i_1 = i$ and $i_{t+1} \ne i_t^*$. Observe that if $T \in M_r(\mathbb {C})$ is a matrix with singular values $\| T \| = s_1 \geq \ldots \geq s_r \geq 0$ and right singular vectors $u_1, \ldots , u_r$, then
$$
\| T \varphi \|^2 = \sum_{i=1}^r s_i ^2 | \langle u_i , \varphi \rangle |^2.
$$
Hence, if $\varphi$ is uniformly distributed on $S^{r-1}$, we find
$$
\mathbf {E} \| T \varphi \|^2 = \frac 1 r \sum_{i=1}^r s_i ^2 \geq \frac{\| T \|^2 }{r} .
$$
In particular, from what precedes,
$$
\rho_{k}^{2k} \geq \frac{(2d-1)}{r} \max_{i \in \INT{2d}} \sum \NRM{\prod_{t=1}^k a_{i_t} }^2 .
$$
We find from \eqref{eq:defrho} that for all $k \geq k_0$,
$$
\sum \NRM{\prod_{t=1}^k a_{i_t} }^2 \leq \frac{2d r }{2d-1} \rho^{2k},
$$
where the sum is over all $(i_1, \cdots, i_k)$ such that and $i_{t+1} \ne i_t^*$.
Up to a large multiplicative constant on the right-hand side (depending on $\max_i \|a_i\|$, $k_0, r, d$), this last inequality is valid for all $k \geq 1$. This is the first claimed statement.
For the second statement, we apply Cauchy-Schwarz inequality and get:
$$
\sum \NRM{\prod_{t=1}^k a_{i_t} } \leq \PAR{2d (2d-1)^{k-1}}^{\frac 1 2} \PAR{\sum \NRM{\prod_{t=1}^k a_{i_t} }^2 }^{\frac 1 2}.
$$
Modifying the constant $c$, we obtain the second statement. The last statement is an immediate consequence of the first statement.
\end{proof}
We may now prove Lemma \ref{le:agamma}.
\begin{proof}[Proof of Lemma \ref{le:agamma}] We start by introducing a new decomposition of a path $\gamma \in P_\gamma$. This decomposition is tailored to use the three bounds given by Lemma \ref{le:agamma0} but keeping in mind that for each use of Lemma \ref{le:agamma0}, there is a cost of a constant factor $c$ for the first and third bounds and an exponential cost $c^k$ for the second bound. The conclusion of Lemma \ref{le:agamma} will ultimately follow from an application of the first and last inequality $O(\chi + q \theta)$ times and the second $O(b)$ times.
As usual, we write $\gamma \in P_{\ell,\theta}$ with $\gamma^{\alpha}_t = (x^{\alpha}_t, i^{\alpha}_t) \in \INT{n}^q \times \INT{2d}$,
$x^{\alpha}_t = (x^{\alpha}_{t,p})_{p \in \INT{q}}$. Let $T = \INT{2\theta} \times \INT{\ell}$ and let $T_{\geq 3}$ be the set of $ (\alpha,t) \in T$
such that for some $p \in \INT{q}$, $x^\alpha_{t,p}$ has a degree at least $3$ in $G_\gamma$ (recall that the degree of a vertex was defined in
Definition \ref{def:colorgraph}). We let $T_1$ be the set of $(\alpha,t) \in T$ such that $x^\alpha_{t,p} \in \{ x^{\alpha'}_{t',p'} : \alpha'\in \INT{2\theta}, t' \in \{1,\ell+1\}, p' \in \INT{q} \}$. In words, $T_1$ is the set of indices $(\alpha,t)$ such that $x^\alpha_{t,p}$ reaches a boundary vertex for some $p \in \INT{q}$. We note that, in particular, $(\alpha,1) \in T_1$.
We finally set $T_\star = T _{\geq 3} \cup T_1 $. We write $l = |T_\star |$,
$T_\star = \{ (\alpha,t_j^\alpha) : j \in \INT{l_\alpha} \}$ with $\sum_\alpha l_\alpha=l$, $(t_j^\alpha)_j$ increasing with $t^\alpha_1 = 1$ and by convention we set $t^\alpha_{l_\alpha+1} = \ell+1$.We write $a(\gamma)$
defined below \eqref{eq:tr1} as
\begin{equation}\label{eq:ag10}
a(\gamma) = \prod_{\alpha =1}^{2\theta} \NRM{ \prod_{j = 1}^{l_{\alpha} } \prod_{t=t^\alpha_j}^{t^ \alpha_{j+1} -1 } a^{\varepsilon_\alpha}_{ {i^{\alpha}_{t}} } }.
\end{equation}
We start with an upper bound on $|T_{\star}| $. Similarly to \eqref{eq:m3bd}, let $m_{\geq 3} = \sum_e \IND(m_e \geq 3) m_e$ be the sum of multiplicities at least $3$, i.e. $m_{ \geq 3}$ is the number of indices $(\alpha,t,p) \in \INT{2\theta} \times \INT{\ell} \times \INT{q}$ where the path visits an edge visited at least $3$ times. Let $v_k$ be the number of
vertices of $V_\gamma$ with degree $k$ and $d_{\ge 3} = \sum_{k \geq 3} k v_k$. By definition, $d_{\geq 3}$ is the number of edges neighboring a vertex of degree at least $3$. Observe that if $(\alpha,t) \in T_{\geq 3}$, then there exists $p \in \INT{q}$ such that $x^{\alpha}_{t,p} $ has degree at least $3$ and the edge $[x^{\alpha}_{t,p} ,i^\alpha_t , x^{\alpha}_{t+1,p}]$ either is visited for the first or second time or is visited for the third time or more. Similarly, if $(\alpha,t) \in T_1 \backslash T_{\geq 3}$, then there exists $p \in \INT{q}$ such that $x^{\alpha}_{t,p} $ is a boundary vertex of degree $1$ or $2$ and the edge $[x^{\alpha}_{t,p} ,i^\alpha_t , x^{\alpha}_{t+1,p}]$ either is visited for the first or second time or is visited for the third time or more. Since there are at most $2 q \theta$ boundary vertices,
we get the inequality
\begin{equation}\label{eq:T3T3}
|T_{\star}| = |T_{\geq 3} \cup T_1| \leq 2 d_{\geq 3} + m_{\geq 3} + 4 \times 2 q \theta.
\end{equation}
Arguing as in \eqref{eq:m3bd}, we claim that $m_{\geq 3} \leq 6 \chi$. Indeed, if $e_2$ is the number of edges of multiplicity $2$, we get
$$
e_1 + 2 e_2 + m_{\geq 3} = 2 q \ell \theta \quad \mathrm{ and } \quad e_1 + e_2 + \frac{m_{\geq 3}}{3} \geq e \geq v - q,
$$
where the last inequality is \eqref{eq:genus0}. Cancelling $e_2$, we get
$$
-e_1 + \frac{m_{\geq 3} }{3} \leq 2q \ell \theta - 2e.
$$
Hence $m_{\geq 3} / 3 \leq 2q \ell \theta - 2v + 2 q + e_1 = 2 \chi$, as claimed.
We may also similarly estimate $d_{\geq 3}$. We write
$$
v_1 + v_2 + \frac{d_{\geq 3}}{3} \geq \sum_k v_k = v \quad \mathrm{ and } \quad v_1 + 2 v_2 + d_{\geq 3} = \sum_{k} k v_k = 2 e,
$$
where $e = |E_\gamma|$ is the number of edges. Canceling $v_2$, we find
\begin{equation}\label{eq:sge3}
\frac{d_{\geq 3} }{3} \leq 2 ( e - v) + v_1 \leq 2 (q \ell \theta + e_1/ 2 - v) + 2 q \theta = 2 \chi + 2 q(\theta-1),
\end{equation}
where, in the second inequality, we have used \eqref{eq:bdee1} and the bound $ v_1 \leq 2q \theta $ is a consequence of the assumption \eqref{eq:NBi}: only the vertices
$\gamma^{2 \alpha}_{1,p} = \gamma^{2\alpha+1}_{1,p} $ and $\gamma^{2\alpha}_{\ell+1,p} = \gamma^{2\alpha +1}_{\ell+1,p}$ with
$p \in \INT{q}$ can possibly be of degree $1$ in $G_\gamma$. Therefore, we deduce from \eqref{eq:T3T3} that
\begin{equation}\label{eq:boundl}
l \leq 18 \chi +20 q\theta.
\end{equation}
We now proceed to the bound of the factors in \eqref{eq:ag10} when we sum over all $\gamma' \sim \gamma$. We perform this by building a partition of the factors in \eqref{eq:ag10} and then summing within each block of the partition over all $\gamma' \sim \gamma$.
Consider the color set $\{-1,1\}$ equipped with the trivial involution $c^* = c$. We then define a colored graph, say $\Gamma$, on the vertex set
$T = \INT{2\theta} \times \INT{\ell}$ and on the color set $C$, by placing an edge $[(\alpha,t),1,(\alpha',t')]$ between $(\alpha,t) \ne (\alpha',t')$ with
color $1$ if there exists $(p,p')$ in $\INT{q}$ such that
$(x^{\alpha}_{t,p},i^{\alpha}_{t},x^{\alpha}_{t+1,p}) = (x^{\alpha'}_{t',p'},i^{\alpha'}_{t'},x^{\alpha'}_{t'+1,p'})$ and the edge
$[(\alpha,t),-1,(\alpha',t')]$ with color $-1$ if
$(x^{\alpha}_{t,p},i^{\alpha}_{t},x^{\alpha}_{t+1,p}) = (x^{\alpha'}_{t'+1,p'},{i^{\alpha'}_{t'}}^*,x^{\alpha'}_{t',p'})$.
In words, we place an edge between two elements of $T$ if they share a common edge in $G_\gamma$, the color encoding the orientation. In particular, by definition:
{\em Claim 1. } $(\alpha,t) \in T$ has degree $0$ in $\Gamma$ iff $[x^{\alpha}_{t},i^{\alpha}_{t},x^{\alpha}_{t+1}]$ is an isolated bracket.
Summing over all possibilities for $\gamma'$, we find from \eqref{eq:ag10} that
\begin{equation}
\label{eq:ag101}
I = \sum_{\gamma' : \gamma'\sim \gamma} a(\gamma') \leq (2d -1)^{\theta} n^v \sum \prod_{\alpha= 1}^{2\theta} \NRM{ \prod_{j = 1}^{l_{\alpha} } \prod_{t=t^{\alpha}_j}^{t^\alpha_{j+1} -1 } a_{ {i^{\alpha}_{t}} } },
\end{equation}
where the sum is over all $(i^{\alpha}_t)_{(\alpha,t) \in T}$ in $\INT{2d}$ such that for all $(\alpha,t), (\alpha',t')$:
\begin{equation}\label{eq:sumindex}
\hbox{ $i^{\alpha}_{t+1} \ne {i^{\alpha}_t}^*$, $i^{\alpha}_{t} = i^{\alpha'}_{t'}$ if $[(\alpha,t),1,(\alpha',t')] \in \Gamma$ and
$i^{\alpha}_{t} =(i^{\alpha'}_{t'})^*$ if $[(\alpha,t),-1,(\alpha',t')] \in \Gamma$.}
\end{equation}
The factor $(2d-1)^\theta$ in \eqref{eq:ag101} comes from the choices of $i^\alpha_{\ell+1}$.
For ease of notation, for $r = (\alpha,t)$ and $s = (\alpha , t')$ with $t < t'$ in $\INT{\ell+1}$ and $\alpha \in \INT{2\theta}$, we will denote
by $[r,s)$ the sequence $((\alpha,t),(\alpha,t+1), \cdots, (\alpha,t' - 1))$ and write $u \in [r,s)$ if $u = (\alpha,t'')$ with $t \leq t'' < t'$.
Such sequence $[r,s)$ will be called an {\em interval} of {\em length} $t'-t$. If $[r_1,s_1)$ and $[r_2,s_2)$ are two intervals,
$r_i = (\alpha_i , t_i)$, $s_i = (\alpha_i , t'_i)$ we say that $[r_1,s_1)$ and $[r'_2,s'_2)$ are {\em $1$-paired} if they have the same length
and, for all $0\leq \delta < t'_i - t_i$, either: $[ (\alpha_1,t_1 + \delta) , 1 ,(\alpha_2,t_2+\delta)]$ is an edge of $\Gamma$ or
$[ (\alpha_1,t_1 + \delta) , -1 ,(\alpha_2,t'_2-1-\delta)]$ is an edge of $\Gamma$, see Figure \ref{fig:1-1}. We say that there are {\em $2$-paired} if there exists a
third interval which is $1$-paired to both of them. Iteratively, we define {\em $k$-paired} intervals. Finally, we say that two intervals
are {\em paired} if they are $k$-paired for some $k \geq 1$. The key property is that if two intervals $[r_1,s_1)$ and $[r_2,s_2)$ are paired then,
with the above notation, either
\begin{equation}\label{eq:pairing}
\hbox{for all $0\leq \delta < t'_i - t_i$: } i^{\alpha_1}_{t_1 + \delta} = i^{\alpha_2}_{t_2 + \delta} \quad \hbox{ or } \quad \hbox{for all $0\leq \delta < t'_i - t_i$: } i^{\alpha_1}_{t_1 + \delta} = (i^{\alpha_2}_{t'_2 - \delta-1})^*.
\end{equation}
\begin{figure}
\begin{center}
\begin{tikzpicture}[xscale=0.8]
\draw[-,cyan,ultra thick] (0,0) -- (4,0) ;
\draw (0.5,0) node[above]{$a$} ;
\draw (1.5,0) node[above]{$b$} ;
\draw (2.5,0) node[above]{$c$} ;
\draw (3.5,0) node[above]{$d$} ;
\draw (0,0) node[left] {$t_{1}$};
\draw (4,0) node[right] {$t'_{1}$};
\foreach \x in {0, ...,4}
\draw[shift={(\x,0)},color=black] (0pt,3pt) -- (0pt,-3pt) ;
\draw (0.5,-1) node[above]{$a$} ;
\draw (1.5,-1) node[above]{$b$} ;
\draw (2.5,-1) node[above]{$c$} ;
\draw (3.5,-1) node[above]{$d$} ;
\draw (0,-1) node[left] {$t_{2}$};
\draw (4,-1) node[right] {$t'_{2}$};
\draw[-,cyan,ultra thick] (0,-1) -- (4,-1) ;
\foreach \x in {0, ...,4}
\draw[shift={(\x,-1)},color=black] (0pt,3pt) -- (0pt,-3pt) ;
\draw (0.5,-2) node[above]{$d^*$} ;
\draw (1.5,-2) node[above]{$c^*$} ;
\draw (2.5,-2) node[above]{$b^*$} ;
\draw (3.5,-2) node[above]{$a^*$} ;
\draw (0,-2) node[left] {$t_{3}$};
\draw (4,-2) node[right] {$t'_{3}$};
\draw[-,cyan,ultra thick] (0,-2) -- (4,-2) ;
\foreach \x in {0, ...,4}
\draw[shift={(\x,-2)},color=black] (0pt,3pt) -- (0pt,-3pt) ;
\end{tikzpicture}
\caption{Paired intervals of length $4$. The letters on the lines represent the different brackets that are equal. The first two are paired with color $1$, and the third with color $-1$. \label{fig:1-1}}
\end{center}
\end{figure}
Our second claim asserts that the intervals $[t^{\alpha}_j,t^{\alpha}_{j+1})$ are paired unless they are made of isolated brackets.
{\em Claim 2. } For any $(\alpha,j)$, the degree of $(\alpha,t) \in T$ in $\Gamma$ is constant for all $t \in [t^{\alpha}_j,t^{\alpha}_{j+1})$, including the degree counted by colors. Moreover, if the degree is at least one, then the interval $ [t^{\alpha}_j,t^{\alpha}_{j+1})$ is paired to a subinterval of $ [t^{\alpha'}_{j'},t^{\alpha'}_{j'+1})$ for some $(\alpha',j') \ne (\alpha,j)$.
Let us check this claim. Assume that $t\in [t^{\alpha}_j,t^{\alpha}_{j+1})$ and assume that $(\alpha,t)$ has a colored edges with color $1$ in $\Gamma$. Then, there exists $ (\alpha',t') \ne (\alpha,t)$ in $T$ and $(p,p') \in \INT{q}^2$ such that $(x^{\alpha}_{t,p},i^{\alpha}_{t},x^{\alpha}_{t+1,p}) = (x^{\alpha'}_{t',p'},i^{\alpha'}_{t'},x^{\alpha'}_{t'+1,p'})$. If $ t+1 \in [t^{\alpha}_j,t^{\alpha}_{j+1})$ then $x^{\alpha}_{t+1,p} = x^{\alpha'}_{t'+1,p}$ is not a boundary vertex and has degree exactly $2$. Hence, $t+1 , t'+1 \le \ell$, $(\alpha',t'+1) \notin T_\star$ and from \eqref{eq:NBi}, the two adjacent edges of $x^{\alpha}_{t+1,p} = x^{\alpha'}_{t'+1,p}$ are $[x^{\alpha}_{t,p}, i^{\alpha}_{t,p}, x^{\alpha}_{t+1,p}] = [x^{\alpha'}_{t',p'}, i^{\alpha}_{t',p'}, x^{\alpha}_{t'+1,p'}]$ and $[x^{\alpha}_{t+1,p}, i^{\alpha}_{t+1,p}, x^{\alpha}_{t+2,p}] = [x^{\alpha'}_{t'+1,p'}, i^{\alpha}_{t'+1,p'}, x^{\alpha}_{t'+2,p'}]$ (the last equality follows from the constraint that the degree of $x^{\alpha}_{t+1,p}$ is exactly $2$). We thus have proved that if $ t+1 \in [t^{\alpha}_j,t^{\alpha}_{j+1})$ then $ (\alpha,t+1)$ shares a colored edge with color $1$ in $\Gamma$ with $(\alpha',t'+1)$ and that $((\alpha,t),(\alpha,t+1))$ is paired with $((\alpha',t'),(\alpha',t'+1))$. Similarly, if $ t- 1 \in [t^{\alpha}_j,t^{\alpha}_{j+1})$ then $x^{\alpha}_{t,p} = x^{\alpha'}_{t',p}$ is not a boundary vertex and has degree exactly $2$. We deduce as above that $ (\alpha,t-1)$ shares a colored edge with color $1$ in $\Gamma$ with $ (\alpha',t'-1)$ and that $((\alpha,t-1),(\alpha,t))$ is paired with $((\alpha',s'-1),(\alpha',s'))$. The same argument applies to edges of color $-1$. By induction, this proves the claim.
In view of Claim 2, our goal is then to find paired subintervals of $[t^{\alpha}_j,t^{\alpha}_{j+1})$ of long length in \eqref{eq:ag101} and then use Lemma \ref{le:agamma0}
for each of them. More precisely, we claim that there exists a partition of the set $T$ made of $m$ intervals $[r_i,r_{i+1}), i \in \INT{m},$ with
$m \leq 8 l$ and a classification of the intervals $[r_i,r_{i+1})$ into $3$ types. An interval $[r_i,r_{i+1})$ is of type (1) iif all $u \in [r_i,r_{i+1})$
are of degree $0$ in $\Gamma$. There is a matching of intervals of type (2) such that two matched intervals of type (2) are paired together (by
`matching', we mean a pair partition, we use the word matching to make a distinction between `matched intervals' and `paired intervals' defined
above). Finally, any $(\alpha,t) \in T$ in an interval of type (3) is connected in $\Gamma$ to an element in an interval of type (2).
Let us first assume the existence of such partition and conclude the proof of Lemma \ref{le:agamma}. For $c \in \{1,3\}$, we let $m_c $ be
the number of intervals of type (c) and $k_{c,j}$ be the length of the $j$-th interval. Let $2m_2$ be the number of intervals of type (2) and
$k_{2,j}$ be the common length of the two $j$-th paired intervals. We get from \eqref{eq:ag101} and \eqref{eq:pairing}
\begin{align*}
I &\leq (2d)^\theta n^v \prod_{j=1}^{m_1} \BRA{ \sum_{\substack{1 \leq t \leq k_{1,j} \\ i_{t+1} \ne i^*_t }} \NRM{ \prod_{t = 1}^{k_{1,j}} a_{ {i_{t}} } }}\prod_{j=1}^{m_2} \BRA{ \sum_{\substack{1 \leq t \leq k_{2,j} \\ i_{t+1} \ne i^*_t }} \NRM{ \prod_{t = 1}^{k_{2,j}} a_{ {i_{t}} } }^2} \prod_{j=1}^{m_3} \BRA{\max_{\substack{1 \leq t \leq k_{3,j} \\ i_{t+1} \ne i^*_t }} \NRM{ \prod_{t = 1}^{k_{3,j}} a_{ {i_{t}} } } } \\
&= (2d)^\theta n^v \prod_{c=1}^3 \prod_{j=1}^{m_c} I_{c,j}.
\end{align*}
We note that to obtain the above expression, we have used Cauchy-Schwarz inequality for intervals of type (2) associated with the color $-1$:
\begin{eqnarray*}
\sum_{\substack{1 \leq t \leq k \\ i_{t+1} \ne i^*_t }} \NRM{ \prod_{t = 1}^{k} a_{ {i_{t}} } }\NRM{ \prod_{t = 1}^{k} a_{ {i^*_{k-t+1}} } } \leq \sqrt{\sum_{\substack{1 \leq t \leq k \\ i_{t+1} \ne i^*_t }} \NRM{ \prod_{t = 1}^{k} a_{ {i_{t}} } }^2 } \sqrt{ \sum_{\substack{1 \leq t \leq k \\ i_{t+1} \ne i^*_t }}\NRM{ \prod_{t = 1}^{k} a_{ {i^*_{k-t+1}} } }^2} = \sum_{\substack{1 \leq t \leq k \\ i_{t+1} \ne i^*_t }} \NRM{ \prod_{t = 1}^{k} a_{ {i_{t}} } }^2.
\end{eqnarray*}
Also, to get the above bound on $I$, we have used that for an interval of type (3), the $i^\alpha_t$'s associated with the indices $(\alpha,t)$ in the interval are determined by the $i^\alpha_t$'s associated the indices $(\alpha,t)$ in intervals of type (2).
Then, by Lemma \ref{le:agamma0}, we have
$I_{1,j} \leq c^{k_{1,j}}\rho^{k_{1,j}}$, $I_{2,j} \leq c^{2}\rho^{2 k_{2,j}}$ and $I_{3,j} \leq c \rho^{k_{3,j}}$. So finally, we deduce that
$$
I \leq (2d)^\theta n^v c^{b} c^{m} \rho^{2 \ell\theta}.
$$
We are have used that $2 m_2 + m_3 \leq m$, $\sum_j k_{1,j} = b$ (by Claim 1) and $\sum k_{1,j} + 2 \sum k_{2,j} + \sum k_{3,j} = 2 \ell\theta $. Thus,
from the assumption $m \leq 8 l$ and from \eqref{eq:boundl}, up to modifying the constant $c$, we deduce that the conclusion of
Lemma \ref{le:agamma} holds provided the existence of the partition $[r_j,r_{j+1}), j \in \INT{m}$, of $T$.
The rest of the proof is devoted to the construction of the partition. To that end, we define a natural greedy pairing algorithm. At a given step
$i \geq 0 $ of the algorithm, there will be a partition of $T$ into intervals $[r^{(i)}_j,r^{(i)}_{j+1})$. Some intervals of the current partition will be
said to be {\em active}, and the others are {\em frozen} (frozen intervals will remain unchanged in all subsequent partitions). The frozen intervals
have been classified into types; the active intervals have not. The algorithm stops when all intervals are frozen. At step $i=0$, the initial partition
is $[t^{\alpha}_j,t^{\alpha}_{j+1})$, $\alpha \in \INT{2\theta}, j \in \INT{l_\alpha},$ and all intervals are active. At each step of the algorithm,
the number of active intervals decreases until there is none. Thus, the algorithm stops after at most $l$ steps.
We have the following induction hypothesis: each active interval $[r^{(i)}_j ,r^{(i)}_{j+1})$ is contained in $[t^\alpha_{j'},t^{\alpha}_{j'+1})$ for some
$\alpha \in \INT{2m}$
and $j' \in \INT{l_\alpha}$. Moreover, any $(\alpha,t)$ in a frozen interval of type (3) is connected in $\Gamma$ to an element in a frozen interval
of type (2). We now describe the induction step. We pick the longest active sequence, say $[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$. There are several
cases to consider:
\begin{enumerate}[(i)]
\item {\em All $u \in T$ with $u \in [r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$ have common degree $0$}. Then, we leave the partition unchanged, the interval
$[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$ is classified as type $(1)$ and becomes frozen.
\item \label{lbii}{\em All $u \in T$ with $u \in [r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$ have common degree at least $1$, then we pick an interval $[s,s')$
such that $[s,s')$ and $[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$ are $1$-paired (its existence is guaranteed by the induction hypothesis and Claim 2). The interval
$[s,s')$ intersects a finite number, say $\alpha$, of intervals, $\mathcal{I} = \{ [r^{(i)}_{j_1 + k - 1 },r^{(i)}_{j_1 + k}), k \in \INT{ \alpha} \}$. We assume
that the corresponding color is $1$.
} There are further subcases to consider:
\item[(ii-a)]{\em The intervals in $\mathcal{I}$ are all active and $\mathcal{I}$ does not contain $[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$.} Then, we merge the intervals in
$\mathcal{I}$ into $3$ intervals $[r^{(i)}_{j_1}, s)$, $[s,s')$ and $ [s',r^{(i)}_{j_1 + \alpha})$ (the two extreme intervals may be empty). The two extreme
intervals are active. The intervals $[s,s')$ and $ [r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$ are frozen, they are classified as type (2), and they are matched
together. The rest of the partition remains unchanged at step $i+1$. Note that if $\alpha = 1$, then the two extreme intervals are empty since we
have picked the longest interval.
\item[(ii-b)]{\em The intervals in $\mathcal{I}$ are all active and $\mathcal{I}$ contains $[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$.} We then have that $\alpha \geq 2$ (the case
$\alpha = 1$ is ruled out by the assumption that the corresponding color is $1$) and $[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$ is one of the two extreme intervals
of $\mathcal{I}$. Assume, for example, that $j_1 = j_0$. We denote by $k$ and $k_0$ the lengths of $[r^{(i)}_{j_0} ,s )$ and $[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$.
Then, $1 \leq k \leq k_0$ and $h =k_0+k$ is the total length of $[r^{(i)}_{j_1} ,s')$. By construction if $[(\alpha,t),(\alpha,t'))$ and
$[(\alpha,t+ k),(\alpha,t'+ k))$ are contained $[r^{(i)}_{j_1} ,s')$ then they are paired. By induction, we deduce that for any integer $a \geq 1$, if $[(\alpha,t),(\alpha,t'))$ and
$[(\alpha,t + a k),(\alpha,t'+ a k))$ are contained $[r^{(i)}_{j_1} ,s')$ then they are paired.
We let $k' = a k$ where $a \geq 1$ is the largest integer
such that $a k \leq h/2$. We have $k'\geq h/4$ since otherwise, $k ' + k \leq 2 k' < h/2$. Assume, for example, that $h/4 \leq k' < h/3$. Let $r =h - 3 k'$.
We divide the sequence $(0,\ldots,h-1)$, into $(p k',\ldots,p k' + r-1)$ for $p \in \{0,1,2,3\}$ and $(qk' + r , \ldots, (q +1) k'-1)$, for $q \in \{0,1,2\}$.
Mapping the interval $[r^{(i)}_{j_1} ,s')$ into the sequence $(0,\ldots, h-1)$, we have split $[r^{(i)}_{j_1} ,r^{(i)}_{j_1+\alpha})$ into $7$ frozen intervals
at step $i+1$ (see Figure \ref{fig:7} for an illustration). The two intervals corresponding to $p \in \{0,1\}$ are of type $(2)$ and are matched together, similarly for the two intervals associated
with $p \in \{2,3\}$. Finally, the three intervals corresponding to $q \in \{0,1,2\}$ are mutually paired; two of them are classified as type (2) and the
remaining one as type (3). The interval $[s',r^{(i)}_{j_1 + \alpha})$ becomes an active interval (if not empty). The rest of the partition remains
unchanged at step $i+1$. The case $h/3 \leq k' \leq h/2$ is treated similarly with $5$ intervals, $3$ of length $r$ and $2$ of length $k' -r$
where $r = h - 2k'$ (if $r = 0$, $2$ intervals are sufficient), see Figure \ref{fig:7}.
\begin{figure}
\begin{center}
\begin{tikzpicture}[xscale=0.8]
\draw[-,cyan,ultra thick] (0,0) -- (20,0) ;
\foreach \x in {0,6,12,18}
\draw (\x+0.5,0) node[above]{$a$} ;
\foreach \x in {0,6,12,18}
\draw (\x+1.5,0) node[above]{$b$} ;
\foreach \x in {0,6,12}
\draw (\x+2.5,0) node[above]{$c$} ;
\foreach \x in {0,6,12}
\draw (\x+3.5,0) node[above]{$d$} ;
\foreach \x in {0,6,12}
\draw (\x+4.5,0) node[above]{$e$} ;
\foreach \x in {0,6,12}
\draw (\x+5.5,0) node[above]{$f$} ;
\foreach \x in {0, ...,20}
\draw[shift={(\x,0)},color=black] (0pt,3pt) -- (0pt,-3pt) node[below] {$\x$};
\foreach \x in {0, ...,20}
\draw[shift={(\x,0.5)},color=black] (0pt,3pt) -- (0pt,-3pt) ;
\foreach \x in {0,6,12,18}
\draw[shift={(\x,0)},-,color = black ,ultra thick] (-0.0,0.5) -- (1.0,0.5) ;
\foreach \x in {0,6,12}
\draw[shift={(\x,0)},-,color = black ,ultra thick] (2,0.5) -- (5.0,0.5) ;
\draw (0,1) node {$r_{j_0}^{(i)}$};
\draw (6,1) node {$s$};
\draw (14,1) node {$r_{j_0+1}^{(i)}$};
\draw (20,1) node {$s'$};
\draw[-,cyan,ultra thick] (0,-3) -- (20,-3) ;
\foreach \x in {0,3,6,9,12,15,18}
\draw (\x+0.5,-3) node[above]{$a$} ;
\foreach \x in {0,3,6,9,12,15,18}
\draw (\x+1.5,-3) node[above]{$b$} ;
\foreach \x in {0,3,6,9,12,15}
\draw (\x+2.5,-3) node[above]{$c$} ;
\foreach \x in {0, ...,20}
\draw[shift={(\x,-3)},color=black] (0pt,3pt) -- (0pt,-3pt) node[below] {$\x$};
\foreach \x in {0, ...,20}
\draw[shift={(\x,-2.5)},color=black] (0pt,3pt) -- (0pt,-3pt) ;
\foreach \x in {0,9,18}
\draw[shift={(\x,-3)},-,color = black ,ultra thick] (-0.0,0.5) -- (1.0,0.5) ;
\foreach \x in {0,9}
\draw[shift={(\x,-3)},-,color = black ,ultra thick] (2,0.5) -- (8.0,0.5) ;
\draw (0,-2) node {$r_{j_0}^{(i)}$};
\draw (3,-2) node {$s$};
\draw (17,-2) node {$r_{j_0+1}^{(i)}$};
\draw (20,-2) node {$s'$};
\end{tikzpicture}
\caption{Above : The splitting into $7$ intervals in the case $k=6$, $k_0 = 14$, $h = 20$, $k' = 6$, $r = 2$. The letters on the line represent the different brackets. Below : the splitting into $5$ intervals in the case $k=3$, $k_0 = 17$, $h=20$, $k' = 9$, $r=2$. \label{fig:7}}
\end{center}
\end{figure}
\item[(ii-c)]{\em The intervals in $\mathcal{I}$ are all frozen.} Note that by the induction hypothesis, the intervals of $ \mathcal{I}$ are of type (2) or (3), and all
elements in $\mathcal{I}$ are connected in $\Gamma$ to an element in a frozen interval of type (2). The interval $[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$ becomes
a frozen interval of type (3). The rest of the partition remains unchanged at step $i+1$.
\item[(ii-d)]{\em The set $\mathcal{I}$ contains both active and frozen intervals.} We split $\mathcal{I}$ into at most three-set of contiguous intervals, say
$\mathcal{I}_1, \ldots, \mathcal{I}_{k}$ with $k \in \{2,3\}$, such that $\mathcal{I}_1$ contains only active intervals and the one or two others contain only frozen intervals.
This is possible since, at each step, we have chosen the longest active interval: it follows that the union of frozen intervals is a union of disconnected
intervals such that each of these disconnected intervals has a length larger or equal than the longest active interval. We also split
$[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$ into two or three intervals, $J_1, \ldots, J_k$, such that for all $p \in \INT{k}$, $J_p$ is paired with $I_p$ which
is contained in the union of the intervals in $\mathcal{I}_p$. For each $p$, we may apply one of the above steps.
\item \label{lbiii}{\em As in \eqref{lbii} but the corresponding color is $-1$.} Up to minor modifications on the indices, the same splitting strategy \eqref{lbii} also works for
color $-1$.
\end{enumerate}
At each step, the number of active intervals decreases. It follows that the algorithm stops after at most $l$ steps with a partition and a classification
by types with all desired properties. Since $[r^{(i)}_{j_0} ,r^{(i)}_{j_0+1})$ is split into at most $7+1 = 8$ parts, the total number of intervals
$m$ is bounded by $8l$. This concludes the proof of Lemma \ref{le:agamma}.
\end{proof}
All ingredients are gathered to prove Theorem \ref{th:FKB}.
\begin{proof}[Proof of Theorem \ref{th:FKB}]
We fix $\varepsilon >0$. For some small $\delta>0$ to be fixed later on, we set $$\theta = \lceil n^{\delta/q} \rceil.$$
We start by assuming that (for $n \geq 3$),
\begin{equation}\label{eq:choixql}
q \leq \frac{1}{2^7} \frac{\ln n}{\ln \ln n} \quad \mathrm{ and } \quad C_0 q \leq \ell \leq 2C_0 q
\end{equation}
for some constant $C_0$ to be also fixed. In this proof, we write $c$ for a universal constant and $C$ for a constant which depends on $d$,
$\varepsilon$ and $\max_i \| a_i\|$.
From \eqref{eq:tr1}, with $A(\gamma)$ as in Lemma \ref{le:agamma} and $\mathcal {P}_{\ell,\theta} (v,e_1)$ as in Lemma \ref{le:Plclass}, we find
\begin{equation*}
\mathbf {E} \PAR{ \| B^{\ell} \|^{2\theta} } \leq \sum_{v =1}^{q(\ell\theta +1)} \sum_{e_1 = 0}^\infty | \mathcal {P}_{\ell,\theta} (v,e_1)| \max_{\gamma \in \mathcal {P}_{\ell,\theta}(v,e_1)} p(\gamma)A(\gamma),
\end{equation*}
where we have used, by Lemma \ref{le:pgamma} that $p(\gamma) = 0$ unless $v \leq q (\ell\theta +1)$. Set $\chi = q (\ell \theta +1) +e_1/2 - v$.
By Lemma \ref{le:pgamma} and Lemma \ref{le:agamma}, we have,
\begin{equation}\label{eq:pa00}
p(\gamma)A(\gamma) \leq n^{ - q \ell \theta} \PAR{ (c q \ell \theta)^{q/2} n^{-1/8}}^{b + \frac{e_1}{q}} \left( c q\ell \theta \right) ^{ 2\chi } n^v C^{b + q\theta + \chi} \rho^{2 \ell\theta},
\end{equation}
where $b$ is the number of isolated brackets of $\gamma$. For our choice of parameters in \eqref{eq:choixql}, for any constant $c >0$, we have
for all $n$ large enough: $(cq)^q \leq n^{1/2^7}$. It follows that, if $\delta$ is small enough, for all $n$ large enough,
$$
(c q \ell \theta)^{q/2} n^{-1/8} \leq n^{1/2^7 + \delta/2 -1/8} \leq n^{-1/10}.
$$
We deduce that for all $n$ large enough, the factor in $b$ in \eqref{eq:pa00} is bounded by $1$. Therefore, for all $n$ large enough, we have
$$
\max_{\gamma \in \mathcal {P}_{\ell,\theta}(v,e_1)} p(\gamma)A(\gamma) \leq n^{ - q \ell \theta } n^{-\frac{e_1}{10q}} \left( c q \ell \theta \right) ^{ 2\chi } n^v C^{ q\theta + \chi} \rho^{2 \ell\theta}.
$$
We set $r = q (\ell \theta +1) -v$ and use Lemma \ref{le:Plclass}. Since $v \leq 2 q \ell \theta$, we obtain, for some new constant $C >0$,
\begin{equation*}
\mathbf {E} \PAR{ \| B^{\ell} \|^{2\theta} } \leq \rho^{2 \ell\theta } C^{q\theta} ( (q \ell \theta)^9 n)^{ q} \sum_{r=0}^{\infty} \left( \frac{ ( C q \ell \theta )^8}{ n}\right)^r \sum_{e_1 = 0}^\infty \left( \frac{ (C q\ell \theta) ^{4 } } {n^{\frac{ 1}{10q}}} \right)^{e_1}.
\end{equation*}
For our choices of parameters in \eqref{eq:choixql}, for all $n$ large enough,
$$
(C q\ell \theta) ^{4 q} \leq (2C C_0 q^2 \theta )^{4q} \leq n^{1/2^4 + 4 \delta}.
$$
This last expression is bounded by $n^{1/12}$ for $\delta$ small enough. In particular, the above geometric series in $e_1$ converges and
$$
\sum_{e_1 = 0}^\infty \left( \frac{ (C q\ell \theta) ^{4 } } {n^{\frac{ 1}{10q}}} \right)^{e_1} \leq \sum_{e_1 = 0}^\infty n^{-\frac{ e_1 }{60q} } \leq \frac{1}{1 - n^{-\frac{1}{60q}}} \leq 2.
$$
for all $n$ large enough. Similarly, if $\delta$ is small enough, the series in $r$ converges and is bounded by $2$. We fix such choice of $\delta >0$; we deduce that for all $n$ large enough,
$$
\mathbf {E} \PAR{ \| B^{\ell} \|^{2\theta} } \leq 4 \rho^{2 \ell\theta } C^{q\theta} ( (q \ell \theta)^9 n)^{ q}.
$$
We fix the constant $C_0$ in \eqref{eq:choixql} such that $C^{q\theta/(2\ell\theta)} \leq C^{1/(2C_0)} \leq 1+\varepsilon$. We note that
$$
n^{q/(2\ell \theta)} \leq \exp \PAR{\frac{\ln n}{2C_0 n^{\delta/q}} }.
$$
If we further assume that $q \leq \max (\delta/2 , 2^{-7}) \ln n / (\ln \ln n)$, we get that $n^\theta \geq n^{\delta / q} \geq (\ln n)^2$. It follows that for these choices of parameters, for all $n$ large enough,
$$
4 ( (q \ell \theta)^9 n)^{ q} \leq ( 1+ \varepsilon)^{2\ell \theta}.
$$
We thus have checked that
$$
\mathbf {E} \PAR{ \| B^{\ell} \|^{2\theta} } \leq \PAR{ \rho (1+\varepsilon)^2 }^{2 \ell \theta}.
$$
We have $\rho(B_\star) \leq \| B_\star\| \leq C$ for some $ C >0$. Also, from \eqref{eq:defrho}, we have $\rho \leq \rho(B_\star) + \varepsilon$. We get $$\mathbf {E} \PAR{ \| B^{\ell} \|^{2\theta} } \leq (\rho(B_\star) + \varepsilon')^{2\ell \theta},$$
where $\varepsilon'$ can be taken arbitrarily small if $\varepsilon$ is small enough. From Markov inequality, we obtain that
$$
\mathbb{P} \PAR{ \| B^{\ell} \|^{1/\ell} \geq (1+\varepsilon')(\rho(B_\star) + \varepsilon') }\leq (1+\varepsilon')^{-2\ell\theta}.
$$
We easily obtain the required statement by adjusting the value of $\varepsilon$ and the constants.
\end{proof}
\subsection{Proof of Theorem \ref{th:main}}
The orthogonal projection of $A$ defined in \eqref{eq:defA} onto $H_r^\perp$ can be written as
\begin{equation}\label{eq:def[A]}
[A] = a_0 \otimes 1 + \sum_{i=1}^{2d} a_i \otimes [V_i].
\end{equation}
We fix $\varepsilon >0$ and let $\delta = \delta(\varepsilon)$ be as in Theorem \ref{prop:edgeAB}(ii). In view of Theorem \ref{prop:edgeAB}, the event
$$
\BRA{ \| A_{|H_r^\perp} \| \geq \| A_\star \| + \varepsilon }
$$
is contained in the event
$$
{\mathcal{E}}_\varepsilon = \bigcup_{a = (a_1,\ldots a_{2d}) \in S^{2d}_\varepsilon} {\mathcal{E}}_{\delta}(a) ,
$$
where
$$
{\mathcal{E}}_\delta (a) = \{ \rho(B) \geq \rho(B_\star) + \delta \},
$$ with $B = B(a)$ is an in \eqref{eq:defBV2}, $B_\star(a) = B_\star$ is as in \eqref{eq:defBstar} and
$$S_\varepsilon = \{ b \in M_r(\mathbb {C}) : \| b \| \leq \varepsilon^{-1} \}.$$
To prove Theorem \ref{th:main} it is thus sufficient to check that for any $\varepsilon >0$, for all $n$ large enough,
\begin{equation}\label{eq:tdb0}
\mathbb{P} ( {\mathcal{E}}_\varepsilon) \leq n^{-2}.
\end{equation}
To prove \eqref{eq:tdb0}, we need to use a net on $S_\varepsilon^{2d}$. A similar argument appears in \cite{MR4024563}.
Due to the lack of uniform continuity of spectral radii, we perform the net
argument with operator norms. From \eqref{eq:defBV2}, for $a \in M_r(\mathbb {C})^{2d}$, the map $a \mapsto B(a)$ is linear and
$ \| B(a ) \| \leq (2d-1) \| a \|
$,
where
$$
\| a \| = \sum_{i =1}^{2d} \| a_i \|.
$$
Let $\ell = \lfloor C q \rfloor$ be as in Theorem \ref{th:FKB}. The map $a \mapsto B^\ell (a)$ satisfies a deviation inequality
\begin{eqnarray}
\| B^{\ell} (a ) - B^{\ell} (a') \| & \leq & \ell \max( \| B (a) \| , \|B (a') \| )^{\ell-1} \| B (a - a') \| \nonumber \\
&\leq &\ell (2d-1)^\ell \max( \| a \| , \| a' \| ) ^{\ell-1} \| a - a' \|. \label{eq:LIPB}
\end{eqnarray}
For a given $\eta >0$, the net $N_\eta$ of $S_\varepsilon^{2d}$ is built as follows. First, since all matrix norms are equivalent and
$M_r( \mathbb {C}) \simeq \mathbb {R}^{2r^2}$, there exists a subset $N^1 _{\eta} \subset \{ b \in M_r(\mathbb {C}) : \| b\| \leq \varepsilon^{-1}\}$ of cardinality
at most $( c / (\varepsilon \eta) )^{2r^2}$ such that for
any $b \in S_\varepsilon$, there exists $b' \in N_\eta$ with $\| b - b' \| \leq \eta$ (the constant $c$ depends
on $r$). We set $N_\eta = (N^1_\eta)^{2d}$. From \eqref{eq:LIPB}, for some new constant $c >0$ (depending on $\varepsilon,r,d$), for any
$a \in S_\varepsilon$, there exists $a' \in N_\eta$ such that
$$
\| B^{\ell} (a ) - B^{\ell} (a') \| \leq \ell (2d-1) \PAR{\frac{ 2d-1}{ \varepsilon}}^{\ell -1} \eta \leq c^\ell \eta.
$$
Besides, from \eqref{eq:src}, for all $\eta \leq \eta_0$ small enough,
\begin{equation}\label{eq:contrhoB}
\left| \rho(B_\star(a)) - \rho(B_\star (a')) \right|\leq \frac \delta 3,
\end{equation}
If $\eta=\min(\eta_0, (\delta / 3 c)^\ell )$ and $\mathcal E_{\delta / 3} (a') $ does not hold, we deduce that
\begin{eqnarray*}
\| B^{\ell} (a) \| & \leq & \| B^{\ell} (a')\| + \| B^{\ell} (a ) - B^{\ell} (a') \| \\
& < & \PAR{ \rho(B_\star (a')) + \frac \delta 3 }^\ell + \PAR{\frac \delta 3 }^\ell \\
& < & \PAR{ \rho(B_\star (a')) +\frac {2\delta} 3 }^\ell \\
& < & \PAR{ \rho(B_\star (a)) +\delta }^\ell,
\end{eqnarray*}
where we have used \eqref{eq:contrhoB} at the last line.
We find that, for our choice of $\eta$,
$$
{\mathcal{E}}_\varepsilon = \bigcup_{a \in S_\varepsilon^{2d}} \mathcal E_\delta (a) \subset \bigcup_{a \in N_\eta} \mathcal E _{\delta / 3} (a) ,
$$
and, for some $c_1 >0$ (depending on $\varepsilon$, $r$ and $d$),
\begin{equation*}\label{eq:epsnet}
| N_\eta| \leq c_1 ^\ell .
\end{equation*}
We may now use the union bound to obtain an estimate of \eqref{eq:tdb0}:
\begin{eqnarray*}
\mathbb{P} \PAR{ \mathcal E_\varepsilon } & \leq & \sum_{a \in N_\eta} \mathbb{P}\PAR{ \mathcal E _{\delta / 3} (a)} \\
& \leq & |N_\eta | C \exp ( -(\ln n)^2 ) ,
\end{eqnarray*}
where at the second line, we have used Theorem \ref{th:FKB} (the constant $C$ depends on $\varepsilon,d,r$). For our choice of $\ell$, we have
$|N_{\eta}| \leq c_1^\ell \leq n$ for all $n$ large enough.
The bound \eqref{eq:tdb0} follows. \qed
\section{Appendix: Effective linearization for unitaries}
\label{sec:appendix}
To prove the strong convergence, we rely on the following variant of a Theorem by Pisier:
\begin{theorem}[\cite{MR1401692}, Proposition 6]\label{thm-pisier}
Let $A,B$ be two unital $C^*$-algebras generated respectively by unitaries $\{U_i,i\in \INT{d}\}$ and
$\{V_i,i\in \INT{d}\}$. For convenience, define $U_{-i}:=U_i^*$, $V_{-i}:=V_i^*$, $U_0=1_A$, $V_0=1_B$.
Assume that, for any $k\in\mathbb{N}_*$ and for any $a_{-d},\ldots ,a_d\in M_k(\mathbb {C})$,
$$||\sum_{i=-d}^da_i\otimes U_i||=||\sum_{i=-d}^da_i\otimes V_i||,$$
then the map $U_i\to V_i$ extends to a unital $*$-isomorphism between $A$ and $B$.
\end{theorem}
In particular, it means that given a non-commutative polynomial in $U_i$, its norm is determined by the collection of
all $\{||\sum_{i=-d}^da_i\otimes U_i||,a_i\in M_k(\mathbb {C}),i\in\{-d,\ldots, d\}, k\in\mathbb{N}_*\}$. However, the theorem does not make this dependence explicit. This section aims to make this dependence explicit and robust enough to allow concrete estimates for convergence speeds
for any non-commutative polynomial.
Before, let us outline how we use Theorem \ref{thm-pisier}. We replace $A$ by a sequence of unital $C^*$-algebras $A^{(n)}$ and
$U_i$ by $U_i^{(n)}$.
Let $\mathbb{C}[{\mathbb{F}_d}]$ be the $*$-algebra of the free group and $P\in \mathbb{C}[{\mathbb{F}_d}]$, where the canonical generators of ${\mathbb{F}_d}$ are $u_1,\ldots , u_d$. The map $u_i\mapsto U_i$ generates a $*$-homomorphism $\mathbb{C}[{\mathbb{F}_d}]\to A$
which we denote by $P\mapsto P_U$. Likewise, we have a $*$-homomorphism $\mathbb{C}[{\mathbb{F}_d}]\to A$
which we denote by $P\mapsto P_V$.
If, for any $k\in\mathbb{N}_*$ and for any $a_{-d},\ldots ,a_d\in M_k(\mathbb {C})$,
$$\lim_n||\sum_{i=-d}^da_i\otimes U_i^{(n)}||=||\sum_{i=-d}^da_i\otimes V_i||,$$
then
$$\lim_n||P_{U^{(n)}}||=||P_V||.$$
Indeed, if consider the $C^*$-algebra $A=\prod_{n\ge 1}A^{(n)}/\mathcal{I}$
where $\mathcal{I}$ is the two-sided ideal of sequences of elements in $A^{(n)}$ whose norm tends to zero
as $n\to \infty$, and call $U_i$ the image of $(U_i^{(n)})_{n\ge 1}$ under taking the quotient.
Then, since finite-dimensional matrix algebras are exact, the $C^*$ subalgebra of $A$ generated by $U_i$ satisfies the hypotheses of Pisier's proposition. Thus, the isomorphism of the conclusion and the definition of $A$ ensures that
$\lim_n||P_{U^{(n)}}||=||P_V||$ for any $P\in \mathbb{C}[{\mathbb{F}_d}]$.
As mentioned above,
Pisier's theorem is non-constructive. Let us outline the idea of its proof. In the language of operator spaces, the map
$U_i\to V_i$ is a completely contractive map from the operator space $\text{span} (U_i)$ to $\text{span} (V_i)$.
Therefore, the Arveson-Wittstock extension theorem applies and yields an unital completely contractive map from
$A$ to $B$. The classification of completely contractive maps from $A$ to $B$ (Stinespring theorem) and the fact
that the generators are unitaries implies that this complete contraction must be an isometric homomorphism,
which concludes the proof.
Let us digress and compare with Haagerup and Thorbj{\o}rnsen's result in \cite{MR2183281}, where the first statement of a linearization argument can be found for selfadjoint elements. Two remarks are in order. As for Pisier's result, the initial version is not constructive.
A constructive version has been found later, see, e.g., \cite{MR3585560}, Chapter 10 for a comprehensive explanation.
Interestingly, Pisier's assumptions are simpler because only an equality on the norm of linear pencils is required. In contrast, for Haagerup and Thorbj{\o}rnsen, an equality of spectrum is needed (which looks stronger at first sight).
As far as we can tell, a constructive proof of Pisier has not been given yet. We fill in this gap by providing a linearization. It is to be noted that this linearization does not require knowledge about the inverse of matrices.
The first lemma proves that it is enough to restrict to selfadjoint pencils and that the result also extends to
all rectangular matrix coefficients.
\begin{lemma}\label{lem-selfadjoint-is-enough}
The following are equivalent:
\begin{enumerate}
\item For any $a_{-d},\ldots ,a_d\in M_k(\mathbb {C})$,
$$||\sum_{i=-d}^da_i\otimes U_i||=||\sum_{i=-d}^da_i\otimes V_i||.$$
\item For any $a_{-d},\ldots ,a_d\in M_{N,k}(\mathbb {C})$,
$$||\sum_{i=-d}^da_i\otimes U_i||=||\sum_{i=-d}^da_i\otimes V_i||.$$
\item For any $a_{-d},\ldots ,a_d\in M_k(\mathbb {C})$, $a_{-d},\ldots ,a_d\in M_k(\mathbb {C})$, $a_i=a_{-i}^*$.
$$||\sum_{i=-d}^da_i\otimes U_i||=||\sum_{i=-d}^da_i\otimes V_i||.$$
\end{enumerate}
\end{lemma}
\begin{proof}
The equivalence between the first point and the second point follows from the fact that padding a rectangular matrix
with zeros into a square matrix does not change its operator norm.
The equivalence between the first point and the third point follows from the fact that for any operator $A: L(H_1,H_2)$,
$$||A||=
||\begin{pmatrix}
0 & A \\
A^* & 0
\end{pmatrix}||.$$
The conclusion follows.
\end{proof}
The second lemma proves that it is enough to restrict to positive pencils.
\begin{lemma}\label{lem-positive-is-enough}
The following are equivalent:
\begin{enumerate}
\item For any $a_{-d},\ldots ,a_d\in M_k(\mathbb {C})$,
$$||\sum_{i=-d}^da_i\otimes U_i||=||\sum_{i=-d}^da_i\otimes V_i||.$$
\item For any $a_{-d},\ldots ,a_d\in M_k(\mathbb {C})$ such that
$\sum_{i=-d}^da_i\otimes u_i\ge 0$ is positive as an element of $\mathbb{C}[{\mathbb{F}_d}]$ (i.e. a finite sum of selfadjoint squares), then
$$||\sum_{i=-d}^da_i\otimes U_i||=||\sum_{i=-d}^da_i\otimes V_i||.$$
\end{enumerate}
\end{lemma}
\begin{proof}
The above follows from the following fact from spectral theory:
for a selfadjoint element $P$ in a unital $C^*$-algebra $A$,
for any $c>0$, if one knows $||x1+P||$ for any $x\in \mathbb{R}, |x|\ge c$, then one knows
$||P||$. This concludes the proof, along with the fact that $||P||=||-P||$.
\end{proof}
We also need the following crucial technical lemma
\begin{lemma}\label{lem-specific-to-unitaries}
Let $G\subset F_d$ be a finite, symmetric subset of the free group (i.e., $G^{-1}=G$). Consider the set
$G^2=\{gh, g\in G, h\in G\}$.
Let $Q=\sum_{g\in G^2}a_g\otimes \lambda_g \in M_k(\mathbb {C} )\otimes \mathbb{C}[{\mathbb{F}_d}]$ be a self-adjoint element (equivalently,
$a_g^*=a_{g^{-1}}$).
Then, there exists $N\ge k$ such that for all $c >0$ large enough, there exists $P = \sum_{g \in G} b_g\otimes \lambda_g \in M_{N,k}(\mathbb {C} )\otimes \mathbb{C}[{\mathbb{F}_d}]$ which satisfies
$P^*P=Q+c1$.
\end{lemma}
\begin{proof}
Given $Q=\sum_{g\in G^2}a_g\otimes \lambda_g \in M_k(\mathbb {C} )\otimes C[{\mathbb{F}_d}]$, pick
a selfadjoint element
$\tilde Q\in \mathrm{End} (\underbrace{\mathbb{C}^k\oplus \cdots\oplus\mathbb{C}^k}_{\text{$|G|$ times}})$
selfadjoint
that satisfies the following property: writing $\tilde Q=(\tilde Q_{g,h})_{g,h\in G}$,
$\sum_{g,h\in G, g^{-1} h=r}\tilde Q_{g,h}=a_r$.
Such a $\tilde Q$ can always be found
(for example, take $\tilde Q_{g,h}= a_{g^{-1} h} / |\{(g',h')\in G^2, (g')^{-1}h'=g^{-1} h\}|$). Let $c>0$ be large enough to ensure that $\tilde Q+ c1$ is positive and let
$\tilde P$ be a selfadjoint square root of $\tilde Q+ c1$.
Write $\tilde P= (\cdots |\tilde P_g |\cdots )_{g\in G}$ where $\tilde P_g\in M_{k|G|,k}(\mathbb {C} ) $ is a ``block column vector'', i.e.
an operator $\mathbb{C}^k\to \underbrace{\mathbb{C}^k\oplus \cdots\oplus\mathbb{C}^k}_{\text{$|G|$ times}}$.
Then, the operator $P=\sum_{g\in G} \tilde P_g\otimes \lambda_g$ in $M_k(\mathbb {C} )\otimes \mathbb{C}[{\mathbb{F}_d}]$ satisfies
$$P^*P= \sum_{g,h \in G} \tilde P_g^* \tilde P_h \otimes \lambda(g^{-1} h) = \sum_{g,h \in G} ( \tilde Q_{g,h} + c \IND_{g = h} \cdot 1_k) \otimes \lambda(g^{-1} h) = Q + c |G| 1,$$
and this concludes the proof.
\end{proof}
Following notations of operator systems, we call
$$S_k(U)=\text{span} \{\sum_{i=-d}^da_i\otimes U_i, a_i\in M_k(\mathbb {C})\},$$
and $S_k(U)^{(l)}$ the vector space spanned by the product of all $l$-tuples of elements in $S_k(U)$.
By definition, $S_k(U)=S_k(U)^{(1)}$. In addition, one checks that $S_k(U)^{(l)}\subset S_k(U)^{(l+1)}$ and
$\cup_{l\ge 1} S_k(U)^{(l)}$ is $M_k(\mathbb {C})$ tensored by the $*$-algebra generated by $U_i$.
For what follows, we extend the homomorphism $\mathbb{C}[F_d]\to A$ by tensorising by matrices, and we keep the same notation
$P\mapsto P_U$ (resp. $P\mapsto P_V$ for the homomorphism $M_k(\mathbb {C} )\otimes\mathbb{C}[{\mathbb{F}_d}]\to M_k(\mathbb {C} )\otimes A$).
The following lemma concludes the proof
\begin{lemma}
Consider the map $S_k(U)^{(l)}\subset M_k(\mathbb {C} )\otimes A$ to $S_k(V)^{(l)}\subset M_k(\mathbb {C} )\otimes B$
defined as follows: for $\tilde Q\in S_k(U)^{(l)}$, pick an
element $Q=\sum_{g\in G^{2l}}a_g\otimes \lambda_g \in M_k(\mathbb {C} )\otimes \mathbb{C}[{\mathbb{F}_d}]$ such that
$Q_U=\tilde Q$. Then, the image of $\tilde Q$ is $Q_V$. Assume that this map is well-defined, linear, and
isometric for all $k$, then the same result holds if one replaces $l$ by $2l$.
\end{lemma}
\begin{proof}
Take $G=\{u_{-d},\ldots \ldots, u_d\}$ where $u_1, \ldots, u_d$ are generators
of the free group $F_d$ and $u_{-i}=u_i^{-1}, u_0=e$.
Consider a self-adjoint element $Q=\sum_{g\in G^{2l}}a_g\otimes \lambda_g \in M_k(\mathbb {C} )\otimes \mathbb{C}[{\mathbb{F}_d}]$.
Thanks to Lemma \ref{lem-selfadjoint-is-enough} (point 3), it will be enough to prove that $||Q_U||=||Q_V||$.
Given $Q\in M_k(\mathbb {C} )\otimes\mathbb{C}[{\mathbb{F}_d}]$ its images are $Q_U\in A$ and $Q_V\in B$.
Then by Lemma \ref{lem-specific-to-unitaries}, there exists an integer $N$ such that for all $c$ large enough,
$$Q+c1=P^*P,$$ where $P=\sum_{g\in G^{l}}b_k \otimes \lambda_g
\in M_{N,k}(\mathbb{C})\otimes \mathbb{C}[{\mathbb{F}_d}]$.
Then, by Lemma \ref{lem-selfadjoint-is-enough} (point 2) we have $||P_U||=||P_V||$. The same is true for $-Q$. Therefore, by the $C^*$ norm axiom,
$||Q_U+c1_A||=||Q_V+c1_B||$ for any $|c|$ large enough. Thanks to Lemma \ref{lem-positive-is-enough}, this
implies that $||Q_U||=||Q_V||$ and we have extended the isometry.
To conclude, the representations $\mathbb{C}[{\mathbb{F}_d}]\to A$ and $\mathbb{C}[{\mathbb{F}_d}]\to B$, given respectively by $Q\mapsto Q_U$ and $Q\mapsto Q_V$
yield the $C^*$-norm on $\mathbb{C}[{\mathbb{F}_d}]$, therefore the map $U_i\mapsto V_i$ extends to an isomorphism $A\to B$. \end{proof}
Two remarks are in order. Firstly, the above argument is specific to unitaries. Indeed, if one replaces unitaries with selfadjoint elements,
the proof of the lemma \ref{lem-specific-to-unitaries} breaks down. For example, it does not a priori allow to construct
$X_1^2-X_2^2 +c1$ as a square of polynomials.
Secondly, by following the same reasoning, the lemma can be quantitatively improved as follows. If instead
of an equality, assume we have an inequality of the following type: there exists $\varepsilon_n >0$ tending to zero as $n\to\infty$ such that
$||P_{U^{(n)}}||\le ||P_V||(1+\varepsilon_n)$. Then,
$||Q_U+c1_A|| \le ||Q_V+c1_B||(1+\varepsilon_n)^2$.
Since an upper bound on an appropriate $c$ can be given as a function of the norm of each matrix coefficient
attached to a group element, there exists a constant $C$ that depends explicitly on
$P\in M_k(\mathbb {C} )\otimes\mathbb{C}[{\mathbb{F}_d}]$ (and on nothing else)
such that for $n$ large enough,
$$||Q_{U^{(n)}}|| \le ||Q_V||(1+C \varepsilon_n).$$
Naturally, an inequality in the opposite direction can be obtained similarly.
Iterating, we obtain the following
\begin{proposition}
If there exists a sequence tending to zero $\varepsilon_n$ such that
any $k\in\mathbb{N}_*$ and for any $a_{-d},\ldots ,a_d\in M_k(\mathbb {C})$,
$$||\sum_{i=-d}^da_i\otimes U_i^{(n)}||-||\sum_{i=-d}^da_i\otimes V_i||=O(\varepsilon_n),$$
then the same holds at the level of polynomials:
$$||P_{U^{(n)}}||-||P_V||=O(\varepsilon_n).$$
\end{proposition}
In addition, uniform bounds in the $a_i$'s on the speed of convergence of $||\sum_{i=-d}^da_i\otimes U_i^{(n)}||$ are
not needed, and the final estimate $||P_{U^{(n)}}||-||P_V||=O(\varepsilon_n)$ can be made explicit as a function of $P$ and the $a_i$'s.
\bibliographystyle{abbrv}
| f8a33f627e353ac19be3ac39e5557c394cae23de | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Consider a compact Riemannian manifold $(M,g)$, an open subset $\omega\subseteq M$ and the \textit{Laplace-Beltrami} operator (or simply the Laplacian) $\Delta$ on $M$ associated to the metric $g$. Since $-\Delta$ is a non-negative self-adjoint operator on $L^2(M)$, it is possible to define the fractional Laplacian $(-\Delta)^{\alpha/2}$ for $\alpha>0$ using the spectral theorem.
In this article we are interested in understanding the validity of various observability estimates associated to $(-\Delta)^{\alpha/2}$ or, more generally, to the fractional Schrödinger operator $(-\Delta)^{\alpha/2}+V$, where $V$ is a bounded real potential.
We first consider the fractional Schrödinger evolution equation:
\begin{equation}\label{e:fschro}
\left\{
\begin{array}{ll}
i\partial_t u(t,x)=(-\Delta )^{\alpha/2}u(t,x)+V(x)u(t,x),& (t,x)\in\IR\times M,\medskip\\
u|_{t=0}=u^0\in L^2(M),&
\end{array}
\right.
\end{equation}
where $\alpha>0$ and $V\in\cC^\infty(M;\IR)$.
The fractional Schrödinger evolution is said to be observable from an open set $\omega\subseteq M$ and time $T>0$ provided a constant $C=C_{T,\omega}>0$ exists such that for every initial datum $u^0\in L^2(M)$ the corresponding solution to \eqref{e:fschro} satisfies the \textit{observability estimate}
\begin{equation}\label{e:obss}\tag{$O(T,\omega)$}
||u^0||_{L^2(M)}^2\leq C \int_0^T\int_\omega |e^{-it((-\Delta )^{\alpha/2}+V)}u^0(x)|^2dx\,dt.
\end{equation}
The observability property is crucial in establishing controllability results for the fractional Schrödinger evolution or uniform stabilization for the semi-groups generated by bounded, non self-adjoint perturbations of $(-\Delta )^{\alpha/2}$, see \cite{LionsBook}.
We are also interested in understanding the validity of a weaker observability property, regarding the eigenfunctions of the fractional Schrödinger operator. Since $M$ is compact, $(-\Delta)^{\alpha/2}+V$ has compact resolvent and therefore its spectrum is discrete and bounded below, and there exists an orthonormal basis of $L^2(M)$ consisting of eigenfunctions of $(-\Delta)^{\alpha/2}+V$.
We say that the eigenfunctions of the fractional Schrödinger operator are observable from an open subset $\omega\subseteq M$ if a constant $C=C_{\omega}>0$ exists such that
\begin{equation}\label{e:obse}\tag{$O_\mathrm{E}(\omega)$}
||\varphi||_{L^2(M)}\leq C_\omega \|\varphi\|_{L^2(\omega)},
\end{equation}
holds for every $L^2(M)$-eigenfunction of the Schrödinger operator:
\[
((-\Delta)^{\alpha/2}+V(x))\varphi(x)=\lambda\varphi(x),\quad x\in M.
\]
Note that it is crucial in this definition that the constant $C_\omega$ is required to be uniform with respect to the eigenvalue $\lambda$.
If $\varphi$ is an eigenfunction of $(-\Delta)^{\alpha/2}+V$ with $\|\varphi\|_{L^2(M)}=1$ and eigenvalue $\lambda$ then:
$u(t,\cdot)=e^{-it\lambda}\varphi$ is a solution to the fractional Schrödinger equation \eqref{e:fschro}.
Therefore \eqref{e:obss} implies \eqref{e:obse}. The observability property for eigenfunctions is interesting as it describes localization properties of high-energy eigenstates, a fundamental question in mathematical quantum mechanics. It has been shown in \cite{PTZ} that this property is related to controllability properties of wave equations with random initial data.
It is known since the seventies that both \eqref{e:obss} and \eqref{e:obse} hold under a geometric condition that relates the geometry of the observability region $\omega$ and the geodesics of the manifold $(M,g)$; more detailed references will be given after the statement of our first result.
Let us first recall the precise definitions. The open subset $\omega\subseteq M$ satisfies the \textit{Geometric Control Condition (GCC)} provided that every geodesic of $(M,g)$ intersects $\omega$. Given $T>0$, we say that $\omega$ satisfies GCC$_T$ whenever every geodesic segment of length smaller than $T$ intersects $\omega$.
Since $M$ is compact, $\omega$ satisfies GCC if and only if it satisfies GCC$_T$ for some $T>0$.
\begin{theo}\label{t:suff}
Let $(M,g)$ be a compact smooth Riemannian manifold without boundary, $T>0$ and $\omega\subseteq M$ open. Then the following results hold.
\begin{itemize}
\item if $0<\alpha<1$ and $\overline{\omega}\neq M$ then \eqref{e:obss} does not hold no matter how large $T>0$ is.
\item if $\alpha=1$ and $\omega$ satisfes GCC$_T$ then \eqref{e:obss} holds. Conversely, if $M\setminus\overline{\omega}$ contains a closed geodesic segment of length $T$ then \eqref{e:obss} does not hold.
\item if $\alpha>1$ and $\omega$ satisfes GCC then \eqref{e:obss} holds for every $T>0$.
\end{itemize}
\end{theo}
We give a unified proof of this result in Section \ref{s:suf}. Some of the statements in Theorem \ref{t:suff} hold for time-dependent potentials; Theorem \ref{t:tdep} in Section \ref{s:suf} presents their precise formulation.
The case $\alpha=1$ corresponds to the wave equation and was first proved by Rauch and Taylor in \cite{RauchTaylor75}. The necessary condition is related to the existence of highly localized solutions to the wave equation known as Gaussian beams or wave-packets, see \cite{RalstonGB, MZ02}. Note that the necessary condition for observability is slightly weaker than GCC$_T$, since geodesics are just required to intersect the larger set $\overline{\omega}$; the sufficiency of this less restrictive condition depends in general of the particular configuration of geodesics that intersect $\overline{\omega}$ precisely at the boundary $\partial\omega$ (the so-called \textit{grazing rays}).
When $\alpha=2$, problem \eqref{e:fschro} is the usual Schrödinger equation. The sufficiency of GCC in this case was established by Lebeau \cite{Lebeau1992}, in the more general setting of manifolds with boundary. We present here a rather direct proof in the boundaryless setting that remains valid for every $\alpha>1$.
Finally, note that when $\alpha>2$ (or $\alpha>1$ but $T$ large enough so that GCC$_T$ holds) and $V=0$ , the sufficiency of GCC can also be deduced from that of the case $\alpha=2$ from abstract, functional-theoretic arguments, see \cite[Theorem 3.5]{Mi12}.
In general, GCC is far from being necessary for the observability property when $\alpha=2$. The simplest setting where this kind of behavior takes place is the flat torus $\IT^d=\IR^d/\IZ^d$, where \eqref{e:obss} holds for every open set $\omega\subseteq \IT^d$. This was first proved by Jaffard \cite{Jaffard1990} for the free Laplacian when $d=2$, and generalized to the multidimensional case by Komornik \cite{Komornik1992}. The addition of a potential does not change the final result: see \cite{BZ12, BBZ14} for results on the two-dimensional case $d=2$ and \cite{AM14} for the general case. The article \cite{AFM15} provides more general results that encompass the general case $\alpha>1$ on the torus $\IT^d$.
Other situations where observability for the Schrödinger equation has been established include Zoll manifolds (\textit{i.e.} manifolds all of whose geodesics are closed) \cite{MaciaDispersion, MacRiv16} and the Euclidean disk \cite{ALM16,ALMCras}. On the disk, a necessary and sufficient condition for observability is that the observability region intersects the boundary of the disk in an open arc. Note that this situation is intermediate between the torus (for which observability holds unconditionally) and the sphere, for which it turns out that GCC is necessary. This is the content of the next result.
\begin{theo}\label{t:nes}
Suppose $(M,g)=(\IS^d,\mathrm{can})$ is the sphere equipped with its canonical metric. Let $\omega\subseteq \IS^d$ be an open set such that $M\setminus \overline{\omega}$ contains a geodesic. Then, for every $\alpha>0$ the following results hold.
\begin{itemize}
\item The observability estimate \eqref{e:obss} does not hold for any $T>0$;
\item If in addition $V=0$, then \eqref{e:obse} fails as well.
\end{itemize}
\end{theo}
The proof of this theorem is given in Section \ref{s:nec}, and covers the case of time-dependent potentials for the first part of the statement. It is a variation of the results in \cite{MaciaDispersion, MacRiv16} for $\alpha=2$, and here we follow closely the strategy of the proofs in those references. The statement concerning the observability of eigenfunctions for the free Laplacian on the sphere is well-known, although we give here a complete proof for the sake of completeness.
It turns out that the conditions under which \eqref{e:obse} holds change dramatically when a perturbation $V$ is present. This was first analyzed in \cite{MacRiv19}, where it is shown that, on the $d$-dimensional sphere, \eqref{e:obse} holds provided $\omega\subseteq \IS^d$ satisfies a geometric condition involving the potential $V$ (what we call here $V$-GCC, see Section \ref{s:vgcc}) that is in general much weaker than GCC. Here we present a proof for fractional Schrödinger operators based on an observability estimate over long times for the fractional Schrödinger evolution, see Theorem \ref{t:lt} in Section \ref{s:vgcc} which is of independent interest. This is based on the detailed study of the long-time dynamics of Schrödinger flows presented in \cite{MaciaAv, MacRiv16}.
Our last result, valid in two dimensions, gives an explicit construction of a family of smooth potentials on the sphere such that \eqref{e:obse} holds when $\omega$ is an arbitrarily small neighborhood of two fixed points on the sphere. This is in stark contrast with the same situation for the free Laplacian, for which GCC is a necessary condition for \eqref{e:obse}.
\begin{theo}\label{t:minimal}
Suppose that $(M,g)=(\IS^2,\mathrm{can})$ and $\alpha>0$.There exist a non-empty family of potentials $\cT\subset\cC^\infty(\IS^2;\IR)$ such that, for every $V\in\cT$ there exist two distinct points $p,q\in\IS^2$ such that \eqref{e:obse} holds for any open set $\omega\subseteq \IS^2$ that contains $p,q$.
\end{theo}
This result is proved in Section \ref{s:potential}. As it will be clear from the proof, the size of $V$ in any reasonable norm can be arbitrarily small. The result follows from the analysis of the flow of a certain Hamiltonian vector field on the sphere constructed from the potential $V$.
\section{Observability for the fractional Schrödinger evolution on a compact manifold. Proof of Theorem \ref{t:suff}}\label{s:suf}
Most of the proofs we present remain valid when the potentials are time-dependent. In this section we will consider the fractional Schrödinger equation:
\begin{equation}\label{e:fracschro}
\left\{
\begin{array}{ll}
i\partial_t u(t,x)=(-\Delta )^{\alpha/2}u(t,x)+V(t,x)u(t,x),& (t,x)\in\IR\times M,\medskip\\
u|_{t=0}=u^0\in L^2(M),&
\end{array}
\right.
\end{equation}
where $\alpha>0$ and $V\in\cC^\infty_c(\IR\times M;\IR)$ is bounded together with all its derivatives.
The following observability results still hold in this more general context.
\begin{theo}\label{t:tdep}
Let $(M,g)$ be a smooth compact Riemannian manifold without boundary and $\omega\subseteq M$ an open set. Suppose that the observability estimate
\begin{equation}\label{e:obst}
||u^0||_{L^2(M)}^2\leq C \int_0^T\int_\omega |u(t,x)|^2dx\,dt.
\end{equation}
holds for every solution $u$ to \eqref{e:fracschro}. Then,
\begin{itemize}
\item when $0<\alpha<1$ the set $\omega$ must be dense in $M$.
\item when $\alpha=1$ the set $\overline{\omega}$ must intersect all closed geodesic segments of length $T$.
\end{itemize}
\end{theo}
The proof of this result, together with that of Theorem \ref{t:suff}, are presented at the end of this section. Its starting point is a compactness argument based on the analysis of semiclassical defect measures. This type of approach goes back to \cite{Lebeau1996}. It will require several preparatory steps.
Given $\chi\in\cC^\infty_c((0,\infty))$ and $h>0$ define:
\[
\cF^h_\chi\,:\,L^2(M)\To C(\IR;L^2(M))\,:\,u^0\longmapsto \chi(-h^2\Delta) u,
\]
where $u$ solves \eqref{e:fracschro} with initial datum $u^0$.
\begin{lem}\label{l:ape}
Let $\chi\in\cC^\infty_c((0,\infty))$ and $\sigma_\alpha\in\cC^\infty_c((0,\infty))$ such that $\sigma_\alpha(s)=s^{\alpha/2}$ for $s\in\supp\chi$. Then for every $u^0\in L^2(M)$ and $h>0$ the function $u_h:=\cF^h_\chi u^0$ solves
\begin{equation}\label{e:sfracschro}
\left\{
\begin{array}{l}
ih^\alpha\partial_t u_h(t,x)=\sigma_\alpha(-h^2\Delta )u_h(t,x)+h^\alpha V(t,x)u_h(t,x)+h^{1+\alpha}F_h(t,x), \medskip\\
u_h|_{t=0}=\chi(-h^2\Delta)u^0,
\end{array}
\right.
\end{equation}
and there exists $C>0$ such that for every $t\in\mathbb{R}$ and $h\in (0,1]$,
\begin{equation}\label{e:estFh}
\|F_h(t,\cdot)\|_{L^2(M)}\leq C\|u^0\|_{L^2(M)}.
\end{equation}
If $u$ is the solution to \eqref{e:fracschro} with initial datum $u^0$ then
\begin{equation}\label{e:estuuh}
\|u(t,\cdot)-u_h(t,\cdot)\|_{L^2(M)}\leq \|(1-\chi(-h^2\Delta) )u^0\|_{L^2(M)}+Ch|t|\|u^0\|_{L^2(M)}.
\end{equation}
\end{lem}
\begin{proof}
Apply $\chi(-h^2\Delta)$ to both sides of equation \eqref{e:fracschro} to find that $u_h$ solves:
\[
i\partial_t u_h=(-\Delta )^{\alpha/2}u_h+ Vu_h+[\chi(-h^2\Delta),V]u.
\]
By definition
\[
(-h^2\Delta )^{\alpha/2}\chi(-h^2\Delta)=\sigma_\alpha(-h^2\Delta )\chi(-h^2\Delta),
\]
and, using the commutator identity \eqref{e:comm} and the functional calculus for the Laplacian \eqref{e:fc},
\[
[\chi(-h^2\Delta),V]u=\frac{h}{i}\Op_h(r)u,
\]
for some $r\in S^{0}(T^*M)$. This proves claims \eqref{e:sfracschro} and \eqref{e:estFh}. Estimate \eqref{e:estuuh} follows from Duhamel's identity.
\end{proof}
The rest of this section will show how the dynamics of $u_h$ when $h$ is small are related to the geodesic flow on the manifold $(M,g)$. We introduce the following notation for the squared Riemannian norm on the cotangent bundle
\[
p(x,\xi):=\|\xi\|^2_x,\quad (x,\xi)\in T^*M.
\]
The geodesic flow $\phi_t$ on $T^*M\setminus\{0\}$ is the flow of the Hamiltonian vector field (taken with respect to the canonical symplectic form) associated to the Hamiltonian $p^{1/2}$. Projections of the trajectories of $\phi_t$ onto $M$ are geodesics of $(M,g)$ parametrized by arc-length.
We next deduce a transport-type equation for the Wigner distributions of solutions to equation \eqref{e:sfracschro} involving the Hamiltonians $p^{\alpha/2}$. Recall that the Wigner distribution $W^h_v\in\cD'(T^*M)$ of a function $v\in L^2(M)$ is defined by
\[
\left\la W^h_v,a\right\ra = (\Op_h(a) v\,|\,v)_{L^2(M)},\quad \forall a\in\cC^\infty_c(T^*M),
\]
where $\Op_h(a)$ stands for the Weyl semiclassical pseudo-differential operator of symbol $a$. Appendix \ref{a:pdo} reviews the facts of the theory of Wigner distributions and semiclassical analysis that are needed here. In what follows, $\{a,b\}$ will denote the Poisson bracket of two functions $a,b$ defined on $T^*M$.
\begin{lem}\label{l:wie}
Let $W_h(t)\in\cD'(T^*M)$ denote the Wigner distribution of the function $\cF^h_\chi u^0(t,\cdot)$ for some $u^0\in L^2(M)$. Then, for every $a\in\cC^\infty_c(T^*M)$:
\begin{equation}\label{e:wignereq}
\frac{d}{dt}\left\la W_h(t),a\right\ra = h^{1-\alpha}\left\la W_h(t),\{\sigma_\alpha\circ p,a\}\right\ra+h^\beta R^h_a(t),
\end{equation}
where $\beta:=\min\{1,2-\alpha\}$ and $|R^h_a(t)|\leq C\|a\|_{\cC^N(T^*M)} \|u^0\|^2_{L^2(M)}$ for some $N>0$ large enough, independent of $a$.
\end{lem}
\begin{proof}
Set $u_h:=\cF^h_\chi u^0$. By definition of the Wigner distribution and equation \eqref{e:sfracschro}:
\begin{equation}\label{e:wignertem}
\frac{d}{dt}\left\la W_h(t),a\right\ra =\frac{1}{ih^\alpha}([\Op_h(a),\sigma_\alpha(-h^2\Delta )] u_h(t,\cdot)|u_h(t,\cdot))_{L^2(M)}+D^h_a(t),
\end{equation}
with
\[
D^h_a(t)=i([V,\Op_h(a)]u_h(t,\cdot)|u_h(t,\cdot))_{L^2(M)}+2h\mathrm{Im}(\Op_h(a)F_h(t,\cdot)|u_h(t,\cdot))_{L^2(M)}.
\]
Now using the symbolic calculus of semiclassical pseudo-differential operators \eqref{e:comm} and the functional calculus \eqref{e:fc}, we deduce that:
\[
[\Op_h(a),\sigma_\alpha(-h^2\Delta )] = \frac{h}{i}\Op_h(\{a,\sigma_\alpha\circ p\})+h^2\Op(r),
\]
for some symbol $r\in S^{0}(T^*M)$ (keep in mind that symbol $\sigma_\alpha\circ p$ is a function in $\cC^\infty_c(T^*M)$). Substituting this in \eqref{e:wignertem} gives \eqref{e:wignereq} with
\[
R^h_a(t)=h^{-\beta}D^h_a(t)+ih^{2-\alpha-\beta}(\Op_h(r)u_h(t,\cdot)|u_h(t,\cdot))_{L^2(M)}.
\]
The estimate then follows from the Calderón-Vaillancourt theorem \eqref{e:cv} and again from \eqref{e:comm}, that allows us to estimate $|D^h_a(t)|\leq Ch \|a\|_{\cC^N(T^*M)} \|u^0\|^2_{L^2(M)}$ for some $N\geq N_d$ ($N_d$ appears in \eqref{e:cv}), which is independent of $a$.
\end{proof}
This result will allow us to characterize semiclassical measures of solutions to \eqref{e:fracschro}. Recall that a semiclassical measure of a sequence $(v_h)_{0<h\leq 1}$ that is bounded in $L^2(M)$ is an accumulation point in $\cD'(T^*M)$ of the corresponding sequence of Wigner distributions $(W^h_{v_h})_{0<h\leq 1}$.
\begin{prop}\label{p:scm}
Let $(u^0_h)_{0<h\leq 1}$ be bounded in $L^2(M)$ and $u_h:=\cF^h_\chi u^0_h$. Denote by $(W_h(t))_{0<h\leq 1}$ the sequence of Wigner distributions of $(u_h(t,\cdot))_{0<h\leq 1}$. The following hold:
\begin{itemize}
\item if $\alpha\in (0,1]$ then there exists a subsequence $(h_n)$ converging to zero along which $(W_{h_n}(t))$ converges in $\cD'(T^*M)$ for every $t\in\IR$ to a semiclassical measure $\mu_t$. In addition:
\[
\mu_t=\left\{ \begin{array}{ll}
\mu_0 & \text{ if }0<\alpha<1,\medskip\\
(\phi_t)_*\mu_0 & \text{ if } \alpha =1;
\end{array} \right.
\]
\item if $\alpha>1$ and $(W_{h_n})$ converges in $\cD'(\IR\times T^*M)$ to a semiclassical measure $\mu_t$ as in \eqref{e:tscm} then, for almost every $t\in\IR$:
\[
\mu_t=(\phi_s)_*\mu_t,\quad \forall s\in\IR.
\]
\end{itemize}
\end{prop}
\begin{proof}
When $\alpha\in (0,1]$, equation \eqref{e:wignereq} shows that $(W_{h})_{0<h\leq 1}$ is uniformly bounded in $\cC^1(\IR;\cD'(T^*M))$. Therefore, there exists a subsequence $(h_n)$ converging to zero such that $(W_{h_n}(t))$ converges for every $t\in\IR$. The resulting accumulation points $\mu_t$ satisfy, as a consequence of \eqref{e:wignereq}, that for every $a\in\cC^\infty_c(T^*M)$
\[
\frac{d}{dt}\int_{T^*M}a(x,\xi)\mu_t(dx,d\xi)=0,\quad \text{ for }0<\alpha<1,
\]
which concludes the proof in this case, and
\[
\frac{d}{dt}\int_{T^*M}a(x,\xi)\mu_t(dx,d\xi)=
\int_{T^*M}\{a,\sigma_1\circ p\}(x,\xi)\mu_t(dx,d\xi),
\quad\text{ for }\alpha=1.
\]
Notice that, by construction of the sequence $(u_h)_{0<h\leq 1}$, on the support of $\mu_t$ one has $\sigma_1\circ p=p^{1/2}$. Therefore $\mu_t$ is transported along the Hamiltonian flow $p^{1/2}$ which is the geodesic flow $\phi_t$.
Suppose now that $\alpha>1$. In this case $(W_{h})_{0<h\leq 1}$ is only bounded in $\cC(\IR;\cD'(T^*M))$, and therefore one cannot expect to have pointwise convergence in $t\in\IR$. Still $(W_{h})_{0<h\leq 1}$ has accumulation points in $\cD'(\IR\times T^*M)$ (see Remark \ref{r:tbdd}), that are described by \eqref{e:tscm}. In our case this yields:
\[
\int_\IR\theta(t)\left\la W_{h_n}(t),a\right\ra dt \Tend{n}{\infty} \int_\IR\int_{T^*M}\theta(t)a(x,\xi)\mu_t(dx,d\xi)dt,
\]
for all $\theta\in \cC^\infty_c(\IR)$, $a\in\cC^\infty_c(T^*M)$.
Note that, after multiplication by $\theta\in \cC^\infty_c(\IR)$ and integration by parts, \eqref{e:wignereq} can be rewritten as:
\[
-h^{\alpha-1}_n\int_\IR\theta'(t)\left\la W_{h_n}(t),a\right\ra dt=\int_\IR\theta(t)\left\la W_{h_n}(t),\{\sigma_\alpha\circ p,a\}\right\ra dt+\cO(h_n).
\]
Taking limits and noticing that, as before, $\sigma_\alpha\circ p=p^{\alpha/2}$ on the support of $\mu_t$ for almost every $t\in\IR$, we deduce that:
\[
\int_\IR\int_{T^*M}\theta(t)\{a, p^{\alpha/2}\}(x,\xi)\mu_t(dx,d\xi)dt=0,\quad \forall \theta\in \cC^\infty_c(\IR),\,\forall a\in\cC^\infty_c(T^*M).
\]
This identity implies that $\mu_t$ is, for a.e. $t\in\IR$, invariant by the Hamiltonian flow associated to $p^{\alpha/2}$, which is merely a reparametrization of the geodesic flow $\phi_t$. The claim then follows.
\end{proof}
Let $(x_0,\xi_0)\in S^*M$ (recall that this means that $\|\xi_0\|_{x_0}=1$). A wave-packet centered at $(x_0,\xi_0)$ is a family $(u_h^{(x_0,\xi_0)})_{0<h\leq 1}$ of functions supported in a coordinate chart $(U,\varphi)$ of $M$ with $x_0\in U$ of the form:
\[
u_h^{(x_0,\xi_0)}(x)=\frac{1}{h^{d/4}}\rho\left(\frac{\varphi(x)-\varphi(x_0)}{\sqrt{h}}\right)e^{i\frac{(d\varphi_{x_0}^{-1})^T\xi_0}{h}\cdot \varphi(x)},\quad x\in U,
\]
where $\rho\in \cC^\infty_c(\varphi(U))$ and normalized to have $\|u^{(x_0,\xi_0)}_h\|_{L^2(M)}=1$.
Let $\chi\in\cC^\infty_c((0,\infty);[0,1])$ such that $\chi|_{[1,2]}=1$, and $\chi(s)=0$ for $s<1/2$ or $s>5/2$ and write $\Pi_h:=\chi(-h^2\Delta)$.
\begin{lem}\label{l:wp}
For every $(x_0,\xi_0)\in S^*M$, the sequence $(W^{(x_0,\xi_0)}_h)_{0<h\leq 1}$ of Wigner distributions of a frequency-localized wave-packet $(\Pi_h u_h^{(x_0,\xi_0)})_{0<h\leq 1}$ converges in $\cD'(T^*M)$ to the Dirac mass
$\delta_{(x_0,\xi_0)}$.
\end{lem}
\begin{proof}
A direct computation shows that the Wigner distributions
of $u_h^{(x_0,\xi_0)}$ converge to $\delta_{(x_0,\xi_0)}$. Identities \eqref{e:prod} and \eqref{e:fc} imply that $(W^{(x_0,\xi_0)}_h)_{0<h\leq 1}$ converges to
\[
(\chi\circ p)^2\delta_{(x_0,\xi_0)}=\delta_{(x_0,\xi_0)},
\]
as claimed.
\end{proof}
We now give the proofs of Theorems \ref{t:tdep} and \ref{t:suff}.
\begin{proof}[Proof of Theorem \ref{t:tdep}]
We start with the case $0<\alpha<1$. Let $x_0$ be a point in the open set $M\setminus\overline{\omega}$, and
\[
u^0_h:=\Pi_h u_h^{(x_0,\xi_0)},\quad \xi_0\in S^*_{x_0}M.
\]
Let $u_h$ be the solution to \eqref{e:fracschro} issued from $u^0_h$. If $\chi$ is the cut-off used to define $\Pi_h$ then $\Pi_h u_h=\cF^h_\chi u^0_h$ and Lemma \ref{l:ape} implies that:
\begin{equation}\label{e:aeq}
\lim_{h\to 0^+}\|u_h(t,\cdot)-\Pi_h u_h(t,\cdot)\|_{L^2(M)}=0,\quad \forall t\in\IR.
\end{equation}
Let $(h_n)$ be a sequence given by applying Proposition \ref{p:scm} to $(u^0_h)_{0<h\leq 1}$ with the cut-off $\chi$ and let $\mu_t$ the corresponding semiclassical measure. By Lemma \ref{l:wp} and Proposition \ref{p:scm} we know that $\mu_t=\delta_{(x_0,\xi_0)}$ for every $t\in\IR$. Combining this with \eqref{e:proj} and \eqref{e:aeq} we find that, for every $T>0$ and every $b\in\cC(M;[0,1])$ that equals $1$ on $\overline{\omega}$ and vanishes in a neighborhood of $x_0$,
\begin{align*}
\lim_{n\to\infty}\int_0^T\int_\omega|u_{h_n}(t,x)|^2dx\,dt\leq & \lim_{n\to\infty}\int_0^T\int_M b(x)|\Pi_{h_n} u_{h_n}(t,x)|^2dx\,dt\\=&
\int_0^T \int_{T^*M} b(x)\delta_{(x_0,\xi_0)}(dx,d\xi)dt=0.
\end{align*}
Since $\|u^0_{h_n}\|_{L^2(M)}=1$, this shows that no constant $C>0$ can exist such that \eqref{e:obst} holds.
Let us now consider the case $\alpha=1$. The proof follows closely the lines of that of the preceding case with few modifications. Choose now $(x_0,\xi_0)\in S^*M$ such that $\phi_t(x_0,\xi_0)\not\in\omega$ for $t\in [0,T]$. Proposition \ref{p:scm} now shows that the semiclassical measures of the corresponding wave-packets are $\mu_t=\delta_{\phi_t(x_0,\xi_0)}$. The same argument we presented above shows that these wave-packets violate any observability inequality of the form \eqref{e:obst}.
\end{proof}
\begin{proof}[End of the proof of Theorem \ref{t:suff}]
It remains to prove that GCC$_T$ (resp. GCC) are sufficient for \eqref{e:obss} when $\alpha=1$ (resp. $\alpha>1$). Since $V$ does not depend on time, we can use frequency localization and unique continuation for eigenfunctions of Schrödinger operators as in \cite[Proof of Theorem4.1]{Lebeau1992} (see also \cite[Proposition 4.1]{BZ12}) to show that \eqref{e:obss} can be deduced from the \textit{a priori} weaker semiclassical estimate: there exist $h_0>0$ such that, for every $u^0\in L^2(M)$ and $0<h<h_0$,
\begin{equation}\label{e:sobs}
\|\Pi_h u^0\|_{L^2(M)}^2\leq C\int_0^T\int_\omega | e^{-it((-\Delta)^{\alpha/2}+V)} \Pi_h u^0|^2dx\,dt.
\end{equation}
We prove that \eqref{e:sobs} holds by contradiction. If \eqref{e:sobs} fails, then it is possible to find a sequence $(h_n)$ that tends to zero and functions $u^0_{h_n}\in L^2(M)$ such that:
\begin{equation}\label{e:contr}
\|\Pi_{h_n} u^0_{h_n}\|_{L^2(M)}=1,\quad \lim_{n\to\infty}\int_0^T\|e^{-it((-\Delta)^{\alpha/2}+V)} \Pi_{h_n} u^0_{h_n}\|_{L^2(\omega)}=0.
\end{equation}
Modulo the extraction of a subsequence, we can assume that $(W^{h_n}_{u_{h_n}})$ converges to a semiclassical measure $\mu_t$, which by \eqref{e:proj}, \eqref{e:prob} and \eqref{e:contr} satisfies:
\[
\int_0^T\mu_t(T^*M\setminus\{0\})dt=T,\quad\int_0^T\int_{T^*M}b(x)\mu_t(dx,d\xi)dt=0,\quad \forall b\in\cC_c(\omega).
\]
Using Lemma \ref{l:ape} and Proposition \ref{p:scm} we deduce, writing:
\[
F^T_\omega:=\bigcup_{t\in [0,T]}\{\phi_t(x,\xi)\,:\, (x,\xi)\in T^*\omega\setminus\{0\}\},
\]
that $\mu_0(F^T_\omega)=0$ when $\alpha=1$ and $\mu_t(F^s_\omega)=0$ for every $s>0$ and almost every $t\in [0,T]$ when $\alpha>1$. If $\alpha=1$ and $\omega$ satisfies GCC$_T$ this implies $\mu_0(T^*M\setminus\{0\})=0$; whereas if $\alpha>1$ and $\omega$ satisfies GCC it gives $\mu_t(T^*M\setminus\{0\})=0$ for almost every $t\in [0,T]$. This is a contradiction, and the result is proved.
\end{proof}
\section{Observability on the sphere. Proof of Theorem \ref{t:nes}}\label{s:nec}
Here we focus on the particular case $(M,g)=(\IS^d,\mathrm{can})$ and prove, at the end of this section, Theorem \ref{t:nes}. The geodesics on the sphere are great circles, obtained as the intersection of $\IS^d$ with planes through the origin. We normalize the Riemannian metric in order that all the geodesics of the sphere have length equal to $2\pi$. Therefore, the geodesic flow $\phi_t$ is periodic of period $2\pi$.
The \textit{Radon transform} or \textit{X-Ray transform} of a function $a\in\cC^\infty(T^*\IS^d)$ is defined by averaging $a$ along orbits of the geodesic flow:
\begin{equation}\label{e:rt}
\cI(a)(x,\xi)=\frac{1}{2\pi}\int_0^{2\pi}a(\phi_s(x,\xi)) ds, \quad (x,\xi)\in T^*\IS^d\setminus\{0\}.
\end{equation}
As soon as $a\in\cC^\infty_c(T^*\IS^d\setminus\{0\})$ (in particular, $a$ vanishes in a neighborhood of the zero section $\{\xi=0\}$) one can identify $\cI(a)$ to a function in $\cC^\infty_c(T^*\IS^d\setminus\{0\})$ as well.
Recall that the spectrum of $-\Delta$ is:
\[
\Spec(-\Delta)=\{k(k+d-1)\,:\,k\in\mathbb{N}\cup\{0\}\}.
\]
Let $A$ be a positive, self-adjoint operator such that
\[
A^2:= -\Delta+\frac{(d-1)^2}{4};
\]
then $A=\Op_\hbar(\sqrt{p+(d-1)^2/4})+O(h)$ and the spectrum of $A$ equals $\IN+(d-1)/2$.
Then the unitary flow generated by $A$ is essentially periodic of period $2\pi$:
\begin{equation}\label{e:ap}
e^{2i\pi A}=e^{i\pi(d-1)}\text{Id}.
\end{equation}
Given $a$ in $\ml{C}^{\infty}_c(T^*\IS^d\setminus\{0\})$, we define following Weinstein \cite{WeinsteinZoll} the quantum average of the operator $\Op_h(a)$:
$$\cI_{\text{qu}}(\Op_h(a)):=\frac{1}{2\pi}\int_0^{2\pi}e^{-is A}\Op_h(a)e^{isA}ds.$$
Then, it follows from \eqref{e:ap} that
\begin{equation}\label{e:commav}
\left[\ml{I}_{\text{qu}}(\Op_h(a)),A\right]=\left[\ml{I}_{\text{qu}}(\Op_h(a)),(-\Delta)^{\alpha/2}\right]=0.
\end{equation}
We define the averaged Wigner distribution $\cI^*(W^h_u)$ of a function $u\in L^2(\IS^d)$ as:
\[
\left\la \cI^*(W^h_u), a \right\ra :=(\cI_{\text{qu}}(\Op_h(a)) u\,|\, u)_{L^2(\IS^d)} , \quad \forall a\in\ml{C}^{\infty}_c(T^*\IS^d\setminus\{0\}).
\]
One has
\begin{equation}\label{e:wws}
| \left\la \cI^*(W^h_u), a \right\ra - \left\la W^h_u, \cI(a) \right\ra |\leq Ch\|\cI(a)\|_{\cC^N(T^*\IS^d)}\|u\|_{L^2(\IS^d)}^2,
\end{equation}
for some $N\geq N_d$, $N_d$ being defined in the Calderón-Vaillancourt Theorem \eqref{e:cv}, that is independent of $a$. This is a consequence of Egorov's theorem \cite{DimassiSjostrand,F14,Zwobook}, which implies:
\begin{equation}\label{e:egorov}
\ml{I}_{\text{qu}}(\Op_h(a))=\Op_h(\ml{I}(a))+h\Op_h(r),
\end{equation}
for some $r\in S^0(T^*\IS^d)$.
\begin{lem}\label{l:ws}
Let $u^0\in L^2(\IS^d)$ and $u$ be the solution to \eqref{e:fracschro} with initial datum $u^0$. Then, for every $t\in\IR$:
\begin{equation}
\left\la \cI^*(W^h_{u(t,\cdot)}), a \right\ra = \left\la \cI^*(W^h_{u^0}), a \right\ra + hR^h_a(t) ,
\end{equation}
where $|R^h_a(t)|\leq C|t|\|\cI(a)\|_{\cC^N(T^*\IS^d)}\|u^0\|_{L^2(\IS^d)}^2$ for some $N>0$ which is independent of $a$.
\begin{proof}
Using \eqref{e:commav} we deduce that, for every $t\in\IR$ and $\ml{C}^{\infty}_c(T^*\IS^d\setminus\{0\})$,
\begin{align*}
\frac{d}{dt}\left\la \cI^*(W^h_{u(t,\cdot)}),a\right\ra =&\frac{1}{i}([\cI_{\text{qu}}(\Op_h(a)),(-\Delta)^{\alpha/2}+V(t,\cdot)] u(t,\cdot)|u(t,\cdot))_{L^2(M)}\\
=&i([V(t,\cdot),\cI_{\text{qu}}(\Op_h(a))] u(t,\cdot)|u(t,\cdot))_{L^2(M)}.
\end{align*}
To conclude, use \eqref{e:egorov} and \eqref{e:comm} and deduce:
\[
[V(t,\cdot),\cI_{\text{qu}}(\Op_h(a))]=h\Op_h(r_t),
\]
for some $r_t\in S^0(T^*\IS^d)$ uniformly bounded with respect to $t\in\IR$.
\end{proof}
\end{lem}
\begin{proof}[Proof of Theorem \ref{t:nes}]
In order to prove the first statement about the fractional Schrödinger evolution, we take $(x_0,\xi_0)\in S^*M$ such that $\phi_t(x_0,\xi_0)\not\in S^*\omega$ for every $t\in\IR$. Consider the sequence $(u_h)$ of solutions to \eqref{e:fracschro} issued from the initial data $(\Pi_h u^{(x_0,\xi_0)}_h)$. We know by Lemma \ref{l:wp} that $(u_h(0,\cdot))$ has as semiclassical measure $\delta_{(x_0,\xi_0)}$. We use \eqref{e:wws} and Lemma \ref{l:ws} to deduce that the semiclassical measure $\mu_t$ of the sequence $(u_h)$ satisfies for every $\theta\in \cC^\infty_c(\IR)$ and $a\in\cC^\infty_c(T^*M\setminus\{0\})$:
\[
\int_\IR\int_{T^*M}\theta(t)\cI(a)(x,\xi)\mu_t(dx,d\xi)dt=\left(\int_\IR\theta(t) dt\right)\int_{T^*M}\cI(a)(x,\xi)\mu_0(dx,d\xi).
\]
In other words, $\mu_t|_{T^*M\setminus\{0\}}=\cI^*(\mu_0)|_{T^*M\setminus\{0\}}$ for almost every $t\in\IR$.
On the other hand, we know by Proposition \ref{p:scm} that, for almost every $t\in\IR$, $\mu_t$ is invariant by the geodesic flow. Therefore,
\begin{align*}
\int_\IR\int_{T^*M}\theta(t)a(x,\xi)\mu_t(dx,d\xi)dt=&\int_\IR\int_{T^*M}\theta(t)\cI(a)(x,\xi)\mu_t(dx,d\xi)dt\\
=& \left(\int_\IR\theta(t) dt\right)\cI(a)(x_0,\xi_0).
\end{align*}
If $b\in\cC(M;[0,1])$ is such that $b$ vanishes in a neighborhood of the geodesic issued from $(x_0,\xi_0)$ and is equal to one on $\omega$ we conclude, as in the proof of Theorem \ref{t:tdep}, that along some subsequence:
\[
\lim_{n\to\infty}\int_0^T\int_\omega|u_{h_n}(t,x)|^2dx\,dt\leq T\cI(b)(x_0,\xi_0)=0.
\]
Since $\|u_{h_n}(0,\cdot)\|_{L^2(M)}=1$ we conclude that no constant $C>0$ exists such that \eqref{e:obss} holds.\medskip
We now prove the statement concerning eigenfunctions of $(-\Delta)^{\alpha/2}$. Since these eigenfunctions do not depend on $\alpha$, we will assume that $\alpha=2$.
Write the sphere as:
$$\mathbb{S}^d:=\{x\in\IR^{d+1}:|x|=1\}.$$
Let
$$\varphi_k(x)=c_k(x_1+ix_2)^k, \quad \text{ with } c_k:=\sqrt{\frac{\Gamma(k+(d+1)/2)}{2\pi^{\frac{d+1}{2}}k!}}\sim k^{\frac{d-1}{4}}.$$
This function is a spherical harmonic and therefore an eigenfunction of the Laplacian:
$$-\Delta \varphi_k(x)=k(k+d-1)\varphi_k(x),\quad x\in\mathbb{S}^d,\quad ||\varphi_k||_{L^2(\mathbb{S}^d)}=1.$$
Clearly
$$|\varphi_k(x)|^2=(c_k)^2(|x_1|^2+|x_2|^2)^k=(c_k)^2(1-|x'|^2)^k,$$
where $x=(x_1,x_2,x')$. This shows that $|\varphi_k|^2$ concentrates on the equator $\{x'=0\}$.
If $\overline{\omega}\cap\{x'=0\}=\emptyset$ then no constant $C>0$ can exist such that
\begin{equation*}
||\varphi_k||_{L^2(\IS^d)}\leq C \|\varphi_k\|_{L^2(\omega)},
\end{equation*}
holds uniformly in $k\in\IN$, since
$$\lim_{k\to\infty}\int_\omega |\varphi_k(x)|^2dx=0,\quad \text{ and }\quad ||\varphi_k||_{L^2(\mathbb{S}^d)}=1.$$
Since any other geodesic of $\mathbb{S}^d$ can be obtained by applying a rotation to $\{x'=0\}$, and the composition of a spherical harmonic with an Euclidean rotation is again a spherical harmonic, the claim follows.
\end{proof}
\section{Observability over long times and the $V$-GCC}\label{s:vgcc}
As we mentioned in the introduction, observability for eigenfunctions of the fractional Schrödinger operator $(-\Delta)^{\alpha/2}+V$ holds under a geometric assumption on $\omega$ that involves the perturbation $V$. Let us start by describing this new condition.
Identify $V$ to a smooth function on $T^*\IS^d$ that does not depend on the cotangent variable and consider its Radon transform $\cI(V)\in\cC^\infty_c(T^*\IS^d\setminus\{0\})$.
The function $\cI(V)$ defines a Hamiltonian vector field $X_{\cI(V)}$ on $T^*\IS^d\setminus\{0\}$ (with respect to the canonical symplectic form in $T^*\IS^d$). Its flow $\phi^V_s$ commutes with the geodesic flow, since by construction $\cI(V)$ is invariant by the geodesic flow, and therefore $\{\cI(V),p^{1/2}\}=0$.
As a consequence, $\phi^V_s$ maps orbits of the geodesic flow into orbits of the geodesic flow. In other words, for every geodesic $\gamma_0\subseteq T^*\IS^d$
\[
\gamma_s:=\phi^V_s(\gamma_0),
\]
is also a geodesic for every $s\in\IR$. We can thus identify $\phi^V_s$ to a function acting on the space of geodesics on $T^*\IS^d\setminus\{0\}$.
This flow on the space of geodesics induces a new geometric condition on $\omega$, that we name the $V$-Geometric Control Condition, that holds provided that
\begin{equation}\label{e:vgcc}\tag{$V$-GCC$_T$}
K^{V}_{T,\omega}:=\{\gamma_0\text{ geodesic} \;:\; \phi_s^V(\gamma_0)\cap T^*\omega\neq \emptyset,\;\forall s\in (0,T)\}=\emptyset.
\end{equation}
In other words $\omega$ satisfies \eqref{e:vgcc} provided that, given any geodesic $\gamma_0$ one can find $s\in (0,T)$ such that $\gamma_s\cap \omega\neq \emptyset$.
This new condition is sufficient in order to have observability for the fractional Schrödinger evolution over long times.
\begin{theo}\label{t:lt}
Suppose $(M,g)=(\IS^d,\mathrm{can})$ and that $\omega\subseteq \IS^d$ is open and $\alpha>0$. A constant $C>0$ exists such that the time-frequency observability estimate
\begin{equation}\label{e:lsobs}
\|\Pi_h u^0\|_{L^2(M)}^2\leq \frac{C}{T_h}\int_0^{T_h}\int_\omega | e^{-it((-\Delta)^{\alpha/2}+V)} \Pi_h u^0|^2dx\,dt,
\end{equation}
holds for every $u^0_h\in L^2(M)$ and $h\in (0,h_0]$ in these two cases:
\begin{itemize}
\item $T_h=T/h$ and $\omega$ satisfies $V$-$GCC_{T}$.
\item $hT_h\to \infty$ as $h\to 0 ^+$ and $\omega$ satisfies $V$-$GCC_{T}$ for some $T>0$.
\end{itemize}
\end{theo}
Since solutions issued from an eigenfunction are periodic in time, we obtain the following consequence, which is proved in \cite{MacRiv19} when $\alpha=2$.
\begin{coro}\label{c:eig}
Suppose that $\omega$ satisfies $V$-$GCC_{T}$ for some $T>0$. Then \eqref{e:obse} holds.
\end{coro}
\begin{rem}
If $V$ is odd (meaning $V(x)=-V(-x)$) then $\cI(V)=0$ and $V$-GCC$_{T}$ is equivalent to GCC. However, in \cite{MacRiv19} it is shown that a similar result to Corollary \ref{c:eig} holds under a new geometric condition, in which the Radon transform of the potential is replaced by a different nonlinear transform of $V$, whose expression is a bit more complicated. This term could be constant again, and in general one gets a geometric condition related to the Hamiltonian flow of the first non-vanishing term in a Quantum Birkhoff Normal Form (see \cite{ArM20} for a precise account on the closely related case of the Harmonic Oscillator). Up to our knowledge, it is not known whether or not the vanishing of all those terms implies that $V$ is constant.
This yields the following question: suppose that \eqref{e:obse} holds for every eigenfunction of $-\Delta+V$ if and only if $\omega$ satisfies GCC. Does this imply that $V$ is constant?
\end{rem}
\begin{proof}[Proof of Theorem \ref{t:lt}]
We argue again by contradiction: suppose \eqref{e:lsobs} is not true, this means that there exist a sequence $(h_n)$ that tends to zero and functions $u^0_{h_n}\in L^2(M)$ such that $\|\Pi_{h_n} u^0_{h_n}\|_{L^2(M)}=1$ and
\begin{multline}
\lim_{n\to\infty}\frac{1}{T_{h_n}}\int_0^{T_{h_n}}\|e^{-it((-\Delta)^{\alpha/2}+V)}\Pi_{h_n} u^0_{h_n}\|^2_{L^2(\omega)}dt\\=\lim_{n\to\infty}\int_0^1\|e^{-itT_{h_n}((-\Delta)^{\alpha/2}+V)} \Pi_{h_n} u^0_{h_n}\|^2_{L^2(\omega)}dt=0.
\end{multline}
Consider the semiclassical measure $\mu_t$ of the sequence $(u_{h_n})$ where
\[
u_{h_n}(t,\cdot):=e^{-itT_{h_n}((-\Delta)^{\alpha/2}+V)}\Pi_{h_n} u^0_{h_n}.
\]
A straightforward modification of the proof of Proposition \ref{p:scm} gives that, for almost every $t\in\IR$, the measures $\mu_t$ are invariant by the geodesic flow. Moreover, it is proved in \cite[Proposition 2.2]{MacRiv16} that the following hold when $\alpha=2$:
\begin{itemize}
\item if $T_h=T/h$ then $\mu_t=(\phi^V_{tT})\cI^*(\mu_0)$, where $\mu_0$ stands for the semiclassical measure of $(u_{h_n}(0,\cdot))$;
\item if $hT_h\to\infty$ then $(\phi^V_t)_*\mu_t=\mu_t$ for almost every $t\in\IR$.
\end{itemize}
To see why this holds for every $\alpha>0$, simply recall the proof of Lemma \ref{l:ws}, taking now into account that time derivatives of the Wigner distributions give a new factor $T_{h_n}$:
\[
\frac{d}{dt}\left\la \cI^*(W^{h_n}_{u_{h_n}(t,\cdot)}),a\right\ra=T_{h_n}h_n\left\la W^{h_n}_{u_{h_n}(t,\cdot)},\{V,\cI(a)\}\right\ra+\cO(T_{h_n}h_n^2).
\]
By taking limits, it follows that, for every $t\in\mathbb{R}$,
\[
\left\la \mu_t,\cI(a)\right\ra-\left\la \mu_0,\cI(a)\right\ra=T\int_0^t \left\la \mu_s,\{V,\cI(a)\}\right\ra ds, \quad\text{ when }T_{h_n}h_n=T,
\]
and, for every $\theta\in\cC^\infty_c(\IR)$,
\[
0=\int_\IR\theta(t)\left\la \mu_t,\{V,\cI(a)\}\right\ra dt, \quad\text{ when }T_{h_n}h_n\to\infty.
\]
Since the measure $\mu_t$ is invariant by the geodesic flow,
\[
\left\la \mu_t,a\right\ra=\left\la \mu_t,\cI(a)\right\ra,
\]
and
\[
\left\la \mu_t,\{V,\cI(a)\}\right\ra=\left\la \mu_t,\{\cI(V),\cI(a)\}\right\ra.
\]
From here the claimed invariance and transport properties of $\mu_t$ follow (for further details see the proofs of \cite[Proposition 2.2]{MacRiv16} and Proposition \ref{p:scm} in Section \ref{s:suf}).
Now, since $\mu_t(T^*\omega\setminus\{0\})=0$, $t\in [0,1]$ and $\omega$ satisfies $V$-GCC$_{T}$ it follows that $\mu_t=0$ for almost every $t\in\IR$, which is a contradiction.
\end{proof}
\section{Proof of Theorem \ref{t:minimal}}\label{s:potential}
The Radon transform of the potential $\cI(V)$ is always a zero-homogeneous smooth function on $T^*M\setminus \{0\}$; as such, it can be identified to a unique element in $\cC^\infty(S^*\IS^d)$. In addition, since $\cI(V)$ is invariant by the geodesic flow, it can be identified to a function on $G(\IS^2)$, the space of oriented geodesics. Recall that $G(\IS^d)$ can be constructed as the quotient space of $S^*\IS^d$ in which $(x,\xi)$ and $(x',\xi')$ are equivalent if and only if they belong to the same orbit of the geodesic flow.
The two-dimensional sphere has the nice feature that its space of oriented geodesics $G(\IS^2)$ can be identified to the sphere $\IS^2$ itself and the symplectic form on $T^*\IS^2$ induces a symplectic structure on $G(\IS^2)$ (which must necessarily be a non-zero multiple of the volume form on $\IS^2$). The mapping:
\[
\Phi: G(\IS^2)\To \IS^2 : \gamma \longmapsto x\times \xi,
\]
where $(x,\xi)\in \gamma$ and $\times$ denotes the vector product in $\IR^3$, is well defined and bijective. To see this, note that every geodesic in $\IS^2$ is obtained by intersecting the sphere by the linear plane spanned by $x$ and $\xi$, where $x$ is a point on the geodesic and $\xi$ a unitary cotangent vector to the geodesic at $x$. The two unit normal vectors of this plane are obtained as $x\times \xi$, depending on the choice of orientation of $\xi$. For instance, if $\gamma^\pm$ denotes the geodesic $\{x_3=0\}$ parametrized positively/negatively from the point of view of an observer located at $(0,0,1)$ then:
\[
\Phi(\gamma^\pm)=(0,0,\pm 1).
\]
The set $G_{x_0}$ of all geodesics issued from the same point $x_0\in\IS^2$ is then mapped via $\Phi$ to the geodesic in $\IS^2$ that lies in the plane through the origin that is orthogonal to $x_0$. $G(\IS^2)$ has natural smooth and symplectic structures inherited from $T^*M$, which are preserved by $\Phi$.
With this in mind, the Radon transform, when restricted to functions of $\cC^\infty(S^*\IS^2)$ that only depend on $x$, can be identified to an operator:
\[
\tilde{\cI}\,:\, C^\infty(\IS^2)\To C^\infty(\IS^2),
\]
where $\tilde{\cI}=(\Phi^*)^{-1}\circ \cI$. Then, see \cite{Guillemin1976},
\[
\ker \tilde{\cI} = \cC^\infty_{\rm odd}(\IS^2) :=\{u\in C^\infty(\IS^2)\,:\, u(-x)=-u(x),\;\forall x\in\IS^2\},
\]
whereas
\[
\tilde{\cI}(C^\infty(\IS^2))=\cC^\infty_{\rm even}(\IS^2):=\{u\in C^\infty(\IS^2)\,:\, u(-x)=u(x),\;\forall x\in\IS^2\}.
\]
Therefore,
\begin{equation}\label{e:ibij}
\tilde{\cI}:\cC^\infty_{\rm even}(\IS^2)\longrightarrow \cC^\infty_{\rm even}(\IS^2)\text{ is bijective.}
\end{equation}
Analogously, the Hamiltonian vector field $X_{\cI(V)}$ can be identified to a vector field on $\IS^2$ that is Hamiltonian with respect to the new symplectic form. In particular, its flow $\tilde{\phi}^V_s:=\phi^V_s\circ\Phi^{-1}$ satisfies
\begin{equation}\label{e:level}
\tilde{\phi}^V_s:\IS^2\To\IS^2,\quad \text{ and }\quad \tilde{\cI}(V)\circ\tilde{\phi}^V_s=\tilde{\cI}(V),\quad \forall s\in \IR.
\end{equation}
The strongest obstruction to $V$-GCC$_T$ comes from the fact that $\cI(V)$ always has critical points:
\[
\cC(V)=\{\gamma \in G(\IS^2)\,:\, d\cI(V)_\gamma=0\}=\Phi^{-1}(\{p\in\IS^2\,:\, d\tilde{\cI}(V)_p=0\})\neq\emptyset.
\]
If $\gamma_0\in \cC(V)$ then $\phi^V_s(\gamma_0)=\gamma_0$ for every $s\in\IR$. Therefore,
if $\omega$ satisfies $V$-GCC$_T$ then it must necessarily intersect the projection of $\gamma_0$ onto $\IS^2$.
Let us now define the class of potentials $\cT$. Let:
\[
Q_{(a,b,c)}(x)=ax^2_1+bx_2^2+cx_3^2,\quad x\in\IR^3.
\]
Then $Q_{(a,b,c)}|_{\IS^2}\in \cC^\infty_{\mathrm{even}}(\IS^2;\IR)$ and we define, using \eqref{e:ibij},
\[
\cT:=\tilde{\cI}^{-1}(\{Q_{(a,b,c)}|_{\IS^2}\,:\, 0<a<b<c\}).
\]
For any $V\in\cT$, the function $\tilde{\cI}(V)=Q_{(a,b,c)}|_{\IS^2}\in \cC^\infty_{\mathrm{even}}(\IS^2;\IR)$ has exactly six (non-degenerate) critical points:
\[
\{c_1^\pm:=(\pm 1,0,0),\;c_2^\pm:=(0,\pm 1, 0),\; c_3^\pm:=(0,0,\pm 1)\},
\]
and
\[
\min_{\IS^2}\tilde{\cI}(V)=a,\quad \max _{\IS^2}\tilde{\cI}(V)=c.
\]
The orbits of $\tilde{\phi}^V_t$ are contained in the connected components of the level sets $\tilde{\cI}(V)^{-1}(E)$, $E\in [a,c]$, by \eqref{e:level}. The set of orbits is invariant by the symmetries $x_i\mapsto -x_i$, for $i=1,2,3$; the precise description of the orbits is:
\begin{itemize}
\item The equilibrium points $c_1^\pm$ when $E=a$.
\item Two closed orbits around $c_1^\pm$ when $E\in (a,b)$.
\item The equilibrium points $c_2^\pm$ and four orbits connecting $c_2^+$ to $c_2^-$, when $E=b$.
\item Two closed orbits around $c_3^\pm$ when $E\in (b,c)$.
\item The equilibrium points $c_3^\pm$ when $E=c$.
\end{itemize}
The six oriented geodesics mapped by $\Phi$ to the critical points $c_i^\pm$ of $\tilde{\cI}(V)$ are $\gamma_1^\pm,\gamma_2^\pm,\gamma_3^\pm$, which correspond to the non-oriented geodesics:
\[
\gamma_i=\{x_i=0\},\quad i=1,2,3.
\]
Suppose that $\omega\subseteq\IS^2$ is an open set that contains $p=(0,0,1)\in\gamma_1^+\cap\gamma_2^+$ and $q=(0,1,0)\in\gamma_1^+\cap\gamma_3^+$. This means that the set of all geodesics that intersect $\omega$ contains $G_p\cup G_q$ (recall that this denotes the union of all geodesics issued from $p,q$). Now, $\Phi$ maps:
\[
\Phi(G_p)=\IS^2\cap \{x_3=0\}\quad\text{ and }\quad \Phi(G_q)=\IS^2\cap \{x_2=0\}.
\]
The set $\Phi(G_p)$ is parameterized by $(\cos t,\sin t,0)$, and $\Phi(G_q)$ by $(\cos t,0,\sin t)$; evaluating along $Q_{(a,b,c)}$ shows that
\[
(\Phi(G_p)\cup \Phi(G_q))\cap Q_{(a,b,c)}^{-1}(E)\neq \emptyset, \quad \forall E\in [a,c],
\]
and, moreover,
\[
\quad c_2^\pm\in \Phi(G_p),\quad \Phi(G_q)\cap (Q_{(a,b,c)}^{-1}(b)\setminus\{c_2^\pm\})\neq \emptyset.
\]
Since the set of orbits of $\tilde{\phi}^V_t$ and both sets $\Phi(G_p),\Phi(G_q)$ are invariant by the symmetries $x_i\mapsto -x_i$, for $i=1,2,3$, we conclude that $\Phi(G_p)\cup \Phi(G_q)$ has non-empty intersection with all the orbits of $\tilde{\phi}^V_t$. Therefore, $V$-GCC$_T$ is satisfied for some $T>0$ and the result follows from Corollary \ref{c:eig}.
Let us mention that six is the least number of critical points an even Morse function on $\IS^2$ may have. This is due to the fact that any such function induces a Morse function on the projective plane $\IP$. Since the Euler characteristic of $\IP$ is equal to one, the Poincaré-Hopf theorem implies:
\[
1=\chi(\IP)=\sum_{j=0}^2 (-1)^j\#\{\gamma\in\cC(V)\,:\,\gamma\text{ has index }j\}.
\]
There are at least one critical point of index zero and one of index two, therefore one must have also at least one saddle point. The number of critical points of $\cI(V)$ when viewed as a function of $\IP$ must be at least three, hence the claim.
| ca76a6c57d4762f40d17c150ef82b416ad8c5d5a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Abstract}
The Generalised Differential Image Motion Monitor (GDIMM) was proposed a few years ago as a new generation instrument for turbulence monitoring. It measures integrated parameters of the optical turbulence, i.e the seeing, isoplanatic angle, scintillation index, coherence time and wavefront coherence outer scale. GDIMM is based on a fully automatic small telescope (28cm diameter), equipped with a 3-holes mask at its entrance pupil. The instrument is installed at the Calern observatory (France) and performs continuous night-time monitoring of turbulence parameters. In this communication we present long-term and seasonnal statistics obtained at Calern, and combine GDIMM data to provide quantities such as the equivalent turbulence altitude and the effective wind speed.
\section{INTRODUCTION}
\label{par:intro}
The Generalized Differential Image Monitor (GDIMM) is an instrument designed to monitor integrated parameters of the atmospheric turbulence above astronomical observatories. It provides the seeing $\epsilon_0$, the isoplanatic angle $\theta_0$, the scintillation index $s_0$, the coherence time $\tau_0$ and the spatial coherence outer scale ${\cal L}_0$. The 3 parameters $\epsilon_0$, $\theta_0$ and $\tau_0$ are of fundamental importance for adative optics (AO) correction. The scintillation has a strong impact on photometric signals from astronomical sources as well as on optical telecommunications with satellites. The outer scale ${\cal L}_0$ has a significant effect for large diameter telescopes (8m and above) and impacts low Zernike mode such as tip-tilt\cite{Winker91}. The GDIMM is at the moment the only monitor to provide simultaneously all these parameters.
GDIMM was proposed in 2014\cite{Aristidi14} to replace the old-generation turbulence monitor GSM\cite{Ziad00}. It is a compact instrument very similar to a DIMM\cite{Sarazinroddier90}, with 3 sub-apertures of different diameters. GDIMM observes bright single stars up to magnitude $V\sim 2$, at zenith distances up to 30$^\circ$. After a period of developpement and tests in 2013--2015, the GDIMM is operational since the end of 2015, as a part of the Calern atmospheric Turbulence Station (C\^ote d'Azur Observatory -- Calern site, France, UAI code: 010, Latitude=$43^\circ 45' 13''$~N, Longitude=$06^\circ 55' 22''$~E). GDIMM provides continuous monitoring of 5 turbulence parameters ($\epsilon_0$, $s_0$, $\theta_0$, $\tau_0$ and ${\cal L}_0$) above the Calern Observatory. Data are displayed in real time through a website ({\tt cats.oca.eu}), as a service available to all observers at Calern. The other objective is that Calern becomes an operational on-sky test platform for the validation of new concepts and components in order to overcome current limitations of high angular resolution existing systems. Several activities regarding adaptive optics are operated at the M\'eO\cite{Samain08} and C2PU\cite{Bendjoya12} telescopes and they benefit of the data given by the CATS station.
The present communication is an update of previous papers\cite{Aristidi14, Aristidi18, Aristidi19}. We describe last
improvements and the present status of the instrument in Sect.~\ref{par:obs}. Long-term statistics (5 years) of turbulence above the Plateau de Calern are presented in Sect.~\ref{par:stats}.
\section{OBSERVATIONS AND DATA PROCESSING}
\label{par:obs}
GDIMM is based on a small commercial telescope (diameter 28cm) equipped with a 3 apertures pupil mask. 2 sub-pupils have a diameter of 6cm and are equipped with glass prisms oriented to give opposite tilts to the incident light. The 3rd sub-aperture is circular, its diameter is 10cm and it has a central obstruction of 4cm: this particular geometry is required to estimate the isoplanatic angle from the scintillation index\cite{Looshogge79}. At the focus of the telescope images are recorded by a fast camera allowing a framerate of 100 frames per second. Turbulence parameters are estimated on data cubes containing 2048 images. The first half of these cubes are taken with an exposure time of 5ms, the second half with 10ms, to allow compensation of the bias due to the exposure time. A set of parameter is calculated every 2mn and sent to a database accessible worldwide via a web interface ({\tt cats.oca.eu}). More details are given in previous papers\cite{Aristidi19}.
The dataset analyzed here covers the period March 2015 to September 2020. The number of data points collected each month of the period is displayed in Fig.~\ref{fig:nbdata} (left). As the robotization of the instrument progressed during the years 2015--2016, the number of data increased. At the beginning of 2019, for a few months, we reduced the delay between successive acquisitions from 2 mn to 1mn, this resulted in a larger number of data (up to $\sim 10^4$ in Februrary 2019). Fig.~\ref{fig:nbdata} (right) shows the total number of measurements per month. Two periods in the year present a smaller number of data: April-May and November. They correspond to technical failures or adverse weather (in particular the spring 2018 was very bad and rainy, and the gap in April 2020 was due to the covid-19 lockdown).
\begin{figure*}
\begin{center}
\includegraphics[width=70mm]{nb_data_hist-eps-converted-to.pdf}
\includegraphics[width=70mm]{nb_data_mois-eps-converted-to.pdf}
\end{center}
\caption{Left: number of data points collected each month since 2015. Right: monthly number of data.}
\label{fig:nbdata}
\end{figure*}
\section{STATISTICS AT THE CALERN OBSERVATORY}
\label{par:stats}
\begin{table}
\begin{center}
\begin{tabular}{l|c|c|c|c|c|c|c}\hline
& $\epsilon_0$ & $s_0$ & $\theta_0$ & $\tau_0$ & ${\cal L}_0$ & $\bar h$ & $\bar v$ \\
& [$''$] & [\%] & [$''$] & [ms] & [m] & [m] & [m/s] \\ \hline
Median & 1.19 & 2.99 & 1.61 & 2.02 & 24.50 & 3394 & 13.31\\
Mean & 1.33 & 3.63 & 1.73 & 2.78 & 35.67 & 3645 & 13.84 \\
Std. dev. & 0.58 & 1.93 & 0.59 & 1.66 & 28.50 & 1473 & 5.37\\
$1^{st}$ quartile & 0.86 & 1.99 & 1.26 & 1.25 & 12.40 & 2520 & 9.71 \\
$3^{rd}$ quartile & 1.64 & 4.52 & 2.06 & 3.39 & 48.80 & 4511 & 17.08\\
$1^{st}$ centile & 0.46 & 0.60 & 0.66 & 0.44 & 2.70 & 1176 & 3.21\\
Last centile & 3.55 & 11.78 & 3.94 & 13.13 & 141.50 & 8839 & 27.80\\ \hline
Paranal & 0.81 & 1.51 & 2.45 & 2.24 & 22 & 3256 & 17.3 \\
La Silla & 1.64 & 4.63 & 1.25 & 1.46 & 25.5 & 3152 & 13.1 \\
Mauna Kea & 0.75 & 1.11 & 2.94 & 2.43 & 24 & 2931 & 17.2 \\ \hline
\end{tabular}
\end{center}
\caption{Statistics of turbulence parameters measured at Calern (at the wavelength $\lambda=0.5\mu$m) during the period March 2015--Sept 2020. Paranal, La Silla and Mauna Kea values are from the GSM database.}
\label{table:paramstat}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=17cm]{params_hist-eps-converted-to.pdf}
\caption{Histograms of the seeing, isoplanatic angle, coherence time and outer scale at Calern, calculated at the wavelength $\lambda=0.5\mu$m for the period June 2015--September 2020.}
\label{fig:paramshisto}
\end{center}
\end{figure}
We collected about 148\,000 turbulence parameter sets during the 5~year period. The number for the outer scale ${\cal L}_0$ is lower (54\,000): this parameter is sensitive to telescope vibrations and a reliable value is not always available\cite{Aristidi19}. Statistics are presented in Table~\ref{table:paramstat} for the 5 turbulence parameters ($\epsilon_0$, $s_0$, $\theta_0$ $\tau_0$, ${\cal L}_0$). Note that the scintillation index $s_0$ is measured through a 10cm diameter sub-pupil: this causes a low-pass spatial filtering and values are lower than the actual scintillation defined for a zero-diameter pupil\cite{Roddier81}.
We also provide values for the equivalent turbulence altitude $\bar h$ and the effective wind speed $\bar v$ (weighted average of the wind speed over the whole atmosphere). A definition of these quantities can be found in papers by Roddier\cite{Roddier81, Roddier82}, and we estimated them by the following equations:
\begin{equation}
\bar{v}=0.31 \frac{r_0}{\tau_0}
\end{equation}
and
\begin{equation}
\bar{h}=0.31 \frac{r_0}{\theta_0}
\end{equation}
Histograms are displayed in Fig.~\ref{fig:paramshisto} and show a classical log-normal shape for all parameters. We did not display the scintillation histogram since is is directly deduced from $\theta_0$ by an exact analytic relation. A comparison with other astronomical sites in the world (examples for Paranal, La Silla and Mauna Kea are given in Table~\ref{table:paramstat}) show that the Calern plateau is an average site.
Fig.~\ref{fig:params_vs_month} displays the monthly evolution of parameters. The seeing is slightly lower in summer, we measured a median value of $1.06''$ in July and August (the median winter seeing during the period November--March is 1.34$''$). As a consequence, the median coherence time is higher in summer (2.24ms in July--August, 1.78ms in November--March). Also, Fig.~\ref{fig:hourlyseeingtau0} shows that, in Summer, there is a dependence of $\epsilon_0$ and $\tau_0$ with time. The median seeing decreases below 0.8$''$ at the end of the night (while the coherence time increases). Nothing similar was observed during the winter. The isoplanatic angle and the outer scale didn't show any noticeable time dependence, both in summer and winter. The outer scale ${\cal L}_0$ has values similar to other sites such as Mauna Kea or La Silla.
Sequences of several hours of good seeing were sometimes observed, which is a good point for this site (and already known by ``old'' observers on interferometers during the 80's and 90's). An exemple is shown on Fig.~\ref{fig:seeingtsgood}. It is a time series recorded on Jan. 1st, 2020, showing a seeing below 1$''$ during 12~hours, the median seeing for that night was 0.69$''$.
\begin{figure}
\begin{center}
\includegraphics[width=17cm]{params_vs_month-eps-converted-to.pdf}
\caption{Monthly median values of the four parameters ($\epsilon_0$, $\theta_0$, $\tau0$, ${\cal L}_0$), calculated at the wavelength $\lambda=0.5\mu$m for the period June 2015--September 2020. On each graph, the red curve is the median value, green curves are the first and third quartiles.}
\label{fig:params_vs_month}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{summer_hourly_seeingtau0-eps-converted-to.pdf}
\caption{Hourly variations of the seeing $\epsilon_0$ and coherence time $\tau_0$ during the summer (July-August). Symbols (triangle/circle) are the median values. Error rectangles lower and upper edges are the 1st and 3rd quartiles. The data set represents a total of 27300 measurements for 5 summers (2015 to 2019).}
\label{fig:hourlyseeingtau0}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=15cm]{seeing_tau0_2020-01-01-eps-converted-to.pdf}
\caption{Exceptional conditions observed on Jan. 1st, 2020. Left: seeing time series (the median value for this night was 0.70$''$), right: coherence time (median value 4.3ms).}
\label{fig:seeingtsgood}
\end{center}
\end{figure}
Fig.~\ref{fig:histseeingfit} displays seasonal seeing histograms, calculated for the summer (July and August) and the winter (November--March). They appear to be well modelled by a sum of two log-normal functions (they appear as dashed curves on the plots, their sum is the solid line). This is an evidence of the existence of two regimes: a ``good seeing'' distribution $\phi_1$ with a median value $\epsilon_1$ and a ``medium seeing'' distribution $\phi_2$ with a median value $\epsilon_2$. In summer, we have $\epsilon_1=0.68''$ (the good seeing distribution contains 20\% of the data) and $\epsilon_2=1.04''$ (80\% of the data). In winter we have $\epsilon_1=0.76''$ (21\% of the data) and $\epsilon_2=1.25''$ (79\% of the data). The time series displayed in Fig.~\ref{fig:seeingtsgood} corresponds to a realisation of the good seeing distribution, during 12~hours.
\begin{figure*}
\begin{center}
\includegraphics[width=8cm]{histoseeing_fit_summer-eps-converted-to.pdf}
\includegraphics[width=8cm]{histoseeing_fit_winter-eps-converted-to.pdf}
\caption{Left: seeing histogram for the summer (July--August). Right: seeing histogram for the winter (Nov--March). Superimposed curves $\phi_1$ and $\phi_2$ are a least-square fit by a sum of two log-normal distributions (individual log-normal curves are dashed lines).}
\label{fig:histseeingfit}
\end{center}
\end{figure*}
\subsection*{Temporal stability of parameters}
The stability of turbulence parameters is an important concern. Temporal fluctuations were studied by different authors\cite{Racine96, Ziad99}. The characteristic time of stability is generally of a few minutes for temperate sites (e.g. Mauna Kea and La Silla).
We define here an interval of stability as a continuous period of time $t_s$ in which a parameter is {\em better} than a given threshold $x_0$. The term {\em better} means {\em lower} for the seeing, as low values of the seeing correspond to high Strehl ratio. It means {\em higher} for the coherence time and isoplanatic angle. In this interval, we allow the parameter to be better than $x_0$ during a few minutes (we took 4 mn, i.e. two sampling intervals of parameter time series). For a given value of $x_0$, we calculate the distribution of $t_s$ on the whole dataset. Its median value gives the characteristic time of stability. This approach was the one used for the site characterization of Dome C by means of DIMMs\cite{Aristidi09, Fossat10} and Single Star Scidar\cite{Giordano12}.
We performed this analysis for the 3 parameters $\epsilon_0$, $\tau_0$ and $\theta_0$, for the summer and the winter periods. We could not obtain results for the outer scale, since it is not always measurable because of telescope vibrations (data sequences show large gaps that hamper the calculation of $t_s$). Results are displayed in Fig~\ref{fig:stability}. For the seeing, the characteristic time increases with the seeing value and saturates. It is of the order of 20mn for seeing values around $1''$, and is longer during the summer. For $\tau_0$ and $\theta_0$ we observe that $t_s$ decrease with the threshold $x_0$, down to a value of $\sim$10mn. Here again, summer conditions appear to be slightly more stable.
Note that the curves are likely to be biased by interruptions of the observations (change of target star, clouds, day-night cycle). But the mean value of uninterrupted sequences duration is $\sim 150$mn, much larger than the characteristic stability times.
\begin{figure*}
\begin{center}
\includegraphics[width=60mm]{stabtemp_seeing-eps-converted-to.pdf}\hskip -5mm
\includegraphics[width=60mm]{stabtemp_tau0-eps-converted-to.pdf}\hskip -5mm
\includegraphics[width=60mm]{stabtemp_isop-eps-converted-to.pdf}
\caption{Temporal stability of parameters, given by the continuous time a parameter is better than a given value (``better'' means lower for the seeing, and higher for the coherence time and the isoplanatic angle). Curves are plotted for the summer (July--August) and winter (November to Mars) datasets.}
\label{fig:stability}
\end{center}
\end{figure*}
\section{CONCLUSION}
We have presented statistics of optical turbulence parameters above the plateau de Calern, measured for five years with the GDIMM monitor. GDIMM is a part of the CATS station which aims at monitoring atmospheric turbulence parameters and vertical profiles on the site of Calern. The station is fully automatic, using informations from a meteo station and an All-Sky camera to replace human interventions.
Seeing and isoplanatic measurements are given by differential motions and scintillation, according to well known techniques. The coherence time is derived from AA structure functions. The method is simple and has proven to give satisfactory results\cite{Ziad12}. It may be used as well with a classical DIMM providing that the acquisition camera is fast enough (typically 100 frames/sec) to properly sample the AA correlation times. The outer scale is derived from ratios of absolute to relative AA motions\cite{Aristidi19}. This parameter is the most sensitive to telescope vibrations, and requires a good stability of the telescope mount and pillar.
Simultaneous observations of GDIMM and another instrument, the Profiler of Moon Limb, which estimates the vertical profile of turbulence, gave concordant results\cite{Ziad19, Chabe20}. A portable version of the GDIMM has been developped
in parallel to the Calern one, to perform turbulence measurements at any site on the world. Discussions with the ESO
(European Southern Observatory) are currently in progress to make observations at Paranal and Armazones.
\section*{Acknowledgments}
This CATS project has been done under thefinancial support of CNES, Observatoire de la C\^ote d'Azur,
Labex First TF, AS-GRAM, Federation Doeblin, Universit\'e de Nice-Sophia Antipolis and Re\'gion Provence Alpes
C\^ote d'Azur. We would like thanks all the technical and administrative staff of the Observatoire de la C\^ote
d'Azur and our colleagues from the Astrog\'eo team from G\'eoAzur laboratory for their help and support all along
the project namely Pierre Exertier, Etienne Samain, Dominique Albanese, Mourad Aimar, Jean-Marie Torre,
Emmanuel Tric, Thomas Lebourg and Sandrine Bertetic.
| 5e24d0609b050e50da1bed3ba2a13718f06ff674 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Many generations of astronomers
have been seraching for
the possible existence of
exoplanets before the end of twentieth century
(Briot \& Schneider 2018)
with the discovery of 51 Peg b (Mayor \& Queloz 1995).
Since then, exoplanetary science has witnessed
a boom that has made it
one of the most studied branches of astronomy.
Although the initial success in discovering exoplanets came from
the results of the Doppler method, the transit method has
played the most dominating role in discovering new exoplanets.
This is due to space-based surveys like
Kepler (Borucki et al 2010), K2 (Howell et al. 2014),
Convection, Rotation and planetary Transits
(CoRoT, Baglin et al. 2006) and also, the recently launched
Transiting Exoplanet Survey Satellite
(TESS, Ricker et al. 2015).
However, transit surveys from ground-based observing facilities
have also contributed significantly in discovering transiting exoplanets
owing to the surveys such as
the Hungarian-made Automated Telescope Network (HATNet, Bakos et al. 2004),
The Hungarian-made Automated Telescope Network-South (HATSouth, Bakos et al. 2013),
Trans-Atlantic Exoplanet Survey (TrES, Alonso et al. 2004),
Super Wide Angle Search for Planets (SuperWASP, Pollacco et al. 2006),
Kilodegree Extremely Little Telescope (KELT, Pepper et al. 2007),
Qatar Exoplanet Survey (QES, Alsubai et al. 2013),
Multi-site All-Sky CAmeRA (MASCARA, Talens et al. 2017) survey
to name a few.
These surveys preferably detect short period, close-in planets.
Ground-based telescopes also
provide follow-up observations to confirm the transiting nature
of exoplanets discovered from space surveys.
The additional contributions from ground-based telescopes
are to cover a large field-of-view and to carry out
extensive follow-up observation programs which improve the
orbital parameters of a planetary system.
These observations can also be used
for the analysis of
the transit timing variations (TTVs)
over a longer time baseline.
In essence, a TTV is the
transit time deviation from a linear ephemeris
which provides clues about
the existence of another planet in the system
(Agol et al. 2005, Agol \& Fabrycky 2017,
Sun et al. 2017, Linial et al. 2018,
Ba\c{s}t\"urk et al. 2019).
In fact, TTVs have not only led to the discoveries of new exoplanets
(Nesvorn\'y et al. 2012, Ioannidis et al. 2014,
Fox \& Wiegert 2019, Sun et al. 2019),
but also become a tool to characterize the bulk composition of
exoplanets (Jontof-Hutter et al. 2015, Kipping et al. 2019).
Motivated by these important results,
Holczer et al. (2016) constructed a transit timing catalog of 2599
Kepler Objects of Interest, which will be very useful for further TTV studies.
As discussed by Ba\c{s}t\"urk et al. (2019), Saturn mass planets
are interesting for their densities and orbital properties.
HAT-P-12b is a low-density, moderately irradiated,
sub-Saturn mass ($m_p= 0.211\pm0.012 M_J$)
transiting exoplanet whose discovery was reported by
Hartman et al. (2009) using the HAT-5 telescope (located in Arizona)
of HATNet (Bakos et al. 2004).
HAT-P-12b orbits a moderately bright ($V\sim12.8$),
metal-poor K4 dwarf within a period of $\sim$3.21 days
(Hartman et al. 2009). By the time of its discovery, it was
the least massive H/He-dominated gas giant planet.
Because the above characteristics are
very different from Jupiter-mass close-in exoplanets,
HAT-P-12b has been studied through the methods of
photometric transit observations,
radial velocity measurements, and transmission spectroscopy
by many groups.
Lee et al. (2012) used follow-up observations to
improve the ephemeris of the system.
Likewise, Sada et al. (2012) published $J$-band transit light curves
for HAT-P-12b.
Todorov et al. (2013) observed the secondary eclipses using the
IRAC instrument
on the {\it Spitzer Space Telescope}. They did not detect eclipses
at either 4.5 $\mu$m or 3.6 $\mu$m wavelengths.
The radial velocity measurements of the planet were
produced by Kunston et al. (2014) and Ment et al. (2018).
Mancini et al. (2018) used HARPS-N high-precision radial velocity measurements
to analyze the Rossiter-McLaughlin effect. They determined the
sky-projected obliquity ($\lambda=-54^{\circ}\,^{+41^{\circ}}_{-13^{\circ}}$)
for HAT-P-12b.
Sada \& Ramon-Fox (2016) combined publicly available light curves with radial
velocity measurements and determine physical and orbital parameters
for HAT-P-12b.
Spectroscopically,
Line et al. (2013) presented an NIR transmission spectrum for the system
using HST WFC-3.
They found a lack of water absorption feature for
a hydrogen-dominated atmosphere.
Alexoudi et al. (2018) performed a homogeneous analysis
which included published data from Sing et al. (2016) and their own data
and obtained a transmission spectrum with a low-amplitude spectral slope.
The above discussion shows that
the determination of updated orbital parameters is very important
as small deviations in these values could lead to different
physical parameters and structures of exoplanets.
Motivated by this, through a homogeneous
long baseline TTV analysis, here
we present a comprehensive study of HAT-P-12b
with observations combining our new
observations with the publicly available published light curves.
We include the light curves from the discovery paper
of HAT-P-12b (Hartman et al. 2009), up to very recent
observations, in order to cover a large range of 1160 epochs, where the
entire data has a time baseline of $\sim10.2$ years.
The data used in this work and its basic reduction procedure
are given in Section~\ref{OBS}.
The analysis of the light curves using
Markov-Chain Monte-Carlo (MCMC) techniques is
described in Section~\ref{TAP}.
Section~\ref{TTV} describes a new ephemeris using linear fitting
as well as a frequency analysis and corresponding $O-C$ diagrams.
Section~\ref{dynamics} presents the dynamical two-planet model.
Finally, the conclusions of this study are presented in Section~\ref{CONC}.
\section{The Data and Reduction Procedures}
\label{OBS}
\begin{table*}
\caption{The observational log of the new data used in this work.}
\small
\centering
\begin{tabular}{ccccccc}
\hline\hline
Run & UT Date & Instrument & Filter & Interval & Exposure & Number of \\
& & & & ($BJD_{TDB}-2450000$) & (s) & images \\
\hline
1 & 2011 March 29 & Tenagra & $R$ & 5649.712644-5649.930803 & 75 & 192 \\
2 & 2011 April 17 & PM0 & Sloan $r$ & 5668.988182-5669.165950 & 50 & 159 \\
3 & 2011 April 27 & Tenagra & $R$ & 5678.633962-5678.848666 & 75 & 144 \\
4 & 2014 March 19 & P60 & $R$ & 6735.744668-6735.888703 & 15 & 276 \\
5 & 2014 April 17 & P60 & $R$ & 6764.676953-6764.805016 & 20 & 216 \\
6 & 2015 July 14 & P60 & $R$ & 7217.695978-7217.850202 & 24 & 186 \\
\hline
\label{log}
\end{tabular}
\end{table*}
Among our data, we used three transit observations
from the 60 inch telescope (P60) installed at the
Palomar Observatory in California, USA.
Two light curves were observed with 32 inch
telescope at Tenagra Observatory in Arizona, USA.
The Purple Mountain Observatory's
40 inch Near-Earth Object Survey Telescope at the Xuyi Station
provided another light curve's data used in this study.
The log of the observations is listed in Table~\ref{log}.
The `run' in the table is in accordance to the date of observation.
The observed CCD images first went through some
standard procedures such as bias subtraction, flat-fielding,
dark frames (when needed) and cosmic
rays removal with Image Reduction and Analysis Facility
(IRAF\footnote{IRAF is distributed by the National Optical
Astronomical Observatory which is operated by the Association
of Universities for Research in Astronomy, under contact
with the National Science Foundation.}).
Before conducting photometry of the images, the images were first aligned
using the `xregister' task of IRAF.
The photometry of the `cleaned' images is conducted using
the `apphot' task in the `digiphot' routine.
The initial step in aperture photometry is to find/detect the stars in the image.
IRAF task `daofind' finds stars in the image and lists them in a file.
The next step in aperture photometry gives the flux value of the stars.
IRAF task `phot' serves this purpose.
Once we have the fluxes of stars, we conduct differential photometry.
In differential photometry,
the target star's flux (or magnitude) is presented
with respect to one or multiple comparison stars
(e.g. Sariya et al. 2013, Jiang et al. 2013, 2016).
The selected comparison stars should not be of a variable nature.
Differential photometry cancels out the corrections required for
the airmass and exposure time.
It is also useful when the observing conditions are not the best.
For the HAT-P-12b data, we selected the comparison stars having the
same instrumental magnitude and neighboring position
to the target star (HAT-P-12) in the CCD frames.
For the TTV analysis, it is always best to include the
published light curves with the new observations
as a longer time baseline assures a better ephemeris.
We have, therefore, used three light curves from
Hartman et al. (2009), three light curves from Lee et al. (2012),
ten light curves from Mancini et al. (2018)
and nine light curves from Alexoudi et al. (2018).
The total time duration covered by the data thus becomes
slightly more than a decade.
We did not simply use the mid-transit times
for the published light curves given in the respective papers.
Instead, we applied the same procedure on those light curves
that we applied to our data. This approach removes any systematics
while performing parameter fitting and provides more consistent
inputs for the TTV analysis.
The light curves were then processed through a normalization routine
to get rid of the effects caused by the airmass.
For this purpose, we adopted the procedure described by
Murgas et al. (2014) wherein a third degree polynomial is used
to model the airmass.
The observed flux of a light curve $F_0(t)$ can be represented as:
\begin{equation}
F_0(t)=F(t)\mathcal{P}(t),
\end{equation}
where $F(t)$ is the normalized flux of the light curve
which will be used in the further analysis
and $\mathcal{P}(t) = a_0 + a_1 t + a_2 t^2+ a_3 t^3$
is a third degree polynomial.
A python code is used to numerically calculate the best values of
the parameters $a_0$, $a_1$, $a_2$, and $a_3$ so that the
out-of-transit part of $F(t)$ is close to unity.
As for the timing scheme for the light curves,
we took the time from the headers of the individual images.
To make sure that the criterion used for the time is uniform,
first we calculated observation time for the mid-exposure for every image.
Further, it is essential to bring all the mid-exposure times to a
common time stamp for a consistent fitting.
Hence, all the individual times of observations were converted to
the Barycentric Julian Date in
Barycentric Dynamical Time (BJD$_{TDB}$)
following Eastman et al. (2010).
\section{The Light-Curve Analysis}
\label{TAP}
\begin{table*}
\caption{The settings of initial values and conditions for running the TAP.
The values of $P$, $i$,
$a/{R_*}$ and $R_{p}/R_{*}$
are taken from Hartman et al. (2009).
The value of eccentricity ($e$) is taken from Knutson et al. (2014).
}
\centering
\begin{tabular}{ccc}
\hline\hline
Parameter & Initial Value & Condition during MCMC Chains \\
\hline
period ($P$,day) & 3.2130598 & Gaussian penalty with $\sigma = 0.0000021$\\
orbital inclination ($i$, $^{\circ}$) & 89.0 & Gaussian penalty with $\sigma = 0.4$ \\
scaled semi-major axis ($a/R_{*}$) & 11.77 & free \\
planet to star radius ratio ($R_{p}/R_{*}$) & 0.1406 & free \\
mid-transit time ($T_m$) & TAP calculations & free, linked only for the same transit event\\
linear limb darkening ($u_1$) & Table~\ref{LD} & Gaussian penalty with $\sigma = 0.05$ \\
quadratic limb darkening ($u_2$) & Table~\ref{LD} & Gaussian penalty with $\sigma = 0.05$ \\
orbital eccentricity ($e$) & 0.026 & Gaussian penalty with $\sigma = 0.022$ \\
longitude of periastron ($\varpi,^{\circ}$)& 0.0 & locked\\
\hline
\label{input}
\end{tabular}
\end{table*}
The transit light curves (6 new+25 published) were analyzed using
the Transit Analysis Package (TAP, Gazak et al. 2012).
TAP has previously been used by our group for
TrES-3b (Jiang et al. 2013, Mannaday et al. (2020),
WASP-43b (Jiang et al. 2016)
and Qatar-1b (Su et al., {\it submitted}).
TAP is an IDL based graphical user-interface driven software package
which employs the MCMC approach to
fit the light curves using the analytic model given by Mandel \& Agol (2002)
and wavelet-based likelihood function by Carter \& Winn (2009).
TAP involves a set of nine parameters that the user has to input.
These parameters are:
orbital period of the planet ($P$),
orbital inclination
on the sky plane ($i$),
scaled semi-major axis ($a/R_{\ast}$),
the planet-to-star radius ratio ($R_{\rm p}/R_{\ast}$),
the mid-transit time ($T_{\rm m}$),
the linear limb darkening coefficient ($u_1$),
the quadratic limb darkening coefficient ($u_2$),
orbital eccentricity ($e$)
and
the longitude of periastron ($\varpi$).
For the input parameters mentioned above, one has to
define one of the three conditions while running the MCMC chain of TAP.
According to the conditions, a parameter can be one of the following:
(1) completely free
(2) completely locked, or
(3) varying according to a Gaussian function.
As discussed in Section~\ref{OBS}, we dropped
the first publicly available light curve from
Hartman et al. (2009). So, the epoch zero in this study was defined by the
second of the four publicly available light curves
from Hartman et al. (2009).
In order to define the initial input values,
we considered most of the values mentioned in Hartman et al. (2009),
as their paper presents the maximum number of required input parameters
and it is better for the consistency to use
input parameters from the same source.
For the eccentricity, we considered the initial input value from
Knutson et al. (2014).
The orbital period ($P$) was defined as 3.2130598
with a Gaussian penalty of 0.0000021.
The scaled semi-major axis ($a/R_{\ast}$)
and the planet-to-star radius ratio ($R_{\rm p}/R_{\ast}$)
were chosen to be completely free and their input values were
11.77 and 0.1406, respectively.
We also allowed the mid-transit time ($T_{\rm m}$) to be completely free
and did not input any value for it.
The longitude of periastron ($\varpi$)
was set to 0$^\circ$ and was completely locked.
Also set with a Gaussian penalty,
the orbital inclination on the sky plane was
set as 89$^\circ$ with a sigma of 0.4$^\circ$.
The value of eccentricity is listed as 0.026$^{+0.026}_{-0.018}$
by Knutson et al. (2014), where we
input the value ($e$=0.026) as a Gaussian with a sigma of 0.022,
where the sigma
was calculated by taking the mean of errors in positive and negative directions.
The values of limb darkening coefficients were chosen
to be Gaussian with a sigma ($\sigma$) value of 0.05.
Table~\ref{input} contains the information about the input parameters
and the condition chosen for them while running the MCMC chains.
\begin{table*}
\caption{The values of quadratic limb darkening coefficients}
\centering
\begin{tabular}{ccc}
\hline\hline
Filter & $u_1$ & $u_2$ \\
\hline
$B$ & 0.93774724 & -0.083432883 \\
$R$ & 0.57122572 & 0.14770584 \\
$I$ & 0.44099208 & 0.18460748 \\
sloan $g$ & 0.86437516 & -0.029412285 \\
sloan $r$ & 0.60995392 & 0.13478272 \\
sloan $i$ & 0.47080896 & 0.17786368 \\
sloan $z$ & 0.38368084 & 0.19694884 \\
Str\"omgren $u$ & 1.2500255 & -0.37429263 \\
\hline
\label{LD}
\end{tabular}
\end{table*}
The limb darkening is a filter dependent quantity.
All of our new light curves are in the
Cousin $R$ band, except one light curve being in the
Sloan $r$ band. However, the light curves we use from
the published literature come from various filters.
These filters include Johnson $B$, Cousin $RI$,
Sloan $griz$, Gunn $gr$ and Str\"omgren $u$ band.
But the issue is that the published papers
do not always provide the
numerical values of limb darkening coefficients they used.
Because we want to determine the mid-transit time
values using TAP instead of directly taking
them from the concerned papers,
we decided to calculate the limb darkening coefficients
even for the published light curves.
We used the EXOFAST routine (Eastman et al. 2013)
which incorporates the quadratic limb darkening tables of
Claret \& Bloemen (2011).
This tool requires
some input values which were picked from
Hartman et al. (2009) as :
effective temperature ($T_{\rm eff}$) = 4650 K,
surface gravity (log$g$) = 4.61 ${\rm cm/s^{2} }$ and
metallicity $[Fe/H] = -0.29$.
The values of the resulting limb darkening coefficients are
listed in Table~\ref{LD}.
As mentioned previously,
these values were defined with a Gaussian penalty
and a $\sigma$ of 0.05 while running the MCMC chains.
EXOFAST did not output the values of
limb darkening coefficients for the Gunn-$g$ and Gunn-$r$ bands.
So, for these filters,
we used the limb darkening coefficients
obtained for the Sloan-$g$ and Sloan-$r$ bands instead.
One can also choose a parameter to be `linked'
among different light curves if it is not completely locked.
In the present study, we have some light curves that represent the
same transit event, and hence, the same epoch.
We have linked the light curves representing the same epoch
together while calculating the mid-transit time for
such light curves. If the filters were different for those
light curves, we defined the values of limb darkening coefficients accordingly.
For each individual TAP run, five MCMC chains were calculated
and were added together to provide the final results.
The results from TAP for the mid-transit times are given in
Table~\ref{TAPresults}.
Please remember that all the light curves
corresponding to the same transit event are represented by a single epoch
in the table.
Epoch numbers 346, 446 and 1144 represent multiple light curves
(see Table~\ref{TAPresults} for more information).
Owing to this reason, Table~\ref{TAPresults} contains
25 epochs for the 31 light curves we have used.
The errors in the mid-transit time determined in this study
for the published light curves
are consistent with the
errors mentioned in Mallonn et al. (2015)
and Alexoudi et al. (2018) for the common light curves.
We also present the results for the photometric parameters
$a/R_{\ast}$
and $R_{\rm p}/R_{\ast}$
in Table~\ref{TAPresults} for individual epochs.
These parameters are also in agreement with the literature values.
Using radial velocity observations,
Knutson et al. (2014) listed the value of planet's mass $m_P$,
where they mention to have used
the sky-plane inclination $i$ from Hartman et al. (2009).
Using those, we calculated the corresponding value of $m_P {\rm sin} i$
for Knutson et al. (2014).
Using this $m_P {\rm sin} i$
and our TAP outputs for inclination during TAP runs,
we obtained the results of planet's mass according to our analysis.
Table~\ref{TAPresults} contains the TAP results for
eccentricity, inclination and planet's mass.
\begin{table*}
\centering
\caption{The results obtained from the TAP for mid-transit times
and some photometric parameters for individual light curves.
The calculated values of planet's mass are also presented here.
Epoch here is the sequential number of transit
with respect to the reference transit light curve from Hartman et al. (2009).
The notations for the data source imply:
(a)-- Hartman et al. (2009);
(b)-- Mancini et al. (2018);
(c)-- Lee et al. (2012);
(d)-- the present work;
and
(e)-- Alexoudi et al. (2018).
Among the data sources noted with asterisks,
the epoch number 346 represents 4 light curves from Mancini et al. (2018).
Also, 3 light curves from
Alexoudi et al. (2018) were observed during the same night
and thus, they belong to the common epoch number 1144.
One of our light curves from Tenagra observatory happens
to be observed on the same night as a published light curve.
}
\vspace{0.4cm}
\tiny
\begin{tabular}{cccccccc}
\hline\hline
Epoch & Data Source & $T_m$($BJD_{TDB}-2450000$) & $a$/$R_{\ast}$ & $R_{\rm p}$/$R_{\ast}$ & $e$ & $i$ & $m_P$ \\
& & day & & & & ($^{\circ}$) & ($M_J$)\\
\hline
0 & (a) & 4216.77244$^{+0.00023}_{-0.00022}$ & 11.84$^{+0.15}_{-0.19}$ & 0.1400$^{+0.0017}_{-0.0017}$ & 0.029$^{+0.021}_{-0.018}$ & 89.08$^{+0.37}_{-0.37}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
203 & (a) & 4869.02413$^{+0.00057}_{-0.00055}$ & 11.47$^{+0.23}_{-0.26}$ & 0.1450$^{+0.0032}_{-0.0030}$ & 0.029$^{+0.020}_{-0.017}$ & 88.96$^{+0.40}_{-0.39}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
212 & (a) & 4897.94185$^{+0.00084}_{-0.00089}$ & 12.19$^{+0.33}_{-0.34}$ & 0.1387$^{+0.0060}_{-0.0056}$ & 0.029$^{+0.021}_{-0.017}$ & 89.04$^{+0.39}_{-0.40}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
346 & (b$^{*}$) & 5328.49039$^{+0.00021}_{-0.00022}$ & 11.86$^{+0.11}_{-0.12}$ & 0.1389$^{+0.0011}_{-0.0010}$ & 0.026$^{+0.011}_{-0.011}$ & 89.12$^{+0.18}_{-0.18}$ & 0.20889$^{+0.01000}_{-0.00970}$ \\
446 & (c,d) & 5649.79746$^{+0.00019}_{-0.00020}$ & 11.72$^{+0.13}_{-0.15}$ & 0.1406$^{+0.0014}_{-0.0014}$ & 0.027$^{+0.015}_{-0.014}$ & 89.02$^{+0.27}_{-0.26}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
451 & (c) & 5665.86206$^{+0.00032}_{-0.00034}$ & 11.72$^{+0.20}_{-0.23}$ & 0.1438$^{+0.0022}_{-0.0020}$ & 0.030$^{+0.021}_{-0.018}$ & 88.95$^{+0.39}_{-0.37}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
452 & (d) & 5669.07486$^{+0.00077}_{-0.00082}$ & 11.77$^{+0.17}_{-0.17}$ & 0.1410$^{+0.0013}_{-0.0013}$ & 0.030$^{+0.020}_{-0.018}$ & 89.11$^{+0.35}_{-0.34}$ & 0.20889$^{+0.01000}_{-0.00970}$ \\
455 & (d) & 5678.71382$^{+0.00041}_{-0.00041}$ & 11.85$^{+0.20}_{-0.24}$ & 0.1370$^{+0.0019}_{-0.0020}$ & 0.029$^{+0.020}_{-0.018}$ & 89.05$^{+0.38}_{-0.39}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
460 & (c) & 5694.78087$^{+0.00023}_{-0.00023}$ & 11.83$^{+0.13}_{-0.16}$ & 0.1406$^{+0.0016}_{-0.0017}$ & 0.029$^{+0.021}_{-0.017}$ & 89.23$^{+0.33}_{-0.32}$ & 0.20889$^{+0.01000}_{-0.00970}$ \\
553 & (b) & 5993.59516$^{+0.00037}_{-0.00035}$ & 11.41$^{+0.24}_{-0.28}$ & 0.1388$^{+0.0028}_{-0.0029}$ & 0.029$^{+0.021}_{-0.017}$ & 88.63$^{+0.45}_{-0.43}$ & 0.20893$^{+0.01000}_{-0.00970}$ \\
698 & (b) & 6459.48810$^{+0.00019}_{-0.00020}$ & 11.66$^{+0.18}_{-0.22}$ & 0.1376$^{+0.0017}_{-0.0017}$ & 0.029$^{+0.021}_{-0.017}$ & 88.90$^{+0.39}_{-0.36}$ & 0.20891$^{+0.01000}_{-0.00970}$ \\
783 & (b) & 6732.59758$^{+0.00015}_{-0.00015}$ & 11.80$^{+0.10}_{-0.14}$ & 0.1453$^{+0.0011}_{-0.0010}$ & 0.030$^{+0.020}_{-0.017}$ & 89.29$^{+0.31}_{-0.33}$ & 0.20888$^{+0.01000}_{-0.00970}$ \\
784 & (d) & 6735.81023$^{+0.00029}_{-0.00029}$ & 11.69$^{+0.22}_{-0.25}$ & 0.1375$^{+0.0031}_{-0.0029}$ & 0.029$^{+0.020}_{-0.017}$ & 88.87$^{+0.42}_{-0.39}$ & 0.20891$^{+0.01000}_{-0.00970}$ \\
792 & (b) & 6761.51588$^{+0.00013}_{-0.00013}$ & 11.87$^{+0.11}_{-0.14}$ & 0.1414$^{+0.0009}_{-0.0009}$ & 0.029$^{+0.020}_{-0.017}$ & 89.26$^{+0.33}_{-0.33}$ & 0.20888$^{+0.01000}_{-0.00970}$ \\
793 & (d) & 6764.72820$^{+0.00040}_{-0.00040}$ & 11.99$^{+0.22}_{-0.25}$ & 0.1277$^{+0.0037}_{-0.0035}$ & 0.029$^{+0.021}_{-0.018}$ & 89.00$^{+0.38}_{-0.37}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
928 & (b) & 7198.49137$^{+0.00026}_{-0.00025}$ & 11.76$^{+0.18}_{-0.22}$ & 0.1473$^{+0.0028}_{-0.0030}$ & 0.029$^{+0.021}_{-0.018}$ & 89.04$^{+0.36}_{-0.36}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
934 & (d) & 7217.76896$^{+0.00059}_{-0.00055}$ & 11.77$^{+0.15}_{-0.15}$ & 0.1404$^{+0.0012}_{-0.0012}$ & 0.029$^{+0.020}_{-0.017}$ & 88.91$^{+0.36}_{-0.31}$ & 0.20891$^{+0.01000}_{-0.00970}$ \\
1027 & (e) & 7516.58280$^{+0.00025}_{-0.00025}$ & 11.66$^{+0.18}_{-0.22}$ & 0.1389$^{+0.0020}_{-0.0019}$ & 0.029$^{+0.020}_{-0.018}$ & 89.00$^{+0.38}_{-0.37}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
1045 & (b) & 7574.41888$^{+0.00025}_{-0.00024}$ & 11.74$^{+0.18}_{-0.22}$ & 0.1406$^{+0.0016}_{-0.0016}$ & 0.029$^{+0.020}_{-0.018}$ & 88.98$^{+0.40}_{-0.37}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
1125 & (e) & 7831.46347$^{+0.00032}_{-0.00031}$ & 11.72$^{+0.19}_{-0.22}$ & 0.1352$^{+0.0026}_{-0.0025}$ & 0.029$^{+0.021}_{-0.018}$ & 89.10$^{+0.37}_{-0.37}$ & 0.20889$^{+0.01000}_{-0.00970}$ \\
1126 & (e) & 7834.67580$^{+0.00051}_{-0.00049}$ & 11.98$^{+0.25}_{-0.28}$ & 0.1259$^{+0.0036}_{-0.0038}$ & 0.029$^{+0.020}_{-0.018}$ & 89.01$^{+0.40}_{-0.40}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
1139 & (e) & 7876.44492$^{+0.00022}_{-0.00021}$ & 11.84$^{+0.15}_{-0.17}$ & 0.1358$^{+0.0015}_{-0.0015}$ & 0.029$^{+0.019}_{-0.018}$ & 89.22$^{+0.32}_{-0.32}$ & 0.20889$^{+0.01000}_{-0.00970}$ \\
1144 & (e$^{*}$) & 7892.51130$^{+0.00014}_{-0.00015}$ & 11.81$^{+0.14}_{-0.15}$ & 0.1390$^{+0.0012}_{-0.0012}$ & 0.026$^{+0.012}_{-0.012}$ & 88.91$^{+0.25}_{-0.24}$ & 0.20891$^{+0.01000}_{-0.00970}$ \\
1149 & (e) & 7908.57577$^{+0.00026}_{-0.00026}$ & 11.72$^{+0.23}_{-0.26}$ & 0.1337$^{+0.0020}_{-0.0020}$ & 0.030$^{+0.020}_{-0.018}$ & 88.91$^{+0.40}_{-0.39}$ & 0.20891$^{+0.01000}_{-0.00970}$ \\
1159 & (e) & 7940.70745$^{+0.00037}_{-0.00038}$ & 11.81$^{+0.24}_{-0.28}$ & 0.1217$^{+0.0025}_{-0.0025}$ & 0.030$^{+0.021}_{-0.018}$ & 88.94$^{+0.41}_{-0.44}$ & 0.20890$^{+0.01000}_{-0.00970}$ \\
\hline
\label{TAPresults}
\end{tabular}
\end{table*}
\begin{table*}
\caption{A sample of the photometric light curve data of this work.
The TDB here demonstrates Barycentric Dynamical Time which is originated
from the French term `Temps Dynamique Barycentrique'.}
\centering
\begin{tabular}{cccc}
\hline\hline
Run & Epoch & TDB-based BJD & Relative Flux \\
\hline
1 & 446 & 2455649.712644 & 0.998728 \\
& & 2455649.713767 & 0.998597 \\
& & 2455649.714889 & 1.001292 \\
& & 2455649.716012 & 0.999719 \\
-& - & - & - \\
\hline
2 & 452 & 2455668.988182 & 0.996761 \\
& & 2455668.992052 & 0.998422 \\
& & 2455668.993144 & 0.996509 \\
& & 2455668.994237 & 0.991236 \\
-& - & - & - \\
\hline
-& - & - & - \\
\hline
\label{ourdata}
\end{tabular}
\end{table*}
Our normalized observational light curves
and the corresponding TAP fitting
with $x-$axis adjusted for the
mid-transit time are shown in Fig.~\ref{lightcurves}.
The light curves are in the sequence of the `run' defined in
Table~\ref{log}.
A few lines of our photometric observations
with $BJD_{TDB}$ and normalized relative flux are given in Table~\ref{ourdata}.
The full version of this table will be provided in
the machine readable format with this paper.
\section{The TTV Analysis}
\label{TTV}
\subsection{The Linear Fit and a New Ephemeris}
Once we have all the mid-transit times in
BJD$_{TDB}$, we can determine a new ephemeris
by $\chi^2$ minimization of the following linear relation:
\begin{equation}
T^C_{\rm m} (E)=T_0+PE,
\end{equation}
where $P$ and $E$ are period and epoch.
The reference time $T_0$ was arbitrarily chosen to be at epoch $E=0$.
For an individual epoch $E$, $T^C_{\rm m} (E)$ is the calculated mid-transit time.
Using linear fitting, we obtain
$T_{0} = 2454216.773311 \pm 0.000293 $ ($BJD_{TDB}$) and
$P = 3.21305762 \pm 0.00000036 $ (day).
If $\sigma_i$ is the mean of the error in the positive and
negative directions of an observed mid-transit time
given by TAP, then using observed
and calculated values of mid-transit times,
the $\chi^2$ of the fitting is determined using the formula:\\
\begin{equation}
\chi^2 =\sum_{i=1}^{N} \frac{({O}_i-{C}_i)^2}{{\sigma}_i^2},
\end{equation}
where ${O}_i$ is an observed mid-transit time,
${C}_i$ is a calculated mid-transit times, and $N$ is the
number of included epochs.
The value of the $\chi^2$ for the linear fitting is
182.49.
There are 23 degrees of freedom in our model, so
the reduced $\chi^2$
becomes,
$\chi^2_{red}(23)$ = 7.93.
This large value of the $\chi^2_{red}$ in the linear fitting
can be
indicative of the presence of TTVs.
Ideally, when there is no TTV,
the time between any two adjacent transit events
should be exactly equal to the orbital period.
The $O-C$ diagram for the linear fitting is presented in
Fig.~\ref{OCLinear}, which shows deviation between
the observed mid-transit time and the one predicted by
a simple two-body orbit.
\subsection{The Frequency Analysis}
\label{FREQ}
We searched for possible frequencies which might be causing
variation in the data using
generalized Lomb-Scargle periodogram
(Lomb 1976, Scargle 1982, Zechmeister \& Kuerster 2009).
This procedure considers the error bars while determining the periodogram.
The periodogram is shown in the Fig.~\ref{periodogram}.
If $f$ is the frequency related to the highest peak of power
in the periodogram, then the possible TTVs are tested by $\chi^2$
minimization of the following equation that consists of
both linear and sinusoidal terms:
\begin{equation}
T_S (E) = P E+b+z\sin(2\pi f E -\phi).
\end{equation}
In the equation above, the predicted mid-transit time at a given epoch
$E$ is $T_S (E)$
while $P, b,$ amplitude $z$
and phase $\phi_1$ are the fitting parameters.
The frequency corresponding to the highest power peak
($f= 0.00790059461$ epoch$^{-1}$,
allows us to determine the fitting parameters:
$P= 3.21305803 \pm 0.00000019 $ day,
$b= 2454216.773065 \pm 0.000145 $ day,
$z= -0.000754 \pm 0.000080 $ day,
and
$\phi= 4.163 \pm 0.109$ rad.
The value of the $\chi^2$ is 88.02.
For 25 data points, we are determining four
parameters from fitting.
This model has 21 degrees of freedom, where
the value of the reduced $\chi^2$ decreases to 4.19.
The $O-C$ diagram as a function of epoch $E$,
for one frequency scenario is given in Fig.~\ref{sinefitallfreq}.
The $O-C$ value shown in the curve
depicts the value of $T_S (E)$
minus the linear term ($P E+b$).
The data points representing the light curves
are also adjusted according to the fitting and are shown in the figure.
The false-alarm probability (FAP)
was determined following the procedure explained in
Press et al. (1992).
As can be seen in Fig.~\ref{periodogram}, the FAP
for the frequency with maximum power is 61\%.
Only the peaks with a significantly high signal-to-noise power ratio
should be considered from a periodogram
(Breger et al. 1993; Kuschnig et al. 1997).
Since no other peak shown in Fig.~\ref{periodogram} has a high S/N ratio,
we did not consider any other peak for frequnecy analysis.
To conclude, considering that the $\chi^2_{red}$ value is around 4 and
the FAP is large, the possible TTVs are probably of the non-sinusoidal nature.
\section{The Two-Planet Model}
\label{dynamics}
The values of $\chi^2_{red}$ in the above analysis indicate
possible non-sinusoidal TTVs in HAT-P-12 planetary system.
In order to probe a physical scenario for the explanation
of these TTVs, we explore the possibility to have an additional
exoplanet (HAT-P-12c) in this system (see, Nesvorn\'y et al. 2012 for example).
In the approach we used, by feeding some assumed initial input values of
the parameters for both the planets, HAT-P-12b and HAT-P-12c,
the theoretical TTVs are produced through the dynamical calculations of
the {\it TTVFast} code (Deck et al. 2014).
These theoretical TTVs are used to fit our observational mid-transit times.
The best-fit model can be obtained
through a MCMC sampling code {\it MC3} (Cubillos et al. 2017).
Before running the MCMC sampling, we first need to set
the distributions and the ranges of numerical values of photometric parameters
for both planets.
For HAT-P-12b, the parameters are already determined in
the previous sections of this paper.
So, these parameters were set as either fixed values or with a
certain range around the previously determined values.
For example, since the orbital period can vary slightly
during the orbital integration, we provide a total interval width of
0.2 day for the orbital period of HAT-P-12b.
The orbital eccentricity and inclination of HAT-P-12b
are taken as the mean values
of the results shown in Table~\ref{TAPresults}.
In order to search for the best-fit model for the new exoplanet
HAT-P-12c, the initial input values were set within larger
ranges and are set to be uniformly distributed.
Also note that the mass of central star (HAT-P-12) is set to be
0.733 $M_\odot$ according to Hartman et al. (2009).
Table~\ref{MC3input} gives a summary of the input parameters for {\it TTVFast}.
For the parameters with a defined range of input values,
the values are given inside the brackets, [ ].
\begin{table*}
\caption{The parameter setting of the two-planet model.
The notations and units of the parameters are also given in column 1.
The range of input values for some parameters are defined in [ ]. }
\centering
\begin{tabular}{ccc}
\hline\hline
Parameter & HAT-P-12b & HAT-P-12c \\
\hline
mass ($m_{\rm p}$, $M_{\rm J}$) & [0.205, 0.213] & [0.0001, 1] \\
period ($P$,day) & [3.1130, 3.3130] & [3.3, 16.5] \\
orbital eccentricity ($e$) & $0.02898^{(\#)}$ & [0.0, 0.2] \\
orbital inclination ($i$,$^{\circ}$) & $89.02176^{(\#)}$ &[59.02176, 119.02176] \\
longitude of ascending node ($\Omega$, $^{\circ}$)& $0.0^{(\#)}$ & [-30, 30] \\
argument of pericenter ($\omega$, $^{\circ}$) & $0.0^{(\#)}$ & [0, 360] \\
mean anomaly ($M$, $^{\circ}$) & $(\#\#)$ & [-180, 180] \\
\hline
\multicolumn{3}{l}
{\footnotesize Remark $(\#)$ indicates that the parameter is fixed.}\\
\multicolumn{3}{l}
{\footnotesize Remark $(\#\#)$ indicates that the mean anomaly of HAT-P-12b
is determined by other parameters.}\\
\label{MC3input}
\end{tabular}
\end{table*}
As shown in Table~\ref{MC3input}, nine parameters can change
their values during the MCMC sampling,
with a total number of samples being $2\times 10^{7}$.
After we obtain the above result,
in order to have more MCMC samples
with parameters closer to the best-fit model,
the MCMC sampling is executed again while nine parameters are now
within smaller ranges as shown in Fig.~\ref{figMCMC}.
Fig.~\ref{figMCMC} presents
the MCMC posterior distributions of these nine parameters.
Both two-dimensional and one-dimensional projections
are plotted.
Those parameters with subscript `$b$' are for the exoplanet HAT-P-12b,
and those with subscript `$c$' are for the exoplanet HAT-P-12c.
These distributions give the probabilities that
particular numerical values are employed during the MCMC sampling.
The color panels are the pairwise distributions.
The histograms are the one-dimensional distributions,
where grey areas denote $68\%$ highest
posterior density regions of the distributions.
The dotted lines in the figure
indicate the values of the best-fit model.
The corresponding results for the parameter values
according to this best-fit model are given in Table~\ref{bestmodel}.
Since nine parameters are being determined, there are
16 degrees of freedom.
The reduced $\chi^2$ of this best-fit model thus becomes
$\chi^2_{red}(16)$ = 2.09, which is much smaller than
the reduced $\chi^2$ values during linear fitting and frequency analysis.
The theoretical TTVs of this best-fit, two-planet model
are plotted as the curve shown in Fig.~\ref{figdynamicalTTV}.
The data points with error bars in this figure are
$O-C$ values for the observational data (same as Fig.~\ref{OCLinear}).
The bottom panel in Fig.~\ref{figdynamicalTTV} shows
the residuals of fitting the model to the $O-C$ data points.
The standard deviation of the residuals is 0.61 min
while the average value of the means of error bars for
$O-C$ values is $\sim$0.49 min.
It is evident from Fig.~\ref{figdynamicalTTV} that
the theoretical curve lies within the error bars of
observational data for most of the epochs.
Therefore, judging from a reasonably good data fitting
and a smaller value of the reduced $\chi^2$,
we deduce that this two-planet model
could explain the observational TTV of HAT-P-12 planetary system.
Considering the orbital period and inclination of HAT-P-12c
(Table~\ref{bestmodel}),
it is impossible to have transits for this exoplanet
unless its radius is larger than 40 times that of Jupiter.
This explains why there are no observed transit events for this exoplanet.
\begin{table*}
\caption{Results from the best-fit model for two-planet scenario.
The subscripts $b$ or $c$ are added to distinguish between the two planets.}
\vspace{0.5mm}
\centering
\begin{tabular}{ccccccccc}
\hline\hline
$m_{\rm pb}$ & $P_{\rm b}$ & $m_{\rm pc}$ & $P_{\rm c}$ & $e_{\rm c}$ &
$i_{\rm c}$ & $\Omega_{\rm c}$ & $\omega_{\rm c}$ & $M_{\rm c}$\\
($M_{\rm J})$ & (day) & ($M_{\rm J}$) & (day) & &
($^{\circ}$) & ($^{\circ}$) & ($^{\circ}$) & ($^{\circ}$) \\
\hline
0.212 & 3.2134 &
0.218 & 8.8530 & 0.15499 & 73.49569 & -5.58 & 52.785 & 18.892\\
\hline
\label{bestmodel}
\end{tabular}
\end{table*}
\section{Conclusions}
\label{CONC}
Using the telescopes from three observatories,
we present six new light curves of the transiting exoplanet
HAT-P-12b. These observations were combined with 25 light curves from
published papers to further enrich the baseline of data to 1160 epochs.
A self-consistent homogeneous analysis was carried out for all
the light curves to make sure
that our TTV results are not affected by any systematics.
The photometric parameters determined by us are in agreement to their
values in earlier published works.
We determined new ephemeris for the HAT-P-12b system
by a linear fit and sinusoidal curve fitting for different frequencies.
The values of reduced $\chi^2$ from the linear fitting is
7.93
while from the sine-curve fitting
for the highest power frequency,
the value of $\chi^2_{red}$ is obtained as
4.19.
These values and the large FAP indicate that the
TTV could be non-sinusoidal.
Finally, through a MCMC sampling, a two-planet model
is found to be able to produce a theoretical TTV
which could explain the observations to a satisfactory level with
a value of $\chi^2_{red}$ = 2.09.
Therefore, a scenario
with a non-transiting exoplanet might explain the
TTV of HAT-P-12b.
To conclude, our results show the existence of non-sinusoidal
TTVs. Though a two-planet model could lead to a better fitting,
the validation of a new exoplanet is out of the scope and not provided here.
Hopefully, the nature of this system could be further understood
in the future.
\begin{acknowledgements}
We are thankful to the referee of this paper for very helpful suggestions.
This work is supported by the grant from
the Ministry of Science and Technology (MOST), Taiwan.
The grant numbers are
MOST 105-2119-M-007 -029 -MY3
and MOST 106-2112-M-007 -006 -MY3.
Devesh P. Sariya is grateful to
Chiao-Yu Lee, Dr. Chien-Yo Lai, and Dr. Arti Joshi
for useful discussion on data reduction,
and also the Crimean Astrophysical Observatory (CrAO) for
the hospitality and exchange of knowledge during his research visit.
P.T. and V.K.M. acknowledge the University Grants
Commission (UGC), New Delhi, for providing the financial
support through Major Research Project no. UGC-MRP 43-521/2014(SR).
P.T. expresses his sincere thanks to IUCCA, Pune, India for providing
the supports through IUCCA Associateship Programme.
D. Bisht is financially supported by
the Natural Science Foundation of China
(NSFC-11590782, NSFC-11421303).
JJH is supported by the B-type Strategic Priority Program of the Chinese
Academy of Sciences (Grant No. XDB41000000), the National Natural Science
Foundation of China (Grant No. 11773081), CAS Interdisciplinary Innovation
Team, Foundation of Minor Planets of the Purple Mountain Observatory.
\end{acknowledgements}
| 3215bc29aa4630186b95d793de6caff288507543 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\subsection{Classical story}
Given a negative integer $d$ with $d\equiv 0 \text{ or } 1 \bmod 4$, let $h(d)$ be the proper ideal class number of the imaginary quadratic order $\mathcal{O}_d$ with discriminant $d$. Put $w(d):=\#(\mathcal{O}_d^\times)/2$.
The classical Kronecker-Hurwitz class number relation says that for a non-square $n \in {\mathbb N}$,
\begin{eqnarray}\label{eqn: HCNR}
\sum_{\subfrac{t \in {\mathbb Z}}{t^2< 4n}}\left(\sum_{\subfrac{d \in {\mathbb N}}{d^2\mid (t^2-4n)}}\frac{h\big((t^2-4n)/d^2\big)}{w\big((t^2-4n)/d^2\big)}\right)&=&\sum_{\subfrac{m \in {\mathbb N}}{m\mid n}} \max(m,n/m).
\end{eqnarray}
One can derive the above identity via ``modular polynomial'', i.e.\ the defining equation of the graph of the Hecke correspondence $T_n$, $n \in {\mathbb N}$, on the modular curve $X$ of full level (cf.\ \cite{G-K}).
In particular, the quantity in \eqref{eqn: HCNR} is equal to the ``finite part'' of the intersection number of the divisors $T_1$ and $T_n$ on the surface $X\times X$. Taking the ``infinite part'' (from cuspidal intersections) into account, the total intersection number of $T_1$ and $T_n$ becomes
$$T_1\cdot T_n = 2 \sigma(n),$$
where $\sigma(n) := \sum_{m\mid n} m$ is precisely the $n$-th Fourier coefficient of the weight-two Eisenstein series (normalized so that the first Fourier coefficient equals to $1$).
This provides a very concrete example in the following connections:
$$
\xymatrix{
\{\text{Class numbers}\}\ar@{<->}[rr]\ar@{<->}[dr] & & \{\text{Intersections}\} \ar@{<->}[dl] \\
&\{\text{Fourier coefficients}\} &
}
$$
In the celebrated work of Hirzebruch and Zagier \cite{H-Z}, the whole theory on the ground of the Hilbert modular surfaces associated with real quadratic fields is well-established.
More precisely, they express the intersections of certain special divisors in terms of Hurwitz class numbers, and show that the generating function associated with these intersection numbers is actually a particular Eisenstein series with nebentypus.
The interpretations for the Fourier coefficients of Eisenstein series, which have been generalized to the ``Kudla-Millson'' theta integrals (cf.\ \cite{K-MI} and \cite{K-M}) on the quotients of symmetric spaces for orthogonal and unitary groups, are viewed as \textit{geometric Siegel-Weil formula} and have various applications (cf.\ \cite{Kud2}, \cite{Cog}, \cite{K-MII}, \cite{Kud}, and \cite{Fun} ).
Moreover, connections with the class numbers make it possible to compute explicitly the intersections in question (cf.\ \cite{H-Z}, \cite{Fun}, and \cite{G-Y}).
The purpose of this paper is to attempt an exploration of this phenomenon in the function field setting, and to derive a Hirzebruch-Zagier style geometric interpretation for the class number relations in the world of positive characteristic.
\subsection{Function Field analogue}
Let $A = {\mathbb F}_q[\theta]$, the polynomial ring with one variable $\theta$ over a finite field ${\mathbb F}_q$ with $q$ elements, and let $k$ be the field of fractions of $A$.
Let $k_\infty$ be the completion of $k$ with respect to the ``degree valuation'' (cf.\ Section~\ref{sec: BS}), and denote by ${\mathbb C}_\infty$ the completion of a chosen algebraic closure of $k_\infty$.
The \textit{Drinfeld half plane} is $\mathfrak{H}:= {\mathbb C}_\infty - k_\infty$, equipped with the M\"obius left action of $\operatorname{GL}_2(k_\infty)$.
Let $B$ be a quaternion algebra over $k$ which is split at $\infty$ (i.e.\ $B\otimes_k k_\infty \cong \operatorname{Mat}_2(k_\infty)$), and $O_B$ be an Eichler $A$-order in $B$ of type $(\mathfrak{n}^+,\mathfrak{n}^-)$ (cf.\ Section~\ref{sec: TSP}).
Then the embedding $B^\times \hookrightarrow \operatorname{GL}_2(k_\infty)$ induces an action of $\Gamma(\mathfrak{n}^+,\mathfrak{n}^-):= O_B^\times$ on $\mathfrak{H}$.
The quotient space $$
X(\mathfrak{n}^+,\mathfrak{n}^-)\ :=\
\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)\backslash \mathfrak{H}
$$
is called the \textit{Drinfeld-Stuhler modular curve for $\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)$}.
When $B = \operatorname{Mat}_2(k)$, the group $\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)$ coincides (up to conjugations) with the congruence subgroup
$$
\Gamma_0(\mathfrak{n}^+):=
\left\{\begin{pmatrix} a&b \\ c&d\end{pmatrix} \in \operatorname{GL}_2(A)\ \bigg|\ c \equiv 0 \bmod \mathfrak{n}^+\right\},
$$
and the compactification of $X(\mathfrak{n}^+,\mathfrak{n}^-)$ is the so-called \textit{Drinfeld modular curve for $\Gamma_0(\mathfrak{n}^+)$}.
\begin{rem}
As in the classical case, the study of Drinfeld modular polynomials in \cite{Bae}, \cite{B-L}, and \cite{Hsia} give an analogue of the Kronecker-Hurwitz class number relation for ``imaginary'' quadratic $A$-orders (cf. \cite{JKYu} and \cite{W-Y}). Also, the connection with the intersections of the Hecke correspondence on the Drinfeld modular curves is derived in \cite{JKYu} when $q$ is odd.
Moverover, these intersection numbers appear in the Fourier expansion of the ``improper'' Eisenstein series on $\operatorname{GL}_2(k_\infty)$ which is introduced by Gekeler (cf.\ \cite{Gek1} and \cite{Gek2}). Thus, a parallel story for the Kronecker-Hurwitz case over rational function fields is developed. We may also expect to see these connections when the base field $k$ is an arbitrary global function field.
\end{rem}
From now on, we always assume that $q$ is \textbf{odd}.
Fix a monic square-free $\mathfrak{d} \in A$ with even degree. Then the quadratic field $F:= k(\sqrt{\mathfrak{d}})$ is \textit{real} over $k$, (i.e.\ the infinite place of $k$ is split in $F$). The embedding $F\hookrightarrow F\otimes_k k_\infty \cong k_\infty\times k_\infty$ induces
$$\operatorname{GL}_2(F) \hookrightarrow \operatorname{GL}_2(k_\infty) \times \operatorname{GL}_2(k_\infty),$$
providing an action of $\operatorname{GL}_2(F)$ on $\mathfrak{H}_F:= \mathfrak{H} \times \mathfrak{H}$.
Let $O_F$ be the integral closure of $A$ in $F$.
Given a monic $\mathfrak{n} \in A$, put
$$
\Gamma_{0,F}(\mathfrak{n}) := \left\{
\begin{pmatrix} a&b \\ c&d \end{pmatrix} \in \operatorname{GL}_2(O_F)
\ \bigg|\ ad-bc \in {\mathbb F}_q^\times \text{ and } c \equiv 0 \bmod \mathfrak{n}\right\}.
$$
The \textit{Drinfeld-Stuhler modular surface for $\Gamma_{0,F}(\mathfrak{n})$} is
$$\mathcal{S}_{0,F}(\mathfrak{n}):= \Gamma_{0,F}(\mathfrak{n}) \backslash \mathfrak{H}_F,$$
which is a coarse moduli scheme for the so-called \textit{Frobenius-Hecke sheaves} (with additional ``level-$\mathfrak{n}$ structure'') introduced by Stuhler in \cite{Stu} (see also \cite{Ozg}).
We are interested in the intersections between the ``Hirzebruch-Zagier-type divisors'' on $\mathcal{S}_{0,F}(\mathfrak{n})$ defined as follows.
For $x = \begin{pmatrix}a&b \\ c&d\end{pmatrix} \in \operatorname{Mat}_2(F)$, we put
$$
\bar{x} := \begin{pmatrix}d&-b \\ -c&a\end{pmatrix}
\quad \text{ and } \quad
x' := \begin{pmatrix} a'&b' \\ c'&d'\end{pmatrix},
$$
where for every $\alpha \in F$, $\alpha'$ is the conjugate of $\alpha$ under the action of the non-trivial element in $\operatorname{Gal}(F/k)$.
Consider the involution $*$ on $\operatorname{Mat}_2(F)$ defined by
$$ x^* \ :=\ \begin{pmatrix}0&1/\mathfrak{n} \\ 1&0\end{pmatrix} \bar{x}'\begin{pmatrix}0&1 \\ \mathfrak{n} & 0 \end{pmatrix}, \quad \forall x \in \operatorname{Mat}_2(F).
$$
Let
$$
\Lambda:= \{x \in \operatorname{Mat}_2(O_F)\mid x^* = x\}.
$$
We have a left action of $\Gamma:= \Gamma_{0,F}(\mathfrak{n})$ on $\Lambda$ by
$$ \gamma \star x := \gamma x \gamma^* \cdot (\det \gamma )^{-1}, \quad \gamma \in \Gamma_{0,F}(\mathfrak{n}) \text{ and } x \in \Lambda.
$$
For each $x$ in $\Lambda$ with $\det x \neq 0$, let
$$B_x:= \{b \in \operatorname{Mat}_2(F) \mid x b^* = \bar{b} x\}
\quad \text{ and } \quad
\Gamma_x := B_x^\times \cap \Gamma.
$$
From Lemma~\ref{lem: B_x}, we know that $B_x$ is an \textit{indefinite} quaternion algebra over $k$ (i.e.\ unramified at the infinite place of $k$), and so the quotient $\mathcal{C}_x:= \Gamma_x \backslash \mathfrak{H}$ becomes the Drinfeld-Stuhler modular curve for $\Gamma_x$ (cf.\ Section~\ref{sec: CMBES}).
Put
$$S_x := \begin{pmatrix}0&1 \\ \mathfrak{n} &0\end{pmatrix} \bar{x}.$$
The embedding from $\mathfrak{H}$ into $\mathfrak{H}_F$ defined by $(z \mapsto (z,S_x z))$ gives rise to a (rigid analytic) morphism $f_x: \mathcal{C}_x \rightarrow \mathcal{S}_{0,F}(\mathfrak{n})$, and we set
$$\mathcal{Z}_x := f_{x,*}(\mathcal{C}_x),$$
the push-forward divisor of $f_x$ on $S_{0,F}(\mathfrak{n})$.
For non-zero $a \in A$, the \textit{Hirzebruch-Zagier divisor of discriminant $a$} is:
$$\mathcal{Z}(a) := \sum_{x\in\Gamma \backslash \Lambda_a}\mathcal{Z}_x, \quad \text{ where } \Lambda_a:= \{x \in \Lambda \mid \det(x) = a\}.$$
${}$
Notice that we may identify $B_1$ with the quaternion algebra
$$\left(\frac{\mathfrak{d}, \mathfrak{n}}{k}\right) := k+k\mathbf{i}+k\mathbf{j}+k\mathbf{ij} \quad \text{ with } \mathbf{i}^2 = \mathfrak{d},\ \mathbf{j}^2 = \mathfrak{n}, \text{ and } \mathbf{ji} = -\mathbf{ij}.
$$
In particular, suppose $\mathfrak{n}$ is square-free and coprime to $\mathfrak{d}$.
Write $\mathfrak{n} = \mathfrak{n}^+\cdot \mathfrak{n}^-$ and $\mathfrak{d} = \mathfrak{d}^+ \cdot \mathfrak{d}^-$, where for each prime factor $\mathfrak{p}$ of $\mathfrak{n}^{\pm}$ (resp.\ $\mathfrak{d}^{\pm}$) we have the Legendre quadratic symbol $\left(\frac{\mathfrak{d}}{\mathfrak{p}}\right) = \pm 1$ (resp.\ $\left(\frac{\mathfrak{n}}{\mathfrak{p}}\right) = \pm 1$).
Then $B_1$ is ramified precisely at the prime factors of $\mathfrak{d}^-\mathfrak{n}^-$,
and
$O_{B_1}:= B_1\cap \operatorname{Mat}_2(O_F)$ is an Eichler $A$-order of type $(\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-)$ in $B_1$.
In other words, $\mathcal{C}_1$ is actually the Drinfeld-Stuhler modular curve for $\Gamma(\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-)$.
We pick $\mathcal{Z}_1$ as our ``base'' divisor on $\mathcal{S}_{0,F}(\mathfrak{n})$, and determine the intersection number of $\mathcal{Z}_1$ and $\mathcal{Z}(a)$ for non-zero $a \in A$ in the following:
\begin{thm}\label{thm: MT1}
Suppose $\mathfrak{n} \in A_+$ is square-free with coprime to $\mathfrak{d}$ and $\deg (\mathfrak{d}^-\mathfrak{n}^-) >0$.
The intersection number of $\mathcal{Z}_1$ and $\mathcal{Z}(a)$ for non-zero $a \in A$ is equal to
$$
\mathcal{Z}_1\cdot \mathcal{Z}(a) = 2\cdot \sum_{\subfrac{t \in A}{t^2-4a\preceq 0}}H^{\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}(\mathfrak{d}(t^2-4a)).
$$
Here for $d \in A$, we write $d\preceq 0$ if $d=0$ or $k(\sqrt{d})$ is an ``imaginary'' quadratic extension of $k$ (i.e.\ the infinite place of $k$ does not split), and $H^{\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}(d)$ is the modified Hurwitz class number in {\rm Definition~\ref{defn: HCN}}
and Remark~{\rm \ref{rem: H(0)}}.
\end{thm}
We point out that the self-intersection number of $\mathcal{Z}_1$ is defined to be
an analogue of the ``Euler characteristic'' of $\mathcal{Z}_1$ in Definition~\ref{defn: SIN}.\\
To establish the equality in Theorem~\ref{thm: MT1}, the main bridge is the (adelic) theta integral $I(\cdot;
\varphi_\Lambda)$ associated with a particular chosen Schwartz function $\varphi_\Lambda$ in Section~\ref{sec: TSN-PSF}.
Our strategy is briefly sketched as follows.
Notice that using adelic language, we may express very naturally the $a$-th Fourier coefficient of $I(\cdot;\varphi_\Lambda)$ for a given non-zero $a \in A$ in terms of the modified Hurwitz class numbers (cf.\ Theorem~\ref{thm: TINFC}).
On the other hand, the strong approximation theorem (for the indefinite quaternion algebra ramified precisely at the prime factors of $\mathfrak{n}^-$) leads to an alternative expression of the $a$-th Fourier coefficient of $I(\cdot;\varphi_\Lambda)$ (cf.\ Theorem~\ref{thm: AFCN}), which enables us to connect the Fourier coefficient with the intersection number $\mathcal{Z}_1\cdot \mathcal{Z}(a)$ (Theorem~\ref{thm: CN-Int} and Corollary~\ref{cor: TINFC}).
This completes the proof.\\
The theta integral $I(\cdot;\varphi_\Lambda)$ has nice invariant property and transformation law (cf.\ Propostion~\ref{prop: TLN}).
In particular, the choice of the ``infinite component'' of $\varphi_\Lambda$ in \eqref{eqn: phi_infty} is most crucial in bridging two sides of the equality in Theorem~\ref{thm: MT1}, and ensures as well the ``harmonicity'' of $I(\cdot;\varphi_\Lambda)$ (cf.\ Lemma~\ref{lem: har}).
This allows us to extend $I(\cdot;\varphi_\Lambda)$ to a ``Drinfeld-type'' automorphic form on $\operatorname{GL}_2(k_\infty)$ (an analogue of weight-$2$ modular forms over function fields, see \textit{Remark}~\ref{rem: wt-2} and \cite{G-R}) with nebentypus character $\left(\frac{\cdot}{\mathfrak{d}}\right)$ for $\Gamma_0^{(1)}(\mathfrak{d}\mathfrak{n}):= \Gamma_0(\mathfrak{d}\mathfrak{n})\cap \operatorname{SL}_2(A)$, cf.\ Theorem~\ref{thm: Ext-DTAF}.
In other words, we have the following theorem (cf.\ \textit{Remark}~\ref{rem: DTS-FC}):
\begin{thm}\label{thm: MT2}
Under the assumptions in \text{\rm Theorem~\ref{thm: MT1}}, there exists a Drinfeld-type automorphic form $\vartheta_\Lambda$ on $\operatorname{GL}_2(k_\infty)$ with nebentypus character $\left(\frac{\cdot}{\mathfrak{d}}\right)$ for the congruence subgroup $\Gamma_0^{(1)}(\mathfrak{d}\mathfrak{n})$ whose Fourier expansion is given by: for $(x,y) \in k_\infty \times k_\infty^\times$,
$$
\vartheta_\Lambda\begin{pmatrix} y & x \\ 0&1 \end{pmatrix} = -\text{\rm vol}(\mathcal{Z}_1)\beta_{0,2}(y) + \sum_{0 \neq a \in A}\big(\mathcal{Z}_1\cdot \mathcal{Z}(a)\big) \cdot \big(\beta_{a,2}(y) \psi_\infty(ax)\big).
$$
Here
\begin{itemize}
\item $\beta_{a,s}(y)$ is the following truncated analogue of the Bessel function:
$$\beta_{a,s}(y):=|y|_\infty^{s/2}\cdot
\begin{cases} 1, & \text{ if $\deg a + 2 \leq \operatorname{ord}_\infty(y)$;}\\
0, & \text{ otherwise;}
\end{cases}$$
\item $|\cdot|_\infty$ is the absolute value on $k_\infty$ normalized so that $|\theta| = q$, \item $\psi_\infty: k_\infty \rightarrow {\mathbb C}^\times$ is a fixed additive character on $k_\infty$ defined in {\rm Section~\ref{sec: TMk}},
\item
$$\text{\rm vol}(\mathcal{Z}_1)
:= 2 H^{\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}(0).
$$
\end{itemize}
\end{thm}
\begin{rem}
(1) Our approach for the geometric interpretation of the Fourier coefficients of $\vartheta_\Lambda$ as the corresponding intersection numbers is basically in the framework of \cite{K-M} and \cite{Fun}, while the rigid analytic geometry is utilized and the test function constructed by Kudla-Millson at the archimedean place is replaced by a proper Schwartz function in \eqref{eqn: phi_infty} for this positive-characteristic setting.
(2) The technical assumption ``$\deg(\mathfrak{d}^-\mathfrak{n}^-)>0$'' in Theorem~\ref{thm: MT1} implies that the Drinfeld-Stuhler modular curve $\mathcal{C}_1$ has no ``cusps''.
Therefore there are no contributions of the ``cuspidal intersections'' to $\mathcal{Z}_1 \cdot \mathcal{Z}(a)$ in Theorem~\ref{thm: MT1} and Theorem~\ref{thm: MT2}.
When $\mathfrak{d}^-\mathfrak{n}^- = 1$,
this argument would need to be adjusted by ``regularizing the theta integral $I(\cdot;\varphi_\Lambda)$'' as in \cite{Fun}, and taking the cuspidal intersections into account for $\mathcal{Z}_1\cdot \mathcal{Z}(a)$ in a suitable ``compactification of the surface $\mathcal{S}_{0,F}(\mathfrak{n})$''.
However, due to a lack of studies in the literature for these two technical issues in the function field context, we make this assumption in Theorem~\ref{thm: MT1} first.
The general case will be explored in future work.
(3) Given non-zero $d \in A$, a point $\boldsymbol{z}$ in $\mathcal{C}_1 = X(\mathfrak{d}^+ \mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-)$ is \textit{CM with discriminant $d$} if it has a representative $z \in \mathfrak{H}$ and there exists $x_z \in O_{B_1}$ with $x_z^2 = d$ and $x_z \cdot z = z$
(cf.\ Definition~\ref{defn: CMP}).
As $z \notin k_\infty$, the existence of such a CM point $\boldsymbol{z}$ ensures that $d \prec 0$. Moreover, for non-zero $a \in A$, the intersection of $\mathcal{Z}_1$ and $\mathcal{Z}(a)$, excluding possible self-intersection of $\mathcal{Z}_1$, is actually supported by the image of the CM points on $\mathcal{C}_1$ with discriminant $\mathfrak{d}(t^2-4a)$ for $t \in A$ and $t^2-4a \prec 0$ (cf.\ \textit{Remark}~\ref{rem: CM-sup}).
\end{rem}
We also study the ``mass'' over CM points on the Drinfeld-Stuhler modular curves via ``metaplectic'' theta series.
Let $\mathfrak{n}^+, \mathfrak{n}^- \in A_+$ be two square-free polynomials with $\text{gcd}(\mathfrak{n}^+,\mathfrak{n}^-) = 1$, and the number of prime factors of $\mathfrak{n}^-$ is positive and even.
Given a point $\boldsymbol{z}$ on $X(\mathfrak{n}^+,\mathfrak{n}^-)$ represented by $z \in \mathfrak{H}$, let $\operatorname{Stab}_{\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)}(z)$ be the stablizer of $z$ in $\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)$, and put
$$w(\boldsymbol{z}):= \frac{q-1}{
\#\big(\text{Stab}_{\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)}(z)\big)}.$$
For non-zero $d \in A$,
the \textit{mass} over the CM points with discriminant $d$ on $X(\mathfrak{n}^+,\mathfrak{n}^-)$ is:
$$
\mathcal{M}(d) := \sum_{\subfrac{\text{CM } \boldsymbol{z} \in X(\mathfrak{n}^+,\mathfrak{n}^-)}{\text{with discriminant $d$}}}\frac{1}{w(\boldsymbol{z})}.
$$
Comparing with the classical story, it is natural to expect that this ``mass'' amounts to generating function of a metaplectic form.
We derive that
(cf.\ Theorem~\ref{thm: CNo-geo} and Corollary~\ref{cor: TSo-FC}):
\begin{thm}\label{thm: Intro.MS-HCN}
Let $\mathfrak{n}^+, \mathfrak{n}^- \in A_+$ be two square-free polynomials with $\text{gcd}(\mathfrak{n}^+,\mathfrak{n}^-) = 1$, and the number of prime factors of $\mathfrak{n}^-$ is positive and even.
Given non-zero $d \in A$, the following equality holds:
$$\mathcal{M}(d) = \begin{cases}
H^{\mathfrak{n}^+,\mathfrak{n}^-}(d), & \text{ if $d \prec 0$;}\\
0, & \text{ otherwise.}
\end{cases}$$
Moreover, if we put
$\mathcal{M}(0):= H^{\mathfrak{n}^+,\mathfrak{n}^-}(0)$,
then there exists a ``weight-$\frac{3}{2}$'' metaplectic form $\vartheta^o$ on $\widetilde{\operatorname{GL}}_2(k_\infty)$ (introduced in {\rm Section~\ref{sec: Ext-PTI}}) with the following Fourier expansion: for $(x,y) \in k_\infty\times k_\infty^\times$,
$$
\vartheta^o\left(\begin{pmatrix} y&x\\ 0&1\end{pmatrix},1\right)
=
\sum_{d \in A}
\mathcal{M}(d) \cdot \big(\beta_{d,\frac{3}{2}}(y)\psi_\infty(-dx)\big).
$$
\end{thm}
\begin{rem}
${}$
\begin{itemize}
\item[(1)] When the number of prime factors of $\mathfrak{n}^-$ is odd, we are still able to extend the generating function
$$
\sum_{d \in A, \ d \preceq 0}H^{\mathfrak{n}^+,\mathfrak{n}^-}(d)\cdot \big(\beta_{d,\frac{3}{2}}(y)\psi_\infty(-dx)\big), \quad \forall (x,y) \in k_\infty\times k_\infty^\times
$$
to a weight-$\frac{3}{2}$ metaplectic form (coming from the theta series associated with ``definite'' pure quaternions ramified precisely at the prime factors of $\mathfrak{n}^-$ and $\infty$ (cf.\ \cite[Proposition 2.2]{Wei} and \cite[Corollary 5.4]{CLWY}).
\item[(2)] When $\mathfrak{n}^- = 1$, we would expect that, after further work, the above generating function becomes the ``mock'' part of a particular metaplectic automorphic form as in the classical case (cf.\ \cite{Coh} and \cite{Zag})
by ``regularizing'' the theta integral $I(\cdot; \varphi^o)$ in Section~\ref{sec: PTI1}.
This will be studied in a subsequent paper.
\end{itemize}
\end{rem}
\subsection{Content}
The contents of this paper go as follows:
\begin{itemize}
\item (\textit{Preliminaries.}) We set up basic notations used throughout this paper in Section~\ref{sec: BS}.
The modified Hurwitz class number and the needed properties are reviewed in Section~\ref{sec: HCN}.
The Tamagawa measures on the groups appearing in this paper are given in Section~\ref{sec: TMk}, \ref{sec: HCN}, and \ref{sec: TMB}, respectively.
The definition of the Weil representation and theta series are recalled in Section~\ref{sec: WRTS}.
\item (\textit{Metaplectic theta series and mass over CM points.}) In Section~\ref{sec: TSP}, we focus on the theta integral associated with a particular Schwartz function $\varphi^o$ on the quadratic space of pure quaternions in an indefinite quaternion algebra over $k$.
The Fourier coefficients of the theta integral associated with $\varphi^o$ are expressed via two different ways in Theorem~\ref{thm: FC1} and Corollary~\ref{cor: AFC}, respectively.
In Section~\ref{sec: CMBES}, we study the mass over the CM points with a given discriminant on a Drinfeld-Stuhler modular curve, and establish
the equality between $\mathcal{M}(d)$ and $H^{\mathfrak{n}^+,\mathfrak{n}^-}(d)$ in Theorem~\ref{thm: CNo-geo}.
\item (\textit{Class number relations and intersections.}) In Section~\ref{sec: TSN},
we take a particular Schwatz function $\varphi_\Lambda$ associated with the $A$-lattice $\Lambda$, and express the Fourier coefficients of the theta integral $I(\cdot;\varphi_\Lambda)$ explicitly in terms of the modified Hurwitz class numbers in Theorem~\ref{thm: TINFC}.
In Section~\ref{sec: Ext-TI}, we show the harmonicity of $I(\cdot;\varphi_\Lambda)$ and extend it to a Drinfeld-type automorphic form $\vartheta_\Lambda$ on $\operatorname{GL}_2(k_\infty)$ in Theorem~\ref{thm: Ext-DTAF}.
On the other hand, the intersection between $\mathcal{Z}_1$ and Hirzebruch-Zagier-type divisors are described in Section~\ref{sec: Intersection}.
Via the alternative expression of the Fourier coefficients of $I(\cdot;\varphi_\Lambda)$ in Theorem~\ref{thm: AFCN}, we prove Theorem~\ref{thm: MT1} and Theorem~\ref{thm: MT2} in the end.
\item (\textit{Appendix: local optimal embeddings.}) The needed results in Eichler's theory of local optimal embeddings are recalled in Appendix~\ref{sec: App-LOE}, and we express the technical local integrals used in Theorem~\ref{thm: FC1} by the number of local optimal embeddings in Appendix~\ref{sec: App-SLI}.
\end{itemize}
\subsection*{Acknowledgements}
This work is initiated during the conference ``2019 Postech-PMI and NCTS Joint workshop on number theory'', at the Pohang University of Science and Technology, Korea.
The authors are very grateful to Professor Jeehoon Park and Professor Chia-Fu Yu for organizing such a wonderful conference. The authors would also like to thank Jing Yu for many helpful suggestions and comments.
The first author is supported by the Ministry of Science and Technology
(grant no.\ 109-2115-M-002-017-MY2).
The second author is supported by the Ministry of Science and Technology
(grant no.\ 107-2628-M-007-004-MY4 and 109-2115-M-007-017-MY5)
and the National Center for Theoretical Sciences.
\section{Preliminaries}
\subsection{Basic settings}\label{sec: BS}
Let ${\mathbb F}_q$ be a finite field with $q$ elements.
Throughout this paper, we always assume $q$ to be \textbf{odd}.
Let $A:={\mathbb F}_q[\theta]$, the polynomial ring with one variable $\theta$ over ${\mathbb F}_q$, and $k:= {\mathbb F}_q(\theta)$, the field of fractions of $A$.
Let $\infty$ be the infinite place of $k$, i.e.\ the place corresponding to the \lq\lq degree\rq\rq\ valuation $\operatorname{ord}_\infty$ defined by
$$\operatorname{ord}_\infty\left(\frac{a}{b}\right) := \deg b- \deg a, \quad \forall a,b \in A \text{ with $b \neq 0$.}
$$
The associated absolute value on $k$ is normalized by $|\alpha|_\infty := q^{-\operatorname{ord}_\infty(\alpha)}$ for every $\alpha \in k$.
Let $k_\infty$ be the completion of $k$ with respect to $|\cdot|_\infty$, which can be identified with the Laurent series field ${\mathbb F}_q(\!(\theta^{-1})\!)$.
Put $\varpi := \theta^{-1}$, a fixed uniformizer at $\infty$, and $O_\infty:= {\mathbb F}_q[\![\varpi]\!]$,
the valuation ring in $k_\infty$.
Let $A_+$ be the set of monic polynomials in $A$.
By abuse of notations, we identify $A_+$ with the set of non-zero ideals of $A$.
In particular, for $\mathfrak{a} \in A_+$ we put
$$\Vert \mathfrak{a} \Vert \ :=\ \#(A/\mathfrak{a}) \quad (= |\mathfrak{a}|_\infty).$$
Given a non-zero prime ideal $\mathfrak{p}$ of $A$,
the normalized absolute value associated with $\mathfrak{p}$ is:
$$|\alpha|_\mathfrak{p} := \Vert\mathfrak{p}\Vert^{-\operatorname{ord}_\mathfrak{p}(\alpha)}, \quad \forall \alpha \in k.
$$
Here $\operatorname{ord}_\mathfrak{p}(\alpha)$ is the order of $\alpha$ at $\mathfrak{p}$ for every $\alpha \in k$.
The completion of $k$ with respect to $|\cdot|_\mathfrak{p}$ is denoted by $k_\mathfrak{p}$, and put $O_\mathfrak{p}$ the valuation ring in $k_\mathfrak{p}$. We also refer the non-zero prime ideals of $A$ to the finite places of $k$. \\
Let $k_\AA:= \prod_v' k_v$, the adele ring of $k$.
The maximal compact subring of $k_\AA$ is denoted by $O_\AA$.
The adelic norm $|\cdot|_\AA$ on the idele group $k_\AA^\times$ is:
$$
|(\alpha_v)_v|_\AA := \prod_v |\alpha_v|_v, \quad \forall (\alpha_v)_v \in k_\AA^\times.
$$
\subsubsection{Additive character and Tamagawa measure}\label{sec: TMk}
Let $p$ be the characteristic of $k$ and $\psi_\infty: k_\infty \rightarrow {\mathbb C}^\times$ be the additive character defined by:
$$
\psi_\infty\left(\sum_i a_i \varpi^i\right) := \exp\left(\frac{2\pi \sqrt{-1}}{p} \cdot \text{Trace}_{\, {\mathbb F}_q/{\mathbb F}_p}(-a_1)\right), \quad \forall \sum_i a_i \varpi^{i} \in k_\infty.
$$
The conductor of $\psi_\infty$ is $\varpi^2 O_\infty$ and $\psi_\infty(A) = 1$.
Since
$$
k_\AA = k + \left(k_\infty \times \prod_{\mathfrak{p} } O_\mathfrak{p}\right)
\quad \text{ and } \quad
k \cap \left(\prod_{\mathfrak{p}} O_\mathfrak{p}\right) = A,
$$
we may extend $\psi_\infty$ uniquely to an additive character $\psi: k_\AA \rightarrow {\mathbb C}^\times$
so that $\psi(\alpha) = 1$ for all $\alpha \in k + \big((\varpi^2 O_\infty) \times \prod_{\mathfrak{p}} O_\mathfrak{p}\big)$ and $\psi\big|_{k_\infty} = \psi_\infty$.
Put $\psi_\mathfrak{p} := \psi\big|_{k_\mathfrak{p}}$ for each finite place $\mathfrak{p}$ of $k$, which is a non-trivial additive character on $k_\mathfrak{p}$ with trivial conductor.\\
For each place $v$ of $k$, let $dx_v$ be the ``self-dual'' Haar measure on $k_v$ with respect to $\psi_v$, i.e.\
$$\text{vol}(O_\mathfrak{p},dx_\mathfrak{p}) = 1 \text{ for each finite place $\mathfrak{p}$ of $k$,} \quad \text{and}
\quad
\text{vol}(O_\infty,dx_\infty) = q.
$$
Define the Haar measure $d^\times x_v$ on $k_v^\times$ by
$$d^\times x_\mathfrak{p}:= \frac{\Vert\mathfrak{p}\Vert}{\Vert \mathfrak{p} \Vert-1} \cdot \frac{dx_\mathfrak{p}}{|x_\mathfrak{p}|_\mathfrak{p}}\quad
\text{ and }\quad
d^\times x_\infty:= \frac{q}{q-1} \cdot \frac{dx_\infty}{|x_\infty|_\infty}.
$$
The \textit{Tamagawa measure} on $k_\AA^\times$ (with respect to $\psi$) is $d^\times x = \prod_v d^\times x_v$.
\subsection{Imaginary quadratic fields and class numbers}\label{sec: HCN}
A quadratic field extension $K/k$ is called \textit{imaginary} if the infinite place of $k$ does not split in $K$.
Let $K_\AA := K\otimes_k k_\AA$ and $\text{T}_{K/k}:= K_\AA \rightarrow k_\AA$ be the trace map.
Then the Tamagawa measure on $K_\AA^\times$ (with respect to the additive character $\psi \circ \text{T}_{K/k}$)
and the one on $k_\AA^\times$ induce a Haar measure $d^\times \alpha$ on the quotient group $K_\AA^\times/k_\AA^\times$.
More precisely,
let $O_K$ (resp.\ $O_{K_\infty}$) be the integral closure of $A$ (resp.\ $O_\infty$) in $K$ (resp.\ $K_\infty:= K\otimes_k k_\infty$).
For each non-zero prime ideal $\mathfrak{p}$ of $A$, put $K_\mathfrak{p} := K \otimes_k k_\mathfrak{p}$ and $O_{K_\mathfrak{p}}:= O_K \otimes_A O_\mathfrak{p}$.
We normalize the Haar measure $d^\times \alpha_v$ on $K_v^\times/k_v^\times$ for each place of $v$ by
$$\text{vol}(O_{K_\mathfrak{p}}^\times/O_\mathfrak{p}^\times)
= \Vert\mathfrak{p}\Vert^{-1+1/e_\mathfrak{p}(K/k)} \quad
\text{ and } \quad
\text{vol}(O_{K_\infty}^\times/O_\infty^\times)
= q^{1/e_\infty(K/k)}.
$$
Here $e_v(K/k)$ is the ramification index of the place $v$ of $k$ in $K$.
Then $d^\times \alpha = \prod_v d^\times \alpha_v$.
\begin{prop}\label{prop: VOL-CN}
Let $K$ be an imaginary quadratic field over $k$, and $O_K$ be the integral closure of $A$ in $K$.
Let $\Delta(O_K/A)$ be the discriminant ideal of $O_K$ over $A$, $h(O_K)$ be the class number of $O_K$, and put
$w(O_K):= \#(O_K^\times)/(q-1)$.
We have
\begin{eqnarray}
\text{\rm vol}(K^\times \backslash K_\AA^\times/k_\AA^\times)
&=& \frac{h(O_K)}{w(O_K)}
\cdot e_\infty(K/k) \cdot \prod_v \text{\rm vol}(O_{K_v}^\times/O_v^\times) \nonumber \\
&=& \frac{h(O_K)}{w(O_K)} \cdot e_\infty(K/k) \cdot q^{1/e_\infty(K/k)}
\cdot \Vert \Delta(O_K/A)\Vert^{-1/2}. \nonumber
\end{eqnarray}
\end{prop}
\begin{proof}
As $A$ is a principal ideal domain, one gets
$$k_\AA^\times = k^\times \cdot (k_\infty^\times \times \prod_{\mathfrak{p}} O_\mathfrak{p}^\times).
$$
Thus the exact sequence
$$
1 \longrightarrow \frac{O_K^\times}{{\mathbb F}_q^\times} \longrightarrow \frac{K_\infty^\times \times \prod_\mathfrak{p} O_{K_\mathfrak{p}}^\times}{k_\infty^\times \times \prod_\mathfrak{p} O_\mathfrak{p}^\times} \rightarrow
\frac{K^\times \cdot (K_\infty^\times \times \prod_\mathfrak{p} O_{K_\mathfrak{p}}^\times)}{K^\times \cdot (k_\infty^\times \times \prod_\mathfrak{p} O_\mathfrak{p}^\times)} \longrightarrow 1
$$
implies
\begin{eqnarray}
\text{vol}(K^\times \backslash K_\AA^\times/k_\AA^\times)
&=&\text{vol}\Big(K^\times \backslash K_\AA^\times/(k_\infty^\times \times \prod_{\mathfrak{p} } O_\mathfrak{p}^\times)\Big) \nonumber \\
&=& \frac{\#\big(K^\times \backslash K_\AA^\times/(K_\infty^\times \times \prod_\mathfrak{p} O_{K_\mathfrak{p}}^\times)\big)}{\#(O_K^\times/{\mathbb F}_q^\times)} \cdot \text{vol}\left(\frac{K_\infty^\times \times \prod_\mathfrak{p} O_{K_\mathfrak{p}}^\times}{k_\infty^\times \times \prod_{\mathfrak{p}} O_\mathfrak{p}^\times}\right).
\nonumber
\end{eqnarray}
The result follows from
$$
\frac{\#\big(K^\times \backslash K_\AA^\times/(K_\infty^\times \times \prod_\mathfrak{p} O_{K_\mathfrak{p}}^\times)\big)}{\#(O_K^\times/{\mathbb F}_q^\times)}
= \frac{h(O_K)}{w(O_K)}
$$
and
\begin{eqnarray}
\text{vol}\left(\frac{K_\infty^\times \times \prod_\mathfrak{p} O_{K_\mathfrak{p}}^\times}{k_\infty^\times \times \prod_{\mathfrak{p}} O_\mathfrak{p}^\times}\right)
&=&
\text{vol}(K_\infty^\times/k_\infty^\times) \cdot \prod_{\mathfrak{p}} \text{vol}(O_{K_\mathfrak{p}}^\times/O_\mathfrak{p}^\times) \nonumber \\
&=&
e_\infty(K/k)\cdot q^{1/e_\infty(K/k)}
\cdot \Vert \Delta(O_K/A)\Vert^{-1/2}.
\nonumber
\end{eqnarray}
\end{proof}
\begin{rem}
Let $\varsigma_K : k^\times\backslash k_\AA^\times \rightarrow \{\pm 1\}$ be the quadratic Hecke character associated with $K/k$, and let $L(s,\varsigma_K)$ be the $L$-function of $\varsigma_K$.
It is known that (cf.\ \cite[Section 2.2]{CWY}, see also \cite[Theorem 5.9]{Ros})
$$
L(1,\varsigma_K)
= \frac{\#(O_K^\times)}{\#({\mathbb F}_q^\times)} \cdot q \cdot \Big(q^{(1-e_\infty(K/k))/2} \cdot \Vert \Delta(O_K/A)\Vert^{-1/2}\Big) \cdot \frac{h(O_K)}{2/e_\infty(K/k)}.
$$
The above proposition says in particular that
$$
\text{vol}(K^\times \backslash K_\AA^\times/k_\AA^\times) = 2 \cdot L(1,\varsigma_K).
$$
\end{rem}
Recall the following fact (cf.\ \cite[Chapter I (12.12) Theorem]{Neu}):
\begin{lem}\label{lem: CN}
For each $A$-order $\mathcal{O}$ in an imaginary quadratic extension $K$ of $k$, let $h(\mathcal{O})$ be the proper ideal class number of $\mathcal{O}$ and $w(\mathcal{O}) := \#(\mathcal{O}^\times)/(q-1)$. Then
$$
\frac{h(\mathcal{O})}{w(\mathcal{O})} = \frac{h(O_K)}{w(O_K)} \cdot
\prod_{\mathfrak{p}} \#\left(\frac{O_{K_\mathfrak{p}}^\times}{\mathcal{O}_\mathfrak{p}^\times}\right).
$$
Here $\mathcal{O}_\mathfrak{p} := \mathcal{O} \otimes_A A_\mathfrak{p}$ for every non-zero prime ideal $\mathfrak{p}$ of $A$.
\end{lem}
We write $d \prec 0$ for $d \in A$ if the quadratic extension $k(\sqrt{d})$ is imaginary over $k$.
Given $d \in A$ with $d \prec 0$,
let $\mathcal{O}_d := A[\sqrt{d}]$, $h(d):=h(\mathcal{O}_d)$, and $w(d) := w(\mathcal{O}_d)$.
\begin{defn}\label{defn: HCN}
For square-free $\mathfrak{n}^+,\mathfrak{n}^- \in A_+$ with $\text{gcd}(\mathfrak{n}^+,\mathfrak{n}^-) = 1$, we recall the following \textit{modified Hurwitz class number}
$$
H^{\mathfrak{n}^+,\mathfrak{n}^-}(d) := \sum_{\subfrac{\mathfrak{c} \in A_+}{\mathfrak{c}^2 \mid d}} \frac{h(d/\mathfrak{c}^2)}{w(d/\mathfrak{c}^2)}\cdot \prod_{\mathfrak{p} \mid \mathfrak{n}^+}\left(1+\left\{\frac{d/\mathfrak{c}^2}{\mathfrak{p}}\right\}\right)\prod_{\mathfrak{p} \mid \mathfrak{n}^-}\left(1-\left\{\frac{d/\mathfrak{c}^2}{\mathfrak{p}}\right\}\right).
$$
Here
$$\left\{\frac{d}{\mathfrak{p}}\right\} :=
\begin{cases}
1, & \text{ if either $\mathfrak{p}$ split in $k(\sqrt{d})$ or $\mathfrak{p}^2 \mid d$;} \\
-1, & \text{ if $\mathfrak{p}$ is inert in $k(\sqrt{d})$ and $\operatorname{ord}_\mathfrak{p}(d) = 0$;} \\
0, & \text{ if $\operatorname{ord}_\mathfrak{p}(d) = 1$.}
\end{cases}
$$
\end{defn}
Write $$d = d_0 \cdot \prod_{\mathfrak{p}} \mathfrak{p}^{2c_{\mathfrak{p}}},$$
where $d_0 \in A$ is square-free
(and $c_\mathfrak{p} = 0$ for almost all irreducible $\mathfrak{p} \in A_+$).
For each irreducible $\mathfrak{p} \in A_+$ and integer $\ell_\mathfrak{p}$ with $0\leq \ell_\mathfrak{p} \leq c_\mathfrak{p}$, put
$$
e^{\mathfrak{n}^+,\mathfrak{n}^-}_\mathfrak{p}(\ell_\mathfrak{p}) :=
\begin{cases}\displaystyle 1 \pm \left\{\frac{d_0 \mathfrak{p}^{2\ell_\mathfrak{p}}}{\mathfrak{p}}\right\},
& \text{ if $\mathfrak{p} \mid \mathfrak{n}^{\pm}$;}\\
1, & \text{ otherwise.}
\end{cases}
$$
We obtain the following expression for the modified Hurwitz class numbers in later use:
\begin{prop}\label{prop: HCN}
Given $d \in A$ with $d \prec 0$, write $d = d_0 \prod_\mathfrak{p} \mathfrak{p}^{2c_\mathfrak{p}}$.
Then
$$
H^{\mathfrak{n}^+,\mathfrak{n}^-}(d)
= \frac{h(d_0)}{w(d_0)} \cdot \prod_\mathfrak{p} \left[
\sum_{0\leq \ell_\mathfrak{p} \leq c_\mathfrak{p}}\#\left(\frac{\mathcal{O}_{d_0,\mathfrak{p}}^\times}{\mathcal{O}_{d_0\mathfrak{p}^{2\ell_\mathfrak{p}},\mathfrak{p}}^\times}\right) \cdot e^{\mathfrak{n}^+,\mathfrak{n}^-}_\mathfrak{p}(\ell_\mathfrak{p})
\right].
$$
\end{prop}
\begin{proof}
For $\boldsymbol{\ell} = (\ell_\mathfrak{p})_\mathfrak{p} \in \prod_\mathfrak{p} {\mathbb Z}$ with $0\leq \ell_\mathfrak{p}\leq c_\mathfrak{p}$,
put $d_0(\boldsymbol{\ell}):= d_0 \prod_\mathfrak{p} \mathfrak{p}^{2\ell_\mathfrak{p}}$.
Then
$$
\left\{\frac{d_0(\boldsymbol{\ell})}{\mathfrak{p}}\right\} = \left\{\frac{d_0\mathfrak{p}^{2\ell_\mathfrak{p}}}{\mathfrak{p}}\right\}
\quad \text{ and } \quad
\mathcal{O}_{d_0(\boldsymbol{\ell}),\mathfrak{p}} = \mathcal{O}_{d_0\mathfrak{p}^{2\ell_\mathfrak{p}}, \mathfrak{p}}.
$$
Therefore
\begin{eqnarray}
H^{\mathfrak{n}^+,\mathfrak{n}^-}(d)
&=& \sum_{\subfrac{\boldsymbol{\ell} \in \prod_\mathfrak{p} {\mathbb Z}}{0\leq \ell_\mathfrak{p} \leq c_\mathfrak{p}}}
\frac{h(d(\boldsymbol{\ell}))}{w(d(\boldsymbol{\ell}))}\cdot
\prod_{\mathfrak{p} \mid \mathfrak{n}^+}\left(1+\left\{\frac{d(\boldsymbol{\ell})}{\mathfrak{p}}\right\}\right) \cdot
\prod_{\mathfrak{p} \mid \mathfrak{n}^-}\left(1-\left\{\frac{d(\boldsymbol{\ell})}{\mathfrak{p}}\right\}\right) \nonumber \\
&=&
\frac{h(d_0)}{w(d_0)} \cdot \sum_{\subfrac{\boldsymbol{\ell} \in \prod_\mathfrak{p} {\mathbb Z}}{0\leq \ell_\mathfrak{p} \leq c_\mathfrak{p}}}
\left[\prod_\mathfrak{p} \#\left(\frac{O_{d_0,\mathfrak{p}}^\times}{\mathcal{O}_{d_0\mathfrak{p}^{2\ell_\mathfrak{p}},\mathfrak{p}}^\times}\right) \cdot \prod_\mathfrak{p} e_\mathfrak{p}^{\mathfrak{n}^+,\mathfrak{n}^-}(\ell_\mathfrak{p})\right] \nonumber \\
&=&
\frac{h(d_0)}{w(d_0)} \cdot
\prod_\mathfrak{p} \left[
\sum_{0\leq \ell_\mathfrak{p} \leq c_\mathfrak{p}}\#\left(\frac{\mathcal{O}_{d_0,\mathfrak{p}}^\times}{\mathcal{O}_{d_0\mathfrak{p}^{2\ell_\mathfrak{p}},\mathfrak{p}}^\times}\right) \cdot e^{\mathfrak{n}^+,\mathfrak{n}^-}_\mathfrak{p}(\ell_\mathfrak{p})
\right]. \nonumber
\end{eqnarray}
\end{proof}
\begin{rem}\label{rem: H(0)}
For convention, we put
$$
H^{\mathfrak{n}^+,\mathfrak{n}^-}(0)
:=
-\frac{1}
{q^2-1} \cdot \prod_{\mathfrak{p}\mid \mathfrak{n}^+}(\Vert \mathfrak{p} \Vert + 1) \prod_{\mathfrak{p} \mid \mathfrak{n}^-}(\Vert \mathfrak{p} \Vert -1).
$$
This is related to a volume quantity with respect to the ``Tamagawa measure'' on quaternion algebras discussed below.
\end{rem}
\subsection{Tamagawa measure on quaternion algebras}\label{sec: TMB}
Let $B$ be an \textit{indefinite} quaternion algebra over $k$ (i.e.\ $B_\infty := B\otimes_k k_\infty$ is not division).
Put $B_\AA := B\otimes_k k_\AA$.
Let $\operatorname{Tr}: B_\AA \rightarrow k_\AA$ be the reduced trace map.
Choose a Haar measure $db = \prod_v db_v$ on $B_\AA$ which is self-dual with respect to the additive character $\psi \circ \operatorname{Tr}$.
More precisely, for each non-zero prime ideal $\mathfrak{p}$ of $A$,
let $R_\mathfrak{p}$ be a maximal $O_\mathfrak{p}$-order in $B_\mathfrak{p}:= B\otimes_k k_\mathfrak{p}$.
Then
$$
\text{vol}(R_\mathfrak{p}, db_\mathfrak{p})
= \begin{cases} 1/ \Vert \mathfrak{p} \Vert, & \text{ if $B$ is ramified at $\mathfrak{p}$;}\nonumber \\
1, & \text{ otherwise.}
\end{cases}
$$
Let $O_{B_\infty}$ be a maximal $O_\infty$-order in $B_\infty$.
Then
$\text{vol}(O_{B_\infty},db_\infty) = q^4$.
Let $\operatorname{Nr} : B_\AA^\times \rightarrow k_\AA^\times$ be the reduced norm map.
For each non-zero prime ideal $\mathfrak{p}$ of $A$,
we take the Haar measure $d^\times b_\mathfrak{p}$ on $B_\mathfrak{p}^\times$ defined by
$$
d^\times b_\mathfrak{p}
:= \frac{\Vert \mathfrak{p} \Vert}{\Vert \mathfrak{p} \Vert -1}\cdot \frac{db_\mathfrak{p}}{|\operatorname{Nr}(b_\mathfrak{p})|_\mathfrak{p}}.
$$
In particular,
$$
\text{vol}(R_\mathfrak{p}^\times, d^\times b_\mathfrak{p}) =
\left(1-\frac{1}{\Vert \mathfrak{p} \Vert^{2}}\right) \cdot
\begin{cases}
1/(\Vert \mathfrak{p} \Vert -1), & \text{ if $B$ is ramified at $\mathfrak{p}$;} \\
1, & \text{ otherwise.}
\end{cases}
$$
Similarly, put
$$d^\times b_\infty := \frac{q}{q-1} \cdot \frac{db_\infty}{|\operatorname{Nr}(b_\infty)|_\infty}.
$$
Then
$$
\text{vol}(O_{B_\infty}^\times, d^\times b_\infty)
= q^4-q^2.
$$
The \textit{Tamagawa measure} $d^\times b$ on $B_\AA^\times$ is the Haar measure satisfying that for every compact open subgroup $\mathcal{K} = \prod_v \mathcal{K}_v$ of $B_\AA^\times$, one has
$$
\text{vol}(\mathcal{K}, d^\times b)
= \prod_v \text{vol}(\mathcal{K}_v, d^\times b_v).
$$
Let $\mathfrak{n}^- \in A_+$ be the product of the primes at which $B$ is ramified, and
$\mathfrak{n}^+ \in A_+$ be a square-free polynomial coprime to $\mathfrak{n}^-$.
Let $O_B$ be an Eichler $A$-order of type $(\mathfrak{n}^+,\mathfrak{n}^-)$ in $B$,
and put $O_{B_\mathfrak{p}} := O_B \otimes_A O_\mathfrak{p}$ for non-zero prime $\mathfrak{p}$ of $A$.
Let $O_{B_\AA} := \prod_v O_{B_v}$.
Then:
\begin{lem}\label{lem: vol-hcn}
The Tamagawa measures on $B_\AA^\times$ and $k_\AA^\times$ induces a Haar measure on $B_\AA^\times/k_\AA^\times$ so that
$$
\text{\rm vol}(O_{B_\AA}^\times/O_\AA^\times) = \frac{(q-1)(q^2-1)}{\prod_{\mathfrak{p} \mid \mathfrak{n}^+}(\Vert \mathfrak{p} \Vert + 1) \prod_{\mathfrak{p} \mid \mathfrak{n}^-}(\Vert \mathfrak{p} \Vert -1)}
= -\frac{q-1}{H^{\mathfrak{n}^+,\mathfrak{n}^-}(0)}.
$$
\end{lem}
\begin{proof}
For each non-zero prime ideal $\mathfrak{p}$ of $A$, let $R_\mathfrak{p}$ be a maximal $O_\mathfrak{p}$-order containing $O_{B_\mathfrak{p}}$.
As $\mathfrak{n}^+$ is square-free, we have
$$
\#(R_\mathfrak{p}^\times/O_{B_\mathfrak{p}}^\times) =
\begin{cases}
\Vert \mathfrak{p} \Vert + 1, & \text{ if $\mathfrak{p} \mid \mathfrak{n}^+$;}\\
1, & \text{ otherwise.}
\end{cases}
$$
Thus
$$
\text{vol}(O_{B_\mathfrak{p}}^\times)
= \left(1-\frac{1}{\Vert \mathfrak{p} \Vert^2}\right)\cdot
\begin{cases}
1/(\Vert \mathfrak{p} \Vert -1), & \text{ if $\mathfrak{p} \mid \mathfrak{n}^-$;} \\
1/(\Vert \mathfrak{p} \Vert + 1), & \text{ if $\mathfrak{p} \mid \mathfrak{n}^+$;}\\
1, & \text{ otherwise.}
\end{cases}
$$
Notice that
$$
\prod_\mathfrak{p}\left(1-\frac{1}{\Vert \mathfrak{p} \Vert^s}\right)^{-1}
= \frac{1}{1-q^{1-s}}, \quad \operatorname{Re}(s)>1.
$$
Therefore we obtain
$$\prod_\mathfrak{p}\left(1-\frac{1}{\Vert \mathfrak{p} \Vert^2}\right) = 1-q^{1-2} = \frac{q-1}{q},
$$
and
\begin{eqnarray}
\text{vol}(O_{B_\AA}^\times/O_\AA^\times)
&=& \frac{\text{vol}(O_{B_\infty}^\times)}{\text{vol}(O_\infty^\times)} \cdot \prod_\mathfrak{p}\frac{\text{vol}(O_{B_\mathfrak{p}}^\times)}{\text{vol}(O_\mathfrak{p}^\times)} \nonumber \\
&=& \frac{q^4-q^2}{q} \cdot \prod_\mathfrak{p}\left(1-\frac{1}{\Vert \mathfrak{p} \Vert^2}\right) \prod_{\mathfrak{p} \mid \mathfrak{n}^+}\left(\frac{1}{\Vert \mathfrak{p} \Vert + 1}\right) \prod_{\mathfrak{p} \mid \mathfrak{n}^-}\left(\frac{1}{\Vert \mathfrak{p} \Vert -1}\right) \nonumber \\
&=&
\frac{(q-1)(q^2-1)}{\prod_{\mathfrak{p} \mid \mathfrak{n}^+}(\Vert \mathfrak{p} \Vert + 1) \prod_{\mathfrak{p} \mid \mathfrak{n}^-}(\Vert \mathfrak{p} \Vert -1)} \nonumber \\
&=& - \frac{q-1}{H^{\mathfrak{n}^+,\mathfrak{n}^-}(0)}. \nonumber
\end{eqnarray}
The last equality follows directly from the definition of $H^{\mathfrak{n}^+,\mathfrak{n}^-}(0)$.
\end{proof}
\begin{rem}
The Haar measure on $B_\AA^\times/k_\AA^\times$ induced by the Tamagawa measures on $B_\AA^\times$ and on $k_\AA^\times$ satisfies (cf.\ \cite[Theorem 3.3.1]{Weil2})
$$
\text{vol}(B^\times \backslash B_\AA^\times/k_\AA^\times)
= 2.
$$
\end{rem}
\subsection{Weil representation and theta series}\label{sec: WRTS}
Given a place $v$ of $k$,
let $\widetilde{\operatorname{SL}}_2(k_v)$ be the metaplectic cover of $\operatorname{SL}_2(k_v)$, i.e.\ the non-trivial central externsion of $\operatorname{SL}_2(k_v)$ by $\{\pm 1\}$ associated with the following ``variant'' Kubota $2$-cocycle:
$$\sigma_v^1(g,g'):= \left(\frac{X(gg')}{X(g)}, \frac{X(gg')}{X(g')}\right)_v \cdot \frac{s_v^1(g)s_v^1(g')}{s_v^1(gg')},$$
where
\begin{itemize}
\item $(\cdot,\cdot)_v$ is the (quadratic) Hilbert symbol;
\item
$$
X\begin{pmatrix} a&b \\c & d \end{pmatrix}:=
\begin{cases} c & \text{ if $c \neq 0$;}\\
d & \text{ if $c = 0$;}
\end{cases}
$$
\item
$$
s_v^1\begin{pmatrix}a&b \\ c&d\end{pmatrix} :=
\begin{cases}
(c,d)_v, & \text{ if $\operatorname{ord}_v(c)$ is odd and $d \neq 0$;}\\
1, & \text{ otherwise.}
\end{cases}
$$
\end{itemize}
Identifying $\widetilde{\operatorname{SL}}_2(k_v)$ with $\operatorname{SL}_2(k_v) \times \{\pm 1\}$ as sets, the group law of $\widetilde{\operatorname{SL}}_2(k_v)$ is given by
$$(g_1,\xi_1)\cdot (g_2,\xi_2) := \big(g_1g_2, \xi_1\xi_2 \sigma_v^1(g_1,g_2)\big),\quad
\forall (g_1,\xi_1), (g_2,\xi_2) \in \widetilde{\operatorname{SL}}_2(k_v).
$$
It is known that (cf.\ \cite[Section 2.3]{Gel}) the inclusion map $\operatorname{SL}_2(O_v)\hookrightarrow \widetilde{\operatorname{SL}}_2(k_v)$ defined by
$(\kappa_v \mapsto \tilde{\kappa}_v := (\kappa_v, 1))$
is actually a group homomorphism.
Set
$$\sigma^1(g,g'):= \prod_v(g_v,g'_v), \quad \forall g = (g_v)_v, g' = (g_v')_v \in \operatorname{SL}_2(k_\AA),$$
which is a well-defined $2$-cocycle of $\operatorname{SL}_2(k_\AA)$.
This induces a non-trivial central extension $\widetilde{\operatorname{SL}}_2(k_\AA)$ of $\operatorname{SL}_2(k_\AA)$ by $\{\pm 1\}$, called the metaplectic cover of $\operatorname{SL}_2(k_\AA)$.
Moreover, the inclusions $\operatorname{SL}_2(k)\hookrightarrow \widetilde{\operatorname{SL}}_2(k_\AA)$ and $\operatorname{SL}_2(O_\AA)\hookrightarrow \widetilde{\operatorname{SL}}_2(k_\AA)$ defined by
$$\gamma \longmapsto \tilde{\gamma}:=\big(\gamma, \prod_vs_v(\gamma)\big), \quad \forall \gamma \in \operatorname{SL}_2(k)
\quad \text{ and } \quad \kappa \longmapsto \tilde{\kappa}:= (\kappa,1), \quad \forall \kappa \in \operatorname{SL}_2(O_\AA)
$$
are group homomorphisms.\\
Let $(V,Q_V)$ be a non-degenerated quadratic space over $k$ with $\dim_k(V) = n$.
For each place $v$ of $k$, let $V(k_v) := V\otimes_k k_v$ and $S(V(k_v))$ be the space of Schwartz function on $V(k_v)$.
The {\it Weil representation} $\omega_v^V$ of $\widetilde{\operatorname{SL}}_2(k_v) \times \operatorname{O}(V)(k_v)$ on $S(V(k_v))$, where $\operatorname{O}(V)$ is the orthogonal group of $V$, is given by (cf.\ \cite[Theorem 2.22]{Gel}):
\begin{eqnarray}
(1) && \omega^{V}_v(h) \phi(x)= \phi(h^{-1} x), \text{ $h \in \operatorname{O}(V)(k_v)$}, \phi\in S(V(k_v)); \nonumber \\
(2) && \omega^{V}_v(1,\xi) \phi(x)= \xi^n \cdot \phi(x), \text{ $\xi \in \{\pm 1\}$}; \nonumber \\
(3) && \omega^{V}_v \left(\begin{pmatrix}1&u\\0&1\end{pmatrix},1\right)\phi(x) = \psi_v(u Q_{V}(x))\cdot \phi(x), \text{ $u \in k_v$};\nonumber \\
(4) && \omega^{V}_v \left(\begin{pmatrix}a_v&0\\0&a_v^{-1}\end{pmatrix},1\right)\phi(x)=|a_v|_v^{\frac{n}{2}} \cdot (a_v,a_v)^n_v\cdot \frac{\varepsilon^{V}_v(a_v)}{\varepsilon^{V}_v(1)} \cdot \phi(a_v x), \text{ $a_v \in k_v^{\times}$;} \nonumber \\
(5) && \omega^{V}_v\left(\begin{pmatrix}0&1\\-1&0\end{pmatrix},1\right)\phi(x)= \varepsilon^{V}_v(1) \cdot \widehat{\phi}(x). \nonumber
\end{eqnarray}
Here:
\begin{itemize}
\item
$\varepsilon^V_v(a_v)$ is the following \textit{Weil index}:
$$\varepsilon^{V}_v(a_v):= \int_{L_{v}} \psi_v(a_v Q_{V}(x))d_{a_v}x, \quad \forall a_v \in k_v^{\times},$$
where $L_{v}$ is a sufficiently large $O_v$-lattice in $V(k_v)$, and the Haar measure $d_{a_v}x$ is self-dual with respect to the pairing
$$(x,y)\mapsto \psi_v(a_v \cdot \langle x, y \rangle_V),\quad \forall x,y \in V(k_v),$$
with $\langle \cdot,\cdot \rangle_V$ the bilinear form on $V$ associated with $Q_V$;
\item $\widehat{\phi}(x)$ is the Fourier transform of $\phi$ (with respect to the self-dual Haar measure):
$$\widehat{\phi}(x) :=\int_{V(k_v)} \phi(y)\psi_v(\langle x, y\rangle_V)dy.$$
\end{itemize}
The (global) Weil representation of $\widetilde{\operatorname{SL}}_2(k_\AA)\times \operatorname{O}(V)(k_\AA)$ on the Schwartz space $S(V(k_\AA))$, where $V(k_\AA) := V\otimes_k k_\AA$, is $\omega^V:= \otimes_v \omega^V_v$.
\begin{rem}\label{rem: sections}
${}$
\begin{itemize}
\item[(1)]
Set $V_0 = k$ and $Q_0(x):= x^2$ for $x \in V_0$.
For convention we put
$$\varepsilon_v(a_v) := \varepsilon_v^{V_0}(a_v), \quad \forall a_v \in k_v^\times.$$
It is known that for $a_v,a_v' \in k_v^\times$ one has
$$\frac{\varepsilon_v(a_v)\varepsilon_v(a_v')}{\varepsilon_v(a_va_v')\varepsilon_v(1)} = (a_v,a_v')_v.$$
\item[(2)] Let $(V,Q_V)$ be a non-degenerate quadratic space over $k$ with $\dim_k(V) = n$.
Choose an orthogonal basis $\{x_1,...,x_n\}$ of $V$ and let $\alpha_i = Q_V(x_i)$, $i=1,...,n$.
Then for every place $v$ of $k$ and $a_v \in k_v^\times$, one has
$$\varepsilon_v^V(a_v) = \prod_{i=1}^n \varepsilon_v(a_v \alpha_i).$$
\item[(3)] For $y = (y_v)_v \in k_\AA^\times$, let
$$
\varepsilon^V(y) := \prod_v \varepsilon_v^V(y_v).
$$
The Weil reciprocity says (cf.\ \cite[Proposition 5]{Weil3}):
$$ \varepsilon^V(\alpha) = 1, \quad \forall \alpha \in k^\times.$$
\item[(3)] When $\dim_k(V)$ is even, the local and global Weil representations actually factor through $\operatorname{SL}_2(k_v)$ and $\operatorname{SL}_2(k_\AA)$, respectively.
\end{itemize}
\end{rem}
Given $\varphi \in S(V(k_\AA))$, the \textit{theta series associated with $\varphi$} is:
$$
\Theta(\tilde{g},h;\varphi):= \sum_{x \in V}\big(\omega^V(\tilde{g},h)\varphi\big)(x), \quad \forall (\tilde{g},h) \in \widetilde{\operatorname{SL}}_2(k_\AA)\times \operatorname{O}(V)(k_\AA).
$$
For every $\gamma \in \operatorname{SL}_2(k)$, $\tilde{g} \in \widetilde{\operatorname{SL}}_2(k_\AA)$, and $\varphi \in S(V(k_\AA))$ we have
$$
\Theta(\tilde{\gamma}\tilde{g};\varphi) = \Theta(\tilde{g};\varphi).
$$
Given $a \in k$ and $y \in k_\AA^\times$,
let:
$$
\Theta^*(a,y;h;\varphi) := \int_{k\backslash k_\AA} \Theta\left(\Bigg(\begin{pmatrix}y&uy^{-1}\\0&y^{-1}\end{pmatrix},1\Bigg) ,h;\varphi\right)\psi(-au)du,
$$
where the Haar measure $du$ is normalized so that $\text{vol}(k\backslash k_\AA,du) = 1$.
For $u \in k_\AA$, one has the following Fourier expansion
$$
\Theta\left(\Bigg(\begin{pmatrix}y&uy^{-1}\\0&y^{-1}\end{pmatrix},1\Bigg),h;\varphi\right)
= \sum_{a \in k}\Theta^*(a,y;h;\varphi) \psi(au).
$$
We shall focus on particular quadratic spaces with degree $3$ and $4$, and study the Fourier coefficients of the theta integrals associated with special Schwartz functions.
\section{The case of pure quaternions}\label{sec: TSP}
Let $B$ be an indefinite division quaternion algebra over $k$, and $B^o$ be the subspace of pure quaternions in $B$, i.e.\
$$B^o:= \{b \in B \mid \operatorname{Tr}(b) = 0\}.$$
Let $$(V^o, Q^o):= (B^o, \operatorname{Nr}),$$
where $\operatorname{Nr}:B \rightarrow k$ is the reduced norm map.
Then $V^o$ is an anisotropic quadratic space of degree $3$ over $k$.
Note that we have the following exact sequence:
$$
1\longrightarrow k^\times \longrightarrow B^\times \longrightarrow \text{SO}(V^o)\longrightarrow 1,
$$
where the map from $B^\times$ to $\text{SO}(V^o)$ is defined by
$$b \longmapsto h_b:= (x \mapsto bxb^{-1}), \quad \forall b \in B^\times.$$
Thus the chosen Haar measure on $B_\AA^\times/k_\AA^\times$ in Section~\ref{sec: TMB} gives us a Haar measure $dh$ on $\text{SO}(V^o)(k_\AA)$ satisfying
$$\text{vol}(\text{SO}(V^o)(k)\backslash \text{SO}(V^o)(k_\AA),dh) = \text{vol}(B^\times \backslash B_\AA^\times/k_\AA^\times, d^\times b) = 2.$$
For $\varphi \in S(V^o(k_\AA))$, we are interested in the following theta integral:
$$
I(\tilde{g};\varphi):= \int_{\text{SO}(V^o)(k)\backslash \text{SO}(V^o)(\AA)} \Theta(\tilde{g},h;\varphi) dh,
\quad \forall \tilde{g} \in \widetilde{\operatorname{SL}}_2(k_\AA),
$$
Given $a \in k$, the $a$-th Fourier coefficients of $I(\cdot;\varphi)$ is:
$$
I^*(a,y;\varphi):=
\int_{k\backslash k_\AA}I\left(\bigg(\begin{pmatrix} y & uy^{-1} \\ 0 & y^{-1} \end{pmatrix},1\bigg);\varphi\right) \psi(-a u) du.
$$
\begin{lem}\label{lem: FCo}
For $a \in k$ and $y \in k_\AA^\times$,
we have $I^*(a,y;\varphi) = 0$ unless there exists $x \in B^o$ such that $x^2 = -a$.
In this case,
\begin{eqnarray}
I^*(a,y;\varphi)
&=&
|y|_\AA^{3/2} \cdot (y,y)_\AA \cdot \varepsilon^{V^o}(y) \cdot \text{\rm vol}(K_{x}^\times \backslash K_{x,\AA}^\times/ k_\AA^\times) \cdot \int_{K_{x,\AA}^\times \backslash B_\AA^\times} \varphi(y b^{-1} x b) d^\times b. \nonumber
\end{eqnarray}
Here $(y,y)_\AA := \prod_v(y_v,y_v)_v$ for $y = (y_v)_v \in k_\AA^\times$, $K_{x}$ is the centralizer of $x$ in $B$, and $K_{x,\AA} := K_x \otimes_k k_\AA$.
\end{lem}
\begin{proof}
By definition, we get
\begin{eqnarray}
&& I^*(a,y;\varphi) \nonumber \\
&=& \int_{k\backslash k_\AA}
\left[\int_{B^\times \backslash B_\AA^\times/ k_\AA^\times} \Theta\left(\bigg(\begin{pmatrix} y & uy^{-1} \\ 0 & y^{-1} \end{pmatrix},1\bigg);\varphi\right) d^\times b\right] \psi(-au) du \nonumber \\
&=& \int_{B^\times \backslash B_\AA^\times/ k_\AA^\times}
\left[
\int_{k\backslash k_\AA} \left( \sum_{x \in V^o} \bigg(\omega^{V^o}\bigg(\begin{pmatrix} y & uy^{-1} \\ 0 & y^{-1}\end{pmatrix},1\bigg) \varphi\bigg)(b^{-1}xb)\right) \psi(-au) du\right] d^\times b. \nonumber
\end{eqnarray}
For $x \in V^o$ and $b \in B_\AA^\times$, it is straightforward that
$$
\Bigg(\omega^{V^o}\bigg(\begin{pmatrix} y & uy^{-1} \\ 0 & y^{-1}\end{pmatrix},1\bigg) \varphi\Bigg)(b^{-1}xb)
= \psi(uQ^o(x)) \cdot |y|_\AA^{3/2} \cdot (y,y)_\AA \cdot \varepsilon^{V^o}(y) \cdot \varphi(y \cdot b^{-1}xb).
$$
Since
$$
\int_{k\backslash k_\AA} \psi\big(uQ^o(x)\big) \cdot \psi(-au) du
= \begin{cases}
1, & \text{ if $Q^o(x) = a$;}\\
0, & \text{ otherwise,}
\end{cases}
$$
we have
\begin{eqnarray}
&& \int_{k\backslash k_\AA}
\left( \sum_{x \in V^o} \bigg(\omega^{V^o}\bigg(\begin{pmatrix} y & uy^{-1} \\ 0 & y^{-1}\end{pmatrix},1\bigg) \varphi\bigg)(b^{-1}xb)\right) \psi(-au) du \nonumber \\
&=& |y|_\AA^{3/2} \cdot (y,y)_\AA \cdot \varepsilon^{V^o}(y) \cdot \sum_{x \in V_a^o} \varphi(y \cdot b^{-1}xb), \nonumber
\end{eqnarray}
where $V_a^o := \{x \in V^o \mid Q^o(x) = -x^2 = a\}$.
Therefore,
$$
I^*(a,y;\varphi) =
|y|_\AA^{3/2} \cdot (y,y)_\AA \cdot \varepsilon^{V^o}(y) \cdot \sum_{x \in V_a^o} \int_{B^\times \backslash B_\AA^\times/ k_\AA^\times} \varphi(y\cdot b^{-1}xb) d^\times b.
$$
In particular, $V_a^o$ is empty unless $a = 0$ or $-a \notin (k^\times)^2$ and $k(\sqrt{a})$ embeds into $B$.
This shows the first assertion of this lemma.
Suppose $V_a^o$ is non-empty, say $x \in V_a^o$.
Note that the conjugation action of $B^\times$ on $V_a^o$ is transitive, and the stablizer of $x$ in $B^\times$ is exactly $K_x^\times$.
Hence
\begin{eqnarray}
I^*(a,y;\varphi) &=&
|y|_\AA^{3/2} \cdot (y,y)_\AA \cdot \varepsilon^{V^o}(y) \cdot
\int_{K_x^\times \backslash B_\AA^\times/ k_\AA^\times}\varphi(y\cdot b^{-1}xb) d^\times b \nonumber \\
&=& |y|_\AA^{3/2} \cdot (y,y)_\AA \cdot \varepsilon^{V^o}(y) \cdot
\text{vol}(K_x^\times \backslash K_{x,\AA}^\times/ k_\AA^\times)
\cdot \int_{K_{x,\AA}^\times \backslash B_\AA^\times} \varphi(y\cdot b^{-1}xb) d^\times b, \nonumber
\end{eqnarray}
and the proof is complete.
\end{proof}
Note that for non-zero $x \in B^o$, we have $K_x = k(x)$, the quadratic field generated by $x$ in $B$.
In particular, $I^*(a,y;\varphi) = 0$ unless $a = 0$ or $-a \notin (k^\times)^2$
and the quadratic field $k(\sqrt{-a})$ embeds into $B$.
\begin{rem}
Suppose $\varphi$ is a pure-tensor, i.e.\ $\varphi = \otimes_v \varphi_v$, where $\varphi_v \in S(V^o(k_v))$.
Then for $x \in V^o$, the following equality holds:
$$
\int_{K_{x,\AA}^\times \backslash B_\AA^\times} \varphi(yb^{-1}xb) d^\times b
= \prod_v \int_{K_{x,v}^\times \backslash B_v^\times} \varphi_v(y_vb_v^{-1}x b_v) d^\times b_v.
$$
We shall choose a particular pure-tensor Schwartz function $\varphi^o = \otimes_v \varphi^o_v \in S(V^o(k_\AA))$ so that the associated Fourier coefficients can be expressed in terms of modified Hurwitz class numbers.
\end{rem}
\subsection{Particular theta integral}\label{sec: PTI1}
Let $\mathfrak{n}^- \in A_+$ be the product of the primes at which $B$ is ramified, and take a square free $\mathfrak{n}^+ \in A_+$ coprime to $\mathfrak{n}^-$.
Let $O_B$ be an Eichler $A$-order of type $(\mathfrak{n}^+,\mathfrak{n}^-)$ in $B$, and take $\Lambda^o:= O_B \cap B^o$.
For each non-zero prime $\mathfrak{p}$ of $A$, take
$\varphi^o_v := \mathbf{1}_{\Lambda^o_\mathfrak{p}}$, the characteristic function of $\Lambda^o_\mathfrak{p} := \Lambda^o\otimes_A O_\mathfrak{p}$.
To choose $\varphi^o_\infty$, we first identify $B_\infty:= B\otimes_k k_\infty$ with $\operatorname{Mat}_2(k_\infty)$, and take $O_{B_\infty}:= \operatorname{Mat}_2(O_\infty)$.
Let $L_\infty := \pi_\infty O_{B_\infty}$ and
$$L_\infty':= \left\{\begin{pmatrix} a&b \\ c&d \end{pmatrix} \in L_\infty \ \bigg|\ c \in \pi_\infty^2 O_\infty\right\}.$$
Put $L_\infty^o := L_\infty \cap B^o$ and $L_\infty^{\prime,o}:= L_\infty' \cap B^o$.
We take
$$\varphi_\infty^o := \mathbf{1}_{L_\infty^o} - \frac{q+1}{2} \cdot \mathbf{1}_{L_\infty^{\prime,o}}.$$
\begin{prop}\label{prop: TLo}
Let $\mathfrak{n} := \mathfrak{n}^+\mathfrak{n}^-$ and
$$
\mathcal{K}_0^{1}(\mathfrak{n}\infty) := \left\{ \begin{pmatrix} a&b \\c&d \end{pmatrix} \in \operatorname{SL}_2(O_\AA)\ \bigg|\ c \equiv 0 \bmod \mathfrak{n}\infty\right\}.
$$
Then for $\gamma \in \operatorname{SL}_2(k)$, $\tilde{g} \in \widetilde{\operatorname{SL}_2}(k_\AA)$, and $\kappa \in \mathcal{K}_0^{1}(\mathfrak{n} \infty)$, we have
$$I(\tilde{\gamma} \tilde{g} \tilde{\kappa}; \varphi^o) = I(\tilde{g};\varphi^o).$$
Here $\tilde{\gamma}$ for $\gamma \in \operatorname{SL}_2(k)$ and $\tilde{\kappa}$ for $\kappa \in \operatorname{SL}_2(O_\AA)$ are introduced in {\rm \textit{Remark}~\ref{rem: sections}}.
\end{prop}
\begin{proof}
It suffices to show that
$\omega^{V^o}(\kappa) \varphi^o = \varphi^o$
for every $\kappa = (\kappa_v)_v \in \mathcal{K}_0^1(\mathfrak{n}\infty)$, i.e.\
\begin{eqnarray}\label{eqn: Inv}
\omega_v^{V^o}(\kappa_v) \varphi^o_v = \varphi^o_v \quad \text{ for every place $v$ of $k$}.
\end{eqnarray}
Given a place $v$ of $k$, notice that $\operatorname{SL}_2(O_v)$ is generated by $\begin{pmatrix} 0 & 1 \\ -1&0\end{pmatrix}$ and $\begin{pmatrix}1& u \\ 0&1\end{pmatrix}$, $u \in O_v$, and for $\begin{pmatrix} a&b \\ c&d \end{pmatrix} \in \operatorname{SL}_2(O_v)$ with $c \equiv 0 \bmod v$, one has $d \in O_v^\times$ and
$$
\begin{pmatrix}
a&b \\ c&d
\end{pmatrix}
=
\begin{pmatrix}
1 & bd^{-1} \\ 0 & 1
\end{pmatrix}
\begin{pmatrix}
d^{-1} & 0 \\ 0 & d
\end{pmatrix}
\begin{pmatrix}
0&1 \\ -1&0
\end{pmatrix}
\begin{pmatrix}
1 & -d^{-1} c \\ 0&1
\end{pmatrix}
\begin{pmatrix}
0&-1 \\ 1&0
\end{pmatrix}
.
$$
The equality~\eqref{eqn: Inv} then follows from straightforward calculations.
\end{proof}
The above invariant property of $I(\cdot;\varphi^o)$ implies in particular that for every $\alpha \in k^\times$, $\varepsilon \in O_\AA^\times$, and $u \in O_\AA$, the Fourier coefficient $I^*(a,y;\varphi^o)$ satisfies
$$
I^*(\alpha^{-2}a,\alpha y \varepsilon;\varphi^o) = (y, \alpha \varepsilon)_\AA \cdot I^*(a,y;\varphi^o)
\quad \text{ and } \quad
I^*(a,y;\varphi^o) = I^*(a,y;\varphi^o)\cdot \psi(a y^2 u).
$$
Since
$$k_\AA^\times = k^\times \cdot (k_\infty^\times \times \prod_{\mathfrak{p}\neq \infty} O_\mathfrak{p}^\times),$$
it suffices to consider $I^*(a,y;\varphi^o)$ for $y \in k_\infty^\times$, and in this case we get
$$I^*(a,y;\varphi^o) = 0 \quad \text{ unless
$a \in A$ with $\deg a + 2 \leq 2 \operatorname{ord}_\infty(y)$.}
$$
Moreover, we may express $I^*(a,y;\varphi^o)$ via modified Hurwitz class numbers in the following:
\begin{thm}\label{thm: FC1}
Given $a \in A$ and $y \in k_\infty^\times$ with $\deg a + 2 \leq 2\operatorname{ord}_\infty(y)$, we have
$$
I^*(a,y;\varphi^o) =
\text{\rm vol}(O_{B_\AA}^\times/O_\AA^\times) \cdot \tilde{\beta}(y) \cdot \begin{cases}
H^{\mathfrak{n}^+,\mathfrak{n}^-}(-a), & \text{ if $-a\preceq 0$;}\\
0, & \text{ otherwise.}
\end{cases}
$$
Here
$$
\tilde{\beta}(y) := |y|_\infty^{3/2}\cdot
\frac{\varepsilon_\infty(y)}{\varepsilon_\infty(1)}, \quad \forall y \in k_\infty^\times.
$$
\end{thm}
\begin{proof}
Notice that for each place $v$ of $k$ and $y_v \in k_v^\times$, one has
$$\varepsilon_v(y_v)\cdot \varepsilon_v^{V^o}(y_v)
= \begin{cases}
-1 & \text{ if $B$ is ramified at $v$;} \\
1 & \text{ otherwise.}
\end{cases}
$$
Thus the product formula of the Weil index in \textit{Remark}~\ref{rem: sections} (3) shows that for $y \in k_\infty^\times$,
$$|y|_\AA^{3/2}\cdot (y,y)_\AA \cdot \varepsilon^{V^o}(y) = |y|_\infty^{3/2} \cdot (y,y)_\infty \cdot \frac{\varepsilon_\infty(1)}{\varepsilon_\infty(y)} = \tilde{\beta}(y).
$$
If $a = 0$, then $K_x = B$, and Lemma~\ref{lem: FCo} implies
\begin{eqnarray}
I^*(0,y;\varphi^o)
&=& \tilde{\beta}(y) \cdot \text{vol}(B^\times k_\AA^\times\backslash B_\AA^\times) \cdot \varphi^o(0) \nonumber \\
&=& \tilde{\beta}(y) \cdot 2 \cdot \frac{1-q}{2} \nonumber \\
&=& \text{\rm vol}(O_{B_\AA}^\times/O_\AA^\times) \cdot \tilde{\beta}(y) \cdot H^{\mathfrak{n}^+,\mathfrak{n}^-}(0). \nonumber
\end{eqnarray}
The last equality follows from Lemma~\ref{lem: vol-hcn}.\\
Suppose $a \neq 0$.
From Lemma~\ref{lem: FCo} we get
\begin{eqnarray}
I^*(a,y;\varphi^o)
&=& \tilde{\beta}(y) \cdot \text{vol}(K_x^\times \backslash K_{x,\AA}^\times/ k_\AA^\times) \cdot
\int_{K_{x,\infty}^\times \backslash B_\infty^\times} \varphi_\infty^o(y b_\infty^{-1}xb_\infty)d^\times b_\infty
\nonumber \\
&& \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \cdot
\prod_\mathfrak{p} \int_{K_{x,\mathfrak{p}}^\times \backslash B_\mathfrak{p}^\times} \varphi_\mathfrak{p}^o(b_\mathfrak{p}^{-1}xb_\mathfrak{p})d^\times b_\mathfrak{p}. \nonumber
\end{eqnarray}
Write $a = a_0 \prod_\mathfrak{p} \mathfrak{p}^{2c_\mathfrak{p}}$, where $a_0$ is square-free.
Applying Corollary~\ref{cor: OE1} and \ref{cor: OE2}, we get
$$
\int_{K_{x,\mathfrak{p}}^\times \backslash B_\mathfrak{p}^\times} \varphi_\mathfrak{p}^o(b_\mathfrak{p}^{-1}xb_\mathfrak{p})d^\times b_\mathfrak{p}
= \frac{\text{vol}(O_{B_\mathfrak{p}}^\times/O_\mathfrak{p}^\times)}{\text{vol}(\mathcal{O}_{-a_0,\mathfrak{p}}^\times/O_\mathfrak{p}^\times)} \cdot
\sum_{\ell_\mathfrak{p} = 0}^{c_{\mathfrak{p}}}
\#\left(\frac{\mathcal{O}_{-a_0,\mathfrak{p}}^\times}{\mathcal{O}_{-a_0\mathfrak{p}^{2\ell_\mathfrak{p}},\mathfrak{p}}^\times}\right)\cdot
e(\mathcal{O}_{-a_0\mathfrak{p}^{2\ell_\mathfrak{p}},\mathfrak{p}},O_{B_\mathfrak{p}}),
$$
and
\begin{eqnarray}
\int_{K_{x,\infty}^\times \backslash B_\infty^\times} \varphi_\infty^o(y b_\infty^{-1}xb_\infty)d^\times b_\infty
&=& \frac{1}{e_\infty(K_x/k)} \cdot
\frac{\text{vol}(O_{B_\infty}^\times/O_\infty^\times)}{\text{vol}(O_{K_{x,\infty}}^\times/O_\infty^\times)}
\nonumber \\
&& \cdot
\begin{cases}
1, & \text{ if $k(\sqrt{-a})/k$ is imaginary;} \\
0, & \text{ otherwise.}
\end{cases}
\nonumber
\end{eqnarray}
Thus $I^*(a,y;\varphi^o) = 0$ unless $-a \prec 0$.
In this case,
by Proposition~\ref{prop: VOL-CN} and Lemma~\ref{lem: CN} one has
$$
\text{vol}(K_x^\times \backslash K_{x,\AA}^\times/k_\AA^\times)
= \frac{h(-a_0)}{w(-a_0)} \cdot e_\infty(K_x/k) \cdot \text{vol}(O_{K_{x,\infty}}^\times/O_\infty^\times) \cdot \prod_{\mathfrak{p}} \text{vol}(\mathcal{O}_{-a_0,\mathfrak{p}}^\times/O_\mathfrak{p}^\times).
$$
Finally, notice that for each non-zero prime ideal $\mathfrak{p}$ of $A$, from Lemma~\ref{lem: OES1} and \ref{lem: OES2} we have
$$
e(\mathcal{O}_{-a_0\mathfrak{p}^{2\ell_\mathfrak{p}},\mathfrak{p}},O_{B_\mathfrak{p}})\ = \
e^{\mathfrak{n}^+,\mathfrak{n}^-}_\mathfrak{p}(\ell_\mathfrak{p}).
$$
Hence we conclude that when $-a \prec 0$,
\begin{eqnarray}
I^*(a,y;\varphi^o) &=& \beta(y) \cdot \text{vol}(O_{B_\AA}^\times/O_\AA^\times)
\nonumber \\
&& \cdot \frac{h(-a_0)}{w(-a_0)} \cdot \prod_\mathfrak{p} \left[\sum_{\ell_\mathfrak{p} = 0}^{c_\mathfrak{p}}
\#\left(\frac{\mathcal{O}_{-a_0,\mathfrak{p}}^\times}{\mathcal{O}_{-a_0\mathfrak{p}^{2\ell_\mathfrak{p}},\mathfrak{p}}^\times}\right)\cdot e^{\mathfrak{n}^+,\mathfrak{n}^-}(\ell_\mathfrak{p})\right]\nonumber \\
&=& \text{vol}(O_{B_\AA}^\times/O_\AA^\times) \cdot \tilde{\beta}(y) \cdot H^{\mathfrak{n}^+,\mathfrak{n}^-}(-a). \nonumber
\end{eqnarray}
The last equality follows from Proposition~\ref{prop: HCN}.
\end{proof}
\subsection{Alternative expression of the Fourier coefficients}\label{sec: AFC}
Given $a \in A$, recall that $V^o_a:=\{x \in B^o\mid x^2 = a\}$.
For $y \in k_\infty^\times$ with $\deg a + 2 \leq 2 \operatorname{ord}_\infty(y)$,
in the proof of Lemma~\ref{lem: FCo} we have the following expression:
$$
I^*(-a,y;\varphi^o)
= \tilde{\beta}(y) \cdot \sum_{x \in V^o_a} \int_{B^\times \backslash B_\AA^\times/k_\AA^\times}
\varphi^o(yb^{-1}xb)d^\times b.
$$
Put $O_{\widehat{B}}:= O_B\otimes_A \widehat{A}$ with $\widehat{A} := \prod_{\mathfrak{p}} O_\mathfrak{p}$.
From the strong approximation theorem one has the following bijection:
$$
O_B^\times \backslash B_\infty^\times/k_\infty^\times \longleftrightarrow B^\times \backslash
B_\AA^\times /k_\AA^\times O_{\widehat{B}}
^\times.$$
Let $\Lambda^o_a:=B^o_a \cap \Lambda^o$.
Then we may write the Fourier coefficient $I^*(-a,y;\varphi^o)$ as
\begin{eqnarray}\label{eqn: AFC}
I^*(-a,y;\varphi^o)
&=& \tilde{\beta}(y)\cdot \text{vol}(O_{\widehat{B}}^\times/\widehat{A}) \cdot \sum_{x \in \Lambda_a^o}
\int_{O_B^\times \backslash B_\infty^\times/k_\infty^\times} \varphi_\infty^o(y b_\infty^{-1} xb_\infty) d^\times b_\infty. \nonumber \\
&=&
\tilde{\beta}(y)\cdot \text{vol}(O_{\widehat{B}}^\times/\widehat{A}) \cdot \sum_{x \in O_B^\times \backslash \Lambda_a^o}
\int_{(O_B^\times\cap K_x^\times) \backslash B_\infty^\times/k_\infty^\times} \varphi_\infty^o(y b_\infty^{-1} xb_\infty) d^\times b_\infty.
\end{eqnarray}
\begin{lem}\label{lem: Inf}
Given $a \in A$ with $a \prec 0$ and $x \in \Lambda_a^o$, we have
$$
\int_{(O_B^\times\cap K_x^\times) \backslash B_\infty^\times/k_\infty^\times} \varphi_\infty^o(y b_\infty^{-1} xb_\infty) d^\times b_\infty
=\text{\rm vol}(O_{B_\infty}^\times/O_\infty^\times) \cdot \frac{q-1}{\#(O_B^\times \cap K_x^\times)}.
$$
\end{lem}
\begin{proof}
Observe that
\begin{eqnarray}
\int_{(O_B^\times\cap K_x^\times) \backslash B_\infty^\times/k_\infty^\times} \varphi_\infty^o(y \cdot b_\infty^{-1} xb_\infty) d^\times b_\infty
&=& \text{vol}\big((O_B^\times\cap K_x^\times)\backslash K_{x,\infty}^\times/k_\infty^\times\big) \nonumber \\
&& \cdot
\int_{K_{x,\infty}^\times \backslash B_\infty^\times}
\varphi_\infty^o(y \cdot b_\infty^{-1} x b_\infty) d^\times b_\infty. \nonumber
\end{eqnarray}
Since
$$
\text{vol}\big((O_B^\times\cap K_x^\times)\backslash K_{x,\infty}^\times/k_\infty^\times\big)
= \frac{q-1}{\#(O_B^\times \cap K_x^\times)} \cdot e_\infty(K_{x,\infty}/k_\infty) \cdot \text{vol}(O_{K_{x,\infty}}^\times/O_\infty^\times)
$$
and by Corollary~\ref{cor: OE2} (as $a \prec 0$) one has
$$
\int_{K_{x,\infty}^\times \backslash B_\infty^\times}
\varphi_\infty^o(y \cdot b_\infty^{-1} x b_\infty) d^\times b_\infty
= \frac{1}{e_\infty(K_{x,\infty}/k_\infty)} \cdot
\frac{\text{vol}(O_{B_{\infty}}^\times/O_\infty^\times)}{\text{vol}(O_{K_{x,\infty}}^\times /O_\infty^\times)},
$$
the result holds.
\end{proof}
The equation~\eqref{eqn: AFC} and Lemma~\ref{lem: Inf} show that for $a \in A$ with $a \prec 0$ and $y \in k_\infty^\times$ with $\deg a + 2 \leq 2 \operatorname{ord}_\infty(y)$, we get
$$
I^*(-a,y;\varphi^o)
= \text{\rm vol}(O_{B_\AA}^\times/O_\AA^\times) \cdot \tilde{\beta}(y) \cdot
\sum_{x \in O_B^\times \backslash \Lambda_a^o}
\frac{q-1}{\#(O_B^\times \cap K_x^\times)}.
$$
Moreover,
consider the (right) action of $\{\pm 1\}$ on $O_B^\times \backslash \Lambda_a^o$ by scalar multiplication.
For $[x] \in O_B^\times \backslash \Lambda_a^o$ (represented by $x \in \Lambda_a^o$), let $ [x]\cdot \{\pm 1\}$ be the $\{\pm 1\}$-orbit of $[x]$.
Then
$$
\#\big([x] \cdot \{\pm 1\} \big) =
\begin{cases}
1, & \text{ if there exists $\gamma \in O_B^\times$ so that $\gamma^{-1} x \gamma = -x$;} \\
2, & \text{ otherwise.}
\end{cases}
$$
For $x \in \Lambda_a^o$, put $\nu(x) := \#\big([x] \cdot \{\pm 1\}\big)$.
We conclude that:
\begin{cor}\label{cor: AFC}
Given $a \in A$ with $a \prec 0$ and $y \in k_\infty^\times$ with $\deg a + 2 \leq 2 \operatorname{ord}_\infty(y)$, we have
$$
I^*(-a,y;\varphi^o)
= \text{\rm vol}(O_{B_\AA}^\times/O_\AA^\times) \cdot \tilde{\beta}(y) \cdot
\sum_{x \in O_B^\times \backslash \Lambda_a^o/\{\pm 1\}}
\frac{q-1}{\#(O_B^\times \cap K_x^\times)} \cdot \nu(x).
$$
\end{cor}
This expression allows us to connect $I^*(-a,y;\varphi^o)$ with a weighted sum over the ``CM points'' on the ``Drinfeld-Stuhler modular curves'' in Section~\ref{sec: CMBES}.
\subsection{Extension of $I(g;\varphi^o)$}\label{sec: Ext-PTI}
Notice that the Kubota $2$-cocycle $\sigma^1_\infty$ can be extended to $\operatorname{GL}_2(k_\infty)$ by:
$$\sigma_\infty(g,g'):=
\left(\frac{X(gg')}{X(g)},\frac{X(gg')}{\det(g)X(g')}\right)_\infty \cdot \frac{s_\infty(g) s_\infty(g')}{s_\infty(gg')}
$$
where
$$
s_\infty\left(\begin{pmatrix}a&b \\ c&d \end{pmatrix}\right):=
\begin{cases} \big(c,d/(ad-bc)\big)_\infty, &\text{ if $\operatorname{ord}_\infty(c)$ is odd and $d \neq 0$;}\\
1, & \text{ otherwise.}
\end{cases}
$$
Let $\widetilde{\operatorname{GL}}_2(k_\infty)$ be the central extension of $\operatorname{GL}_2(k_\infty)$ by ${\mathbb C}^1:=\{z \in {\mathbb C}^\times \mid z\bar{z} = 1\}$ associated with $\sigma_\infty$, i.e.\ $\widetilde{\operatorname{GL}}_2(k_\infty)$ is identified with $\operatorname{GL}_2(k_\infty) \times {\mathbb C}^1$ as sets with the following group law:
$$(g_1,\xi_1)\cdot (g_2,\xi_2) := \big(g_1g_2, \xi_1\xi_2 \sigma_\infty (g_1,g_2)\big), \quad \forall (g_1,\xi_1),(g_2,\xi_2) \in\widetilde{\operatorname{GL}}_2(k_\infty).
$$
Then $\widetilde{\operatorname{SL}}_2(k_\infty)$ becomes a subgroup of $\widetilde{\operatorname{GL}}_2(k_\infty)$.
We remark that $\sigma_\infty$ splits on $\operatorname{GL}_2(O_\infty)$, and the inclusion $k_\infty^\times \hookrightarrow \widetilde{\operatorname{GL}}_2(k_\infty)$ defined by
$$ y \longmapsto \tilde{y}:= \left(y, \frac{\varepsilon_\infty(1)}{\varepsilon_\infty(y)}\right) \in \widetilde{\operatorname{GL}}_2(k_\infty)$$
is a group homomorphism.
Moreover, for $y \in k_\infty^\times$ and $\tilde{g} \in \widetilde{\operatorname{SL}}_2(k_\infty)$, it is checked that
$$ \tilde{y} \cdot \tilde{g} = \tilde{g} \cdot \tilde{y}.$$
Let
$$
\mathcal{K}_\infty^+ := \left\{\begin{pmatrix} a&b \\ c&d \end{pmatrix} \in \operatorname{GL}_2(O_\infty)\ \bigg|\ c \equiv 0 \bmod \pi_\infty \text{ and } (\pi_\infty,d)_\infty = 1\right\},
$$
and let $\widetilde{\mathcal{K}}_\infty^+$ be the corresponding subgroup in $\widetilde{\operatorname{GL}}_2(k_\infty)$.
We introduce the following \textit{weight-$s$} operator $T_{\infty,s}$ on the space of functions $f$ on $\widetilde{\operatorname{GL}}_2(k_\infty)/\widetilde{\mathcal{K}}_\infty^+$ (cf.\ \cite[p.\ 34]{CLWY}):
$$
T_{\infty,s} f (\tilde{g})
:= q^{s/2-1}\cdot \sum_{\epsilon \in {\mathbb F}_q} f\left(\tilde{g}\cdot \left(\begin{pmatrix} \pi_\infty & \epsilon \\ 0&1\end{pmatrix},1\right)\right), \quad \forall \tilde{g} \in \widetilde{\operatorname{GL}}_2(k_\infty)/\widetilde{\mathcal{K}}_\infty^+.
$$
\begin{defn}
Suppose $s \in \frac{1}{2}{\mathbb Z}$. A function $f$ on $\widetilde{\operatorname{GL}}_2(k_\infty)/\widetilde{\mathcal{K}}_\infty^+$ is called a \textit{weight-$s$ metaplectic form} if
$$ f((1,\xi)\cdot \tilde{g}) = \xi^{2s}\cdot f(\tilde{g})
\quad \text{ and } \quad
T_{\infty,s} f (\tilde{g}) = f(\tilde{g}) \quad \forall \tilde{g} \in \widetilde{\operatorname{GL}}_2(k_\infty).
$$
\end{defn}
We shall extend $I(\cdot;\varphi^o)$ to a weight-$\frac{3}{2}$ metaplectic form.
Given $(g,\xi) \in \widetilde{\operatorname{GL}}_2(k_\infty)$,
suppose $\operatorname{ord}_\infty(\det(g)) \equiv 0 \bmod 2$.
Write $\det(g) = u \cdot \pi_\infty^{2\ell}$ with $\ell \in {\mathbb Z}$ and $u \in O_\infty^\times$.
One gets
$$
g^{(1)}:= \pi_\infty^{-\ell} \cdot g \cdot \begin{pmatrix} u^{-1} & 0 \\ 0 & 1 \end{pmatrix} \in \operatorname{SL}_2(k_\infty).
$$
Thus we can decompose $(g,\xi)$ into
$$
(g,\xi) = \tilde{\pi}_\infty^{\ell}\cdot (g^{(1)},1) \cdot \left(\begin{pmatrix}u&0 \\ 0&1 \end{pmatrix},1\right) \cdot (1,\xi_1),$$
where
$$
\xi_1 := \xi\cdot \big(\pi_\infty^{\ell}, X(g)\big)_\infty \cdot \frac{\varepsilon_\infty(\pi_\infty^{\ell})}{\varepsilon_\infty(1)} \in {\mathbb C}^1.
$$
When $\operatorname{ord}_\infty(\det(g)) \equiv 1 \bmod 2$, for $\epsilon \in {\mathbb F}_q$ we put
$$(g_\epsilon,\xi_\epsilon):= (g,\xi) \cdot \left(\begin{pmatrix} \pi_\infty & \epsilon \\ 0 & 1 \end{pmatrix},1\right).$$
In particular, one has $\operatorname{ord}_\infty(g_\epsilon) \equiv 0 \bmod 2$.
We now define
$$
\vartheta^o(g,\xi):=\frac{1}{\text{vol}(O_{B_\AA}^\times/O_\AA^\times)}\cdot
\begin{cases}
\xi_1^3\cdot I\big((g^{(1)},1);\varphi^o\big), & \text{ if $\operatorname{ord}_\infty(\det(g)) \equiv 0 \bmod 2$;} \\
& \\
\displaystyle q^{-1/4}\cdot \sum_{\epsilon \in {\mathbb F}_q} \vartheta^o(g_\epsilon,\xi_\epsilon), & \text{ if $\operatorname{ord}_\infty(\det(g)) \equiv 1 \bmod 2$.}
\end{cases}
$$
Let
$$\Gamma_0^{(1)}(\mathfrak{n}):= \left\{\begin{pmatrix} a&b \\ c&d \end{pmatrix} \in \operatorname{SL}_2(A)\ \bigg|\ c \equiv 0 \bmod \mathfrak{n}\right\}.$$
Given $a \in A$ and $s \in {\mathbb R}$, recall the following analogue of Bessel function introduced in Theorem~\ref{thm: MT2}:
for $y \in k_\infty^\times$,
$$\beta_{a,s}(y):=|y|_\infty^{s/2}\cdot
\begin{cases} 1, & \text{ if $\deg a + 2 \leq \operatorname{ord}_\infty(y)$;}\\
0, & \text{ otherwise.}
\end{cases}$$
We derive that:
\begin{thm}\label{thm: FET0}
The function $\vartheta^o$ is a well-defined weight-$\frac{3}{2}$ metaplectic form on $\widetilde{\operatorname{GL}}_2(k_\infty)/\mathcal{K}_\infty^+$ satisfying that
$$
\vartheta^o(\tilde{z} \tilde{\gamma} \tilde{g}) = \vartheta^o(\tilde{g}),
\quad \forall z \in k_\infty^\times \text{ and }\gamma \in \Gamma_0^{(1)}(\mathfrak{n}).$$
Moreover, for $(x,y) \in k_\infty\times k_\infty^\times$, we have the following Fourier expansion
\begin{eqnarray}\label{eqn: FE1}
\vartheta^o\left(\begin{pmatrix}y&x \\ 0&1\end{pmatrix},1\right)
&=& \sum_{d \in A,\ d \preceq 0} H^{\mathfrak{n}^+,\mathfrak{n}^-}(d) \cdot \big(\beta_{d,\frac{3}{2}}(y)\psi_\infty(-dx)\big).
\end{eqnarray}
\end{thm}
\begin{proof}
The definition of $\vartheta^o$ assures that
$$\vartheta^o(\tilde{z}\tilde{g}) = \vartheta^o(\tilde{g}) \quad \text{ for every $z \in k_\infty^\times$ and $\tilde{g} \in \widetilde{\operatorname{GL}}_2(k_\infty)$.}$$
Moreover, the invariant property of $I(\cdot;\varphi^o)$ in Proposition~\ref{prop: TLo} guarantees that
$$\vartheta^o(\tilde{\gamma}\tilde{g}\tilde{\kappa}) = \vartheta^o(\tilde{g}) \quad \text{ for every } \gamma \in \Gamma_0^{(1)}(\mathfrak{n}),\ \tilde{g} \in \widetilde{\operatorname{GL}}_2(k_\infty),\ \kappa \in \widetilde{\mathcal{K}}_\infty^+.
$$
Notice that for $(x,y) \in k_\infty\times k_\infty^\times$ with $\operatorname{ord}_\infty(y) \equiv 0 \bmod 2$, we write $y = u \pi_\infty^{2\ell}$ with $\ell \in {\mathbb Z}$, and get
$$
\left(\begin{pmatrix}y&x \\ 0&1 \end{pmatrix},1\right)
= \tilde{\pi}_\infty^\ell \cdot \left(\begin{pmatrix} \pi_\infty^{\ell}& x \pi_\infty^{-\ell} \\ 0 & \pi_\infty^{-\ell}\end{pmatrix},1\right)
\cdot \left(1, \frac{\varepsilon_\infty(\pi_\infty^\ell)}{\varepsilon_\infty(1)}\right).
$$
Since $\varepsilon_\infty(y)^4/\varepsilon_\infty(1)^4 = 1$, from Theorem~\ref{thm: FC1} the equality~\eqref{eqn: FE1} holds.
The case when $\operatorname{ord}_\infty(y) \equiv 1 \bmod 2$ is similar.
Finally, for each $\tilde{g} \in \widetilde{\operatorname{GL}}_2(k_\infty)$, there exists $z \in k_\infty^\times$, $\gamma \in \Gamma_0^{(1)}(\mathfrak{n})$, $(x,y) \in k_\infty\times k_\infty^\times$, $\kappa \in \mathcal{K}_\infty^+$, and $\xi \in {\mathbb C}^1$ such that
$$\tilde{g} = \tilde{z}\cdot \tilde{\gamma} \cdot \left(\begin{pmatrix} y&x \\ 0&1 \end{pmatrix},1\right) \cdot \tilde{\kappa} \cdot (1,\xi).
$$
To show the weight-$3/2$ property of $\vartheta^o$, it suffices to assume that
$$\tilde{g} = \left(\begin{pmatrix}y&x \\ 0&1 \end{pmatrix},1\right), \quad \text{ for } (x,y) \in k_\infty\times k_\infty^\times.
$$
Therefore the equality $T_{\infty,\frac{3}{2}} \vartheta^o = \vartheta^o$ holds by comparing the Fourier expansions on both sides.
\end{proof}
\section{CM points on the Drinfeld-Stuhler modular curves}\label{sec: CMBES}
Let ${\mathbb C}_\infty$ be the completion of a chosen algebraic closure of $k_\infty$.
The \textit{Drinfeld half plane} is
$$\mathfrak{H}:= {\mathbb C}_\infty - k_\infty,$$
which equipped with the M\"obius action of $\operatorname{GL}_2(k_\infty)$:
$$\begin{pmatrix} a&b \\ c&d\end{pmatrix} \cdot z := \frac{az+b}{cz+d}, \quad \forall \begin{pmatrix} a&b \\ c&d\end{pmatrix} \in \operatorname{GL}_2(k_\infty),\ z \in \mathfrak{H}.
$$
We recall the analytic construction of Drinfeld-Stuhler modular curves as follows.
Let $B$ be an indefinite quaternion algebra over $k$, and $\mathfrak{n}^- \in A_+$ be the product of the primes at which $B$ is ramified.
Take a square-free $\mathfrak{n}^+ \in A_+$ coprime to $\mathfrak{n}^-$, and let $O_B$ be an Eichler $A$-order in $B$ of type $(\mathfrak{n}^+,\mathfrak{n}^-)$.
Fix an isomorphism $B\otimes_k k_\infty \cong \operatorname{Mat}_2(k_\infty)$, which embeds $\Gamma(\mathfrak{n}^+,\mathfrak{n}^-):= O_B^\times$ into $\operatorname{GL}_2(k_\infty)$ as a discrete subgroup.
This induces an action of $O_B^\times$ on the Drinfeld half plane $\mathfrak{H}$.
Let
$$X(\mathfrak{n}^+,\mathfrak{n}^-):= \Gamma(\mathfrak{n}^+,\mathfrak{n}^-)\backslash \mathfrak{H},$$ which is a rigid analytic space (compact if $B$ is division).
From the moduli interpretation of $X(\mathfrak{n}^+,\mathfrak{n}^-)$ (parametrizing the ``$\mathscr{B}$-ellptic sheaves with additional level-$\mathfrak{n}^+$ structure'', cf.\ \cite{LRS} and \cite{Pak}),
we may identify $X(\mathfrak{n}^+,\mathfrak{n}^-)$ (rigidly analytically) with the ${\mathbb C}_\infty$-valued points of a smooth curve (projective if $B$ is division) over ${\mathbb C}_\infty$,
called the \textit{Drinfeld-Stuhler modular curve for $\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)$}.
For our purpose, we shall only use the analytic description of $X(\mathfrak{n}^+,\mathfrak{n}^-)$.
Notice that when $B = \operatorname{Mat}_2(k)$, every Eichler $A$-order $O_B$ of type $(\mathfrak{n}^+,1)$ is equal (up to conjugation) to
$$\left\{\begin{pmatrix} a&b \\ c&d\end{pmatrix} \in \operatorname{Mat}_2(A)\ \bigg|\ c \equiv 0 \bmod \mathfrak{n}^+\right\},
$$
and so $\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)$ coincides with the congruence subgroup
$$
\Gamma_0(\mathfrak{n}^+)
:= \left\{\begin{pmatrix} a&b \\ c&d \end{pmatrix} \in \operatorname{GL}_2(A)\ \bigg|\ c \equiv 0 \bmod \mathfrak{n}^+\right\}.
$$
The ``compactification''
$$ X_0(\mathfrak{n}^+) := \Gamma_0(\mathfrak{n})\backslash \Big(\mathfrak{H}\cup {\mathbb P}^1(k)\Big)$$
is called the \textit{Drinfeld modular curve for $\Gamma_0(\mathfrak{n}^+)$}.\\
Recall that $B^o$ consists of all the pure quaternions in $B$, $\Lambda^o:=O_B\cap B^o$, and $$\Lambda^o_a := \{x \in \Lambda^o_a\mid Q^o(x) = -x^2 = a\}.$$
\begin{defn}\label{defn: CMP}
Given non-zero $d \in A$, a point $\boldsymbol{z} \in X(\mathfrak{n}^+,\mathfrak{n}^-)$ is \textit{CM with discriminant $d$} if there exists a representative $z \in \mathfrak{H}$ of $\boldsymbol{z}$ and $x \in \Lambda^o_d$
so that $x\cdot z = z$.
\end{defn}
\begin{rem}
The CM points on $X_0(\mathfrak{n}^+,\mathfrak{n}^-)$ can be viewed as analogue of CM points on Shimura curves in the classical case.
Indeed, when $\mathfrak{n}^+ = 1$ (i.e.\ $O_B$ is a maximal $A$-order in $B$), the CM points are in bijection with the isomorphism classes of the Drinfeld-Stuhler $O_B$-modules with ``complex multiplications'', see \cite[Theorem 4.10 and \textit{Remark} 4.11]{Pak}.
\end{rem}
Let $\boldsymbol{z} \in X(\mathfrak{n}^+,\mathfrak{n}^-)$ be a CM point with discriminant $d \in A$ which is represented by $z \in \mathfrak{H}$.
Set
$$
w(z) := \frac{q-1}{\#\big(\text{Stab}_{\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)}(z)\big)}.$$
Here $\operatorname{Stab}_{\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)}(z)$ is the stablizer of $z$ in $\Gamma(\mathfrak{n}^+,\mathfrak{n}^-)$.
Then $w(z)$ only depends on the point $\boldsymbol{z} \in X(\mathfrak{n}^+,\mathfrak{n}^-)$.
Put $w(\boldsymbol{z}) = w(z)$.
For non-zero $d \in A$, we are interested in the following \textit{mass}:
$$
\mathcal{M}(d):= \sum_{\subfrac{\text{CM } \boldsymbol{z} \in X(\mathfrak{n}^+,\mathfrak{n}^-)}{\text{with discriminant $d$}}}
\frac{1}{w(\boldsymbol{z})}.
$$
Let
$$
S_d:=\{z \in \mathfrak{H} \mid \text{ there exists } x \in \Lambda_d^o \text{ so that } x \cdot z = z\}.
$$
Then $S_d$ is empty unless $-d \prec 0$.
In this case, we have $k_\infty(z) = k_\infty(\sqrt{-d})$ for each $z \in S_d$.
Identifying $\operatorname{Gal}(k_\infty(\sqrt{-d})/k_\infty)$ with $\{\pm 1\}$, we have a (right) action of $\{\pm 1\}$ on $S_d$ commuting with the (left) action of $O_B^\times$.
This then induces an action of $\{\pm 1\}$ on $O_B^\times \backslash S_d$.
Denote by $\bar{z} \in k_\infty(\sqrt{-d})$ as the conjugate of $z$.
For $z \in S_d$, put
$$
\nu(z):=
\begin{cases}
1, & \text{ if there exists $\gamma \in O_B^\times$ so that $\gamma \cdot z = \bar{z}$;}\\
2, & \text{ otherwise.}
\end{cases}
$$
Then we may express $\mathcal{M}(d)$ in the following:
\begin{prop}\label{prop: M(a)}
Given $d \in A$ with $d \prec 0$, we have
$$
\mathcal{M}(d) = \sum_{z \in O_B^\times \backslash S_d/\{\pm 1\}} \frac{\nu(z)}{w(z)}.
$$
\end{prop}
\begin{proof}
It is from the definition that
$$
\mathcal{M}(d) = \sum_{z \in O_B^\times \backslash S_d} \frac{1}{w(z)}.
$$
Moreover, for each element in $O_B^\times \backslash S_d$ represented by $z \in S_d$, the cardinality of its $\{\pm 1\}$-orbit equals to $\nu(z)$.
Therefore the result follows.
\end{proof}
We have the following bijection
$$S_d/\{\pm 1\} \quad \longleftrightarrow \quad \Lambda_d^o/\{\pm 1\}$$
by sending $[z] \in S_d/\{\pm 1\}$ to $[x_z] \in \Lambda_d^o/\{\pm 1\}$, where
$x_z \cdot z = z$.
To see this is, for $z \in S_d$ and $x_z \in \Lambda_d^o$ with $x_z \cdot z = z$,
one has $x_z \cdot \bar{z} = \bar{z}$, which says that two vectors $(z,1)^t$ and $(\bar{z},1)^t$ are distinct eigenvectors of $x_z$.
If $x \in \Lambda_d^o$ with $x \cdot z = z$, then $x$ and $x_z$ share the same eigenvectors, which implies that $x \in K_{x_z} = k(x_z)\subset B$. Since $x^2 = -d = x_z^2$, we must have $x = \pm x_z$, i.e.\
$$
\{x \in \Lambda_d^o \mid x \cdot z = z\} = \{\pm x_z\}.
$$
On the other hand for $x \in \Lambda_d^o$ and $z_x \in \mathfrak{H}$ with $x\cdot z_x = z_x$,
we have
$$\{z \in \mathfrak{H} \mid x \cdot z = z\} = \{z_x, \bar{z}_x\}.$$
Therefore the above map is well-defined and bijective.
It is clear that the bijection is actually $O_B^\times$-equivariant.
Moreover:
\begin{lem}
Given $d \in A$ with $d \prec 0$, $z \in S_d$ and $x_z \in \Lambda^o_d$ with $x_z\cdot z = z$, we have
$$
w(z) = \frac{q-1}{\#(O_B^\times \cap K_{x_z}^\times)}
\quad \text{ and } \quad
\nu(z) = \nu(x_z).
$$
\end{lem}
\begin{proof}
We first show that $O_B^\times \cap K_{x_z}^\times = \text{Stab}_{O_B^\times}(z)$.
To see this, for $\gamma \in O_B^\times$, one has
$\gamma \in K_{x_z}$ if and only if $\gamma$ and $x_z$ share the same eigenvectors, which is equivalent to that $\gamma$ fixes $z$ (and $\bar{z}$), i.e.\ $\gamma \in \text{Stab}_{O_B^\times}(z)$.
To prove the second assertion, suppose
$\nu(z) = 1$, i.e.\ there exists $\gamma \in O_B^\times$ so that $\gamma z = \bar{z}$.
In this case, we obtain that
$$(\gamma x_z \gamma^{-1})\cdot z = z,$$
which implies $\gamma x_z \gamma^{-1} = \pm x_z$.
If $\gamma x_z \gamma^{-1} = x_z$, i.e.\ $\gamma \in K_{x_z}$, then $\gamma \cdot z = z$, a contradiction.
Hence $\gamma x_z \gamma^{-1} = -x_z$, showing that $\nu(x_z) = 1$.
Conversely, if $\nu(x_z) = 1$, i.e.\ there exists $\gamma \in O_B^\times$ so that $\gamma x_z \gamma^{-1} = -x_z$.
Then $$x_z \cdot (\gamma \cdot z) = \gamma \cdot z,$$
showing that $\gamma \cdot z = z$ or $\bar{z}$.
If $\gamma \cdot z = z$, then $\gamma$ and $x_z$ share the same eigenvectors. Thus $\gamma x_z \gamma^{-1} = x_z$, a contradiction.
Therefore $\gamma \cdot z = \bar{z}$, i.e.\ $\nu(z) = 1$.
\end{proof}
From Proposition~\ref{prop: M(a)}, the above lemma enables us to rewrite $\mathcal{M}(d)$ as
$$
\mathcal{M}(d) = \sum_{x \in O_B^\times\backslash \Lambda_d^o/\{\pm 1\}}\frac{q-1}{\#(O_B^\times \cap K_{x}^\times)} \cdot \nu(x).
$$
For convention, let
$$\mathcal{M}(0) := H^{\mathfrak{n}^+,\mathfrak{n}^-}(0).$$
Together with Corollary~\ref{cor: AFC}, we arrive at:
\begin{thm}\label{thm: CNo-geo}
Let $\mathfrak{n}^+,\mathfrak{n}^- \in A_+$ be two square-free polynomials with $\text{\rm gcd}(\mathfrak{n}^+,\mathfrak{n}^-) = 1$.
Suppose the number of prime factors of $\mathfrak{n}^-$ is positive and even.
Given $d \in A$ and $y \in k_\infty^\times$ with $\deg d + 2 \leq 2 \operatorname{ord}_\infty(y)$, the following equality holds:
$$
I^*(-d,y;\varphi^o) = \text{\rm vol}(O_{B_\AA}^\times/O_\AA^\times) \cdot \tilde{\beta}(y) \cdot \mathcal{M}^{\mathfrak{n}^+,\mathfrak{n}^-}(d).
$$
Consequently, the mass $\mathcal{M}^{\mathfrak{n}^+,\mathfrak{n}^-}(d)$ over the CM points with discriminant $d$ coincides with $H^{\mathfrak{n}^+,\mathfrak{n}^-}(d)$ for every $d \in A$ with $d \prec 0$.
\end{thm}
Recall that $\vartheta^o$ is a weight-$\frac{3}{2}$ metaplectic form on $\widetilde{\operatorname{GL}}_2(k_\infty)$ extended from $I(\cdot;\varphi^o)$.
From Theorem~\ref{thm: FET0}, the above theorem leads us to:
\begin{cor}\label{cor: TSo-FC}
For $(x,y) \in k_\infty \times k_\infty^\times$, the Fourier expansion of $\vartheta^o$ can be written via masses:
$$
\vartheta^o\left(\begin{pmatrix} y&x\\ 0&1\end{pmatrix},1\right)
=
\sum_{d \in A}
\mathcal{M}^{\mathfrak{n}^+,\mathfrak{n}^-}(d)\cdot \big(\beta_{d,\frac{3}{2}}(y) \psi_\infty(-dx)\big).
$$
\end{cor}
\begin{rem}
(1) Using Eichler's theory of optimal embeddings (cf.\ \cite[p.\ 94]{Vie}), the equality
$$H^{\mathfrak{n}^+,\mathfrak{n}^-}(d) = \mathcal{M}^{\mathfrak{n}^+,\mathfrak{n}^-}(d) \quad \forall d \in A,\ d\prec 0 $$
also holds for $\mathfrak{n}^- = 1$.\\
(2) When the number of prime factors of $\mathfrak{n}^-$ is odd, the generating function
$$
\sum_{d \in A, \ d \prec 0}
H^{\mathfrak{n}^+,\mathfrak{n}^-}(d) \cdot \big(\beta_{d,\frac{3}{2}}(y)\psi_\infty(-dx)\big), \quad \forall (x,y) \in k_\infty\times k_\infty^\times
$$
extends as well to a ``weight-$3/2$'' metaplectic form on $\widetilde{\operatorname{GL}}_2(k_\infty)$ (cf.\ \cite[Proposition 2.2]{Wei} and \cite[Corollary 5.4]{CLWY}).
\end{rem}
\section{Theta series with nebentypus}\label{sec: TSN}
Fix a square-free $\mathfrak{d} \in A_+$ with $\deg \mathfrak{d}$ even.
Let $F = k(\sqrt{\mathfrak{d}})$.
For each $\alpha \in F$, the Galois conjugate of $\alpha$ (over $k$) is denoted by $\alpha'$.
Given $x = \begin{pmatrix} a&b \\ c&d\end{pmatrix} \in \operatorname{Mat}_2(F)$, put
$$
\bar{x} := \begin{pmatrix} d&-b \\ -c&a\end{pmatrix} \quad \text{ and } \quad
x' := \begin{pmatrix} a'&b' \\ c'&d'\end{pmatrix}.
$$
Given $\mathfrak{n} \in A_+$,
let $*$ be the involution on $\operatorname{Mat}_2(F)$
defined by: for $x \in \operatorname{Mat}_2(F)$,
$$
x^* := \begin{pmatrix}0& 1/\mathfrak{n} \\ 1&0 \end{pmatrix} \bar{x}' \begin{pmatrix} 0 & 1 \\ \mathfrak{n}&0 \end{pmatrix}.
$$
Let
$$
V:= \{x \in \operatorname{Mat}_2(F)\mid x^* = x\} \quad \text{ and } \quad Q_V := \det\big|_V.
$$
Then $(V,Q_V)$ is a quadratic space with degree $4$ over $k$.
In concrete terms, we have
$$
V = \left\{\begin{pmatrix}a&\beta \\ -\mathfrak{n}\beta' & d \end{pmatrix}\ \bigg|\
a,d \in k,\ \beta \in F\right\}.
$$
Let
$$B_1 := \{b \in \operatorname{Mat}_2(F)\mid b^* = \bar{b}\}
=
\left\{\begin{pmatrix}\alpha&\beta \\ \mathfrak{n}\beta' & \alpha' \end{pmatrix}\ \bigg|\
\alpha,\beta \in F\right\}.
$$
We may identify $B_1$ with the quaternion algebra
$$\left(\frac{\mathfrak{d},\mathfrak{n}}{k}\right) := k+k\mathbf{i}+k\mathbf{j}+k\mathbf{ij}, \quad
\text{where }
\mathbf{i}^2 = \mathfrak{d},\ \mathbf{j}^2 = \mathfrak{n}, \mathbf{ij}=-\mathbf{ji},
$$
where $\mathbf{i}$ corresponds to $\begin{pmatrix} \sqrt{\mathfrak{d}}&0 \\ 0& - \sqrt{\mathfrak{d}} \end{pmatrix}$
and $\mathbf{j}$ corresponds to $\begin{pmatrix}0&1 \\ \mathfrak{n} &0\end{pmatrix}$.
From now on, we make the following assumptions:
\begin{apn}\label{apn: level}
${}$
\begin{itemize}
\item[(1)] The polynomial $\mathfrak{n} \in A_+$ is square-free and coprime to $\mathfrak{d}$.
\item[(2)] Write $\mathfrak{n} = \mathfrak{n}^+\cdot \mathfrak{n}^-$ (resp.\ $\mathfrak{d} = \mathfrak{d}^+\cdot \mathfrak{d}^-$), where each prime factor $\mathfrak{p}$ of $\mathfrak{n}^{\pm}$ (resp.\ $\mathfrak{d}^{\pm}$) satisfies that the Legendre quadratic symbol $\left(\frac{\mathfrak{d}}{\mathfrak{p}}\right) = \pm 1$
(resp.\ $\left(\frac{\mathfrak{n}}{\mathfrak{p}}\right) = \pm 1$). Then $\deg (\mathfrak{d}^-\mathfrak{n}^-) > 0$.
\end{itemize}
\end{apn}
\begin{rem}
Under the above assumptions, we observe that $B_1$ is the indefinite division quaternion algebra over $k$ ramified precisely at prime factors of $\mathfrak{d}^-\mathfrak{n}^-$.
\end{rem}
Consider the following left exact sequence
$$
1 \longrightarrow k^\times \longrightarrow B_1^\times \longrightarrow \text{SO}(V),
$$
where the map from $B_1^\times$ into $\text{SO}(V)$ is defined by
$$
b \longmapsto h_b := (x \mapsto bxb^{-1}), \quad \forall b \in B_1^\times.
$$
Given $\varphi \in S(V(k_\AA))$, we are interested in the following theta integral:
$$I(g;\varphi) := \int_{B_1^\times \backslash B_{1,\AA}^\times/k_\AA^\times} \Theta(g,h_b;\varphi) d^\times b, \quad \forall g \in \operatorname{SL}_2(\AA).
$$
For $a \in k$ and $y \in k_\AA^\times$, let
$$
I^*(a,y;\varphi):=
\int_{k\backslash k_\AA}I\left(\bigg(\begin{pmatrix} y & uy^{-1} \\ 0 & y^{-1} \end{pmatrix},1\bigg);\varphi\right) \psi(-a u) du.
$$
Put $V_a:= \{x \in V\mid Q_V(x) = a\}$.
Adapting the proof of Lemma~\ref{lem: FCo}, we obtain:
\begin{lem}\label{lem: FCV}
For $a \in k$ and $y \in \AA^\times$, we have
\begin{eqnarray}
I^*(a,y;\varphi)
&=& |y|_\AA^2 \cdot \sum_{x \in B_1^\times \backslash V_a} \text{\rm vol}(K_x^\times \backslash K_{x,\AA}^\times/k_\AA^\times) \cdot
\int_{K_{x,\AA}^\times \backslash B_{1,\AA}^\times} \varphi(yb^{-1}xb) d^\times b. \nonumber
\end{eqnarray}
Here $K_x$ is the centralizer of $x$ in $B_1$, and $K_{x,\AA} = K_x \otimes_k k_\AA$.
\end{lem}
We shall take a special pure-tensor Schwartz function and determine the Fourier coefficients of the associated theta integral.
\subsection{Particular Schwartz function}\label{sec: TSN-PSF}
Note that the trace map $\operatorname{Tr}: \operatorname{Mat}_2(F) \rightarrow F$ restricting to $V$ gives a $k$-linear functional on $V$.
For $x \in V$, put
$$x^{\natural}:= \left(x - \frac{\operatorname{Tr}(x)}{2}\right) \cdot \sqrt{\mathfrak{d}} \in B_1^o, $$
where
$B_1^o$ is the space of of pure quaternions in $B_1$, i.e.\
$$B_1^o = \{b \in B \mid \operatorname{Tr}(b) = 0\}.$$
Then
$$
K_x = \begin{cases}
k(x^\natural), & \text{ a quadratic field over $k$ if $x^\natural \neq 0$,}\\
B_1, & \text{otherwise.}
\end{cases}
$$
For $a \in k$, notice that two elements $x_1$ and $x_2$ in $V_a$ belong to the same orbit of $B_1^\times$ (under the conjugation action) if and only if $\operatorname{Tr}(x_1) = \operatorname{Tr}(x_2)$.
This follows from
$$
a = Q_V(x) = \frac{\operatorname{Tr}(x)^2}{4} - \frac{(x^\natural)^2}{\mathfrak{d}}, \quad \forall x \in V_a.
$$
Take
$$\Lambda := \operatorname{Mat}_2(O_F) \cap V =
\left\{\begin{pmatrix}a&\beta \\ -\beta'\mathfrak{n} & d \end{pmatrix} \bigg|\ a,d \in A,\ \beta \in O_F\right\}
$$
and
$$O_{B_1} := \operatorname{Mat}_2(O_F) \cap B_1 =
\left\{\begin{pmatrix} \alpha & \beta \\ \beta' \mathfrak{n} & \alpha' \end{pmatrix}\bigg|\ \alpha,\beta \in O_F\right\}.
$$
It is checked that $O_{B_1}$ is an Eichler $A$-order in $B_1$ of type $(\mathfrak{d}^+\mathfrak{n}^+, \mathfrak{d}^-\mathfrak{n}^-)$, and $u^{-1}xu \in \Lambda$ for every $x \in \Lambda$ and $u \in O_{B_1}^\times$.
For each non-zero prime ideal $\mathfrak{p}$ of $A$,
put $\Lambda_\mathfrak{p}:= \Lambda\otimes_A O_\mathfrak{p}$ and $\Lambda_\mathfrak{p}^\natural := O_{B_{1,\mathfrak{p}}}^o = \{b \in O_{B_{1,\mathfrak{p}}}\mid \operatorname{Tr}(b) = 0\}$.
Then:
\begin{lem}
For $x_\mathfrak{p} \in V(k_\mathfrak{p})$ with $Q_V (x_\mathfrak{p}) \in O_\mathfrak{p}$, we have
$$
x_\mathfrak{p} \in \Lambda_\mathfrak{p} \quad \text{ if and only if } \quad
\operatorname{Tr}(x_\mathfrak{p}) \in O_\mathfrak{p} \text{ and } x_\mathfrak{p}^\natural \in \Lambda_\mathfrak{p}^\natural.
$$
\end{lem}
\begin{proof}
It is straightforward that when $x_\mathfrak{p} \in \Lambda_\mathfrak{p}$, one has $\operatorname{Tr}(x_\mathfrak{p}) \in O_\mathfrak{p}$ and $x_\mathfrak{p}^\natural \in \Lambda_\mathfrak{p}^\natural$.
Conversely, suppose $\operatorname{Tr}(x_\mathfrak{p}) \in O_\mathfrak{p}$ and $x_\mathfrak{p}^\natural \in \Lambda_\mathfrak{p}^\natural$.
Write
$$x_\mathfrak{p}^\natural = \begin{pmatrix} a\sqrt{\mathfrak{d}} & \beta \\ \beta'\mathfrak{n} & -a\sqrt{\mathfrak{d}} \end{pmatrix} \quad \text{ with } a \in O_\mathfrak{p},\ \beta \in O_{F,\mathfrak{p}} := O_F \otimes_A O_\mathfrak{p}, \quad \text{ and } \quad t = \operatorname{Tr}(x_\mathfrak{p}) \in O_\mathfrak{p}.
$$
Then
$$
Q_V(x_\mathfrak{p}) = \frac{t^2}{4} - \frac{(x_\mathfrak{p}^\natural)^2}{\mathfrak{d}}
= \frac{t^2}{4} + \frac{a \mathfrak{d} + \operatorname{Nr}_{F/k}(\beta)\mathfrak{n}}{\mathfrak{d}} \in O_\mathfrak{p}.
$$
Since $\mathfrak{n}$ is coprime to $\mathfrak{d}$, we obtain that
$$\frac{\operatorname{Nr}_{F/k}(\beta)}{\mathfrak{d}} \in O_\mathfrak{p},
$$
which shows that $\beta = \sqrt{\mathfrak{d}} \cdot \tilde{\beta}$ for some $\tilde{\beta} \in O_{F,\mathfrak{p}}$ (as $\mathfrak{d}$ is square-free).
Therefore
$$x_\mathfrak{p} = \frac{t}{2} + \frac{x_\mathfrak{p}^\natural}{\sqrt{\mathfrak{d}}}
= \begin{pmatrix}
t/2 + a & \tilde{\beta} \\ -\tilde{\beta}\mathfrak{n} & t/2 - a
\end{pmatrix} \in \Lambda_\mathfrak{p}.
$$
\end{proof}
Take
$$
\varphi_{\Lambda,\mathfrak{p}} := \mathbf{1}_{\Lambda_\mathfrak{p}} \in S(V(k_\mathfrak{p})),
$$
and
$$
\varphi^{\natural}_\mathfrak{p} := \mathbf{1}_{\Lambda_\mathfrak{p}^\natural} \in S(B_{1,\mathfrak{p}}^o).
$$
The above lemma says that for $x_\mathfrak{p} \in V(k_\mathfrak{p})$, we have
\begin{eqnarray}\label{eqn: compare-phi}
\varphi_{\Lambda,\mathfrak{p}}(x_\mathfrak{p}) = 1 \quad \text{ if and only if } \quad
\operatorname{Tr}(x) \in O_\mathfrak{p} \text{ and } \varphi^\natural_\mathfrak{p} (x^\natural_\mathfrak{p}) = 1.
\end{eqnarray}
As $\mathfrak{d}$ is monic with even degree, the field $F$ is \textit{real} over $k$, i.e.\ the infinite place $\infty$ of $k$ splits in $F$.
Fix an embedding $F\hookrightarrow k_\infty$, which induces a $k_\infty$-algebra isomorphism
$$B_{1,\infty} = B_1\otimes_k k_\infty \cong \operatorname{Mat}_2(k_\infty).$$
On the other hand, the natural inclusion $V \hookrightarrow \operatorname{Mat}_2(F)$ also induces an isomorphism $V(k_\infty) \cong \operatorname{Mat}_2(k_\infty)$ (as quadratic spaces over $k_\infty$).
Recall in Section~\ref{sec: PTI1} that we take
$$L_\infty := \varpi \cdot \operatorname{Mat}_2(O_\infty),$$
and
$$L_\infty' := \left\{\begin{pmatrix} a & b \\ c & d \end{pmatrix}
\in L_\infty \ \bigg|\ c \in \varpi^2 O_\infty\right\}.
$$
Via the identification $V(k_\infty) \cong \operatorname{Mat}_2(k_\infty)$,
we may view $L_\infty$ and $L_\infty'$ as two $O_\infty$-lattices in $V_\infty$.
Choose
$$
\varphi_{\Lambda,\infty} := \mathbf{1}_{L_\infty} - \frac{q+1}{2} \cdot \mathbf{1}_{L_\infty'} \in S(V(k_\infty))
\quad \text{ and } \quad
\varphi_\infty^\natural := \mathbf{1}_{L_\infty^o} - \frac{q+1}{2} \cdot \mathbf{1}_{L_\infty^{\prime,o}} \in S(B_{1,\infty}^o).
$$
It is straightforward that:
\begin{lem}
For $x \in V(k_\infty)$ with $ \operatorname{Tr}(x) \in \varpi O_\infty$, we have
$$
\varphi_{\Lambda,\infty}( x) =
\varphi_\infty^\natural(\sqrt{\mathfrak{d}} \cdot x^\natural).
$$
\end{lem}
Let
$$\varphi_\Lambda := \otimes_\mathfrak{p} \varphi_{\Lambda,\mathfrak{p}} \otimes \varphi_{\Lambda,\infty} \in S(V(\AA)).$$
Similar to Proposition~\ref{prop: TLo}, we have the following transformation law of $I(g;\varphi_\Lambda)$:
\begin{prop}\label{prop: TLN}
Given $\gamma \in \operatorname{SL}_2(k)$, $g \in \operatorname{SL}_2(k_\AA)$ and $\kappa \in \mathcal{K}_0^1(\mathfrak{d} \mathfrak{n} \infty)$, we have
$$
I(\gamma g \kappa; \varphi_\Lambda)
= \chi_\mathfrak{d}(\kappa) \cdot I(g;\varphi_\Lambda).
$$
Here for $\kappa = \begin{pmatrix} a&b \\ c&d\end{pmatrix} \in \mathcal{K}_0^1(\mathfrak{d} \mathfrak{n} \infty)$
with $d = (d_v)_v \in O_\AA$ $($and so $d_\mathfrak{p} \in O_\mathfrak{p}^\times$ for each prime factor $\mathfrak{p}$ of $\mathfrak{d})$, we put
$$
\chi_\mathfrak{d}(\kappa) := \prod_{\mathfrak{p} \mid \mathfrak{d}}\left(\frac{d_\mathfrak{p}}{\mathfrak{p}}\right).
$$
\end{prop}
This transformation law implies in particular that for $a \in k$, $y \in k_\AA^\times$, $\varepsilon \in O_\AA^\times$, and $u \in O_\AA$, we have
$$
I^*(a,y;\varphi_\Lambda) = I^*(\alpha^{-2}a, \alpha y \varepsilon; \varphi_\Lambda)=
I^*(a,y;\varphi_\Lambda) \cdot \psi(ay^2 u).
$$
Thus it suffices to consider $y \in k_\infty^\times$, and in this case we get
$$I^*(a,y;\varphi_\Lambda) = 0
\quad \text{ unless $a\in A$ with $\deg a + 2 \leq 2 \operatorname{ord}_\infty(y)$}.
$$
Next, we shall express $I^*(a,y;\varphi_\Lambda)$ in terms of the modified Hurwitz class numbers.
\subsection{Fourier coefficients of $I(g;\varphi_\Lambda)$}
For $y \in k_\infty^\times$ and $a \in A$ with $\deg a + 2 \leq 2 \operatorname{ord}_\infty(y)$, we get
\begin{eqnarray}
I^*(a,y;\varphi_\Lambda)
&=& |y|_\infty^2 \cdot \sum_{x \in B_1^\times \backslash V_a} \text{vol}(K_x^\times \backslash K_{x,\AA}^\times/k_\AA^\times) \nonumber \\
&& \cdot \left(\prod_\mathfrak{p} \int_{K_{x,\mathfrak{p}}^\times \backslash B_{1,\mathfrak{p}}^\times}\varphi_{\Lambda,\mathfrak{p}}(b_\mathfrak{p}^{-1}xb_\mathfrak{p}) d^\times b_\mathfrak{p}\right)
\cdot \int_{K_{x,\infty}^\times \backslash B_{1,\infty}^\times} \varphi_{\Lambda,\infty}(yb_\infty^{-1}xb_\infty) d^\times b_\infty.
\nonumber
\end{eqnarray}
Thus for $x \in V_a$,
$$
\prod_\mathfrak{p} \int_{K_{x,\mathfrak{p}}^\times \backslash B_{1,\mathfrak{p}}^\times}\varphi_{\Lambda,\mathfrak{p}}(b_\mathfrak{p}^{-1}xb_\mathfrak{p}) d^\times b_\mathfrak{p} = 0
\quad \text{ unless \ \ $t = \operatorname{Tr}(x) \in A$.}
$$
In this case,
$$
K_x = \begin{cases}
k(\sqrt{\mathfrak{d}(t^2-4a)}), & \text{ if $x \notin k$;}\\
B_1, & \text{ otherwise.}
\end{cases}
$$
Moreover, for $x \in V_a$ so that $K_x/k$ is an imaginary quadratic extension, the condition $\deg a + 2 \leq 2 \operatorname{ord}_\infty(y)$ implies that $y\cdot \operatorname{Tr}(x) \in \varpi O_\infty$.
Hence by \eqref{eqn: compare-phi} we have
\begin{eqnarray}
&& \prod_\mathfrak{p} \int_{K_{x,\mathfrak{p}}^\times \backslash B_{1,\mathfrak{p}}^\times}\varphi_{\Lambda,\mathfrak{p}}(b_\mathfrak{p}^{-1}xb_\mathfrak{p}) d^\times b_\mathfrak{p}
\cdot \int_{K_{x,\infty}^\times \backslash B_{1,\infty}^\times} \varphi_{\Lambda,\infty}(yb_\infty^{-1}xb_\infty) d^\times b_\infty \nonumber \\
&=&
\prod_\mathfrak{p} \int_{K_{x,\mathfrak{p}}^\times \backslash B_{1,\mathfrak{p}}^\times}\varphi_{\mathfrak{p}}^\natural(b_\mathfrak{p}^{-1}x^\natural b_\mathfrak{p}) d^\times b_\mathfrak{p}
\cdot \int_{K_{x,\infty}^\times \backslash B_{1,\infty}^\times} \varphi^\natural_\infty\big((y\sqrt{\mathfrak{d}}) \cdot b_\infty^{-1}x^\natural b_\infty) d^\times b_\infty.
\nonumber
\end{eqnarray}
For $x \in V_a$ with $\operatorname{Tr}(x) = t \in A$, one has $$(x^\natural)^2 = \mathfrak{d}(t^2-4a).$$
As the Eichler $A$-order $O_{B_1}$ is of type $(\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-)$, similar to Theorem~\ref{thm: FC1} we obtain
\begin{eqnarray}
&& \text{vol}(K_x^\times \backslash K_{x,\AA}^\times/k_\AA^\times) \cdot \left(\prod_\mathfrak{p} \int_{K_{x,\mathfrak{p}}^\times \backslash B_{1,\mathfrak{p}}^\times}\varphi_{\Lambda,\mathfrak{p}}(b_\mathfrak{p}^{-1}xb_\mathfrak{p}) d^\times b_\mathfrak{p}\right)
\nonumber \\
&& \quad \quad \quad \quad \quad \quad \quad \quad \,
\cdot \int_{K_{x,\infty}^\times \backslash B_{1,\infty}^\times} \varphi_{\Lambda,\infty}(yb_\infty^{-1}xb_\infty) d^\times b_\infty \nonumber \\
&=&
\text{vol}(K_x^\times \backslash K_{x,\AA}^\times/k_\AA^\times) \cdot \left(\prod_\mathfrak{p} \int_{K_{x,\mathfrak{p}}^\times \backslash B_{1,\mathfrak{p}}^\times}\varphi_{\mathfrak{p}}^\natural(b_\mathfrak{p}^{-1}x^\natural b_\mathfrak{p}) d^\times b_\mathfrak{p}\right)
\nonumber \\
&& \quad \quad \quad \quad \quad \quad \quad \quad \,
\cdot \int_{K_{x,\infty}^\times \backslash B_{1,\infty}^\times} \varphi^\natural_\infty\big((y\sqrt{\mathfrak{d}}) \cdot b_\infty^{-1}x^\natural b_\infty) d^\times b_\infty \nonumber \\
&=& \text{vol}(O_{B_\AA}^\times/O_{\AA}^\times) \cdot
\begin{cases}
H^{\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}(\mathfrak{d}(t^2-4a)), & \text{ if $\mathfrak{d}(t^2-4a) \preceq 0$;}\\
0, & \text{ otherwise.}
\end{cases} \nonumber
\end{eqnarray}
Notice that two elements $x_1, x_2 \in V_a$ belong to the same $B_1^\times$-orbit if and only if $\operatorname{Tr}(x_1) = \operatorname{Tr}(x_2)$.
From the above discussion, we conclude that:
\begin{thm}\label{thm: TINFC}
Given $a \in A$ and $y \in k_\infty^\times $ with $\deg a +2 \leq 2\operatorname{ord}_\infty(y)$, the following equality holds:
$$
I^*(a,y;\varphi_\Lambda)
= \text{\rm vol}(O_{{B_1},\AA}^\times/O_\AA^\times) \cdot |y|_\infty^2 \cdot \sum_{\subfrac{t \in A}{t^2\preceq 4a}} H^{\mathfrak{d}^+ \mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}(\mathfrak{d}(t^2-4a)).
$$
\end{thm}
\subsection{Alternative expression of the Fourier coefficients}
For $y \in k_\infty^\times$ and $a \in A$,
one may express $I^*(a,y;\varphi_\Lambda)$ as
$$
I^*(a,y;\varphi_\Lambda)
= |y|_\infty^2 \cdot \sum_{x \in V_a} \int_{B_1^\times \backslash B_{1,\AA}^\times/k_\AA^\times} \varphi_\Lambda(y b_\infty ^{-1} x b_\infty ) d^\times b_\infty .
$$
As in Section~\ref{sec: AFC}, we apply the strong approximation theorem and get
$$
I^*(a,y;\varphi_\Lambda)
= |y|_\infty^2 \cdot \text{vol}(O_{\widehat{B}_1}^\times/\widehat{A}) \cdot \sum_{x \in \Lambda_a}
\int_{O_{B_1}^\times \backslash B_{1,\infty}^\times/k_\infty^\times} \varphi_{\Lambda,\infty}(y b_\infty ^{-1} xb_\infty ) d^\times b_\infty .
$$
Let
$$\Gamma = \Gamma_{0,F}(\mathfrak{n})
:= \left\{\begin{pmatrix}
a&b \\ c&d \end{pmatrix} \in \operatorname{GL}_2(O_F)\ \bigg|\ ad-bc \in {\mathbb F}_q^\times,\ c \equiv 0 \bmod \mathfrak{n} \right\}.
$$
Define an action $\star$ of $\Gamma$ on $\Lambda$ by:
$$\gamma \star x := \gamma x \gamma^* \cdot \det(\gamma)^{-1}, \quad \forall \gamma \in \Gamma,\ x \in \Lambda.$$
Given a non-zero $a \in A$ and $x \in \Lambda_a$, let
$$B_x:= \{b \in \operatorname{Mat}_2(F) \mid x b^* = \bar{b} x\}, \text{ and }
\Gamma_x := B_x^\times \cap \Gamma.
$$
Then $\Gamma_x$ is actually the stablizer of $x$ in $\Gamma$, and:
\begin{lem}\label{lem: B_x}
Given a non-zero $a\in A$ and $x \in \Lambda_a$, $B_x$ is a quaternion algebra over $k$ which is isomorphic to:
$$
\left(\frac{\mathfrak{d}, a\mathfrak{n}}{k}\right)
:= k +k\mathbf{i}+k\mathbf{j}+k\mathbf{ij},
\quad
\text{ where $\mathbf{i}^2 = \mathfrak{d}$, $\mathbf{j}^2 = a\mathfrak{n}$, and $\mathbf{ji} = -\mathbf{ij}$.}
$$
\end{lem}
\begin{proof}
Write $x = \begin{pmatrix} d_1 & \beta \\ -\mathfrak{n} \beta' & d_2 \end{pmatrix}$
where $d_1,\ d_2 \in A$ and $\beta \in O_F$ with $d_1d_2 + \mathfrak{n} \beta\beta' = a$.
Take
$$
U:= \begin{cases}
\begin{pmatrix}1 & 0 \\ \mathfrak{n} \beta' & d_1 \end{pmatrix}, & \text{ if $d_1 \neq 0$;} \\
\begin{pmatrix} d_2 & -\beta \\ 0 & a \end{pmatrix}, & \text{ if $d_1 = 0$ and $d_2 \neq 0$ } \\
\begin{pmatrix} -1 & \beta \\ \mathfrak{n} \beta & \mathfrak{n} \beta^2 \end{pmatrix}, & \text{ if $d_1 = d_2 = 0$.}
\end{cases}
$$
Then
$$
x_U:= U x U^* = \begin{pmatrix} 1& 0 \\ 0 &a \end{pmatrix} \cdot
\begin{cases}
d_1, &\text{ if $d_1 \neq 0$;}\\
a d_2, &\text{ if $d_1 = 0$ and $d_2 \neq 0$;}\\
2 a, & \text{ if $d_1 = d_2 = 0$.}
\end{cases}
$$
It is straightforward that
$B_x = U^{-1} B_{x_U} U$
and
$$
B_{x_U}
= \left\{\begin{pmatrix} \alpha & \beta \\ a \mathfrak{n} \beta' & \alpha' \end{pmatrix}
\ \bigg|\ \alpha, \beta \in F\right\}.
$$
Thus
$$
B_x \cong B_{x_U} \cong
\left(\frac{\mathfrak{d},a\mathfrak{n}}{k}\right).
$$
\end{proof}
\begin{rem}
Observe that $B_x = B_1$ if and only if $x \in k^\times$.
In this case, $a$ is a square in $A$, and $\Gamma_x = O_{B_1}^\times$.
\end{rem}
The following lemma is straightforward.
\begin{lem}
For $a \in A$ and $y \in k_\infty^\times$, we have
\begin{eqnarray}
&& I^*(a,y;\varphi_\Lambda) \nonumber \\
&=& |y|^2_\infty \cdot \text{\rm vol}(O_{\widehat{B}_1}^{\times}/\widehat{A})
\cdot
\sum_{x \in \Gamma\backslash \Lambda_a}
\sum_{\gamma \in \Gamma_1 \backslash \Gamma / \Gamma_x}
\int_{(\Gamma_1 \cap \Gamma_{\gamma \star x})\backslash B_{1,\infty}^\times /k_\infty^\times} \varphi_{\Lambda,\infty}(yb_\infty^{-1}(\gamma \star x)b_\infty)d^\times b_\infty .
\nonumber
\end{eqnarray}
\end{lem}
The integral inside the above sum expression is determined by:
\begin{lem}
Let $a \in A$ and $y \in k_\infty^\times$ with $\deg a + 2 \leq 2 \operatorname{ord}_\infty(y)$. Take $x \in V_a$.
\begin{itemize}
\item[(1)]
If $B_{x} = B_1$, then
$\Gamma_{x} = \Gamma_1$ and
$$
\int_{(\Gamma_1\cap \Gamma_{x})\backslash B_{1,\infty}^\times /k_\infty^\times} \varphi_{\Lambda,\infty}(yb_\infty ^{-1}x b_\infty )d^\times b_\infty
= \frac{1-q}{2} \cdot \text{\rm vol}(\Gamma_1 \backslash B_{1,\infty}^\times/k_\infty^\times).
$$
\item[(2)]
If $B_x \neq B_1$,
then $B_{x} \cap B_1 = K_{x}$, and
\begin{eqnarray}
&& \int_{(\Gamma_1 \cap \Gamma_{x})\backslash B_{1,\infty}^\times /k_\infty^\times} \varphi_{\Lambda,\infty}(yb_\infty ^{-1}x b_\infty )d^\times b_\infty \nonumber \\
&=& \text{\rm vol}(O_{B_{1,\infty}}^\times/O_\infty^\times)\cdot
\begin{cases}
\displaystyle\frac{q-1}{\#(\Gamma_1 \cap \Gamma_{x})}, & \text{ if $K_{x}/k$ is imaginary;}\\
0, & \text{ otherwise.}
\end{cases}
\nonumber
\end{eqnarray}
\end{itemize}
\end{lem}
\begin{proof}
When $B_x = B_1$, we get $x \in k^\times$ with $x^2 = a$.
Thus the condition $\deg a + 2 \leq 2\operatorname{ord}_\infty(y)$ implies
$$
\varphi_{\Lambda,\infty}(yb_\infty^{-1}xb_\infty) = 1-\frac{q+1}{2} = \frac{1-q}{2}, \quad \forall b_\infty \in B_{1,\infty}^\times.
$$
Hence the assertion (1) holds.\\
For (2), the integral vanishes unless $K_x/k$ is imaginary.
In this case, $\Gamma_1\cap \Gamma_x$ is a finite subgroup of $K_x^\times$,
and
\begin{eqnarray}
&& \int_{(\Gamma_1 \cap \Gamma_{x})\backslash B_{1,\infty}^\times /k_\infty^\times} \varphi_{\Lambda,\infty}(yb_\infty ^{-1}x b_\infty )d^\times b_\infty \nonumber \\
&=& \frac{\text{vol}(K_{x,\infty}^\times/k_\infty^\times)}{\#(\Gamma_1\cap \Gamma_x)}
\cdot \int_{K_{x,\infty}^\times \backslash B_{1,\infty}^\times} \varphi_\infty^{\natural}\big((y\sqrt{\mathfrak{d}}) b_\infty^{-1} x^\natural b_\infty\big) d^\times b_\infty \nonumber \\
&=&
\text{\rm vol}(O_{B_{1,\infty}}^\times/O_\infty^\times)\cdot
\frac{q-1}{\#(\Gamma_1 \cap \Gamma_{x})}. \nonumber
\end{eqnarray}
The last equality follows from Corollary~\ref{cor: OE2}.
\end{proof}
For $x \in \Lambda$, put
$$i(x):=
\begin{cases}
-H^{\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}(0), & \text{ if $B_x = B_1$;} \\
\displaystyle\frac{q-1}{\#(\Gamma_1\cap \Gamma_x)}, & \text{ if $K_x/k$ is imaginary;} \\
0, & \text{ otherwise.}
\end{cases}
$$
Define
$$
\mathcal{I}(x) := \sum_{\gamma \in \Gamma_1 \backslash \Gamma/\Gamma_x} i(\gamma\star x).
$$
We then obtain:
\begin{thm}\label{thm: AFCN}
Given $a \in A$ and $y \in k_\infty^\times$ with $2\operatorname{ord}_\infty(y)+2 \geq \deg a$, we have:
$$
I^*(a,y;\varphi_\Lambda)
= \text{\rm vol}(O_{B_\AA}^\times/O_\AA^\times)
\cdot |y|_\infty^2 \cdot \sum_{x \in \Gamma \backslash \Lambda_a} \mathcal{I}(x).
$$
\end{thm}
The above theorem enables us to connect the Fourier coefficients of the theta integral $I(g;\varphi_\Lambda)$ with the intersection numbers of the ``Hirzebruch-Zagier-type divisors'' on the ``Drinfeld-Stuhler modular surfaces'' in Section~\ref{sec: Intersection}.
\subsection{Extension of $I(g;\varphi_\Lambda)$}
\label{sec: Ext-TI}
Let
$$
\mathcal{K}_\infty := \left\{ \begin{pmatrix} a&b \\ c&d \end{pmatrix} \in \operatorname{GL}_2(O_\infty) \ \bigg|\ c \equiv 0 \bmod \varpi\right\}
$$
and
$$
\Gamma_0(\mathfrak{d} \mathfrak{n}) :=
\left\{ \begin{pmatrix} a&b \\ c&d \end{pmatrix} \in \operatorname{GL}_2(A) \ \bigg|\ c \equiv 0 \bmod \mathfrak{d} \mathfrak{n} \right\}.
$$
Put $\mathcal{K}_\infty^1 := \mathcal{K}_\infty \cap \operatorname{SL}_2(k_\infty)$ and $\Gamma_0^1(\mathfrak{d} \mathfrak{n}) := \Gamma_0(\mathfrak{d}\mathfrak{n}) \cap \operatorname{SL}_2(A)$.
From the strong approximation theorem, the natural embedding $\operatorname{SL}_2(k_\infty)\hookrightarrow \operatorname{SL}_2(k_\AA)$ induces the following bijection
$$ \Gamma^1_0(\mathfrak{d}\mathfrak{n}) \backslash \operatorname{SL}_2(k_\infty)/\mathcal{K}^1_\infty \longleftrightarrow
\operatorname{SL}_2(k)\backslash \operatorname{SL}_2(k_\AA)/\mathcal{K}_0^1(\mathfrak{d}\mathfrak{n}\infty).
$$
This allows us to view $I(g;\varphi_\Lambda)$ as a function on $\operatorname{SL}_2(k_\infty)/\mathcal{K}^1_\infty$ satisfying
$$
I(\gamma g_\infty;\varphi_\Lambda)
= \chi_\mathfrak{d}(\gamma) I(g_\infty;\varphi_\Lambda), \quad \forall g_\infty \in \operatorname{SL}_2(k_\infty) \text{ and } \gamma \in \Gamma_0^1(\mathfrak{d}\mathfrak{n}).
$$
We shall extend $I(\cdot;\varphi_\Lambda)$ to a function $\vartheta_\Lambda$ on $\operatorname{GL}_2(k_\infty)/k_\infty^\times \mathcal{K}_\infty$ which is ``Drinfeld-type'', i.e.\ the following \textit{harmonic property} holds:
for $g_\infty \in \operatorname{GL}_2(k_\infty)$ we have
$$
\vartheta_\Lambda(g) + \vartheta_\Lambda\left(g \begin{pmatrix} 0&1\\ \varpi&0\end{pmatrix}\right) = 0 =
\sum_{\kappa \in \operatorname{GL}_2(O_\infty)/\mathcal{K}_\infty}\vartheta_\Lambda(g\kappa).
$$
\begin{rem}\label{rem: wt-2}
Let $f$ be a Drinfeld-type automorphic form on $\operatorname{GL}_2(k_\infty)/k_\infty^\times \mathcal{K}_\infty$.
The harmonicity of $f$ implies that for $g \in \operatorname{GL}_2(k_\infty)$ we have
$$
\sum_{\epsilon \in {\mathbb F}_q} f\left(g\begin{pmatrix} \pi_\infty &\epsilon \\ 0&1 \end{pmatrix}\right) = f(g).
$$
Extending $f$ to a function on the metaplectic group $\widetilde{\operatorname{GL}}_2(k_\infty)$ (introduced in Section~\ref{sec: Ext-PTI}) by
$$f(g,\xi) := \xi^4 \cdot f(g), \quad \forall \xi \in {\mathbb C}^1.$$
The harmonicity implies in particular that $f$ is actually a weight-$2$ form, i.e.\ (cf.\ \cite[Example 4.2]{CLWY})
$$
f((1,\xi)\tilde{g}) = \xi^4 \cdot f(\tilde{g})
\quad \text{ and } \quad T_{\infty,2}f(\tilde{g}) = f(\tilde{g}), \quad \forall \xi \in {\mathbb C}^1 \text{ and } \tilde{g} \in \widetilde{\operatorname{GL}}_2(k_\infty).
$$
Viewed as
analogue to classical weight-two modular forms, Drinfeld-type automorphic forms are objects of great interest in the study of
function field arithmetic.
We refer the readers to \cite{G-R}, \cite{CLWY}, \cite{CWY}, and \cite{Wei}) for further discussions.
\end{rem}
Let $w_\infty:= \begin{pmatrix} 0&1\\ \varpi&0\end{pmatrix}$.
We first prove that:
\begin{lem}\label{lem: har}
Given $g_\infty \in \operatorname{SL}_2(k_\infty)$, the following equality holds:
$$
\sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}^1_\infty} I(g_\infty \kappa;\varphi_\Lambda) = 0 =
\sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}^1_\infty} I(g_\infty w_\infty^{-1} \kappa w_\infty ;\varphi_\Lambda).
$$
\end{lem}
\begin{proof}
Notice that $\omega_\infty^V(\kappa)\mathbf{1}_{L_\infty} = \mathbf{1}_{L_\infty}$ for every $\kappa \in \operatorname{SL}_2(O_\infty)$, and
$$
\sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}^1_\infty}\omega_\infty^V(\kappa)\mathbf{1}_{L_\infty'}
= \mathbf{1}_{L_\infty} + \mathbf{1}_{w_\infty L_\infty w_\infty'}.
$$
Thus
\begin{eqnarray}
&& \sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}_\infty^1}I(g_\infty\kappa;\varphi_\Lambda) \nonumber \\
&=&
(q+1)\cdot I(g_\infty;\otimes_\mathfrak{p} \varphi_{\Lambda,\mathfrak{p}} \otimes \mathbf{1}_{L_\infty})
- \frac{q+1}{2}\cdot
I\Big(g_\infty;\otimes_\mathfrak{p} \varphi_{\Lambda,\mathfrak{p}} \otimes \left(\mathbf{1}_{L_\infty}+\mathbf{1}_{w_\infty L_\infty w_\infty^{-1}}\right)\Big) \nonumber \\
&=&
(q+1)\cdot I(g_\infty;\otimes_\mathfrak{p} \varphi_{\Lambda,\mathfrak{p}} \otimes \mathbf{1}_{L_\infty})
- \frac{q+1}{2}\cdot 2 \cdot
I(g_\infty;\otimes_\mathfrak{p} \varphi_{\Lambda,\mathfrak{p}} \otimes \mathbf{1}_{L_\infty}) \nonumber \\
&=& 0. \nonumber
\end{eqnarray}
Similarly, let
$$
L_\infty'':= \begin{pmatrix} \varpi O_\infty & O_\infty \\ \varpi O_\infty & O_\infty \end{pmatrix}.
$$
Then
$$
\sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}^1_\infty}\omega_\infty^V\left(w_\infty \kappa w_\infty^{-1}\right) \mathbf{1}_{L_\infty}
=
q \cdot \sum_{\kappa \in SL_2(O_\infty)/\mathcal{K}^1_\infty}
\mathbf{1}_{\kappa L_\infty'' \kappa^{-1}},
$$
and
$$
\sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}^1_\infty}\omega_\infty^V(w_\infty \kappa w_\infty^{-1}) \mathbf{1}_{L_\infty'}
= q\cdot \left(\mathbf{1}_{L_\infty''} + \mathbf{1}_{w_\infty L_\infty'' w_\infty^{-1}}\right).
$$
Therefore
\begin{eqnarray}
&& \sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}_\infty^1}I\left(g_\infty w_\infty \kappa w_\infty^{-1};\varphi_\Lambda\right) \nonumber \\
&=&
\sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}_\infty^1}
q\cdot I\left(g_\infty;
\otimes_\mathfrak{p} \varphi_{\Lambda,\mathfrak{p}} \otimes \mathbf{1}_{\kappa L_\infty''\kappa^{-1}}\right) \nonumber \\
&& - \frac{q+1}{2}\cdot q\cdot
I\Big(g_\infty;\otimes_\mathfrak{p} \varphi_{\Lambda,\mathfrak{p}} \otimes \left(\mathbf{1}_{L_\infty''}+\mathbf{1}_{w_\infty L_\infty'' w_\infty^{-1}}\right)\Big) \nonumber \\
&=&
(q+1)\cdot q \cdot I\left(g_\infty;\otimes_\mathfrak{p} \varphi_{\Lambda,\mathfrak{p}} \otimes \mathbf{1}_{L_\infty''}\right)
- \frac{q+1}{2}\cdot q \cdot 2 \cdot
I\left(g_\infty;\otimes_\mathfrak{p} \varphi_{\Lambda,\mathfrak{p}} \otimes \mathbf{1}_{L_\infty''}\right) \nonumber \\
&=& 0. \nonumber
\end{eqnarray}
\end{proof}
Let $$\operatorname{GL}_2^+(k_\infty):= \{g \in \operatorname{GL}_2^+(k_\infty)\mid \operatorname{ord}_\infty(\det g) \equiv 0 \bmod 2\}.$$
The natural inclusion $\operatorname{SL}_2(k_\infty) \hookrightarrow \operatorname{GL}_2(k_\infty)$ gives a bijection $$\operatorname{SL}_2(k_\infty)/\mathcal{K}^1_\infty \longleftrightarrow \operatorname{GL}_2^+(k_\infty)/k_\infty^\times \mathcal{K}_\infty.$$
Thus $I(\cdot;\varphi_\Lambda)$ can be viewed as a function on
$\operatorname{GL}_2^+(k_\infty)/k_\infty^\times \mathcal{K}_\infty$.
For $g_\infty \in \operatorname{GL}_2(k_\infty)$,
define $\vartheta_\Lambda(g_\infty)$ by:
$$
\vartheta_\Lambda(g_\infty) := \frac{2}{\text{vol}(O_{B_\AA}^\times/O_\AA^\times)} \cdot
\begin{cases}
I(g_\infty;\varphi_\Lambda), & \text{ if $g_\infty \in \operatorname{GL}_2^+(k_\infty)$;}\\
-I(g_\infty w_\infty ;\varphi_\Lambda), & \text{ otherwise.}
\end{cases}
$$
The above lemma implies immediately that:
\begin{thm}\label{thm: Ext-DTAF}
The function $\vartheta_\Lambda$ on $\operatorname{GL}_2(k_\infty)/k_\infty^\times \mathcal{K}_\infty$ satisfies the harmonic property, i.e.\ for $g_\infty \in \operatorname{GL}_2(k_\infty)$,
$$
\vartheta_\Lambda(g_\infty) + \vartheta_\Lambda(g_\infty w_\infty) = 0 =
\sum_{\kappa \in \operatorname{GL}_2(O_\infty)/\mathcal{K}_\infty}\vartheta_\Lambda(g_\infty \kappa).
$$
Moreover, for $\gamma \in \Gamma_0^{(1)}(\mathfrak{d}\mathfrak{n})$ we have
$$
\vartheta_\Lambda(\gamma g_\infty) = \chi_\mathfrak{d}(\gamma) \cdot \vartheta_\Lambda(g_\infty), \quad \forall g_\infty \in \operatorname{GL}_2(k_\infty).
$$
Here for $\gamma = \begin{pmatrix} a&b \\ \mathfrak{d}\mathfrak{n} c & d \end{pmatrix} \in \Gamma_0^{(1)}(\mathfrak{d}\mathfrak{n})$,
$\chi_\mathfrak{d}(\gamma)$ is equal to the Legendre quadratic symbol $\left(\frac{d}{\mathfrak{d}}\right)$.
\end{thm}
\begin{proof}
The second assertion follows directly from Proposition~\ref{prop: TLN}.
To show the harmonicity of $\vartheta_\Lambda$, by definition we get immediately that
$$\vartheta_\Lambda(g_\infty) + \vartheta_\Lambda(g_\infty w_\infty) = 0.
$$
Moreover, suppose $g_\infty \in \operatorname{GL}_2^+(k_\infty)$.
Then by Lemma~\ref{lem: har}, one has
$$\sum_{\kappa \in \operatorname{GL}_2(O_\infty)/\mathcal{K}_\infty}\vartheta_\Lambda(g_\infty \kappa)
= \frac{2}{\text{vol}(O_{B_\AA}^\times/O_\AA^\times)} \cdot \sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}_\infty^1} I(g_\infty \kappa; \varphi_\Lambda) = 0.
$$
When $g_\infty \notin \operatorname{GL}_2^+(k_\infty)$,
by Lemma~\ref{lem: har} again we get
$$\sum_{\kappa \in \operatorname{GL}_2(O_\infty)/\mathcal{K}_\infty}\vartheta_\Lambda(g_\infty \kappa)
= \frac{-2}{\text{vol}(O_{B_\AA}^\times/O_\AA^\times)} \cdot \sum_{\kappa \in \operatorname{SL}_2(O_\infty)/\mathcal{K}_\infty^1} I\big((g_\infty w_\infty) w_\infty^{-1} \kappa w_\infty; \varphi_\Lambda\big) = 0.
$$
Therefore the proof is complete.
\end{proof}
\begin{rem}\label{rem: DTS-FC}
In conclusion, we extend the theta integral $I(\cdot;\varphi_\Lambda)$ to a Drinfeld-type automorphic form $\vartheta_\Lambda$ on $\operatorname{GL}_2(k_\infty)$ for the congruence subgroup $\Gamma_0^1(\mathfrak{d}\mathfrak{n})$ with nebentypus $\chi_\mathfrak{d}$, whose Fourier expansion is:
for $(x,y) \in k_\infty\times k_\infty^\times$,
$$
\vartheta_\Lambda\begin{pmatrix} y & x \\ 0 & 1\end{pmatrix}
= 2 \cdot \sum_{a \in A} \left(\sum_{\subfrac{t \in A}{t^2 \preceq 4a}}H^{\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}\big(\mathfrak{d}(t^2-4a)\big)\right) \cdot \big(\beta_{a,2}(y)\psi_\infty(a x)\big).
$$
\end{rem}
\section{Intersections of the Hirzebruch-Zagier-type divisors}\label{sec: Intersection}
Here we follow the notations in the last section.
Let $\mathfrak{d} \in A_+$ be square-free with $\deg \mathfrak{d}$ even and $F = k(\sqrt{\mathfrak{d}})$.
Identifying $F_\infty := F\otimes_k k_\infty \cong k_\infty \times k_\infty$,
we denote the image of $\alpha \in F$ in $k_\infty^2$ by $(\alpha, \alpha')$ (i.e.\ $\alpha'$ is the Galois conjugate of $\alpha$ over $k$).
Let $\mathfrak{H}_F:= \mathfrak{H}\times \mathfrak{H}$, equipped
with the M\"obius action of $\operatorname{GL}_2(k_\infty)^2$.
The above embedding $F\hookrightarrow k_\infty\times k_\infty$ gives $\operatorname{GL}_2(F) \hookrightarrow \operatorname{GL}_2(k_\infty)^2$, which induces an action of $\operatorname{GL}_2(F)$ on $\mathfrak{H}_F$.
In concrete terms, for $g = \begin{pmatrix}a&b \\ c&d \end{pmatrix} \in \operatorname{GL}_2(F)$ and $\vec{z} = (z_1,z_2) \in \mathfrak{H}_F$, define
$$
g \cdot \vec{z} := \left(\frac{az_1+b}{cz_1+d}\ ,\ \frac{a'z_2+b'}{c'z_2+d'}\right).
$$
For $\mathfrak{n} \in A_+$,
recall that
$$\Gamma_{0,F}(\mathfrak{n})
= \left\{\begin{pmatrix}
a&b \\ c&d \end{pmatrix} \in \operatorname{GL}_2(O_F)\ \bigg|\ ad-bc \in {\mathbb F}_q^\times,\ c \equiv 0 \bmod \mathfrak{n} \right\}.
$$
The \textit{Drinfeld-Stuhler modular surface for $\Gamma_{0,F}(\mathfrak{n})$} is $$\mathcal{S}_{0,F}(\mathfrak{n}) := \Gamma_{0,F}(\mathfrak{n})\backslash \mathfrak{H}_F.$$
From the work of Stuhler \cite{Stu}, $\mathcal{S}_{0,F}(\mathfrak{n})$ is a moduli space of the so-called ``Frobenius-Hecke sheaves'' (an analogue of the Hilbert-Blumental abelian surfaces in the classical case) with additional ``level-$\mathfrak{n}$ structure''.
This provides the algebraic structure of the surface $\mathcal{S}_{0,F}(\mathfrak{n})$.
For our purpose, we only consider $\mathcal{S}_{0,F}(\mathfrak{n})$ as a rigid analytic space, and study the intersections of the ``Hirzebruch-Zagier-type'' divisors on $\mathcal{S}_{0,F}(\mathfrak{n})$.
\subsection{Hirzebruch-Zagier-type divisors}
Recall in Section~\ref{sec: TSN} that we let
$$V = \{x \in \operatorname{Mat}_2(F) \mid x^* = x\} \quad \text{ and } \quad \Lambda = V \cap \operatorname{Mat}_2(O_F).$$
Given $x \in \Lambda$ with $\det(x) \neq 0$,
let $\mathcal{C}_x := \Gamma_x \backslash \mathfrak{H}$, the Drinfeld-Stuhler modular curve for $\Gamma_x$.
Put $$S_x := \begin{pmatrix}0&1 \\ \mathfrak{n} & 0 \end{pmatrix} \bar{x}.$$
The closed immersion
$\mathfrak{H} \rightarrow \mathfrak{H}_F$ defined by $(z\mapsto (z,S_x z))$ induces a (rigid analytic) proper morphism
$f_x : \mathcal{C}_x \rightarrow \mathcal{S}_{0,F}(\mathfrak{n})$.
We put
$X_x := f_x(\mathcal{C}_x)$ and $\mathcal{Z}_x:= f_{x,*}(\mathcal{C}_x)$, the pushforward divisor of $\mathcal{C}_x$ under $f_x$ on $\mathcal{S}_{0,F}(\mathfrak{n})$.
Let
$$
\widehat{\Gamma}_x := \{\gamma \in \Gamma_{0,F}(\mathfrak{n})\mid \gamma \star x = \pm x\}.
$$
Then $[\widehat{\Gamma}_x : \Gamma_x] = 1 \text{ or } 2$, and:
\begin{lem}
For $x \in \Lambda$ with $\deg(x) \neq 0$, one has
$$\mathcal{Z}_x = [\widehat{\Gamma}_x : \Gamma_x]\cdot X_x.
$$
\end{lem}
\begin{proof}
We need to show that the proper morphism $f_x : \mathcal{C}_x \rightarrow X_x$
has degree equal to $[\widehat{\Gamma}_x :\Gamma_x]$.
Let $\boldsymbol{z}_1, \boldsymbol{z}_2 \in \mathcal{C}_x$ be two points with $f_x(\boldsymbol{z}_1) = f_x(\boldsymbol{z}_2) \in X_x$.
Take representatives $\vec{z}_1 = (z_1,S_x z_1)$ and $\vec{z}_2 = (z_2, S_x z_2)$ of $\boldsymbol{z}_1$ and $\boldsymbol{z}_2$ on $\mathfrak{H}_F$, respectively.
There exists $\gamma \in \Gamma$ so that
$$\vec{z}_1 = \gamma \cdot \vec{z}_2,\ \text{ i.e.\ }
(z_1,S_x z_1) = (\gamma z_2, \gamma' S_x z_2).
$$
Thus
$$
z_1 = \gamma z_1 = \gamma ((\gamma'S_x)^{-1}S_x) z_1
= \big((\gamma \star x) \bar{x}\big) z_1.
$$
When $z_1$ is in ``general position'', e.g.\ the stablizer of $z_1$ in $\operatorname{GL}_2(F)$ is $F^\times$,
one has
$(\gamma\star x) \bar{x} \in F^\times$.
Taking the determinant of $(\gamma\star x) \bar{x}$, we obtain
$\gamma \star x = \pm x$, which says that $\gamma \in \widehat{\Gamma}_x$.
Therefore the result holds.
\end{proof}
Suppose now that $\mathfrak{n}$ satisfies Assumption~\ref{apn: level}.
We shall study the number of intersections of $\mathcal{Z}_1$ and $\mathcal{Z}_x$ by lifting to a ``fine covering'' of $\mathcal{S}_{0,F}(\mathfrak{n})$.
More precisely, for $\mathfrak{m} \in A_+$, we let
$$
\Gamma_F(\mathfrak{m}) := \left\{ \gamma \in \operatorname{GL}_2(O_F)\ \bigg|\ \gamma \equiv \begin{pmatrix} 1&0 \\ 0&1 \end{pmatrix} \bmod \mathfrak{m}\right\}.
$$
Choose $\mathfrak{m}$ so that $\mathfrak{n}^2$ divides $\mathfrak{m}$.
Then
$$\Gamma_F(\mathfrak{m})^*:= \{ \gamma^* \mid \gamma \in \Gamma_F(\mathfrak{m})\} \subset \Gamma_F(\mathfrak{n}).$$
Consider the finite morphism
$$\pi:\Gamma_F(\mathfrak{m})\backslash \mathfrak{H}_F =: \mathcal{S}_F(\mathfrak{m}) \twoheadrightarrow \mathcal{S}_{0,F}(\mathfrak{n}).$$
For $x \in \Lambda$ with $\det(x) \neq 0$,
let $\mathfrak{H}_x := \{(z,S_x z)\mid z \in \mathfrak{H}\} \subset \mathfrak{H}_F$ and
$\widetilde{X}_x$ be the image of $\mathfrak{H}_x$ in $\mathcal{S}_F(\mathfrak{m})$ under the canonical map from $\mathfrak{H}_F$ onto $\mathcal{S}_F(\mathfrak{m})$.
Let $\Gamma_x(\mathfrak{m}) := \Gamma_x \cap \Gamma_F(\mathfrak{m})$, and put
$\widetilde{\mathcal{C}}_x := \Gamma_x(\mathfrak{m}) \backslash \mathfrak{H}$.
We have:
\begin{lem}
Assume $\mathfrak{n}^2 \det(x)$ divides $\mathfrak{m}$.
Then the identification between $\mathfrak{H} \cong \mathfrak{H}_x$ induces an isomorphism $\tilde{f}_x: \widetilde{\mathcal{C}}_x \cong \widetilde{X}_x$.
\end{lem}
\begin{proof}
Notice that the defining equation of $\mathfrak{H}_x$ in $\mathfrak{H}_F$ makes it smooth everywhere.
As each point in $\mathfrak{H}_F$ has trivial stablizer in $\Gamma_F(\mathfrak{m})$, we may identify a sufficiently small admissible open neighborhood of a given point in $\mathfrak{H}_F$ with the corresponding affinoid subdomains in $\mathcal{S}_F(\mathfrak{m})$.
This assures the smoothness of $\widetilde{X}_x$.
Therefore it suffices to show that the morphism
from $\tilde{f}_x: \widetilde{C}_x\rightarrow \widetilde{X}_x$ is a bijection.
The surjectivity of $\tilde{f}_x$ comes directly from the definition.
On the other hand, let $\tilde{\boldsymbol{z}}_1$ and $\tilde{\boldsymbol{z}}_2$ be two points on $\widetilde{\mathcal{C}}_x$ so that $\tilde{f}_x(\tilde{\boldsymbol{z}}_1) = \tilde{f}_x(\tilde{\boldsymbol{z}}_2)$.
Take representatives $\vec{z}_1 = (z_1,S_x z_1)$ and $\vec{z}_2 = (z_2, S_x z_2)$ of $\tilde{\boldsymbol{z}}_1$ and $\tilde{\boldsymbol{z}}_2$ on $\mathfrak{H}_x$, respectively.
Then there exists $\gamma \in \Gamma_F(\mathfrak{m})$ so that
$\vec{z}_1 = \gamma \cdot \vec{z}_2$, i.e.\
$$
(z_1, S_x z_1 ) = (\gamma z_2, \gamma' S_x z_2).
$$
Thus
$$ z_1 = \gamma z_2 = \big(\gamma (\gamma' S_x)^{-1} S_x\big) z_1 = \big((\gamma \star x)\bar{x}\big) z_1.
$$
Since $\mathfrak{n}^2 \det (x)$ divides $\mathfrak{m}$, we obtain that
$$
(\gamma \star x)x^{-1} = \gamma x \gamma^* x^{-1} \cdot \det(\gamma)^{-1} \in \Gamma_F(\mathfrak{n}), \quad \text{i.e. }
(\gamma \star x)x^{-1} \equiv \begin{pmatrix} 1&0 \\ 0 &1 \end{pmatrix} \bmod \mathfrak{n}.
$$
As it fixes $z_1 \in \mathfrak{H}$, we obtain that
$$
(\gamma \star x)x^{-1} = 1, \quad \text{i.e.} \quad
\gamma \star x = x.
$$
Hence $\gamma \in \Gamma_x \cap \Gamma_F(\mathfrak{m}) = \Gamma_x(\mathfrak{m})$.
In other words, the morphism $\tilde{f}_x$ is bijective, and the proof is complete.
\end{proof}
\subsection{Formula of intersections}
Let $x \in \Lambda$ with $\det(x) \neq 0$, and $\mathfrak{m} \in A_+$ with $\mathfrak{n}^2 \det(x) \mid \mathfrak{m}$.
We first verify the transversality of the intersections of $\widetilde{X}_1$ and $\widetilde{X}_x$ on $\mathcal{S}_F(\mathfrak{m})$.
\begin{lem}
Suppose $\widetilde{X}_1 \neq \widetilde{X}_x$.
Then $\widetilde{X}_1$ and $\widetilde{X}_x$ intersect transversally.
\end{lem}
\begin{proof}
It suffices to check that the preimages of $\widetilde{X}_1$ and $\widetilde{X}_x$ in $\mathfrak{H}_F$ intersect transversally.
Since $\gamma \mathfrak{H}_x = \mathfrak{H}_{\gamma \star x}$ for every $\gamma \in \Gamma$, it is reduced to show the transversality of the intersection of $\mathfrak{H}_1$ and $\mathfrak{H}_x$ when $x \notin A$.
Suppose $\vec{z} (z,S_1z) = (z,S_x z) \in \mathfrak{H}_1 \cap \mathfrak{H}_x$.
Write $x = \begin{pmatrix} d_1 & \beta \\ -\mathfrak{n} \beta' & d_2\end{pmatrix}$ with $d_1,d_2 \in k$ and $\beta \in F$,
and put $a := \det(x) = d_1d_2 + \mathfrak{n} \beta\beta' \neq 0$.
Then $\bar{x} z = z$, i.e.\
$$
\frac{d_2 z - \beta}{\mathfrak{n}\beta z + d_1} = z.
$$
Thus $\mathfrak{n} \beta z^2 +(d_1-d_2)z +\beta = 0$.
Multiplying $\beta$ on both sides we get
\begin{eqnarray}\label{eqn: T1}
(a-d_1d_2)z^2 + (d_1-d_2)\beta z + \beta^2 = 0.
\end{eqnarray}
On the other hand, the tangent vectors of $\vec{z}$ along $\mathfrak{H}_1$ and $\mathfrak{H}_x$, respectively, are
$$
\left(1, -\frac{1}{\mathfrak{n} z^2}\right) \quad \text{ and }\quad
\left(1, \frac{-a}{\mathfrak{n}(d_2z-\beta)^2}\right).
$$
If these two coincide, we get
$a z^2 = (d_2z-\beta)^2$, which says that
\begin{eqnarray}\label{eqn: T2}
(a-d_2^2)z^2 + 2d_2 \beta z - \beta^2 = 0.
\end{eqnarray}
As $z \in \mathfrak{H}$, comparing the coefficients of the equations~\eqref{eqn: T1} and \eqref{eqn: T2} we get
$$
(a-d_1d_2)\cdot 2d_2 \beta = (a-d_2^2) \cdot (d_1-d_2)\beta \quad \text{ and } \quad
(a-d_1d_2)\cdot (-\beta^2) = (a-d_2^2)\cdot \beta^2.
$$
Hence either $\beta = 0$ or
$$2ad_2 - 2d_1d_2^2 = ad_1 -ad_2 - d_1d_2^2 + d_2^3 \quad \text{ and } \quad
d_1d_2-a = a-d_2^2.$$
In both cases we can obtain $d_1 =d_2$ and $\beta = 0$, which says that $x \in A$ and $\mathfrak{H}_x = \mathfrak{H}_1$.
Therefore as $\mathfrak{H}_x \neq \mathfrak{H}_1$, the two tangent vectors much be different, i.e.\
the intersection of $\mathfrak{H}_x$ and $\mathfrak{H}_1$ at $\vec{z}$ must be transversal.
\end{proof}
Let $\widetilde{\mathcal{Z}}_x$ be the prime divisor associated with $\widetilde{X}_x$ on $\mathcal{S}_F(\mathfrak{m})$. We get
$$
\pi_*(\widetilde{\mathcal{Z}}_x) = [\Gamma_x : \Gamma_x(\mathfrak{m})\cdot {\mathbb F}_q^\times] \cdot \mathcal{Z}_x.$$
From the above lemmas, the intersection number of $\mathcal{Z}_1$ and $\mathcal{Z}_x$ is determined in the following:
\begin{prop}\label{prop: projection formula}
Given $x \in \Lambda$ with $\det(x) \neq 0$, suppose $\mathcal{Z}_1 \neq \mathcal{Z}_x$. Choose $\mathfrak{m} \in A_+$ so that $\mathfrak{n}^2 \det(x) \mid \mathfrak{m}$.
The intersection number of $\mathcal{Z}_1$ and $\mathcal{Z}_x$ is equal to
$$\mathcal{Z}_1\cdot \mathcal{Z}_x = \frac{q-1}{[\Gamma_1:\Gamma_1(\mathfrak{m})]\cdot[\Gamma_x:\Gamma_x(\mathfrak{m})]}
\cdot \sum_{\gamma \in \Gamma/\Gamma_F(\mathfrak{m})} \widetilde{\mathcal{Z}}_1 \cdot \gamma \widetilde{\mathcal{Z}}_x.
$$
\end{prop}
\begin{proof}
Observe that $\mathcal{Z}_1$ is a ${\mathbb Q}$-Cartier divisor on $\mathcal{S}_{0,F}(\mathfrak{n})$.
Thus the result is a rigid-analytic version of the projection formula for the intersection of $\mathcal{Z}_1$ and $\mathcal{Z}_x = \pi_*(\widetilde{\mathcal{Z}}_x)$.
We include the argument here for completeness.
Let $\Gamma = \Gamma_{0,F}(\mathfrak{n})$.
Given $x \in \Lambda$ with $\det(x) \neq 0$,
the irreducible curve
$X_x$ has normalization isomorphic to $\widehat{\Gamma}_x \backslash \mathfrak{H}$.
Notice that
for each $\boldsymbol{z} \in X_1 \cap X_x$,
take $\tilde{\boldsymbol{z}}_1 $ and $\tilde{\boldsymbol{z}}_x $ be two lifts of $\boldsymbol{z}$ in $\widetilde{X}_1$ and $\widetilde{X}_x$, respectively.
The intersection multiplicity of $X_1$ and $X_x$ at $\boldsymbol{z}$ is actually equal to
$$
m_{\boldsymbol{z}}(X_1,X_x)
= \frac{(q-1)\cdot \#\operatorname{Stab}_{\Gamma}(\tilde{\boldsymbol{z}}_1)}{\#\operatorname{Stab}_{\widehat{\Gamma}_1}(\tilde{\boldsymbol{z}}_1) \cdot \#\operatorname{Stab}_{\widehat{\Gamma}_x}(\tilde{\boldsymbol{z}}_x)}.
$$
Now, consider the disjoint union
$$
\Phi:= \coprod_{\gamma \in \Gamma/\widehat{\Gamma}_x\cdot \Gamma_F(\mathfrak{m})} \widetilde{X}_1 \cap \gamma \widetilde{X}_x
$$
which maps surjectively to $X_1 \cap X_x$ via the finite morphism $\pi$ on each component (we denote this surjection $\Phi \twoheadrightarrow X_1\cap X_x$ by $\tilde{\pi}$).
For each point $\boldsymbol{z} \in X_1 \cap X_x$, the pre-image of $\boldsymbol{z}$ in $\Phi$ has cardinality equal to
$$[\widehat{\Gamma}_1 : \operatorname{Stab}_{\widehat{\Gamma}_1}(\tilde{\boldsymbol{z}}_1)\cdot \Gamma_1(\mathfrak{m})] \cdot
\frac{\#\operatorname{Stab}_{\Gamma}(\tilde{\boldsymbol{z}}_1)}{ \#\operatorname{Stab}_{\widehat{\Gamma}_x}(\tilde{\boldsymbol{z}}_x)}.
$$
Thus the cardinality of $\Phi$ can be expressed as
\begin{eqnarray}
\sum_{\gamma \in \Gamma/\widehat{\Gamma}_x\cdot \Gamma_F(\mathfrak{m})} \#(\widetilde{X}_1 \cap \gamma \widetilde{X}_x)
&=&
\sum_{\boldsymbol{z} \in X_1\cap X_x} \#\big(\tilde{\pi}^{-1}(\boldsymbol{z})\big) \nonumber \\
&=&
\sum_{\boldsymbol{z} \in X_1 \cap X_x}
[\widehat{\Gamma}_1 : \operatorname{Stab}_{\widehat{\Gamma}_1}(\tilde{\boldsymbol{z}}_1)\cdot \Gamma_1(\mathfrak{m})] \cdot
\frac{\#\operatorname{Stab}_{\Gamma}(\tilde{\boldsymbol{z}}_1)}{ \#\operatorname{Stab}_{\widehat{\Gamma}_x}(\tilde{\boldsymbol{z}}_x)} \nonumber \\
&=& [\widehat{\Gamma}_1 : \Gamma_1(\mathfrak{m}){\mathbb F}_q^\times] \cdot
\sum_{\boldsymbol{z} \in X_1 \cap X_x}
\frac{(q-1)\cdot \#\operatorname{Stab}_{\Gamma}(\tilde{\boldsymbol{z}}_1)}{\#\operatorname{Stab}_{\widehat{\Gamma}_1}(\tilde{\boldsymbol{z}}_1) \cdot \#\operatorname{Stab}_{\widehat{\Gamma}_x}(\tilde{\boldsymbol{z}}_x)} \nonumber \\
&=& [\widehat{\Gamma}_1 : \Gamma_1(\mathfrak{m}){\mathbb F}_q^\times] \cdot
\sum_{\boldsymbol{z} \in X_1 \cap X_x} m_{\boldsymbol{z}}(X_1,X_x).\nonumber
\end{eqnarray}
Therefore
\begin{eqnarray}
\mathcal{Z}_1\cdot \mathcal{Z}_x
&=& [\widehat{\Gamma}_1:\Gamma_1]\cdot [\widehat{\Gamma}_x:\Gamma_x] \cdot \sum_{\boldsymbol{z} \in X_1\cap X_x}m_{\boldsymbol{z}}(X_1,X_x) \nonumber \\
&=& \frac{[\widehat{\Gamma}_1:\Gamma_1]\cdot [\widehat{\Gamma}_x:\Gamma_x]}{[\widehat{\Gamma}_1 : \Gamma_1(\mathfrak{m}){\mathbb F}_q^\times]} \cdot
\sum_{\gamma \in \Gamma/\widehat{\Gamma}_x\cdot \Gamma_F(\mathfrak{m})} \#(\widetilde{X}_1 \cap \gamma \widetilde{X}_x) \nonumber \\
&=&
\frac{q-1}{[\Gamma_1:\Gamma_1(\mathfrak{m})]\cdot[\Gamma_x:\Gamma_x(\mathfrak{m})]}
\cdot \sum_{\gamma \in \Gamma/\Gamma_F(\mathfrak{m})} \widetilde{\mathcal{Z}}_1 \cdot \gamma \widetilde{\mathcal{Z}}_x, \nonumber
\end{eqnarray}
where the last equality holds as $\widetilde{X}_1$ and $\gamma \widetilde{X}_x (= \widetilde{X}_{\gamma \star x})$ intersect transversally for every $\gamma \in \Gamma$.
\end{proof}
\begin{prop}\label{prop: Intersection}
Let $x \in \Lambda$ with $\det(x) \neq 0$, and $\mathfrak{m} \in A_+$ so that $\mathfrak{n}^2 \det(x)$ divides $\mathfrak{m}$.
Given $\gamma \in \Gamma$, suppose $\widetilde{\mathcal{Z}}_1 \neq \gamma \widetilde{\mathcal{Z}}_x$.
We have
$$
\widetilde{\mathcal{Z}}_1 \cdot \gamma \widetilde{\mathcal{Z}}_x
= \sum_{\gamma_0 \in \Gamma_1(\mathfrak{m}) \backslash \Gamma_F(\mathfrak{m})/\Gamma_{\gamma \star x}(\mathfrak{m})} \#(\mathfrak{H}_1 \cap \gamma_0 \gamma \mathfrak{H}_x).
$$
\end{prop}
\begin{proof}
It suffices to show that the union
$$\bigcup_{\gamma_0 \in \Gamma_1(\mathfrak{m}) \backslash \Gamma_F(\mathfrak{m})/\Gamma_{\gamma \star x}(\mathfrak{m})} \mathfrak{H}_1 \cap \gamma_0 \gamma \mathfrak{H}_x \ (\subset \mathfrak{H}_F)$$
is disjoint and in bijection with the intersection points of $\widetilde{X}_1$ and $\gamma \widetilde{X}_x$ under the canonical map $\mathfrak{H}_F \rightarrow \mathcal{S}_F(\mathfrak{m})$.
The surjectivity is straightforward.
On the other hand, given $\gamma_1, \gamma_2 \in \Gamma_F(\mathfrak{m})$ and $\vec{z}_i \in \mathfrak{H}_1 \cap \gamma_i\gamma \mathfrak{H}_x$ for $i = 1,2$,
write
$$\vec{z}_i = (z_i, S_1 z_i) = (\gamma_i\gamma w_i, \gamma_i'\gamma' S_x w_i) \quad \text{ with } z_i,w_i \in \mathfrak{H} \text{ for $i=1,2$.}
$$
Suppose the image of $\vec{z}_1$ and $\vec{z}_2$ coincides in $\mathcal{S}_F(\mathfrak{m})$, i.e.\ there exists $\gamma_0 \in \Gamma_F(\mathfrak{m})$ so that
$(z_1, S_1 z_1) = (\gamma_0 z_2, \gamma_0' S_1 z_2)$.
Then
$$
S_1 z_1 = \gamma_0' S_1 z_2 = \gamma_0' S_1 \gamma_0^{-1} z_1,
$$
which says
$(\gamma_0 \gamma_0^*) z_1 = z_1$.
From our choice of $\mathfrak{m}$, we get $\gamma_0 \gamma_0^* \in \Gamma_F(\mathfrak{n}\det(x))$ which fixes $z_1$. This implies $\gamma_0 \gamma_0^* = 1$.
Thus $\gamma_0 \in \Gamma_1 \cap \Gamma_F(\mathfrak{m}) = \Gamma_1(\mathfrak{m})$.
Moreover, let
$$\gamma_3 = \gamma^{-1}\gamma_1^{-1} \gamma_0 \gamma_2 \gamma \in \Gamma_F(\mathfrak{m}) \quad \text{ (as $\Gamma_F(\mathfrak{m})$ is normal in $\Gamma$)}.$$
We get $(w_1,S_x w_1) = (\gamma_3 w_2, \gamma_3' S_x w_2)$, which says
$$(\gamma_3 \cdot \bar{x}^{-1} \gamma_3^*\bar{x}) w_1 = w_1.$$
Similarly, from our choice of $\mathfrak{m}$ we get $\gamma_3 \cdot \bar{x}^{-1} \gamma_3^*\bar{x} \in \Gamma_F(\mathfrak{n})$ and fixes $w_1$.
Thus
$$\gamma_3 \cdot \bar{x}^{-1} \gamma_3^*\bar{x} = 1, \quad \text{ which shows that $x \gamma_3^* = \bar{\gamma}_3 x$.}$$
Therefore $\gamma_3 \in \Gamma_x \cap \Gamma_F(\mathfrak{m}) = \Gamma_x(\mathfrak{m})$.
In conclusion, we have
$$\gamma_1 \cdot (\gamma \gamma_3 \gamma^{-1}) = \gamma_0 \gamma_2,$$
i.e.\ $\gamma_1$ and $\gamma_2$ represents the same double cosets in
$\Gamma_1(\mathfrak{m})\backslash \Gamma_F(\mathfrak{m})/\Gamma_{\gamma \star x}(\mathfrak{m})$.
This assures the injectivity and completes the proof.
\end{proof}
\begin{lem}\label{lem: Intersection}
Given $x \in \Lambda$ with $\det(x) \neq 0$.
For $\gamma \in \Gamma$ with $\gamma \mathfrak{H}_x \neq \mathfrak{H}_1$ one has
$$
\mathfrak{H}_1 \cap \gamma \mathfrak{H}_x
= \{\vec{z} = (z,S_1 z)\mid (\gamma \star x) \cdot z = z\}.
$$
Consequently,
put
$$
\tilde{\iota}(x):= \begin{cases} 1 & \text{ if $K_{x}/k$ is an imaginary quadratic field extension;}\\
0 & \text{ otherwise.}
\end{cases}
$$
Then
$$
\#(\mathfrak{H}_1\cap \gamma\mathfrak{H}_x) = 2 \cdot \tilde{\iota}(\gamma \star x).
$$
\end{lem}
\begin{proof}
Given $\vec{z} \in \mathfrak{H}_1 \cap \gamma \mathfrak{H}_x$, write $\vec{z} = (z,S_1 z) = (\gamma w, \gamma'S_x w)$ for $z,w \in \mathfrak{H}$.
We get
$$
\gamma'S_x \gamma^{-1} z = \gamma'S_x w = S_1 z.
$$
Thus
\begin{eqnarray}
z &=& \gamma S_x^{-1} (\gamma')^{-1} S_1 z \nonumber \\
&=& \gamma x (S_1^{-1} \bar{\gamma}' S_1) \cdot z \nonumber \\
&=& (\gamma \star x) \cdot z. \nonumber
\end{eqnarray}
Conversely, given $z \in \mathfrak{H}$ so that $(\gamma \star x)\cdot z = z$,
we obtain
$$
\gamma'S_x \gamma^{-1} z = S_1 z.
$$
Let $w = \gamma^{-1} z$. Then
$$(z,S_1 z) = (\gamma w, \gamma' S_x w).$$
Hence $\vec{z} = (z,S_1 z) \in \mathfrak{H}_1 \cap \gamma \mathfrak{H}_x$.
This shows the first equality of (1).
Note that from the assumption that $\gamma \mathfrak{H}_x \neq \mathfrak{H}_1$, the element $\gamma \star x \notin k^\times$.
Observe that $(\gamma \star x)\cdot z = z$ if and only if the column vector $(z,1)^t$ is an eigen-vector of $\gamma \star x$.
This guarantees the second equality of (1).
\end{proof}
\begin{rem}\label{rem: CM-sup}
For non-zero $a \in A$, $x \in \Lambda_a$, and $\gamma \in \Gamma$ with $\gamma \mathfrak{H}_x \neq \mathfrak{H}_1$, let $t = \operatorname{Tr}(\gamma \star x)$.
Then we get $2(\gamma \star x)^\natural \in O_{B_1}^o$ with
$$
\big(2 (\gamma \star x)^{\natural}\big)^2
= \mathfrak{d}(t^2-4a).
$$
Moreover, for $z \in \mathfrak{H}$ with $(\gamma \star x) \cdot z = z$, we must have
$\big(2(\gamma \star x)^\natural\big) \cdot z = z$.
Thus the above lemma tells in particular that the intersection of $\mathcal{Z}_1$ and $\mathcal{Z}_x$, when $\mathcal{Z}_1 \neq \mathcal{Z}_x$, are supported by the image of the CM points of $\mathcal{C}_1$ with discriminant $\mathfrak{d}(t^2-4a)$ for some $t \in A$ with $t^2-4a \prec 0$.
\end{rem}
Observe that for $\gamma \in \Gamma$, $\gamma_1 \in \Gamma_1$, and $\gamma_x \in \Gamma_x$, one has
$$\tilde{\iota}((\gamma_1\gamma \gamma_x) \star x) = \tilde{\iota}(\gamma \star x) = \#(\Gamma_1 \cap \Gamma_{\gamma \star x}) \cdot \iota(\gamma \star x).$$
We are now able to express the intersection number $\mathcal{Z}_1\cdot \mathcal{Z}_x$ as follows:
\begin{thm}\label{thm: CN-Int}
Given $x \in \Lambda$ with $\det x \neq 0$. Suppose $\mathcal{Z}_1 \neq \mathcal{Z}_x$, or equivalently, $\gamma \star x \notin k$ for every $\gamma \in \Gamma$.
Then
\begin{eqnarray}
\mathcal{Z}_1\cdot \mathcal{Z}_x
&=&
2 \cdot \sum_{\gamma \in \Gamma_1\backslash \Gamma / \Gamma_x} \iota(\gamma \star x). \nonumber
\end{eqnarray}
\end{thm}
\begin{proof}
We have
\begin{eqnarray}
\mathcal{Z}_1 \cdot \mathcal{Z}_x &=&
\frac{q-1}{[\Gamma_1:\Gamma_1(\mathfrak{m})]\cdot[\Gamma_x:\Gamma_x(\mathfrak{m})]}
\cdot \sum_{\gamma \in \Gamma/\Gamma_F(\mathfrak{m})} \widetilde{\mathcal{Z}}_1 \cdot \gamma \widetilde{\mathcal{Z}}_x \nonumber \\
&=&
\frac{q-1}{[\Gamma_1:\Gamma_1(\mathfrak{m})]\cdot [\Gamma_x:\Gamma_x(\mathfrak{m})]}
\cdot \sum_{\gamma \in \Gamma_1(\mathfrak{m}) \backslash \Gamma / \Gamma_x(\mathfrak{m})} 2 \cdot \tilde{\iota}(\gamma \star x) \nonumber \\
&=&
\sum_{\gamma \in \Gamma_1\backslash \Gamma / \Gamma_x} \frac{q-1}{\#(\Gamma_1 \cap \Gamma_{\gamma \star x})} \cdot 2 \cdot \tilde{\iota}(\gamma \star x) \nonumber \\
&=&
2 \cdot \sum_{\gamma \in \Gamma_1\backslash \Gamma / \Gamma_x} \iota(\gamma \star x). \nonumber
\end{eqnarray}
\end{proof}
We now define the self-intersection number of $\mathcal{Z}_1$ (following \cite[p.\ 84]{H-Z}).
First, put
$$\text{vol}(X_1)\ := \ \frac{2}{[\widehat{\Gamma}_1:\Gamma_1]} \cdot
H^{\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}(0)
$$
and
$$
\text{vol}(\mathcal{Z}_1) \ := \ [\widehat{\Gamma}_1:\Gamma_1] \cdot \text{vol}(X_1) \ =\ 2H^{\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}(0).
$$
For each point $\boldsymbol{z} \in X_1$,
take a lift $\tilde{\boldsymbol{z}} \in \mathfrak{H}_1$, and let
$$r_{\boldsymbol{z}} := \frac{\#\big(\operatorname{Stab}_\Gamma(\tilde{\boldsymbol{z}})\big)}{\#\big(\operatorname{Stab}_{\widehat{\Gamma}_1}(\tilde{\boldsymbol{z}})\big)}.
$$
We set the following ``Plücker-type'' number:
$$
\mu_{\boldsymbol{z}}(X_1)
:= \frac{q-1}{\#(\operatorname{Stab}_{\Gamma}(\tilde{\boldsymbol{z}}))} \cdot \big(r_{\boldsymbol{z}}(r_{\boldsymbol{z}}-1)\big).
$$
\begin{defn}\label{defn: SIN}
The self-intersection number of $\mathcal{Z}_1$ is then defined to be:
\begin{eqnarray}
\mathcal{Z}_1\cdot \mathcal{Z}_1 := [\widehat{\Gamma}_1:\Gamma_1]^2 \cdot\left( - \text{vol}(X_1) + \sum_{\boldsymbol{z} \in X_1} \mu_{\boldsymbol{z}}(X_1)\right). \nonumber
\end{eqnarray}
\end{defn}
\begin{lem}\label{lem: SIN}
We may express the self-intersection number of $\mathcal{Z}_1$ as follows:
$$
\mathcal{Z}_1\cdot \mathcal{Z}_1 = 2 \cdot \sum_{\gamma \in \Gamma_1 \backslash \Gamma/\Gamma_1} \iota(\gamma \star 1).
$$
\end{lem}
\begin{proof}
Given $\gamma \in \Gamma$, notice that $\gamma \star 1 \in k$ if and only if $\gamma \in \widehat{\Gamma}_1$.
As $\Gamma_1$ is normal in $\widehat{\Gamma}_1$, one has
\begin{eqnarray}
2\cdot \sum_{\gamma \in \Gamma_1\backslash \Gamma/\Gamma_1} \iota(\gamma \star 1)
&=& 2\cdot \sum_{\gamma_1 \in \widehat{\Gamma}_1/\Gamma_1} \iota(\gamma \star 1)
+ 2 \cdot \sum_{\subfrac{\gamma \in \Gamma_1\backslash \Gamma/\Gamma_1}{\gamma \notin \widehat{\Gamma}_1}} \iota(\gamma \star 1) \nonumber \\
&=& [\widehat{\Gamma}_1:\Gamma_1]\cdot \big(-2H^{\mathfrak{d}^+\mathfrak{n}^+,\mathfrak{d}^-\mathfrak{n}^-}(0)\big)
+ 2 \cdot \sum_{\subfrac{\gamma \in \Gamma_1\backslash \Gamma/\Gamma_1}{\gamma \notin \widehat{\Gamma}_1}} \iota(\gamma \star 1). \nonumber
\end{eqnarray}
Hence the result holds if we show
\begin{eqnarray}\label{eqn: angle part}
[\widehat{\Gamma}_1:\Gamma_1]^2 \cdot \sum_{\boldsymbol{z} \in X_1}\mu_{\boldsymbol{z}}(X_1) \ = \ 2 \cdot \sum_{\subfrac{\gamma \in \Gamma_1\backslash \Gamma/\Gamma_1}{\gamma \notin \widehat{\Gamma}_1}} \iota(\gamma \star 1).
\end{eqnarray}
Take $\mathfrak{m} \in A_+$ with $\mathfrak{n}^2 \mid \mathfrak{m}$.
Adapting the proof of Propostion~\ref{prop: projection formula} we get
$$
[\widehat{\Gamma}_1:\Gamma_1]^2 \cdot \sum_{\boldsymbol{z} \in X_1}\mu_{\boldsymbol{z}}(X_1)
= \frac{q-1}{[\Gamma_1:\Gamma_1(\mathfrak{m})]^2}\cdot \sum_{\subfrac{\gamma \in \Gamma/\Gamma_F(\mathfrak{m})}{\gamma \notin \widehat{\Gamma}_1}} \widetilde{\mathcal{Z}}_1 \cdot \gamma \widetilde{\mathcal{Z}}_1.
$$
From Proposition~\ref{prop: Intersection} and Lemma~\ref{lem: Intersection}, we have
\begin{eqnarray}
\sum_{\subfrac{\gamma \in \Gamma/\Gamma_F(\mathfrak{m})}{\gamma \notin \widehat{\Gamma}_1}} \widetilde{\mathcal{Z}}_1 \cdot \gamma \widetilde{\mathcal{Z}}_1
&=& \sum_{\subfrac{\gamma \in \Gamma_1(\mathfrak{m})\backslash \Gamma/\Gamma_1(\mathfrak{m})}{\gamma \notin \widehat{\Gamma}_1}} 2 \cdot \tilde{\iota}(\gamma \star 1) \nonumber \\
&=&\frac{[\Gamma_1:\Gamma_1(\mathfrak{m})]^2}{q-1} \cdot
\sum_{\subfrac{\gamma \in \Gamma_1\backslash \Gamma/\Gamma_1}{\gamma \notin \widehat{\Gamma}_1}} \frac{q-1}{\#(\Gamma_1\cap \Gamma_{\gamma \star 1})} \cdot 2 \cdot \tilde{\iota}(\gamma \star 1) \nonumber \\
&=& \frac{[\Gamma_1:\Gamma_1(\mathfrak{m})]^2}{q-1} \cdot 2 \cdot
\sum_{\subfrac{\gamma \in \Gamma_1\backslash \Gamma/\Gamma_1}{\gamma \notin \widehat{\Gamma}_1}} \iota(\gamma \star 1). \nonumber
\end{eqnarray}
Therefore the equality~\eqref{eqn: angle part} follows and the proof is complete.
\end{proof}
For nonzero $a \in A$, consider the following \textit{Hirzebruch-Zagier-type divisor}
$$\mathcal{Z}(a) := \sum_{\Gamma \backslash \Lambda_a} \mathcal{Z}_x.$$
From Theorem~\ref{thm: AFCN}, Theorem~\ref{thm: CN-Int}, and Lemma~\ref{lem: SIN}, we finally arrive at:
\begin{cor}\label{cor: TINFC}
Given nonzero $a \in A$ and $y \in k_\infty^\times$ with $\deg a \leq 2 \operatorname{ord}_\infty(y)+2$, we have
$$\text{\rm vol}(O_{B_\AA}^\times/O_\AA)^{-1} \cdot I^*(a,y;\varphi_\Lambda) = \frac{|y|_\infty^2}{2} \cdot \sum_{x \in \Gamma \backslash \Lambda_a} \mathcal{Z}_1 \cdot \mathcal{Z}(a),
$$
and
$$
\text{\rm vol}(O_{B_\AA}^\times/O_\AA)^{-1} \cdot I^*(0,y;\varphi_\Lambda) = - \frac{|y|_\infty^2}{2} \cdot \text{\rm vol}(\mathcal{Z}_1).
$$
\end{cor}
\begin{rem}
From \textit{Remark}~\ref{rem: DTS-FC}, we may express the Fourier expansion of the Drinfeld-type automorphic form $\vartheta_\Lambda$ in terms of the corresponding intersection numbers: for $(x,y) \in k_\infty^\times \times k_\infty$,
$$
\vartheta_\Lambda\begin{pmatrix} y & x \\ 0&1\end{pmatrix}
= -\text{vol}(\mathcal{Z}_1)\beta_{0,2}(y) + \sum_{0 \neq a \in A}\big(\mathcal{Z}_1\cdot \mathcal{Z}(a)\big) \cdot \big(\beta_{a,2}(y)\psi_\infty(ax)\big).
$$
\end{rem}
| ad06f19292c9a7cdd4bbeff3b0379bd6bedaaab5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Dendrogram analysis}
In addition to visual identification of substructures towards the prestellar core G205-M3 (see Fig~\ref{fig2}, panel d),
we also show the identification of the substructures with the dendrogram analysis, which is often employed in identifying and labeling clumpy structures \citep{Rosolowsky2008,Pineda2015}.
We ran the dendrogram algorithm with min\_value =4$\sigma$ (minimum intensity considered in the analysis) and min\_delta= 2$\sigma$(minimum spacing between isocontours).
Two substructures are clearly identified and presented in Fig.~\ref{dendrogram} (red contours). The separation between the blobs, B1 \& B2 as obtained from the dendrogram analysis, based on the peak position of the blobs is 1148 AU. If we consider the position of the peaks of B1 and B2 as obtained from Gaussian fittings (presented in Table~\ref{tab:1}) then the separation is given by 1360 $\pm$ 140 AU, considering the fitting uncertainties. So, in general we consider the separation $\sim$ 1200 AU, as mentioned in the text.
The image presented in Fig.~\ref{dendrogram} corresponds to the combined data of ALMA (ACA, TM2 \& TM1 configurations) with a uv-taper of 150~k$\lambda$ ( $\sim$0\farcs8). The dust emission of the five dense prestellar cores which were detected in ALMA-TM2 ($\sim$1.2\arcsec), resolved out at the highest observing resolution (ALMA-TM1; $\sim$ 0\farcs3). So, to identify the substructures from the dust continuum images, we applied a range of uv-tapering on the combined ALMA data from 100~k$\lambda$ to 200~k$\lambda$ (corresponds to resolution from 0\farcs6 to 1.0\arcsec). The substructures of G205.46-14.56M3 appear most prominent
in ALMA combined data at $\sim$0\farcs8 - 1.0\arcsec resolution (with uv-taper of 100 - 150~k$\lambda$), but at higher resolutions (beyond 200~k$\lambda$) their emission was resolved out.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{dendro_plot.pdf}
\caption{ The red-contours show the condensations/blobs as identified using dendrogram analysis. White ellipses corresponds to FWHM size of the two component Gaussian fitting of the condensations B1 and B2. Dotted contours corresponds to 1.3mm dust continuum at 0.8$''$ resolution with the levels 4,7,10,13$\sigma$ where the rms noise is given by $\sigma=0.1$ mJy/beam.}
\label{dendrogram}
\end{figure}
\section{Introduction }
Stars form from the gravitational collapse of dense molecular cloud cores in the interstellar medium.
Characterization of the dense cores is therefore of great importance for gaining insights into the initial conditions and evolutionary stages in star formation. Starless cores, which are condensations of matter without any embedded young stellar objects (YSOs), are considered as the earliest phase of star formation. A subset of the starless cores, called prestellar cores, are gravitationally bound and will presumably collapse to form YSOs \citep{andre2014}. The evolution of prestellar cores to YSOs, and whether they form single or multiple stellar systems, however, is far from being understood.
While extensive efforts have been made previously to investigate the physical and chemical properties of starless and prestellar cores with single element telescopes \citep[e.g.,][]{Caselli2011}, the densest and presumably innermost 1000~au of such cores are yet to be characterized due to the limited angular, hence spatial, resolution.
This scale is however crucial for testing theories of fragmentation to form multiple systems.
To this end,
survey observations of starless cores in nearby molecular clouds (including, for example, Perseus, Ophiuchus, and Chamaeleon) with modern (sub)millimeter interferometers, such as the Combined Array for Research in Millimeter-wave Astronomy (CARMA), the Submillimeter Array (SMA), and the Atacama Large Millimeter and submillimeter Array (ALMA) were conducted \citep[e.g][]{Schnee2010,Schnee2012a,Dunham2016,Kirk2017}.
In these survey observations, very few localized sources were recovered
because of the insensitivity of interferometers to small fluctuations in extended emission.
Among those detected sources, some turned out to be protostellar, while others remained starless with no additional substructure in their (sub)millimeter continuum emission.
Such results suggest that the density profiles of starless cores are
predominantly flat,
consistent with the earlier suggestions \citep[e.g.,][]{Ward-Thompson1994,Ward-Thompson1999,Shirley2000,Caselli2019}.
A profile having nearly constant density
at small radii ($\sim$ few 1000 au)
is reminiscent of the ``flat zone" in Bonnor-Ebert spheres \citep{Ebert1955,Bonnor1956}.
Meanwhile, observations indicate that the multiplicity fraction and the companion star fraction are highest in Class 0 protostars and decrease in more evolved protostars \citep{chen2013,Tobin2016}. This implies that multiple systems may develop at an even earlier phase.
Indeed, some theoretical works suggest that substructures in prestellar cores can be produced by turbulent fragmentation at a scale of 1000 au, which will form wide-multiple stellar systems \citep{Offner2010}, and these should be visible using sensitive interferometers \citep{Offner2012}.
However, none of the observational surveys described in the previous paragraph found secondary/multiple substructures within dense cores at a scale of 1000 au.
Although a few studies report cases of substructure and fragmentation in specific starless cores \citep[e.g.,][]{Kirk2009,chen2010,Nakamura2012,Takahashi2013,Pineda2015,Friesen2014,ohashi2018,Tatematsu2020}
they are associated with
high-density condensation,
and not related to the substructures (1000 au scale) in the central dense region of prestellar cores.
In a recent ALMA survey, \citet{Tokuda2020} reported the presence of very low mass ($\sim 10^{-2}$ M$_\odot$ ,
$n_{\mbox{\scriptsize H$_2$}}$ $\sim 10^{5}~$cm$^{-3}$) substructure toward starless cores at a scale of 1000 au; these substructures are not
massive enough to produce multiple stars. In another observational effort to study the inner dense region of the prestellar core L1544 \citep{Caselli2019},
no substructures were detected inside the central dense region ($\sim 2000$ au with $n_{\mbox{\scriptsize H$_2$}}$ $\geq 10^{6}$cm$^{-3}$) of the core, where stellar multiplets are expected to potentially develop.
Recently, we performed a survey \citep[ALMA Survey of Orion Planck Galactic Cold Clumps or ALMASOP;][]{Dutta2020} with ALMA, in which we targeted 72 cores toward the Orion Molecular Clouds \citep[distance $\sim $ 400 pc;][]{Kounkel2018}.
An overview of the ALMASOP is presented in \citet{Dutta2020}. Among the 72 targets in ALMASOP observations, 23 were previously classified as (candidate) starless cores.
While all these starless cores were detected in the SCUBA-2 observations (with a core scale of $\sim 0.1$ pc or 20000 au), only 16 of them are detected by the ALMASOP ACA observations (scale $\sim 0.03$ pc or 6000 au).
In this letter, we highlight the detection of a central compact dense structure toward a sample of five cores and the discovery of substructures within one of them using the ALMA 12-m array.
\section{Observations\label{sec:Obs}}
The observations of ALMASOP (project ID:2018.1.00302.S. ; PI: Tie Liu) were carried out with ALMA in Cycle 6 toward 72 fields during 2018 October to 2019 January.
The observations were executed in four blocks in three different array configurations: 12m C43-5 (TM1), 12m C43-2 (TM2), and 7m ACA, resulting in a resolution ranging 0\farcs34 - 5.5\arcsec in the 1.3 mm band.
The correlator was configured into four spectral windows with 1.875 GHz bandwidth, which provides a spectral resolution of 1.129 MHz,
corresponding to velocity resolutions between 1.465 and 1.563 kms$^{-1}$.
We adopted this coarse velocity resolution to facilitate efficient continuum observations and to maximize the spectral line coverage.
The spectral set-up covers the continuum emission at 233.0 GHz and 216.6 GHz, and offers
simultaneous coverage of the molecular lines CO (2-1), \chem{C^{18}O} (2-1), \chem{N_2D^+} (3-2), SiO (5-4) and other hot corino tracers as well \citep[see][]{Hsu2020}.
The remaining details of the observational parameters are presented in \citet{Dutta2020}.
The calibrated visibility data were obtained using CASA 5.4 \citep[Common Astronomy Software Applications package;][]{McMullin2007} pipeline script as delivered by the observatory. The visibility data for different configurations and executions corresponding to the 72 sources were then separated into continuum and spectral data, and imaged jointly. The 1.3~mm continuum images of the sources are generated through CASA’s tclean task with the `automask’ on, the hogbom deconvolver, and a robust weighting of 0.5.
\begin{figure*}[ht!]
\centering
\includegraphics[width=18cm]{fig1.pdf}
\caption{The central dense region of the prestellar cores G208.68-19.20N2 (top-left), G209.29-19.65S1 (top-right), G209.94-19.52N (bottom-left), and G212.10-19.15N1 (bottom-right). The contours represent the 1.3 mm continuum observed by the ACA with levels 5,10,15$\sigma$..., where $\sigma$=3.4, 1.9, 2.5, 1.6~mJy/beam, respectively for the sources; and the color scales show the ALMA TM2 detection of the centrally dense region.
Beam sizes (typically 1.2") for each source are shown in the lower left of each panel.}
\label{fig1}
\end{figure*}
\section{Results}
\subsection{Detection of compact dense structure inside prestellar cores}
We present in Fig.~\ref{fig1} and Fig.~\ref{fig2} (c) the detection of the 1.3~mm (dust) continuum emission in contours toward the five targets (G208.68-19.20N2, G209.29-19.65S1, G209.94-19.52N, G212.10-19.15N1, and G205.46-14.46M3 (G205-M3 hereafter)) with the ALMA ACA+TM2 configurations at a resolution of $\sim$ 1\farcs2.
As introduced in Section 1, 16 of the 23 candidate starless cores were detected by the ACA observations at a resolution of $\sim$ 6\arcsec and the 5 targets are among the 16 cores.
Compared with the other 11 detections, the 5 cores discussed in this paper
have further compact dense features within the structures seen by the ACA.
However, their dust emission is not like the YSOs, which have point-like compact emission features as imaged using ALMA-TM1 configuration ($\sim$ 0\farcs3).
None of the five cores displays an outflow signature when observed in CO ($J$=2-1), its isotopologue lines and SiO ($J$=5-4) \citep{Dutta2020}.
Neither near-infrared nor mid-infrared emission is present within these targets based on our archival search\footnote{\url{(https://irsa.ipac.caltech.edu/irsaviewer/)}}.
Meanwhile, toward the central region of these five cores, CO appears fully depleted while \chem{N_2D^+} traces well the 1.3 mm (dust) continuum (see Fig.~\ref{n2dp}).
This is in accordance with their intense \chem{N_2D^+} emission and high [\chem{N_2D^+}]/[\chem{N_2H^+}] abundance ratio of $\geq 0.1$ seen by the NRO 45m telescope at a larger ($\sim 0.05$ pc to 0.1 pc) scale \citep{Kim2020}.
All these lines of evidence suggest that the five cores are genuinely chemically evolved prestellar cores. Note that, all of these sources were earlier classified as `starless' by \citet{Yi2018}, although \citet{Kirk2016} using JCMT survey results classified G205-M3 as YSO because of the presence of a YSO \citep{Megeath2012,Stutz2013} within the very extended ($\sim 65$\arcsec) envelope surrounding the core. The nearest YSO is $\sim 27$\arcsec\ away from the position of G205-M3 core (see Fig~\ref{fig2}, panel 1), and therefore not directly associated.
The flux density of the observed structure toward each of the five cores can be estimated from 2-dimensional (2D) Gaussian fitting.
We have neither gas nor dust temperature measurements on the observed angular scales.
Nevertheless, given their prestellar nature, we assume the temperature of the compact component is at most 10~K and can go as low as 6.5~K \citep{crapsi2007,keto2010}.
By adopting a (dust) specific absorption coefficient (per mass, with a gas-to-dust mass ratio of 100) $\kappa$ = 0.009 cm$^2$ g$^{-1}$ at the observing wavelength for grains with thick icy mantles after 10$^5$~yr of coagulation \citep{Ossenkopf1994}, we infer the source masses.
The effective radius of the cores ($R_{\mbox{\scriptsize{core}}}$) is estimated as $\rm{\sqrt{(major \times minor)}/2}$, where `major' and `minor' correspond to the two axes of the ellipse obtained from the 2D Gaussian fitting, and
the gas column density can be found as $N_{\mbox{\scriptsize H$_2$}} = M/(\pi \mu_{\mbox{\scriptsize H$_2$}} m_{\mbox{\scriptsize H}} R_{\mbox{\scriptsize{core}}}^2 )$.
By assuming a spherical geometry, the gas volume density can also be calculated from $n_{\mbox{\scriptsize H$_2$}} = 3M/(4\pi \mu_{\mbox{\scriptsize H$_2$}} m_{\mbox{\scriptsize H}} R_{\mbox{\scriptsize{core}}}^3)$, where $ \mu_{\mbox{\scriptsize H$_2$}}$ (=2.8) is the
molecular weight per hydrogen molecule \citep{Kauffmann2008}, and $m_{\mbox{\scriptsize H}}$ is the proton mass. The mass, size, column density, and volume density of the cores inferred from both the ACA observations and the ACA+TM2 observations
are presented in Table~\ref{tab:1}.
\begin{figure*}[ht!]
\centering
\includegraphics[width=18cm]{fig2.pdf}
\caption{The structure of G205.46-14.56M3 as revealed at different wavelengths and scales. Panel (a) shows 1.3 mm continuum observed by SCUBA-2 in contours: 0.3,0.5,0.7,0.9$\times$peak (where peak = 671.4 mJy/beam corresponds to the peak flux of closest protostellar core) overplotted on the 70 $\mu$m mid-infrared image taken by Herschel PACS. Panel (b) shows the SCUBA-2 850 $\mu$m continuum in contours (6$\sigma$, $\sigma =38$ mJy/beam) and the ACA 1.3 mm continuum map in color. Both beams are shown in left bottom corner.
Panel(c) shows the highly dense central region as revealed by the ALMA ACA+TM2 dust continuum map (synthesized beam $\sim 1.2''$). Plotted in contours are the ACA continuum with levels 3,5,10,15$\sigma$.., where $\sigma =1.6$ mJy/beam. Panel (d) zooms toward the central region at a higher resolution (synthesized beam $0.8''$) where the substructures (B1 \& B2) are visible. The contours are at 4,6,10 and 13 $\sigma$ levels where $\sigma$ = 0.1 mJy/beam. The ellipses correspond to the FWHM sizes inferred from the two-component Gaussian fitting. }
\label{fig2}
\end{figure*}
\begin{deluxetable*}{ccccccccccc}
\tabletypesize{ \scriptsize}
\tablecaption{ALMA-ACA and ACA+TM2 configuration 1.3 mm continuum results of five presetellar cores \label{tab:1}}
\tablewidth{2pt}
\tablehead{
\colhead{Source } & \colhead{RA}& \colhead{DEC} &
\colhead{FWHM} & \colhead{ S$_\nu$(1.3 mm)} &
\colhead{M$_{gas}$} &
\colhead{n$_{H2}$} & \colhead{N$_{H2}$}& \colhead{Diameter} & \colhead{$L_J$}\tablenotemark{a} & \colhead{$\alpha$ (FWHM)} \tablenotemark{b}\\
\colhead{} & \colhead{(J2000)}& \colhead{(J2000)} &
\colhead{($\arcsec$)} & \colhead{ (mJy)} &
\colhead{(M$_{\sun}$)} &
\colhead{(cm$^{-3}$)} & \colhead{(cm$^{-2}$)}& \colhead{(au)} & \colhead{(au)} & \colhead{ - (kms$^{-1}$)} }
\startdata
\multicolumn{11}{c}{\emph{ACA results}}\\
G205.46-14.56M3* & 05:46:05.99 & -00:09:32.37 & $8.4\times7.0$ & 79.8 & 0.76-1.69 & $6.4-14.2 \times 10^6$ & $1.9-4.3 \times 10^{23}$ & 3067 & 1561 & 0.68 (0.40)\\
G208.68-19.20N2 & 05:35:20.72 & -05:00:54.09 & $26.7\times7.7$ & 724.0 & 6.93-15.36 & $8.9-19.7\times10^6$ & $5.1-11.3\times10^{23}$ & 5735 & 1325 & 0.17 (0.49)\\
G209.29-19.65S1 & 05:34:56.04 & -05:46:05.28 &$21.5\times7.5$ & 266.0 & 2.55-5.64 & $4.7-10.4\times10^6$ & $2.4-5.4\times 10^{23}$ & 5087 & 1826 & 1.12 (0.97)\\
G209.94-19.52N & 05:36:11.39 & -06:10:45.93 &$14.2\times 7.2$ & 129.1 & 1.24-2.74 & $4.5- 9.9\times 10^6$ & $1.8-4.0\times10^{23}$ & 4059 & 1868 & 0.83 (0.57)\\
G212.10-19.15N1 & 05:41:21.27 & -07:52:27.01 &$11.6\times6.0$ & 47.1 & 0.45-1.00 & $2.9-6.5\times 10^6$ & $1.0-2.2\times10^{23}$ & 3337 & 2306 & 1.80 (0.55)\\
\hline
\multicolumn{11}{c}{\emph{Combined ACA+TM2 results}} \\
G205.46-14.56M3* & 05:46:05.96 & -00:09:32.45 & $6.0 \times 4.8$ & 53.2 & 0.51-1.13 & $1.2 -2.8\times 10^7$ & $2.7 - 5.9 \times 10^{23}$ & 2146 & 1119 & 0.71 (0.40) \\
G208.68-19.20N2 & 05:35:20.76 & -05:00:55.21 & $16.3 \times 3.6$ & 325.0 & 3.11-6.89 & $2.6-5.7 \times 10^7$ & $7.9 -17.0 \times 10^{23}$ & 3087 & 781 & 0.21 (0.49)\\
G209.29-19.65S1 & 05:34:55.84 & -05:46:04.81 & $7.8 \times 3.9$ & 98.6 & 0.94 - 2.09 & $2.1 -4.7\times 10^7$ & $4.7-10.0 \times 10^{23}$ & 2205 & 856 & 1.32 (0.97)\\
G209.94-19.52N & 05:36:11.38 & -06:10:45.65 & $10.7 \times 6.5$ & 89.3 & 0.86-1.89 & $0.6-1.2 \times 10^7$ & $1.8 - 4.1 \times 10^{23}$ & 3344 & 1680 & 0.99 (0.57)\\
G212.10-19.15N1 & 05:41:21.28 & -07:52:27.50 & $7.7 \times 4.2$ & 31.8 & 0.30 - 0.67 & $0.6 - 1.4 \times 10^7$ & $1.4 - 3.2 \times 10^{23}$ & 2264 & 1568 &1.83 (0.55)\\
\hline
\multicolumn{11}{c}{\emph{Substructure in G205-M3}}\\
B1 &05:46:06.008 & -00.09.32.812 & $5.5\times3.5$ & 39.4 & 0.38-0.84 & $1.9 - 4.2 \times 10^7$ & $3.22 - 7.14\times 10^{23}$ & 1755 & 1014 & 0.79 (0.40) \\
B2 &05:46:05.795 & -00.09.31.659 & $2.8\times1.5$ & 8.8 & 0.08-0.19 & $3.7 - 8.2 \times 10^7$ & $3.03 - 6.71\times 10^{23}$ & 820 & 730 & 1.76 (0.40)
\enddata
\tablenotetext{}{ RA and DEC correspond to the peak positions of 2D Gaussian fitting. The range for M$_{gas}$, n$_{H2}$, N$_{H2}$ corresponds to the estimation, assuming \mbox{$T_K$}= 10~K and 6.5~K respectively.}
\tablenotetext{*}{The source name in JCMT survey \citep{Yi2018} is G205.46-14.56N1.}
\tablenotetext{a}{Jeans lengths (L$_J$) calculated at 10 K; at 6.5~K the values will be lower}
\tablenotetext{b}{{ Virial parameters calculated} at 10 K with FWHMs obtained by \citet{Kim2020}; at 6.5~K the values will be lower; see section 4.3 for details }
\end{deluxetable*}
\subsection{Substructures in the prestellar core G205-M3}
G205-M3 is the only core where we detect substructures at a scale of 1000 au inside the compact dense structure of the core (also see section 4.1.1).
We show in Fig.~2 the (dust) continuum images of the G205-M3 core.
Different $uv$--tapering of the visibility data has been employed to highlight the intricate features of the core seen at different angular scales.
The G205-M3 core size in SCUBA-2 observation (beam size $\sim 14''$) is on the order of 0.05 pc or 10000 au \citep{Yi2018}.
The core is detected by ACA (beam $\sim 6\arcsec$, see Fig. 2(b)) with an overall size on the order of 4000 au and moreover, the faint substructure can be readily discerned.
At a higher (1\farcs2) resolution (see Fig. 2(c)) an asymmetric structure can be seen with an average core size of $\sim 2000$ au.
Further zooming into the source with an angular resolution of 0\farcs8 (
corresponding to a uv-taper of 150~k$\lambda$; Fig.~2-d), the inner 2000~au region of G205-M3 is then resolved into two noticeable substructures.
These substructures are named B1 and B2 ( see Fig~\ref{fig2} -panel d).
These substructures were also identified using a dendrogram analysis, see Fig.~\ref{dendrogram}.
They were fitted simultaneously with 2D Gaussian fitting.
The two components are, respectively, 5\farcs5 $\times$ 3\farcs5\ and 2\farcs8 $\times$ 1\farcs5 in size and separated roughly by 1200 au (3\arcsec).
The masses of B1 and B2 are 0.38-0.84 M$_\odot$ and 0.08-0.19 M$_\odot$, respectively, if a (dust) temperature range of 6.5-10~K is assumed, with larger masses associated with the lower temperatures.
Their corresponding volume densities, considering a spherical geometry, are listed in Table~\ref{tab:1}.
\begin{comment}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{dendro_plot.pdf}
\caption{ The red-contours show the condensations/blobs as identified using dendrogram. White ellipses corresponds to FWHM size of the two component Gaussian fitting of the condensations B1 and B2. Dotted contours corresponds to 1.3mm dust continuum at 0.8$''$ resolution with the levels 4,7,10,13$\sigma$ where $\sigma$(rms)=0.1 mJy/beam.}
\label{dendrogram}
\end{figure}
\end{comment}
\begin{figure*}
\centering
\includegraphics[width=18cm]{fig3_n2dp.pdf}
\caption{ Integrated \chem{N_2D^+} emission in contours overplotted on the ALMA ACA+TM2 dust continuum map.
The contours are 3$\sigma$, 5$\sigma$, 7$\sigma$..., where $\sigma$ (rms)=20 and 17 mJy/beam~kms$^{-1}$\ for G205.46-14.56M3 and G212.10-19.15N1 respectively. For G208.68-19.20N2, G209.29-19.65S1, G209.94-19.52N the rms ($\sigma$)=32,20,21 mJy/beam~kms$^{-1}$, respectively and contours are 5,10,15$\sigma$. }
\label{n2dp}
\end{figure*}
\section{Discussion}
Summarizing the results, a series of observations of increasingly better resolution and sensitivity has revealed increasingly compact, denser structures in a small subset of starless cores. In this section, we discuss the physical properties, the stability, and the dynamical state of these structures.
\subsection{Detection of compact structures
In the ACA/ALMA maps, the flux densities of G205-M3 enclosed in apertures with an equivalent radius
$r$ of 0\farcs4, 1\farcs1, 2\farcs2, 5\farcs7, and 14\farcs9 centered at the peak position of B1 are 1.4~mJy, 7.5~mJy, 24.8~mJy, 52.2~mJy,and 99.5~mJy respectively. Assuming the gas temperature and dust property are uniform within these scales, we find the corresponding enclosed masses will scale as $r^{1.2}$, which hints at a density profile of $r^{-1.8}$ between 300~au and 10000~au scale.
Using similar method, for other cores (except G208-N2) the masses scale as $r^{1.2} - r^{1.3}$, and therefore their density profiles vary as
$r^{-1.7} - r^{-1.8}$.
The actual density profile could be even steeper if the dust temperature decreases toward the center of the core.
These structures are thus not consistent with Bonnor-Ebert spheres unless the inner region of constant density is very small \citep{Dunham2016}.
Such systems are generally very close to the initiation of collapse or they have begun to collapse.
Based on their statistical analysis on the lifetime of starless cores, \citet{Dunham2016} and also \citet{Kirk2017} argued that only evolved cores with short lifetime ($\sim 10^4$ year) can be detected by interferometers. Therefore, the five cores detected at high resolution ($\sim$ 1\farcs2) likely represent an evolved starless stage, just before the onset of star formation.
Such dense regions can be compared with the well known prestellar core `kernel' in L1544 \citep{Caselli2019}.
The average and peak densities of the five prestellar cores are at least 10 times higher than that of the L1544 core, although the sizes of the regions are comparable.
\subsubsection{Search for substructures at 1000~au scale}
Do the five cores have substructures inside the compact dense regions?
We use different methods, e.g., uv-tapering of the image, slicing along possible substructure features to
identify the intensity variation, and the dendrogram technique. This helps us identify the real and prominent substructures which are not related to
imaging artifacts \citep[e.g.,][]{Caselli2019} and weak substructure features. Prominent signatures
of substructures are found only towards G205-M3 (section 3.2). Among the other four cores, the G208-N2
core appears elongated and similar to a filamentary structure (More details of this core will be presented
by Hirano et al. in preparation).
However, the intensity fluctuations along the apparent substructures are at most comparable to the noise level, therefore we see no evident signs of substructure towards G208-N2. Similarly, we do not find
substructures toward the other remaining cores, and their detailed physical evolutionary status and density structure will be described in a future work (Sahu et al. in preparation).
\subsection{Thermal stability}
Assuming first an isothermal molecular core, being supported only by its thermal pressure against gravity without considering non-thermal gas motions, one can estimate the Jeans length scale beyond which gravitational collapse would prevail.
The Jeans length is defined by: $ L_J=\sqrt{\frac{\pi {C_s}^2}{G \rho_0}}$,
where $G$ is the gravitational constant, $\rho_0$ is mass density and \chem{C_s} is the isothermal sound speed, \chem{C_s = (k_B T/\mu_p m_H)^{0.5}}, with $\mu_p = 2.37$ \citep{Kauffmann2008}.
\citet{Yi2018} found that at the scale of the SCUBA-2 observation, these five cores are unstable based on the Jeans analysis.
We find similar results for the dense structures seen at the ACA (6$''$) and TM2($\sim 1.2''$ scales (see Table~\ref{tab:1}).
The dense structures in the five cores have sizes significantly greater than their correspondingly Jeans lengths, implying that they are Jeans unstable.
The adopted gas temperature is a major uncertainty in the above analysis.
While we have adopted a gas temperature of 10~K for all sources, \citet{Kim2020} found from \chem{N_2H^+} and \chem{N_2D^+} line observations that under LTE conditions the gas temperatures of the five cores are in the range of 10.8-17.3~K at larger scales with an average density ($\sim 10^5$cm$^{-3}$).
Nevertheless, the cores remain unstable even if we adopt a gas temperature of 20~K.
In fact,
the temperatures of the central compact dense regions are likely to be lower than 10~K given their high density.
A good example is L1544, for which \citet{Caselli2019} assumed a core temperature to be 6.5 K for its high density ($\sim 10^6$cm$^{-3}$).
The average density of the central dense regions in our core sample is at least ten times higher than L1544 and a gas temperature of 6.5 K is therefore viable.
At this temperature all the core structures detected by ACA are Jeans unstable, further highlighting their prestellar nature.
\subsection{Dynamical state}
We particularly focus on the dynamical state towards the G205-M3 core, where substructures are detected.
To assess the dynamical state of the regions, we use molecular line data to determine thermal and non-thermal contributions, using \chem{{\sigma_v}^2={\sigma_{th}}^2 +{\sigma_{nt}^2}}, where \chem{\sigma_{th} = (k_B T/m)^{0.5}}, with $m$ the mass of the observed molecule, and \chem{\sigma_{nt}} are the thermal and non-thermal components, respectively. The non-thermal contributions could include infall, rotation, and streaming motions, but they are usually attributed to turbulence. Ideally, we would use the line data on the scales of the substructures, but our velocity resolution is insufficient. We instead consider information from NH$_3$ with an FWHM linewidth (\mbox {$\Delta v$}) of 0.88 kms$^{-1}$ \citep{Cesaroni1994}
on a scale of 40 arcseconds ($2\ee4$ au)
and from N$_2$D$^+$\ with $\mbox {$\Delta v$} = 0.4$ kms$^{-1}$\ on a scale of 8000 au. Using $\mbox {$\Delta v$} = (8 ln 2 )^{0.5}\rm{\sigma_v}$, we obtain \chem{\sigma_v = 0.37} kms$^{-1}$\ for NH$_3$ and 0.17 kms$^{-1}$\ for N$_2$D$^+$. If we assume $\mbox{$T_K$} = 10$ K, \chem{\sigma_{th} = 0.068} kms$^{-1}$\ for NH$_3$ and 0.053 kms$^{-1}$\ for N$_2$D$^+$, resulting in \chem{\sigma_{nt} = 0.37} kms$^{-1}$\ for NH$_3$ or \chem{\sigma_{nt} = 0.16} for N$_2$D$^+$, with the smaller value more likely relevant to the individual cores and indicating nearly equal contributions from turbulence and thermal broadening. The isothermal sound speed, \chem{C_s}
is 0.19 kms$^{-1}$\ at $\mbox{$T_K$} = 10$ K, indicating that the turbulence is supersonic,
with Mach number of 2 based on NH$_3$ or transonic, with Mach number of 0.8 based on N$_2$D$^+$. Taken at face value, these results suggest that the turbulence has decayed on small scales. The effective sound speed, \chem{c_{s,eff} = (C_{ s}^2 + \sigma_{nt}^2)^{0.5}}, is 0.247 kms$^{-1}$, using the data from N$_2$D$^+$.
The fate of the substructures can be assessed from their virial parameters. The parameter $\alpha_{\rm vir} = 2a E_{\rm k}/|E_{\rm g}|$. where $E_{\rm k}$ is the kinetic energy and $E_{\rm g}$ is the gravitational potential energy, and $a = 2\pm1$ for a wide range of geometries and density structures \citep{Kauffmann2013}.
The virial parameter can be written as $\alpha_{\rm vir} = 5.6\ee{-3} a c_{\rm s, eff}^2 R(\rm au) M(M_\odot)^{-1}$. Taking the sizes and masses and the effective sound speed of 0.247 kms$^{-1}$, we obtain $\alpha_{\rm vir} = 0.36a$ to $0.79a$ for B1 and $\alpha_{\rm vir} = 0.74a$ to $1.76a$ for B2. This calculation suggests that the substructures are close to being gravitationally bound, even if all the non-thermal component is entirely due to turbulence. Other motions, such as infall, may be present. It is at least plausible that the sub-structures are collapsing to form separate objects. Using a similar calculations, we found that the four other dense cores are also close to being gravitationally bound (see Table~\ref{tab:1}). This again supports the evolved prestellar nature of the cores.
\subsection{Substructures toward G205-M3 and stellar multiplicity}
We have detected, for the first time, substructures within the central compact dense region ($\sim$ 2000 au) of the prestellar core G205-M3. What will be the fate of these kinds of substructures?
If the substructures are in free-fall, they are likely to form separate objects that could be members of a wide binary. They are unlikely
to coalesce before they collapse. The minimum time to coalesce is the crossing time. Using the projected separation of 1148 au ( also see the appendix section) and the effective sound speed of 0.247 kms$^{-1}$, we calculate $t_{cross} = 5.0$ yr $s/c_{\rm s, eff} = 2.3\ee4 $ yr, where $s$ is the projected separation in au and $c_{\rm s, eff}$ is in kms$^{-1}$. This estimate assumes that there is no separation along the line of sight and that they are headed on a collision course. In contrast, the free- fall time is
$\sqrt{3\pi/32 \rho_0 G }$ (where, $\rho_0=n_{H_2}\mu_{H_2}m_{H}$ with $\mu_{H_2}$=2.8), for a minimum H$_2$ number density (n$_{H_2}$) of 1.9\ee7 cm$^{-3}$, $t_{\rm ff} < 7.06\ee3$ yr, which is much shorter than $t_{cross}$. On the other hand, a binary system will be bound if the internal energy: $E_{int}=1/2\mu v^2 - G m_1 m_2/r$ $< 0$, where $\mu$ is the reduced mass of the two bodies. Considering the mass of B1 and B2 substructures/sub-cores, the system will be bound for a velocity difference ($v$) of 0.8-1.2 kms$^{-1}$. From the low resolution spectral data ( \chem{N_2D^+}), we find that the velocity difference is at most 1.465 kms$^{-1}$. Therefore, it is very plausible that the sub-cores may eventually form a wide binary system. Notably, \citet{Karnath2020} reported substructures associated with protostellar candidates. Those substructures appear to trace an early stage of protostellar evolution, during which substructures are associated with collapsing fragments and individual components may be optically thick hydrostatic cores. The detected structures toward the prestellar core (G205-M3) are therefore not some transient features; they likely persist into the protostellar phase as well.
\subsection{Physical explanations for fragmentations}
In the `gravoturbulent fragmentation’ \citep{Palau2018} scenario fragmentation takes place in a self-gravitating turbulent medium. In this case, the density is determined by enhancements created by turbulence \citep[e.g.,][]{Padoan2002,Fisher2004,Goodwin2004,Offner2010}. Based on this theory, \citet{Offner2012} predicted that prestellar core fragmentation can be observable at 1000 au scale. It is plausible that fragmentation towards the G205-M3 core represents such a case.
However, the non-detections of sub-cores in other prestellar cores possibly imply a younger stage than the G205-M3 core, and they may fragment in a later period of their evolution.
\section{Summary}
We present 1.3 mm dust continuum observations using different configurations of ALMA, resulting in different synthesized beams to study five highly dense ($> 10^7$cm$^{-3}$) prestellar cores in the Orion molecular cloud. We found that in addition to detection using the ALMA-ACA-configuration ($6''$), these cores are also detected using the ALMA-TM2-configuration ($1.2''$) which imply that the cores have a centrally dense region of size $\sim 2000$ au.
No NIR/MIR emission has been detected towards these cores, signifying that the cores are starless/prestellar in nature.
The cores are found to be gravitationally unstable, and at the onset of star formation
We found two substructures of sizes ranging from 800-1700 au and masses $ >0.08-0.84$ M$_\odot$ towards the core G205-M3. Considering that the free fall time is shorter than the coalescence time of the substructures, and they are likely bound within a separation of $\sim$ 1200 au, we speculate
that this core will produce a wide binary or multiple star system.
\acknowledgments
\input{ACKNOWLEDGMENT}
| c8b5c89f667d7093f8279c03890b68990dc1c8eb | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Author Biographies}
\begingroup
\setlength\intextsep{0pt}
\begin{minipage}[t][6.3cm][t]{1.0\textwidth}
\begin{wrapfigure}{L}{0.25\textwidth}
\includegraphics[width=0.25\textwidth]{john_smith.eps}
\end{wrapfigure}
\noindent
{\bfseries John Smith} received his BSc (Mathematics) in 2000 from The University of Maryland. His research interests include lasers and optics.
\end{minipage}
\begin{minipage}{1.0\textwidth}
\begin{wrapfigure}{L}{0.25\textwidth}
\includegraphics[width=0.25\textwidth]{alice_smith.eps}
\end{wrapfigure}
\noindent
{\bfseries Alice Smith} also received her BSc (Mathematics) in 2000 from The University of Maryland. Her research interests also include lasers and optics.
\end{minipage}
\endgroup
}{}
\end{document}
| adebe1ff4826c447ba96473adc4b570af79eb92a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Acknowledgments}
\label{sec:acknowledgments}
The authors thank Justin Owen for his insightful feedback.
This paper describes objective technical results and analysis. Any subjective views or opinions that might be expressed in the paper do not necessarily represent the views of the U.S. Department of Energy or the United States Government.
Sandia National Laboratories is a multimission laboratory managed and operated by National Technology and Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International, Inc., for the U.S. Department of Energy's National Nuclear Security Administration under contract DE-NA-0003525.
\section{Implication of Matrix-Element Coding Error}
\label{sec:app}
The purpose of this appendix is to illustrate the implications of a coding error limited to a matrix element.
From~\eqref{eq:zjv_mms}, the original system of equations is
\begin{align}
\mathbf{Z}_\text{MS} \mathbf{J}^h = \mathbf{V}_\text{MS}.
\label{eq:correct}
\end{align}
Introducing a coding error $\delta$ in element $(i,j)$ of $\mathbf{Z}_\text{MS}$, the erroneous matrix is
\begin{align}
\tilde{\mathbf{Z}}_\text{MS} = \mathbf{Z}_\text{MS} + \delta \mathbf{e}_i\mathbf{e}_j^T,
\label{eq:incorrect_Z}
\end{align}
where $\mathbf{e}_i$ denotes the standard unit vector with zero-valued elements, except at element $i$, where the value is unity.
Inserting~\eqref{eq:incorrect_Z} into~\eqref{eq:correct} yields the modified equation
\begin{align}
\tilde{\mathbf{Z}}_\text{MS} \tilde{\mathbf{J}}^h = \mathbf{V}_\text{MS}.
\label{eq:incorrect}
\end{align}
From~\eqref{eq:error}, the error in~\eqref{eq:correct} is
\begin{align}
\|\mathbf{J}^h - \mathbf{J}_n \|\le Ch^p,
\label{eq:correct_error}
\end{align}
whereas the error in~\eqref{eq:incorrect} is
\begin{align}
\|\tilde{\mathbf{J}}^h - \mathbf{J}_n\| \le \tilde{C}h^{\tilde{p}}.
\label{eq:incorrect_error}
\end{align}
Using the Sherman--Morrison formula,
\begin{align*}
\tilde{\mathbf{J}}^h
= \tilde{\mathbf{Z}}_\text{MS}^{-1} \mathbf{V}_\text{MS}
= \left(\mathbf{Z}_\text{MS}^{-1} - \frac{\delta\mathbf{Z}_\text{MS}^{-1}\mathbf{e}_i\mathbf{e}_j^T\mathbf{Z}_\text{MS}^{-1}}{1+\delta\mathbf{e}_j^T\mathbf{Z}_\text{MS}^{-1}\mathbf{e}_i}\right)\mathbf{V}_\text{MS}
= \mathbf{J}^h - \frac{\delta\mathbf{Z}_\text{MS}^{-1}\mathbf{e}_i\mathbf{e}_j^T \mathbf{J}^h}{1+\delta\mathbf{e}_j^T\mathbf{Z}_\text{MS}^{-1}\mathbf{e}_i},
\end{align*}
such that
\begin{align}
\|\tilde{\mathbf{J}}^h - \mathbf{J}^h\|
\le
\frac{\|\delta\mathbf{Z}_\text{MS}^{-1}\mathbf{e}_i\||\mathbf{e}_j^T \mathbf{J}^h|}{|1+\delta\mathbf{e}_j^T\mathbf{Z}_\text{MS}^{-1}\mathbf{e}_i|}.
\label{eq:rate1}
\end{align}
In~\eqref{eq:rate1}, $\|\delta\mathbf{Z}_\text{MS}^{-1}\mathbf{e}_i\|\approx C_1 h^q$ and $|\delta\mathbf{e}_j^T\mathbf{Z}_\text{MS}^{-1}\mathbf{e}_i|\approx C_2 h^q$, where $q$ is negative. If $|\mathbf{e}_j^T \mathbf{J}^h|\le C_3 h^r$, as $h\to 0$,
\begin{align*}
\|\tilde{\mathbf{J}}^h - \mathbf{J}^h\|
\le
C' h^r,
\end{align*}
such that
\begin{align*}
\|\tilde{\mathbf{J}}^h - \mathbf{J}_n \|
\le
\|\tilde{\mathbf{J}}^h - \mathbf{J}^h\| + \|\mathbf{J}^h - \mathbf{J}_n\|
\le
C' h^r + Ch^p
\le
\tilde{C} h^{\min\{r,p\}}.
\end{align*}
Therefore, $\tilde{p}=\min\{r,p\}$ in~\eqref{eq:incorrect_error}.
For a fixed index $j$, as the mesh is refined, the spatial location moves toward the corner of the domain where the normal components of the surface current vanish. Therefore, $r$ is positive.
In Case~\ref{ce:mat_elem}, the choice of $\mathbf{J}_\text{MS}$ results in $r=1$, due to the linear arguments of the sinusoidal functions; therefore, $\tilde{p}=1$.
\section{Conclusions}
\label{sec:conclusions}
As outlined in the introduction, there are several combinations through which the different error sources in the MoM implementation of the EFIE can interact. \reviewerTwo{Combinations that avoid the solution-discretization error are not verifiable, and combinations that verify the numerical-integration error can be impractical due to the difficulty of computing accurate reference solutions for singular integrals}. In this paper, we presented an approach through which the order of the solution-discretization error, Error Source~\ref{err:sol_disc}, can be verified. Through this approach, we manufactured both the surface current and Green's function, and reduced the arising singular system of equations to a set of constraints for an optimization problem that selects the permissible solution closest to the manufactured solution. We demonstrated the effectiveness of this approach for properly and improperly coded examples.
\reviewerOne{Though we demonstrated this approach using a single basis-function choice, this approach can be extended to include other basis-function choices with minimal implications.} This approach can additionally be extended to include the domain-discretization error, Error Source~\ref{err:dom_disc}, to make Combination 4, \reviewerTwo{which accounts for the domain-discretization and solution-discretization errors,} verifiable. Therefore, verifying \reviewerTwo{combinations that account for the numerical-integration error} remains an open challenge for practical problems, requiring an acceptable balance of integration accuracy and computational cost.
\section{The Method-of-Moments Implementation of the EFIE}
\label{sec:efie}
In time-harmonic form, the scattered electric field $\mathbf{E}^\mathcal{S}$ can be computed from the surface current by
\begin{align}\rereading{
\mathbf{E}^\mathcal{S} = -\left(j\omega\mathbf{A}+\nabla\Phi\right)},
\label{eq:Es}
\end{align}
where the magnetic vector potential $\mathbf{A}$ is defined by
\begin{align}
\mathbf{A}(\mathbf{x})= \mu \int_{S'} \mathbf{J} (\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS',
\label{eq:A}
\end{align}
and, \rereading{by employing the Lorenz gauge condition and the continuity equation}, the electric scalar potential $\Phi$ is defined by
\begin{align}
\Phi(\mathbf{x})= \frac{j}{\epsilon\omega} \int_{S'} \nabla'\cdot\mathbf{J}(\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS'.
\label{eq:Phi}
\end{align}
In~\eqref{eq:A} and~\eqref{eq:Phi}, \reviewerOne{the integration domain is the surface $S$ of a perfectly conducting scatterer. Additionally,} $\mathbf{J}$ is the surface current, $\mu$ and $\epsilon$ are the permeability and permittivity of surrounding medium, and $G$ is the Green's function
\begin{align}
G(\mathbf{x},\mathbf{x}') = \frac{e^{-jkR}}{4\pi R},
\label{eq:G}
\end{align}
where $R=|\mathbf{x}-\mathbf{x}'|$, and $k=\omega\sqrt{\mu\epsilon}$ is the wave number. If $S$ is open, the component of $\mathbf{J}$ normal to the boundary of $S$ must vanish on the boundary of $S$ \rereading{to reflect that the total current is zero}.
The total electric field $\mathbf{E}$ is the sum of the incident electric field $\mathbf{E}^\mathcal{I}$ \rereading{(which induces $\mathbf{J}$)} and $\mathbf{E}^\mathcal{S}$. On $S$, the tangential component of $\mathbf{E}$ is zero, such that
\begin{align}\reviewerOne{
\mathbf{E}_t^\mathcal{S}=-\mathbf{E}_t^\mathcal{I}},
\label{eq:tan_BC}
\end{align}
\reviewerOne{%
where the subscript $t$ denotes the tangential component.
Substituting~\eqref{eq:Es} into~\eqref{eq:tan_BC}}, we can compute $\mathbf{J}$ from $\mathbf{E}^\mathcal{I}$:
\begin{align}
\mathbf{E}_t^\mathcal{I} =\left(j\omega\mathbf{A} + \nabla\Phi\right)_t.
\label{eq:tan}
\end{align}
To solve~\eqref{eq:tan} \reviewerOne{for $\mathbf{J}$, we discretize $S$ with a mesh composed of triangular elements and approximate $\mathbf{J}$ with $\mathbf{J}_h$ in terms of the Rao--Wilton--Glisson (RWG) basis functions $\boldsymbol{\Lambda}_{j}(\mathbf{x})$~\cite{rao_1982}:
\begin{align}
\mathbf{J}_h(\mathbf{x}) = \sum_{j=1}^{n_b} J_{j} \boldsymbol{\Lambda}_{j}(\mathbf{x}),
\label{eq:J_approx}
\end{align}}%
where $n_b$ is the \rereading{total number of basis functions.
The RWG basis functions are second-order accurate~\cite[pp.\ 155--156]{warnick_2008}, and are defined for a triangle pair} by
\begin{align*}
\boldsymbol{\Lambda}_{j}(\mathbf{x}) = \left\{
\begin{matrix}
\displaystyle\frac{\ell_{j}}{2A_{j}^+}\boldsymbol{\rho}_{j}^+, & \text{for }\mathbf{x}\in T_{j}^+ \\[1em]
\displaystyle\frac{\ell_{j}}{2A_{j}^-}\boldsymbol{\rho}_{j}^-, & \text{for }\mathbf{x}\in T_{j}^- \\[1em]
\mathbf{0}, & \text{otherwise}
\end{matrix}
\right.,
\end{align*}
where \rereading{$\ell_{j}$ is the length of the edge shared by the triangle pair, and $A_{j}^+$ and $A_{j}^-$ are the areas of the triangles $T_{j}^+$ and $T_{j}^-$ associated with basis function $j$}. $\boldsymbol{\rho}_{j}^+$ denotes the vector from the vertex of $T_{j}^+$ opposite the shared edge to $\mathbf{x}$, and $\boldsymbol{\rho}_{j}^-$ denotes the vector to the vertex of $T_{j}^-$ opposite the shared edge from $\mathbf{x}$.
These basis functions ensure that \reviewerOne{$\mathbf{J}_h$} is tangential to $S$ and has no component normal to the outer boundary of the triangle pair.
Additionally, along \rereading{the shared edge of the triangle pair, the component of $\boldsymbol{\Lambda}_{j}(\mathbf{x})$ normal to that edge is unity. Therefore, for a triangle edge shared by only two triangles,} the component of \reviewerOne{$\mathbf{J}_h$} normal to that edge is $J_j$. \reviewerTwo{Because the normal component is constant along the edge, the} solution is considered most accurate at the midpoint of the edge~\cite[pp.\ 155--156]{warnick_2008}.
Projecting~\eqref{eq:tan} onto $\boldsymbol{\Lambda}_{i}(\mathbf{x})$ yields
\begin{align*}
\int_S \mathbf{E}^\mathcal{I}\cdot \boldsymbol{\Lambda}_{i} dS =j\omega\int_S \mathbf{A}\cdot\boldsymbol{\Lambda}_{i} dS + \int_S \nabla\Phi\cdot\boldsymbol{\Lambda}_{i} dS,
\end{align*}
which can be integrated by parts to obtain
\begin{align}
\int_S \mathbf{E}^\mathcal{I}\cdot \boldsymbol{\Lambda}_{i} dS =j\omega\int_S \mathbf{A}\cdot\boldsymbol{\Lambda}_{i} dS - \int_S \Phi \nabla\cdot\boldsymbol{\Lambda}_{i} dS.
\label{eq:proj}
\end{align}
Inserting~\eqref{eq:A} and~\eqref{eq:Phi} into \eqref{eq:proj} yields
\begin{align}
\int_S \mathbf{E}^\mathcal{I}\cdot \boldsymbol{\Lambda}_{i} dS =&{}%
j\omega\mu \int_S \boldsymbol{\Lambda}_{i}(\mathbf{x})\cdot\int_{S'} \mathbf{J}(\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS'dS
{}-
\frac{j}{\epsilon\omega} \int_S \nabla\cdot\boldsymbol{\Lambda}_{i}(\mathbf{x})\int_{S'} \nabla'\cdot\mathbf{J}(\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS' dS.
\label{eq:proj_exp}
\end{align}
Substituting~\eqref{eq:J_approx} into~\eqref{eq:proj_exp}, we obtain \rereading{the discretized equation}
\begin{align}
\int_S \mathbf{E}^\mathcal{I}\cdot \boldsymbol{\Lambda}_{i} dS ={}&%
j\omega\mu \sum_{j=1}^{n_b} J_{j}\int_S \boldsymbol{\Lambda}_{i}(\mathbf{x})\cdot\int_{S'} \boldsymbol{\Lambda}_{j}(\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS'dS
\nonumber \\
&{}-%
\frac{j}{\epsilon\omega}\sum_{j=1}^{n_b} J_{j}\int_S \nabla\cdot\boldsymbol{\Lambda}_{i}(\mathbf{x})\int_{S'} \nabla'\cdot\boldsymbol{\Lambda}_{j}(\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS' dS.
\label{eq:proj_disc}
\end{align}
\reviewerOne{Letting $\mathbf{J}^h$ denote the vector of coefficients used to construct $\mathbf{J}_h$~\eqref{eq:J_approx}, \eqref{eq:proj_disc} can be written in matrix form as
\begin{align}
\mathbf{Z}\mathbf{J}^h = \mathbf{V},
\label{eq:zjv}
\end{align}}%
where
\begin{align*}
Z_{i,j} &{}= j\omega\mu \int_S \boldsymbol{\Lambda}_{i}(\mathbf{x})\cdot\int_{S'} \boldsymbol{\Lambda}_{j}(\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS'dS - \frac{j}{\epsilon\omega} \int_S \nabla\cdot\boldsymbol{\Lambda}_{i}(\mathbf{x})\int_{S'} \nabla'\cdot\boldsymbol{\Lambda}_{j}(\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS' dS,
\\
\reviewerOne{J_{j}^h} &{}= J_{j},
\\
V_{i} &{}=\int_S \mathbf{E}^\mathcal{I}\cdot \boldsymbol{\Lambda}_{i} dS.
\end{align*}
\section*{References}}
\input{packages.tex}
\usepackage[margin=1in]{geometry}
\setlength{\skip\footins}{1em}
\biboptions{sort&compress}
\begin{document}
\begin{frontmatter}
\title{Manufactured Solutions for the Method-of-Moments Implementation of the Electric-Field Integral Equation}
\author[freno]{Brian A.\ Freno}
\ead{[email protected]}
\author[freno]{Neil R.\ Matula}
\author[freno]{\rereading{William A.\ Johnson}}
\address[freno]{Sandia National Laboratories, Albuquerque, NM 87185}
\begin{abstract}
\input{abstract.tex}
\end{abstract}
\begin{keyword}
method of moments \sep
electric-field integral equation \sep
code verification \sep
manufactured solutions
\end{keyword}
\end{frontmatter}
\input{introduction.tex}
\input{efie.tex}
\input{mms.tex}
\input{results.tex}
\input{conclusions.tex}
\input{acknowledgments.tex}
\section{Introduction}
The method-of-moments (MoM) implementation of the electric-field integral equation (EFIE) is a useful technique for numerically modeling electromagnetic scattering and radiation problems. Through this approach, the surface of the electromagnetic scatterer is discretized using planar or curvilinear mesh elements, and four-dimensional integrals are evaluated over two-dimensional source and test elements. However, the presence of a Green's function in these equations yields scalar- and vector-potential terms with singularities when the test and source elements share one or more edges or vertices and near-singularities when they are otherwise close. Many approaches have been developed to address the singularity and near-singularity for the inner, source-element integral~\cite{graglia_1993,wilton_1984,rao_1982,khayat_2005,fink_2008,khayat_2008,vipiana_2011,vipiana_2012,botha_2013,rivero_2019}, as well as for the outer, test-element integral~\cite{vipiana_2013,polimeridis_2013,wilton_2017,rivero_2019b,freno_em}.
For computational physics codes in general, it is necessary to assess the implementation and the suitability of the underlying models in order to develop confidence in the simulation results. These assessments typically fall into two complementary categories: verification and validation. Validation evaluates the appropriateness of the models instantiated in the code for representing the relevant physical phenomena, and is typically performed through comparison with experimental data. Verification, on the other hand, assesses the correctness of the numerical solutions produced by the code, through comparison with the expected theoretical behavior of the implemented numerical methods. Following Roache~\cite{roache_1998}, Salari and Knupp~\cite{salari_2000}, and Oberkampf and Roy~\cite{oberkampf_2010}, verification can be further divided into the activities of code verification and solution verification. Solution verification involves the estimation of the numerical error for a particular simulation, whereas code verification assesses the correctness of the implementation of the numerical methods within the code. A review of code and solution verification is presented by Roy~\cite{roy_2005}.
Code verification is the focus of this paper. In general, codes that approximately solve systems of differential, integral, or integro-differential equations can only be verified by using them to solve problems with known solutions~\cite{roache_2001}. The discretization of the governing equations necessarily incurs some truncation error, and thus the approximate solutions produced from the discretized equations will incur an associated discretization error. If the solution to the problem is known, a measure of the discretization error (typically a discrete norm thereof) may be evaluated directly from the approximate solution. In the most basic sense of verification, if the discretization error tends to zero as the discretization is refined, the consistency of the code is verified~\cite{roache_1998}. This may be taken a step further by examining not only consistency, but the rate at which the error decreases as the discretization is refined, thereby verifying the order of accuracy of the discretization scheme. The correctness of the numerical-method implementation may then be verified by comparing the expected and observed orders of accuracy obtained from numerous test cases with known solutions.
Exact solutions to systems of engineering interest are rare, and those that do exist often require dramatic simplifications to both the domain geometry and the equations themselves in order to obtain a tractable problem. Hence, manufactured solutions are frequently employed to produce problems of sufficient complexity with known solutions~\cite{roache_2001}. The method of manufactured solutions (MMS) is a general technique for constructing problems of arbitrary complexity with known solutions. One begins this process in reverse by manufacturing the desired solution. In principle, this manufactured solution (MS) may be any function, but several properties are desirable~\cite{salari_2000}:
\begin{enumerate}[leftmargin=*,labelindent=\parindent]
\item \label{prop:elementary} The MS should consist of combinations of elementary functions, such as polynomial, trigonometric, and exponential functions. This not only simplifies derivations and implementation, but ensures that the MS (and its derivatives) will be representable to sufficient precision within the tested code.
\item The MS should be sufficiently smooth, such that the error incurred by the discretization is small on relatively coarse meshes. This ensures that the order of accuracy may be estimated with minimal computational expense.
\item The MS should be general enough that all terms of the governing equations are exercised.
\item The MS should have a sufficient number of nontrivial derivatives, such that the expected order of accuracy of the discretization can be observed. In the most ideal case, the solution will have an infinite number of nontrivial derivatives.
\item Since the robustness of the code is not the primary concern, the MS should not have any features that inhibit the solution of the discretized equations.
\end{enumerate}
Once a solution is manufactured, it is substituted directly into the governing equations.
In general, the MS is not expected to satisfy the governing equations. Instead, a residual term appears, which quantifies the deviation from the satisfaction of the equations. If this residual is added to the governing equations as a source term, the resulting equations will be exactly satisfied by the MS. Concerns are immediately raised regarding uniqueness of the solution to the manufactured problem, but this has rarely been found to cause difficulties in practice~\cite{roache_2001}. The result of this process is a problem, of arbitrary complexity, for which an exact solution is known.
The code to be verified is then modified to support the additional source term and may then be verified by comparing the simulation result for the manufactured problem against the known solution.
\reviewerOne{Ideally, the source term is computed analytically; however, when this is not possible, the source term needs to be computed consistently and at least as accurately as the numerical methods being verified. Otherwise, the error in the source term will overshadow that of the numerical methods, contaminating the verification assessment.}
MMS is purely a mathematical process; the physics of the problem need not be considered, \reviewerTwo{provided the MS remains within the bounds of validity for the underlying algorithms}. This enables the user to avoid difficulties that would normally complicate the solution.
Code verification has been performed on computational physics codes associated with several physics disciplines, including fluid dynamics~\cite{roy_2004,bond_2007,veluri_2010,oliver_2012,eca_2016,freno_2021}, solid mechanics~\cite{chamberland_2010}, fluid--structure interaction~\cite{etienne_2012}, heat transfer in fluid--solid interaction~\cite{veeraragavan_2016}, multiphase flows~\cite{brady_2012}, radiation hydrodynamics~\cite{mcclarren_2008}, electrodynamics~\cite{ellis_2009}, and ablation~\cite{amar_2008,amar_2009,amar_2011,freno_ablation}.
However, existing literature contains few instances of MMS being used in the verification of software for integral equations. This is due to the simple fact that, while analytical differentiation is a straightforward exercise, analytical integration is not always possible. Hence, the residual source term arising from the manufactured solution may not be representable in closed form, and its implementation may be accompanied by numerical techniques that carry their own uncertainties. Furthermore, in many applications, such as the MoM implementation of the EFIE, singular integrals appear, which can further complicate the numerical evaluation of the source term. Therefore, much of the elegance, simplicity, and instilled confidence of MMS is lost when applied to integral equations in a straightforward manner, and, as a result, effective implementation of MMS in the context of boundary element codes is an open subject of research.
The most substantial effort thus far for integral equations in computational electromagnetics has been the work of Marchand et al.~\cite{marchand_2013,marchand_2014}, in which the authors calculate the MMS source terms for the EFIE using numerical techniques. While the quadrature error can be driven low enough that \reviewerTwo{mesh-refinement} studies become feasible, the presence of this additional error often places a lower bound on the discretization error that can be obtained, and therefore limits the scope of the \reviewerTwo{mesh-refinement} study, in addition to being undesirable for the aforementioned reasons. To further complicate matters, for the MoM implementation of the EFIE, the discretized equations can easily become ill-conditioned~\cite{adrian_2019}.
For the MoM implementation of the EFIE, numerical error is introduced from three sources:
\begin{enumerate}[leftmargin=*,labelindent=\parindent]
\item \label{err:dom_disc} \textbf{Domain discretization.} While planar surfaces can be represented exactly by planar elements, the approximation of curved surfaces with planar elements introduces a second-order numerical error~\cite[Chap.~3]{warnick_2008}. This error can be reduced by employing curved elements, but we restrict the scope of this work to planar elements.
\item \label{err:sol_disc} \textbf{Solution discretization.} Common in the solution to differential, integral, and integro-differential equations, the approximation of the solution in terms of a finite number of basis functions, or alternatively the approximation of the underlying equation operators in terms of a finite amount of solution queries, is the most common contributor to the numerical error.
\item \label{err:num_int}\textbf{Numerical integration.}
As previously stated, the analytical evaluation of the integrals in integral equations is not always possible. For well-behaved integrals, quadrature rules or other integration methods can be used, with the expectation that the associated numerical error is at least of the same order as that arising from Error Source~\ref{err:sol_disc}. A less rigorous expectation is that the error from numerical integration decreases as the fidelity of the numerical integration algorithm is increased (e.g., increasing the number of quadrature points). However, for singular or nearly singular integrals, such convergence is not assured~\cite{freno_quad}.
\end{enumerate}
It is important to account for these error sources when performing code verification. Table~\ref{tab:combinations} lists the possible combinations of these error sources.
\begin{table}[!b]
\centering
\begin{tabular}{c c c c c }
\toprule
& \multicolumn{3}{c}{Error Source} \\
\cmidrule(lr){2-4}
Combination &
\hspace{0.5em}\ref{err:dom_disc}\hspace{0.5em} &
\hspace{0.5em}\ref{err:sol_disc}\hspace{0.5em} &
\hspace{0.5em}\ref{err:num_int} \hspace{0.5em} &
Verifiable?\\
\midrule
1 & \cmark & & & \color{red}\ding{55} \\
2 & & \cmark & & \color{green}\cmark \\
3 & & & \cmark & \color{red}\ding{55} \\
4 & \cmark & \cmark & & \color{green}\cmark \\
5 & & \cmark & \cmark & \color{green}\cmark \\
6 & \cmark & & \cmark & \color{red}\ding{55} \\
7 & \cmark & \cmark & \cmark & \color{green}\cmark \\
\bottomrule
\end{tabular}
\caption{Possible combinations of error sources.}
\vskip-\dp\strutbox
\label{tab:combinations}
\end{table}
Error Source~\ref{err:dom_disc} can be eliminated for planar surfaces, and Error Source~\ref{err:num_int} can be eliminated if the numerical integration can be performed exactly. However, for certain solution discretization choices, such as the use of the Rao--Wilton--Glisson basis functions, nontrivial solutions cannot be exactly represented across different meshes.
Therefore, we cannot verify Combinations 1, 3, or 6.
Exact solutions exist for curved surfaces, such as the Mie solution to Maxwell's equations for a sphere, permitting Combination 7 to be verified.
Combination 5 can also be verified; however, while Error Source~\ref{err:num_int} can be easily studied for nonsingular integrands, for singular and nearly singular integrands, Error Source~\ref{err:num_int} is most likely better studied using unit tests, due to the computationally expensive adaptive integration techniques required to compute accurate reference solutions~\cite{freno_quad,freno_em}.
In order for Combinations 2 and 4 to be verified, the numerical integration needs to be exact, which is not possible for the Green's function with nontrivial solutions. Therefore, the purpose of this paper is to present a code-verification technique for verifying Combination 2 in the MoM implementation of the EFIE.
Through this approach, we manufacture the surface current and Green's function and only consider surfaces that can be represented by planes. The planar surfaces avoid Error Source~\ref{err:dom_disc}, and, by manufacturing the two functions, we can quickly and exactly compute the integrals, avoiding Error Source~\ref{err:num_int}. This enables us to isolate Error Source~\ref{err:sol_disc}. Using the approach in this paper, one could verify Combination 4 as well.
Despite permitting analytical evaluation of the integrals, the manufactured Green's function further worsens the conditioning of the discretized equations, permitting multiple solutions. To mitigate this challenge, we reduce the equations to a set of constraints and admit the solution closest to the MS that satisfies these constraints.
This paper is organized as follows. In Section~\ref{sec:efie}, we describe the MoM implementation of the EFIE. In Section~\ref{sec:mms}, we describe the challenges of using MMS with the MoM implementation of the EFIE, and we describe our approach to mitigating them. In Section~\ref{sec:results}, we demonstrate the effectiveness of our approach for cases with and without coding errors. In Section~\ref{sec:conclusions}, we summarize this work and provide an outlook for future work.
\section{Manufactured Solutions}
\label{sec:mms}
Equation~\eqref{eq:proj_exp} can be written in terms of its residual \reviewerOne{as
$\mathbf{r}(\mathbf{J}) = \mathbf{0}$,
where
\begin{align*}
r_{i}(\mathbf{J}) =&{}%
j\omega\mu \int_S \boldsymbol{\Lambda}_{i}(\mathbf{x})\cdot\int_{S'} \mathbf{J}(\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS'dS
{}-
\frac{j}{\epsilon\omega} \int_S \nabla\cdot\boldsymbol{\Lambda}_{i}(\mathbf{x})\int_{S'} \nabla'\cdot\mathbf{J}(\mathbf{x}')G(\mathbf{x},\mathbf{x}')dS' dS -
\int_S \mathbf{E}^\mathcal{I}\cdot \boldsymbol{\Lambda}_{i} dS.
\end{align*}
Similarly, its discretization~\eqref{eq:zjv} can be written as
\begin{align}
\mathbf{r}_h(\mathbf{J}_h) = \mathbf{0},
\label{eq:res_disc}
\end{align}
where $\mathbf{r}_h(\mathbf{J}_h) = \mathbf{Z}\mathbf{J}^h - \mathbf{V}$.}
The method of manufactured solutions modifies~\eqref{eq:res_disc} to be
\begin{align}
\mathbf{r}_h(\mathbf{J}_h) = \mathbf{r}(\mathbf{J}_\text{MS}),
\label{eq:mms}
\end{align}
where $\mathbf{J}_\text{MS}$ is the manufactured solution, and the generally nonzero $\mathbf{r}(\mathbf{J}_\text{MS})$ is computed analytically. \reviewerOne{In~\eqref{eq:mms}, $\int_S \mathbf{E}^\mathcal{I}\cdot \boldsymbol{\Lambda}_{i} dS$ appears on both sides, resulting in its cancellation. However, instead of solving~\eqref{eq:mms}, we could equivalently solve~\eqref{eq:zjv} by inserting $\mathbf{J}_\text{MS}$ into~\eqref{eq:tan} to obtain a new incident electric field, which is projected onto $\boldsymbol{\Lambda}_{i}(\mathbf{x})$ to obtain $\mathbf{V}$.
However, as described in the introduction, integrals containing the Green's function~\eqref{eq:G}, such as those appearing in this hypothetical proposal for $\mathbf{V}$ are not only unable to be computed analytically, but the singularity when $R\to 0$ complicates their accurate approximation, potentially contaminating convergence studies.
}%
\begin{figure}[!t]
\centering
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{G_MS.pdf}
\caption{$G_\text{MS}$~\eqref{eq:G_mms} for different values of $d$.}
\vskip-\dp\strutbox
\label{fig:G_MS}
\end{figure}
Nonetheless, we can still verify the order of accuracy of the solution-discretization error, Error Source~\eqref{err:sol_disc}, by also manufacturing the Green's function, using the form
\begin{align}
G_\text{MS}(\mathbf{x},\mathbf{x}') = \left(1 - \frac{R^2}{R_m^2}\right)^d,
\label{eq:G_mms}
\end{align}
where $R_m=\max_{\mathbf{x},\mathbf{x}'\in S} R$ is the maximum possible distance between two points on $S$, and $d\in\mathbb{N}$. A plot of~\eqref{eq:G_mms} is shown in Figure~\ref{fig:G_MS} for multiple values of $d$. The form of~\eqref{eq:G_mms} is chosen for two reasons: 1) the even powers of $R$ permit the integrals in~\eqref{eq:proj_disc} and
\reviewerOne{$\mathbf{V}$
to be computed analytically for many choices of $\mathbf{J}_\text{MS}$, avoiding contamination from additional error}, and 2) $G_\text{MS}$ increases when $R$ decreases, as with the actual Green's function~\eqref{eq:G}.
Introducing the terms $\mathbf{Z}_\text{MS}$ and $\mathbf{V}_\text{MS}$ to account for $G_\text{MS}$, we write the resulting system of equations as
\begin{align}
\mathbf{r}_\text{MS} = \mathbf{Z}_\text{MS}\mathbf{J}^h - \mathbf{V}_\text{MS} = \mathbf{0},
\label{eq:zjv_mms}
\end{align}
where
\reviewerOne{
\begin{align}
Z_{\text{MS}_{i,j}} &{}= j\omega\mu \int_S \boldsymbol{\Lambda}_{i}(\mathbf{x})\cdot\int_{S'} \boldsymbol{\Lambda}_{j}(\mathbf{x}')G_\text{MS}(\mathbf{x},\mathbf{x}')dS'dS - \frac{j}{\epsilon\omega} \int_S \nabla\cdot\boldsymbol{\Lambda}_{i}(\mathbf{x})\int_{S'} \nabla'\cdot\boldsymbol{\Lambda}_{j}(\mathbf{x}')G_\text{MS}(\mathbf{x},\mathbf{x}')dS' dS, \nonumber \\
V_{\text{MS}_{i}} &{}=\int_S \mathbf{E}_\text{MS}^\mathcal{I}\cdot \boldsymbol{\Lambda}_{i} dS, \label{eq:V_mms}
\end{align}}
and
\begin{align}
\mathbf{E}_\text{MS}^\mathcal{I}(\mathbf{x}) &{}= \frac{j}{\omega\epsilon} \int_{S'}\left( k^2\mathbf{J}_\text{MS} (\mathbf{x}')G_\text{MS}(\mathbf{x},\mathbf{x}') +\nabla'\cdot\mathbf{J}_\text{MS}(\mathbf{x}')\nabla G_\text{MS}(\mathbf{x},\mathbf{x}')\right)dS'.
\label{eq:Ei_mms}
\end{align}
Solving~\eqref{eq:zjv_mms} for $\mathbf{J}^h$ enables us to compute the discretization error
\reviewerTwo{%
\begin{align}
\mathbf{e}_n = \mathbf{J}^h - \mathbf{J}_n
\label{eq:error}
\end{align}%
where $J_{n_j}$ denotes the component of $\mathbf{J}_\text{MS}$ flowing from $T_j^+$ to $T_j^-$. The norm of~\eqref{eq:error} has the property $\|\mathbf{e}_n\|\le C h^p$, where
$C$ is a function of the solution derivatives}, $h$ is representative of the mesh size, and $p$ is the order of accuracy. By performing a mesh-convergence study of \reviewerTwo{the norm of} the discretization error, we can ensure the proper order of accuracy is obtained. For the RWG basis functions, the expectation is second-order accuracy $(p=2)$.
Despite the benefit of yielding analytically integrable integrals in~\eqref{eq:zjv_mms}, $G_\text{MS}$ makes $\mathbf{Z}_\text{MS}$ practically singular, admitting multiple solutions $\mathbf{J}^h$ in~\eqref{eq:zjv_mms}. Therefore, we choose the $\mathbf{J}^h$ that \reviewerOne{both yields $\mathbf{J}_h$ closest} to $\mathbf{J}_\text{MS}$ and satisfies~\eqref{eq:zjv_mms}.
To accomplish this, we perform a pivoted QR factorization of $\mathbf{Z}_\text{MS}^H$, such that
\begin{align}
\mathbf{Z}_\text{MS}^H\mathbf{P} = \left[\mathbf{Q}_1,\,\mathbf{Q}_2\right]\left[\begin{matrix} \mathbf{R}_1 \\ \mathbf{0} \end{matrix}\right] = \mathbf{Q}_1 \mathbf{R}_1,
\label{eq:qr}
\end{align}
where $\mathbf{Z}_\text{MS}\in\mathbb{C}^{n_b\timesn_b}$, $\mathbf{Q}_1\in\mathbb{C}^{n_b\timesm_b}$, $\mathbf{Q}_2\in\mathbb{C}^{n_b\times(n_b-m_b)}$, and $\mathbf{R}_1\in\mathbb{C}^{m_b\timesn_b}$. $\mathbf{P}$ is a permutation matrix, and \rereading{the superscript $H$ denotes conjugated transposition}.
Numerically, the pivoting facilitates the determination of the rank $m_b\len_b$ of $\mathbf{Z}_\text{MS}$. We can express $\mathbf{J}^h$ in terms of the basis $\mathbf{Q}$:
\begin{align}
\mathbf{J}^h = \mathbf{Q}_1 \mathbf{u} + \mathbf{Q}_2 \mathbf{v},
\label{eq:J_qbasis}
\end{align}
where $\mathbf{u}\in\mathbb{C}^{m_b}$ is the vector of coefficients that satisfies~\eqref{eq:zjv_mms} and $\mathbf{v}\in\mathbb{C}^{n_b-m_b}$ is the vector of coefficients that brings $\mathbf{J}_h$ closest to $\mathbf{J}_\text{MS}$, given $\mathbf{u}$.
Substituting~\eqref{eq:qr} and~\eqref{eq:J_qbasis} into~\eqref{eq:zjv_mms} and noting that $\mathbf{Q}$ is unitary, $\mathbf{u}$ can be computed from
\begin{align}
\mathbf{R}_1^H\mathbf{u} = \mathbf{P}^T\mathbf{V}_\text{MS}.
\label{eq:u}
\end{align}
From $\mathbf{u}$ satisfying~\eqref{eq:u}, $\mathbf{J}^h$ satisfies~\eqref{eq:zjv_mms}. With $\mathbf{u}$ known, $\mathbf{v}$ remains to be determined. Let
\begin{align}
\bar{\mathbf{J}}^h = \mathbf{Q}_1 \mathbf{u}.
\label{eq:Jbar}
\end{align}
We solve for $\mathbf{v}$ by minimizing
\begin{align}
\rereading{\|\mathbf{e}_n\|_2^2} = \left(\mathbf{J}^h-\mathbf{J}_n\right)^H\left(\mathbf{J}^h-\mathbf{J}_n\right).
\label{eq:error_L2}
\end{align}
Substituting~\eqref{eq:J_qbasis} into~\eqref{eq:error_L2} and using~\eqref{eq:Jbar}, \eqref{eq:error_L2} can be written as
\begin{align}
\rereading{\|\mathbf{e}_n\|_2^2 = \left[\mathbf{Q}_2\mathbf{v} - \left(\mathbf{J}_n-\bar{\mathbf{J}}^h\right)\right]^H\left[\mathbf{Q}_2\mathbf{v} - \left(\mathbf{J}_n-\bar{\mathbf{J}}^h\right)\right]}.
\label{eq:quadratic}
\end{align}
Equation~\eqref{eq:quadratic} is quadratic, and is minimized when
\begin{align*}
\mathbf{v} = \mathbf{Q}_2^H\left(\mathbf{J}_n-\bar{\mathbf{J}}^h\right).
\end{align*}
Therefore, the expression for $\mathbf{J}^h$ is
\begin{align}
\mathbf{J}^h = \bar{\mathbf{J}}^h + \mathbf{Q}_2\mathbf{Q}_2^H\left(\mathbf{J}_n-\bar{\mathbf{J}}^h\right).
\label{eq:Jh_sol}
\end{align}
Equation~\eqref{eq:Jh_sol} provides us with the solution to~\eqref{eq:zjv_mms} that is closest to $\mathbf{J}_\text{MS}$.
\section{Numerical Examples}
\label{sec:results}
In this section, we demonstrate the effectiveness of the approach described in Section~\ref{sec:mms}.
As stated in the introduction, Error Source~\ref{err:dom_disc} is outside of the scope of this work. Therefore, we restrict our domain to planar regions. Specifically, we consider two unit-square flat plates, as shown in Figure~\ref{fig:two_squares}, with one rotated out of the plane of the other by angle $\theta$.
We manufacture the surface current $\mathbf{J}_\text{MS}(\mathbf{x}) = \{J_\xi(\boldsymbol{\xi}),\,J_\eta(\boldsymbol{\xi})\}$ using sinusoidal functions:
\begin{align}
J_\xi (\boldsymbol{\xi}) &{}= J_0 \cos(\pi \xi/2)\cos(\pi \eta/4), \label{eq:Jxi}\\
J_\eta(\boldsymbol{\xi}) &{}= J_0 \cos(\pi \xi/4)\sin(\pi \eta), \label{eq:Jeta}
\end{align}
where $J_0=1$ A/m and the plate-fixed coordinate system $\boldsymbol{\xi}(\mathbf{x};\theta)$ is given by
\reviewerOne{%
\begin{align*}
\xi(\mathbf{x};\theta) &{}= \frac{1}{L_0}\left\{\begin{matrix}
x, & \text{for } x\le 0 \text{ m} \\
x\cos\theta + z\sin\theta, & \text{for } x>0 \text{ m}
\end{matrix}\right.,
\\
\eta(\mathbf{x}) &{}= y/L_0,
\end{align*}
where $L_0=1$ m.}
At the edges of the domain, the normal component of $\mathbf{J}_\text{MS}(\mathbf{x})$ is zero, satisfying the boundary conditions. Figures~\ref{fig:Jxi} and~\ref{fig:Jeta} provide plots of~\eqref{eq:Jxi} and~\eqref{eq:Jeta}.
\input{fig_two_squares.tex}
\begin{figure}[!b]
\centering
\usebox{\twosquares}%
\caption{Computational domain consisting of two unit-square plates. \reviewerOne{Coordinates are expressed in the plate-fixed coordinate system $\boldsymbol{\xi}(\mathbf{x};\theta)$.}}
\label{fig:two_squares}
\end{figure}
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[scale=.28,clip=true,trim=0in 0in 0in 0in]{uniform_60x30_Jxi.pdf}
\caption{\strut$J_\xi$}
\label{fig:Jxi}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[scale=.28,clip=true,trim=0in 0in 0in 0in]{uniform_60x30_Jeta.pdf}
\caption{\strut$J_\eta$}
\label{fig:Jeta}
\end{subfigure}
\caption{\strut Manufactured surface current $\mathbf{J}_\text{MS}$.}
\vskip-\dp\strutbox
\label{fig:J_MS}
\end{figure}
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[scale=.28,clip=true,trim=0in 0in 0in 0in]{uniform_40x20_mesh.pdf}
\caption{\strut Uniform}
\label{fig:uniform}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[scale=.28,clip=true,trim=0in 0in 0in 0in]{deformed_40x20_mesh.pdf}
\caption{\strut Twisted}
\label{fig:twisted}
\end{subfigure}
\caption{\strut Two different types of meshes, shown with $n_t=1600$.}
\vskip-\dp\strutbox
\label{fig:mesh}
\end{figure}
We consider two types of meshes: a uniform mesh and a twisted mesh, examples of which are shown in Figures~\ref{fig:uniform} and~\ref{fig:twisted} with $n_t=1600$ triangles, as well as $d=1$ and $d=2$ in the manufactured Green's function $G_\text{MS}$~\eqref{eq:G_mms}. The twisted mesh is obtained by transforming the uniform mesh, using the transformation provided in Reference~\cite{freno_2021}.
We measure the $L^\infty$-norm of the error $\mathbf{e}_n$~\eqref{eq:error}:
\begin{align}
\varepsilon = \left\|\mathbf{e}_n\right\|_\infty = \max|\mathbf{J}^h-\mathbf{J}_n|
\label{eq:error_norm}
\end{align}
to determine the maximum error in the surface current normal to the interior edges at the edge midpoints.
Finally, we account for potential disparities in the magnitudes of the contributions to $\mathbf{Z}_\text{MS}$ from $\mathbf{A}$ and $\Phi$:
\begin{align*}
\mathbf{Z}_\text{MS} = \mathbf{Z}^\mathbf{A} + \mathbf{Z}^\Phi,
\end{align*}
where
\begin{align*}
Z_{i,j}^\mathbf{A} &{}= \frac{j k^2}{\epsilon\omega} \int_S \boldsymbol{\Lambda}_{i}(\mathbf{x})\cdot\int_{S'} \boldsymbol{\Lambda}_{j}(\mathbf{x}')G_\text{MS}(\mathbf{x},\mathbf{x}')dS'dS ,
\\
Z_{i,j}^\Phi &{}= - \frac{j}{\epsilon\omega} \int_S \nabla\cdot\boldsymbol{\Lambda}_{i}(\mathbf{x})\int_{S'} \nabla'\cdot\boldsymbol{\Lambda}_{j}(\mathbf{x}')G_\text{MS}(\mathbf{x},\mathbf{x}')dS' dS.
\end{align*}
We consider the contributions $\mathbf{Z}^\mathbf{A}$ and $\mathbf{Z}^\Phi$ together and separately, with $\epsilon=1$ F/m and $\mu=1$ H/m. When we consider the contributions together, we set $k=1$ m$^{-1}$ for $\mathbf{Z}_\text{MS}$. When we consider them separately, we are effectively taking the limits as $k\to\infty$ for $\mathbf{Z}^\mathbf{A}$ and $k\to 0$ for $\mathbf{Z}^\Phi$. We adjust $\mathbf{E}_\text{MS}^\mathcal{I}$~\eqref{eq:Ei_mms} accordingly.
\subsection{Correct Implementation}
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{G1_A000_linf.pdf}
\caption{$\theta=0^\circ$}
\end{subfigure}
\hspace{0.25em}
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{G1_A045_linf.pdf}
\caption{$\theta=45^\circ$}
\label{fig:G1:45}
\end{subfigure}
\\
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{G1_A090_linf.pdf}
\caption{$\theta=90^\circ$}
\end{subfigure}
\hspace{0.25em}
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{G1_A135_linf.pdf}
\caption{$\theta=135^\circ$}
\end{subfigure}
\caption{Error $\varepsilon$~\eqref{eq:error_norm} for $d=1$ in $G_\text{MS}$~\eqref{eq:G_mms}.}
\vskip-\dp\strutbox
\vskip10em
\label{fig:G1}
\end{figure}
We first consider the case of $d=1$ in~\eqref{eq:G_mms}. For this choice of $G_\text{MS}$, $\mathbf{E}_\text{MS}^\mathcal{I}$~\eqref{eq:Ei_mms} can be computed analytically to yield a polynomial expression of degree two, such that the integrand of~\eqref{eq:V_mms} is a polynomial of degree three, which can be integrated exactly using four triangle quadrature points for polynomials~\cite{lyness_1975}.
Figure~\ref{fig:G1} shows plots of $\varepsilon$~\eqref{eq:error_norm} with respect to $n_t$ for four $\theta$ values and the two meshes. The plots indicate that the solution is second-order accurate, as expected. Table~\ref{tab:G1} shows the maximum rank, $\maxm_b$, of~\eqref{eq:zjv_mms} for each of the simulations, across the different mesh sizes. These low values emphasize how poorly conditioned $\mathbf{Z}_\text{MS}$ is.
Next, we consider the case of $d=2$ in~\eqref{eq:G_mms}. For this choice of $G_\text{MS}$, $\mathbf{E}_\text{MS}^\mathcal{I}$~\eqref{eq:Ei_mms} can be computed analytically to yield a polynomial expression of degree four, such that the integrand of~\eqref{eq:V_mms} is a polynomial of degree five, which can be integrated exactly using seven triangle quadrature points for polynomials~\cite{lyness_1975}.
Figure~\ref{fig:G2} shows analogous plots for $\varepsilon$~\eqref{eq:error_norm} with respect to $n_t$ for $d=2$ in~\eqref{eq:G_mms}. These plots also indicate that the solution is second-order accurate. Table~\ref{tab:G2} shows the maximum rank, $\maxm_b$, of~\eqref{eq:zjv_mms} for each of the simulations, across the different mesh sizes. These values are also low, albeit slightly higher than the values in Table~\ref{tab:G1}. The higher values in Table~\ref{tab:G2} are due to the greater relative weight of the matrix diagonal that arises from this choice of $G_\text{MS}$. Higher values of $d$ would be expected to increase $m_b$, but $m_b$ would most likely remain small relative to $n_b$. Additionally, higher values of $d$ require more quadrature points for exact integration, increasing the computational cost and potentially requiring finer meshes for the discretizations to be in the asymptotic region.
\reviewerTwo{Figures~\ref{fig:G1} and~\ref{fig:G2} show that the uniform mesh has a lower error for $\mathbf{Z}^\Phi$, whereas the twisted mesh has a lower error for $\mathbf{Z}_\text{MS}$ and $\mathbf{Z}^\mathbf{A}$. These performance differences are due to fortuitous mesh concentrations in regions with higher gradients. Simulations using the uniform mesh enter the asymptotic region faster, due to the regularity of the mesh.}
Finally, Figure~\ref{fig:res} shows the distribution of the pooled residuals $\mathbf{r}_\text{MS}$~\eqref{eq:zjv_mms} from each of the simulations plotted in Figures~\ref{fig:G1} and~\ref{fig:G2}. These low values confirm~\eqref{eq:zjv_mms} has been satisfied.
\begin{table}[!b]
\centering
\begin{tabular}{c c c c c c c}
\toprule
& \multicolumn{3}{c}{Uniform} & \multicolumn{3}{c}{Twisted} \\
\cmidrule(lr){2-4} \cmidrule(lr){5-7}
$\theta$
& $\mathbf{Z}_\text{MS}$ & $\mathbf{Z}^\mathbf{A}$ & $\mathbf{Z}^\Phi$
& $\mathbf{Z}_\text{MS}$ & $\mathbf{Z}^\mathbf{A}$ & $\mathbf{Z}^\Phi$
\\
\midrule
$\phantom{0}\pz0^\circ$ & \pz8 & \pz8 & \pz2 & \pz8 & \pz8 & \pz2 \\
$\pz45^\circ$ & \rereading{13} & \rereading{13} & \pz3 & \rereading{13} & \rereading{13} & \pz3 \\
$\pz90^\circ$ & \rereading{13} & \rereading{13} & \pz3 & \rereading{13} & \rereading{13} & \pz3 \\
$135^\circ$ & \rereading{13} & \rereading{13} & \pz3 & \rereading{13} & \rereading{13} & \pz3 \\
\bottomrule
\end{tabular}
\caption{Maximum rank of~\eqref{eq:zjv_mms}, $\maxm_b$, across meshes for $d=1$ in $G_\text{MS}$~\eqref{eq:G_mms}.}
\vskip-\dp\strutbox
\label{tab:G1}
\end{table}
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{G2_A000_linf.pdf}
\caption{$\theta=0^\circ$}
\end{subfigure}
\hspace{0.25em}
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{G2_A045_linf.pdf}
\caption{$\theta=45^\circ$}
\end{subfigure}
\\
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{G2_A090_linf.pdf}
\caption{$\theta=90^\circ$}
\end{subfigure}
\hspace{0.25em}
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{G2_A135_linf.pdf}
\caption{$\theta=135^\circ$}
\end{subfigure}
\caption{Error $\varepsilon$~\eqref{eq:error_norm} for $d=2$ in $G_\text{MS}$~\eqref{eq:G_mms}.}
\vskip-\dp\strutbox
\vskip10em
\label{fig:G2}
\end{figure}
\begin{table}[!b]
\centering
\begin{tabular}{c c c c c c c}
\toprule
& \multicolumn{3}{c}{Uniform} & \multicolumn{3}{c}{Twisted} \\
\cmidrule(lr){2-4} \cmidrule(lr){5-7}
$\theta$
& $\mathbf{Z}_\text{MS}$ & $\mathbf{Z}^\mathbf{A}$ & $\mathbf{Z}^\Phi$
& $\mathbf{Z}_\text{MS}$ & $\mathbf{Z}^\mathbf{A}$ & $\mathbf{Z}^\Phi$
\\
\midrule
$\phantom{0}\pz0^\circ$ & 18 & 18 & \pz7 & 18 & 18 & \pz7 \\
$\pz45^\circ$ & \rereading{31} & \rereading{31} & 11 & \rereading{31} & \rereading{31} & 11 \\
$\pz90^\circ$ & \rereading{31} & \rereading{31} & 11 & \rereading{31} & \rereading{31} & 11 \\
$135^\circ$ & \rereading{31} & \rereading{31} & 11 & \rereading{31} & \rereading{31} & 11 \\
\bottomrule
\end{tabular}
\caption{Maximum rank of~\eqref{eq:zjv_mms}, $\maxm_b$, across meshes for $d=2$ in $G_\text{MS}$~\eqref{eq:G_mms}.}
\vskip-\dp\strutbox
\label{tab:G2}
\end{table}
\begin{figure}[!t]
\centering
\includegraphics[scale=.64,clip=true,trim=0in 0in 0.5in 0in]{res.pdf}
\caption{$\mathbf{r}_\text{MS}$~\eqref{eq:zjv_mms} for the simulations in Figures~\ref{fig:G1} and~\ref{fig:G2}.}
\label{fig:res}
\end{figure}
\subsection{Incorrect Implementation}
In addition to showing that the approach of Section~\ref{sec:mms} confirms the expected accuracy in the absence of coding errors, it is \reviewerTwo{important} to show that the approach can detect coding errors as well. Given the low ranks of the constraints in Tables~\ref{tab:G1} and~\ref{tab:G2}, it is important to ensure the admitted solution is not correct and does not achieve the expected order of accuracy in the presence of coding errors.
We consider four coding errors to test the approach:
\begin{enumerate}[leftmargin=*,labelindent=\parindent,label=Case~\arabic*:,ref=\arabic*]
\item \label{ce:k} \textit{Incorrect value of $k$}. $k$ is increased by 1\%, incorrectly weighting the contribution to $\mathbf{Z}_\text{MS}$ from $\mathbf{Z}_\mathbf{A}$.
\item \label{ce:quad} \textit{Incorrect quadrature weights}. The weights are increased by 1\%, making the solutions to the integrals inconsistent.
\item \label{ce:mat_elem} \textit{Incorrect matrix entry}. $Z_{\text{MS}_{1,2}}$ is increased by 1\%.
\item \label{ce:area} \textit{Incorrect triangle areas}. Instead of using the correct areas for the triangle, the areas associated with the uniform mesh are used in the basis function evaluations.
\end{enumerate}
\begin{figure}[!t]
\centering
\includegraphics[scale=.64,clip=true,trim=2.3in 0in 2.8in 0in]{bugs_linf.pdf}
\caption{Error $\varepsilon$~\eqref{eq:error_norm} for multiple coding errors.}
\label{fig:bugs_linf}
\end{figure}
For the twisted mesh, with both contributions to $\mathbf{Z}_\text{MS}$, $d=1$ in $G_\text{MS}$~\eqref{eq:G_mms}, and $\theta=45^\circ$, Figure~\ref{fig:bugs_linf} shows the errors of the four coding errors, as well as the correct result from Figure~\ref{fig:G1:45}. Cases~\ref{ce:k}, \ref{ce:quad}, and \ref{ce:area} compromise the consistency of the governing equations, such that the error does not decrease with mesh refinement. On the other hand, Case~\ref{ce:mat_elem} remains consistent but increases the error and slows its rate of decay to first order. \reviewerTwo{The error for this case is able to decay because, as the mesh is refined, the spatial location of this fixed-column-index error moves toward the corner of the domain, where the normal components of the surface current vanish. Additional details are provided in~\ref{sec:app}. Nonetheless, by decaying at a slower rate than expected, this coding error is detected.}
| 1a22de47b077c0fb0ebd983c020299640a777426 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction
Let $K$ be a global field.
The zeta function of $K$ is an important invariant of $K$, encoding information about the places of $K$.
By classical work of Gassmann, the zeta function of $K$ alone does not determine $K$; for example, Gassmann produces non-isomorphic number fields with the same zeta function \cite{gassmann}. Similarly, Tate has shown that two function fields have the same zeta function if and only if the Jacobians of the corresponding curves are isogenous \cite{tate}. It is natural to ask if additional information of a similar nature helps.
\begin{Question} \label{question1}
Do zeta functions of Galois extensions of $K$ (or $L$-functions for particular characters of the Galois group) suffice to determine $K$ up to isomorphism? What kind of extensions are needed? Can this be done explicitly?
\end{Question}
There has been a number of recent papers addressing instances of this question.
\begin{itemize}
\item For example, \cite{CDL+} proves that a global field can be recovered
from its abelian $L$-functions. This requires a way to identify characters of two fields; more precisely, they prove two global fields are isomorphic if there is an identification of the groups of characters of their Abelianized absolute Galois groups preserving $L$-functions.
\item We have shown that function fields of curves
of genus at least two can be recovered from $L$-functions of unramified characters \cite{BV}.
\item Another approach is to attempt to recover extensions of a fixed global field from a few well-chosen $L$-functions. For number fields, see \cites{CDL+,LiRu,Pi} and the references therein, while \cite{solomatin} considers extensions of a fixed function field. A key difference is that all number fields are natural extensions of ${\mathbb Q}$, while function fields can be given as extensions of the rational function field in different ways and, in general, there is no canonical choice.
\end{itemize}
These results can all be seen as attempts to use limited information about the absolute Galois group of a global field to recover it. (Uchida and Neukirch have shown that the absolute Galois group determines a global field \cites{neukirch,uchida76,uchida77}.)
We will focus on $L$-functions of one-dimensional representations; for the purposes of this paper
a character of a group $G$ is a homomorphism $G \to \mathbb{C}^\times$.
Our goal is to study Question~\ref{question1} for function fields of
curves over finite fields with the goal of explicitly describing an effective,
finite set of $L$-functions of characters that characterizes the curve.
Our perspective is a bit different from the aforementioned results.
We work with a family of affine curves
described in some concrete way and we recover the description using $L$-functions for uniformly-constructed covers.
In other words, we aim to describe a set of $L$-functions that will determine
the member of the family up to some natural equivalence.
\begin{Example} \label{ex:intro}
As we will show later while proving Theorem~\ref{thm:equatoinintro}, we can recover the parameter $\lambda$ in
the Legendre family $y^2=x(x-1)(x - \lambda)$ over ${\mathbb F}_q$ using
the $L$-functions of the extensions $z^p-z=\alpha x(1-y^{q-1})$ with $\alpha$ running through a basis of ${\mathbb F}_q/{\mathbb F}_p$. See Example \ref{ex:legendre}.
\end{Example}
We first tackle the special case of the projective line minus four points in odd characteristic. In this case, the function field is known and the goal is to determine the removed points using $L$-functions up to automorphisms of the projective line. We look at $L$-functions of a suitable class field (a ray class field for the function field giving an unramified cover of the projective line minus four points) and answer an analogue of a conjecture from
\cite{SV} in a strong form. Namely, we prove:
\begin{Theorem} \label{thm:p1intro}
Let $q$ be an odd prime power and $U$ be ${\mathbb P}^1_{{\mathbb F}_q}$ minus four ${\mathbb F}_q$-rational points. Then $U$ can be recovered up to Frobenius twist from the $L$-functions
associated to the maximal abelian extension of ${\mathbb F}_q(x)$ unramified over $U$ and tamely ramified on its complement.
\end{Theorem}
We prove two versions of this theorem. Theorem~\ref{thm:p1distinguished} (and Remark~\ref{remark:q9} for an exceptional case) shows this can be accomplished using the $L$-function of a single well-chosen character.
Furthermore, we can identify which character to use given only the $L$-functions for every character of the extension.
Although we need to know $U$ to compute the $L$-function, the point is that the resulting
value of the $L$-function suffices to determine $U$ uniquely up to Frobenius
twist.
Along slightly different lines,
Theorem~\ref{thm:p1zilber} shows that if $U$ and $U'$ are ${\mathbb P}^1_{{\mathbb F}_q}$ minus four ${\mathbb F}_q$-rational points and there is an isomorphism between the Galois groups of appropriate ray class fields which identifies $L$-functions then $U$ and $U'$ are Frobenius twists.
The proofs rely on obtaining information about the reductions of these $L$-functions modulo primes above $p$ using the Cartier operator, making use of a result of R\"{u}ck \cite{ruck}.
\begin{remark}
\begin{enumerate}
\item The first version of the theorem fits into the approach taken in \cite{LiRu} and \cite{solomatin}, while the second fits into the approach taken in \cites{BV} and \cite{CDL+} as it involves identifying characters.
\item
Many of our arguments extend without change to
characteristic $2$ but a few do not. In some places, we will comment on
the changes that would need to be made to extend the results but we do not
systematically address the issue.
\item Our proof in \cite{BV} used a conjecture of Bogomolov, Korotiaev and Tschinkel about Jacobians of curves of genus at least two which was proven by Zilber \cites{BKT,Z1,Z2}. In Section~\ref{ss:backgroundl}, we extend this conjecture to generalized Jacobians of dimension at least two, and can reverse our argument and use Theorem~\ref{thm:p1zilber} to establish a special case of this extended conjecture.
\end{enumerate}
\end{remark}
Then we tackle the general question of recovering the equation of a curve using $L$-functions of Artin-Schreier extensions.
\begin{Theorem} \label{thm:equatoinintro}
For fixed odd prime power $q=p^r$ and $d\ge 1$, there is an explicit finite
set of polynomials $f \in {\mathbb F}_q[x,y]$ such that we can
recover the coefficients of any absolutely irreducible $F \in {\mathbb F}_q[x,y]$ of degree $d$, defining a function field $K$,
from the $L$-functions of the Artin-Schreier extensions of $K$ given by $z^p-z=f$.
\end{Theorem}
This is Theorem~\ref{thm:equations}. It uses a very large number of $L$-functions, but can be simplified given additional information about the form of $F$ as illustrated in Example~\ref{ex:intro}.
\begin{remark}
There are many irreducible polynomials $F \in {\mathbb F}_q[x,y]$ of degree $d$ giving plane curves with the same function field $K$.
It is again worth emphasizing that in Theorem~\ref{thm:equatoinintro} we use the explicit description of $K$ as the function field of the curve given by $F$ in order to describe Artin-Schreier extensions of $K$ and then recover the description of $K$ by $F$ using $L$-functions of these extensions. This is why we recover a specific $F$ and not just the function field $K$ up to isomorphism.
For example, when $-1$ is a square in ${\mathbb F}_q$ the fibers of the Legendre family over $\lambda$ and $1 -\lambda$ are isomorphic (via $(x,y) \mapsto (1-x,\sqrt{-1} y)$), but Example~\ref{ex:intro} is able to recover $\lambda$ because the Artin-Schreier extensions used depended on the fiber. Identifying the function fields of the two fibers, we end up looking at a different set of $L$-functions!
\end{remark}
\subsection*{Acknowledgements}
The authors were supported by the Marsden Fund Council administered by the Royal Society of New Zealand. We are grateful for the assistance of Z. Brady for his help with Section~\ref{ss:technicalproof}, in particular the proofs of Proposition~\ref{prop:charsdiffer} and Lemma~\ref{lem:witness} and to the referee for the careful reading of the manuscript.
\section{\texorpdfstring{$L$}{L}-functions and Generalized Jacobians}
\subsection{Class Field Theory}
We describe the well-known connection between the class-field theoretic notions of Hilbert class field and ray class field with the language of generalized Jacobians in the function field setting. Since we wish our extensions to have the same constant field (and have finite degree), we use the version of Hilbert and ray class fields introduced in \cites{rosen,hk}.
Let $K/{\mathbb F}_q$ be a function field, $\mathcal{S}_K$ be the set of places of $K$ and $S \subset \mathcal{S}_K$ a finite subset. Let $I_K$ denote the ideles of $K$ and $I_{K,S}$ the subgroup
of the {ideles $(\alpha_v)_{v \in \mathcal{S}_K}$ such that $\alpha_v=1$ for $v \in S$}.
We let ${\mathfrak m}$ be a modulus (a formal sum of places of $K$ or, equivalently, an effective divisor) disjoint from $S$. We allow ${\mathfrak m}=(0)$, the empty divisor. {Define $I_{K,S}^{{\mathfrak m}}$ to be the elements of $I_{K,S}$ which are congruent to $1$ modulo ${\mathfrak m}$.}
The ray class field of $K$ with modulus ${\mathfrak m}$ is defined to be the field corresponding to the subgroup $K^\times I_{K,S}^{{\mathfrak m}} / K^\times$ of $I_K / K^\times$ under the class field theory correspondence. We denote this field by $K_{S}^{\mathfrak m}$.
\begin{Lemma} \label{lem:maximality}
The maximal Abelian extension of $K$ in which the places of $S$ split completely and with conductor dividing ${\mathfrak m}$ is $K_S^{\mathfrak m}$.
\end{Lemma}
\begin{proof}
This is an extension of \cite[Theorem (iv)]{hk}. Let $L$ be an Abelian extension of $K$ in which the places of $S$ split completely and with conductor dividing ${\mathfrak m}$. Under the class field theory correspondence, it corresponds to an open subgroup $W \subset I_K$ containing $K^\times$. For a place $v \in S$, we know that $K_v^\times \subset W$ as $v$ splits completely in $L$. Since the conductor of $L$ divides ${\mathfrak m}$, we know that $$I^{\mathfrak m}_{K,S} \subset W.$$
This shows that $K^\times I_{K,S}^{\mathfrak m} \subset W$, and hence $L \subset K_{S}^{{\mathfrak m}}$.
\end{proof}
Now let $C$ be a smooth projective curve over ${\mathbb F}_q$ with function field $K$. We fix a rational point $M$ of $C$, and take $S = \{v_M\}$ where $v_M$ is the place associated to $M$. Note that the splitting condition implies the constant fields {of $K$ and $K_S^{\mathfrak m}$ are the same}.
We consider an abelian extension of $K$ (abelian covers of $C$) such that $v_M$ splits completely (the cover is defined over ${\mathbb F}_q$ and the fiber above $M$ is rational) subject to additional ramification conditions.
Let $J_{\mathfrak m}$ denote the generalized Jacobian of $C$ with respect to the modulus ${\mathfrak m}$ \cite{Serre}.
Let ${\rm supp}({\mathfrak m})$ denote the support of ${\mathfrak m}$, i.e, the set of places of $K$ occurring with nonzero multiplicity in ${\mathfrak m}$.
The Abel-Jacobi map $C \setminus {\rm supp}({\mathfrak m}) \to J_{\mathfrak m}$ associated with a
divisor $D_1$ of degree one is the map $P \mapsto [P - D_1]$. Let $\Phi$ denote the Frobenius endomorphism of $J_{\mathfrak m}$.
\begin{prop} \label{prop:cftgj}
The ray class field of $K$ with modulus ${\mathfrak m}$ is the function field of the cover
$C'$ of $C$ given by the fiber product of the isogeny $\Phi^* - \operatorname{id} : J_{\mathfrak m} \to J_{\mathfrak m}$ with the Abel-Jacobi map associated to $M$.
This induces an isomorphism $\Gal(K_{S}^{{\mathfrak m}}/K) \simeq J_{\mathfrak m}({\mathbb F}_q) $ and, viewing $P \in \mathcal{S}_K$ as
a divisor of degree $\deg(P)$,
the Frobenius $\operatorname{Frob}_P$ is identified with $[P - \deg(P) D_1] \in J_{\mathfrak m}({\mathbb F}_q)$.
\end{prop}
\begin{proof}
Let $L$ be the function field of $C'$. Observe that $v_M$ splits completely in $L$ as the fiber of the map $\Phi^* - 1$ over $0$ is $J_{\mathfrak m}({\mathbb F}_q)$. We see the cover is Galois, with Abelian Galois group $J_{\mathfrak m}({\mathbb F}_q)$.
Furthermore, the conductor of $L$ divides ${\mathfrak m}$ because the cover $C' \to C \setminus {\rm supp}({\mathfrak m})$ is the pullback of an isogeny $J' \to J_{\mathfrak m}$.
Then Lemma~\ref{lem:maximality} shows that
$L \subset K_{S}^{{\mathfrak m}}$, so it suffices to show that $J_{\mathfrak m}({\mathbb F}_q) \simeq \Gal(K_{S}^{{\mathfrak m}}/K)$.
But class field theory in the form of \cite[Lemma 2 and Theorem]{hk} shows that $\Gal(K_{S}^{{\mathfrak m}}/K)$ is isomorphic to the idele $S$-ray class group modulo ${\mathfrak m}$. There is a natural surjection
from the set of divisors on $C$ with support away from ${\mathfrak m}$ to the set of divisors of degree $0$ on $C$ whose support is disjoint from ${\mathfrak m}$ given by sending a divisor $D$ to $D - \deg(D) M$.
As in \cite[V \S 2]{Serre}, let $C_{\mathfrak m}^0$ denote the set of degree zero divisors on $C$ with support disjoint from ${\mathfrak m}$ up to ${\mathfrak m}$-equivalence; it is a quotient of the group of divisors of degree $0$ on $C$ whose support is disjoint from ${\mathfrak m}$ by the divisors of functions on $C$ which are congruent to one modulo ${\mathfrak m}$. We know that $C_{\mathfrak m}^0 = J_{\mathfrak m}({\mathbb F}_q)$, and
every equivalence classes of divisors on $C$ with support disjoint from ${\mathfrak m}$ can be represented by a divisor whose support is disjoint from $S$ as well. But the idele $S$-ray class group modulo ${\mathfrak m}$ is precisely divisors whose support is disjoint from $S$ and ${\mathfrak m}$ modulo the divisors of functions which are congruent to one modulo ${\mathfrak m}$ and are invertible at the points of $S$. This completes the proof.
\end{proof}
\subsection{\texorpdfstring{$L$}{L}-functions} \label{ss:backgroundl} By a \emph{pointed curve} $(C,D_1)$, we mean a smooth projective curve $C$ over ${\mathbb F}_q$ together with a choice of degree $1$ divisor $D_1$. Such a divisor always exists \cite{Schmidt1931}.
To connect with class field theory, we assume the existence of a rational point $M$ and take $D_1 = [M]$.
Let $\chi$ be a character of $\Gal(K_{S}^{{\mathfrak m}}/K) \simeq J_{\mathfrak m}({\mathbb F}_q)$ with conductor ${\mathfrak m}$; this means ${\mathfrak m}$ is the smallest modulus such that $\chi$ factors through $\Gal(K_{S}^{{\mathfrak m}}/K)$. It is a character of the absolute Galois group of $K$ that is
trivial on the absolute Galois group of ${\mathbb F}_q$.
\begin{Definition}
For a place $P$ of $K$ not in
${\rm supp}({\mathfrak m})$,
let $\operatorname{Frob}_P \in \Gal(K_{S}^{{\mathfrak m}}/K)$ denote the Frobenius at $P$.
For any character $\chi$ with conductor ${\mathfrak m}$ and ${\rm supp}({\mathfrak m})$ disjoint from $S$, the $L$-series (or $L$-function) for $\chi$ is defined as
\[
L(T,C,\chi) := \prod_{P \in \mathcal{S}_C\setminus {\rm supp}({\mathfrak m})} ( 1 - \chi(\operatorname{Frob}_P) T^{\deg P})^{-1}.
\]
\end{Definition}
We sometimes use the notation $L(T,D/C,\chi)$ in the case of a Galois cover
$D/C$ if we can view $\chi$ as a character of $\Gal(D/C)$.
Note that this definition depends on the choice of marked point $M$ as the Abel-Jacobi map determined by $M$ is used to identify the Galois group with $J_{\mathfrak m}({\mathbb F}_q)$.
It is a standard calculation that
\begin{equation}
\label{eq:standard}
\frac{L'(T,C,\chi)}{L(T,C,\chi)} =
\sum_{P \in \mathcal{S}_C}
\frac{(\deg P) \chi(\operatorname{Frob}_P) T^{\deg P -1}}{( 1 - \chi(\operatorname{Frob}_P) T^{\deg P})} =
\frac{d}{dT}\left(\sum_{n=1}^{\infty} S_n(\chi) \frac{T^n}{n}\right)
\end{equation}
where
$\displaystyle S_n(\chi) = \sum_{P \in U({\mathbb F}_{q^n})} \chi(P+\Phi(P)+\cdots+\Phi^{n-1}(P))$ and $U = C \setminus {\rm supp}({\mathfrak m})$.
One of the main purposes of this paper is to study the following
conjecture:
\begin{Conjecture} \label{conj:lfun}
Let $(C,[M])$ and $(C',[M'])$ be smooth irreducible projective pointed curves
over a finite field~${\mathbb F}_q$. Let ${\mathfrak m},{\mathfrak m}'$ be moduli on $C,C'$ disjoint from $M,$ $M'$ respectively. Suppose the corresponding generalized
Jacobians $J_{{\mathfrak m}},J_{{\mathfrak m}'}'$ have dimension at least two, and that there is a set-theoretic map $\psi : J_{{\mathfrak m}'}'(\overline{{\mathbb F}}_q) \to J_{{\mathfrak m}}(\overline{{\mathbb F}}_q)$ inducing an isomorphism of groups between $J_{{\mathfrak m}}({\mathbb F}_{q^n})$ and $J_{{\mathfrak m}'}'({\mathbb F}_{q^n})$
for every $n \ge 1$.
If $L(T, C \otimes {\mathbb F}_{q^n}, \chi) = L(T, C' \otimes {\mathbb F}_{q^n}, \chi \circ \psi|_{
J_{{\mathfrak m}'}'({\mathbb F}_{q^n})})$ for all $n$ and all characters $\chi$ of $J_{{\mathfrak m}}({\mathbb F}_{q^n})$, then
$C$ and $C'$ are Frobenius twists of each other and the map
$\psi$ arises from a morphism of curves composed with a limit of Frobenius maps
which sends $M$ to $M'$ and ${\mathfrak m}$ to ${\mathfrak m}'$.
\end{Conjecture}
\begin{remark}
More precisely, the last condition means that
there exists an integer $m$, an isomorphism $\alpha: \operatorname{Frob}^m(C) \simeq C'$, and a generalized Frobenius $\beta: \overline{{\mathbb F}}_q \to \overline{{\mathbb F}}_q$ restricting to $\operatorname{Frob}^{-m}$ on ${\mathbb F}_q$ such that $\alpha \circ \beta : C \to C'$ sends $M$ to $M'$, ${\mathfrak m}$ to ${\mathfrak m}'$, and such that $\alpha \circ \beta$ induces $\psi$ on generalized Jacobians.
\end{remark}
Conjecture \ref{conj:lfun} was proved in \cite{BV} for ${\mathfrak m} = (0)$ as a consequence of a result conjectured
by Bogomolov, Korotiaev and Tschinkel \cite{BKT} and proved by
Zilber \cites{Z1,Z2}. That result is the ${\mathfrak m} = (0)$ case of the following
conjecture.
\begin{Conjecture} \label{conj:zilber}
Let $(C,[M])$ and $(C',[M'])$ be smooth, projective, irreducible pointed curves over ${\mathbb F}_q$
and ${\mathfrak m},{\mathfrak m}'$ moduli on $C,C'$ respectively. Suppose the generalized
Jacobians $J_{{\mathfrak m}}$ and $J_{{\mathfrak m}'}'$ have dimension at least two, and let $U , U'$ be the complements of the support of ${\mathfrak m},{\mathfrak m}'$ in $C,C'$ respectively. If $\psi : J_{{\mathfrak m}}(\overline{{\mathbb F}}_q) \to J_{{\mathfrak m}'}'(\overline{{\mathbb F}}_q)$ is a group isomorphism such that $\psi(U(\overline{{\mathbb F}}_q)) = U'(\overline{{\mathbb F}}_q)$ (with respect to the Abel-Jacobi embeddings determined by $M$ and $M'$), then $\psi$ arises from a morphism of curves composed with a limit of Frobenius maps which sends ${\mathfrak m}$ to ${\mathfrak m}'$.
More precisely, there exists an integer $m$, an isomorphism $\alpha: \operatorname{Frob}^m(C) \simeq C'$, and a generalized Frobenius $\beta: \overline{{\mathbb F}}_q \to \overline{{\mathbb F}}_q$ restricting to $\operatorname{Frob}^{-m}$ on ${\mathbb F}_q$ such that
$\alpha \circ \beta : C \to C'$ sends ${\mathfrak m}$ to ${\mathfrak m}'$ and $M$ to $M'$, and that
induces $\psi$ on generalized Jacobians.
\end{Conjecture}
The following result relates the two above conjectures.
\begin{Theorem} \label{thm:equivalence}
Conjecture \ref{conj:zilber} implies Conjecture \ref{conj:lfun}. Conversely,
assuming Conjecture \ref{conj:lfun},
the hypotheses of Conjecture \ref{conj:zilber} and, in addition, that
a power of Frobenius commutes with $\psi$, then the conclusion of
Conjecture \ref{conj:zilber} holds.
\end{Theorem}
\begin{proof}
Assume Conjecture \ref{conj:lfun} and, in addition, that the $m$-th
power of Frobenius commutes with $\psi$.
It follows directly that the exponential sums $S_n(\chi)$ for $m|n$ in \eqref{eq:standard} match for each character.
Hence \eqref{eq:standard} gives equality of
$L$-functions over extensions of ${\mathbb F}_{q^m}$ and from that we get the morphism.
The proof in the opposite direction is almost verbatim the proof of
\cite[Theorem 2.5]{BV} replacing linear equivalence of divisors with
the stricter equivalence modulo the fixed modulus. The argument there works without
change unless the curves have genus zero. In this latter case, points
are linearly equivalent to each other so we need to show that they are not equivalent
modulo a fixed modulus under the hypothesis of Conjecture \ref{conj:zilber}. Such a modulus has degree at least two.
If two points in ${\mathbb P}^1$ given by $a, b$ are equivalent modulo ${\mathfrak m}$ and
we assume (without loss of generality) that $\infty$ is in the support of ${\mathfrak m}$, then we conclude that
$(x-a)/(x-b) \equiv 1 \pmod{{\mathfrak m}}$. We distinguish two cases. First assume that there is
a point $c$ other than $\infty$ in the support of ${\mathfrak m}$. As
$(x-a)/(x-b) \equiv 1 \pmod{{\mathfrak m}}$ we see $(c-a)/(c-b) = 1$ and hence $a=b$.
If ${\mathfrak m}$ is supported only at infinity, our hypothesis requires
$(x-a)/(x-b) \equiv 1 \pmod{2\infty}$. But
$$(x-a)/(x-b) = (1-a/x)/(1-b/x) = 1 +(b-a)/x +O(1/x^2)$$
and hence $(x-a)/(x-b) \equiv 1 \pmod{2\infty}$ only if $a=b$.
\end{proof}
\begin{Remark}
In the next section, we will prove a strong form of Conjecture \ref{conj:lfun}
when $C = {\mathbb P}^1$ and ${\mathfrak m}$ is a sum of four distinct places of degree one.
Under these hypotheses
the generalized Jacobians are tori \cite[VI \S 3]{Serre}.
For tori, the ${\mathbb F}_q$-Frobenius is multiplication by $q$ in the group law so it commutes with $\psi$
automatically. Theorem \ref{thm:equivalence} will show that Conjecture \ref{conj:zilber} holds in this case.
In \cite[Theorem 5.11]{BKT} it was proved that a power of Frobenius commutes with $\psi$ in the case
${\mathfrak m} =(0)$ before conjecture \ref{conj:zilber} was proved for ${\mathfrak m}=(0)$. It would be interesting to see if the
argument there can be generalized.
If the generalized Jacobians have dimension one, the condition on the image
of the Abel-Jacobi embedding in Conjecture \ref{conj:zilber} is clearly vacuous and
the condition on the $L$-functions in Conjecture \ref{conj:lfun} can also
be shown to be vacuous. In the special case that the curves have genus one
and both moduli are $(0)$, \cite[Remark 2.7]{BV} provides examples where
the conclusion of both conjectures fail. The other possibility is that
the curves have genus zero and the moduli have degree two. In this case,
the curves are automatically isomorphic under the conjectures' hypotheses
but it is possible to construct
group isomorphisms $\psi$ that do not arise from morphisms of curves.
\end{Remark}
\section{The Projective Line
\label{sec:p1}
Let $q=p^r$ be an odd prime power. In this section, we will show how to recover ${\mathbb P}^1_{{\mathbb F}_q}$ with four marked rational points from $L$-functions of covers unramified away from the marked points.
\subsection{Covers of the Projective Line}
For $\lambda \in {\mathbb F}_q \backslash \{0,1\}$, consider the modulus $${\mathfrak m}(\lambda) = [0] + [1] + [\lambda]$$ on ${\mathbb P}^1_{{\mathbb F}_q}$ and let $U_\lambda \colonequals {\mathbb P}^1_{{\mathbb F}_q} \backslash \{0,1, \lambda\}$.
\begin{Definition}
Let $C_\lambda$ be the projective curve over ${\mathbb F}_q$ given by the projective equation
\begin{equation} \label{eq:clambda}
\lambda x^{q-1} - y^{q-1} + (1-\lambda)z^{q-1} =0.
\end{equation}
Let $\pi : C_\lambda \to {\mathbb P}^1_{{\mathbb F}_q}$ be the map sending $[x,y,z]$ to $[z^{q-1},z^{q-1}-x^{q-1}]$.
\end{Definition}
A straightforward calculation checks that $C_\lambda$ is smooth. Note there is a natural action of $\mu_{q-1}^3$ modulo the diagonal $\mu_{q-1}$, acting by roots of unity on the three coordinates.
\begin{Proposition} \label{prop:rcf}
The cover $\pi: C_\lambda \to {\mathbb P}^1_{{\mathbb F}_q}$ is the ray class field cover of ${\mathbb P}^1_{{\mathbb F}_q}$ with respect to ${\mathfrak m}(\lambda)$ in which the infinite place splits completely. In particular,
\begin{enumerate}
\item \label{rcf1} $\pi$ is a degree $(q-1)^2$ branched Galois cover with Galois group $\mu_{q-1}^2$;
\item \label{rcf2} there are $q-1$ points of $C_\lambda$ with ramification index $q-1$ above each point of ${\mathfrak m}(\lambda)$, and the infinite place splits completely;
\item \label{rcf3} over $U_\lambda$, $\pi$ is the pullback of $\Phi^* - 1 : J_{{\mathfrak m}(\lambda)} \to J_{{\mathfrak m}(\lambda)}$ by the Abel-Jacobi map $U_\lambda \to J_{{\mathfrak m}(\lambda)}$ which sends $t$ to $[t] - [\infty]$ (as before $\Phi$ is the Frobenius map on $J_{\mathfrak m}$).
\end{enumerate}
\end{Proposition}
\begin{proof}
We know that $J_{{\mathfrak m}(\lambda)}$ is isomorphic to $\mathbb{G}_m^3$ modulo the diagonal $\mathbb{G}_m$ \cite[V \S3.13 Prop 7]{Serre};
hence $J_{{\mathfrak m}(\lambda)} \simeq \mathbb{G}_m^2$ by choosing the first coordinate to be $1$. The Abel-Jacobi map sends $t$ to the divisor $[t]-[\infty]$ which is the divisor of the function $f(x)=x-t$. Thus the Abel-Jacobi map is given by sending $t$ to $[(-t,1-t,\lambda-t)]$ in $\mathbb{G}_m^3/\mathbb{G}_m$, or equivalently sending $t$ to $(1-1/t,1-\lambda/t)$ in $J_{{\mathfrak m}(\lambda)} \simeq \mathbb{G}_m^2$.
Letting $x$ and $y$ be the coordinates of $J_{{\mathfrak m}(\lambda)}$, as the Frobenius on $\mathbb{G}_m$ is the $q$th power map pulling back $\Phi^* - 1$ gives the affine curve
\[
x^{q-1} = (1-1/t), \quad y^{q-1} = (1 - \lambda/t).
\]
Eliminating $t$ and writing it as a projective equation gives \eqref{eq:clambda} and the map $\pi$. We directly see the claim about ramification. Note that $\infty$ splits completely over ${\mathbb F}_q$ as the Abel-Jacobi map sends $\infty$ to $(1,1) \in \mathbb{G}_m^2$ and the fiber over $(1,1)$ is $\mu_{q-1} \times \mu_{q-1}$.
This shows that $C_\lambda$ is the ray class field cover of ${\mathbb P}^1_{{\mathbb F}_q}$ with respect to ${\mathfrak m}(\lambda)$ in which the infinite place splits completely. The Galois group is identified with $J_{{\mathfrak m}(\lambda)}({\mathbb F}_q) \simeq \mu_{q-1}^2$ by having the group act via multiplication by roots of unity on $x$ and $y$.
\end{proof}
\begin{Corollary}
The genus of $C_\lambda$ is $\frac{1}{2} (q-2)(q-3)$.
\end{Corollary}
\begin{proof}
This follows from the Riemann-Hurwitz formula, or the fact that $C_\lambda$ is a smooth plane curve of degree $q-1$.
\end{proof}
We now record some additional information about $C_\lambda$ for use later.
Let $G \colonequals \mu_{q-1}^2 \simeq J_{{\mathfrak m}(\lambda)}({\mathbb F}_q)$ be the Galois group of $\pi$, acting on $x$ and $y$ via roots of unity. Let $G_1$ (resp. $G_\lambda$) be copies of $\mu_{q-1}$ embedded into $G$ as the first (resp. second) coordinates, and $G_0$ be a copy of $\mu_{q-1}$ embedded diagonally into $G$.
\begin{Lemma} \label{lem:inertia}
The inertia groups at the $q-1$ points of $C_\lambda$ over $0$ (resp. $1$, $\lambda$) are $G_0$ (resp. $G_1$, $G_\lambda$).
\end{Lemma}
\begin{proof}
Computing using $\pi$, we see the points of $C_\lambda$ over $0$ are those with $z=0$, and hence are fixed by $G_0$. The other branch points are similar.
\end{proof}
\begin{Lemma} \label{lem:differentials}
A basis for the space of regular differentials on the projective curve given by an affine equation $a x^{q-1} + b y^{q-1} + 1=0$ over ${\mathbb F}_q$ is given by $\displaystyle \omega_{i,j} \colonequals \frac{x^i y^j}{y^{q-2}} dx$ where $i,j \geq 0$ and $i+j \leq q-4$.
\end{Lemma}
\begin{proof}
This is standard.
\end{proof}
\subsection{L-Functions and the Cartier Operator} \label{ss:lfunctions}
Recall that the group $G = \mu_{q-1}^2$ is the Galois group of the cover $\pi : C_\lambda \to {\mathbb P}^1_{{\mathbb F}_q}$.
Given a pair $(b,c) \in {\mathbb Z}/(q-1)$, we define a character $\chi_{b,c} : \mu_{q-1}^2 \to \mu_{q-1}$ by
\[
\chi_{b,c}(\zeta_1,\zeta_2) = \zeta_1 ^ b \zeta_2^c .
\]
Any character of $G$ equals $\chi_{b,c}$ for appropriately chosen $(b,c) \in {\mathbb Z}/(q-1)^2$. Let $H_{b,c}$ denote the kernel of $\chi_{b,c}$.
We will use the Cartier operator to obtain information about the $L$-functions for the cover $\pi : C_\lambda \to {\mathbb P}^1_{{\mathbb F}_q}$. For our purposes, the Cartier operator may be viewed as a (semi-linear) operator $\mathscr{C}$ on the space of regular differentials $H^0(C_\lambda, \Omega^1_{C_\lambda})$. If $q = p^r$, the $r$th power of the Cartier operator on $C_\lambda$ is ${\mathbb F}_q$-linear and may be computed explicitly using the following result. We define $\alpha_{m,n}$ to be the multinomial coefficient
\begin{equation} \label{eq:defalpha}
\alpha_{m,n} \colonequals \binom{q-1}{m,n,q-1-m-n} = \frac{(q-1)!}{m!\,n!\,(q-1-m-n)!}.
\end{equation}
\begin{Proposition} \label{prop:cartieraction}
Let $X$ be the projective curve over ${\mathbb F}_q$ determined by the equation $f(x,y)=ax^{q-1}+by^{q-1}+1=0$ with $ab \neq 0$. Writing $q = p^r$, the $r$-th power of the Cartier operator acts diagonalizably on $H^0(X,\Omega^1_X)$ with eigenvalues
\[
\alpha_{i+1,j+1} a^{i+1} b^{j+1} \quad \text{for } 0 \leq i,j \text{ with } i+j \leq q-4.
\]
The corresponding eigenvectors are $\displaystyle \omega_{i,j} = \frac{x^i y^j }{y^{q-2}} dx$.
\end{Proposition}
Note that the eigenvalues are non-zero if $q=p$ but can be zero when
$q \ne p$ as $\alpha_{i+1,j+1}$ may be zero modulo $p$.
\begin{proof}
By \cite[Theorem 1.1 and proof of Theorem 4.1]{StV} the $r$-th power of the Cartier operator on a regular differential $h f_y^{-1} dx$ (with $h \in {\mathbb F}_q[x,y]$)
is given by
$$
\mathscr{C}^{r}\left(h f_{y}^{-1} d x\right)=\left(\left(\frac{\partial}{\partial x}\right)^{(q-1)}\left(\frac{\partial}{\partial y}\right)^{(q-1)}\left(f^{q-1} h\right)\right)^{1/q} f_{y}^{-1} d x
$$
where $\left(\frac{\partial}{\partial x}\right)^{(q-1)}$ and $\left(\frac{\partial}{\partial y}\right)^{(q-1)}$ are partial Hasse derivatives.
The space of regular differentials of $X$ is
$\left\{h f_{y}^{-1} d x | h \in {\mathbb F}_q[x, y] , \deg h \le q-4\right\}$ as in Lemma \ref{lem:differentials}.
We will see that the action of $\mathscr{C}^{r}$ on the basis obtained by taking $h=x^i y^j$ is diagonal.
To see this, notice that $\left(\frac{\partial}{\partial y}\right)^{(q-1)} ( x^i y^j) =0$ unless $j \equiv q-1 \pmod{q}$, with a similar statement for the partial Hasse derivative with respect to $x$. The only term of $f^{q-1} x^i y^j = (a x^{q-1} + b y^{q-1} +1)^{q-1} x^i y^j$ which contributes
is the $x^{(i+1)(q-1)} y^{(j+1) (q-1)}$ term of $ f^{q-1}$. Thus we see that
\begin{align*}
\mathscr{C}^{r} \left( x^i y^j f_y^{-1} dx \right) &= \left( \left(\frac{\partial}{\partial x}\right)^{(q-1)}\left(\frac{\partial}{\partial y}\right)^{(q-1)} (\alpha_{i+1,j+1} a^{i+1} b^{j+1} x^{(i+1)(q-1) +i} y^{(j+1)(q-1)+j} ) \right)^{1/q} f_y^{-1} dx \\
&= \alpha_{i+1,j+1} a^{i+1} b^{j+1} x^i y^j f_y^{-1} dx. \qedhere
\end{align*}
\end{proof}
The Cartier operator gives significant arithmetic information about smooth projective curves over ${\mathbb F}_q$, in particular the curve $C_\lambda$.
\begin{Definition}
The $p$-rank of $C_\lambda$, denoted $f_{C_\lambda}$, is the ${\mathbb F}_q$-dimension of $H^0(C_\lambda,\Omega^1_{C_\lambda})^{\textrm{ss}}$, the subspace of $H^0(C_\lambda,\Omega^1_{C_\lambda})$ where $\mathscr{C}$ is invertible.
More generally, for a character $\chi : G \to \mu_{q-1}$ we define $f_{C_\lambda,\chi} := \dim_{{\mathbb F}_q} H^0(C_\lambda,\Omega^1_{C_\lambda})^{\chi,\textrm{ss}}$ to be the dimension of the subspace of regular differentials where $\mathscr{C}$ acts semisimply and $\Gal(C_\lambda/{\mathbb P}^1_{{\mathbb F}_q} ) =G$ acts via $\chi$.
We also let $P_\lambda(T,\chi)$ denote the numerator of the Artin $L$-function $L(T,C_\lambda/{\mathbb P}^1_{{\mathbb F}_q},\chi)$.
\end{Definition}
Note that $P_\lambda(T,\chi) = L(T,C_{\lambda}/{\mathbb P}^1_{{\mathbb F}_q},\chi)$ unless $\chi$ is trivial, in which case $L(T,C_\lambda/{\mathbb P}^1_{{\mathbb F}_q},1)$ has a denominator.
\begin{Proposition} \label{prop:ruck}
Letting $q=p^r$, we have
\begin{equation} \label{eq:ruck}
P_\lambda(T,\chi) \equiv \det( 1 - \mathscr{C}^r T | H^0(C_\lambda, \Omega^1_{C_\lambda})^{\chi}) \pmod{p}.
\end{equation}
In particular, the degree of $P_\lambda(T,\chi)$ is $f_{C_\lambda,\chi}$.
\end{Proposition}
\begin{proof}
This is a special case of a result of R\"{u}ck \cite[Theorem 4.1 and Corollary 4.1]{ruck}. Note that $P_\lambda(T,\chi)$ can be viewed as a polynomial with coefficients in the ring of Witt vectors $W({\mathbb F}_q)$ as this ring contains the $(q-1)$-st roots of unity, so reducing modulo $p$ makes sense.
The identification of $\mu_{q-1} \subset W({\mathbb F}_q)$ with the complex $(q-1)$-st
roots of unity (used to define the $L$-function) requires an arbitrary choice; we fix one throughout.
\end{proof}
\begin{Remark}
The result of R\"{u}ck generalizes a result of Manin \cite{manin} which relates the numerator of the zeta function with the action of the Cartier operator on the space of regular differentials.
\end{Remark}
\begin{Proposition} \label{prop:charcomputation}
Fix a non-trivial character $\chi_{b,c}$ and let $i \colonequals (b-1 \mod{q-1})$ and $j \colonequals (c-1 \mod{q-1})$.
If $i + j \leq q-4$ and $\alpha_{i+1,j+1} \not \equiv 0 \pmod{p}$ then $f_{C,\chi_{b,c}} =1$ and
\[
P_\lambda(T,\chi_{b,c}) \equiv 1 - \alpha_{i+1,j+1} \left( \frac{ \lambda}{1-\lambda} \right) ^{i+1} \left( \frac{ 1}{\lambda-1} \right)^{j+1} T \pmod{p}.
\]
Otherwise $f_{C_\lambda,\chi_{b,c}} =0$ and $P_\lambda(T,\chi_{b,c}) \equiv 1 \pmod{p}$.
\end{Proposition}
\begin{proof}
Consider the affine patch of $C_\lambda$ obtained by taking $z=1$, given by
\begin{equation}
\frac{\lambda}{1-\lambda} x^{q-1} + \frac{1}{\lambda-1} y^{q-1} + 1 =0.
\end{equation}
A basis for the regular differentials on $C_\lambda$ in this coordinate system is given by
\[
\omega_{m,n} \colonequals \frac{x^m y^n }{y^{q-2}} dx
\]
with $m,n \geq 0$ and $m + n \leq q-4$ by Lemma~\ref{lem:differentials}.
The Galois action of $G = \mu_{q-1}^2$ is acting by roots of unity on $x$ and $y$, so for $(\zeta_1,\zeta_2) \in \mu_{q-1}^2$ we compute that
\[
(\zeta_1,\zeta_2)^* \omega_{m,n} = \frac{(\zeta_1 x)^m (\zeta_2 y)^n }{(\zeta_2 y)^{q-2}} d(\zeta_1 x) = \zeta_1^{m+1} \zeta_2^{n-(q-2)} \omega_{m,n} = \zeta_1^{m+1} \zeta_2^{n+1} \omega_{m,n}.
\]
On the other hand, $\chi_{b,c}(\zeta_1,\zeta_2) = \zeta_1^b \zeta_2^c$. Thus $\omega_{m,n}$ lies in $H^0(\Omega_{C_\lambda}^1)^{\chi_{b,c}}$ if and only if $m +1 \equiv b \pmod{q-1}$ and $n+1 \equiv c \pmod{q-1}$.
Thus we see that $H^0(\Omega_{C_\lambda}^1)^{\chi_{b,c}}$ is one dimensional if and only if $i + j \leq q-4$.
(When this happens, $\omega_{i,j}$ spans the space.)
If $i +j > q-4$, the dimension of $H^0(\Omega_{C_\lambda}^1)^{\chi_{b,c}}$ is zero and hence $f_{C_\lambda,\chi_{b,c}} =0$ and $P_\lambda(t,\chi_{b,c}) =0$ by Proposition~\ref{prop:ruck}. If $i + j \leq q-4$ then Proposition~\ref{prop:cartieraction} shows that
\[
\mathscr{C}^r \omega_{i,j} = \alpha_{i+1,j+1} \left(\frac{\lambda}{1-\lambda} \right)^{i+1} \left(\frac{1}{\lambda-1}\right)^{j+1} \omega_{i,j}.
\]
By Proposition~\ref{prop:ruck}, we conclude that
\[
P_\lambda(T,\chi_{b,c}) \equiv 1 - \alpha_{i+1,j+1} \left(\frac{\lambda}{1-\lambda} \right)^{i+1} \left(\frac{1}{\lambda-1}\right)^{j+1} T \pmod{p}. \qedhere
\]
\end{proof}
\begin{Corollary} \label{cor:congruence}
If $q \neq 3$, we have that $P_\lambda(T,\chi_{1,1}) \equiv 1 + 2 \lambda (\lambda-1)^{-2} T \pmod{p}$. (If $q=3$, $P_\lambda(T,\chi_{1,1}) = 1$.)
\end{Corollary}
\begin{proof}
When $q \neq 3$, take $i = j = 0$ in Proposition~\ref{prop:charcomputation}. In the edge case that $q=3$, by inspection $C_\lambda$ has genus zero so $P_\lambda(T,\chi_{1,1}) = 1$.
\end{proof}
For $\lambda_1$ and $\lambda_2 \in {\mathbb F}_q \backslash \{0,1\}$ we have identified $G$ as the Galois group of $C_{\lambda_1} \to {\mathbb P}^1_{{\mathbb F}_q}$ and of $C_{\lambda_2} \to {\mathbb P}^1_{{\mathbb F}_q}$.
\begin{Corollary} \label{cor:identifyfibers}
If $L(T,C_{\lambda_1}/{\mathbb P}^1_{{\mathbb F}_q},\chi_{1,1}) \equiv L(T, C_{\lambda_2}/{\mathbb P}^1_{{\mathbb F}_q},\chi_{1,1}) \pmod{p}$ then $\lambda_2 = \lambda_1$ or $\lambda_2 = 1/\lambda_1$.
\end{Corollary}
\begin{proof}
Let $f(\lambda) = 2 \lambda (\lambda-1)^{-2}$. As $f(\lambda) = f(1/\lambda)$ and $f(\lambda) = c$ has at most two solutions for fixed $c \in {\mathbb F}_q$, this follows from Corollary~\ref{cor:congruence}.
\end{proof}
We can remove the ambiguity using a second $L$-function.
\begin{Corollary} \label{cor:identifyfibers2}
Suppose $p \geq 5$.
If $L(T,C_{\lambda_1}/{\mathbb P}^1_{{\mathbb F}_q},\chi_{1,1}) \equiv L(T, C_{\lambda_2}/{\mathbb P}^1_{{\mathbb F}_q},\chi_{1,1}) \pmod{p}$ and $L(T,C_{\lambda_1}/{\mathbb P}^1_{{\mathbb F}_q},\chi_{1,2}) \equiv L(T, C_{\lambda_2}/{\mathbb P}^1_{{\mathbb F}_q},\chi_{1,2}) \pmod{p}$, then $\lambda_1 = \lambda_2$.
\end{Corollary}
\begin{proof}
For $b =1, c=2$ we have $i=0$ and $j=1$, so $i+j=1 \leq q-4$ by hypothesis.
Using Proposition~\ref{prop:charcomputation} we see that
\begin{align*}
P_\lambda(T,\chi_{1,2}) &\equiv 1 - \alpha_{1,2} \left( \frac{\lambda}{1-\lambda} \right) \left( \frac{1}{\lambda-1} \right)^2 T \pmod{p} \\
& \equiv 1 - 3 \lambda (\lambda-1)^3 T \pmod{p}
\end{align*}
as $i = q-4$ and $j=0$ and the exponents only matter modulo $q-1$. The linear terms of the two pairs of $L$-functions being congruent modulo $p$ implies that
\begin{equation*}
\frac{\lambda_1}{(\lambda_1-1)^2} \equiv \frac{\lambda_2}{(\lambda_2-1)^2} \pmod{p} \quad \text{and} \quad
\frac{ \lambda_1}{(\lambda_1-1)^3} \equiv \frac{ \lambda_2}{(\lambda_2-1)^3} \pmod{p}.
\end{equation*}
Hence we conclude that $\lambda_1 = \lambda_2$.
\end{proof}
\begin{remark}
These corollaries use the non-canonical identification of $G$ with the Galois group of the cover $C_{\lambda_i} \to {\mathbb P}^1_{{\mathbb F}_q}$ to specify the characters. The identification is natural given the equation \eqref{eq:clambda} (and independent of $\lambda$), but this is not intrinsic to the curve. To distinguish $C_{\lambda_1}$ and $C_{\lambda_2}$ as covers of ${\mathbb P}^1$ without this non-canonical identification, we must intrinsically identify the character $\chi_{1,1}$. We will address this in the next subsection.
\end{remark}
\subsection{Distinguishing Subsets via \texorpdfstring{$L$}{L}-Functions} \label{ss:distinguishing}
Let $U$ be an open subset of ${\mathbb P}^1_{{\mathbb F}_q}$ which is the complement of a finite set of ${\mathbb F}_q$-points $P_1, \ldots, P_n$. If $n \leq 3$, then all such $U$ with $n$ points removed are isomorphic as $\PGL_2({\mathbb F}_q)$ acts three-transitively on ${\mathbb P}^1_{{\mathbb F}_q}$. If $n = 4$, we wish to determine $U$ up to isomorphism (equivalently, the $4$ points up to automorphism of ${\mathbb P}^1_{{\mathbb F}_q}$) using $L$-functions of characters of covers of $U$ (covers of ${\mathbb P}^1_{{\mathbb F}_q}$ unramified away from $P_1, \ldots, P_4$). We will do so using characters of the ray class field with modulus $[P_2] + [P_3] + [P_4]$ in which $P_1$ splits completely. The key technical obstacle is finding a way to specify the character to use.
\begin{Proposition} \label{prop:distinguishedchar}
Continuing the notation of Section \ref{ss:lfunctions}, if $q \neq 9$ and the character $\chi: \Gal(C_\lambda/{\mathbb P}^1_{{\mathbb F}_q}) = G \to \mu_{q-1}$ satisfies
\begin{enumerate}
\item \label{cond1} $\chi$ is surjective;
\item the genus of $C_\lambda / \ker \chi$ is $(q-3)/2$;
\item the cover $C_\lambda \to C_\lambda/ \ker \chi$ is ramified only at the $q-1$ points of $C_\lambda$ lying over $0$;
\item \label{cond4} $f_{C_\lambda,\chi^n} = 1$ if and only if $f_{C_\lambda,\chi_{n,n}} =1$ for all $n$ such that $0< 2n < q-1$;
\end{enumerate}
then $\chi$ is $\chi_{1,1}, \chi_{p,p}, \ldots,$ or $\chi_{p^{r-1},p^{r-1}}$. If $q=9$, $\chi$ could additionally be $\chi_{1,3}$ or $\chi_{3,1}$.
\end{Proposition}
We can reinterpret (\ref{cond4}) using Proposition~\ref{prop:charcomputation}.
For an integer $n$, the condition $f_{C_\lambda,\chi_{n,n}} =1$ is equivalent to the condition $C(n)$ that
\begin{equation} \label{eq:condn}
C(n) : \quad 0<2n < q-1 \quad \text{and} \quad (n \mod{p^i}) < p^i/2 \text{ for each } i \in \{1,\ldots,r-1\}.
\end{equation}
(Note that the first piece of $C(n)$ is almost equivalent to $(n \mod{p^r}) < p^r/2$: the only difference is that $n=(p^r-1)/2$ and $n=0$ are excluded.)
For integers $b$ and $c$, write $b' = (b \mod{q-1})$ and $c' = (c \mod{q-1})$ and let $C(b,c)$ be the condition that
\begin{equation} \label{eq:condbc}
C(b,c) : \quad b' + c' < q-1 \quad \text{ and } \quad \alpha_{b,c} = \frac{(q-1) (q-2) \cdots (q-b'-c')}{b'! \cdot c'!} \not \equiv 0 \pmod{p}.
\end{equation}
Proposition~\ref{prop:charcomputation} shows that $C(nb,nc)$ holds if and only if $f_{C_\lambda, \chi_{b,c}^n}=1$.
\begin{remark} \label{remark:kummer}
The conditions $C(n)$ and $C(b,c)$ may initially appear unconnected. Note that $\alpha_{b,c} \equiv \pm \binom{b'+c'}{b'} \mod{p}$. When $b' + c' < q = p^r$, a theorem of Kummer \cite{kummer} implies this is non-zero modulo $p$ precisely if $(b' \mod p^i) + (c' \mod p^i) < p^i$ for all $1 \leq i \leq r-1$. When $0 < 2n < q-1$, we therefore directly see that $C(n)$ is equivalent to $C(n,n)$.
\end{remark}
We will prove Proposition~\ref{prop:distinguishedchar} relying on the following result:
\begin{Proposition} \label{prop:charsdiffer}
Let $b$ and $c$ be integers that are relatively prime to $q-1$ satisfying $0<b,c < q-1$. For $0 < n < 2q$, suppose that $C(n)$ holds if and only if $C(nb,nc)$ holds. Then $b = c = p^i $ for some $i \in \{0,1,\ldots, r-1\}$, or $q=9$ and $(b,c) = (3,1)$ or $(1,3)$.
\end{Proposition}
We defer the proof of Proposition~\ref{prop:charsdiffer} until Section~\ref{ss:technicalproof}. The proof is elementary but complicated.
\begin{proof}[Proof of Proposition~\ref{prop:distinguishedchar}]
Suppose $\chi$ is a character of $G$ satisfying (\ref{cond1})-(\ref{cond4}).
As $\chi$ is surjective, the degree of the quotient map $C_\lambda \to C_\lambda /\ker \chi$ is $q-1$. As this map is Galois, the ramification index of all of the points of $C_\lambda$ above zero are the same; denote their common value by $e$. Then the Riemann Hurwitz formula implies that
\[
(q-2)(q-3) -2 = (q-1) (q-3 -2) + (q-1)(e-1).
\]
Thus $e=2$.
Write $\chi = \chi_{b,c}$ and $H_{b,c} = \ker \chi_{b,c}$. By considering the intersection of the inertia groups given in Lemma~\ref{lem:inertia} with $H_{b,c}$, we see that the ramification indices of $C_\lambda \to C_\lambda/H_{b,c}$ at the points of $C_\lambda$ above $0$ (resp. $1$, $\lambda$) are $\gcd(b+c,q-1)$ (resp. $\gcd(b,q-1)$, $\gcd(c,q-1)$). Given that $C_\lambda \to C_\lambda /\ker \chi$ is ramified only at the points over $0$ with ramification index $2$, we conclude that
\[
\gcd(b+c,q-1)=2, \quad \gcd(b,q-1)=1, \quad \text{and } \gcd(c,q-1)=1.
\]
Applying Proposition~\ref{prop:charsdiffer} and the reformulation of (\ref{cond4}) in terms of $C(n)$ and $C(n,n)$, we see that $\chi = \chi_{b,c} = \chi_{1,1}^{p^i}$ for $i \in \{0,1,\ldots,r-1\}$ unless we are in the exceptional case when $q=9$.
\end{proof}
Proposition~\ref{prop:distinguishedchar} lets us compare $L$-functions without having to identify the Galois groups.
\begin{Theorem} \label{thm:p1distinguished}
Suppose that $q = p^r$ is odd and not equal to $9$, and let $U$ and $U'$ be ${\mathbb P}^1_{{\mathbb F}_q}$ with four removed rational points. Fix orderings of the removed points ($P_1,\ldots,P_4$ and $P'_1,\ldots,P_4'$). Let $C$ be the ray class field for ${\mathbb P}^1_{{\mathbb F}_q}$ with modulus ${\mathfrak m} = [P_2] + [P_3] +[P_4]$ such that $P_1$ splits completely. Fix a character $\chi : \Gal(C / {\mathbb P}^1_{{\mathbb F}_q}) \to \mu_{q-1}$ such that
\begin{enumerate}
\item $\chi$ is surjective;
\item the genus of the quotient $C/\ker \chi$ is $(q-3)/2$;
\item the cover $C \to C/\ker \chi$ is ramified only at the $q-1$ points of $C$ lying over $P_2$;
\item $f_{C,\chi^n}=1$ if and only if $C(n)$ holds for all integers $0 < 2n < q-1$.
\end{enumerate}
Such a character $\chi$ exists, and is ``unique up to Frobenius'': the characters $\chi,\chi^p, \ldots \chi^{p^{r-1}}$ are the only characters of $\Gal(C/{\mathbb P}^1_{{\mathbb F}_q})$ with these properties.
Let $C'$ and $\chi'$ be defined analogously for $U'$. If the derivatives $L'(0,{\mathbb P}^1_{{\mathbb F}_q},\chi)$ and $L'(0,{\mathbb P}^1_{{\mathbb F}_q},\chi')$
lie in the same Frobenius orbit of ${\mathbb P}^1_{{\mathbb F}_q}$, in other words
there exists an integer $0\leq i <r$ such that
\[
(L'(0, {\mathbb P}^1_{{\mathbb F}_q},\chi) \mod{p}) = (L'(0,{\mathbb P}^1_{{\mathbb F}_q},\chi') \mod{p}) ^{p^i},
\]
then there is an automorphism $\alpha$ of ${\mathbb P}^1_{{\mathbb F}_q}$ such that $\alpha(U) = U'$, $\alpha(P_1)=P'_1$, and $\alpha(P_2)=P'_2$.
\end{Theorem}
\begin{Remark}
Informally, this says there exists a character $\chi : \Gal(C/ {\mathbb P}^1_{{\mathbb F}_q}) \to \mu_{q-1}$ such that $L(T,C/{\mathbb P}^1_{{\mathbb F}_q},\chi)$ determines $U$ up to isomorphism. The choice of character is not intrinsic to $U$ as it depends on distinguishing $P_1$ and $P_2$, and is furthermore only unique up to Frobenius. But choosing $P_1$ and $P_2$ is sufficient to determine $U$ up to isomorphism as the complement of four points in ${\mathbb P}^1_{{\mathbb F}_q}$ with two of them marked. To remove the restriction on marking two points, one can consider the analogous question after permuting the ordering.
\end{Remark}
\begin{proof}
Let $\sigma$ be the fractional linear transformation sending $P_1$ to $\infty$, $P_2$ to $0$, and $P_3$ to $1$. Let $\lambda = \sigma(P_4)$. Then $\sigma$ induces an isomorphism between $C$ and $C_\lambda$ as covers of ${\mathbb P}^1$ as both are ray class fields, and in particular identifies their Galois groups.
Proposition~\ref{prop:distinguishedchar} gives the existence of a character $\chi$ with the required properties and shows it is unique up to $p$th powers as claimed. In particular, there is a $0 \leq j < r$ such that
\[
L(T,{\mathbb P}^1_{{\mathbb F}_q},\chi) = L(T,C_\lambda/{\mathbb P}^1_{{\mathbb F}_q},\chi_{p^j,p^j}).
\]
By Proposition~\ref{prop:charcomputation}, we conclude that
\[
(L'(0,{\mathbb P}^1_{{\mathbb F}_q},\chi) \mod{p}) = 2 \left( \frac{\lambda}{(\lambda-1)^2} \right)^{p^j} .
\]
We can do the same for $U'$, obtaining analogous $\sigma'$, $\lambda'$, and $j'$.
By hypothesis $(L'(0, {\mathbb P}^1_{{\mathbb F}_q},\chi) \mod{p}) =( L'(0,{\mathbb P}^1_{{\mathbb F}_q},\chi') \mod{p})^{p^i}$, so there is an integer $0\leq i' < r$ such that
\[
\left( \frac{\lambda}{(\lambda-1)^2} \right)^{p^{i'}} = \frac{\lambda'}{(\lambda'-1)^2} .
\]
As the function $\lambda \mapsto \lambda (\lambda-1)^{-2}$ is two-to-one and the $p$th power map is an automorphism of ${\mathbb F}_q$, we conclude that $\lambda' = \lambda^{p^{i'}}$ or $1/\lambda' = \lambda^{p^{i'}}$.
Thus there is an automorphism $\beta$ of ${\mathbb P}^1_{{\mathbb F}_q}$ fixing $0$ and $\infty$ and sending $\{1,\lambda\}$ to $\{1,\lambda'\}$; $\beta$ is either the $p^{i'}$-th power map or the $p^{i'}$-th power map composed with the automorphism $x \mapsto x/\lambda$.
Then take $\alpha \colonequals (\sigma')^{-1} \circ \beta \circ \sigma$.
\end{proof}
\begin{remark}
The automorphism $x \mapsto x/\lambda$ sends $\{1,\lambda\}$ to $\{1,1/\lambda\}$, and corresponds to switching the role of the third and fourth points.
\end{remark}
\begin{remark} \label{remark:q9}
In the special case that $q=9$, we can distinguish $U = {\mathbb P}^1_{{\mathbb F}_q} - \{0,1,\infty,\lambda\}$ using $L$-functions as follows. Permuting the four marked points replaces $\lambda$ by $\lambda$, $1/\lambda$, $1-\lambda$, $1/(1-\lambda)$, $(\lambda-1)/\lambda$, and $\lambda/(\lambda-1)$. Thus there are two choices of $U$ up to isomorphism, which occur when $\lambda=-1$ or when $\lambda \in {\mathbb F}_9 - {\mathbb F}_3$. In light of Proposition~\ref{prop:charcomputation}, when $\lambda =-1$ we see that the value $(L'(0,C_\lambda/{\mathbb P}^1_{{\mathbb F}_q},\chi) \mod{p})$ lies in ${\mathbb F}_3$ for any character $\chi$. Otherwise there is some character $\chi$ for which $(L'(0,C_\lambda/{\mathbb P}^1_{{\mathbb F}_q},\chi) \mod{p})$ does not lie in ${\mathbb F}_3$. Therefore we can distinguish these configurations of points using $L$-functions.
\end{remark}
We finally prove a strong version of Conjecture~\ref{conj:lfun} for the projective line with four marked points. As that conjecture deals with pointed curves, the modulus should be supported on three points.
\begin{Theorem} \label{thm:p1zilber}
Suppose that $q = p^r$ is odd and not equal to $9$, and let $U$ (resp. $U'$) be
${\mathbb P}^1_{{\mathbb F}_q}$ with four rational points removed. Fix one of the four points $M$ (resp $M'$) to define the Abel-Jacobi embedding, and let ${\mathfrak m}$ (resp. ${\mathfrak m}'$) be the sum of the other three. Then let $X$ (resp. $X'$) be the ray class field of ${\mathbb P}^1_{{\mathbb F}_q}$ in which $M$ (resp. $M'$) splits completely, and
let $J_{{\mathfrak m}}$ (resp. $J_{{\mathfrak m}'}'$) be the generalized
Jacobian.
Suppose $\psi : J_{{\mathfrak m}}({{\mathbb F}}_q) \to J_{{\mathfrak m}'}'({{\mathbb F}}_q)$ is
an isomorphism of groups. If
$$L(T,X'/{\mathbb P}^1, \chi') = L(T,X/{\mathbb P}^1, \chi' \circ \psi)$$
for all characters $\chi'$ of $J_{{\mathfrak m}'}({\mathbb F}_{q})$, then
$U$ and $U'$ are Frobenius twists of each other and the map
$\psi$ arises from a morphism of curves which sends $M$ to $M'$ and sends ${\mathfrak m}$ to ${\mathfrak m}'$.
\end{Theorem}
\begin{proof}
Using the $\PGL_2({\mathbb F}_q)$-action on ${\mathbb P}^1_{{\mathbb F}_q}$, we may and do assume that $M = M' = \infty$, ${\mathfrak m} = [0]+[1]+[\lambda]$, and ${\mathfrak m}' = [0]+[1]+[\lambda']$ with $\lambda, \lambda' \ne 0, 1$. Note that $\lambda$ and $\lambda'$ are not canonical as they depend on the ordering of the removed points.
Consider a character $\chi'$ of $\Gal(X'/{\mathbb P}^1)$ satisfying conditions (1)-(4) of Theorem~\ref{thm:p1distinguished} (with $P_2 =0 $ say) and let $\chi = \chi' \circ \psi$. Let $Y = X /\ker \chi$ and $Y' = X'/ \ker \chi'$. We know that $Y' \to {\mathbb P}^1$ is a cyclic cover of degree $q-1$ in which $\infty$ splits and that is ramified only above $0, 1,$ and $\lambda'$, and that the genus of $Y'$ is $(q-3)/2$.
Since the zeta functions of $Y$ and $Y'$ are equal by hypothesis, $Y$ also has genus $(q-3)/2$. We claim that $\chi$ also satisfies conditions (1)-(4) of Theorem~\ref{thm:p1distinguished} (with $P_2 \in \{0,1,\lambda\}$), which amounts to showing that $Y \to {\mathbb P}^1_{{\mathbb F}_q}$ is totally ramified at two of $\{0,1,\lambda\}$ and has ramification index $(q-1)/2$ at the remaining point.
To see this, let the ramification indices at the points of $Y$ above $0,1$, and $\lambda$ be $(q-1)/m_i$ (they divide $(q-1)$ since the cover is Galois of degree $q-1$). The Riemann-Hurwitz formula implies that $q-5 = (q-1)(-2) + \sum_{i=1}^3 m_i((q-1)/m_i-1)$
which simplifies to $\sum m_i = 4$. Thus two are $1$ and one is $2$.
Then the theorem follows from Proposition~\ref{prop:charcomputation}
and Proposition~\ref{prop:distinguishedchar} as in the proof of Theorem~\ref{thm:p1distinguished}.
\end{proof}
\subsection{Proof of Proposition~\ref{prop:charsdiffer}} \label{ss:technicalproof}
We continue to work over a finite field of size $q = p^r$, writing $k \colonequals (q-1)/2$ and assuming $p >2$. We thank Z. Brady for help with the proofs of Lemma~\ref{lem:witness} and Proposition~\ref{prop:charsdiffer}; he gave us elegant proofs of the version of these results when $r=1$ (i.e. $q=p$). Our extensions to the case when $r>1$ have become significantly less elegant, for which we apologize.
We begin with a couple of simple observations about the conditions $C(n)$ and $C(b,c)$ introduced in \eqref{eq:condn} and \eqref{eq:condbc}.
\begin{Lemma} \label{lem:condn}
We have:
\begin{enumerate}
\item \label{condn1} condition $C(n)$ holds if and only if $C(k-n)$ holds;
\item \label{condn3} for $0< n < q-1$, the condition $C(n)$ is equivalent to condition $C(n,n)$.
\end{enumerate}
\end{Lemma}
\begin{proof}
The first is elementary. The second was discussed in Remark~\ref{remark:kummer} and is a consequence of a classical theorem of Kummer.
\end{proof}
We next need some elementary yet tricky lemmas.
\begin{Lemma} \label{lem:witness}
For any integer $b$ not congruent to $0,1,p,p^2,\ldots, p^{r-1}$ modulo $q-1$, there exists an integer $n$ satisfying $C(n)$ such that $(n b \mod{q-1}) \geq (q-1)/2$.
\end{Lemma}
\begin{proof}
Without loss of generality, we may assume that $1< b < (q-1)/2$. If $b \geq k=(q-1)/2$ and $q \neq 3$, we take $n=1$ and notice that $C(1)$ holds if $q \neq 3$. (If $q=3$, there do not exist any $b$ as in the statement of the lemma.) We proceed by induction on $r$, but treating some special cases by hand. In this proof only, we will use the notation $C_i(n)$ to denote the condition in \eqref{eq:condn} with $q$ replaced by $p^i$.
For the base case $r=1$, let $n = \lfloor (k-1)/b \rfloor +1$, which satisfies $n < (p-1)/2$ as $b >1$. We directly see that $nb > (k-1) = (q-1)/2-1$ and that $n b \leq (k-1) + b < q-1$ as $b < (q-1)/2$.
Next, suppose that $b \not \equiv 0,1,p,p^2,\ldots, p^{r-2} \pmod{p^{r-1}-1}$. By induction, we may suppose that there exists an integer $n_0$ satisfying $C_{r-1}(n_0)$ such that there are integers $a,c$ for which
\[
n_0 b = a (p^{r-1}-1) + c \quad \text{with} \quad \frac{p^{r-1}-1}{2} \leq c < p^{r-1}-1.
\]
As $C_{r-1}(n_0)$ holds, we know that $n_0 < (p^{r-1}-1)/2$ and hence
\begin{equation*}
a = \frac{n_0 b -c}{p^{r-1}-1} < \frac{b-1}{2}.
\end{equation*}
Then $a(p-1) < b (p-1)/2 - (p-1)/2$ and $(p^r-1)/2 - (p-1)/2 \leq pc$ and hence
\begin{equation} \label{eq:inequality1}
-b \cdot \frac{p-1}{2} + \frac{p^r-1}{2} < -(p-1)a + pc .
\end{equation}
Let $\epsilon$ be the smallest positive integer such that
\[
\frac{p^r-1}{2} \leq -(p-1)a + pc + \epsilon b.
\]
Then \eqref{eq:inequality1} shows that $\epsilon \leq (p-1)/2$.
Therefore $n \colonequals p n_0 + \epsilon$ satisfies $C_r(n)$. We compute
\[
nb = (p^r-1) a - (p-1)a + pc + \epsilon b.
\]
By the choice of $\epsilon$ and as $b < (p^{r}-1)/2$, we conclude that
\[
\frac{p^r-1}{2} \leq -(p-1)a + pc + \epsilon b \leq \frac{p^r-1}{2} +b < p^r-1.
\]
Thus $(nb \pmod{p^{r}-1} ) \geq (p^r-1)/2$ as desired.
Finally, we directly treat the cases that $b \equiv 0,1,p,p^2,\ldots, p^{r-2} \pmod{p^{r-1}-1}$ which the above inductive step does not address.
We first consider the case when $b = a (p^{r-1}-1)$ with $a \in \{1,2,\ldots,p\}$. If $a=1$, take $n=p$. Otherwise take $n = \lfloor (p-1)/(2a) \rfloor +1$. For both, we see that $n$ satisfies $C_r(n)$ and $(nb \mod{p^r-1}) \geq (p^r-1)/2$.
In the other cases, write $b = p^i + a (p^{r-1}-1)$ with $i \in \{0,1,\ldots,r-2\}$ and $0<a\leq p$.
\begin{description}
\item[The case $i=0$ and $a>1$] First suppose $b = 1 + a (p^{r-1}-1)$ with $a>1$. If $a<p/2$, take $n= p$, and observe that $nb \equiv p + a(1-p) \pmod{p^r-1}$ so $(nb \mod{p^r-1}) \geq (p^r-1)/2$. If $a > p/2$, take $n=1$ and observe that $b \geq (p^r-1)/2$. In both cases $C_r(n)$ holds.
\item[The case $a=1$ and $i>0$] Now suppose $b = p^i + (p^{r-1}-1)$ with $i>0$. Taking $n=c p^{r-i-1} + c $, we compute that
\[
nb \equiv 2c p^{r-1} + c p^i - c - c p^{r-i-2}(p-1) \mod{p^r-1}.
\]
If we choose $c = (p-1)/2$, then we see that
\[
2c p^{r-1} + c p^i - c - c p^{r-i-2}(p-1) \leq 2 c p^{r-1} + c p^i \leq (p-1)p^{r-1} + \frac{p-1}{2} p^i \leq p^r -1.
\]
We also compute that
\[
2c p^{r-1} + c p^i - c - c p^{r-i-2}(p-1) \geq (p-1) p^{r-1} - \frac{p-1}{2} - (p-1)^2 p^{r-i-2}/2 \geq (p^r-1)/2.
\]
Thus $(nb \mod{p^r-1}) \geq (p^r-1)/2$ and we directly see that $C_r(n)$ holds.
\item[The case $i>0$ and $a>1$] Now suppose $ b=p^i + a(p^{r-1}-1)$ with $a>1$ and $i >0$. If $a > p/2$, take $n=1$. Otherwise, choose $a' = \lceil (p-1)/(2a) \rceil$ and let $n = p^{r-i-1} + a'$. Notice that $C_r(n)$ holds as $a>1$, and that $(p+1)/2 \leq aa'+1 \leq p-1$. We compute that
\[
nb \equiv p^{r-1} + a a' (p^{r-1}-1) + a p^{r-i-2}(1-p) + a' p^i \mod{p^r-1}.
\]
As $r-i-2< r-2$ and $i < r-1$, we see that the dominant term in the above expression is $(1 + a a') p^{r-1}$. In particular,
\[
p^{r-1} + a a' (p^{r-1} -1)+ a p^{r-i-2}(1-p) + a' p^i \leq (a a' +1 ) p^{r-1}+ a' p^i < p^{r}-1
\]
and
\[
p^{r-1} + a a' (p^{r-1}-1) + a p^{r-i-2}(1-p) + a' p^i \geq (a a' +1 )p^{r-1} + a p^{r-i-2} (1-p) \geq (p^r-1)/2
\]
as $a < p/2$.
Thus $(nb \mod{p^r-1}) \geq (p^r-1)/2$ as desired.
\end{description}
Our argument has covered all of the cases, as we have assumed $b$ is not congruent to $0,1,p,p^2,\ldots, p^{r-1}$ modulo $p^r-1$.
\end{proof}
\begin{Lemma} \label{lem:equalpowers}
Suppose $b = p^{m_1}$ and $c = p^{m_2}$ with $ 0 \leq m_1 < m_2 < r$. For odd $q \neq 9$, there exists an integer $n$ for which $C(n)$ does \emph{not} hold and for which $C(nb,nc)$ holds.
\end{Lemma}
\begin{proof}
Take $n = (p+1)/2$. Notice that $ p/2 < n < p$, so $C(n)$ does not hold. On the other hand,
\begin{align*}
( n b \mod{2k}) + (n c \mod{2k}) & = \frac{p+1}{2} ( p^{m_1} + p^{m_2}) \\
& \leq \frac{p+1}{2} ( p^{r-1} + p^{r-2}) \\
& \leq p^r \left( \frac{1}{2} + \frac{1}{p} + \frac{1}{2p^2} \right) < q-1.
\end{align*}
The last inequality uses that $p>3$ or that $p=3$ and $r>2$. Finally, notice that
\[ \alpha_{nb,nc} \equiv \pm \binom{ (p+1)/2 \cdot p^{m_1} + (p+1)/2 \cdot p^{m_2} }{(p+1)/2 \cdot p^{m_1}} \pmod{p}
\]
and that this binomial coefficient is non-zero modulo $p$ by Kummer's theorem on binomial coefficients \cite{kummer}.
\end{proof}
We are now ready to prove Proposition~\ref{prop:charsdiffer}; we again thank Z. Brady for the key idea.
\begin{proof}[Proof of Proposition~\ref{prop:charsdiffer}]
As $b$ is odd, for any integer $n$
\begin{equation} \label{eq:congruence}
A(b,n) := (nb \mod{2k}) + ((k-n) b \mod{2k}) \equiv k \pmod{2k}.
\end{equation}
In particular, $A(b,n) > 0$.
If $C(n)$ holds, then $C(k-n)$ holds by Lemma~\ref{lem:condn}(\ref{condn1}). By hypothesis
$C(nb,nc)$ and $C((k-n) b, (k-n)c)$ hold as well, so
\begin{align*}
0<A(b,n) + A(c,n) &= (nb \mod{2k}) + ((k-n) b \mod{2k}) \\
&+ (n c \mod{2k}) + ((k-n) c \mod{2k}) < 4k.
\end{align*}
As $A(b,n) + A(c,n) \equiv k + k \equiv 0 \pmod{2k}$
and $0 < A(b,n) + A(c,n)$, we conclude that $A(b,n) + A(c,n) = 2k$ if $C(n)$ holds.
We next claim that if $C(n)$ holds then $(n b \mod{2k}) < k$. If $(n b \mod{2k}) > k$, then \eqref{eq:congruence} shows that $A(b,n) > 2k$. This is impossible as we know that $A(b,n) + A(c,n) = 2k$ and each term is non-negative. Since $b$ is relatively prime to $2k$ and $n <k$ as $C(n)$ holds, the claim follows.
If $b$ were not a power of $p$ or $1$, Lemma~\ref{lem:witness} would give an integer $n$ for which $C(n)$ holds and such that $(nb \mod{2k}) \geq k$, contradicting the claim. Thus $b \in \{1,p,p^2,\ldots, p^{r-1}\}$. Likewise, we see that $c \in \{1,p,p^2,\ldots, p^{r-1}\}$. Then Lemma~\ref{lem:equalpowers} shows that either $b=c$ or we are in the exceptional case with $q=9$. This completes the proof.
\end{proof}
\section{General Affine Curves}
Given an equation $F(x,y) = 0$ defining a smooth projective curve $C/{\mathbb F}_q$, our goal is to recover the coefficients
of the equation from an explicit set of $L$-functions for covers of $C$. We will do so using Artin-Schreier covers. Let $K$ be the function field of $C$ and
$p$ the characteristic of $K$.
In what follows, we will deal with extensions ${\mathbb F}_{q^m}/{\mathbb F}_q$ and consider
both the relative trace $\Tr_m: {\mathbb F}_{q^m} \to {\mathbb F}_q$ and the absolute trace
down to the prime field $\Tr: {\mathbb F}_{q^m} \to {\mathbb F}_p$.
\subsection{Exponential Sums} \label{ss:exp} We begin by reviewing the connection between $L$-functions for Artin-Schreier extensions and exponential sums; a standard reference is \cite[VI]{Bom}.
Artin-Schreier extensions are cyclic degree $p$ extensions of $K$, and are given by the equations $z^p -z =f$ where $f \in K$ is not equal to $w^p - w$ for any $w \in K$. For a fixed $f$, regular on an open subset $U$ of $C$, the conductor of $E/K$ (where $E=K(z)$, with $z^p-z=f$) is bounded above by $(f)_\infty$, the divisor of poles of $f$ on $C$.
There is a character $\chi_f : \Gal(E/K) \to {\mathbb C}^\times$
such that $\chi_f(\operatorname{Frob}_P) = \exp(2\pi i \Tr(f(P))/p)$ for $P \in U$,
where $\Tr$ is the absolute trace $\Tr: {\mathbb F}_{q^m} \to {\mathbb F}_p$.
We assume that $E/K$ is ramified outside $U$ and consider the exponential sums
$$S_m(\chi_f) = S_m(f) \colonequals \sum_{P \in U({\mathbb F}_{q^m})} \exp(2\pi i \Tr(f(P))/p).$$
The values of these sums are determined by $L(T,C,\chi_f)$.
When $p \neq 2$, taking $\varpi = 1 - \exp(2\pi i /p)$ we note that
${\mathbb Z}[\exp(2\pi i /p)]/(\varpi)$ is isomorphic to ${\mathbb F}_p$ and that
\begin{equation}
S_m(f) = \sum_{P \in U({\mathbb F}_{q^m})}(1-\varpi)^{\Tr(f(P))} \equiv
|U({\mathbb F}_{q^m})| - \left( \sum_{P \in U({\mathbb F}_{q^m})}{\Tr(f(P))} \right)\varpi \pmod{\varpi^2}.
\end{equation}
If $p=2$, we instead consider the exponential sums
$$S_m({\chi_{{\mathbf f}}}) = \sum_{P \in U({\mathbb F}_{q^m})} e^{2\pi i T({{\mathbf f}}(P))/4}$$
where ${\mathbf f} = (f,0)$ is a Witt vector of length $2$ and
$T:W_2({\mathbb F}_{q^m}) \to W_2({\mathbb F}_2) \simeq {\mathbb Z}/4{\mathbb Z}$ is the absolute trace (see \cite{VW}).
We can form an $L$-function as before, which will be a factor of the
zeta function of an Artin-Schreier-Witt cover of $C$. Letting $\varpi = 1 - i$, we
get $S_m({{\mathbf f}}) \equiv |U({\mathbb F}_{q^m})| - (\sum_{P \in U({\mathbb F}_{q^m})}{\Tr(f(P))})\varpi \pmod{\varpi^2}$.
\begin{Proposition} \label{prop:determinesums}
Let $\gamma_1, \ldots, \gamma_u$ be a basis for ${\mathbb F}_q$ over ${\mathbb F}_p$.
The $L$-function $L(T,C,\chi_{\gamma_i f})$ determines the value of
\begin{equation}
T_m(f):=\sum_{P \in U({\mathbb F}_{q^m})}{\Tr_m(f(P))} \in {\mathbb F}_q
\end{equation}
for each $m$, where $\Tr_m: {\mathbb F}_{q^m} \to {\mathbb F}_q$ is the relative trace.
\end{Proposition}
\begin{proof}
We have seen that the knowledge of the L-functions $L(T,C,\chi_{\gamma_i f})$ provides us with
\[\sum_{P \in U({\mathbb F}_{q^m})}{\Tr(\gamma_i f(P))} \in {\mathbb F}_p\]
for each $m$. But these determine $T_m(f)$ as the trace pairing is non-degenerate.
\end{proof}
\subsection{Recovering Affine Equations from Exponential Sums}
We now show how to use $L$-function for appropriate Artin-Schreier covers of a curve to recover the coefficients of a defining equation via considering exponential sums.
\begin{Theorem}
\label{thm:equations}
For fixed odd $q$ and $d \ge 1$, there is an explicit finite
set of polynomials $f \in {\mathbb F}_q[x,y]$ such that we can
recover the coefficients of any absolutely irreducible $F \in {\mathbb F}_q[x,y]$ of degree $d$
from the $L$-functions of the Artin-Schreier extensions $z^p-z=f$ of the function field $K$ for the curve determined by $F=0$.
\end{Theorem}
\begin{remark}
When $q$ is even we would need the Artin-Schreier-Witt extensions discussed in Section~\ref{ss:exp}. Our proof does not apply as written, but we expect it to adapt.
\end{remark}
The proof will use the following lemma about the values of symmetric functions and power sums. For a positive integer $n$, let $e_1,\ldots e_n$ be the elementary symmetric functions in $x_1,\ldots, x_n$ viewed as elements of ${\mathbb Z}[x_1,\ldots,x_n]$, and define
\[
p_k = \sum_{i=1}^n x_i^k \in {\mathbb Z}[x_1,\ldots,x_n].
\]
\begin{Lemma} \label{lem:powers}
Fix a field $\mathbf{k}$ of characteristic $p$ and an integer $n>0$.
Given $\alpha_1,\alpha_2, \ldots \alpha_r \in \mathbf{k}$ and multiplicities $n_1, n_2, \ldots , n_r$ with $1 \leq n_i <p$ and $n=\sum_{i=1}^r n_i$, let $\beta_1, \ldots \beta_n$ be these $n$ values, with $\alpha_i$ occurring $n_i$ times. The sequence of values of the power sums
\[
p_j(\beta_1,\ldots, \beta_n) = \sum_{i=1}^r n_i \alpha_i^j
\]
for all $j$ uniquely determine the values of the symmetric functions $e_j(\beta_1,\ldots, \beta_n)$ for $1 \leq j \leq n$. If the field $\mathbf{k}$ is a finite field with $q$ elements, it suffices to know the values of the power sums for $j=1 ,\ldots, q-1$.
\end{Lemma}
We defer the proof to the next subsection.
\begin{Lemma} \label{lem:h}
There exists an integer $m > d$ with $p \nmid m$ (depending on $q$ and $d$ only) and an irreducible polynomial $h \in {\mathbb F}_q[x]$ of degree $m$ such that all solutions to $F(x,y)=h(x)=0$ are defined over ${\mathbb F}_{q^m}$ and that there are $m \deg_y F$ of them. Whether a particular $h$ works can be verified using exponential sums on $C$.
\end{Lemma}
\begin{proof}
Using Chebotarev, for sufficiently large $m$ there exists an irreducible polynomial $h \in {\mathbb F}_q[x]$ of degree $m$
such that all solutions of
$F(x,y)=h(x)=0$ are defined over ${\mathbb F}_{q^m}$ and there are
$m\deg_y F$ of them.
To check this condition for a given irreducible $h \in {\mathbb F}_q[x]$, run through all $s \in {\mathbb F}_q[x]$ with $\deg s < m$ and
take $f=\left(1-h(x)^{q^m-1}\right) \left(1-(y-s(x))^{q^m-1} \right)$. Notice that for $a,b \in {\mathbb F}_{q^m}$ we have $f(a,b)\neq 0$ if and only if $h(a) = 0$ and $b=s(a)$. When $f(a,b)$ is non-zero, it takes on the value $1$. If there exists $a \in {\mathbb F}_{q^m}$ such that $h(a) = 0$ and $F(a,s(a))=0$, it follows that $F(a^{q^j},s(a^{q^j}))=0$ for $j=1,\ldots,m-1$, and hence $F(a,s(a))=0$ for all roots $a$ of $h$ in ${\mathbb F}_{q^m}$. In this case we conclude that $T_m(f) = m^2$ and we have $m$ solutions to $h(x) = F(x,y)=0$ over ${\mathbb F}_{q^m}$ given by $(a^{q^j}, s(a^{q^j}))$ for $j=0 \ldots m-1$. If not, then $T_m(f)=0$.
By varying $s$ so that it takes on different values at the roots of $h$, we can detect whether there are $m \deg_y F$ solutions to $F(x,y)=h(x)=0$ and that are defined over ${\mathbb F}_{q^m}$.
\end{proof}
\begin{Lemma} \label{lem:f}
Choose $h$ as in Lemma~\ref{lem:h}. For any fixed $i,j$, there exists $f \in {\mathbb F}_q[x,y]$ such that for $(a,b) \in {\mathbb F}_{q^m}^2$:
\begin{enumerate}
\item \label{cond:ram} $z^p -z = f$ is irreducible and the Artin-Schreier extension is ramified over every infinite point of $C$;
\item if $h(a) \neq 0$ then $f(a,b) = 0$;
\item if $h(a)=0$ then $f(a,b) = a^i b^j$.
\end{enumerate}
The value of $T_m(f)$ can be computed without knowing an explicit choice of $f$.
\end{Lemma}
\begin{proof}
Consider possible $f$ of the form $f = x^iy^j(1-h^{q^m-1}) (1+(x^{q^m}-x)g)$, for an arbitrary $g \in {\mathbb F}_q[x,y]$. The second and third statements are straightforward to check, and it is clear that the value $f(a,b)$ for $a,b \in {\mathbb F}_{q^m}$ is independent of the choice of $g$. Thus we must show there is a choice of $g$ such that $z^p-z=f$ is irreducible and the Artin-Schreier extension is ramified over every infinite point of $C$.
Take a large integer $D$ and consider all functions on $K$ represented by a
$g \in {\mathbb F}_q[x,y]$ with $\deg g \le D$. There are at least $q^{D\cdot d }$ such functions (where $d = \deg F$ as before). For any such $g$, the resulting $f$ is a polynomial of degree at most $D+O(1)$ and is determined, up to a constant, by the polar parts at the places at infinity. If condition (\ref{cond:ram}) is not satisfied, we must have that the polar part at one of the places at infinity is of the exceptional form $w^p-w$ for some Laurent polynomial $w$ at that place.
Let $v$ denote the valuation at that place and
$m \colonequals \min\{v(x),v(y)\}$; note that $0 > m \ge -d/\deg v$.
Then $v(f) \ge m D$ and if the polar part of $f$ at $v$ is $w^p -w$
then $v(w) \ge mD/p$ and the coefficients of the powers $\pi^j$ of a local
parameter $\pi$ at $v$ occurring in the expansion of $f$ for $j < mD/p$
and $p \nmid j$ must vanish. For $f$ to be of this exceptional form at the place $v$
results in at least $D(1-1/p)^2$ independent linear
conditions on $f$. As there are at most $d$
such places, we see that
there are at most $O(q^{D(1-(1-1/p)^2)})$ polynomials $f$ of degree $D$ for which condition (\ref{cond:ram}) does not hold.
Thus for $D$ sufficiently large (depending just on $d$ and $p$), we see that at least half of the choices of $g$ must give an $f$ satisfying condition (\ref{cond:ram}). We conclude that the value of $T_m(f)$ for a choice of $f$ satisfying the first condition is the most common value obtained by computing $T_m ( x^i y^j (1-h^{q^{m}-1}) (1 + (x^{q^m}-x)g))$ for
all choices of $g$ with $\deg g \le D$ and some suitably large $D$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:equations}]
By Proposition~\ref{prop:determinesums}, the values of the exponential sums on $C$ are determined by the knowledge of the $L$-function.
Choose $m$ and $h$ as in Lemma~\ref{lem:h} using our knowledge of the degree of $F$ and of the values of the exponential sums on $C$. Then choose $f$ as in Lemma~\ref{lem:f}. We compute that
\[
T_m(f) = \sum_{\stackrel{(a,b)\in ({\mathbb F}_{q^m})^2}{F(a,b)=0}} \Tr_m(f(a,b)) = \sum_{a:h(a)=0} \sum_{b:F(a,b)=0} \Tr_m(a^i b^j)
\]
Since Frobenius cyclically permutes the roots of $h(x)$, we see that
\[
T_m(f) = m\sum_{a:h(a)=0} \sum_{b:F(a,b)=0} a^i b^j.
\]
As $p \nmid m$, we can use these values with $i=0,1,\ldots,m-1$ to recover, for each root $a$ of $h$, the value of $\sum_{b:F(a,b)=0} b^j$ as the Vandermonde matrix is invertible. By Lemma~\ref{lem:powers}, these values determine the polynomial $F(a,y) \in {\mathbb F}_q[y]$ for each root $a$. Viewing $F(x,y)$ as a polynomial in $x$ with coefficients in ${\mathbb F}_q[y]$, the polynomials $F(a,y)$ let us determine $F(x,y) \in {\mathbb F}_q[x,y]$ as $m>d$. This completes the proof.
\end{proof}
\begin{Remark}
The proof uses a ridiculously large number of $L$-functions of Artin-Schreier covers, something like $u(q^{m^2}+q^D)$, where $u=[{\mathbb F}_q:{\mathbb F}_p]$, $m$ comes from Lemma \ref{lem:h} and can be explicitly computed applying the effective Chebotarev Theorem from \cite{MS} to the splitting field of $F(x,y)$ over ${\mathbb F}_q(x)$ and $D$ comes from Lemma \ref{lem:f} and can be extracted from the proof there. It would be nice to have a smaller such set. In specific situations, it is often possible to be more efficient.
\end{Remark}
\begin{Example} \label{ex:legendre}
Let us use $L$-functions to distinguish members of the Legendre family of elliptic curves $y^2 = x ( x-1)(x-\lambda)$. We take $g(x,y) = x (1-y^{q-1})$. A direct calculation as in the proof shows that
\[
T_1(g) = \sum_{b : b(b-1)(b-\lambda) =0} b = 1 + \lambda.
\]
Thus the value of this single exponential sum determines $\lambda$.
We now explain how this is a simplification of the proof of Theorem~\ref{thm:equations} for this particular family of curves.
First, note that it suffices to recover the coefficient of $y$ in the polynomial $F(x,y) = x^2 - y(y-1)(y-\lambda)$, so we will focus just on recovering the linear term of $F(0,y)$ instead of the whole polynomial. (Note that we swap the usual role of $x$ and $y$ to conform to the notation in our proof.)
This polynomial has degree $d=3$. The assumption in Lemma~\ref{lem:h} that $m>d$ is irrelevant as we are only interested in the linear term, so let us take $m=1$. Then $h(x) = x$ satisfies the other conditions of Lemma~\ref{lem:h} as $0 = x (x-1)(x-\lambda)$ has three solutions over ${\mathbb F}_q$. Choosing $(i,j) = (0,1)$, we claim that $f = y (1-x^{q-1})$ satisfies the conditions of Lemma~\ref{lem:f}. As $f$ has a pole of order $3q-1$ at infinity, $z^p -z = f$ is irreducible and ramified over the unique point at infinity. If $h(a)=a \neq 0$, then $f(a,b) = b (1 -a^{q-1})=0$ and otherwise $f(a,b) = b$ as desired. (Having knowledge of the behavior of $x$ and $y$ at infinity makes finding a suitable $f$ much easier.) This recovers the exponential sum we used above,
\[
T_1(f) = \sum_{b : b(b-1)(b-\lambda) =0} b = 1 + \lambda.
\]
The proof of Theorem~\ref{thm:equations}, which recovers the whole equation, is more complicated as it needs a larger $m$; it recovers the coefficient of $y$ in $F(0,y) = -y(y-1)(y-\lambda)$ using Artin-Schreier covers for various $(i,j)$. In particular, it recovers the coefficient from recovering the sums of powers of the roots $0$, $1$, and $\lambda$. These power sums in turn determine $F(0,y)$.
\end{Example}
\subsection{Power Sums and Symmetric Functions}
It remains to prove Lemma \ref{lem:powers}. Continuing the notation of the lemma, we have the following elementary recurrences:
\begin{align}
k e_k & = \sum_{i=1}^k (-1)^{i-1} e_{k-i} p_i \quad \text{ when } 1 \leq k \leq n \label{eq:recurrence1}\\
0 & = \sum_{i=k-n}^k (-1)^{i-1} e_{k-i} p_i \quad \text{ when } k>n \label{eq:recurrence2}.
\end{align}
These can be deduced by considering the generating function
\[
F(T) = \prod_{i=1}^n (1- x_i T) = \sum_{i=0}^n (-1)^i e_i T^i.
\]
The logarithmic derivative is
\[
\frac{F'(T)}{F(T)} = \sum_{i=1}^n \frac{-x_i}{1 - x_i T} = \sum_{j=0}^\infty - p_{j+1} T^j.
\]
The recurrences follow by cross-multiplying.
For a field $K$ of characteristic zero, $K[p_1,\ldots,p_n ] = K[e_1,\ldots,e_n] = K[x_1,\ldots, x_n]^{S_n}$. This is false in positive characteristic (and over ${\mathbb Z}$).
\begin{example} \label{ex:power2}
Let $p=2$ and $n=2$. Then $p_1 = e_1$, and the recurrences tell us that $2 e_2 = e_1 p_1 + p_2 = 0$ and that
\[
p_3 = e_1 p_2 - e_2 p_1 \quad \text{i.e.} \quad x_1^3 + x_2^3 =(x_1 + x_2) \left( x_1^2 + x_2^2 - x_1 x_2. \right)
\]
Thus we see that
\[
e_2 = \frac{p_1 p_2 - p_3}{p_1} \quad \text{i.e.} \quad x_1 x_2 = \frac{ (x_1+x_2)(x_1^2 + x_2^2) - x_1^3 - x_2^3}{x_1 + x_2}.
\]
So we can recover the symmetric polynomials as rational functions in $p_1, p_2, p_3, \ldots$. (Actually, we just need $p_1$ and $p_3$, as $p_1^2 = p_2$.)
Evaluating at elements of $\mathbf{k}$, we can recover the values of $e_1$ and $e_2$ provided $p_1(x_1,x_2) \neq 0$. This restriction is necessary: consider the family $x_1 = \alpha$, $x_2 = - \alpha = \alpha$ for $\alpha \in \mathbf{k}$, where we have $p_i(x_1,x_2) =0$ for all $i$ but $e_2(x_1,x_2) = - \alpha^2$.
\end{example}
\begin{lem} \label{lem:field}
For any field $\mathbf{k}$, the fields $\mathbf{k}(p_1,\ldots,p_n, \ldots )$ and $\mathbf{k}(e_1,\ldots, e_n)$ are equal as subfields of $\mathbf{k}(x_1,\ldots,x_n)$.
\end{lem}
\begin{proof}
As the $p_i$ are symmetric, they may be written as polynomials in terms of the $e_i$'s. So it suffices to write the $e_i$ as rational functions in terms of the $p_i$. The key idea is to show that the power series
\[
\sum_{j=0}^\infty - p_{j+1} T^j
\]
is a rational function with coefficients in $k(p_1,p_2,\ldots)$. We already know it is a rational function with coefficients in $k(x_1,\ldots ,x_n)$ from the identity
\[
\frac{F'(T)}{F(T)} = \sum_{j=0}^\infty - p_{j+1} T^j.
\]
As $F(T)$ has no repeated roots, $F(T)$ and $F'(T)$ have no common roots and so if
\[
\frac{A(T)}{B(T)} = \frac{F'(T)}{F(T)}
\]
with $A(T), B(T) \in k(x_1,\ldots,x_n)[T]$ and $A(T)$ of degree at most $n-1$ (in T) and $B(T)$ of degree $n$ (in T), then $A(T) = g F'(T)$ and $B(T) = g F(T)$ for some non-zero $g \in k(x_1,\ldots ,x_n)$. If we require that the constant term of $B(T)$ is one, then $A(T) = F'(T)$ and $B(T) = F(T)$.
Now write $A(T) = \sum_{i=0}^{n-1} a_i T^i$ and $B(T) = \sum_{i=0}^n b_i T^i$, viewing the $a_i$ and $b_i$ as variables, and consider
\[
A(T) = B(T) \cdot \left( \sum_{j=0}^\infty - p_{j+1} T^j \right).
\]
Equating coefficients of $T$, we obtain an (infinite) system of linear equations with variables $\{a_i\}$ and $\{b_i\}$ and coefficients in $\mathbf{k}(p_1,p_2,\ldots)$. This is solvable over the larger field $\mathbf{k}(x_1,\ldots,x_n)$ as we have the identity of rational functions
\[
\frac{F'(T)}{F(T)} = - \sum_{j=0}^\infty - p_{j+1} T^j.
\]
Therefore, the system of linear equations is also solvable over $\mathbf{k}(p_1,p_2,\ldots )$.
In particular, there exists $A(T)$ of degree at most $n-1$ and $B(T)$ of degree $n$ in $\mathbf{k}(p_1,p_2,\ldots)[T]$ such that
\[
\frac{A(T)}{B(T)} = \sum_{j=0}^\infty - p_{j+1} T^j.
\]
We may certainly scale by an element of $\mathbf{k}(p_1,p_2,\ldots)$ so the constant term of $B(T)$ is $1$, in which case we must have that $A(T) = F'(T)$ and $B(T) = F(T)$. Therefore the coefficients of $F(T)$, which up to sign are the elementary symmetric functions, lie in $\mathbf{k}(p_1,p_2,\ldots)$. This completes the proof.
\end{proof}
\begin{remark}
It follows from Lemma~\ref{lem:field} that $\mathbf{k}(p_1,p_2,\ldots)$ is generated by finitely many of the $p_i$'s, although it is not clear which ones do so. As Example~\ref{ex:power2} shows, it is not possible to simply write $e_1,\ldots e_n$ in terms of $p_1,\ldots , p_n$.
\end{remark}
We can now prove Lemma~\ref{lem:powers}.
\begin{proof}[Proof of Lemma~\ref{lem:powers}]
Evaluate $F$ at $x_1 = \beta_1, x_2 = \beta_2, \ldots, x_n = \beta_n$, obtaining
\[
G(T) := \prod_{i=1}^n (1 - \beta_i T ) = \sum_{i=0}^n (-1)^i e_i(\beta_1,\ldots , \beta_n) T^{i}.
\]
The logarithmic derivative is
\[
\frac{G'(T)}{G(T)} = \sum_{i=1}^r \frac{ - n_i \alpha_i}{ 1 - \alpha_i T} = - \sum_{j=0}^\infty p_{j+1}(\beta_1,\ldots,\beta_n) T^j.
\]
So knowledge of the power sums determines the rational function $\frac{G'(T)}{G(T)}$. Note that its partial fraction decomposition is unique.
First assume that none of the $\alpha_i$ are zero. The poles of this rational function are the $\alpha_i$, and the residue at $\alpha_i$ determines $n_i$ modulo $p$. This is enough to recover the values of the symmetric functions.
Otherwise, without loss of generality assume $\alpha_1 = 0$. Then we recover $\alpha_2,\ldots, \alpha_r$ and $n_2,\ldots, n_r$, and can find $n_1$ via $n_1 = n - (n_2 + \ldots + n_r)$.
\end{proof}
\begin{remark}
Without the restriction that $1 \leq n_i <p$, it is not possible to recover the multiset $\{ \beta_1,\ldots, \beta_n\}$, as one could add $p$ to one of the $n_i$ and subtract $p$ from another without changing the values of the power sums or the logarithmic derivative of $G(T)$.
Lemma~\ref{lem:powers} is computationally effective. If all we know is $n$ and the power sums, we can use a continued fraction algorithm on the power series to find the rational function. It is not immediately clear how many power sums are needed in general. If the field $\mathbf{k}$ is the finite field of $q$ elements, then $p_{q-1+i}=p_i$ and thus it's enough to
have $p_i$ for $i=1,\ldots,q-1$.
\end{remark}
\begin{bibdiv}
\begin{biblist}
\bib{BKT}{article}{
author={Bogomolov, Fedor},
author={Korotiaev, Mikhail},
author={Tschinkel, Yuri},
title={A Torelli theorem for curves over finite fields},
journal={Pure Appl. Math. Q.},
volume={6},
date={2010},
number={1, Special Issue: In honor of John Tate.},
pages={245--294},
issn={1558-8599},
}
\bib{Bom}{article}{
AUTHOR = {Bombieri, Enrico},
TITLE = {On exponential sums in finite fields},
JOURNAL = {Amer. J. Math.},
VOLUME = {88},
YEAR = {1966},
PAGES = {71--105},
}
\bib{BV}{article}{
author={Booher, Jeremy},
author={Voloch, Jos\'e Felipe},
title={Recovering algebraic curves from L-functions of Hilbert class fields},
journal={Res. Number Theory},
volume={6},
date={2020},
number={4},
pages={43},
}
\bib{CDL+}{article}{
AUTHOR = {Cornelissen, Gunther},
author={de Smit, Bart},
author={Li, Xin},
author={Marcolli, Matilde},
author={Smit, Harry},
TITLE = {Characterization of global fields by {D}irichlet {$L$}-series},
JOURNAL = {Res. Number Theory},
VOLUME = {5},
YEAR = {2019},
NUMBER = {1},
PAGES = {Art. 7, 15},
}
\bib{gassmann}{article}{
AUTHOR = {Ga{\ss}mann, F.},
TITLE = {Bemerkungen zur Vorstehenden Arbeit von Hurwitz: \"{U}ber {B}eziehungen zwischen den {P}rimidealen eines algebraischen {K}\"{o}rpers und den {S}ubstitutionen seiner
{G}ruppe},
JOURNAL = {Math. Z.},
FJOURNAL = {Mathematische Zeitschrift},
VOLUME = {25},
YEAR = {1926},
NUMBER = {1},
PAGES = {661--675},
ISSN = {0025-5874},
}
\bib{hk}{article}{
author={Halter-Koch, Franz},
title={A note on ray class fields of global fields},
journal={Nagoya Math. J.},
volume={120},
date={1990},
pages={61--66},
issn={0027-7630},
}
\bib{kummer}{article}{
AUTHOR = {Kummer, E. E.},
TITLE = {\"{U}ber die {E}rg\"{a}nzungss\"{a}tze zu den allgemeinen
{R}eciprocit\"{a}tsgesetzen},
JOURNAL = {J. Reine Angew. Math.},
FJOURNAL = {Journal f\"{u}r die Reine und Angewandte Mathematik. [Crelle's
Journal]},
VOLUME = {44},
YEAR = {1852},
PAGES = {93--146},
ISSN = {0075-4102},
MRCLASS = {DML},
MRNUMBER = {1578793},
DOI = {10.1515/crll.1852.44.93},
URL = {https://doi.org/10.1515/crll.1852.44.93},
}
\bib{LiRu}{article}{
author={Li, Wen-Ching Winnie},
author={Rudnick, Zeev},
title={Pair arithmetical equivalence for quadratic fields},
journal={Math. Z.},
volume={299},
date={2021},
number={1-2},
pages={797--826},
issn={0025-5874},
review={\MR{4311619}},
doi={10.1007/s00209-021-02706-w},
}
\bib{manin}{article}{
AUTHOR = {Manin, Ju. I.},
TITLE = {The {H}asse-{W}itt matrix of an algebraic curve},
JOURNAL = {Izv. Akad. Nauk SSSR Ser. Mat.},
FJOURNAL = {Izvestiya Akademii Nauk SSSR. Seriya Matematicheskaya},
VOLUME = {25},
YEAR = {1961},
PAGES = {153--172},
ISSN = {0373-2436},
}
\bib{MS}{article}{
author={Kumar Murty, Vijaya},
author={Scherk, John},
title={Effective versions of the Chebotarev density theorem for function
fields},
journal={C. R. Acad. Sci. Paris S\'{e}r. I Math.},
volume={319},
date={1994},
number={6},
pages={523--528},
}
\bib{neukirch}{article}{
AUTHOR = {Neukirch, J\"{u}rgen},
TITLE = {Kennzeichnung der {$p$}-adischen und der endlichen
algebraischen {Z}ahlk\"{o}rper},
JOURNAL = {Invent. Math.},
FJOURNAL = {Inventiones Mathematicae},
VOLUME = {6},
YEAR = {1969},
PAGES = {296--314},
ISSN = {0020-9910},
}
\bib{Pi}{thesis}{
AUTHOR = {Pintonello, M.},
TITLE = {Characterizing number fields with quadratic L-functions},
NOTE={ALGANT Master Thesis in Mathematics,
Università degli studi di Padova \& Universiteit Leiden, 25 June 2018},
}
\bib{rosen}{article}{
author={Rosen, Michael},
title={The Hilbert class field in function fields},
journal={Exposition. Math.},
volume={5},
date={1987},
number={4},
pages={365--378},
issn={0723-0869},
}
\bib{ruck}{article}{
AUTHOR = {R\"{u}ck, H.-G.},
TITLE = {Class groups and {$L$}-series of function fields},
JOURNAL = {J. Number Theory},
FJOURNAL = {Journal of Number Theory},
VOLUME = {22},
YEAR = {1986},
NUMBER = {2},
PAGES = {177--189},
ISSN = {0022-314X},
}
\bib{Schmidt1931}{article}{
author = {Schmidt, F. K.},
journal = {Mathematische Zeitschrift},
pages = {1--32},
title = {Analytische Zahlentheorie in K\"orpern der Charakteristik $p$},
volume = {33},
year = {1931},
}
\bib{Serre}{book}{
AUTHOR = {Serre, Jean-Pierre},
TITLE = {Algebraic groups and class fields},
SERIES = {Graduate Texts in Mathematics},
VOLUME = {117},
PUBLISHER = {Springer-Verlag, New York},
YEAR = {1988},
PAGES = {x+207}
}
\bib{solomatin}{article}{
author = {Solomatin, P.},
title= {On Artin L-functions and Gassmann equivalence for global function fields},
note = {Preprint (2016). \url{https://arxiv.org/abs/1610.05600}},
}
\bib{StV}{article}{
author={St\"{o}hr, Karl-Otto},
author={Voloch, Jos\'e Felipe},
TITLE = {A formula for the {C}artier operator on plane algebraic
curves},
JOURNAL = {J. Reine Angew. Math.},
VOLUME = {377},
YEAR = {1987},
PAGES = {49--64},
}
\bib{SV}{article}{
author={Sutherland, Andrew V.},
author={Voloch, Jos\'e Felipe},
title={Maps between curves and arithmetic obstructions},
note={preprint, arxiv:1709.05734, Proceedings of AGCT 16, AMS Contemporary Mathematics, to appear.},
}
\bib{tate}{article}{
AUTHOR = {Tate, John},
TITLE = {Endomorphisms of abelian varieties over finite fields},
JOURNAL = {Invent. Math.},
FJOURNAL = {Inventiones Mathematicae},
VOLUME = {2},
YEAR = {1966},
PAGES = {134--144},
ISSN = {0020-9910},
}
\bib{uchida76}{article}{
AUTHOR = {Uchida, K\^{o}ji},
TITLE = {Isomorphisms of {G}alois groups},
JOURNAL = {J. Math. Soc. Japan},
FJOURNAL = {Journal of the Mathematical Society of Japan},
VOLUME = {28},
YEAR = {1976},
NUMBER = {4},
PAGES = {617--620},
ISSN = {0025-5645},
}
\bib{uchida77}{article}{
AUTHOR = {Uchida, K\^{o}ji},
TITLE = {Isomorphisms of {G}alois groups of algebraic function fields},
JOURNAL = {Ann. of Math. (2)},
FJOURNAL = {Annals of Mathematics. Second Series},
VOLUME = {106},
YEAR = {1977},
NUMBER = {3},
PAGES = {589--598},
ISSN = {0003-486X},
}
\bib{VW}{article}{
AUTHOR = {Voloch, Jos\'{e} Felipe},
AUTHOR = {Walker, Judy L.},
TITLE = {Euclidean weights of codes from elliptic curves over rings},
JOURNAL = {Trans. Amer. Math. Soc.},
VOLUME = {352},
YEAR = {2000},
NUMBER = {11},
PAGES = {5063--5076},
}
\bib{Z1}{article}{
AUTHOR = {Zilber, Boris},
TITLE = {A curve and its abstract {J}acobian},
JOURNAL = {Int. Math. Res. Notices},
YEAR = {2014},
NUMBER = {5},
PAGES = {1425--1439},
}
\bib{Z2}{article}{
AUTHOR = {Zilber, Boris},
TITLE = {A curve and its abstract {J}acobian},
NOTE = {Corrected version of \cite{Z1}, preprint, http://people.maths.ox.ac.uk/zilber/JacobianCor.pdf},
YEAR = {2017},
}
\end{biblist}
\end{bibdiv}
\end{document}
| fcf57c7c61870e4f06cf43039c0efb1038302848 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\subsection{Mobile Edge Computing}
\label{MEC}
The exponential growth in smart device adoption is accelerating the ubiquitous Internet of Things (IoT) \cite{7488250, 7786106}, and catalyzing the development of computation-intensive applications, including augmented reality (AR), face recognition, interactive online gaming, and autonomous driving.
However, a computationally tedious task is unlikely to be executed at the local mobile device due to its resource constraint. Instead, such tasks can be offloaded to the remote cloud with abundant computation, storage, and energy resources \cite{5445167, zhang2010cloud, 6897914}. Despite the computational efficiency, the long communication distance between smart devices and the remote cloud inevitably introduces a high transmission latency, resulting in an unsatisfactory user quality of experience (QoE), especially for numerous real-time delay-sensitive applications.
To enable computationally intensive applications with sensitive delay requirements, mobile edge computing (MEC) \cite{hu2015mobile, 7901477, 7883826} has recently emerged as a promising paradigm. Compared with its cloud counterpart, MEC pushes the computing and storage capability to the network edge that is much closer to devices. As a result,
a device can offload its tasks to a proximal MEC server at a base station (BS) or access point (AP), and then collect the subsequent results from the MEC server. This generates the benefits of low latency and reduced mobile device energy consumption. Essentially, task offloading involves joint radio-and-computation resource allocation among multiple users.
A variety of specific task offloading policies have been designed in recent years. Most of them are focused on small-scale synchronous MEC systems, where tasks for a limited number of different users arrive simultaneously.
In this context, the computational resources are relatively abundant, and the policy design was usually formulated as a static centralized optimization problem, where the energy consumption and latency serve as the principal performance indicators \cite{7227025, 8327930, 8611399}.
For example, to minimize the weighted sum of energy consumption and end-to-end delay, the authors in~\cite{7227025} formulated the task offloading problem into a non-convex quadratically constrained quadratic program (QCQP), and semi-definite relaxation and randomization mapping based algorithms were proposed to achieve a near-optimal offloading performance.
A joint radio and computational resource allocation scheme was investigated in \cite{8327930}, and the offloading was formulated as a mixed-integer nonlinear programming problem (MINLP). A game theory-based approach was proposed to handle this problem. Recently, the offloading problem was formulated in~\cite{8611399} as a nonconvex optimization one, which aimed to maximize the weighted sum computation efficiency by imposing the constraints on local computation capability and energy resources.
\subsection{Task Offloading for Large-Scale Asynchronous MEC}
A wide range of emerging massive machine type communication applications, such as industrial automation, and smart transportation~\cite{6774858}, introduce the concept of large-scale asynchronous MEC systems. Typically, these applications involve a variety of tasks each with its own execution deadline.
Furthermore, task arrival patterns for a massive number of users exhibit notable stochasticity, featured by random and asynchronous task arrivals, distinct workloads, and diverse deadlines (see Fig.~\ref{fig: system state}). Due to the deadline constraints on tasks and limited computational resources (with respect to a massive number of users), task offloading policy design for large-scale asynchronous MEC systems becomes extremely challenging.
Task offloading policy should focus on the user selection and be implemented on a dynamic and real-time basis, considering both the task criticality and energy consumption.
To the best of our knowledge, the offloading policy design in a large-scale asynchronous MEC system is still an open challenge.
\textcolor{black}{
In this paper, we aim to address this challenge and propose an index-based task offloading policy. To adapt to the dynamic nature of task arrivals, we deviate from the classical static centralized optimization techniques with hard constraints, and turn to the bandit theory to capture the stochastic behavior in tasks.
We formulate the offloading policy design as a restless multi-armed bandit (RMAB) problem. Mathematically, MAB is a sequential decision model with a set of arms to choose from for the total reward maximization~\cite{5398950, 6035799, whittle1988restless}. At each round, only a subset of arms can be selected and their states will change, while the others remain frozen.
Removing such restrictions in the MAB, the RMAB allows the states of all arms to evolve over time regardless of the actions.
In our setting, we treat each user as an independent restless arm, and the arm state is represented by user task criticality including the remaining number of subtasks and remaining time to deadline.
We then design a reward function to strike a promising balance between two conflicting goals, i.e., minimizing the energy consumption and maximizing the task completion ratio. In this case, ``playing" an arm at each time slot is equivalent to selecting a user to offload its tasks. }
Our goal is to maximize the total discounted reward over the time horizon for the formulated RMAB, resulting in a new task offloading policy. However, the RMAB is generally PSPACE-hard and intractable~\cite{papadimitriou1999complexity}. To address this issue, we develop a novel method based on Whittle index (WI)~\cite{whittle1988restless}, so that multiple arms can be decoupled and the original $N$-dimensional problem reduces to $N$ independent
$1$-dimensional ones.
The key advantage of our WI offloading policy lies in its excellent scalability and low computational complexity, enabling fast user selection in task offloading.
At each time slot, each user only needs to separately calculate its scalar WI in closed form which provides a proxy to measure its task criticality. Each user then reports its WI to the BS, and the users with the highest indices are selected for task offloading. Besides, the WI policy can be implemented in a totally distributed manner.
Specifically, we first consider the scenario where the perfect knowledge of the user offloading energy consumption is available at users.
We exploit the WI theory and rigorously establish the indexability of the RMAB through the inductive method, which theoretically guarantees the existence of WI for our RMAB. On this basis, we derive a closed-form expression in terms of the task state and energy consumption for the WI computation.
Furthermore, when the task completion ratio becomes the focus, the shorter slack time less remaining workload (STLW) priority rule is introduced into the WI offloading policy for performance improvement, referred to as STLW-WI policy.
On the other hand, when the knowledge of user offloading energy consumption is not available prior to the offloading, the WI policy can not be directly applicable. To address this challenge, we develop Bayesian learning-enabled WI policies. In specific, we first integrate the WI policy with the maximum likelihood estimation (MLE) technique. Then, to further improve the performance, we propose a novel Bayesian learning with WI policy (BL-WI) given the conjugate prior. Finally, a refinement mechanism (PSBL-WI) based on prior-swapping is proposed for a fast inference given the non-conjugate prior.
It is verified by simulation that the proposed WI policy can achieve much better performance in terms of the total discounted reward, compared with several existing offloading policies. For the completion ratio-oriented task offloading, our STLW-WI policy achieves a higher task completion ratio.
When the user offloading energy consumption is unknown, our Bayesian learning-enabled WI policy can also achieve a favorable performance compared to the original WI policy.
\color{black}
\subsection{Contribution}
The main contributions of this paper can be summarized as follows.
\begin{itemize}
\item We develop an RMAB framework to enable task offloading for large-scale asynchronous MEC in a realistic setting. We propose a novel WI offloading policy through establishing the WI indexability and deriving a scalable solution with closed-form expression. When the knowledge of user offloading energy consumption is unknown prior to offloading, novel MLE-WI, BL-WI and PSBL-WI offloading policies from the Bayesian learning perspective are developed.
\item The developed WI method offers a potential low-complexity solution to a series of communication/computation resource scheduling problems (e.g., scheduling in a power-aware server farm \cite{5703092}), which typically involves complicated combinatorial optimizations. The developed WI method also addresses a challenging general RMAB problem in a dynamic environment with heterogeneous rewards and non-identical transition probabilities.
\end{itemize}
\subsection{Related Work}
In addition to the work introduced in Section \ref{MEC}, task offloading policy for MEC has been extensively studied in the literature. Two comprehensive surveys on various task offloading policy designs were provided in \cite{8016573, 7879258}. The work in \cite{8638800} designed a new MEC system to satisfy the ultra-reliable low-latency requirements in mission-critical applications. Specifically, a two-timescale association between user and server was proposed by utilizing the Lyapunov optimization and matching theory. For both time division multiple access (TDMA) and orthogonal frequency division multiple access (OFDMA), corresponding offloading policies have been developed in \cite{7762913}, aiming to minimize the user weighted sum energy consumption given the constraints of the user average latency.
As a special case of reinforcement learning \cite{sutton2011reinforcement}, the stateless MAB techniques have been applied to MEC systems \cite{8058414, 8627987, 8790775, 9082866, 8762108, 8887222}. In stateless MAB, the arms do not have any specific state. Each arm, when played, offers an i.i.d random reward drawn from a distribution with an unknown mean.
The authors in \cite{8058414} developed an energy-aware mobility management scheme based on MAB to perform MEC selection. In \cite{8627987}, an adaptive learning task offloading policy was proposed for vehicle edge computing based on the MAB theory.
In \cite{8790775}, the authors considered an edge service replacement problem, where they applied contextual combinatorial MAB to estimate users' demand based on side information.
An MAB online learning algorithm referred to as utility-table learning was proposed in \cite{9082866} to determine the optimal workload balance among MEC servers. In \cite{8762108}, the authors proposed an online task offloading policy based on the non-stationary MAB model, aiming to minimize the long-term total costs including latency, energy consumption and switching cost. Under the MAB framework, a two-stage resource sharing and task offloading strategy were developed in \cite{8887222}.
By contrast, the RMAB in this paper can be categorized into the stateful bandit model \cite{7498076}, where every arm is associated with some finite state space and the state evolves as a Markov process. When an arm is selected, the reward is drawn from some stationary distributions based on the current arm state.
There are several other works leveraging the WI theory first established in \cite{whittle1988restless}. According to \cite{gittins2011multi, ayesta2019unifying}, the WI solution in \cite{whittle1988restless} does not hold in general, and there is no unified solution that can cover all the RMAB problems. Consequently, the establishment of indexability needs to be studied for the individual problem.
For example, the problem formulation in \cite{5605371} suits a restless Bernoulli bandit with a two-state Markov chain, and the establishment of indexability highly depends on the transition probability of wireless channel occupancy. In contrast, the formulation in \cite{8295041} suits an RMAB problem with a static reward and identical state transition probability for each arm. In this paper, we aim to solve a new RMAB problem with heterogeneous rewards and a non-identical transition probability for each arm.
\subsection{Organization}
The rest of the paper is organized as follows. In Sections \ref{section2} and \ref{sec: RMAB Formulation}, we discuss the MEC system model and formulate the offloading problem as an RMAB, respectively.
In Section \ref{sec: WI}, we establish the indexability of the RMAB and develop a WI offloading policy. A Bayesian learning enabled WI offloading policy is proposed in Section \ref{sec: learning}. Simulation results are presented in Section \ref{sec: Numerical Results} followed by conclusions in Section \ref{section 7}.
\color{black}
\textit{Notation}: $\mathcal{N}(\mu,\Sigma)$ denotes the Gaussian distribution with a mean $\mu$ and a variance $\Sigma$. $\Gamma(\cdot)^{-1}$ denotes the Gamma inverse function. $x^{+} = \max(x,0)$. $\mathbbm{1}(\cdot)$ is the indicator function. $\binom{N}{M}$ denotes the combinations of selecting distinct $M$ items out of $N$.
For convenience, we also list most important symbols in Table \uppercase\expandafter{\romannumeral1}.
\begin{table*}[t]
\renewcommand\arraystretch{2}
\centering
\caption{TABLE OF SYMBOLS}
\begin{tabular}{|c|c|}
\hline
Variable & Description \\
\hline
$\beta$ & the discount factor \\
\hline
$B_{i,j}$ & the total number of subtasks for the $i$-th user in the $j$-th task \\
\hline
$b_{i,t}$ & the number of unfinished tasks in the $i$-th user at the $t$-th time slot \\
\hline
$C_i$ & the number of CPU cycles required to process $1$ bit data by the $i$-th user \\
\hline
$\epsilon_i$ & $\epsilon_i \sim \mathcal{N}(0, \Sigma_i)$; the measurement noise with noise variance $\Sigma_i$ \\
\hline
$E_{i}^{\text{loc}}$ & the local computing energy consumption by the $i$-th user \\
\hline
$E_{i, j}^{\text{off}}$ & the offloading energy consumption by the $i$-th user during the $j$-th task \\
\hline
$E_{i, j}^{\text{sav}}$ & $E_{i, j}^{\text{sav}} = k_i E_{i}^{\text{loc}} - E_{i, j}^{\text{off}}$; the energy consumption saving from the offloading for $k_i$ subtasks \\
\hline
$e_{i,j}^{\text{sav}}$ & $e_{i,j}^{\text{sav}} = E_{i,j}^{\text{sav}} + \epsilon$; a noisy observation version of the actual energy saving $E_{i,j}^{\text{sav}}$ \\
\hline
$\gamma_{i,j}$ & the number of observations for the $i$-th user during $j$-th task \\
\hline
$h_{i,j}$ & the channel gain of the $i$-th user during the $j$-th task \\
\hline
$k_i$ & the number of subtasks which can be processed by the MEC server in each time slot \\
\hline
$\kappa_{i,j}$ & small-scale fading channel power gain for the $i$-th user during $j$-th task \\
\hline
$l_{i,t}$ & $l_{i,t} \triangleq \tau_{i,t} - b_{i,t} / k_i$; the slack time of the $i$-th user at the $t$-th time slot \\
\hline
$M$ & the number of MEC servers in the system \\
\hline
$N$ & the number of users in the system \\
\hline
$r_{i,j}$ & the achievable transmission rate of the $i$-th user during $j$-th task \\
\hline
$U_i$ & the CPU frequency of the $i$-th user \\
\hline
$s_{i,t}$ & $s_{i,t} = \left(\tau_{i,t}, b_{i,t} \right)$; the state of the $i$-th user at the $t$-th time slot \\
\hline
$\textbf{S}_t$ & $\textbf{S}_t = \left(s_{1,t}, \cdots, s_{N, t}\right)$; the system state at the $t$-th time slot \\
\hline
$t_{i,j}^{d}$ & the deadline of the $i$-th user's $j$-th task \\
\hline
$\tau_{i,t}$ & $\tau_{i,t} \triangleq t_{i,j}^{d} - t + 1$; the number of remaining time slots to $t_{i,j}^{d}$ \\
\hline
$u_{i,t}$ & $u_{i,t} \in \left\{0, 1\right\}$; the offloading action on the $i$-th user at the $t$-th time slot \\
\hline
$\textbf{u}_t$ & $\textbf{u}_t = \left(u_{1,t}, \cdots, u_{N, t}\right)$; the actions taken by the BS for each user at the $t$-th time slot \\
\hline
$\omega_i(s_{i,t})$ & the WI of the $i$-th user given its current state $s_{i,t}$ \\
\hline
$X_{i,j}(t)$ & $X_{i,j}(t) = \left\{e_{i,j,1}^{\text{sav}}, \cdots, e_{i,j,\gamma_{i,j}}^{\text{sav}}\right\}$ the $i$-th user's energy saving observation set up to the $t$-th time slot \\
\hline
\end{tabular}\\
\label{tab:1}
\end{table*}
\color{black}
\section{System Model}
\label{section2}
\color{black}
\subsection{Large-Scale Asynchronous Task Arrival Model}
\textcolor{black}{Consider a large-scale asynchronous MEC system consisting of a BS and $N$ static users (indexed by $i \in \left\{1, \cdots, N\right\}$) shown in Fig. \ref{fig: system state}, where $N$ is reasonably large.
The system operates in a time-slotted structure, indexed by $t$. The BS is equipped with $M (M < N)$ independent MEC servers, and we assume that each MEC server can serve at most one user at each time slot.
Each user is running computation-intensive and delay-sensitive tasks with stochastic arrival patterns to be elaborated later. Each task is relatively large and can be further partitioned into a number of subtasks to be processed sequentially \cite{8663994}. It is assumed that the local computation capability of each user is not powerful enough to complete a task on time. Therefore, a user seeks assistance from the MEC server by offloading some subtasks for faster execution. As we only have $M$ MEC servers (limited computational resources), at most $M$ users can be selected to perform task offloading at each time slot. The number of possible combinations is $\binom{N}{M}$, which is usually extremely huge to handle\footnote{For example, when $M=30$, $N=100$, we have $\binom{N}{M} \approx 3 \times 10^{25}$.}.}
\begin{figure}[t]
\centering
{\includegraphics[width=0.5\textwidth]{Task_Arrival_Pattern}}
\caption{A large-scale asynchronous MEC system (left) with an illustrative asynchronous task arrival pattern (right). The area of the shaded rectangular indicates the size of the task $B_{i,j}$.}
\label{fig: system state}
\end{figure}
Next, we specify the asynchronous task arrival pattern with an example shown in Fig.~\ref{fig: system state} (a more detailed example is shown in Fig.~\ref{fig: offloading result}). The task arrival time and its deadline vary for each user. For the $i$-th user, at the $t_{i,j}^{a}$-th time slot, a new task $j$ arrives and reveals the number of subtasks $B_{i,j}$ and the task deadline $t_{i,j}^{d}$.
Without loss of generality, a user's subtasks are assumed to have equal size ($l_i$ bits). Upon arrival, the task starts to be processed and will be removed from the user buffer at the end of the $t_{i,j}^{d}$-th time slot. It is worth noting that $t_{i,j}^{a}$, $t_{i,j}^{d}$, and $B_{i,j}$ are discrete random variables. At the beginning of the $t$-th time slot, if the $i$-th user is idle, a new task will arrive with the probability $Q_i$.
\subsection{Computation Model}
Each user is assumed to be able to locally process one subtask in each time slot. For the $i$-th user, the number of CPU cycles required to process $1$ bit data is denoted by $C_i$, which may be different for various users~\cite{miettinen2010energy}. We denote the CPU frequency of the $i$-th user by $U_i$, then the local execution cycles for one subtask with $l_i$ bits can be calculated by $C_{\text{loc}} = C_i l_i / U_i$. Since the user is usually operating at a constant $U_i$ for the sake of energy efficiency~\cite{burd1996processor, 8606442}, the computing power for each CPU cycle can be calculated as $P_{0,i} = \lambda U_i^2$, where $\lambda$ is a power coefficient depending on the chip architecture~\cite{8606442}.
In this case, the local computing energy consumption $E_{\text{loc}, i}$ for the $i$-th user to process one subtask can be calculated as
\begin{equation}
E_{i}^{\text{loc}} = \lambda U_i C_i l_i.
\end{equation}
For MEC servers, we assume that their CPU frequency is a constant $U_{s}$. Accordingly, $k_i = \left \lfloor \frac{U_s}{U_i} \right \rfloor$ is the number of subtasks which can be processed in each time slot by an MEC server. That is, if the $i$-th user is selected to perform task offloading, it can transmit $k_{i}$ subtasks to the server.
\subsection{Communication Model}
Following~\cite{7762913}, the offloading process at each time slot can be divided into three steps: 1) a selected user uploads some subtasks to the MEC server; 2) the MEC server processes the subtasks; 3) the results are transmitted back to the user.
As the results are usually of small size~\cite{7563449}, the downloading time and associated energy consumption can be omitted. Therefore, in the whole process, the user offloading energy consumption mainly comes from the uploading phase.
As at most $M$ users can be selected at each time slot, we adopt OFDMA scheme for data transmission. For the $i$-th user during the $j$-th task, the achievable transmission rate can be calculated as
\begin{equation}
r_{i,j} = W_i \log_2\left(1+\frac{P_{i}^{tx}
h_{i,j}}{N_0 W_i}\right),
\end{equation}
where $W_i$ is the bandwidth, $P_{i}^{tx}$ is the transmission power, $N_0 W_i$ is the noise power. In addition, $h_{i,j} = \kappa_{i,j} g_0 (d_0/d_{i,j})^{\iota}$ denotes the channel gain of the $i$-th user~\cite{7956189} with $\kappa_{i,j}$ being the small-scale fading channel power gain.
In this paper, we adopt a widely used block fading channel model~\cite{ 8611399, 7541539}, and $\kappa_{i,j}$ keeps constant during the $j$-th task but varies independently from task to task, where $g_0$ is the path-loss constant, $\iota$ is the path-loss exponent, $d_0$ is the reference distance, and $d_{i,j}$ is the transmission distance between the $i$-th user to the BS.
Therefore, the required transmission time $t$ for sending $k_i$ subtasks can be calculated as $t_i = (k_i) l_i / r_{i,j} $. We assume that the length of a time slot is relatively large, so that it is always larger than the transmission time $t_i$. As a result, the offloading energy consumption of the $i$-th user during the $j$-th task is calculated as
\begin{equation}
E_{i, j}^{\text{off}} = t_i P_{i, j}^{tx} = \frac{k_i l_i}{r_{i,j}} P_{i, j}^{tx}.
\end{equation}
\color{black}
We need to consider both the task completion ratio and user's energy consumption under the limited computational resource ($M < N$).
At this point, our goal is to strike a promising balance between energy consumption and task completion ratio. Furthermore, the large-scale stochastic task arrivals require the offloading policy to perform user selection dynamically at each time slot with a low computational complexity, which is quite challenging.
\section{Restless Multi-armed Bandit Formulation}
\label{sec: RMAB Formulation}
To address the above challenges, we develop a novel WI policy that enables fast user selection at each time slot, in which offloading priorities are indicated by the value of WI.
For the specific implementation, we first formulate the offloading policy design as an RMAB \cite{whittle1988restless} to capture the randomness in tasks arrivals, workload and their deadlines. Essentially, we treat the remaining number of subtasks and remaining time to deadline of each task as the state of an arm.
The restless nature of the state naturally follows, because the state will also change even though the user is not selected for offloading. Under this formulation, we also create a new reward function as the performance metric, taking into account both deadline requirements and user offloading energy consumption.
In the following subsections, we elaborate on the RMAB with five key factors, including the system state, action, state transition, reward function, and objective function.
\subsubsection{System State}
For the $i$-th user, its state $s_{i,t}$ is represented by its current $j$-th task state at the beginning of the $t$-th time slot, i.e., $s_{i,t} = (\tau_{i,t}, b_{i,t})$, where $\tau_{i,t} \triangleq t_{i,j}^{d} - t + 1$ is the remaining time slots to the task deadline $t_{i,j}^{d}$, and $b_{i,t}$ is the number of the unfinished subtasks.
If there is no task, then $s_{i,t} = (0,0)$. Accordingly, $s_{i,t}$ can be written in a compact form as
\begin{equation}
s_{i,t} =
\begin{cases}
(0,0), & \mbox{no task}; \\
(\tau_{i,t}, b_{i,t}), & \mbox{otherwise}.
\end{cases}
\label{eq:state}
\end{equation}
Collecting the states of $N$ users, the system state ${\bf{S}}_t$ at the $t$-th time slot is denoted by ${\bf{S}}_t \triangleq \left(s_{1,t}, \cdots, s_{N,t}\right)$.
\subsubsection{Action}
At the beginning of each time slot, the action taken by the BS determines $M$ users (among $N$) which could offload their subtasks to the MEC. We define the action as ${\bf{u}}_t = \left(u_{1,t}, \ldots, u_{N,t}\right)$, where $u_{i,t}\in \left\{0,1\right\}$. %
When $u_{i,t} = 0$, task offloading is not allowed. When $u_{i,t} = 1$, the user will be selected to perform task offloading.
\subsubsection{State Transition}
As mentioned before, an MEC server can process at most $k_{i}$ subtasks compared to one subtask processed locally at the $i$-th user each time slot. Therefore, if a user can perform task offloading, the remaining number of subtasks will be reduced by $k_{i}$ maximally.
If the user is idle at the $t$-th time slot, i.e., $s_{i,t}=(0,0)$ , a new task will arrive with probability $Q_i$ at the $(t+1)$-th time slot.
Given the current state $s_{i,t}$ and the action $u_{i,t}$, the next state $s_{i,t+1}$ can be expressed by
\begin{itemize}
\item If $\tau_{i,t} \ge 2$,
\begin{equation}
s_{i,t+1}=
\begin{cases}
\left(\tau_{i,t}-1,(b_{i,t} - k_i)^{+}\right), \hspace{0.5cm} \mbox{if } u_{i,t} = 1; \\
\left(\tau_{i,t}-1,(b_{i,t} - 1)^{+}\right), \hspace{0.65cm} \mbox{if } u_{i,t} = 0; \\
\end{cases}
\end{equation}
\item If $\tau_{i,t} = 1$,
\begin{equation}
s_{i,t+1}=
\begin{cases}
\left(t_{i, j+1}^{d} - t, B_{i, j+1}\right), & \mbox{with Prob. } Q_i \\
(0,0), & \mbox{with Prob. } 1-Q_i;
\end{cases}
\end{equation}
\item If $s_{i,t} = (0,0)$, (assuming the index of the last task is $j$),
\begin{equation}
s_{i,t+1}=
\begin{cases}
\left(t_{i, j+1}^{d} - t, B_{i, j+1}\right), & \mbox{with Prob. } Q_i \\
(0,0), & \mbox{with Prob. } 1-Q_i;
\end{cases}
\end{equation}
\end{itemize}
where $x^{+} = \max(x,0)$. Note that when $\tau_{i,t} = 1$, the task of the $i$-th user will reach its deadline and be removed from the user at the end of the $t$-th time slot. Then, at the beginning of the $(t+1)$-th time slot, the $j+1$-th task with $B_{i, j+1}$ subtasks and deadline $t_{i,j+1}^{d}$ will arrive with probability $Q_i$.
\subsubsection{Reward Function}
Here, we create a reward function in \eqref{eq: Reward} to balance the user offloading energy consumption and deadline requirements,
\begin{figure*}[t]
\begin{equation}
\begin{aligned}
& R(s_{i,t}, u_{i,t}) \\
& =
\begin{cases}
E_{i,j}^{\text{sav}}u_{i,t} , &\mbox{if } \tau_{i,t} > 1, b_{i,t} > 0;
\\
E_{i,j}^{\text{sav}} u_{i,t} - F\left( \left[b_{i,t}-k_{i}u_{i,t} - (1-u_{i,t})\right]^{+}\right),
&\mbox{if } \tau_{i,t} = 1, b_{i,t} > 0;
\\
0, \hspace{5.95cm} &\mbox{otherwise.}
\end{cases}
\label{eq: Reward}
\end{aligned}
\end{equation}
\hrulefill
\end{figure*}
where $E_{i,j}^{\text{sav}} = \left(k_i E_{i}^{\text{loc}} - E_{i,j}^{\text{off}} \right)$ is the energy consumption saving from the offloading for the $k_i$ subtasks.
The penalty function is denoted by $F(x) = \alpha x^2 $ with $x$ indicating the number of unfinished subtasks, and $\alpha$ is the penalty parameter used to adjust penalty for unfinished tasks. \textcolor{black}{\footnote{The penalty function is widely used in Markov decision problem settings with specific form varying from case to case.}}
The key points can be highlighted as follows.
\begin{itemize}
\item When $\tau_{i,t} > 1$ and $b_{i,t} > 0$, the task has not reached its deadline, the reward is related to energy consumption saving $E_{i,j}^{\text{sav}}$ if performing task offloading ($u_{i,t}=1$).
\item When $\tau_{i,t} = 1$, the task will be removed at the end of the $t$-th time slot. If the task cannot be completed by its deadline, a penalty measured by the number of unfinished subtasks is imposed.
\item The benefit of the reward function is to strike a balance between the energy consumption and deadline requirements. For example, putting a priority on the deadline leads to a larger $\alpha$. By contrast, reducing $\alpha$ can increase energy savings for the battery-powered IoT devices to prolong their lifetime.
\end{itemize}
\subsubsection{Objective}
Our objective is to find a policy $\mathcal{G}$ to maximize the expected total discounted system reward with the constraint of the limited computational resources, which is defined by
\begin{equation*}
\begin{aligned}
({\bf P1}) ~~& \max_{\mathcal{G}}~\mathbbm{E}_{\mathcal{G}} \left[\sum_{t=0}^{\infty}\sum_{i=1}^{N}\beta^t R(s_{i,t}, u_{i,t}) \right] \\
& s.t. ~~~~ \sum_{i=1}^{N}u_{i,t} = M, ~ \forall t,
\label{eq:objective}
\end{aligned}
\end{equation*}
where $\beta~(0 < \beta \le 1)$ is the discount factor. The solution to ${\bf P1}$ forms the offloading policy that determines which users are selected to offload in each time slot.
Note that for the task offloading problem under consideration, in every slot, the user task continues to move one slot closer to their deadline, whether or not the task is offloaded in that slot. This makes the task offloading problem a restless bandit one.
However, the formulated RMAB is a PSPACE-hard sequential decision-making problem, which is intractable in general \cite{papadimitriou1999complexity}. The complexity in deriving the optimal solution is exponential with the number of users. This undesirable condition is further exacerbated by the extremely large dimension of the system state space. Therefore, the development of a scalable and low-complexity solution enabling fast and effective user selection at each time slot is a compelling necessity.
\section{Whittle index Based Task offloading Policy}
\label{sec: WI}
Mathematically, WI \cite{whittle1988restless} provides a potential avenue to obtaining an asymptotically optimal solution to a class of RMABs with the knowledge of reward and state information. The key idea is to decouple the arms through Lagrangian relaxation, and then prove that each arm is indexable. On this basis, a complex $N$-dimensional problem can be translated into $N$ independent $1$-dimensional ones, resulting in a scalable solution with a significant reduction in the computational complexity. This motivates us to exploit the WI theory to solve the formulated RMAB when the user offloading energy consumption are available. However, the major challenge lies in how to establish the indexability (existence) and derive the WI in an easily computed form (complexity in computation).
In this section, we first rigorously establish the indexability of the RMAB by considering a single arm reward maximization. Based on the induction method, we then prove that the formulated RMAB can admit a simple WI with closed-form expression. Finally, we elaborate on the practical implementation of the proposed policy.
\subsection{Whittle Relaxation}
A promising method, known as the Whittle relaxation, replaces the hard constraint $\sum_{i=1}^{N}u_{i,t} = M$ in $({\bf P1})$ by a soft one
\begin{equation}
\mathbbm{E}_{\mathcal{G}} \left[\sum_{t=0}^{\infty} \beta^{t} \sum_{i=1}^{N} u_{i,t}\right] = \frac{M}{1-\beta},
\end{equation}
which only requires that the expected discounted number of selected arms is equal to $M$. In other words, the number of selected arms at each time slot can be larger or less than $M$. In this case, the relaxed RMAB can be shown as
\begin{equation*}
\begin{aligned}
({\bf P2})~~ & \max_{\mathcal{G}}~\mathbbm{E}_{\mathcal{G}} \left[\sum_{t=0}^{\infty}\sum_{i=1}^{N}\beta^t R(s_{i,t}, u_{i,t}) \right] \\
& s.t. ~~~~ \mathbbm{E}_{\mathcal{G}} \left[\sum_{t=0}^{\infty} \beta^{t} \sum_{i=1}^{N} u_{i,t}\right] = \frac{M}{1-\beta}.
\label{eq:relaxed equation}
\end{aligned}
\end{equation*}
%
Leveraging the Lagrangian method, we can rewrite ${\bf P2}$ as the following unconstrained problem
\begin{equation}
\begin{aligned}
\max_{\mathcal{G}} \mathbbm{E}_{\mathcal{G}} \left\{\sum_{t=0}^{\infty} \left[ \sum_{i=1}^{N}\beta^t R(s_{i,t}, u_{i,t}) - \right. \right.
\hspace{2cm} \\ \left. \left. \delta \beta^t\left(\sum_{i=1}^{N} u_{i,t} - \frac{M}{1-\beta}\right)\right] \right\},
\label{eq:unconstrined}
\end{aligned}
\end{equation}
where $\delta$ is the Lagrange multiplier and will be referred to as subsidy hereafter. At this point, \eqref{eq:unconstrined} can be readily decoupled into $N$ subproblems (one for each arm) given by
\color{black}
\begin{equation}
\begin{aligned}
& \max_{\mathcal{G}} \mathbbm{E}_{\mathcal{G}} \left\{\sum_{t=0}^{\infty} \beta^t \left[ R(s_{i,t}, u_{i,t}) - \delta \left(u_{i,t} - \frac{M}{1-\beta}\right)\right] \right\}, \forall i.
\\
& = \max_{\mathcal{G}} \mathbbm{E}_{\mathcal{G}} \left\{\sum_{t=0}^{\infty} \beta^t \left[ R(s_{i,t}, u_{i,t}) - \delta u_{i,t} \right] + \delta \frac{M}{1-\beta} \right\}, \forall i.
\label{eq:single unconstrained}
\end{aligned}
\end{equation}
It is clear that the $N$ separate optimization problems interact with each other \eqref{eq:single unconstrained} through a scalar Lagrange multiplier $\delta$. Taking a close look at \eqref{eq:single unconstrained} and neglecting the last constant term
$\delta \frac{M}{1-\beta}$, our objective for each single arm $i$ is to maximize the following objective
\begin{equation}
\max_{\mathcal{G}} \mathbbm{E}_{\mathcal{G}} \left\{\sum_{t=0}^{\infty} \beta^t \left[ R(s_{i,t}, u_{i,t}) - \delta u_{i,t} \right] \right\}, \forall i.
\end{equation}
Then, following \cite{whittle1988restless}, we can define a modified reward of this single arm system as an equivalence to $\left[R(s_{i,t}, u_{i,t}) - \delta u_{i,t}\right]$ as follows
\begin{equation}
R^{\delta}(s_{i,t}, u_{i,t}) = R(s_{i,t}, u_{i,t}) + \delta \mathbbm{1}(u_{i,t}=0),
\label{eq:modified reward}
\end{equation}
\color{black}
where the indicator function $\mathbbm{1}(\cdot)$ gives $1$ if $u_{i,t}=0$.
We can interpret \eqref{eq:modified reward} as follows: (a) we select an arm and obtain an immediate reward $R(s_{i,t}, u_{i,t})$; (b) if the arm is not selected, we do not obtain an immediate reward (i.e., $R(s_{i,t}, u_{i,t})=0$) but receive an immediate subsidy $\delta$ (a virtual compensation from the economic view \cite{whittle1988restless}).
Given the initial state $s_{i,0}$, we use $V_{i,\beta}^{\delta}(s_{i,0})$ to denote the value function that represents the maximum expected total discounted reward with subsidy $\omega$. From the Bellman equation \cite{sutton1998introduction} we have
\begin{equation}
V_{i,\beta}^{\delta}(s_{i,0}) = \max_{u_i \in \left\{ 0,1 \right\}} \left\{ R^{\delta}(s_{i,0}, u_{i}) + \beta Q_{i, \beta}^{\delta} (s_{i,0}, u_{i}) \right\}.
\end{equation}
Here, $Q_{i, \beta}^{\omega} (s_{i,t}, u_{i,t})$ is defined as
\begin{equation}
Q_{i, \beta}^{\delta} (s_{i,t}, u_{i,t}) \triangleq \sum_{s_{i,t+1}^{\prime} \in {\bf{S}}_i} p(s_{i,t+1}^{\prime} |s_{i,t}, u_{i,t})V_{i,\beta}^{\delta}(s_{i,t+1}^{\prime}),
\end{equation}
where $p(s_{i,t+1}^{\prime} |s_{i,t},u_{i,t})$ is the state transition probability from the current state $s_{i,t}$ to the next state $s_{i,t+1}^{\prime}$ given action $u_{i,t}$.
We use $\mathcal{I}(\delta)$ to represent the set of states where the optimal action $u_i^{\star}$ is not selecting the $i$-th arm, i.e.,
\begin{equation}
\mathcal{I}(\delta) \triangleq \left\{s_i:u_i^{\star}(s_i) = 0\right\}.
\end{equation}
\subsection{Whittle Index Based Policy}
\label{Subsection: Indexability and Whittle index}
Now we can formally introduce the concept of indexability and WI.
\begin{definition}[Indexability\cite{whittle1988restless}]
The $i$-th arm is indexable if, as $\delta$ increases from $-\infty$ to $\infty$, $\mathcal{I}(\delta)$ expands monotonically from empty to the entire space. The RMAB problem is indexable if every arm is indexable.
\end{definition}
%
Essentially, the existence of indexability means that there is a priority order on each arm state $s_{i,t}$ in \eqref{eq:state}. Accordingly, when linking $\delta$ to $s_{i,t}$, the WI $\omega_i(s_{i,t})$, to be defined shortly, is used to quantify this order.
\begin{theorem}
The task offloading policy design in the MEC system formulated by the RMAB is indexable.
\end{theorem}
Given the definition of the indexability, we now prove the indexability of the formulated RMAB. The detailed proof can be found in Appendix A.
If the indexability holds, we can assign a WI $\omega_i(s_{i,t})$ for $s_{i,t}$ to measure the criticality of each task (user), which severs as the core indicator for the user task offloading selection. The formal definition of the WI can be provided as follows.
\begin{definition}[Whittle index\cite{whittle1988restless}]
If an indexable arm $i$ is in state $s_{i,t}$ at the $t$-th time slot, its WI $\omega_i(s_{i,t})$ is the least value of $\delta$ for which it is optimal to make the arm passive, that is
\begin{equation}
\begin{split}
& \omega_i(s_{i,t}) \triangleq \\ & \inf_{\delta} \left\{
R(s_{i,t},0) + \delta + \sum_{s_{i,t+1}^{\prime} \in {\bf{S}}_i}\beta p(s_{i,t+1}^{\prime} |s_{i},0)V_{i, \beta}^{\delta}(s_{i,t+1}^{\prime}) \right. \\ & \ge \left. R(s_{i,t},1) + \sum_{s_{i,t+1}^{\prime} \in {\bf{S}}_i}\beta p(s_{i,t+1}^{\prime} |s_{i},1) V_{i, \beta}^{\delta}(s_{i,t+1}^{\prime}) \right\},
\end{split}
\end{equation}
where $p(s_{i,t+1}^{\prime} |s_{i,t},u_{i,t})$ is the state transition probability from the current state $s_{i,t}$ to the next state $s_{i,t+1}^{\prime}$ given action $u_{i,t}$.
\end{definition}
After establishing the indexability of the RMAB and providing the definition of the WI, the remaining problem is how to compute the WI, which usually proves very difficult. For our RMAB, as the deadline and workload information become available once a task arrives, the arm state can be accurately captured. In the following, we will show that the unique structure of the RMAB can result in a closed-form expression for the WI.
\begin{figure*}[t]
\centering
{\includegraphics[width=0.9\textwidth]{Offloading_Result}}
\caption{An example of the proposed WI policy for a large-scale asynchronous MEC system.}
\label{fig: offloading result}
\end{figure*}
\begin{theorem}
The closed-form expression for the WI of the $i$-th user with task state $s_{i,t}=(\tau_{i,t},b_{i,t})$ is calculated as
\begin{equation}
\begin{aligned}
& \omega_i(\tau_{i,t},b_{i,t}) \\
& =
\begin{cases}
0, \hspace{2.5cm} \mbox{if } b_{i,t} = 0;
\\
E_{i,j}^{\text{sav}}, \hspace{2.0cm} \mbox{if } 1 \le b_{i,t} \le (\tau_{i,t} -1) k_{i} + 1;
\\
E_{i,j}^{\text{sav}} + \beta^{\tau_{i,t}-1}F(b_{i,t}-k_{i}\tau_{i,t} + k_{i} - 1),
\\
\hspace{2.7cm} \mbox{if } k_{i} \tau_{i,t} - k_{i} + 2 \le b_{i,t} \le k_{i} \tau_{i,t};
\\
E_{i,j}^{\text{sav}} + \beta^{\tau_{i,t}-1}F(b_{i,t}-k_{i}\tau_{i,t}+ k_{i} -1) \\
~~- \beta^{\tau_{i,t}-1}F(b_{i,t}-k_{i}\tau_{i,t}),
\\
\hspace{2.7cm} \mbox{if } b_{i,t} \ge k_{i} \tau_{i,t} + 1.
\end{cases}
\label{eq: Closed form Whittle}
\end{aligned}
\end{equation}
\label{theorem2}
\end{theorem}
\begin{proof}
Please find the proof of Theorem 2 in Appendix B.
\end{proof}
\color{black}
Here, we provide some insights behind \eqref{eq: Closed form Whittle}.
\begin{itemize}
\item If $b_{i,t} = 0$, it means that the user has no task to offload. The WI is equal to $0$ which is also the minimal value.
\item If $1 \le b_{i,t} \le (\tau_{i,t} - 1)k_{i} + 1$, it means that the $i$-th user's task can be finished at least one time slot ahead of the deadline. The WI is equal to the $i$-th user's energy saving $E_{i,j}^{\text{sav}}$.
\item If $k_{i} \tau_{i,t} - k_{i} + 2 \le b_{i,t} \le k_{i} \tau_{i,t}$, it means that the user should always be selected across all the time slots to finish its task. The WI takes into account both the energy savings and the non-completion penalty $F(b_{i,t}-k_{i}\tau_{i,t}+ k_{i} -1)$.
\item Finally, when it is impossible to finish the task (i.e., $b_{i,t} \ge k_i \tau_{i,t} + 1$), the WI is decreased by subtracting an extra non-completion penalty $F(b_{i,t}-k_{i}\tau_{i,t})$.
\item Note that the selection of tasks (users) depends on the penalty parameter $\alpha$. If we focus on task completion ratio by setting a large $\alpha$, the penalty term will dominate the WI in (17). In this case, those tasks with urgent deadline (i.e., $k_{i} \tau_{i,t} - k_{i} + 2 \le b_{i,t} \le k_{i} \tau_{i,t}$) are given higher priority. On the other hand, when we focus on the energy consumption by setting a small $\alpha$, those tasks with higher energy savings $E_{i,j}^{\text{sav}}$ have higher priority.
\end{itemize}
\color{black}
\subsection{Implementations}
\textbf{Theorem 2} indicates that the developed task offloading can be implemented in a very efficient manner. The whole procedure in its entirety can be summarized as follows.
\begin{enumerate}
\item At each $t$-th time slot, following \eqref{eq: Closed form Whittle} each user first calculates its WI $\omega_i(\tau_{i,t}, b_{i,t})$ based on the task state $\tau_{i,t}$, $b_{i,t}$ and energy saving $E_{i,j}^{\text{sav}}$.
\item In the message-passing phase prior to the computation offloading, each user then reports its WI $\omega_i(\tau_{i,t}, b_{i,t})$ to the BS. As $\omega_i(\tau_{i,t}, b_{i,t})$ is a scalar, the communication overhead is quite small.
\item Finally, the BS selects $M$ users with the largest WI for task offloading.
\end{enumerate}
\color{black}
An example is provided in Fig. \ref{fig: offloading result}, where we have $N=4$ users and $M=2$ MEC servers, and at each time slot only $2$ users can be selected to perform task offloading.
Take a closer look at User 1, it has 3 tasks (indicated by red, black and green) arriving sequentially. Task $1$ with $15$ subtasks (squares) arrives at time slot $1$, revealing time slot $6$ as its deadline\footnote{The square is blank in the time slot $11$ to indicate that not task is at User 1. }. The computation capability of User 1 and the MEC server is $1$ and $4$ subtasks per time slot, respectively\footnote{For different users, the computation capability of an MEC may different. }.
More specifically, from time slot $1$ to time slot $3$, User 1 will process its task locally with 1 subtask per time slot as the other users (User 2 and User 3) have larger WI. Then from the time slot $4$ to time slot $6$, User 1 gets permission for task offloading, thereby $4$ subtasks are processed per time slot through offloading.
\color{black}
The developed policy features very low computational complexity and communication overhead. The key element is to calculate the WI in a distributed manner \eqref{eq: Closed form Whittle}. For the specific implementation, it takes $\mathcal{O}(N)$ time to calculate the Whittle indices of all users in Step 1). In Step 2), the sorting process has the average time complexity $\mathcal{O}(N\log (N))$. Therefore, the total time complexity is only $\mathcal{O}(N\log (N))$. The overhead of collecting Whittle indices and delivering the decision by the MEC are only $\mathcal{O}(N)$ for $N$ users.
\begin{remark}
Another operation is that each user reports its task state $s_{i,t}=\left(\tau_{i,t}, b_{i,t} \right)$ and the transmission cost $E_i$ to the MEC server at each time slot for the WI computation. However, this will incur the additional communication overhead.
\end{remark}
\textcolor{black}{On the other hand, let us denote the average reward of the proposed policy, the optimal solution to $({\bf P1})$, and the solution to $({\bf P2})$ by $R_{\rm{WI}}$, $R_{\rm{Opt}}$, and $R_{\rm Relax}$, respectively. We have the following proposition to qualitatively indicate the performance of the proposed policy.}
\begin{proposition} The reward performance of the proposed policy can be shown as
\begin{equation}
R_{\rm{WI}}\leq R_{\rm{Opt}}\leq R_{\rm Relax}
\end{equation}
\end{proposition}
\begin{proof}
\textcolor{black}{The first inequality naturally holds because $R_{\rm{Opt}}$ corresponds to the optimal solution to the original problem $({\bf P1})$ with hard constraints. In terms of the second inequality, note that $R_{\rm Relax}$ is the average return under the relaxed constraint in $({\bf P2})$ (does not meet the hard constraint).}
\end{proof}
\textcolor{black}{
Given \textbf{Proposition 1}, it is quite difficult to quantify the gap between $R_{\rm{WI}}$ and $R_{\rm{Opt}}$.
However, we can infer their gap by numerically comparing $R_{\rm{WI}}$ with $R_{\rm Relax}$. Due to the relationship of $R_{\rm{WI}}\leq R_{\rm{Opt}}\leq R_{\rm Relax}$ in Proposition 1, if the performance of the WI policy ($R_{\rm{WI}}$) is close to that of the relaxed policy ($R_{\rm Relax}$), we can infer that the performance gap between the proposed policy and the optimal policy $R_{\rm{Opt}}$ is very small.}
\subsection{Completion Ratio-Oriented Task Offloading Policy}
\label{subsec: completion ratio}
In some scenarios, we pay particular attention to the task completion ratio, i.e., the proportion of tasks that can be completed before their deadlines.
In addition to setting a larger $\alpha$ in the penalty function, we notice that the WI \eqref{eq: Closed form Whittle} tends to give higher priority to tasks with less slack time, which is defined as $l_{i,t} \triangleq \tau_{i,t} - b_{i,t}/ k_i$ for the $i$-th user. However, the WI does not distinguish the users whose task states $s_i = (\tau_{i,t}, b_{i,t})$ satisfy $1 \le b_{i,t} \le k_i \tau_{i,t} - k_i + 1 $ (see the second case in \eqref{eq: Closed form Whittle}). This motivates us to further identify the criticality of the tasks in this scenario, thereby accommodating more tasks in a given time duration.
In the reward function \eqref{eq: Reward}, we only consider the deadline breaking by imposing a penalty when $\tau_{i,t} = 1, b_{i,t} > 0$. Although considering the remaining time when $\tau_{i,t} > 1, b_{i,t} > 0$ may further reduce the risk, we may not be able to establish WI indexability and derive a very simple closed-form WI solution as \eqref{eq: Closed form Whittle}.
To address this dilemma, we proposed a priority rule referred to as shorter slack time less remaining workload (STLW).
On this basis, we propose an enhanced WI-based offloading scheduling policy by applying the STLW rule (STLW-WI). The main idea of STLW-WI is to select the users with the highest Whittle indices without violating the STLW rule. The formal definition of the STLW rule can be stated as follows.
\begin{definition}[STLW Rule]
\label{Definition: STLW Principle}
Consider two users $m$ and $n$ with task states $s_{m,t}$ and $s_{n,t}$ at the $t$-th time slot. We define that the $m$-th user has priority over the $n$-th user if user $m$ has shorter slack time and less remaining workload than those of user $n$, i.e., $l_{m,t} \le l_{n,t}$ and $b_{m,t} \le b_{n,t}$, with at least one of the inequalities strictly holding.
\end{definition}
The STLW rule reorders the users based on their task states to ensure that the tasks with shorter slack time and less remaining workload should be given priority. In order to integrate the STLW rule into the WI-based offloading scheduling policy, we generate a directed acyclic graph (DAG) $\mathcal{G}=\left\{\mathcal{V},\varepsilon \right\}$, where $\mathcal{V}$ and $\varepsilon$ represent the vertex set (user set) and the edge set (users' relative priority), respectively.
In the DAG, a directed edge from the $m$-th vertex to the $n$-th vertex indicates that the $m$-th user has the priority over the $m$-th user. \textcolor{black}{The ourdegree of a vertex $m$ is the number of directed edges leaving $m$ while the indegree of $m$ is the number of directed edges entering $m$.
In order to preserve the priorities of users in terms of their Whittle indices whenever it is feasible, we utilize Kahn's algorithm \cite{kahn1962topological} with the largest WI vertex first criterion in the topological sorting \footnote{When no priority is set among users based on the STLW rule, users can still be ranked based on their Whittle indices}. Specifically, unlike the conventional Kahn's algorithm which selects the $0$ indegree vertex arbitrary, we select the $0$ indegree vertex with the largest WI firstly. The detailed topological sorting algorithm is shown in Algorithm~\ref{Alg: Khan's Algorithm}.
When there is no vertex with $0$ indegree, the topological sorting is terminated, and the top $M$ users in the rank list $L_{U}$ will be chosen to offload their subtasks to the MEC server. }
\color{black}
\begin{algorithm}[!t]
\caption{Kahn's Algorithm with largest WI vertex first.}
\label{Alg: Khan's Algorithm}
\begin{algorithmic}[1]
\color{black}
\STATE Initialize rank list $L_{U} = \emptyset$ that will contain the sorted user indices.
\STATE Compute each vertex's indegree, i.e., the number of incoming edges for each vertex.
\STATE Generate a set $S$ that contains all the vertices with $0$ indegree.
\WHILE{$S$ is not empty}
\STATE Select the vertex $m$ in the set $S$ with largest WI and add it to the tail of the rank list $L_{U}$.
\STATE Remove the vertex $m$ from the set DAG.
\FOR{Each vertex $n$ with an edge from the vertex $m$ to vertex $n$}
\STATE Decrease its indegree by $1$.
\IF{The indegree of vertex $n$ == 0}
\STATE Add the vertex $n$ into the set $S$.
\ENDIF
\ENDFOR
\ENDWHILE
\STATE Return rank list $L_{U}$.
\end{algorithmic}
\end{algorithm}
\begin{theorem}[The STLW Performance Analysis]
For every sequence of system state from the $t^{\prime}$-th time slot to the $(t^{\prime}+\Gamma)$-th time slot, the total discounted reward obtained by the policy with the updated STLW rule (denoted by $\mathcal{\tilde{G}}$) is not less than that achieved by the original policy (denoted by $\mathcal{G}$), i.e., we have
\begin{equation}
V_{\tilde{\mathcal{G}}}^{t^{\prime}+\Gamma}({\bf{S}}_{t^{\prime}}) \ge V_{\mathcal{G}}^{t^{\prime}+\Gamma}({\bf{S}}_{t^{\prime}}).
\label{eq: STLW Performance}
\end{equation}
\label{theorem: STLW Performance}
\end{theorem}
\begin{proof}
Please find the proof of Theorem 3 in Appendix C.
\end{proof}
\color{black}
\begin{remark}
It is worth noting that the WI offloading policy in Section IV requires each user to report its WI at each time slot. By contrast, the STLW-WI requires each user to report its current task state $s_{i,t} = (\tau_{i,t}, b_{i,t})$.
\end{remark}
\section{Learn to Offload Task}
\label{sec: learning}
\color{black}
In Section~\ref{Subsection: Indexability and Whittle index}, the proposed WI offloading policy requires the knowledge of user's energy saving $E_{i,j}^{\text{sav}}$ before task offloading. In some cases, however, $E_{i,j}^{\text{sav}}$ might not be available at $i$-th user prior to transmission due to lack of channel state information and offloading energy consumption $E_{i,j}^{\text{off}}$.
In this case, the WI policy is not directly applicable (c.f. \eqref{eq: Closed form Whittle}). To address this issue, in this section, we first integrate the WI policy with the maximum likelihood estimation (MLE). To further improve the performance, we propose a novel Bayesian learning with WI policy (BL-WI) given the conjugate prior, and a refinement algorithm based on prior-swapping suitable for the non-conjugate priors (PSBL-WI).
\subsection{Maximum Likelihood Estimation with WI Policy}
Only after performing task offloading at the $t$-th time slot, the $i$-th user can obtain an estimated energy saving $e_{i,j}^{\text{sav}}$ through equipment measurement for the $j$-th task\footnote{After a successful task offloading, the user can obtain its estimated offloading energy consumption $e_{i,j}^{\text{off}}$ by the equipment measurement, and calculate its estimated energy saving by $e_{i,j}^{\text{sav}} = k_i E_i^{\text{loc}} - e_{i,j}^{\text{off}}$.}.
Due to the energy measurement sensitivity and many other factors, $e_{i,j}^{\text{sav}}$ is a noisy version of the actual energy saving $E_{i,j}^{\text{sav}}$. Therefore,
the observation can be written as $e_{i,j}^{\text{sav}} = E_{i,j}^{\text{sav}} + \epsilon_i$, where $\epsilon_i $ is the measurement noise. Usually the noise is the result of summing a large number of different and independent random variables. From the central limit theorem, we have $\epsilon_i \sim \mathcal{N}(0, \Sigma_i)$, where $\Sigma_i$ is the noise variance. Therefore, $e_{i,j}^{\text{sav}}$ follows a Gaussian distribution written as $e_{i,j}^{\text{sav}} \sim \mathcal{N} \left(E_{i,j}^{\text{sav}}, \Sigma_i\right)$.
In the following, we integrate the WI policy with the maximum likelihood estimation (MLE) technique, where we treat $E_{i,j}^{\text{sav}}$ as an unknown variable.
Suppose that up to the $t$-th time slot for the $j$-th task offloading, the $i$-th user has performed task offloading $\gamma_{i,j}$ times, and obtained the corresponding observations $X_{i,j}(t) = \left\{e_{i,j,1}^{\text{sav}}, \cdots, e_{i, j,\gamma_{i,j}}^{\text{sav}}\right\}$.
The log likelihood is calculated by
\begin{equation}
\begin{aligned}
\ln p \left( X_{i,j}(t) | E_{i,j}^{\text{sav}}, \Sigma_i \right) = - \frac{\gamma_{i,j}}{2} \ln(2 \pi) - \frac{\gamma_{i,j}}{2} \ln |\Sigma_i| \\
- \frac{1}{2} \sum_{n=1}^{\gamma_{i,j}} (e_{i,j,n}^{\text{sav}} - E_{i,j}^{\text{sav}})^2 \Sigma_i^{-1}.
\end{aligned}
\end{equation}
Taking the derivative of the log likelihood with respect to $E_{i,j}^{\text{sav}}$, we obtain
\begin{equation}
\frac{\partial }{\partial E_{i,j}^{\text{sav}}} \ln p \left(X_{i,j}(t) | E_{i,j}^{\text{sav}}, \Sigma_i\right) = \sum_{n=1}^{\gamma_{i,j}} \Sigma_i^{-1} \left(e_{i,j,n}^{\text{sav}} - E_{i,j}^{\text{sav}}\right).
\end{equation}
By setting this derivative to zero, solution for the MLE of the energy saving is calculated as
\begin{equation}
\tilde{E_{i,j}^{\text{sav}}} = \frac{1}{\gamma_{i,j}} \sum_{n=1}^{\gamma_{i,j}} e_{i,j,n}^{\text{sav}}.
\label{eq: ML for E}
\end{equation}
Clearly from \eqref{eq: ML for E}, each user can average its past observations of energy savings to obtain an estimate $\tilde{E_{i,j}^{\text{sav}}}$, and calculate its WI according to \eqref{eq: Closed form Whittle}. However, such simple update may lead to an inaccurate estimate $\tilde{E_{i,j}^{\text{sav}}}$ when the number of observations is not enough.
\subsection{Bayesian Learning with WI Policy}
To further improve the performance, we propose a novel BL-WI policy. The key is to leverage Bayesian learning to obtain the estimated energy saving $\tilde{E_{i,j}^{\text{sav}}}$ rather than just simply averaging the past observations. From the BL perspective, a prior distribution on $\tilde{E_{i,j}^{\text{sav}}}$, obtained from historical observations, can be imposed \cite{lesaffre2012bayesian}. Specifically, an observation $e_{i,j}^{\text{sav}}$ after task offloading is drawn independently from a Gaussian distribution with an unknown mean $E_{i,j}^{\text{sav}}$ and an unknown variance $\Sigma_i$, i.e., $e_{i,j}^{\text{sav}} \sim \mathcal{N} \left(E_{i,j}^{\text{sav}}, \Sigma_i\right)$. We refer to $\theta_{i,j}= \left(E_{i,j}^{\text{sav}}, \Sigma_i\right)$ as the model parameter.
To conduct the Bayesian inference, we place a normal-inverse-gamma (NIG) conjugate prior~\cite{bishop2006pattern} on the model parameter with hyperparameters ${\lambda}_{i,j}$, $\mu_{i,j}$, $\Phi_{i,j}$ and $\nu_{i,j}$.
In specific, the variance $\Sigma_i$ follows an inverse gamma distribution
\begin{equation}
\Sigma_{i,j}|\left\{\Phi_{i,j},\nu_{i,j}\right\} \sim \Gamma^{-1}\left(\Phi_{i,j},\nu_{i,j}\right),
\end{equation}
and the mean $E_{i,j}^{\text{sav}}$ follows a Gaussian distribution
\begin{equation}
E_{i,j}^{\text{sav}}|\left\{\mu_{i,j},\lambda_{i,j},\Sigma_{i,j}\right\} \sim \mathcal{N}\left( {\mu}_{i,j},\frac{1}{\lambda_{i,j}}{\Sigma}_{i,j} \right).
\end{equation}
Note that the Gaussian prior is widely adopted due to a good approximation of different complex parameter distributions.
\begin{algorithm}[!t]
\caption{Bayesian Learning Based Whittle Index}
\begin{algorithmic}[1]
\STATE Initialize $\gamma_{i,j}$, model parameters $\theta_{i,j} = (\tilde{E_{i,j}^{\text{sav}}}, \tilde{\Sigma_i})$, and hyperparameters ${\lambda}_{i,j}$, $\mu_{i,j}$, $\Phi_{i,j}$, $\nu_{i,j}$, observation sets: $X_{i,j} = \emptyset$, $\forall$ $i = 1,...,n $.
\FOR{$t = 0,...,T$}
\FOR{$i = 1, \ldots, n$}
\STATE Each user calculates its WI $\omega_i$ based on its estimated energy saving $\tilde{E_{i,j}^{\text{sav}}}$ according to \eqref{eq: Closed form Whittle}.
\ENDFOR
\STATE All users transmit their $\omega_i$ to the BS.
\STATE The BS selects the top $M$ users based on their indices ${\omega}_{i}$. Denote the selected set as $\mathcal{M}$.
\STATE According to the action, update each user's state according to the predefined state transition.
\FOR{$i = 1, \ldots, n$}
\IF{$i \in \mathcal{M}$ }
\STATE Obtain an observation of energy saving $e_{i,j}^{\text{sav}} \sim \mathcal{N}(E_{i,j}^{\text{sav}}, \Sigma_i)$.
\STATE $\gamma_{i,j} = \gamma_{i,j} + 1$.
\STATE Append current observation into the observation set $X_{i,j}(t) \leftarrow X_{i,j}(t) \cup e_{i,j}^{\text{sav}}$.
\STATE Update hyperparameters ${\lambda}_{i,j}$, $\mu_{i,j}$, $\Phi_{i,j}$ and $\nu_{i,j}$ according to \eqref{eq: hyperpara1}, \eqref{eq: hyperpara2}, \eqref{eq: hyperpara3}, \eqref{eq: hyperpara4}.
\STATE Update $\tilde{\Sigma_i}$ and $\tilde{E_{i,j}^{\text{sav}}}$ according to \eqref{eq: modelpara1} and \eqref{eq: modelpara2}.
\ENDIF
\ENDFOR
\ENDFOR
\end{algorithmic}
\label{Alg: BL-WI}
\end{algorithm}
Given the observation up to the $t$-th time slot $X_{i,j}(t)$ for the $j$-th task, the $i$-th user can obtain its estimated energy saving $\tilde{E_{i,j}^{\text{sav}}}$ by efficient Bayesian inference. On this basis, we propose the BL-WI offloading policy, which is summarized in Algorithm~\ref{Alg: BL-WI}. It consists of three stages: initialization, decision making, and parameter update.
In the initialization stage (Line 1), we initialize the model parameters, hyperparameters of NIG and counter $\gamma_{i,j}$ for each user.
In the decision making (Lines 3-7), according to~\eqref{eq: Closed form Whittle} based on the estimated $\tilde{E_{i,j}^{\text{sav}}}$, each user calculates its WI, and then transmits it to the BS, where the $M$ users with the largest indices are selected to perform task offloading. We denote the selected user set by $\mathcal{M}$.
In the update stage (Lines 8-17), each user first updates its task state according to the state transition defined in Section~\ref{sec: RMAB Formulation}. Then, the user in the selected set $\mathcal{M}$ increases its counter $\gamma_i$ and updates its parameters by the Bayesian inference accordingly.
As the likelihood distribution lies in the exponential family, given the NIG prior on the unknown mean $E_{i,j}^{\text{sav}}$ and the variance ${\Sigma}_i$, we obtain the NIG posterior of $\theta_{i,j} = \left\{E_{i,j}^{\text{sav}}, {\Sigma}_i\right\} $ by the conjugacy property. Specifically, the posterior shares the same form as the prior whose hyperparameters $\lambda_{i,j}^{new}$, ${\mu}_{i,j}^{new}$, $\nu_{i,j}^{new}$, and ${\Phi}_{i,j}^{new}$ are acquired by aggregating the observations $X_{i,t}(t)$ calculated as
\begin{equation}
\lambda_{i,j}^{new} = \lambda_{i,j}^{old} + \gamma_{i,j,t},
\label{eq: hyperpara1}
\end{equation}
\begin{equation}
{\mu}_{i,j}^{new} = \frac{\lambda_{i,j}^{old}{\mu}_{i,j}^{old}+ \gamma_{i,j,t} \overline{E_{i, j,\gamma_i}^{\text{sav}}} }{\lambda_{i,j}^{old} + \gamma_{i,j,t}},
\label{eq: hyperpara2}
\end{equation}
\begin{equation}
{\Phi}_{i,j}^{new} = {\Phi}_{i,j}^{old}+\sum_{n=1}^{\gamma_{i,j,t}}\left( {e}_{i,j,n}-\overline{E_{i,j}} \right)^2 +\frac{\lambda_{i,j}^{old} \gamma_{i,j,t}}{\lambda_{i,j}^{old}+ \gamma_{i,j,t}}\frac{(\overline{E_{i,j}}-{\mu}_{i,j}^{old})^2}{2},
\label{eq: hyperpara3}
\end{equation}
\begin{equation}
\nu_{i,j}^{new} = \gamma_{i,j,t}/2 + \nu_{i,j}^{old},
\label{eq: hyperpara4}
\end{equation}
where $\overline{E_{i,j}}$ is the average of the observations $X_{i,j}(t)$. Finally, a user's estimated noise variance $\tilde{{\Sigma}_i}$ and estimated energy saving $\tilde{E_{i,j}^{\text{sav}}}$ can be sampled with updated hyperparameters as
\begin{equation}
\tilde{{\Sigma}_{i}}|\{{\Phi}_{i,j}^{new},\nu_{i,j}^{new}\} \sim \Gamma^{-1}\left( {\Phi}_{i,j}^{new},\nu_{i,j}^{new} \right),
\label{eq: modelpara1}
\end{equation}
\begin{equation}
\tilde{E_{i,j}^{\text{sav}}}|\{{\mu}_{i,j}^{new},\lambda_{i,j}^{new},{\Sigma}_{i,j}^{\prime}\} \sim \mathcal{N}\left({\mu}_{i,j}^{new},\frac{1}{\lambda_{i,j}^{new}}{\tilde{{\Sigma}_{i}}} \right).
\label{eq: modelpara2}
\end{equation}
\subsection{Refinement with BL-WI Policy}
\label{sec: learning subsection refine}
Although the NIG prior allows for a tractable and convenient Bayesian inference due to the conjugacy property, in practice, the true prior may not be conjugate (e.g., Laplace distribution). Usually, inferring the exact posterior given the non-conjugate prior is intractable, and approximate posterior inference algorithms such as Markov chain Monte Carlo (MCMC) are needed. However, inference by sampling method for the target posterior is very costly and inefficient in an online setting~\cite{zhou2018racing}, as all past observations must be involved to generate the action at each iteration.
Motivated by~\cite{neiswanger2017post}, we adopt the prior swapping (PS) technique to make use of the pre-defined false prior (e.g., Gaussian prior), rather than running standard inference algorithms on the target prior. Here, the original Bayesian inference is divided into two simple steps: we first carry out the closed-form inference with the conjugate prior, and then utilize the PS technique to derive the posterior with the true non-conjugate prior. Hereafter, we refer to this new policy as PSBL-WI policy.
Denote the true prior distribution over the model parameter $\theta_{i,j}$ by $\pi_t({\theta_{i,j}})$. Suppose now we have chosen a conjugate prior distribution $\pi_f(\theta_{i,j})$, which is referred to as the false prior.
To leverage the inferred false posterior for computing the true posterior, we define a prior swapping distribution $p_s(\theta_{i,j})$
\begin{equation}
p_s(\theta_{i,j}) \propto \frac{ \tilde{p_f(\theta_{i,j})} \pi_t(\theta_{i,j}) }{ \pi_f(\theta_{i,j}) },
\end{equation}
where $\tilde{p}_f(\theta_{i,j})$ is the interference result of the false posterior. Note that in our case, $\tilde{p}_f(\theta_{i,j}) = p_f(\theta_i|X_{i,j}(t))$ has an analytic form due to the conjugacy property, and $p_s(\theta_{i,j}) = p(\theta_i|X_{i,j}(t))$ becomes the true posterior density function.
Then our strategy is to use $p_s(\theta_{i,j})$ in random walk Metropolis-Hastings (MH) algorithm~\cite{chib1995understanding} to approximate the true posterior distribution. Unlike the traditional MH whose computational complexity highly depends on the number of observations, the PS technique ensures that each iteration only requires to evaluate a few simple analytic expressions, and the complexity is independent of the number of observations. We denote the proposal distribution by $q (\theta_{i,j,p} | \theta_{i,j,k-1})$, where $\theta_{i,j,p}$ is the proposed sample and $\theta_{i,j,k-1}$ is the old one.
Then the MH ratio (acceptance ratio) is calculated as $\min \left(1,\rho\right)$ with
\begin{equation}
\rho = \frac{ p_s(\theta_{i,j,p}) q(\theta_{i,j,k} | \theta_{i,j,p}) }{p_s(\theta_{i,j,k}) q(\theta_{i,j,p}| \theta_{i,j,k})}.
\label{eq: MH accept ratio}
\end{equation}
Finally, after drawing $K$ samples of model parameters $\theta_{i,j}$, we can average $K$ energy saving samples $\left\{\hat{E}_{i,j,1}, \cdots, \hat{E}_{i,j,K} \right\}$ to obtain the estimated energy saving $\tilde{E_{i,j}^{\text{sav}}} = \frac{1}{K} \sum_{k=1}^{K}{\hat{E}_{i,j,k}}$. The detailed PSBL-WI policy is presented in Algorithm 3.
\begin{algorithm}[!t]
\caption{Prior swapping Bayesian learning Whittle index}
\begin{algorithmic}[1]
\color{black}
\STATE Initialize $\gamma_{i,j}$, model parameters $\theta_{i,j} = \left( \tilde{E_{i,j}^{\text{sav}}}, \tilde{\Sigma_i} \right)$, observation sets: $X_{i,j} = \emptyset$, $\forall$ $i = 1,...,n $, and desired number of samples $K$.
\FOR{$t = 0,...,T$}
\FOR{$i = 1, \cdots, n$}
\FOR{$j = 1,\cdots, K$}
\STATE Sample a new proposal $\theta_{i,j,p} \sim q \left(\theta_{i, j,p} | \theta_{i, j,k-1}\right)$
\STATE Draw $\tilde{u} \sim U(0,1)$
\IF{$\tilde{u} < \min(1, \rho)$}
\STATE accept the proposal $\theta_{i,j,k} \leftarrow \theta_{i,j,p}$
\ELSE
\STATE Reject the proposal $\theta_{i,j,k} \leftarrow \theta_{i,j,k-1}$
\ENDIF
\ENDFOR
\STATE Average samples $\tilde{E_{i,j}^{\text{sav}}} = \frac{1}{K} \sum_{k=1}^{K} \hat{E}_{i,j,k}$
\STATE Each user calculates its WI $\hat{\omega}_i$ as \eqref{eq: Closed form Whittle}.
\ENDFOR
\STATE Same decision stage as Algorithm.~\ref{Alg: BL-WI}
\STATE Same update stage as Algorithm.~\ref{Alg: BL-WI}.
\ENDFOR
\end{algorithmic}
\label{Alg: PSBL-WI}
\end{algorithm}
\color{black}
\section{Numerical Results}
\label{sec: Numerical Results}
In this section, we evaluate the performance of the proposed index-based policies by simulation. In Section~\ref{simulation: first subsection}, we first verify the proposed WI policy when the users have exact information about their energy savings from the task offloading. Then we show the impact of penalty parameter $\alpha$ on the total energy saving and completion ratio. Finally, we verify the performance of the proposed STLW-WI policy when the completion ratio becomes the main performance metric. In Section~\ref{simulation: third subsection}, we present the performance of the Bayesian learning-enabled WI policies when $E_{i,j}^{\text{sav}}$ is not available before transmission, together with its comparison to the WI policy with the knowledge of $E_{i,j}^{\text{sav}}$.
\textcolor{black}{
The common parameters in the simulations are summarized as follows. The total rounds of task offloading is $T = 200$ with discount factor $\beta = 0.99$. The users are randomly located in the MEC system, with distance to the BS $d_{i}$ independently drawn from a uniform distribution $\mathcal{U}(0.1, 0.3)$ in kilometers. The small-scale fading channel power gains are exponentially distributed with unit mean, i.e., $\kappa_{i,j} \sim \text{Exp}(1)$~\cite{8016573}.
We set $\sigma_0^2 = -174$ dbm/Hz, $g_0 = -40$ dB, $d_0 = 1$ m, and $\iota = 4$. Without loss of generality, each allocated sub-channel has the same bandwidth $W_i = 1$ MHz. The transmission power for each user follows a uniform distribution $P_i^{tx} \sim U(20, 25)$ dbm. For local computing, the power efficient is $\lambda = 10^{-28}$. The CPU frequency of each user $U_i$ is selected from the set $\left\{0.2, 0.4 \cdots, 1\right\}$ GHz. The required number of CPU cycles per bit is $C_i \in \left\{1, 2, 3, 4, 5\right\} \times 10^5$ cycles/bit.
The CPU frequencies of all MEC servers are fixed $U_s = 2$ GHz. When the user is idle, the task generation probability is $Q = 0.7$. For the task specification, the duration and size of a task are bounded by $10$ time slots and $30$ subtasks, respectively. The size of a subtask is $l_i \in \left\{100, 150, 200\right\}$ bits. The penalty function in the reward function is set as $P(b) = \alpha + 0. 1b^2$ with a different value of $\alpha$ in different subsections. }
\begin{figure*}
\centering
\subfigure[Constant $M/N = 0.3$]{
\includegraphics[width=0.45\textwidth]{Const_MN3_Reward.eps}
\label{fig: Constant M/N(a)}
}
\subfigure[Constant $M/N = 0.5$]{
\includegraphics[width=0.45\textwidth]{Const_MN4_Reward.eps}
\label{fig: Constant M/N(b)}
}
\caption{Performance comparison in terms of the total discounted reward.}
\label{fig: Constant M/N}
\end{figure*}
\subsection{The Performance of the WI Policy}
\label{simulation: first subsection}
When the perfect knowledge of $E_{i,j}^{\text{sav}}$ is available at the $i$-th user, we compare our WI policy with the following conventional policies.
\begin{itemize}
\item \textit{Earliest Deadline First (EDF)~\cite{liu1973scheduling}:} EDF is a traditional dynamic priority policy where at each time slot the BS always selects $M$ users with minimal task remaining time $\tau_{i,t}$. Each user needs to report its current remaining time $\tau_{i,t}$ to the BS.
\item \textit{Least Slack Time (LST)~\cite{393496}:} LST chooses $M$ users based on their task slack time $l_{i,t} = \tau_{i,t} - b_{i,t}/k_i$. At each time slot, the user needs to transmit its task slack time $l_{i,t}$ to the BS.
\item \textit{Greedy Policy:} The greedy policy selects users according to their immediate reward $R_i$ as defined in \eqref{eq: Reward}, in which $M$ users with the highest rewards are selected. Each user calculates its immediate reward $R_i$ and transmits it to the BS.
\item \textit{Relax Solution:} In addition, we obtain the unrealistic relax solution to (${\bf P2}$) according to the method provided in~\cite{whittle1988restless}. Note that this relaxed solution is the optimal solution to {\bf{P2}}. The maximal expected average reward under relaxed constraint is
\begin{equation}
\bar{R} = \inf_{\delta} \left\{\sum_{i=1}^{N} V^{\delta}_{i,\beta} - \delta(N-M)\right\},
\end{equation}
where $V^{\delta}_{i,\beta}$ is the value function of the $i$-th user with subsidy $\delta$, and $\delta$ can be obtained by an exhaustive search to maximize the $\bar{R}$. Note that this solution does not satisfy the constraint in ${\bf P1}$.
\end{itemize}
To take into account both task deadline and user offloading energy consumption, we set $\alpha = 0.5$.
In comparing the WI policy with the four methods aforementioned, Fig.~\ref{fig: Constant M/N} considers two scenarios with different $M/N$. The total discounted reward is served as the performance metric. It is clearly shown that our WI policy significantly outperforms the other heuristic policies. Taking a closer look at the WI policy and the relaxed solution to $({\bf P2})$, we can infer that the performance of the WI policy is close to the optimal solution according to \textrm{Proposition 1}.
In Fig.~\ref{fig: Constant M/N(a)}, when the number of MEC servers is not relatively enough to the number of users, the total discounted rewards obtained by heuristic policies decrease with the increment of the number of users. While our WI policy can not only achieve a positive reward but also increase with the number of users. It is because our WI policy can fully utilize the system's state information to rank the priorities among users. Compared with Fig.~\ref{fig: Constant M/N(a)} and Fig.~\ref{fig: Constant M/N(b)}, one can see that the total discounted reward increases with the ratio of available MEC servers increasing from $0.3$ to $0.5$. The reason is that more users can be selected to perform task offloading, thereby more tasks can be finished before the deadline.
In Fig.~\ref{fig: Constant N}, we fix the number of users $N=100$ and vary the number of MEC servers $M$. It can be seen that our WI policy outperforms the other policies in terms of the total discounted reward. When the number of available MEC servers is limited (e.g., $M = 25$ with $M/N = 0.25$), the performance gap between different policies is small due to very insufficient computational resources. In fact, there are a large number of tasks that cannot meet their deadlines. With the increasing number of $M$, the performance of all the policies can be improved. Given the adequate computing resources (e.g. $M = 45$ with $M/N = 0.45$), most of the tasks can be accomplished by their deadline in those policies. Therefore, all policies achieve closer performance.
\begin{figure}[t]
\centering{
\includegraphics[width=0.45\textwidth]{Const_N_Reward.eps}
}
\caption{Performance comparison in terms of the total discounted reward with constant $N$.}
\label{fig: Constant N}
\end{figure}
\begin{figure}[t]
\centering{
\includegraphics[width=0.45\textwidth]{Energy_Saving_Complete_Ratio.eps}
}
\caption{The total discounted energy savings and completion ratio versus $\alpha$ in penalty function.}
\label{fig: Various alpha}
\end{figure}
\begin{figure}[t]
\centering{
\includegraphics[width=0.45\textwidth]{EnergySaving_MN.eps}
}
\caption{Performance comparison for energy saving focus case.}
\label{fig: energy saving focus}
\end{figure}
\begin{figure*}
\centering
\subfigure[Completion Ratio]{
\includegraphics[width=0.45\textwidth]{Deadline_MN_CompleteRatio.eps}
\label{fig: deadline focus completion ratio}
}
\subfigure[Reward Per Task]{
\includegraphics[width=0.45\textwidth]{Deadline_MN_TotalReward.eps}
\label{fig: deadline focus reward}
}
\caption{Performance comparison for task completion focus case.}
\label{fig: deadline focus}
\end{figure*}
\begin{figure}[t]
\centering
{\includegraphics[width=0.45\textwidth]{Learn_Compare_Moving.eps}}
\caption{The performance comparison under Gaussian prior.}
\label{fig: Bayesian Learning Performance}
\end{figure}
\begin{figure}[t]
\centering
{\includegraphics[width=0.45\textwidth]{Learn_Compare_NonConjugate.eps}}
\caption{The performance comparison under Laplace prior.}
\label{fig: Bayesian Learning Performance Non-Conjugate}
\end{figure}
Essentially, the penalty parameter $\alpha$ in \eqref{eq: Reward} strikes a tradeoff between the energy saving and the task completion ratio, which is illustrated in Fig.~\ref{fig: Various alpha}. We can find that, when $\alpha$ is small, the WI policy poses an emphasis on the energy savings, resulting in a relatively low task completion ratio. With the increase of $\alpha$, the WI policy becomes task completion ratio-oriented, leading to reduced energy savings and higher task completion ratio. It is worth noting that more computational resources result in a larger $\alpha$ to make the WI policy focus on the task complete ratio.
\textcolor{black}{In Fig.~\ref{fig: energy saving focus}, we compare the performance achieved by different policies in terms of the total energy saving with a small penalty parameter $\alpha = 0.001$. As the greedy policy only selects the task with largest energy savings to offload, it achieves maximum total energy savings. It is clearly shown that the WI-based offloading policy outperforms the EDF and LST policies, and is close to the greedy one.}
In Fig.~\ref{fig: deadline focus completion ratio}, we compare the performance achieved by different policies with a large penalty parameter $\alpha = 5$ in terms of the task completion ratio.
Note that as the computational resource here is limited ($M < N$), not all the tasks can be finished before their deadlines.
It is clearly shown that the WI-based offloading policy outperforms the EDF, LST, and Greedy policies. Furthermore, the STLW-WI policy outperforms the original WI policy because reordering of users with the STLW rule gives urgent tasks higher priorities. When $M/N = 0.45$, the completion ratio of the proposed STLW-WI and WI policies is 82\% and 80\%, compared to 72\%, 70\%, 66\% in LST, EDF and Greedy, respectively. The total discounted reward in Fig.~\ref{fig: deadline focus reward} also demonstrates that applying STLW rule can reduce the penalty of unfinished tasks and improve the performance. This is consistent with the theoretical analysis in Theorem~\ref{theorem: STLW Performance}.
\subsection{The Performance of the BL-enabled WI Policy}
\label{simulation: third subsection}
Next, we evaluate the performance of the proposed BL-enabled WI policy without the knowledge of user energy saving before offloading. The initialized parameters in BL-WI policy are set as follows: $\lambda_{i,j} = 1$, $\mu_{i,j} = 1$, $\Phi_{i,j} = 1$, $\nu_{i,j} = 1$, $\tilde{\Sigma_i} = 0$, $\tilde{E_{i,j}^{\text{sav}}} = 1$, and $\gamma_{i,j} = 0$. After the $i$-th user performing task offloading at the $t$-th time slot for the $j$-th task, the observation of the energy saving is drawn from a Gaussian distribution: $e_{i,j,t} \sim \mathcal{N} (E_{i,j}^{\text{sav}}, \Sigma_i)$, with the observation noise drawn from a uniform distribution $\Sigma_i \sim U(0.5, 1)$. The desired number of samples in PSBL-WI policy is $K = 10$. For reference purpose, we also include the performance of the WI policy with the knowledge of the energy saving.
When the energy saving $E_{i,j}^{\text{sav}}$ follows a Gaussian prior distribution, we evaluate the performance of Algorithm \ref{Alg: BL-WI} in Fig.~\ref{fig: Bayesian Learning Performance}. The prior distribution is $E_{i,j}^{\text{sav}} \sim \mathcal{N}\left(E_{i,j}^{\text{init}}, \Sigma_0\right)$ with mean $E_{i,j}^{\text{init}} = 1$ and variance $\Sigma_{i} = 0.1$.
It is clearly shown that the BL-WI policy can learn faster and more accurate than the MLE-WI policy under various $M/N$. \textcolor{black}{Since the channel gain changes every 20 time slots, the performance of MLE-WI policy is limited by the number of observations collected in 20 time slots.} Additionally, comparing with WI policy, the reward gaps of both BL-WI policy and MLE-WI policy decrease when the number of available MEC servers increases, i.e., $M/N$ increases from $0.3$ to $0.5$,. In particular, the BL-WI policy achieves a much more significant performance improvement than the MLE-WI policy counterpart. It is because more users have opportunities to perform task offloading and obtain more observation samples, accelerating the Bayesian learning process and decreasing the sample bias in MLE.
Next, we evaluate the performance of Algorithm \ref{Alg: PSBL-WI} when the energy saving $E_{i,j}^{\text{sav}}$ has a non-conjugate prior distribution. Specifically, we place a Laplace distribution as the prior distribution $E_{i,j}^{\text{sav}} \sim Laplace\left({E}_{i,j}^{\text{init}}, b_0 \right)$ with location parameter $E_{i,j}^{\text{init}} = 1$ and scale parameter $b_0 = 0.2$. For the BL-WI policy, we still use a conjugate NIG prior to performing exact Bayesian inference.
Fig.~\ref{fig: Bayesian Learning Performance Non-Conjugate} shows that the PSBL-WI policy outperforms the other ones in the non-conjugate case. Although the Gaussian prior brings the convenience in Bayesian inference, comparing the performance of BL-WI policy in Fig.~\ref{fig: Bayesian Learning Performance} and Fig.~\ref{fig: Bayesian Learning Performance Non-Conjugate}, the reward gap between BL-WI policy and WI policy increases due to the false prior assumption.
Similar to the conjugate prior case, when the number of MEC servers increases, the performance of BL-enabled WI policies improve with more samples obtained during the offloading.
\subsection{Discussion}
\color{black}
In our problem, we assume a pre-allocated bandwidth scheme. However, it is possible to include the bandwidth allocation into the formulated problem. To do so, the action at each time slot will be modified as $a_{i,t} = (u_{i,t}, d_{i,t})$, where $u_{i,t} \in \left\{0, 1\right\}$ is the offloading decision and $d_{i,t} \in \left[d_{\min}, d_{\max}\right]$ the bandwidth allocation. Accordingly, two constraints imposed on the question $\sum_{i}^{N} u_{i,t} = M $ (only $M$ users can be selected to perform task offloading), and $\sum_{i} u_{i,t} d_{i,t} = W $ (the sum of bandwidth allocation is $W$).
This new RMAB problem with an extra constraint (bandwidth limitation) makes the establishment of the indexability difficult, and we consider this problem as our future work.
\textcolor{black}{It is worth noting that the task offloading in a large-scale asynchronous MEC system may suffer from Byzantine failure, where
the status of server or users appears to be in failure to some users while functional to other users. Therefore, the system needs to first reach a consensus on whether the user or server has failed, then it can shut down the failure part accordingly. To handle this problem, we may resort to the asynchronous Byzantine fault tolerant protocol proposed in \cite{miller2016honey}.
Under this framework, users receive the offloading history from other users and
store them in their buffer. At the beginning of each epoch, each user selects and provides a subset of the history in its buffer to a randomized agreement protocol which is used to determine whether the target user is in failure.
We will investigate the specific implementation of this method as our future topic.}
\color{black}
Note that due to the limited computational resources, the task completion ratio cannot be further improved by our proposed method. In practical applications (e.g., the task offloading for non-critical wireless sensors such as smart meters reading \cite{7281870, 6574667}), we need other supplementary methods to tolerate high violation probability further. For example, if the task misses its current deadline, it can be stored in the buffer and assigned a new deadline for later task offloading.
\color{black}
\color{black}
\section{Conclusions}
\label{section 7}
We proposed a novel WI-based task offloading policy for a large-scale asynchronous MEC system, which features scalable calculation and simple implementation. We formulated the offloading policy design as an RMAB with the objective to maximize the total discounted reward over the time horizon. Based on the WI theory, we rigorously established the indexability and derived the WI in a closed-form expression.
To achieve a higher task completion ratio, the STLW-WI policy is proposed in the task completion ratio-oriented case.
For the case of unknown user offloading energy consumption prior to offloading, we proposed the BL-WI policy and PSBL-WI policy for the conjugate and non-conjugate prior cases, respectively.
Simulation results verified that the proposed policies significantly outperform the existing policies.
The proposed method provides a potential avenue to the highly efficient task offloading with the upcoming large-scale MEC deployment in the IoT.
\appendices
\section{Proof Of Theorem 1}
\begin{proof}
Without loss of generality, we drop the subscript $i$, $j$, $\delta$ and $t$. Denote the difference between two value functions by $h(\tau,b) = V(\tau, b+k-1) - V(\tau, b)$, and the difference of two actions (offloading and un-offloading) by $g(\tau,b)$. The indexability of the offloading problem depends on the property that $h(\tau,b)$ is piecewise linear in $\delta$ and $\frac{ \partial h(\tau,b)}{ \partial \delta} \ge -1 $, because this property guarantees that $\frac{ \partial g(\tau,b)}{ \partial \delta} = \left[1 - \frac{ \partial h(\tau,b)}{ \partial \delta} \right] \ge 0 $. The induction method is applied to prove this property. Specifically, we first show that the WI $\omega(\tau,b) $ exists for $\tau = 0, 1$, then assuming that the WI exists and $\frac{ \partial h(\tau,b)}{ \partial \delta} \ge -1 $ for $\tau = t - 1$, we show that the WI also exists and $\frac{ \partial h(\tau,b)}{ \partial \delta} \ge -1 $ holds for $\tau = t$.
\begin{enumerate}
\item $\tau=0$: There is no task waiting in the user. The Bellman equation is stated as
\begin{equation}
V(0,0) = \max\left\{\delta+\beta V_e, \beta V_e \right\},
\end{equation}
where $V_e$ is the expected reward of future tasks generation. Therefore, if and only if $\delta>0$, the first term is larger and the un-offloading action is optimal. Thus $\omega(0,0) = 0$.
\item $\tau = 1$: there are four cases.
\begin{enumerate}
\item If $b = 0$, the Bellman equation is stated as
\begin{equation}
V(1,0) = \max\left\{\delta+\beta V_e, \beta V_e \right\}.
\end{equation}
It is same as the case $\tau=0, b=0$, therefore, $\omega(1,0) = 0$.
\item If $ b = 1$, the Bellman equation is stated as
\begin{equation}
V(1,b) = \max\left\{\delta + \beta V_e, E^{\text{sav}} +\beta V_e \right\}.
\end{equation}
If and only if $\delta \ge E^{\text{sav}}$, the un-offloading action is optimal. Thus $\omega(1,b) = E^{\text{sav}}$ when $b = 1$.
\item If $ 1 < b \le k$, the Bellman equation is stated as
\begin{equation}
\begin{aligned}
V(1,b) = \max\left\{\delta +\beta V_e - F(b-1), E^{\text{sav}} + \beta V_e \right\}.
\end{aligned}
\end{equation}
If and only if $\delta \ge E^{\text{sav}} + F(b-1)$, the un-offloading action is optimal. Thus $\omega(1,b) = E^{\text{sav}} + F(b-1)$ when $1 < b \le k$.
\item If $ b > k $, the Bellman equation is stated as
\begin{equation}
\begin{aligned}
V(1,b) = \max\left\{\delta - F(b-1) + \beta V_e, \right.
\\
\left.
E^{\text{sav}} - F(b-k) + \beta V_e \right\}.
\end{aligned}
\end{equation}
If and only if $\delta \ge E^{\text{sav}} + F(b-1) -F(b-k)$, the un-offloading action is optimal. Thus
\begin{equation}
\omega(1,b) = E^{\text{sav}} + F(b-1) -F(b-k).
\end{equation}
\end{enumerate}
Thus the WI for $\tau=1$ exists, and the closed form is given by
\begin{equation}
\omega(1,b) =
\begin{cases}
0, & \mbox{if } b = 0; \\
E^{\text{sav}} & \mbox{if } b = 1; \\
E^{\text{sav}} + F(b-1), & \mbox{if } 1 < b \le k; \\
E^{\text{sav}} + F(b-1) - F(b-k) & \mbox{if } b > k; \\
\end{cases}
\label{WI when tau = 1}
\end{equation}
\end{enumerate}
Now we are ready to show the $\frac{ \partial h^{\delta}(\tau,b)}{ \partial \delta} \ge -1 $ holds when $\tau = 1$.
\begin{enumerate}
\item If $b = 0$, $h(1,0)=V(1,k-1)-V(1,0)$, we have
\begin{equation}
h(1,0) =
\begin{cases}
E^{\text{sav}}, & \mbox{if } \omega < 0; \\
E^{\text{sav}}-\delta, & \mbox{if } 0 \le \delta < \omega(1,k-1);
\\
-F(k-2), & \mbox{if } \delta \ge \omega(1,k-1).
\end{cases}
\end{equation}
\item If $b = 1$, $h(1,1)=V(1,k)-V(1,1)$, we have
\begin{equation}
h(1,1) =
\begin{cases}
0, & \mbox{if } \delta < \omega(1, 1); \\
E^{\text{sav}} - \delta, & \mbox{if } \omega(1, 1) \le \delta < \omega(1, k); \\
-F(k-1), & \mbox{if } \delta \ge \omega(1,k).
\end{cases}
\end{equation}
\item If $2 \le b \le k$, $h(1,b)=V(1,b + k - 1)-V(1,b)$, we have
\begin{equation}
h(1,b) =
\begin{cases}
-F(b-1), \\
\hspace{0.3cm} \mbox{if } \delta < \omega(1, b + k -1); \\
E^{\text{sav}} - \delta, \\
\hspace{0.3cm} \mbox{if } \omega(1,b) \le \delta < \omega(1, b + k -1);
\\
-F(b+k-2)+F(b-1), \\ \hspace{0.3cm} \mbox{if } \delta \ge \omega(1,b).
\end{cases}
\end{equation}
\item If $b > k$, $h(1,b)=V(1,b + k - 1)-V(1,b)$, we have
\begin{equation}
h(1,b) =
\begin{cases}
-F(b-1)+F(b-k),
\\
\hspace{0.3cm} \mbox{if } \delta < \omega(1,b); \\
E^{\text{sav}} - \delta, \\ \hspace{0.3cm} \mbox{if } \omega(1,b) \le \delta < \omega(1, b + k -1); \\
-F(k+b-2)+F(b-1), \\ \hspace{0.3cm} \mbox{if } \delta \ge \omega(1, b + k -1).
\end{cases}
\end{equation}
\end{enumerate}
Therefore, the derivation of $h(1,b)$ on $\delta$ always guarantees that $\frac{\partial h(1,b)}{\partial\delta} \ge -1$, which implies the indexability holds when $\tau=1$.
Then, we show the property of $h(\tau,b)$ holds when $\tau=t$ under the assumption that $\frac{\partial h(\tau,b)}{\partial \delta} \ge -1$ when $\tau = t-1$.
\begin{enumerate}
\item If $b=0$,
\begin{equation}
h(\tau,0) =
\begin{cases}
E^{\text{sav}}, \hspace{0.3cm} \mbox{if } \delta < 0; \\
E^{\text{sav}}-\delta, \\ \hspace{0.3cm} \mbox{if } 0 \le \delta < \omega(\tau,k-1); \\
\beta \left[V(\tau-1,k-2) - V(\tau-1,0) \right],
\\ \hspace{0.3cm} \mbox{if } \delta \ge \omega(\tau,k-1).
\end{cases}
\end{equation}
For the first two cases, it is clearly shown that the gradient of $h(\tau,b)$ with respect to $\delta$ is larger or equal than $-1$. For the third case, we can further expand it by comparing the value of $\delta$ and $\omega(\tau-t^{\prime}, k - t^{\prime}-1)$. Whenever there is a time step $t^{\prime}$ such that $\delta \le \omega(\tau-t^{\prime}, k - t^{\prime}-1)$, $\exists 2 \le t^{\prime} \le \tau$, $h(\tau,0) = \beta^{t^{\prime}} \left(E^{\text{sav}} - \omega \right)$ whose gradient is $-\beta^{t^{\prime}} \ge -1$. On the other hand, if there is no such time step, $h(\tau,0)$ will be calculated by its penalty term whose gradient in terms of $\omega$ is zero.
\item If $ 1 \le b \le k$, $h(\tau,b) = V(\tau,b + k -1)-V(\tau,b)$, we have
\begin{equation}
h(\tau,b) =
\begin{cases}
\beta \left[V(\tau-1, b-1) - V(\tau-1, 0)\right],
\\
\hspace{0.3cm} \mbox{if } \delta <\omega(\tau,b); \\
E^{\text{sav}}-\delta, \\ \hspace{0.3cm} \mbox{if } \omega(\tau,b) \le \delta < \omega(\tau,b + k -1); \\
\beta h\left(\tau-1, b-1\right) \\ \hspace{0.3cm} \mbox{if } \delta \ge \omega(\tau,b+ k -1).
\end{cases}
\end{equation}
For the first case, a similar analysis as the third case in $b=0$ can be carried out here and we have $\frac{\partial h(\tau,b)}{\partial \delta}\ge -1$ as well. Since we have $\frac{\partial h(\tau-1,b-1)}{\partial \delta}\ge -1$ for all $b$ by assumption, we have $\frac{\partial h(\tau,b)}{\partial \delta}\ge -1$ here as well.
\item If $ b > k$, $h(\tau,b) = V(\tau,b + k -1)-V(\tau,b)$, we have
\begin{enumerate}
\item If $\omega(\tau,b + k - 1) > \omega(\tau,b)$, we have
\begin{equation}
h(\tau,b) =
\begin{cases}
\beta h(\tau-1, b-k), \\ \hspace{0.3cm} \mbox{if } \delta<\omega(\tau,b); \\
E^{\text{sav}}-\delta, \\ \hspace{0.3cm} \mbox{if } \omega(\tau,b) \le \delta < \omega(\tau,b + k -1); \\
\beta h(\tau-1, b-1), \\ \hspace{0.3cm} \mbox{if } \delta \ge \omega(\tau,b+ k -1).
\end{cases}
\end{equation}
\item If $\omega(\tau,b + k - 1) \le \omega(\tau,b)$, we have
\begin{equation}
h(\tau,b) =
\begin{cases}
\beta h(\tau-1, b-k), \\ \hspace{0.3cm} \mbox{if } \delta < \omega(\tau,b + k - 1); \\
\delta - E^{\text{sav}} + \beta \left[h(\tau-1, b - 1) \right. \\ \left. + h(\tau-1, b - k)\right], \hspace{0.3cm}
\\ \hspace{0.3cm} \mbox{if } \omega (\tau,b+1) \le \delta < \omega(\tau,b); \\
\beta h(\tau-1, b-1), \\ \hspace{0.3cm} \mbox{if } \delta \ge \omega(\tau,b).
\end{cases}
\end{equation}
\end{enumerate}
\end{enumerate}
Since $\frac{\partial h(\tau-1,b-1)}{\partial \omega}\ge -1$ for all $b$ by assumption, we have $\frac{\partial h(\tau,b)}{\partial \omega}\ge -1$ in all cases.
Thus, the indexability of the RMAB can be established.
\end{proof}
\section{Proof Of Theorem 2}
\begin{proof}
In this section, we derived the closed form of the WI by induction. Since the case for $\tau = 0$ and $\tau=1$ have been proved during the indexability prove, we start from $\tau = 2$ from here.
\begin{enumerate}
\item $\tau = 2$: there are four cases.
\begin{enumerate}
\item If $b = 0$, the Bellman equation is stated as
\begin{equation}
V(2,0) = \max\left\{\delta+\beta V(1,0), \beta V(1,0) \right\}.
\end{equation}
It is same as the case $\tau=0, b=0$, therefore, $\omega(2,0) = 0$.
\item If $b = 1$, the Bellman equation is stated as
\begin{equation}
V(2,b) = \max\left\{\delta + \beta V(1,0), E^{\text{sav}}+\beta V(1,0) \right\}.
\end{equation}
If and only if $\delta \ge E^{\text{sav}}$, the un-offloading action is optimal. Thus $\omega(2,1) = E^{\text{sav}}$.
\item If $ 1 < b \le k$, the Bellman equation is stated as
\begin{equation}
V(2,b) = \max\left\{\delta + \beta V(1,b-1), E^{\text{sav}} +\beta V(1,0) \right\}.
\end{equation}
The difference between actions is
\begin{equation}
\begin{aligned}
g(2, b) & = \delta - E^{\text{sav}} + \beta \left[V(1,b-1) - V(1,0)\right]
\\
& =
\begin{cases}
\delta - E^{\text{sav}} + \beta E^{\text{sav}} \\ \hspace{0.3cm} \mbox{if } \delta<0; \\
\delta - E^{\text{sav}} + \beta \left(E^{\text{sav}} - \delta \right), \\ \hspace{0.3cm} \mbox{if } 0 \le \delta \le \omega(1, b-1) ; \\
\delta - E^{\text{sav}} - \beta F(b-2), \\ \hspace{0.3cm} \mbox{if } \delta > \omega(1, b-1); \\
\end{cases}
\end{aligned}
\end{equation}
The difference equals $0$ when $\delta = E^{\text{sav}} $. Thus $\omega(2,b) = E^{\text{sav}}$ when $ 1 < b \le k$.
\item If $ b = k + 1$, the Bellman equation is stated as
\begin{equation}
V(2,k + 1) = \max\left\{\delta +\beta V(1,k), E^{\text{sav}} +\beta V(1,1) \right\}.
\end{equation}
The difference between actions is
\begin{equation}
\begin{aligned}
g(2, k + 1) & = \delta - E^{\text{sav}} + \beta \left[V(1,k) - V(1,1)\right]
\\
& =
\begin{cases}
\delta - E^{\text{sav}}, \\ \hspace{0.3cm} \mbox{if } \omega< \omega(1,1); \\
\delta - E^{\text{sav}} + \beta \left(E^{\text{sav}} - \delta \right), \\ \hspace{0.3cm} \mbox{if } \omega(1,1) \le \delta \le \omega(1, k); \\
\delta - E^{\text{sav}} - \beta F(k -1), \\ \hspace{0.3cm} \mbox{if }\delta > \omega(1, k); \\
\end{cases}
\end{aligned}
\end{equation}
The difference equals $0$ when $\delta = E^{\text{sav}}$. Thus $\omega(1,b) = E^{\text{sav}} $ when $ b = k+1$.
\item If $k + 1 < b \le 2k $, the Bellman equation is stated as
\begin{equation}
\begin{aligned}
V(1,b) = \max\left\{\delta+\beta V(1,b-1), \right.
\\
\left. E^{\text{sav}} + \beta V(1,b-k) \right\}.
\end{aligned}
\end{equation}
The difference between actions is
\begin{equation}
\begin{aligned}
g(2, b) & = \delta - E^{\text{sav}} + \beta \left[V(1,b-1) - V(1,b-k)\right]
\\
& =
\begin{cases}
\delta - E^{\text{sav}} + \beta \left[- F(b-k-1)\right] \\ \hspace{0.3cm} \mbox{if } \delta <\omega(1, b-k); \\
\delta - E^{\text{sav}} + \beta \left[E^{\text{sav}} - \delta \right], \\ \hspace{0.3cm} \mbox{if } \omega(1, b - k) \le \delta \le \omega(1, b - 1) ; \\
\delta - E^{\text{sav}} + \beta \left[-F(b-2) + F(b-k-1)\right], \\ \hspace{0.3cm} \mbox{if } \delta > \omega(1, b-1); \\
\end{cases}
\end{aligned}
\end{equation}
The difference equals $0$ when $\delta = E^{\text{sav}} + \beta \left[ F(b-k-1)\right]$. Thus $\omega(2, b) = E^{\text{sav}} + \beta \left[ F(b-k-1)\right] $ when $ k + 1 < b \le 2k $.
\item If $b > 2k $, the Bellman equation is stated as
\begin{equation}
\begin{aligned}
V(1,b) = \max\left\{\delta + \beta V(1,b-1), E^{\text{sav}} +\beta V(1,b-k) \right\}.
\end{aligned}
\end{equation}
The difference between actions is
\begin{equation}
\begin{aligned}
g(2, b) & = \delta - E^{\text{sav}} + \beta \left[V(1,b-1) - V(1,b-k)\right]
\\
& =
\begin{cases}
\delta - E^{\text{sav}} + \beta \left[-F(b-k-1) + F(b-2k)\right], \\ \hspace{0.3cm} \mbox{if } \delta <\omega(1, b-k); \\
\omega - E^{\text{sav}}+ \beta \left[E^{\text{sav}}- \delta \right], \\ \hspace{0.3cm} \mbox{if } \omega(1, b-k) \le \delta < \omega(1, b - 1) ; \\
\omega - E^{\text{sav}} + \beta \left[-F(b-2) + F(b-k-1)\right], \\ \hspace{0.3cm} \mbox{if } \delta \ge \omega(1, b - 1); \\
\end{cases}
\end{aligned}
\end{equation}
The difference equals $0$ when $\delta = E^{\text{sav}} + \beta\left[F(b-k-1) - F(b-2k)\right]$. Thus $\omega(1,b) = E^{\text{sav}} + \beta\left[F(b-k-1) - F(b-2k)\right]$ when $ b > 2k $.
\end{enumerate}
Thus the WI for $\tau=2$ exists, and the closed form is calculated as
\begin{equation}
\omega(2,b) =
\begin{cases}
0, \hspace{1.5cm} \mbox{if } b = 0; \\
E^{\text{sav}}, \hspace{1.1cm} \mbox{if } 1 \le b \le k + 1; \\
E^{\text{sav}} + \beta \left[ F(b-k-1)\right], \\ \hspace{0.3cm} \mbox{if } k + 1 < b \le 2 k; \\
E^{\text{sav}} + \beta\left[F(b-k-1) - F(b-2k)\right] \\ \hspace{0.3cm} \mbox{if } b > 2 k; \\
\end{cases}
\end{equation}
\end{enumerate}
Next we show the closed-form of the WI for the case of $\tau \ge 3$, assuming (1) holds for $\tau-1$.
\begin{enumerate}
\item If $b = 0$, the Bellman equation is stated as
\begin{equation}
V(\tau,0) = \max\left\{\delta + \beta V(\tau-1,0), \beta V(\tau-1,0) \right\}.
\end{equation}
Therefore, $\omega(\tau,0) = 0$.
\item If $b= 1$, the Bellman equation is stated as
\begin{equation}
\begin{aligned}
V(\tau,1) = \max\left\{\delta + \beta V\left(\tau-1,0\right), E^{\text{sav}}+\beta V\left(\tau-1, 0 \right) \right\}.
\end{aligned}
\end{equation}
Thus $\omega(\tau,1) = E^{\text{sav}}$.
\item If $ 2 \le b \le k$, the Bellman equation is stated as
\begin{equation}
\begin{aligned}
V(\tau,b) = \max\left\{\omega + \beta V\left(\tau-1,b-1\right),
\right. \\
\left.
E^{\text{sav}}+\beta V\left(\tau - 1, 0 \right) \right\}.
\end{aligned}
\end{equation}
The difference between actions is:
\begin{equation}
\begin{aligned}
g(\tau,b) & = \delta - E^{\text{sav}} + \beta \left[V(\tau-1, b-1) - V(\tau-1, 0)\right]
\\
& =
\begin{cases}
\delta - E^{\text{sav}} + \beta^2 E^{\text{sav}}, \\ \hspace{0.3cm} \mbox{if } \delta < 0; \\
\delta - E^{\text{sav}} + \beta \left[E^{\text{sav}} - \delta \right], \\ \hspace{0.3cm} \mbox{if } 0 \le \delta \le \omega(\tau-1, b - 1); \\
\omega - E^{\text{sav}} + \beta^2 \left[V(\tau-2, b-2) - V(\tau-2, 0)\right] \\ \hspace{0.3cm} \mbox{if } \delta > \omega(\tau-1, b - 1); \\
\end{cases}
\end{aligned}
\end{equation}
The difference equals $0$ when $\delta = E^{\text{sav}}$. Thus $\omega(\tau,b) = E^{\text{sav}}$ when $ 2 \le b \le k$.
\item If $ k < b \le k (\tau-2) + 2$, the Bellman equation is stated as
\begin{equation}
\begin{aligned}
V(\tau, b) = \max\left\{\delta+\beta V\left(\tau-1,b-1\right), E^{\text{sav}} + \beta V\left(\tau-1,b-k\right) \right\}.
\end{aligned}
\end{equation}
The difference between actions is
\begin{equation}
\begin{aligned}
g(\tau,b) & = \delta - E^{\text{sav}} + \beta \left[V\left(\tau-1,b-1\right)- V\left(\tau-1,b-k\right)\right]
\\
& =
\begin{cases}
\delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-2,b-1-k\right)- \right. \\ \left. V\left(\tau-1,(b-2 k)^{+}\right)\right], \hspace{0.5cm} \mbox{if } \delta < E^{\text{sav}}; \\
\delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-2,b-2\right)- \right. \\ \left. V\left(\tau-2,b- k -1\right)\right], \hspace{0.5cm} \mbox{if } \delta \ge E^{\text{sav}}; \\
\end{cases}
\end{aligned}
\end{equation}
Since we have $\omega(\tau-1, b - k -1) = E^{\text{sav}}$ when $1 \le b - k -1 \le k(\tau-2) + 1$
So $\omega(\tau,b) = E^{\text{sav}}$ when $b= k+1$.
\item If $ k (\tau-2) + 2 < b \le k (\tau-1) + 1$, the Bellman equation is stated as
\begin{equation}
\begin{aligned}
V(\tau, b) = \max\left\{\delta+\beta V\left(\tau-1,b-1\right), \right. \\ \left. E^{\text{sav}}+\beta V\left(\tau-1,b-k\right) \right\}.
\end{aligned}
\end{equation}
Therefore, the difference between actions is
\begin{equation}
\begin{aligned}
g(\tau,b) & = \delta - E^{\text{sav}} + \beta \left[V\left(\tau-1,b-1\right)- V\left(\tau-1,b-k\right)\right]
\\
& =
\begin{cases}
\delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-2,b-1-k\right)- \right. \\ \left. V\left(\tau-1,(b-2 k)\right)\right], \hspace{0.3cm} \mbox{if } \delta < \omega(\tau-1, b - k);
\\
\delta - E^{\text{sav}} + \beta \left[\delta - E^{\text{sav}}\right] \\ \hspace{0.3cm} \mbox{if } \omega(\tau-1, b - k) \le \delta < \omega(\tau-1, b-1);
\\
\delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-2,b-2\right)- \right. \\ \left. V\left(\tau-2,b- k -1\right)\right], \hspace{0.3cm} \mbox{if } \delta \ge \omega(\tau-1, b-1); \\
\end{cases}
\end{aligned}
\end{equation}
It equals $0$ when $\delta = E^{\text{sav}}$. Thus $\omega (\tau,b) = E^{\text{sav}}$ when $ k (\tau-2) + 2 < b \le k (\tau-1) + 1$.
\item If $ k(\tau-1)+2 \le b \le k \tau$, the Bellman equation is stated as
\begin{equation}
\begin{aligned}
V(\tau, b) = \max\left\{\delta +\beta V\left(\tau-1,b-1\right), \right. \\ \left. E^{\text{sav}}+\beta V\left(\tau-1,b-k\right) \right\}.
\end{aligned}
\end{equation}
Therefore, the difference between actions is
\begin{equation}
\begin{aligned}
g(\tau,b) & = \delta - E^{\text{sav}} + \beta \left[V\left(\tau-1,b-1\right)- V\left(\tau-1,b-k\right)\right]
\\
& =
\begin{cases}
\delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-2,b-1-k\right)- \right. \\ \left. V\left(\tau-2,b-2 k\right)\right], \hspace{0.3cm} \mbox{if } \omega < \omega(\tau-1, b-k);
\\
\delta - E^{\text{sav}} + \beta \left[\delta - E^{\text{sav}}\right] \\ \hspace{0.3cm} \mbox{if } \omega(\tau-1, b - k) \le \delta < \omega(\tau-1, b-1);
\\
\delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-2,b-2\right)- \right. \\ \left. V\left(\tau-2,b- k -1\right)\right], \hspace{0.3cm} \mbox{if } \delta \ge \omega(\tau-1, b-1); \\
\end{cases}
\end{aligned}
\label{18}
\end{equation}
In the first case since $\delta < \omega(\tau-1, b-k)$, according to the equation
\begin{equation}
\begin{aligned}
\delta & < \omega(\tau-1 -\tau^{\prime}, b-k-k\tau^{\prime}) \\
& < \omega(\tau-1-\tau^{\prime}, b-1-k\tau^{\prime}),
\end{aligned}
\end{equation}
the difference can be further written as
\begin{equation}
\begin{aligned}
& \delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-2,b-1-k\right)- V\left(\tau-2,b-2 k\right)\right] \\
& = \delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-3,b-1-2k\right)- V\left(\tau-3,b-3 k\right)\right] \\
& = \cdots \\
& = \delta - E^{\text{sav}} - \beta^{\tau-1} \left[F(b - (\tau-1)k -1)\right].
\end{aligned}
\end{equation}
Therefore, when $\delta = E^{\text{sav}} + \beta^{\tau-1} \left[F(b-(\tau-1)k-1)\right]$, the first case in \eqref{18} equals $0$. Accordingly, when $k(\tau-1)+2 \le b \le k \tau$, the WI is calculated as:
\begin{equation}
\omega(\tau, b) = E^{\text{sav}} + \beta^{\tau-1} \left[F(b-(\tau-1)k-1)\right]
\end{equation}
\item If $ b \ge k\tau + 1$,
\begin{equation}
\begin{aligned}
V(\tau, b) = \max\left\{\delta+\beta V\left(\tau-1,b-1\right), \right. \\ \left. E^{\text{sav}}+\beta V\left(\tau-1,b-k\right) \right\}.
\end{aligned}
\end{equation}
Therefore, the difference between actions is
\begin{equation}
\begin{aligned}
g(\tau,b) & = \delta - E^{\text{sav}} + \beta \left[V\left(\tau-1,b-1\right)- V\left(\tau-1,b-k\right)\right]
\\
& =
\begin{cases}
\delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-2,b-1-k\right)- \right. \\ \left. V\left(\tau-2,b-2 k\right)\right], \hspace{0.3cm} \mbox{if } \delta < \omega(\tau-1, b-k);
\\
\delta - E^{\text{sav}} + \beta \left[\delta - E^{\text{sav}}\right] \\ \hspace{0.3cm} \mbox{if } \omega(\tau-1, b-k) \le \delta < \omega(\tau-1, b-1);
\\
\delta - E^{\text{sav}} + \beta^2 \left[V\left(\tau-2,b-2\right)- \right. \\ \left. V\left(\tau-2,b- k -1\right)\right], \hspace{0.3cm} \mbox{if } \delta \ge \omega(\tau-1, b-1); \\
\end{cases}
\end{aligned}
\label{19}
\end{equation}
Similar with the previous case, the difference equals $0$ when $\delta = E^{\text{sav}} + \beta^{\tau-1} \left[F(b-(\tau-1)k+1) + F(b - k\tau)\right].$
Accordingly, when $b \ge k \tau + 1$, the WI is calculated as:
\begin{equation}
\omega(\tau, b) = E^{\text{sav}} + \beta^{\tau-1} \left[F(b-(\tau-1)k-1) + F(b - k\tau)\right]
\end{equation}
\end{enumerate}
Therefore, the closed-form expression for the WI (17) holds.
\end{proof}
\section{Proof of Theorem 3}
To prove Theorem 3, for any given offloading scheduling policy that violates the STLW rule, we construct an updated policy that meets the STLW rule. Then we need to show that this updated policy can increase the reward, compared with the original one.
Assume that the $i$-th user has priority over the $j$-th user based on the STLW rule at the $t^{\prime}$-th time slot with the system state ${\bf{S}}_t^{\prime}$. Let $\Gamma \triangleq \max \left\{\tau_{i,t}, \tau_{j,t}\right\} - 1$, assume we have a policy $\mathcal{G} = \left\{{\bf{u}}_{t^{\prime}}, {\bf{u}}_{t^{\prime}+1}, \ldots, {\bf{u}}_{t^{\prime} + \Gamma} \right\}$ violates the STLW rule and selects the $j$-th user instead of the $i$-th user at the $t^{\prime}$-th time slot. Then we construct an updated policy $\tilde{\mathcal{G}} = \left\{{\bf{\tilde{u}}}_{t^{\prime}}, {\bf{\tilde{u}}}_{{t^{\prime}}+1}, \ldots, {\bf{\tilde{u}}}_{t^{\prime} + \Gamma} \right\}$ as follows.
\begin{enumerate}
\item At the ${t^{\prime}}$-th time slot, $\tilde{\mathcal{G}}$ selects the $i$-th user instead of $j$. That is, ${\bf{\tilde{u}}}_{t^{\prime}}$ is same as ${\bf{u}}_{t^{\prime}}$ except that its $i$-th component is $1$ and the $j$-th component is $0$.
\item
Denote the set of time slots that the policy $\mathcal{G}$ selects the $i$-th user instead of the $j$-th user after the $t^{\prime}$-th time slot by ${\Pi}\left(t\right) \subseteq \left\{{t^{\prime}}+1, \ldots, \min \left\{d_i, d_j\right\}-1\right\}$.
\begin{itemize}
\item If the set ${\Pi}\left(t\right)$ is empty, let $\tilde{\mathcal{G}}$ take the same actions as the $\mathcal{G}$ in the following time slot, that is, ${\bf{\tilde{u}}}_q = {\bf{u}}_q$, for $k = t^{\prime}+1, \ldots, t^{\prime}+ \Gamma$.
\item If the set ${\Pi}\left(t\right)$ is not empty, denote the minimal time slot of the set by $t_{\min}$. For the time slots $q = t^{\prime}+1, \ldots, t_{\min}-1$, let ${\bf{\tilde{u}}}_q = {\bf{u}}_q$. However, at the $t_{\min}$-th time slot, the new policy $\mathcal{\tilde{G}}$ selects the $j$-th user instead of $i$. That is, ${\bf{\tilde{u}}}_{t^{\prime}}$ is the same as ${\bf{u}}_{t^{\prime}}$ except that the $j$-th component is $1$ and the $i$-th component is $0$.
\end{itemize}
\label{Definition: STLW Interchaning Policy}
\end{enumerate}
Take a closer look between policies $\mathcal{G}$ and $\mathcal{\tilde{G}}$, after the $t^{\prime}$-th time slot, we can always find a sequence $\left\{{\bf{\bar{S}}}_k \right\}_{q=t^{\prime}+1}^{t^{\prime}+\Gamma}$ in $\mathcal{\tilde{G}}$ as a comparison to the sequence $\left\{{\bf{S}}_k \right\}_{q=t^{\prime}+1}^{t^{\prime}+\Gamma}$ in $\mathcal{G}$, which satisfies the following condition:
\begin{itemize}
\item If the set $\Pi(t)$ is empty, then we have $\left\{{\bf{\bar{S}}}_q \right\}_{q=t^{\prime}+1}^{t^{\prime}+\Gamma} = \left\{{\bf{S}}_q \right\}_{q=t^{\prime}+1}^{t^{\prime}+\Gamma}$.
\item Otherwise, we have
\begin{equation}
\left\{{\bf{\bar{S}}}_q \right\}_{q=t^{\prime}+1}^{t^{\prime}+\Gamma} = \left\{{\bf{\bar{S}}}_{t^{\prime} + 1}, \ldots, {\bf{\bar{S}}}_{t_{\min}}, {\bf{{S}}}_{t_{\min}+1}, \ldots, {\bf{{S}}}_{t_{\prime} + \Gamma} \right\},
\end{equation}
\end{itemize}
where $t_{\min}$ is the minimal time slot in the set $\Pi(t)$.
Since two policies $\mathcal{G}$ and $\tilde{\mathcal{G}}$ always select an equal number of users to perform offloading and will be identical after the $\left(t^{\prime} + \Gamma\right)$-th time slot. Therefore, to arrive at the result in Theorem 3, we only need to show that
\begin{equation}
\begin{aligned}
R\left( {{\bf{S}}_t}, {\bf{u}}_t\right) & + \sum_{q = t^{\prime}+1}^{t^{\prime}+\Gamma} R\left( {\bf{S}}_q, {\bf{u}}_q\right)
\\ & \le R\left( {{\bf{S}}_t}, {\bf{\tilde{u}}}_t\right) + \sum_{q = t^{\prime}+1}^{t^{\prime}+\Gamma} R\left( {\bf{\tilde{S}}}_q, {\bf{\tilde{u}}}_q\right).
\end{aligned}
\label{eq: modified SLSW difference}
\end{equation}
To verify \eqref{eq: modified SLSW difference}, we consider the following two cases.
\begin{enumerate}
\item When the set ${\Pi}\left(t\right)$ is not empty, for every pair of system state sequence, $\left\{{\bf{S}}_q \right\}_{q=t^{\prime}}^{t^{\prime} + \Gamma}$ and $\left\{{\bf{\bar{S}}}_q \right\}_{q=t^{\prime}}^{t^{\prime} + \Gamma}$, both policies will result in the same result, i.e., the equality holds in \eqref{eq: modified SLSW difference}.
\item When the set ${\Pi}\left(t\right)$ is empty. Whenever the policy $\mathcal{G}$ selects the $i$-th user, it must also select the $j$-th user, for $q = t^{\prime}+1, \ldots, \min \left\{d_i, d_j\right\}-1$.
Denote the remaining workload of the $i$-th user after its deadline by $\delta_i$ under the policy $\mathcal{G}$. Similarly, $\tilde{\delta}_i$ is the remaining workload under the policy $\mathcal{\tilde{G}}$. Since the $i$-th user has priority over the $j$-th user at system state ${\bf{S}}_{t^{\prime}}$, it implies that $\tilde{\delta}_i = \delta_i -1$ and $\tilde{\delta}_j = \delta_j + 1$ (according to the Definition \ref{Definition: STLW Interchaning Policy}). Therefore, we have the reward difference under two policies calculated as
\begin{equation}
\begin{aligned}
& V_{\tilde{\mathcal{G}}}^{t^{\prime}+\Gamma} - V_{{\mathcal{G}}}^{t^{\prime}+\Gamma}
\\
& = \alpha \left\{- F(\tilde{\delta}_i) - F(\tilde{\delta}_j) - \left[- F(\delta_j) - F(\delta_i)\right] \right\}
\\
& = \alpha \left\{ \left[{\delta}_i^2 - \tilde{\delta}_i^2 + {\delta}_j^2 - \tilde{\delta}_j^2 \right] \right\}
\\
& = 2 \alpha \left(\delta_j - \delta_i\right) - 2
\end{aligned}
\label{eq: reward difference}
\end{equation}
Note that according to the STLW rule, we have $0 \le \delta_i < \delta_j$, where both $\delta_i$ and $\delta_j$ are integers. Therefore, when $\alpha \ge 1$ (i.e. focus on task completion), this reward difference is always no less than 0, which implies that the constructed policy $\mathcal{\tilde{G}}$ can achieve more rewards than the original policy $\mathcal{G}$.
\end{enumerate}
| 46fb92bf60432c0ab5e74759bb42a49f625d35f8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Neural Machine Translation (NMT) \cite{Bahdanau2015} has achieved state of the art in various MT systems, including rich and low resource language pairs \cite{Edunov2018, Gu2019, ngo2019}.
However, the quality of low-resource MT is quite unpretentious due to the lack of parallel data while it has achieved better results on systems of the available resource. Therefore, low-resource MT is one of the essential tasks investigated by many previous works \cite{ Ha2016, Lee2016, senrich2019}.
Recently, some works present MT systems that have achieved remarkable results for low-resource language \cite{Gu2019, Roee2020}. Inspired by these works, we collect data from the TED Talks domain, then attempt to build multilingual MT systems from French, English-Vietnamese. Experiments demonstrate that both language pairs: French-Vietnamese and English-Vietnamese have achieved significant performance when joining the training.
Although multilingual MT can reduce the sparse data in the shared space by using word segmentation, however, rare words still exist, evenly they are increased more if languages have a significant disparity in term vocabulary. Previous works suggested some strategies to reduce rare words such as using translation units at sub-word and character levels or generating a universal representation at the word and sentence levels \cite{ Lee2016, Gu2019}. These help to downgrade the dissimilarity of tokens shared from various languages. However, these works require learning additional parameters in training, thus increasing the size of models.
Our paper presents two methods to augment the translation of rare words in the source space without modifying the architecture and model size of MT systems: (1) exploiting word similarity. This technique has been mentioned by previous works \cite{luong2015, Li2016, trieu, ngo2019}. They employ monolingual data or require supervised resources like a bilingual dictionary or WordNet, while we leverage relation from the multilingual space of MT systems. (2) Adding a scalar value to the rare word embedding in order to facilitate its translation in the training process.
Due to the fact that NMT tends to have bias in translating frequent words, so rare words (which have low frequency) often have less opportunity to be considered. Our ideal is inspired by the works of \cite{Toan2017, ngo2019, Gu2019}. \cite{Toan2017} and \cite{ngo2019} proposed various solutions to urge for translation of rare words, including modification embedding in training. They only experimented with recurrent neural networks (RNNs) while our work uses the state-of-the-art transformer architecture. \cite{Gu2019} transforms the word embedding of a token into the universal space, and they learn plus parameters while our method does not. We apply our strategies in our fine-tuning processes, and we show substantial improvements of the systems after some epochs only.
Monolingual data are widely used in NMT to augment data for low-resource NMT systems \cite{Sennrich2015, zhang2016,lample2018unsupervised, wu2019, siddhant2020leveraging}. Back-translation \cite{Sennrich2015} is known as the most popular technique in exploiting target-side monolingual data to enhance the translation systems while the self-learning method \cite{zhang2016} focuses on utilizing source-side monolingual data. Otherwise, the dual-learning strategy \cite{wu2019} also suggests using both source- and target-side monolingual data to tackle this problem. Our work investigates the self-learning method \cite{zhang2016} on the low-resource multilingual NMT systems specifically related to Vietnamese. Besides, monolingual data are also leveraged in unsupervised\cite{lample2018unsupervised} or zero-shot translation\cite{lample2018unsupervised}.
The main contributions of our work are:
\vspace*{-0.2cm}
\begin{itemize}
\setlength{\itemsep}{0pt}
\item We first attempt to build a multilingual system for two low-resource language pairs: French-Vietnamese and English-Vietnamese.
\item We propose two simple techniques to encourage the translation of rare words in multilingual MT to upgrade the systems.
\item We investigate the quality translation of the low-resource multilingual NMT systems when they are reinforced synthetic data.
\item We release more datasets extracted from the TED Talks domain for the research purpose: French-Vietnamese and English-Vietnamese.
\end{itemize}
In section 2, we review the transformer architecture used for our experiments. The brief of multilingual translation is shown in section 3. Section 4 presents our methods to deal with rare words in multilingual translation scenarios. The exploitation of monolingual data for low-resource multilingual MT is discussed in section 5. Our results are described in section 6, and related work is shown in section 7. Finally, the paper ends with conclusions and future work.
\section{Transformer-based NMT}
Transformer architecture for machine translation is mentioned for the first time by \cite{Vaswani2017}. This is based on the sequence to sequence framework \cite{Sutskever2014} which includes an encoder to transform information of the source sentence $X=(x_1, x_2,...,x_n)$ into continuous representation and a decoder to generate the target sentence $Y=(y_1,y_2,...,y_m)$.
Self-attention is an important mechanism in the transformer architecture. It enables the ability to specify the relevance of a word with the remaining words in the sentence through the equation:
\begin{equation}
\begin{aligned}
\text{Self-Attn}(\vect{Q},\vect{K},\vect{V}) =\text {Softmax} (\displaystyle \frac{\vect{Q}\vect{K}^T}{d}) \vect{V}
\end{aligned}
\label{eq:att}
\end{equation}
where \textit{K} (key),\textit{ Q} (query), \textit{V }(value) are the representations of the input sentence and \textit{d} is the size of the input.
The attention mechanism \cite{Luong2015a} bridges between the source sentence in the encoder and the target sentence in the decoder. Furthermore, the feed-forward networks are used to normalize the outputs on both encoder and decoder.
The MT system is trained to minimize the maximum likelihood of \textit{K} parallel pairs:
\begin{equation}
\begin{aligned}
\mathcal{L} (\theta)= \frac{1}{K} \sum_{k=1}^{k=K} logp(Y^k|X^k; \theta)
\end{aligned}
\label{lagra}
\end{equation}
\section{Multilingual NMT}
Multilingual NMT systems can translate between many language pairs, even in the zero-shot issue. Previous works investigate multilingual translation in many fashions: (1) Many to many \cite{Ha2016, Roee2020}: from many sources to many target languages; (2) Many to one \cite{Gu2019}: from many source languages to a target language; (3) One to many \cite{Wang2018}: from one source language to many target languages. In cases (1) and (3), an artificial token is often added to the beginning of the source sentence to specify the predicted target language. Our MT systems are the same as the case (2), so we do not add any artificial token to the texts.
In a multilingual NMT system from many to one with $M$ language pairs and $K$ sentence pairs for each one, the objective function uses maximum likelihood estimation on the whole parallel pairs $\left\{ X^{(m,k)}, Y^{(m,k)}\right\} _{k=1..K}^{m=1...M} $ as:
\begin{equation}
\begin{aligned}
\mathcal{L} (\theta)= \frac{1}{K} \sum_{m=1}^{m=M} \sum_{k=1}^{k=K} logp(Y^{(m,k)}|X^{(m,k)}; \theta)
\end{aligned}
\label{lagra}
\end{equation}
where $K=\sum_{m=1}^{m=M} K_m$ is the total number of sentences of the whole corpus.
The vocabulary of the source side is mixed from all source languages: $V=\sum_{m=1}^{m=M} V_m$.
\cite{Gu2019} has shown that if the languages shared the same alphabet and had many similar words, such system will get many advantages from multilingual MT. In fact, different words from many languages can differ in form, but they may share the same subwords. This significantly reduces the number of rare words in the MT systems. Nevertheless, the rare word issue is still a challenge in NMT. We choose English and French are source languages in our experiment with the hope that they can share many tokens even though we do not have much data of those translation directions.
\section{Augmenting Rare Word Translation}
\label{methods}
\subsection{Learning multilingual word similarity}
\label{ws}
We assume that a rare word or rare token (which has a low frequency in the training data) from one source language may be similar to another word in a shared multilingual space. Similar words can belong to several languages and they can be replaced by the others.
Our method replaces rare tokens with their similar tokens in shared space. The replacements are learned dynamically in the training NMT system. To avoid slowing down the training speed, we only compute similar tokens after each epoch. In the experiments, we attempt to replace rare tokens from French with similar tokens in English and French.
Our method is described as follows:
Firstly, we extract the lists of all tokens from the English - $\{A\}$ corpus, and the most \textit{k} common words from the vocabulary of the source side of the French - $\{B\}$. We set \textit{k=15} thousand words in the experiments.
Secondly, we compute the similarity score between the embedding of a rare token $ t_i $, $ \forall t_i \notin \{ A \cup B \} $ and each embedding of the tokens $ t_j $, $ \forall t_j \in \{ A \cup B \} $ as follows:
\begin{equation}
\begin{aligned}
score_i= min(d_j(\vect{e_i},\vect{e_j}) \cdot e^{\cos(\vect{e_i}, \vect{e_j})})
\end{aligned}
\label{score}
\end{equation}
where $j=1..M$ with $M$ is the number of tokens of $ {A \cup B} $; \textit{d} is the Euclidean distance between embedding $e_i$ of token $t_i$ and embedding $e_j$ of token $t_j$.
The last, the token $ t_i $ is replaced by its similar tokens. The scores are computed iteratively after each epoch during the training process. It may have more tokens similar to a rare token, so we experimentalize in the case of random selection a token from the similar tokens. To accrete the effectiveness of the method, we use a threshold to neglect similar pairs that have scores close to 0 or too large. In the experiments, we choose the scores in $[2.4, 2.72]$ to warrant similar pairs alike in terms of distance as well as direction.
\subsection{Updating source embedding}
\label{se}
In this approach, we assume that the embedding $ e_i $ of token $t_i$, $ \forall t_i \notin \{ A \cup B \} $ is represented by the approximate embedding vector as following:
\begin{equation}
\begin{aligned}
\vect{e_i}= \vect{e_i} + d \\
\end{aligned}
\label{score}
\end{equation}
where $d$ is the difference between embedding $ e_i $ and the average of the all embeddings $ e_j $ of token $t_j$, $ \forall t_j \in \{ A \cup B \} $:
\begin{equation}
\begin{aligned}
d = \vect{e_i} - \frac{\sum_{j=1}^{j=M} \vect{e_j}}{M}
\end{aligned}
\label{score}
\end{equation}
where M is the number of tokens of $ \{A \cup B \} $.
These embeddings are then updated during the training. The average of embeddings is only estimated after each epoch to avoid slowing down the training speed. We observe the improvements in both language pairs in the experiments.
\section{Exploiting monolingual data for low-resource multilingual NMT}
\label{monolingual}
Similar to the idea suggested in \cite{zhang2016}, we leverage monolingual data from the source-side to generate synthetic bilingual data. Instead of using monolingual data from all source languages, we only attempt to exploit monolingual data of English.
Firstly, we train the multilingual NMT system from English, French $ \rightarrow $ Vietnamese based on bilingual data from the TED talks with the approaches mentioned in section \ref{methods}. The best system is then used to translate English to Vietnamese.
Lastly, the synthetic parallel data are mixed with original bilingual data in the normal training scheme.
\section{Experiments}
\subsection{Datasets}
We extracted data from TED Talks domain\footnote{\url{https://www.ted.com/}} for two language pairs English-Vietnamese and French-Vietnamese. The details of those datasets are described in Table \ref{tab1}. For the English-Vietnamese, we used standard datasets like {\tt tst2012} and {\tt tst2013} from \cite{cettolo2016iwslt} as dev and test sets for validation and evaluation. For French-Vietnamese, we separate a subset from collected data for the same purposes.
\vspace{0.4cm}
\begin{table} [h]
\vspace*{-0.1cm}
\centerline{
\begin{tabular}{|c|c|c|c|}
\hline
Datasets & Training & dev & test \\
\hline
English-Vietnamese & 231K & 1553 & 1268 \\
\hline
French-Vietnamese & 203K & 1007 & 1049 \\
\hline
\end{tabular}}
\caption{\label{tab1} {The bilingual datasets in our experiments}}
\vspace*{-0.4cm}
\end{table}
\vspace{0.2cm}
To generate synthetic bilingual data, we sampled 1.2 millions English monolingual sentences from the European Parliament English-French corpus\footnote{\url{https://www.statmt.org/europarl}}. After inferring from the multilingual MT system, we obtained two sets of pseudo bilingual data: English - Vietnamese, French - Vietnamese.
\subsection{Preprocessing}
English and French texts were tokenized and true-cased using Moses's scripts, and then they are applied to Sennrich's BPE \cite{Sennrich2016}. 30000 operators are learned to generate BPE codes for both languages.
For Vietnamese texts, we only did tokenization and true-casing using Moses's scripts.
We extracted a list of all tokens in English (A) and another list of the 15K most frequency of tokens in French (B). All lists were then used for the mentioned strategies in section~\ref{methods}.
\subsection{Systems and Training}
We implement our NMT systems using the framework {\tt NMTGMinor}\footnote{\url{https://github.com/quanpn90/NMTGMinor}}. The same settings are used for all experiments. The system includes 4 layers for both encoder and decoder, and the embedding size is 512. For the systems that adapted monolingual data, we use 6 layers. Adam optimizer is set with the initial learning rate at 1.0 for baseline and the multilingual systems and 0.5 for the fine-tuned systems. The size of a mini-batch is 128, and the vocabulary size is set to be the top 50K most frequent tokens. Training and development sets of both language pairs are concatenated prior to the training of our multilingual systems.
We modified this framework to apply our ideals proposed in section \ref{methods}. To speed up the training, we compute the similarity scores and find out similar tokens for rare tokens or the mean of all tokens in $ \{ A \cup B \} $ after each epoch. We replace rare tokens or update their embeddings in each batch. We do not use these techniques for the decoding process, so the system's performance is not affected.
The baseline and multilingual systems are trained for 70 epochs. Our methods are then used to fine-tune the systems for 15 epochs. We choose the five best models to decode the test sets independently for residual systems despite the baseline systems. The beam size is 10, and we try different values of \textit{alpha}: $0.2, 0.4, 0.8, 1.0$. Other settings are the default settings of {\tt NMTGMinor}.
\subsection{Results}
\begin{center}
\begin{table*}[t]
\vspace*{-0.1cm}
{\small
\hfill{}
\begin{tabular}{|c|l|c|c|}
\hline
\textbf{Datasets} & \textbf{Systems} & \textbf{ dev} & \textbf{test} \\
\hline
\multirow{5}{*}{English $\rightarrow$ Vietnamese} &
Bilingual Baseline & 31.74 & 35.13 \\
\cline{2-4}
& Multilingual & 31.66 (-0.08) & \textbf{36.18 (+1.05)}\\
\cline{2-4}
& Multilingual + fine-tuning & \textbf{31.88 (+0.14)} & \textbf{36.56 (+1.43)} \\
\cline{2-4}
& Multilingual + fine-tuning with similarity & \textbf{31.93 (+0.19)} & \textbf{36.75 (+1.62)} \\
\cline{2-4}
& Multilingual + fine-tuning with updated embedding & \textbf{32.11 (+0.37)} & \textbf{36.74 (+1.61)}\\
\cline{2-4}
& Multilingual + mixing pseudo bilingual data & 30.86 (-0.88) & 35.09 (-0.04) \\
\hline
\hline
\multirow{5}{*} { French $\rightarrow$ Vietnamese} & Bilingual Baseline & 23.07 & 23.03 \\
\cline{2-4}
& Multilingual & \textbf{24.49 (+1.42)} & \textbf{24.22 (+1.19)} \\
\cline{2-4}
& Multilingual + fine-tuning & \textbf{24.51 (+1.44)} & \textbf{24.86 (+1.83)} \\
\cline{2-4}
& Multilingual + fine-tuning with similarity & \textbf{24.37 (+1.30)} & \textbf{24.70 (+1.63)}
\\
\cline{2-4}
& Multilingual + fine-tuning with updated embedding & \textbf{24.60 (+1.53)} & \textbf{24.96 (+1.93)}
\\
\cline{2-4}
& Multilingual + mixing pseudo bilingual data & \textbf{25.59 (+2.52)} & \textbf{25.57 (+2.54)} \\
\cline{2-4}
& Pseudo bilingual data translation & 19.00 & 18.71 \\
\hline
\end{tabular}}
\hfill{}
\caption{\label{tab2} {The results of our MT systems are measured in BLEU. We evaluate the best model for the baseline systems and the average scores on the five best models for the multilingual and pseudo systems.}}
\end{table*}
\end{center}
We evaluate the quality of systems on two translation tasks: French to Vietnamese and English to Vietnamese, using on different approaches mentioned in previous sections. The {\tt multi-BLEU} from Moses's scripts\footnote{\url{https://github.com/moses-smt/mosesdecoder/tree/master/scripts}} is used. The results have shown in the Table~\ref{tab2}.
\textbf{(1) Bilingual baseline systems.} We train the systems based on separate bilingual data of each language pair for 70 epochs. The best model is used to decode the test data for comparison purposes in our experiments.
\textbf{(2) Multilingual systems.} We concatenate training and development sets in order to construct the new sets: French, English $\rightarrow$ Vietnamese, and then train the system using those data for the same number of epochs as for the baseline systems. We observe an improvement of +1.05 BLEU points on English $\rightarrow$ Vietnamese translation task and another one of +1.19 BLEU points on French $\rightarrow$ Vietnamese translation task compared to the baseline systems.
\textbf{(3) Multilingual fine-tuning systems.} The multilingual system is fine-tuned from the baseline for further 15 epochs with an initial learning rate of 0.05. We see the improvements of +1.43 and +1.83 BLEU points on both translation tasks, respectively.
\textbf{(4) Multilingual fine-tuning with similarity systems.} The systems from (2) are fine-tuned with the strategy mentioned in section \ref{ws} using the modified framework. We obtained a bigger gain of +1.62 BLEU points on the English $\rightarrow$ Vietnamese translation task whilst the French $\rightarrow$ Vietnamese translation task has achieved a lower improvement than the systems in (3). We show that the English $\rightarrow$ Vietnamese translation task has more advantages when rare tokens from French are replaced by similar tokens in the multilingual space. In the future, we would attempt the inverse replacement.
\textbf{(5) Multilingual fine-tuning with updated embedding systems.} We use the modified framework to fine-tune the systems in (2) with the method mentioned in section \ref{se}. The greater improvements can be found at +1.61 and +1.93 on both translation tasks compared to the systems which do not use our methods.
\textbf{(6) Multilingual with mixing of pseudo bilingual data.}
We use 400K synthetic bilingual sentence pairs for each of the language pairs: English-Vietnamese and French-Vietnamese. We train the multilingual NMT system on a mix of pseudo and real bilingual data mentioned in section \ref{monolingual} for 50 epochs. And then it is fine-tuned on the actual parallel data for 20 epochs. We observed a bigger improvement of \textbf{+2.54} BLEU points on the French $\rightarrow$ Vietnamese system while the English $\rightarrow$ Vietnamese system has achieved less improvement compared to previous systems. We speculate that the English $\rightarrow$ Vietnamese translation task may be affected by the French $\rightarrow$ Vietnamese pseudo bilingual data. In future work, we would leverage the data selection methods in order to equip better synthetic data for our systems.
\textbf{(7) Pseudo bilingual data translation.}
We train the French $\rightarrow$ Vietnamese NMT system relied on only 1.2 thousands pseudo bilingual data mentioned in section \ref{monolingual} for 26 epochs. We achieve 18.71 BLEU points on the averaged model from our five best models. Thus, we can generate synthetic parallel data for a low-resource language pair from another language pair with a bigger bilingual resource.
\section{Related Work}
Due to the unavailability of the parallel data for low-resource language pairs or zero-shot translation, previous works focus on the task to have more data such as leveraging multilingual translation \cite{Ha2016, ha2017effective, Wang2018, Gu2019, Roee2020} or using monolingual data with back-translation, self-learning \cite{Sennrich2015,zhang2016, wu2019} or mix-source \cite{Ha2016} technique.
For leveraging multilingual translation, \cite{Ha2016} added language code and target forcing in order to learn the shared representations of the source words and specify the target words. \cite{Wang2018} demonstrated a one-to-many multilingual MT with three different strategies which modify their architecture. \cite{Gu2019} built many-to-one multilingual MT systems by adding a layer to transform the source embeddings and representation into a universal space to augment the translation of low resource language, which is similar to ours. \cite{Roee2020} implemented a massive many-to-many multilingual system, employing many low-resource language pairs. All of the mentioned works have shown substantial improvements in low-resource translation, however, they are less correlative to our translation tasks.
Although multilingual MT equips a shared space with many advantages, rare word translation is still the issue that needs to be considered. The task of dealing with rare words has been mentioned in previous works. \cite{luong2015} copied words from source sentences by words from target sentences after the translation using a bilingual dictionary. \cite{Li2016} and \cite{trieu} learned word similarity from monolingual data to improve their systems. Our approach is similar to these works, but we only learn similarity from the shared multilingual space of MT systems. \cite{ngo2019} addressed the rare word problem by using the synonyms from WordNet.
\cite{Toan2017} and \cite{ngo2019} presented different solutions to solve rare word situation by transforming the embeddings during the training of their RNN-based architecture. Those solutions cannot be applied to the transformer architecture. In \cite{Gu2019}, the embeddings of rare tokens and universal tokens are jointly learned through a plus parameter while we only add a scalar value to the embeddings.
Monolingual data is used to generate synthetic bilingual data in sparsity data issues. \cite{Sennrich2015} proposed back-translation method that uses a backward model to get the source data from the monolingual target data. In contrast, \cite{zhang2016} shown the self-learning technique by employing a forward model to translate monolingual source data into the target data. \cite{wu2019} incorporated both mentioned techniques into their NMT systems. Monolingual data is also demonstrated its efficiency in unsupervised machine translation\cite{lample2018unsupervised} or in zero-shot multilingual NMT \cite{siddhant2020leveraging, ha2017effective}. In our work, we use the self-learning method to produce pseudo bilingual data, and it is then used to train our low-resource multilingual NMT systems.
\section{Conclusion and Future Work}
We have built multilingual MT systems for two low-resource language pairs: English-Vietnamese and French-Vietnamese, and proposed two approaches to tackle rare word translation. We show that our approaches bring significant improvements to our MT systems. We find that the pseudo bilingual can furthermore enhance a multilingual NMT system in case of French $\rightarrow$ Vietnamese translation task. In the future, we would like to use more language pairs in our systems and to combine proposed methods in order to evaluate the effectiveness of our MT systems.
| bd1319a7598785faa7c4ede498c61feeee24f0eb | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In all areas, such as banks, government applications, the pharmaceutical
sector, military organisations, educational establishments, etc., security issues are growing today. Government institutions set guidelines, pass regulations, and compel organizations and
agencies to conform with these standards, with wide-ranging implications of non-compliance.
In these various and varied industries with a common weak link being passwords, there are many challenges when it comes to security issues. To verify the identity of the user,
most applications today rely on static passwords. These keys, though, come with serious security issues for administrators. Users prefer to use easy-to-guess passwords, use different accounts with the same password, write passwords or save them on their computers unencrypted. Moreover, although dedicated systems, called password managers, can offer secure password storage and retrieval, only a small fraction of users use them \cite{zhang2019people}. In addition, hackers have the choice of using many password stealing methods, such as shoulder surfing, snooping, sniffing, guessing, etc. Several best practices have been suggested for the use of passwords.
Some of them are very difficult to use and
others do not fulfill the security needs of the organization.
To overcome the password problem, two factor authentication using devices such as tokens and
ATM cards has been suggested and has been shown to be difficult to hack \cite{twoFactor}. There are several limitations of two-factor authentication, including the
cost of purchasing, issuing, and handling tokens or cards.
From the point of view of the user, having more than one two-factor authentication
methods demands the purchase of several tokens/cards that are likely to be misplaced or stolen.
Traditionally, cell phones have been considered
a device for making phone calls.
But today, the use of cell phones has been generalized to send
calls, review addresses, shop contacts, etc., provided the developments in hardware and software. Also, opportunities for smartphone access have expanded.
Cell phones combine infra-red, Bluetooth, 3G,
and WLAN connectivity, on top of normal GSM connectivity.
For contact purposes, most of us, if not all of us, hold cell phones.
Several accessible mobile banking services take advantage of mobile computer enhancement capabilities. From the ability to collect account balance information in the form of SMS messages to the use of WAP and
Java along with GPRS to allow fund transfers between accounts, stock trading, and direct payment confirmation through the phone's micro browser.
The principle of using passwords and smart cards to authenticate customers is an old idea going back 40 years now. Since then many systems with two-factor authentication mechanisms were developed. However since the smart card may be intercepted and the data contained in the smart card may be duplicated, the reliability of two-factor authentication may be breached, and the number of potential passwords can be limited and users could forget or lose their passwords.
Biometric authentication was adopted to authenticate users by using their biometric characteristics due to those issues.
Scholars have suggested biometric authentication system since back in 1999 which enhances some facets of two-factor authentication since biometric features have greater entropy and can not be missed and are rarely lost. One drawback, though is that biometric characteristics are not entirely confidential since one can "steal" biometric characteristics from others for example, the fingerprint can be retrieved from a mug used by the suspect and the facial features can be obtained from an image of a user. Combining all these three variables together is a way to mitigate these concerns. This technique is often referred to as three-factor authentication, and has been greatly adapted by cloud-based applications.\cite{yu2014efficient}
SIM cards are available in varying storage sizes. Related memory utilization of the SIM card connected with it plays a part in deciding the effectiveness of cloning the SIM card, more memory stored on the original SIM card than the longer the Ki A8 algorithm cracking process on the SIM card. Problems resulting from the above perspective relating to the inclusion of the A8 algorithm inserted in any SIM card used by telecommunications users to duplicate or replicate the SIM card are detrimental to the privacy and protection of cell phone users on either side.
The purpose of the SIM card cloning research is to provide an alert to consumer safety and provide a dedicated SIM card to tackle SIM card cloning criminal investigations along with their abuse of data.Subscriber Authentication Based on IMSI (Stored on SIM) and Andom Number Generator/RAND (Provided by Network), SIM card cloning authentication will be further investigated by comparing the network login response of the customer to the mobile service network. The Random Number Generator (RAND) includes an algorithm A3 (Provided by Network) such that RAND participates in the process of cloning the SIM card in order to adapt the algorithms contained in the SIM card A8 to A3 algorithms contained in the user data of the connected network authentication. \cite{anwar2016forensic}
Scholars have already demonstrated that by launching a cross-platform infection attack, an attacker is able to compromise another device, either a PC or a cell phone. Prototypes of proof-of-concept demonstrate that such attacks are feasible and thus it is not fair to preclude them from the mobile 2FA scheme adversary model. The intruder will snatch all authentication tokens and impersonate the rightful user when both 2FA devices are infected, regardless of what individual smartphone 2FA instantiation is used.We carry out attacks against various instantiations of mobile 2FA schemes implemented by banks and common Internet service providers to help our argument.
Schemes with 2FA OTPs created on the client side, such as Google Authenticator (GA), depend on pre-shared secrets. The configuration process of the GA app, used by hundreds of providers, including Google Mail, Facebook and Outlook.com, was evaluated. When the user allows GA-based authentication in his account settings, the GA initialization begins. A QR code is created by the service provider and shown to the user (on the PC) and scanned by the user's smartphone. All the information required to initialize GA with user-specific account details and pre-shared secrets is stored in the QR code. During the initialization process, scholars analysed the QR code submitted by Facebook and Google and defined the structure of the QR code. This includes information such as the scheme sort (counter-based vs. time-based), the service and account identifier, the counter (counter-based mode only the generated OTP duration and the mutual secret identifier.
In addition, all this material is provided in plain text. To check if GA supports any alternate initialization system, scholars \cite{dmitrienko2014security} reverse engineered the app with the JEB Decompiler and evaluated the internal apps. We have not found any alternate initialization routines, suggesting that this initialization protocol is used by all 32 service providers using GA.
The initialization message may be intercepted by a PC-residing malware (clear text encoded as an QR code). The attacker will then initialize its own version of the GA and can produce legitimate OTPs.
The use of 'honeywords' was introduced in order to detect whether or not the password file was stolen, i.e. a series of false passwords that are combined with the original password of the user and the hash values of these passwords (real passwords and honeywords) are contained in the password file.
The adversary also does not know which one is the true password if this file is corrupted and all the hash values in the file are cracked. Note that LS identity and password are submitted by the customer or the adversary to request login.LS then checks if a password submitted is among the honeywords of a user, but even if this search succeeds, LS needs to review another protected component, HC, to see if the index of the honeyword retrieved corresponds to the actual password of the user. HC warns the administrator otherwise, as a honeyword signal has been detected that the password file might have been corrupted \cite{genc2017examination}.
Based on these findings and trying to combine the strengths of honeywords and 2FAs while at the same time keeping the system simple and easily integrated in any existing platform or system, we present in this paper a prototype of a novel security mechanism. We develop and propose an innovative security mechanism for web applications that produces both passwords and QR codes covering different login modes. The proposed system entitled "Two-Factor HoneyToken Authentication (2FHA)", combines the strengths of two-factor authentication and Honeyword technologies. In the developed prototype a sms with 3 OTP passwords that correspond to 3 QR codes is sent to the user. Only one of these three elements is the correct token that can be used in order to continue. This induces an extra layer of security adding more safety to the system. The proposed system offers enhanced security to the user while at the same time is simple and doesn't impose additional overhead during login.
The rest of the article is structured as follows. Section \ref{2FA} presents two-factor authentication principles and limitations. Section \ref{sec:honey} discusses honeywords principles. Section \ref{sec:prototype} presents the proposed system architecture and protopype and Section \ref{sec:concl} concludes the article and discusses future work.
\section{Two factor authentication}\label{2FA}
Two-factor authentication (2FA) is a security mechanism in which users use two separate authentication keys to validate themselves, often referred to as two step verification or dual-factor authentication. This process is undertaken to help secure both the credentials of
the user and the tools that can be used by the user.
Two-factor authentication offers a higher degree of protection than one-factor authentication (SFA)-dependent authentication systems, in which the user only provides one factor, normally a password or passcode. Two-factor authentication strategies rely on a password-providing mechanism, as well as a second factor, typically
either a safety token or a biometric factor, such as a fingerprint or facial scan.
Two-factor authentication brings to the authentication process an extra layer of security by making it more difficult for criminals to obtain access to computers or online accounts of an individual since it is not enough to know the victim's password alone to pass the authentication check. To monitor access to confidential applications and files, two-factor authentication has long been used and online service providers are gradually using 2FA to secure
the identities of their customers from being used by hackers who have compromised a password database or used phishing campaigns to acquire user passwords\cite{twobirds}.
\subsection{What are authentication factors?}
There are many different ways in which more than one authentication mechanisms are used to authenticate anyone.
Most authentication mechanisms usually rely on factors of information, such as a traditional password,
whereas two-factor authentication methods incorporate either a possession factor or a factor of inherence \cite{ferrag2017authentication}.
Authentication factors, listed in approximate order of adoption for computing, include the following:
\begin{enumerate}
\item A knowledge factor is when The user knows something, such as a password, a personal
identification number (PIN) or some other sort of mutual secret.
\item A possession factor is when a user has To accept authentication requests, the user has something, such as an ID card,
a protection key, a cell phone, a mobile computer or a smartphone app.
\item An inherence factor refers to anything intrinsic to the physical self of the
individual is more generally considered a biometric element.
This may be personal characteristics, such as fingerprints authenticated by a fingerprint scanner, are mapped to physical features. Facial and speech recognition are other widely used inherence variables.
There are also the biometrics of behavior, such
as keystroke dynamics, variations of gait or voice.
\item A location factor typically denoted by the location from which an authentication attempt is made, can be implemented by restricting authentication attempts to specific devices in a specific location or more commonly, by monitoring the geographical source of an authentication attempt based on the Internet Protocol (IP) source address or some other geolocation detail, such as data from the Global Positioning System (GPS),
\item A time factor limits user authentication to a fixed time frame where it is allowed to log in and limits access to the device beyond that window.
\end{enumerate}
It should be remembered that the vast majority of two-factor authentication mechanisms rely on the first three authentication factors, while multifactor authentication
(MFA), which may rely on two or more separate passwords for more reliable authentication, can be used by systems that demand greater security.
\subsection{How does two-factor authentication work?}
In this section we briefly describe the process of a typical two factor authentication system \cite{ferrag2017authentication}.
\begin{itemize}
\item The user is asked by the program or by the website to log in.
\item The user enters what he or she knows—usually a username and password. Then a match is made by the site's server and the user is remembered.
\item The website creates a special authentication key for
the user for processes that don't need passwords.
The authentication function processes the key and
it is checked by the site's server.
\item Then the site asks the user
to start the second stage of login.
While a variety of ways can be taken through this step, users must show that they only
have what they will have, such as an identification key, ID card, smartphone or other mobile device.
This is the
factor for ownership.
\item During phase four, the
user enters a one-time code created.
\item The customer is authenticated and given access to the program or website after supplying all variables.
\end{itemize}
In technical terms, two authentication factors are required to obtain access to a device or facility at any point.
Using two variables from the same group, though, would not constitute 2FA; for instance, it is always called SFA
to require a password and a mutual secret since both belong to the same class of authentication factor: information.
The user ID and password are not the
most reliable as far as SFA services.
One concern with password-based authentication is that generating
and recalling good passwords requires awareness and diligence.
Passwords need protection against many internal attacks, such as carelessly kept
login credential sticky notes, old hard drives and vulnerabilities in social engineering.
Passwords are often vulnerable to external threats, such as hackers using brute-force, dictionary or
rainbow table attacks.
An intruder will typically break password-based protection mechanisms and steal corporate
data, including personal information of users, provided ample time and money.
Because of their low cost, ease of execution and
familiarity, passwords have remained the most common type of SFA.
Depending on how they are applied, several challenge-response questions can provide more security,
and stand-alone biometric authentication approaches can also provide a more reliable SFA process.
\subsection{Types of two-factor authentication products}
There are several different 2FA deployment equipment and utilities — from tokens to radio frequency identification (RFID) cards to applications for smartphones \cite{ferrag2020authentication}.
It is possible to separate two-factor authentication devices into two categories: tokens that are provided to users to use
while signing in and infrastructure or software that detects and authenticates entry for users who correctly use their tokens.
Physical devices, such as key fobs or smart cards, may be authentication keys, or
they may exist in applications like mobile or web apps that produce authentication PIN codes \cite{limbasiya2018advanced}.
These authentication codes are normally created by a server, often known as one-time
passwords (OTPs), and can be recognized by an authentication system or app as authentic.
The authentication code is a short sequence connected to a specific computer, user
or account that can be used once as part of an authentication process.
To accept, process and authorize — or reject — access to
users who authenticate with their tokens, organisations need to install a framework.
This may be implemented in the form of cloud applications, a dedicated hardware server, or supplied by a third-party provider as a service.
An significant feature of 2FA is ensuring that the authenticated user is granted
access to all services the user is allowed for — and only those resources.
As a consequence, one of 2FA's main functions is to
connect the authentication method with the authentication data of an entity.
Microsoft offers some of the required infrastructure for Windows 10 2FA service organisations through Windows
Hello, and will work with Microsoft accounts, as well as authenticate users with Microsoft Active Dii.
\subsection{How 2FA hardware tokens work}
Hardware tokens for 2FA are
available that support numerous authentication approaches \cite{reynolds2020empirical}.
The YubiKey, a small Universal Serial Bus (USB) system that supports OTPs, public key encryption and authentication,
and the Universal 2nd Factor (U2F) protocol developed by the FIDO Alliance, is a common hardware token.
YubiKey tokens are sold by
Palo Alto, California-based Yubico Inc.
When YubiKey users log in to an OTP-supported online site, such as Gmail, GitHub, or WordPress, they insert their
YubiKey into their device's USB port, enter their password, select the YubiKey field, and then tap the YubiKey icon.
YubiKey produces and inputs
an OTP into the field. The OTP is a 44-character, single-use password; a special ID defining
the authentication key associated with the account is the first 12 characters.
The remaining 32 characters contain information that is encrypted using a key only known to the
computer and the servers of Yubico that was generated during the initial registration of the account.
An OTP is submitted from an online
service to Yubico for verification of authentication.
The Yubico authentication server sends back a message verifying that this
is the correct token for this user until the OTP is checked. Two authentication criteria have been given by the user: the information
factor is the password, and the possession factor is the YubiKey.
\subsection{Two-factor authentication for mobile device authentication}
For 2FA, smartphones provide a number of possibilities,
encouraging organizations to choose what suits best for them.
A built-in camera can be used for face recognition or iris detection, and the
microphone can be used for speech recognition. Certain applications are able to recognise fingerprints.
GPS-equipped smartphones will check the
location as an extra consideration.
Also, Speech or Short Message Service (SMS)
may be used as an out-of-band authentication channel.
For receiving authentication codes by text message or automatic
phone call, a trustworthy phone number may be used.
To participate in 2FA, a person needs
to check at least one trustworthy phone number.
Both applications that support 2FA are available for Apple iOS, Google Android and Windows 10,
allowing the phone itself to function as the physical interface to satisfy the ownership aspect.
Duo Defense, headquartered in Ann Arbor, Mich., and acquired for \$2.35 billion by Cisco in
2018, is a 2FA software provider whose solution allows 2FA consumers to use their trusted products.
Before checking that the mobile device can still be trusted to
authenticate the customer, Duo's platform first determines that a user is trusted.
The need to acquire an authentication code through text,
voice call or email is replaced by authenticator apps.
For example, users type in their username and password to access a
website or web-based application that supports Google Authenticator — a knowledge factor.
Users are then asked to
type a number of six digits.
Instead of having to wait a few seconds to answer
a text message, an Authenticator produces the number for them.
Every 30 seconds, these numbers alter
and are different with every login.
Users complete the authentication process by entering the correct number
and show custody of the correct unit — an ownership element.
\subsection{Is two-factor authentication secure?}
There are several limitations of two-factor authentication, including the
cost of purchasing, issuing, and handling tokens or cards.
From the point of view of the user, having more than one two-factor authentication
method allows several tokens/cards to be held that are likely to be misplaced or stolen.
Although two-factor authentication improves security—because access privileges are no longer dependent solely on a password's strength,—two-factor authentication systems are just as reliable as their weakest part.
Hardware tokens, for instance, depend on
the security of the issuer or manufacturer.
In 2011, when the technology firm RSA Security announced its SecurID authentication tokens had
been stolen, one of the most high-profile examples of a compromised two-factor device occurred.
If it is used to circumvent two-factor authentication, the account recovery mechanism itself can often be subverted because it sometimes resets the
existing password of a user and e-mails a new password to allow the user to log in again, bypassing the 2FA process.
The corporate Gmail accounts of the chief
executive of Cloudflare were compromised in this way.
Although 2FA is cheap, simple to implement and user-friendly based on SMS, it is vulnerable to multiple attacks.
In its special publication 800-63-3, the National Institute of Standards and
Technology (NIST) has discouraged the use of SMS in the 2FA services \cite{grassi2017draft}.
Due to cell phone number portability attacks, such as the Signaling System 7 hack, against the mobile phone network and malware, such as Eurograbber, that can be used to intercept or divert text messages, NIST concluded that OTPs sent via SMS are too vulnerable.From all the above factors the idea of 2HFA is created.
\section{Honeywords}\label{sec:honey}
The fundamental principle behind the Honeywords scheme is to adjust the password storage mechanism in such
a way that a password and a series of false passwords are associated with each account \cite{honeywords}.
The phony passwords are called honeywords. Sweetwords are
the union of both honeywords and the password.
As soon as the password is entered during the authentication
process, the password database is immediately detected to have been compromised. Therefore unlike traditional schemes, implementations focused on
honeywords can effectively detect violations of password databases.
\begin{figure}
\centering
\includegraphics{figures/hashes.png}
\caption{Credentials database of aLSin the Honey-words system}
\label{fig:honey1}
\end{figure}
\begin{figure}
\centering
\includegraphics{figures/passwordindex.png}
\caption{Data stored on aHC}
\label{fig:honey2}
\end{figure}
The method of Honeyword is as follows.
During the authentication process, users select a username and a password, as with many traditional schemes. The
Login Server (LS) then produces honeywords for the password and maintains a record in the database of passwords.
The ordering of the sweetwords is randomly
selected by the LS in each record.
In addition, LS sends the corresponding user ID and actual password index
to Honeychecker (HC), the auxiliary server built to store the password index.
Let ui and H() denote respectively the user name of user I and the
hash function used in the method. H(swi,j) denotes the hash of user i. jth sweetword. A standard example of a table
of qualifications is illustrated in Figure \ref{fig:honey1}.
HC saves the user IDs and the password index between the honeywords. During the authentication, no username or password itself is sent to HC. In comparison, HC is built as a hardened server that can only be reached by LS.
A standard structure of the
HC data is seen in Figure \ref{fig:honey2}.
Notice that only two kinds of messages
are accepted by HC: Check and Set
To verify {\bf if j=ci, check(i, j)} implies that uf j=ci, HC returns True, otherwise False
is returned and a warning is activated.
The command set is structured as:
{\bf Set (I j) indicates setting ci=j}.
The user submits its username and password.LStries during the authentication process
to locate the corresponding record for that username in the credentials database.
If a record exists, LS computes the hash of the password
sent and attempts to find a match in the sweetword hashes. If no match occurs, then the password
sent is incorrect and access is refused.
LS sends the respective user ID and the
corresponding index to HC if there is a match.
First, HC seeks the record that fits the user ID and
compares the index value obtained with the one stored in its database. If the outcome is
valid, then access is provided.
Otherwise the HC returns incorrect, generates
an alert and notifies the administrators of the device policy.
Originally, the Honeywords scheme was constructed with the expectation that the opponent
could steal the hashed passwords and invert the hashes to obtain the passwords.
It is therefore presumed that both LS and HC will
not be abused by the attacker within the same time frame. The Honeywords mechanism defends passwords from brute-force and dictionary attacks mentioned in Section \ref{2FA}. The method attempts to prevent violations of the password database and seeks to prevent only offline dictionary attacks where the adversary is believed to have taken the hashes of the password and abandoned the system.
\section{The proposed Two-Factor HoneyToken Authentication (2FHA) Mechanism}\label{sec:prototype}
In this article we introduce an alternative authentication method, for enhancing systems' security. The system combines two factor authentication with honeywords in order to make impossible for an attacker to bypass the authentication mechanism of the system. Even in the occasion that the attacker has access to the device that receives the token, e.g. by sim cloning, the proposed 2FHA method makes the authentication bypass unfeasible if not impossible.
\begin{figure*}[h]
\centering
\includegraphics[width=0.85\linewidth]{figure.jpg}
\caption{Architecture of the 2FHA protoype}
\label{fig:my_label}
\end{figure*}
In order to demonstrate the proposed system we created a website that includes a login page and have developed a prototype. The user in order to enter the system must fill the correct username and password, which is the first authentication factor. Then the system sends to the user a number $M$ that indicates the token that is correct on every login attempt in the future. When logging into the system from a new device, the user must enter the correct OTP. The user receives a number of tokens $N$. He can choose with what platform wants to be alerted for the token, to get it (e-mail, sms, phone call etc.).
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{figures/login.png}
\caption{The login page}
\label{fig:login}
\end{figure}
Then we must enter the second authentication factor. The prototype of the 2FHA mechanism produces 3 qrcodes\cite{qrcodes}, each one of those is represented with a password and sends an sms message\cite{website} to the mobile phone of the user. The sms includes all 3 OTPs (One Time Password) passwords corresponding to each of the qrcodes \cite{atms}. One is the correct and the others 2 are fake. The user now has to choose what it’s more suitable method for him to continue in order to fill the OTP box and proceed in the website\cite{security}. We ahve to highlight here that the number of produced tokens is kept to 3 only for demonstrating purposes but can be generalized to a number $N$.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{figures/qrcodes.png}
\caption{The produced qrcodes}
\label{fig:qrcodes}
\end{figure}
If the user chooses to scan the qrcodes \cite{QRCODES1}, the process is simple. He scans the correct qrcode and then he fills the OTP box. The qrscanner is free software and most of them are suitable for any device. If the user doesn’t have qrscanner then the option of sms is more convenient for him. The sms message as presented in Figure \ref{fig:SMS}, will be sent to the user the time he logins to the system. As you can see in Figure \ref{fig:SMS} the message contains 3 OTP passwords(OTP, OTP1, OTP2). These are the produced from the qr codes. Each user knows that only one of the 3 qrcodes is the correct while the other 2 are fake.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{figures/sms.jpg}
\caption{OTP passwords sent as an sms message}
\label{fig:SMS}
\end{figure}
If the user fills the OTP box correctly, he will continue to the system. If not, then he will be sent back to the initial login page and has to follow the procedure again. Also for precaution reasons the account of the user can be suspended. The OTPs must follow some rules when created; they can't be very similar among them in order to avoid mispelling mistakes.
\section{Conclusion - Discussion}\label{sec:concl}
In this paper we have taken actions to strengthen the security of a system against stolen tokens and penetration attempts. The proposed mechanism combines 2FA and Honeyword principles and can be integrated in any existing platform or web application. We plan to improve the system in the future by producing a higher number of qrcodes and passwords that will increase the security. In the prototype of the proposed system OTP's are sent them through SMS. In the near future we plan to integrate the proposed 2FHA with google and microsoft authenticators. We also plan to enhance the registration phase in order to make it more secure by encrypting the initial information.
\balance
\bibliographystyle{IEEEtran}
| da581cefbc7751d72fa17b9ad04f5ca43698c069 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
At the end of this year, the largest national optical telescope, Big Telescope Azimuthal (BTA), with a 6-meter primary mirror marks its 45th anniversary, counting from the date of signature of the certificate by the interagency acceptance committee on December 30, 1975. Figure~1 shows the copies of the first images that were obtained with the BTA and confirmed the compliance of its parameters with the design characteristics. Until 1993, the BTA had been the largest telescope in the world, now it remains among the top twenty largest instruments. It is still the country`s largest telescope and a joint-use instrument for Russian and foreign scientists.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig1.png}
\caption{The copies of the first images obtained with the BTA.}
\end{figure}
During its operation, the telescope has been constantly developed and upgraded. In the first decades, the main manufacturers of the telescope, among them LOMO (Leningrad Optical Mechanical Association), LZOS (Lytkarino Optical Glass Plant), and GOI (Vavilov State Optical Institute), had been providing assistance and patronage; recently, this task has almost completely fallen on the SAO~RAS personnel. Among the main results that contributed to maintaining a high level of research at the BTA, we can highlight the creation of the new automated control system (BTA ACS), introduction of remote observations for most scientific equipment units, implementation of a program for the creation of modern light detectors, and, finally, the development of observing methods that cover the entire spectrum of scientific problems solved with the telescope.
Summing up the intermediate results, we should also note a number of our failures, among those a technological error in the primary mirror repolishing procedure and the lack of methods for observing in the infrared range at the BTA. Among the reasons that had a negative impact are insufficient funding for complex technological projects, technological gap in the Russian industry, and the deficit of human resources.
Rather a modest place of our telescope in the current world rankings, new large ground-based projects expected in this decade---the Extremely Large Telescope (39.3-m mirror, consortium of European countries) and the Giant Magellan Telescope (25.4~m, consortium led by the USA)---force us to search for a niche where our telescopes are going to be in demand and competitive at the world level.
It can be argued that in recent decades the BTA has not been able to compete with new-generation world telescopes in addressing most challenging observational problems. In addition to purely technological reasons---the absence of active and adaptive optics on the telescope in the presence of a ``thick'' glass mirror with an average, by contemporary standards, diameter---there is the astroclimate, which is dramatically inferior to the best foreign sites both in seeing and in the number of clear nights.
Nevertheless, the search for our ``ecological niche'' does not seem to us a hopeless venture. The BTA remains a fairly large instrument, it has a unique geographical location in latitude and longitude among the world`s telescopes. Not least, the telescope is equipped and is constantly being reequipped with modern scientific instruments and light detectors, which allows astronomers to obtain world-class scientific results. In addition, an essential function of SAO~RAS is the preservation of personnel potential, the staff of highly qualified specialists having experience in creating original research and observing methods and their implementation at a large telescope.
\section{Scientific Equipment and Methods of Observation}
The history of the observatory and the BTA is inextricably associated with development of observing methods and scientific equipment. We can say with confidence that the current observing methods used at the telescope are organically related to the scientific interests of SAO~RAS scientists, and thus already reflect BTA characteristics and naturally denote a niche for our research. It can be designated as (1)~conducting research by classical means: high and low resolution spectroscopy and photometry of objects difficult to access for other large instruments and (2)~the use of technically sophisticated and not widely used methods: high-accurate spectropolarimetry, fast photometry, panoramic spectroscopy, and speckle interferometry. Of course, the development and support of such methods should be continued in the future.
\begin{figure}
\centerline{\includegraphics[width=0.45\columnwidth]{Fig2.png}}
\caption{The universal spectrograph SCORPIO-2.}
\end{figure}
\subsection{Spectroscopy and Photometry of Faint Objects}
The multipurpose spectrographs of the BTA primary focus: SCORPIO-1 \citep{sco1} and SCORPIO-2, developed in the Laboratory of Spectroscopy and Photometry of Extragalactic Objects under the leadership of Chief Researcher V.\,L.~Afanasiev,
implement classical methods of spectroscopy and photometry of faint objects in a $6'\times6'$ field of view in the whole visible wavelength range (from 360 to 1000~nm) with resolution $R=\lambda/\Delta\lambda$ from 500 to 2500. SCORPIO-2 (shown in Fig.~2) perform in addition highly accurate polarimetric and spectropolarimetric observations \citep{sco2_pol} and is equipped with a set of scanning Fabry--Perot interferometers \citep{sco2_fpi2002, sco2_fpi2008, sco2_fpi2015} with resolution $R$ in the interval from 250 to 16\,000, allowing panoramic spectroscopy in the same wide field ($6'\times6'$) but in a narrow spectral range (smaller than 10~nm). The developers also intend to introduce an integral-field spectroscopy unit \citep{sco2_ifu}, which is going to implement spectroscopy of extended objects with angular resolution better than $1''$ and spectral resolution $R$ in the range of 1000--2800 within a $16''\times16''$ field. SCORPIO-1 and SCORPIO-2 are equipped with large-format light detectors based on CCDs up to $2048\times4600$~pixels, manufactured in recent years by the Advanced Design Laboratory (ADLab) of SAO RAS.
\subsection{Stellar Spectroscopy}
The list of scientific equipment designed to obtain stellar spectra at the BTA with resolution greater than 10\,000 is represented by two instruments: the Nasmyth Echelle Spectrograph (NES), created under the leadership of V.\,E.~Panchuk in the Astrospectroscopy Laboratory, and the Main Stellar Spectrograph (MSS), the only instrument that remained in the BTA equipment out of those created by LOMO and GOI during the construction of the telescope.
\begin{figure}
\includegraphics[width=\columnwidth]{Fig3.png}
\caption{A planetary nebula spectrum obtained with NES in the 5700--8500\,\AA\ range.}
\end{figure}
NES \citep{nes} obtains spectra of stars up to $10^m$ with\linebreak \mbox{$R\sim40\,000$--$50\,000$} and a simultaneously recorded spectral range of about\linebreak 300~nm. To increase efficiency, the spectrograph has 3-fragment image slicers with slit widths of $0\farcs4$ and $0\farcs6$, there are also classical ($0\farcs6\times2\farcs5$) and adjustable \mbox{($0\farcs2$--$2''\times0''$--$40''$)} slits. The instrument is equipped with two interchangeable cross-dispersers for operation in the $\lambda$ ranges of 305--600~nm and 480--1000~nm. Figure~3 shows an example of a NES spectrum obtained in the near infrared wavelength range.
The Main Stellar Spectrograph has been repeatedly upgraded by V.\,E.~Pan\-chuk, G.\,A.~Chountonov, and I.\,D.~Naidenov \citep{mss} in order to install large-sized diffraction gratings, introduce CCD systems, and implement spectropolarimetry (circular polarization analyzers). The spectrograph obtains spectra of stars brighter than $12^m$ with resolution up to $R\sim15\,000$ and simultaneously recorded ranges of 55~nm for wavelengths shorter than 500 nm and more than 80~nm for the 560-900~nm range. The instrument operates mainly in the spectropolarimetric mode with a circular polarization analyzer combined with an image slicer producing 7~fragments for each polarization.
Both spectrographs are equipped with large-format light detectors developed in SAO~RAS ADLab based on the $2048\times4608$~px E2V CCD42-90 chips.
\subsection{Speckle Interferometry with the Diffraction-Limited Angular~Resolution}
The Group of High Resolution Methods in Astronomy led by Yu.\,Yu.~Balega has been developing and constantly improving the speckle interferometry method \citep{speckle} since the 80s. The method is aimed at studying binary and multiple star systems, single stars, asteroids, and other celestial bodies with an angular resolution close to the diffraction limit of the telescope. This can be achieved by Fourier analysis of interference patterns from sub-apertures of the primary mirror, detected with exposures of 5--30 ms (the time of ``freezing'' of atmospheric fluctuations). The instrument operates in the 500-900~nm range with a set of narrow interference filters, its field of view with the currently used $512\times512$~px Andor EMCCD iXon Ultra 897 light detector is 4\farcs4, 7\farcs1, or 28\farcs2 depending on a microlense used, the limiting magnitude is $15^m$, the angular resolution is $0\farcs02$.
\subsection{Superfast Photometry}
For studies of optical variability of astrophysical objects with high temporal resolution (up to 1~ms), i.e., stellar mass black hole candidates, pulsars, flare stars, etc., the Group of Relativistic Astrophysics under the leadership of Leading Researcher G.\,M.~Beskin designed the Multimode Panoramic Photopolarimeter \citep{mppp}. The method can use two types of detectors depending on required temporal resolution: either two position-sensitive detectors (to achieve a temporal resolution of the order of 1~ms) or an EMCCD camera with charge amplification (for times of the order of 0.1~s). The field of view is $1'$ in the $UBVR$ panoramic photometry mode, there is also a slit mode with adjustable slit widths and heights \mbox{($0''$--$10''\times10''$--$60''$)}. Linear polarization measurements (using a double Wollaston prism) and spectral modes with an Abbe prism and a diffraction grating are available.
\subsection{Light Detectors}
The Advanced Design Laboratory began development of optical light detectors for telescopes in the 80s under the leadership of Ph.D. in Engineering Science S.\,V.~Markelov, and since the beginning of the 90s, as technical characteristics improved, the detectors began to be introduced into regular BTA observations, gradually replacing, where possible, photographic plates and detectors based on image intensifiers. Improvement of the quality of semiconductor crystals and development of the signal processing technique allowed us by the turn of the century to move from $256\times256$~px systems with quantum efficiency of 30\%--40\%\ in the red part of the spectrum and additional noises of the order of hundreds of electrons to $2048\times2048$~px systems with quantum efficiency of almost 90\%\ and noises of several electrons.
For the last 15~years, SAO telescopes have almost everywhere used systems sized from $2048\times2048$~px to greater formats, and all further development followed mainly the path of increasing the stability of signal detection, expanding the spectral range to the red part, and eliminating effects of light interference in detector layers. To date, the observatory has mastered the production of detection systems based on the E2V $4096\times4096$ chips, which have almost ideal properties: high quantum efficiency and a low level of interference noise in the near IR range. As part of joint work with Russian enterprises, ADLab has developed a technology for creating mosaic detection systems of an almost arbitrary format limited only by price and the size of the cryostat. Figure~4 shows a CCD system designed to equip the BTA fiber-fed spectrograph.
\begin{figure}
\centerline{\includegraphics[width=0.45\columnwidth]{Fig4.png}}
\caption{A CCD system based on a $4128\times4112$~px E2V CCD 231-84 detector, created for the BTA fiber-fed spectrograph, was accepted for observations in the summer of 2020.}
\end{figure}
\section{BTA Projects under Development}
\renewcommand{\labelitemi}{\textbullet}
Among the equipment at various stages of implementation, the appearance of which is expected in the near future, we can name: \begin{itemize}
\item Fiber-fed optic spectrograph for the BTA \citep{ffs}. It was manufactured as part of the SAO RAS project ``Evolution of Stars from Their Birth to the Origin of Life'' supported by the Russian Science Foundation in 2014--2018. The spectrograph is installed on the telescope and equipped with a light detector, test observations are carried out with a temporary optical camera, first results have been obtained, the manufacturer is adjusting the permanent optical camera.
\item Photospectropolarimeter in the Nasmyth-1 focus. The instrument is intended for operating in the semi-automatic mode for observations of alert transients. A $1024\times1024$~px Andor EMCCD system with subsecond temporal resolution is planned as a detector.
\item To increase the recording rate in panoramic observations at the BTA,\linebreak SAO~RAS ADLab started in 2019 the development of systems based on CMOS sensors within the project ``Development of the Large Unique Scientific Facility Big Telescope Alt-Azimuthal'' supported by the Ministry of Science and Higher Education in 2019-2020 within the federal target program ``Research and Development in Priority Areas of Development of the Scientific and Technological Complex of Russia for 2014--2020.\!'' The laboratory have created a prototype of a fast photodetector based on the CMOS sensor Gpixel GSENSE4040CMN and developed design documentation for a photodetector based on the astronomical-class large-format back-illuminated sensor GSENSE6060.
\item The BTA Multiobject Fiber Spectrograph (MOFS) is being upgraded. It is planned to develop a new camera, improve the technology for making masks, upgrade the control computer and software of the instrument.
\item The observatory is planning to introduce the first, experimental, adaptive system in the BTA Nasmyth-2 focus, which should improve the seeing in observations with high spectral resolution methods. The purchased adaptive system for correction of large-scale wavefront aberrations will perform aberration correction up to the 4th order at a frequency of 200~Hz with an accuracy of $\lambda/15$.
\item The work on upgrading the BTA Echelle Spectropolarimeter for Primary Focus (ESPriF) continues; the instrument is planned to be equipped with a $2048\times2048$~px Andor iKon-L 936 light detector.
\end{itemize}
\section{The BTA As a Unique Scientific Facility}
The status of Russia's ``Unique Scientific Facilities'' held by the BTA and\linebreak \mbox{RATAN-600} of SAO RAS has in recent years become increasingly important in maintaining the provision of regular observations and financial stability of the observatory as a whole. Before the reform of the state academies of sciences in Russia, this status often remained nominal without significantly affecting the terms of funding. With a new regulatory framework and after the transfer of economic and administrative management functions from the Russian Academy of Sciences to the jurisdiction of the Federal Agency for Scientific Organizations in 2013, the situation with shared research facility centers (SRFCs) and the unique scientific facilities (USFs) has drastically changed.
For instance, the basic funding of SAO RAS is calculated without taking into account the needs of the telescopes, as for an average scientific institute, and even allowing for the regional location (which, for example, reduces the estimated amount of the salary fund). In this situation, the USF status and the presence of external users make it possible to request and receive additional funds for operation of the telescopes, as well allowing participation in project competitions of the Ministry of Science and Higher Education to attract additional funding for development and renovation of scientific equipment. The loss of the USF status will inevitably lead to the loss of funds for operation of the telescopes.
Among the main target indicators established by a special decree of the Russian Government on the requirements for SRFCs and USFs are the equipment workload (the ratio of the actual USF operating time to the maximum possible operating time per year), which must be 70\% at least, the presence of external users, the number of organizations applying for a USF, and publications in indexed journals. Therefore, SAO RAS is extremely interested in maintaining high efficiency of the telescopes and attracting external users. In 2019, the BTA workload was 96\%, the workload in the interests of external users 36\%, we had 13~applying organizations, 41 papers were published based on observations obtained with the telescope in the journals indexed in the Web of Science.
Even though it should be said that the principles of financing the science in Russia are far from ideal---for example, the specific character of development and production of unique scientific equipment is poorly taken into account, most project competitions are aimed at instant scientific return, amid significant investments in ensuring the work of scientists, inadequate attention is paid to the support of scientific and technical personnel---in general, the government aims at development of the scientific sphere, invests heavily in science and even, to some extent, provides feedback communication, which gives hope for gradual improvement of policy in this area.
\section{The BTA in the SAO RAS Development Program until~2024}
In conclusion, we present some points of the SAO RAS development program for 2019--2023 created in the framework of the implementation of the national ``Science'' project. Surely, the percentage of completion of these plans will be determined by the availability of funding, but the program shows well the directions along which SAO RAS scientists are going to progress in the coming years. The observatory supposes to develop the existing and create new effective methods for dealing most topic astronomical problems: studying the large-scale structure of the Universe, active galactic nuclei, detecting fast radio bursts, identifying optical transients associated with gamma-ray bursts and gravitational events, and exoplanet research. In particular, the development of the BTA complex involves the following projects:
\begin{itemize}
\item development of new scientific instruments to increase technical capabilities of the BTA and expand the observed electromagnetic range: a near-infrared spectrograph for the 0.8--2.5 {\textmu}m wavelength interval and new-generation 2D spectrographs and spectropolarimeters for the primary focus;
\item development of new high-speed large-format light detectors and mosaic systems for optical telescopes;
\item development of adaptive optics for partial compensation of atmospheric distortions and vibrations of the telescope mount;
\item creation of a system for photometric control of the reflective layer in applying the coating onto the BTA primary mirror.
\end{itemize}
\section*{Acknowledgements} Observations with the SAO RAS telescopes are supported by the Ministry of Science and Higher Education of the Russian Federation (including agreement No.\,05.619.21.0016, project~ID RFMEFI61919X0016). The renovation of telescope equipment is currently provided within the national ``Science'' project. The authors would like to thank the Russian Foundation for Basic Research for the consistent support of the RAS large telescope user conferences and the Russian Telescope Time Allocation Committee sessions during many years up to 2018 (e.g., RFBR projects 17-02-20559, 18-02-20136). We should also appreciate the support from the Russian Science Foundation, which allows the observatory to develop new scientific equipment, particularly the BTA fiber-fed spectrograph (RSF project No.~14-50-00043).
| 983be83260143f99a5e7ccc1b53982dda3d24e55 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{\label{sec:level1}Introduction}
\begin{figure}
\vspace{-5mm}
\centering
\includegraphics[width=3in]{HOM_cancelling_4.pdf}
\caption{Hong-Ou-Mandel effect. Two identical photons are sent into different beam splitter input ports. There are four possible outcomes, two photons leaving one port, two photons leaving the other port, each photon reflecting to give single photons at each exit, and both transmitting to give single photons at each port. The coincidence terms cancel out since they are identical but enter with opposite sign. The final state is a superposition of two outcomes, each with both photons clustered together at the same exit port.}
\label{fig:HOM_effect}
\end{figure}
\section{Introduction}
\vspace{-10px}
The Hong-Ou-Mandel (HOM) effect is one of the most recognized quantum two-photon interference effects \cite{hong1987measurement}.
When two indistinguishable photons arrive simultaneously at different inputs of a 50:50 beam splitter (BS), single-photon amplitudes at each output cancel, resulting in quantum superposition of two-photon states appearing at each output port, as in Fig. \ref{fig:HOM_effect}.
This traditional HOM method, observed on a BS having two input and two output ports, always has the two-photon state simultaneously occupying both output spatial modes, leaving no room to engineer control of propagation direction.
Various types of studies on quantum state transformations in multiport devices have been performed such as two photon propagation in a multimode system \cite{weihs1996two,zukowski1997realizable},
quantum interference effects using a few photons \cite{meany2012non,de2014coincidence,tichy2011four,campos2000three}, and propagation of multi-photons \cite{lim2005generalized,tillmann2015generalized,menssen2017distinguishability}. Internal degrees of freedom are also incorporated to enhance communication capacity \cite{walborn2003multimode,poem2012two,zhang2016engineering}.
Systems and procedures using multi-photon states, such as boson sampling, have been analyzed using multiport beam splitters both theoretically and experimentally \cite{aaronson2011computational,tillmann2013experimental,spring2013boson,bentivegna2015experimental,he2017time,wang2019boson}.
The HOM effect plays an important role in the field of quantum metrology when two-photon $|2002\rangle$-type states are extended to $N$-photon $N00N$ state \cite{dowling2008quantum,motes2015linear}.
Additionally, coherent transport of quantum states has been attracting attention, where single- and two-photon discrete-time quantum walk schemes are employed to transfer and process quantum states \cite{bose2003quantum,perez2013coherent,lovett2010universal,chapman2016experimental,nitsche2016quantum}.
A quantum routing approach has been proposed to transfer unknown states in 1D and 2D structures to assist quantum communication protocols \cite{zhan2014perfect,vstefavnak2016perfect,bartkiewicz2018implementation}.
Photon propagation control is especially crucial in a large optical network to distribute quantum states between two parties.
The network can be formed by combining multiple copies of four-port devices.
The state manipulation schemes we present can be integrated in quantum communication protocols since state retrieval timing can be chosen at will.
In this manuscript, we propose two-photon quantum state engineering and transportation methods with a linear-optical system which allows manipulation of photon amplitudes by using linear-optical devices such as optical multiports, beam splitters, and phase shifters.
Previously, such multiports have been introduced to demonstrate a two-photon clustering effect in quantum walks when multiple multiport devices are connected to form a chain \cite{simon2020quantum}. Clustering of two photons means that after encountering a multiport, the input two-photon amplitude separates into a superposition of a right-moving and a left-moving two-photon amplitude, with no amplitude for the photons to move in opposite directions.
By utilizing this separation, a higher-dimensional unitary transformation enables flexible quantum engineering designs of possible travel path combinations by switching relative phases within right moving and left moving amplitudes independently.
When two or more multiports are combined, this control of quantum amplitudes in a two-photon state allows demonstration of a “delayed" HOM effect engaging also time-bin modes in addition to spatial modes.
To perform this delayed effect, two or more multiports are required, and relative phase shifts between two rails can reflect the incoming amplitudes.
This controllable reflection without mirrors can also be seen as an additional state manipulation feature.
We introduce two distinct systems. The first system utilizes direct transformation of two-photons by the four-port device using circulators.
The second case does not have circulators in the system.
The photons are sent from the left side of the beam splitters, then the amplitudes encounter the multiport device.
This second system has not been analyzed in the past. Specific input states for both distinguishable and indistinguishable photons are redistributed between two parties coherently. Therefore, this system is particularly useful in quantum routing type applications.
This article is organized as follows. In Sec. II, we introduce the main optical components used in this manuscript to perform quantum state transformation. These basic linear optical devices are used to show HOM effect engaging in spatial modes and time-bin modes, and is addressed in Sec. III. In Sec. IV, we show redistribution of two photon states using the devices introduced that are presented in Sec. II. The summary of the results are given in Sec. V.
\begin{figure}[htp]
\subfloat[]{%
\includegraphics[clip,width=0.8\columnwidth]{multiport_systematic.pdf}%
}
\vspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=0.8\columnwidth]{multiport_systematic_photons.pdf}%
}
\vspace{-10px}
\caption{(a) A possible experimental realization of a directionally-unbiased linear-optical four-port consists of four beam splitters, four mirrors, and four phase shifters. A photon can enter any of the four ports, and exit at any of the four ports (labeled as $a,b,c$, and $d$). With a specific choice of phase settings, a Grover matrix can be realized by coherently summing all possible paths to each output \cite{osawa2019directionally}. A schematic symbol for this device is shown on the right. (b) Single multiport transformation of a two-photon input state. The input state of two correlated photons entering from the left is depicted as $ab$ $(\ket{1,1})$. After scattering by a Grover multiport, the state transforms into $-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2$, which has clear separation of right- and left-moving two-photon amplitudes. No cross-terms with photons moving in opposite directions occur.}
\label{fig:combined_four_port}
\vspace{-10pt}
\end{figure}
\section{Photonic state transformations via linear optical devices}
\vspace{-10px}
In this section, we consider photonic state transformations in higher-dimensional spatial modes using a unitary four-dimensional Grover matrix \cite{grover1996fast} in place of the beam splitter.
In this section, we introduce the main systems that will be used for linear state transformations, followed by the basic photonic devices to implement them.
Beam splitters and the four-dimensional Grover matrix are the central system component.
We mainly use photon number representation to describe states through out the manuscript.
The general beam splitter transformation matrix is
\begin{eqnarray}
\label{eqn:BS}
&\begin{pmatrix}
\hat{c}
\\
\hat{d}
\end{pmatrix}=
\frac{1}{\sqrt{2}}
\begin{pmatrix}
1 & 1\\
-1 & 1
\end{pmatrix}
\begin{pmatrix}
\hat{a}\\
\hat{b}
\end{pmatrix}
\end{eqnarray}
where $\hat{a}$,$\hat{b}$,$\hat{c}$, and $\hat{d}$ are used to describe the input photon state transformation.
The labels are generalized here, therefore the specific location dependent beam splitter transformations are redefined in later sections.
We use photon number states to describe the system unless otherwise specified. The input state is denoted as $\hat{a}\hat{b}$ where $\hat{a}$ and $\hat{b}$ are respectively creation operators for the spatial modes $a$ and $b$.
The hat notation is dropped henceforth.
For a photon in spatial mode $a$ with horizontally polarized photon is denoted as $a_{H}$ and horizontally polarized photon in mode $b$ is denoted as $b_{H}$.
We omit polarization degrees of freedom when identical photons are used through out the system.
Photonic implementations of the Grover matrix can be readily realized \cite{carolan2015universal,crespi2013anderson,spagnolo2013three,fan1998channel,nikolopoulos2008directional}.
To be concrete, we use directionally-unbiased linear-optical four-ports (Fig. \ref{fig:combined_four_port} (a)) as an example.
Consider sending two indistinguishable photons into a four-dimensional multiport device realization of a Grover matrix.
This Grover operator, the multiport, described by the unitary matrix
\begin{equation}
\label{eqn:Grover}
Grover =
\frac{1}{2}
\begin{pmatrix}
-1&1&1&1\\
1&-1&1&1\\
1&1&-1&1\\
1&1&1&-1
\end{pmatrix},
\end{equation}
has equal splitting ratios between all input-output combinations and generalizes the BS transformation matrix given below in Eq.\eqref{eqn:BS}.
In general, photons in modes $a$ and $b$ are transformed in the following manner,
\begin{eqnarray}
&a \xrightarrow{M} \frac{1}{2}(-a+b+c+d)\; \mbox{ and }\;\\ \nonumber
&b \xrightarrow{M} \frac{1}{2}(a-b+c+d).
\end{eqnarray}
Theoretical analysis of the reversible Grover matrix has been performed by linear-optical directionally-unbiased four-ports \cite{simon2016group,simon2018joint,osawa2019directionally}, which consist of four beam splitters, four phase shifters, and four mirrors as indicated in Fig. \ref{fig:combined_four_port}(a).
They are represented schematically by the symbol in Fig. \ref{fig:combined_four_port}(b).
The three-port version of this device has been experimentally demonstrated using bulk optical devices \cite{osawa2018experimental}.
To have better and precise control of phases, miniaturization of the device is highly preferred to realize the four-port especially when several multiport devices are required to carry out an experiment.
In general, directional unitary devices such as those of the Reck and some other unitary matrix decomposition models \cite{reck1994experimental,su2019hybrid,clements2016optimal,de2018simple,motes2014scalable} can also realize a Grover matrix.
However, directionally-unbiased devices are advantageous when designing the delayed HOM effect, as well as requiring fewer optical resources.
Identical photons are sent into two of the four input-output ports from the left side (indicated in Fig. \ref{fig:combined_four_port}(b)).
We used multiport devices and beam splitters to form two systems for state propagation.
The photons are sent from the left side of the system through out the manuscript.
The first BS multiport composite system is denoted as subscript 0 and the other half is denoted as subscript 1.
The result differs depending on the input location of photons.
Consider a system consisting of two multiports and two beam splitters.
There are several ways to insert photons in the system, however we choose two specific ones in this manuscript.
To be able to send a photon into the middle of the system, the setup needs to be supplied with circulators, is shown in Fig. \ref{fig:multiport_circ}.
Another setup requires no circulators to propagate input photons.
The photons experience an extra transformation by a beam splitter upon photon entrance.
The system is graphically supplied in Fig. \ref{fig:multiport_no_circ}.
It needs to be noted that the number of multiports in the system does not change the final outcome.
We are using two multiports as an example, however, the result is the same when the system has a single multiport or more than two multiports as long as the devices are assumed to be lossless during the propagation.
Brief comments on the mathematical structure of the transformations carried out by the configurations given in Figs. \ref{fig:multiport_circ} and \ref{fig:multiport_no_circ} are given in the Appendix.
\vspace{-15px}
\subsection{Photon propagation using circulators}
\vspace{-10px}
\begin{figure}
\vspace{-5mm}
\centering
\includegraphics[width=3.5in]{multiport_two_circulator.pdf}
\caption{A system setup with input photons supplied by circulators. The system consists of two beam splitters, two multiport devices, and two circulators. These circulators allow us to send photons from the left side of the multiport device without experiencing a beam splitter transformation before entering the multiport device. The input state split into right moving and left moving amplitudes (shown as dotted arrows) upon multiport transformation.}
\label{fig:multiport_circ}
\end{figure}
This method is used to distribute HOM pair between the right and the left side of the system.
The original input state $a_0b_0$ transforms to:
\begin{eqnarray} \label{eq:trans}
a_0b_0 &\xrightarrow{M} \frac{1}{2}(-a_0+b_0+c_0+d_0)\frac{1}{2}(a_0-b_0+c_0+d_0) \nonumber \\
&=-\frac{1}{4}(a_0^2+b_0^2)+\frac{1}{2}a_0b_0+\frac{1}{4}(c_0^2+d_0^2)+\frac{1}{2}c_0d_0\nonumber \\
&= -\frac{1}{4}(a_0-b_0)^2+\frac{1}{4}(c_0+d_0)^2,
\end{eqnarray}
where we have used the commutation relation $ab = ba$ since the photons are identical and in different spatial locations.
Eq.\eqref{eq:trans} shows that correlated photons are split into right moving $\frac{1}{4}(c_0+d_0)^2$ and left moving $-\frac{1}{4}(a_0-b_0)^2$ amplitudes, with no cross terms.
This absolute separation of propagation direction without mixing of right moving and left moving amplitudes is important because the photon pairs remain distinctly localized and clustered at each step \cite{simon2020quantum}.
The right moving amplitude is translated to $\frac{1}{4}(a_1+b_1)^2$ and propagates without changing its form.
$\frac{1}{4}(a_1+b_1)^2 \xrightarrow{M} \frac{1}{4}(c_1+d_1)^2$.
The left moving amplitude $-\frac{1}{4}(a_0-b_0)^2$ stays the same until BS transformation.
The controlled HOM effect can be observed in higher-dimensional multiports assisted by extra beam splitters.
Imagine beam splitters inserted in the system as in Fig. \ref{fig:multiport_circ}.
Input state $ab$ is now transformed into $-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2$ as indicated above, then further transformed by beam splitters to obtain HOM pairs between the right side and left side of the system.
The right and left sides of the system each have two output ports, and the exit port of the photon pair can be controlled by varying phase shift settings before the beam splitters.
A phase shift on the left side of the system does not affect the result of the right side amplitude, and vice versa.
This system, having circulators at the beginning of the system, is denoted as transformation pattern I, and the detailed discussions of its transformation are in Sec. \ref{sec:pattern_I}.
\vspace{-15px}
\subsection{Photon propagation without circulators}
\vspace{-10px}
\begin{figure}
\vspace{-5mm}
\centering
\includegraphics[width=3.5in]{multiport_two.pdf}
\caption{System setup without circulators. The input photons are subjected to a beam splitter before they enter the multiport. The input state is transformed and propagated in one direction (shown as dotted arrows). The BS transformed input state is transformed again by the first multiport devices.}
\label{fig:multiport_no_circ}
\end{figure}
This method allows to redistribute input states between right and left side of the system without changing amplitudes.
Consider sending two photons from the left side of the beam splitter as indicated in fig. \ref{fig:multiport_no_circ} then transform the output state by the multiport device.
We only consider the first multiport transformation here.
The rest of the transformation is given in sec. \ref{sec:pattern_II}.
\begin{equation}
e_0f_0\xrightarrow{BS}-\frac{1}{2}(a_0^2-b_0^2)
\xrightarrow{M}-\frac{1}{2}(a_0-b_0)(c_0+d_0)
\end{equation}
The final state has cross-terms, and it is different from the case with circulators in a sense that the output state is \textit{coupled}.
The state does not provide clear separation between right moving and left moving amplitudes.
Even though, the state does not have clear distinction between right moving and left moving, we still refer the amplitudes right and left moving amplitudes unless special attention is required.
This system having no circulators is denoted as transformation pattern II, and the detailed discussions of its transformation are in Sec. \ref{sec:pattern_II}
\vspace{-10px}
\section{Transformation pattern I: directionally-controllable HOM effect in higher-dimensional spatial and temporal modes} \label{sec:pattern_I}
\vspace{-10px}
In this section we discuss the transformation pattern I.
The higher dimensional HOM effect is generated by the multiport-based linear optics system with circulators at the inputs.
The propagation direction control and delays between amplitudes are discussed in subsections.
We use a single multiport device to show the control effect and we introduce two multiport devices in the system for delayed effect.
\begin{figure}[htp]
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator1.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator2.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator3.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator4.pdf}%
}
\caption{Higher dimensional HOM effect with directional control.
Correlated photons, $a_0b_0$, are sent in from the circulators into the first multiport.
After the first multiport interaction, the incoming photon pair splits into right-moving and left-moving two-photon amplitudes.
The separately-moving amplitudes are bunched at the beam splitters on right and left sides.
We can controllably switch between four different output sites, and where the clustered output photons appear depends on the location of the phase shifter $P$.
In (a), no phase plates are introduced, and the output biphoton amplitudes leave $f_0$ and $e_1$. The final state is $\frac{1}{\sqrt{2}}(\ket{0,2}_0+\ket{2,0}_1)$, meaning superposition of two photons in mode $f_0$ and two in mode $e_1$. In case (b), the phase shifter $P=\pi$ is to the left, changing the relative phase between upper and lower arms.
Similarly in (c) and (d), other locations for the phase shifters cause biphotons to leave in other spatial modes.} \label{fig:HOM_control_fig}
\end{figure}
\vspace{-15px}
\subsection{Control of propagation direction}
\vspace{-10px}
Given that one two-photon amplitude must exit left and one right, there are four possible combinations of outgoing HOM pairs as indicated in Fig. \ref{fig:HOM_control_fig}.
The combinations are, (a): $(f_0^2,e_1^2)$, (b): $(e_0^2,e_1^2)$, (c): $(e_0^2,f_1^2)$, and (d): $(f_0^2,e_1^2)$.
This means, in the case of (a) for example, the left-moving two-photon amplitude leaves in mode f, and the right-moving amplitude leaves in mode e.
Directional control of the four cases is readily demonstrated, as follows. In case (a) there is only a beam splitter transformation after the multiport, giving
\begin{eqnarray}
&-\frac{1}{4}(a_0-b_0)^2\xrightarrow{BS} -\frac{1}{4\sqrt{2}}(e_0-f_0-e_0-f_0)^2 = \frac{1}{2}f_0^2, \nonumber \\
&\frac{1}{4}(c_1+d_1)^2 \xrightarrow{BS} -\frac{1}{4\sqrt{2}}(e_1-f_1+e_1+f_1)^2 = \frac{1}{2}e_1^2.
\end{eqnarray}
The final output state is,
\begin{eqnarray}
\frac{1}{2}(f_0^2+e_1^2)=\frac{1}{\sqrt{2}}(\ket{0,2}_0+\ket{2,0}_1).
\end{eqnarray}
In case (b), a phase plate is inserted in the lower arm of the left side to switch the exit port from $d$ to $c$. All the phase shifters P are set to $\pi$, therefore transforming $b \rightarrow -b$.
\begin{eqnarray}
&-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2 \xrightarrow{P} -\frac{1}{4}(a+b)^2+\frac{1}{4}(c+d)^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(-e_0^2+e_1^2) = \frac{1}{\sqrt{2}}(-\ket{2,0}_0+\ket{2,0}_1).
\end{eqnarray}
Compared to case (a), the exit port is switched from f to e. In (c), phase plates are inserted in the lower arms of both right and left sides.
Photons in modes $b$ and $d$ are transformed to $-b$ and $-d$, respectively.
\begin{eqnarray}
&-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2 \xrightarrow{P} -\frac{1}{4}(a+b)^2+\frac{1}{4}(c-d)^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(-e_0^2-f_1^2) = -\frac{1}{\sqrt{2}}(\ket{2,0}_0+\ket{0,2}_1).
\end{eqnarray}
In (d), a phase plate is inserted in the lower arm of the right side. A photon in mode $d$ is transformed to $-d$.
\begin{eqnarray}
&-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2 \xrightarrow{P} -\frac{1}{4}(a-b)^2+\frac{1}{4}(c-d)^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(f_0^2-f_1^2) = \frac{1}{\sqrt{2}}(\ket{0,2}_0-\ket{0,2}_1).
\end{eqnarray}
This demonstrates complete directional control of biphoton propagation direction using only linear optical devices.
Directional control does not require changing splitting ratios at each linear optical device (BS and multiport), and occurs in a lossless manner since no post-selection is required.
\begin{figure*}[htp]
\centering
\subfloat[][Delayed HOM effect without reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_circulator_arrays.pdf}\label{}}
\subfloat[][Delayed HOM effect with reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_circulator_arrays_phase_plate.pdf}\label{}}
\caption{Delayed HOM effect. The two-photon amplitude transformation progresses in time from top to bottom. The distance traveled in a single time step is indicated by vertical dashed lines. The original photons as well as photons in the target state are indicated using red circles. The green striped circles indicate intermediate transformed state. The total number of photons are always two through out the transformations. (a) Two multiports and beam splitters {\it without} phase shifters between the multiports. At the first step, the behavior is the same as for a single multiport with beam splitters. The right-moving amplitude propagates through the second multiport, and left-moving amplitude propagates through the beam splitter. The right moving amplitude is delayed by one additional multiport transformation before a two-photon observation probability will become available in spatial modes on the right. (b) Two multiports and beam splitters {\it with} a phase shifter P set at $\pi$ between multiports. When the P is present, the right-moving amplitude gains a relative phase between modes $
a_1$ and $b_1$. Reflection occurs at the multiport when the relative phase between the two is $\pi$. Therefore, the transformed amplitude reflects upon a second multiport encounter, going back to the original state with opposite propagation direction. Reflection does not occur on this transformed left-moving amplitude, therefore it continues to propagate leftward. The original left-moving amplitude becomes available for detection earlier than the transformed left-moving amplitude.}
\label{fig:delayed_HOM}
\end{figure*}
\subsection{Delayed HOM effect}
\vspace{-10px}
\subsubsection{Delayed HOM effect without reflection}
\vspace{-10px}
We introduce a phase shifter between two multiports as in Fig. \ref{fig:delayed_HOM} (b). Without the phase plate between two multiport devices, the photons behave exactly the same as in the previous subsection. However, the phase shifter can change propagation direction of right moving amplitude to the left. This reflection results in detecting HOM pairs only on the left side, but with some delay between the two exiting amplitudes. We start with the case without the phase shifter. The photon insertion is the same as the previous case, coming from the left side of the first multiport.
\vspace{-10px}
\begin{eqnarray}
&{a_{0}b_{0}}_R\xrightarrow{M}-\frac{1}{4}(a_{0}-b_{0})_L^2+ \frac{1}{4}(c_{0}+d_{0})_R^2 \nonumber \\
&\xrightarrow{T+BS,\mbox{ }T} -\frac{1}{2}f_{0L}^2 + \frac{1}{4}(a_{1}+b_{1})_R^2 \nonumber \\
&\xrightarrow{M} -\frac{1}{2}f_{0L}^2 + \frac{1}{4}(c_{1}+d_{1})_R^2 \xrightarrow{BS} -\frac{1}{2}f_{0L}^2 + \frac{1}{2}e_{1R}^2,
\end{eqnarray}
where M, T, BS represents multiport, translation and beam splitter transformation respectively. We use subscript $R$ and $L$ to illustrate amplitudes propagating to the right or left. $T$ translates a photon amplitude by a single time step (for example, $\frac{1}{4}(c_{0}+d_{0})^2 \rightarrow \frac{1}{4}(a_{1}+b_{1})^2$). The second transformation $T+BS$, $T$ is read as applying $T+BS$ on the first term and $T$ on the second term.
The final state is,
\begin{eqnarray}
-\frac{1}{2}f_{0L}^2 + \frac{1}{2}e_{1R}^2=-\frac{1}{\sqrt{2}}(\ket{0,2}_{0{T_0}L}-\ket{2,0}_{1{T_1}R}),
\end{eqnarray}
where $T_0$ is the time when the first biphoton amplitude leaves the system and $T_1$ is the exit time of the second.
The right moving amplitude stays in the system longer than the left moving amplitude because of the extra multiport device in the system, leading to time delay $\Delta T = T_1-T_0$.
\vspace{-10px}
\subsubsection{Delayed HOM effect with reflection}
\vspace{-10px}
When a $\pi$-phase shifter is inserted on one path between the multiports, the right-moving amplitude gets reflected upon the second multiport encounter.
Instead of having two-photon amplitudes on the right and left sides of the system, both photon amplitudes end up leaving from the left.
The HOM effect still occurs but now with some delay between the two amplitudes at the end of the BS.
This is indicated in Fig. \ref{fig:delayed_HOM} (b).
\vspace{-10px}
\begin{eqnarray}
&{a_{0}b_{0}}_R\xrightarrow{M}-\frac{1}{4}(a_{0}-b_{0})_L^2+ \frac{1}{4}(c_{0}+d_{0})_R^2 \nonumber \\
&\xrightarrow{T+BS,\mbox{ }T+P} -\frac{1}{2}f_{0L}^2 + \frac{1}{4}(a_{1}-b_{1})_R^2.
\end{eqnarray}
The second transformation $T+BS$, $T+P$ is read as applying $T+BS$ on the first term and $T+P$ on the second term.
Left-moving photons leave before right-moving photons.
\begin{eqnarray}
\xrightarrow{M}&\frac{1}{4}(a_{1}-b_{1})_L^2 \xrightarrow{P+T} \frac{1}{4}(c_{0}+d_{0})_L^2 \nonumber \\
&\xrightarrow{M} \frac{1}{4}(a_{0}+b_{0})_L^2 \xrightarrow{BS} \frac{1}{2}e_{0L}^2.
\end{eqnarray}
The final state,
\begin{eqnarray}
-\frac{1}{2}f_{0L}^2+\frac{1}{2}e_{0L}^2=-\frac{1}{\sqrt{2}}(\ket{0,2}_{0T_0L}-\ket{2,0}_{0T_2L}),
\end{eqnarray}
is now two HOM pair amplitudes, both on the left side of the system, at output ports $e_0$ and $f_0$, with some time delay $\Delta T = T_2 - T_0$ between them. The first amplitude leaves port $f_0$ at $T_0$, then the second leaves $e_0$ and the time labeled $T_2$.
\vspace{-10px}
\section{Transformation pattern II: state redistribution in higher-dimensional spatial and temporal modes} \label{sec:pattern_II}
\vspace{-10px}
\subsection{State transformation and propagation}
\vspace{-10px}
We have considered the case where the input photon state is transformed by the multiport device right after photon insertion in the previous section.
Instead of using circulators, we can transform the input state by the BS in advance and then transform the state by using the multiport device.
Even though the Grover matrix spreads the input state equally in four directions, the end result preserves the original form of the input state.
We demonstrate a state redistribution property using distinguishable and indistinguishable photons, meaning the input state gets redistributed between right and left side without changing amplitudes.
The propagation result is different from the previous case.
Consider sending two indistinguishable photons in the system.
The input two photons have the same polarization to make them indistinguishable.
The input photons are inserted from the left side of the beam splitter.
The beam splitter transforms the input state and propagates from the left side to right side of the device without any reflections.
The amplitudes are transformed by the multiport device after the beam splitter transformation.
This transformation splits input photons into coupled right- moving and left-moving amplitudes.
The coupled left moving amplitudes reflected from the first multiport counter propagates and transformed by the first beam splitter from the right to the left.
The right moving amplitude is transmitted without changes in amplitude.
This amplitude gets transmitted by the right side beam splitter at the end.
\vspace{-15px}
\subsubsection{Indistinguishable photons}
\vspace{-10px}
We examine the mathematical details on indistinguishable photons in the system without circulators first.
We consider three cases by sending photons in spatial modes e and f.
First, we consider indistinguishable a pair of single photons from spatial mode e and f.
\vspace{-5px}
\begin{eqnarray}
&e_{H0}f_{H0}\xrightarrow{BS}-\frac{1}{2}(a_{H0}^2-b_{H0}^2)\nonumber \\
&\xrightarrow{M} -\frac{1}{2}(a_{H0}-b_{H0})(c_{H0}+d_{H0}) \nonumber \\
&\xrightarrow{BS} -e_{H0}e_{H1}.
\end{eqnarray}
HOM state with relative phase between two amplitudes equal to +1 is considered here.
\begin{eqnarray}
&\frac{1}{2}(e_{H0}^2+f_{H0}^2)\xrightarrow{BS}\frac{1}{2}(a_{H0}^2+b_{H0}^2) \nonumber \\
&\xrightarrow{M} \frac{1}{4}(a_{H0}-b_{H0})^2+\frac{1}{4}(c_{H0}+d_{H0})^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(e_{H0}^2+f_{H1}^2).
\end{eqnarray}
The input state is redistributed in a sense that one amplitude is on the right side of the system and the other amplitude is on the left side while maintaining the original structure of the state.
HOM state with relative phase between two amplitudes equal to -1 is considered here.
\vspace{-5px}
\begin{eqnarray}
&\frac{1}{2}(e_{H0}^2-f_{H0}^2) \xrightarrow{BS} -a_{H0}b_{H0} \nonumber \\
&\xrightarrow{M} \frac{1}{4}(a_{H0}-b_{H0})^2-\frac{1}{4}(c_{H0}+d_{H0})^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(e_{H0}^2-e_{H1}^2).
\end{eqnarray}
In both cases, the output state is identical to the input state except for the spatial modes.
\vspace{-10px}
\subsubsection{Distinguishable photons}
\vspace{-10px}
Now, we examine the case of distinguishable two photon input.
The procedure is identical to the the previous case.
We begin with two distinguishable photons at each modes without superposition.
\begin{eqnarray}
&e_{H0}f_{V0} \xrightarrow{BS} \frac{1}{2}(a_{H0}-b_{H0})(a_{V0}+b_{V0})\nonumber \\
&\xrightarrow{M} -\frac{1}{2}(a_{H0}-b_{H0})(c_{V0}+d_{V0})\nonumber\\
&\xrightarrow{BS} - e_{H0} e_{V1}.
\end{eqnarray}
We examine the case of HOM states.
\begin{eqnarray}
&\frac{1}{2}(e_{H0}^2 \pm f_{V0}^2) \xrightarrow{BS}\frac{1}{4}(a_{H0}-b_{H0})^2\pm\frac{1}{4}(a_{V0}+b_{H0})^2\nonumber\\
&\xrightarrow{M}\frac{1}{4}(a_{H0}-b_{H0})^2\pm\frac{1}{4}(c_{V0}+d_{H0})^2\nonumber\\
&\xrightarrow{BS} \frac{1}{2}(e_{H0}^2 \pm e_{V1}^2).
\end{eqnarray}
The control of exit location can be performed as well in this scheme by introducing phase shifters in the system as indicated in Fig. \ref{fig:control_without_circ}.
This procedure does not destroy the redistribution property.
There are four potential spatial modes and by switching the phase shift before beam splitters, the direction of propagation switches.
The combinations are, (a): $(e_0,f_0)\rightarrow(e_0,e_1)$, (b): $(e_0,f_0)\rightarrow(e_0,f_1)$, (c): $(e_0,f_0)\rightarrow(f_0,e_1)$, and (d): $(e_0,f_0)\rightarrow(f_0,f_1)$.
The result from the system with circulators is summarized in Table. \ref{tab:table_1}, the system without them is in Table. \ref{tab:table_2}.
In the case of indistinguishable photons, the results are cyclic in a sense that all three states can be produced by using the other system.
However, there is a significant difference when distinguishable photons are considered.
\begin{figure}[htp]
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator2.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator3.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator4.pdf}%
}
\caption{Quantum state redistribution with control of propagation direction. We performed the same analysis as the higher dimensional HOM effect with direction control. By introducing phase shifters in the system before beam splitters, we can change the exit direction of the amplitudes. The starting state is $e_0f_0$. The first beam splitter transforms the input state, then they enter the multiport device. The multiport transformed state goes through beam splitters on the right and left side. The final outcome has the same form as the input state.}
\label{fig:control_without_circ}
\end{figure}
\begin{table}
\begin{tabular}{P{0.4\linewidth}P{0.6\linewidth}}
\hline \hline
\noalign{\vskip 1ex}
\multicolumn{2}{c}{State transformation with circulators}\\
\noalign{\vskip 1ex}
\hline \hline
\noalign{\vskip 1ex}
Indistinguishable photons & $a_{H0}b_{H0} \rightarrow -\frac{1}{2}(e_{H0}^2 - e_{H1}^2)$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with +1 relative phase & $\frac{1}{2}(a_{H0}^2 + b_{H0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 + e_{H1}^2)$\\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with $-1$ relative phase & $\frac{1}{2}(a_{H0}^2 - b_{H0}^2) \rightarrow -e_{H0}e_{H1}$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable photons & $a_{H0}b_{V0} \rightarrow -\frac{1}{2}(e_{H0}-e_{H1})(e_{V0}+e_{V1})$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable HOM pair & $\frac{1}{2}(a_{H0}^2 \pm b_{V0}^2) \rightarrow \frac{1}{4}\{(e_{H0}-e_{H1})^2\pm(e_{V0}-e_{V1})^2\}$ \\
\noalign{\vskip 1ex}
\hline \hline
\end{tabular}
\caption{State transformations in a system with circulators.
The first three states deal with indistinguishable photons by giving them the same polarization.
A state consisting of two single photons will become an HOM state.
We analyzed HOM states as an initial state, and they become either the HOM state or a two single-photon state.
Distinguishable photons are also analyzed by introducing orthogonal polarizations.
The output states become coupled states meaning the original states are not preserved.
} \label{tab:table_1}
\end{table}
\begin{table}
\begin{tabular}{P{0.4\linewidth}P{0.6\linewidth}}
\hline \hline
\noalign{\vskip 1ex}
\multicolumn{2}{c}{State transformation without circulators} \\
\noalign{\vskip 1ex}
\hline \hline
\noalign{\vskip 1ex}
Indistinguishable photons & $e_{H0}f_{H0} \rightarrow -e_{H0}e_{H1}$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with +1 relative phase & $\frac{1}{2}(e_{H0}^2 + f_{H0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 + e_{H1}^2)$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with $-1$ relative phase & $\frac{1}{2}(e_{H0}^2 - f_{H0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 - f_{H1}^2)$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable photons & $e_{H0}f_{V0} \rightarrow -e_{H0}e_{V1}$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable HOM pair & $\frac{1}{2}(e_{H0}^2 \pm f_{V0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 \pm f_{V1}^2)$ \\
\noalign{\vskip 1ex}
\hline \hline
\end{tabular}
\caption{State transformations in a system without circulators.
The structure of the table is the same as the Table. I.
The first three states deal with indistinguishable photons by giving them the same polarization.
The last two states handle distinguishable photons.
The output states preserve the same form as the input state.
We start the transformation from the system location 0, then the transformed states are redistributed between location 0 and location 1.
The result shows coherent transportation of input states.} \label{tab:table_2}
\end{table}
\vspace{-10px}
\subsection{Delayed state redistribution}
\vspace{-10px}
\begin{figure*}[htp]
\centering
\subfloat[][State redistribution without reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_array.pdf}\label{}}
\subfloat[][State redistribution with reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_array_phase_plate.pdf}\label{}}
\caption{Delayed state redistribution.
The two-photon amplitude transformation progresses in time from top to bottom.
The distance traveled in a single time step is indicated by vertical dashed lines. The total photon numbers are two in the system through out the propagation. At the first step for both cases, the input two-photon state is transformed by the BS.
The transformed state becomes the HOM state, and it is indicated as red transparent overlapped circles occupying both modes.
The initial and the final transformed state are indicated using solid red circles, and intermediate states are indicated in striped yellow circles.
(a) Two multiports and beam splitters {\it without} phase shifters between the multiports.
The HOM state enters the multiport and transformed taking the form of $-\frac{1}{2}(a_{0}-b_{0})(c_{0}+d_{0})$.
The amplitudes are coupled, however, they propagate without changing its amplitude. After several steps, the amplitudes occupying two rails converges to a single mode state after transformation by beam splitters.
The final state has the same form as the input state.
(b) Two multiports and beam splitters {\it with} a phase shifter P set at $\pi$ between multiports.
When the P is present, the right-moving coupled amplitude gains a relative phase between modes $a_1$ and $b_1$.
Reflection occurs at the multiport when the relative phase between the two is $\pi$.
Therefore, the transformed amplitude reflects upon a second multiport encounter, going back to the original state with opposite propagation direction.
Reflection does not occur on this transformed coupled left-moving amplitude, therefore it continues to propagate leftward.
The original left-moving amplitude becomes available for detection earlier than the transformed left-moving amplitude.}
\label{fig:delayed_state_dist}
\end{figure*}
We introduce the temporal delay effect as the higher dimensional HOM case by introducing a phase shifter between two multiports.
\vspace{-10px}
\subsubsection{Without reflection}
\vspace{-10px}
When there is no phase shifter between the two multiports, the result is identical to the system with a single multiport from the previous section.
The state transformation and propagation is provided schematically in Fig. \ref{fig:delayed_state_dist} (a).
The photons are initially sent from the left side of the BS.
The correlated photons are transformed to HOM state through the BS.
\vspace{-10px}
\begin{eqnarray}
&{e_{0}f_{0}}_R \xrightarrow{BS} \frac{1}{2}(a_{0}^2-b_{0}^2)_R \nonumber\\
&\xrightarrow{M}-\frac{1}{2}(a_{0}-b_{0})_L(c_{0}+d_{0})_R.
\end{eqnarray}
The HOM state is transformed by the multiport device.
This state is in a coupled state because right moving and left moving amplitudes are not separated.
We propagate this state through the BS on the left and translate the amplitudes moving to the right.
\begin{eqnarray}
&\xrightarrow{BS,T}-\frac{1}{\sqrt{2}}f_{0L}(a_{1}+b_{1})_R \xrightarrow{M} -\frac{1}{\sqrt{2}}f_{0L}(c_{1}+d_{1})_R\nonumber\\
&\xrightarrow{BS} -f_{0T_0L}e_{1T_1R}
\end{eqnarray}
The left moving amplitude is transformed by the left BS while right moving amplitude propagates to the second multiport device.
We introduced temporal difference between the right moving and the left moving photons.
\vspace{-15px}
\subsubsection{With reflection}
\vspace{-10px}
Reflection of amplitudes are introduced when there is a phase shifter between two multiport devices as indicated in fig. \ref{fig:delayed_state_dist} (b).
\begin{eqnarray}
&{e_{0}f_{0}}_R \xrightarrow{BS} \frac{1}{2}(a_{0}^2-b_{0}^2)_R
\xrightarrow{M} -\frac{1}{2}(a_{0}-b_{0})_L(c_{0}+d_{0})_R \nonumber \\
&\xrightarrow{BS,T+P}-\frac{1}{\sqrt{2}}f_{0L}(a_{1}-b_{1})_R
\end{eqnarray}
The right moving amplitude gains relative phase between upper and lower rails, and this relative phase allows the amplitude to get reflected upon multiport encounter.
\begin{eqnarray}
&\xrightarrow{M} -\frac{1}{\sqrt{2}}f_{0L}(a_{1}-b_{1})_L \xrightarrow{T+P} -\frac{1}{\sqrt{2}}f_{0L}(c_{0}+d_{0})_L\nonumber\\
&\xrightarrow{M} -\frac{1}{\sqrt{2}}f_{0L}(a_{0}+b_{0})_L \xrightarrow{BS} -f_{0T_{0}L} e_{0T_{2}L}
\end{eqnarray}
The input photons do not have any delays between the two at the beginning. The delay $\Delta T = T_2 - T_0$ is introduced from the reflection in the system.
\vspace{-10pt}
\section{Conclusion}
\vspace{-10px}
We demonstrated higher dimensional quantum state manipulation such as the HOM effect and state redistribution by applying linear-optical four-ports realizing four-dimensional Grover matrix accompanied by beam splitters and phase shifters.
Identical photons are sent into two of the four input-output ports and split into right-moving and left-moving amplitudes, with no cross terms to observe the HOM effect.
This absolute separation of propagation direction without mixing of right-moving and left-moving amplitudes insures the photons remain clustered as they propagate through the system.
Variable phase shifts in the system allow the HOM photon pairs to switch between four spatial output destinations, which can increase information capacity.
Time delays between emerging parts of the clustered two-photon state illustrating “delayed” HOM effect can be engineered using two multiports.
In addition, depending on the phase shifter position, the propagation direction can be reversed so that the right moving amplitude can get reflected at the second multiport, resulting in HOM pairs always leaving only from the left side of the system and with a particular time-bin delay.
The same situations have been investigated in a system without circulators. This system allows to redistribute the input state between the right and the left side of the system without changing amplitudes.
The HOM effect and clustered photon pairs are widely used in quantum information science. The approach introduced here adds extra degrees of freedom, and paves the way for new applications that require control over the spatial and temporal modes of the HOM amplitudes as they move through one- and two-dimensional networks.
We have demonstrated two photon amplitude control in both spatial and temporal modes. This two photon system can be extended to multiphoton input states, and manipulation of more complex entangled states would be the next milestones to be achieved.
\vspace{-15pt}
\section*{Appendix}
\section{\label{sec:level1}Introduction}
\begin{figure}
\vspace{-5mm}
\centering
\includegraphics[width=3in]{HOM_cancelling_4.pdf}
\caption{Hong-Ou-Mandel effect. Two identical photons are sent into different beam splitter input ports. There are four possible outcomes, two photons leaving one port, two photons leaving the other port, each photon reflecting to give single photons at each exit, and both transmitting to give single photons at each port. The coincidence terms cancel out since they are identical but enter with opposite sign. The final state is a superposition of two outcomes, each with both photons clustered together at the same exit port.}
\label{fig:HOM_effect}
\end{figure}
\section{Introduction}
\vspace{-10px}
The Hong-Ou-Mandel (HOM) effect is one of the most recognized quantum two-photon interference effects \cite{hong1987measurement}.
When two indistinguishable photons arrive simultaneously at different inputs of a 50:50 beam splitter (BS), single-photon amplitudes at each output cancel, resulting in quantum superposition of two-photon states appearing at each output port, as in Fig. \ref{fig:HOM_effect}.
This traditional HOM method, observed on a BS having two input and two output ports, always has the two-photon state simultaneously occupying both output spatial modes, leaving no room to engineer control of propagation direction.
Various types of studies on quantum state transformations in multiport devices have been performed such as two photon propagation in a multimode system \cite{weihs1996two,zukowski1997realizable},
quantum interference effects using a few photons \cite{meany2012non,de2014coincidence,tichy2011four,campos2000three}, and propagation of multi-photons \cite{lim2005generalized,tillmann2015generalized,menssen2017distinguishability}. Internal degrees of freedom are also incorporated to enhance communication capacity \cite{walborn2003multimode,poem2012two,zhang2016engineering}.
Systems and procedures using multi-photon states, such as boson sampling, have been analyzed using multiport beam splitters both theoretically and experimentally \cite{aaronson2011computational,tillmann2013experimental,spring2013boson,bentivegna2015experimental,he2017time,wang2019boson}.
The HOM effect plays an important role in the field of quantum metrology when two-photon $|2002\rangle$-type states are extended to $N$-photon $N00N$ state \cite{dowling2008quantum,motes2015linear}.
Additionally, coherent transport of quantum states has been attracting attention, where single- and two-photon discrete-time quantum walk schemes are employed to transfer and process quantum states \cite{bose2003quantum,perez2013coherent,lovett2010universal,chapman2016experimental,nitsche2016quantum}.
A quantum routing approach has been proposed to transfer unknown states in 1D and 2D structures to assist quantum communication protocols \cite{zhan2014perfect,vstefavnak2016perfect,bartkiewicz2018implementation}.
Photon propagation control is especially crucial in a large optical network to distribute quantum states between two parties.
The network can be formed by combining multiple copies of four-port devices.
The state manipulation schemes we present can be integrated in quantum communication protocols since state retrieval timing can be chosen at will.
In this manuscript, we propose two-photon quantum state engineering and transportation methods with a linear-optical system which allows manipulation of photon amplitudes by using linear-optical devices such as optical multiports, beam splitters, and phase shifters.
Previously, such multiports have been introduced to demonstrate a two-photon clustering effect in quantum walks when multiple multiport devices are connected to form a chain \cite{simon2020quantum}. Clustering of two photons means that after encountering a multiport, the input two-photon amplitude separates into a superposition of a right-moving and a left-moving two-photon amplitude, with no amplitude for the photons to move in opposite directions.
By utilizing this separation, a higher-dimensional unitary transformation enables flexible quantum engineering designs of possible travel path combinations by switching relative phases within right moving and left moving amplitudes independently.
When two or more multiports are combined, this control of quantum amplitudes in a two-photon state allows demonstration of a “delayed" HOM effect engaging also time-bin modes in addition to spatial modes.
To perform this delayed effect, two or more multiports are required, and relative phase shifts between two rails can reflect the incoming amplitudes.
This controllable reflection without mirrors can also be seen as an additional state manipulation feature.
We introduce two distinct systems. The first system utilizes direct transformation of two-photons by the four-port device using circulators.
The second case does not have circulators in the system.
The photons are sent from the left side of the beam splitters, then the amplitudes encounter the multiport device.
This second system has not been analyzed in the past. Specific input states for both distinguishable and indistinguishable photons are redistributed between two parties coherently. Therefore, this system is particularly useful in quantum routing type applications.
This article is organized as follows. In Sec. II, we introduce the main optical components used in this manuscript to perform quantum state transformation. These basic linear optical devices are used to show HOM effect engaging in spatial modes and time-bin modes, and is addressed in Sec. III. In Sec. IV, we show redistribution of two photon states using the devices introduced that are presented in Sec. II. The summary of the results are given in Sec. V.
\begin{figure}[htp]
\subfloat[]{%
\includegraphics[clip,width=0.8\columnwidth]{multiport_systematic.pdf}%
}
\vspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=0.8\columnwidth]{multiport_systematic_photons.pdf}%
}
\vspace{-10px}
\caption{(a) A possible experimental realization of a directionally-unbiased linear-optical four-port consists of four beam splitters, four mirrors, and four phase shifters. A photon can enter any of the four ports, and exit at any of the four ports (labeled as $a,b,c$, and $d$). With a specific choice of phase settings, a Grover matrix can be realized by coherently summing all possible paths to each output \cite{osawa2019directionally}. A schematic symbol for this device is shown on the right. (b) Single multiport transformation of a two-photon input state. The input state of two correlated photons entering from the left is depicted as $ab$ $(\ket{1,1})$. After scattering by a Grover multiport, the state transforms into $-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2$, which has clear separation of right- and left-moving two-photon amplitudes. No cross-terms with photons moving in opposite directions occur.}
\label{fig:combined_four_port}
\vspace{-10pt}
\end{figure}
\section{Photonic state transformations via linear optical devices}
\vspace{-10px}
In this section, we consider photonic state transformations in higher-dimensional spatial modes using a unitary four-dimensional Grover matrix \cite{grover1996fast} in place of the beam splitter.
In this section, we introduce the main systems that will be used for linear state transformations, followed by the basic photonic devices to implement them.
Beam splitters and the four-dimensional Grover matrix are the central system component.
We mainly use photon number representation to describe states through out the manuscript.
The general beam splitter transformation matrix is
\begin{eqnarray}
\label{eqn:BS}
&\begin{pmatrix}
\hat{c}
\\
\hat{d}
\end{pmatrix}=
\frac{1}{\sqrt{2}}
\begin{pmatrix}
1 & 1\\
-1 & 1
\end{pmatrix}
\begin{pmatrix}
\hat{a}\\
\hat{b}
\end{pmatrix}
\end{eqnarray}
where $\hat{a}$,$\hat{b}$,$\hat{c}$, and $\hat{d}$ are used to describe the input photon state transformation.
The labels are generalized here, therefore the specific location dependent beam splitter transformations are redefined in later sections.
We use photon number states to describe the system unless otherwise specified. The input state is denoted as $\hat{a}\hat{b}$ where $\hat{a}$ and $\hat{b}$ are respectively creation operators for the spatial modes $a$ and $b$.
The hat notation is dropped henceforth.
For a photon in spatial mode $a$ with horizontally polarized photon is denoted as $a_{H}$ and horizontally polarized photon in mode $b$ is denoted as $b_{H}$.
We omit polarization degrees of freedom when identical photons are used through out the system.
Photonic implementations of the Grover matrix can be readily realized \cite{carolan2015universal,crespi2013anderson,spagnolo2013three,fan1998channel,nikolopoulos2008directional}.
To be concrete, we use directionally-unbiased linear-optical four-ports (Fig. \ref{fig:combined_four_port} (a)) as an example.
Consider sending two indistinguishable photons into a four-dimensional multiport device realization of a Grover matrix.
This Grover operator, the multiport, described by the unitary matrix
\begin{equation}
\label{eqn:Grover}
Grover =
\frac{1}{2}
\begin{pmatrix}
-1&1&1&1\\
1&-1&1&1\\
1&1&-1&1\\
1&1&1&-1
\end{pmatrix},
\end{equation}
has equal splitting ratios between all input-output combinations and generalizes the BS transformation matrix given below in Eq.\eqref{eqn:BS}.
In general, photons in modes $a$ and $b$ are transformed in the following manner,
\begin{eqnarray}
&a \xrightarrow{M} \frac{1}{2}(-a+b+c+d)\; \mbox{ and }\;\\ \nonumber
&b \xrightarrow{M} \frac{1}{2}(a-b+c+d).
\end{eqnarray}
Theoretical analysis of the reversible Grover matrix has been performed by linear-optical directionally-unbiased four-ports \cite{simon2016group,simon2018joint,osawa2019directionally}, which consist of four beam splitters, four phase shifters, and four mirrors as indicated in Fig. \ref{fig:combined_four_port}(a).
They are represented schematically by the symbol in Fig. \ref{fig:combined_four_port}(b).
The three-port version of this device has been experimentally demonstrated using bulk optical devices \cite{osawa2018experimental}.
To have better and precise control of phases, miniaturization of the device is highly preferred to realize the four-port especially when several multiport devices are required to carry out an experiment.
In general, directional unitary devices such as those of the Reck and some other unitary matrix decomposition models \cite{reck1994experimental,su2019hybrid,clements2016optimal,de2018simple,motes2014scalable} can also realize a Grover matrix.
However, directionally-unbiased devices are advantageous when designing the delayed HOM effect, as well as requiring fewer optical resources.
Identical photons are sent into two of the four input-output ports from the left side (indicated in Fig. \ref{fig:combined_four_port}(b)).
We used multiport devices and beam splitters to form two systems for state propagation.
The photons are sent from the left side of the system through out the manuscript.
The first BS multiport composite system is denoted as subscript 0 and the other half is denoted as subscript 1.
The result differs depending on the input location of photons.
Consider a system consisting of two multiports and two beam splitters.
There are several ways to insert photons in the system, however we choose two specific ones in this manuscript.
To be able to send a photon into the middle of the system, the setup needs to be supplied with circulators, is shown in Fig. \ref{fig:multiport_circ}.
Another setup requires no circulators to propagate input photons.
The photons experience an extra transformation by a beam splitter upon photon entrance.
The system is graphically supplied in Fig. \ref{fig:multiport_no_circ}.
It needs to be noted that the number of multiports in the system does not change the final outcome.
We are using two multiports as an example, however, the result is the same when the system has a single multiport or more than two multiports as long as the devices are assumed to be lossless during the propagation.
Brief comments on the mathematical structure of the transformations carried out by the configurations given in Figs. \ref{fig:multiport_circ} and \ref{fig:multiport_no_circ} are given in the Appendix.
\vspace{-15px}
\subsection{Photon propagation using circulators}
\vspace{-10px}
\begin{figure}
\vspace{-5mm}
\centering
\includegraphics[width=3.5in]{multiport_two_circulator.pdf}
\caption{A system setup with input photons supplied by circulators. The system consists of two beam splitters, two multiport devices, and two circulators. These circulators allow us to send photons from the left side of the multiport device without experiencing a beam splitter transformation before entering the multiport device. The input state split into right moving and left moving amplitudes (shown as dotted arrows) upon multiport transformation.}
\label{fig:multiport_circ}
\end{figure}
This method is used to distribute HOM pair between the right and the left side of the system.
The original input state $a_0b_0$ transforms to:
\begin{eqnarray} \label{eq:trans}
a_0b_0 &\xrightarrow{M} \frac{1}{2}(-a_0+b_0+c_0+d_0)\frac{1}{2}(a_0-b_0+c_0+d_0) \nonumber \\
&=-\frac{1}{4}(a_0^2+b_0^2)+\frac{1}{2}a_0b_0+\frac{1}{4}(c_0^2+d_0^2)+\frac{1}{2}c_0d_0\nonumber \\
&= -\frac{1}{4}(a_0-b_0)^2+\frac{1}{4}(c_0+d_0)^2,
\end{eqnarray}
where we have used the commutation relation $ab = ba$ since the photons are identical and in different spatial locations.
Eq.\eqref{eq:trans} shows that correlated photons are split into right moving $\frac{1}{4}(c_0+d_0)^2$ and left moving $-\frac{1}{4}(a_0-b_0)^2$ amplitudes, with no cross terms.
This absolute separation of propagation direction without mixing of right moving and left moving amplitudes is important because the photon pairs remain distinctly localized and clustered at each step \cite{simon2020quantum}.
The right moving amplitude is translated to $\frac{1}{4}(a_1+b_1)^2$ and propagates without changing its form.
$\frac{1}{4}(a_1+b_1)^2 \xrightarrow{M} \frac{1}{4}(c_1+d_1)^2$.
The left moving amplitude $-\frac{1}{4}(a_0-b_0)^2$ stays the same until BS transformation.
The controlled HOM effect can be observed in higher-dimensional multiports assisted by extra beam splitters.
Imagine beam splitters inserted in the system as in Fig. \ref{fig:multiport_circ}.
Input state $ab$ is now transformed into $-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2$ as indicated above, then further transformed by beam splitters to obtain HOM pairs between the right side and left side of the system.
The right and left sides of the system each have two output ports, and the exit port of the photon pair can be controlled by varying phase shift settings before the beam splitters.
A phase shift on the left side of the system does not affect the result of the right side amplitude, and vice versa.
This system, having circulators at the beginning of the system, is denoted as transformation pattern I, and the detailed discussions of its transformation are in Sec. \ref{sec:pattern_I}.
\vspace{-15px}
\subsection{Photon propagation without circulators}
\vspace{-10px}
\begin{figure}
\vspace{-5mm}
\centering
\includegraphics[width=3.5in]{multiport_two.pdf}
\caption{System setup without circulators. The input photons are subjected to a beam splitter before they enter the multiport. The input state is transformed and propagated in one direction (shown as dotted arrows). The BS transformed input state is transformed again by the first multiport devices.}
\label{fig:multiport_no_circ}
\end{figure}
This method allows to redistribute input states between right and left side of the system without changing amplitudes.
Consider sending two photons from the left side of the beam splitter as indicated in fig. \ref{fig:multiport_no_circ} then transform the output state by the multiport device.
We only consider the first multiport transformation here.
The rest of the transformation is given in sec. \ref{sec:pattern_II}.
\begin{equation}
e_0f_0\xrightarrow{BS}-\frac{1}{2}(a_0^2-b_0^2)
\xrightarrow{M}-\frac{1}{2}(a_0-b_0)(c_0+d_0)
\end{equation}
The final state has cross-terms, and it is different from the case with circulators in a sense that the output state is \textit{coupled}.
The state does not provide clear separation between right moving and left moving amplitudes.
Even though, the state does not have clear distinction between right moving and left moving, we still refer the amplitudes right and left moving amplitudes unless special attention is required.
This system having no circulators is denoted as transformation pattern II, and the detailed discussions of its transformation are in Sec. \ref{sec:pattern_II}
\vspace{-10px}
\section{Transformation pattern I: directionally-controllable HOM effect in higher-dimensional spatial and temporal modes} \label{sec:pattern_I}
\vspace{-10px}
In this section we discuss the transformation pattern I.
The higher dimensional HOM effect is generated by the multiport-based linear optics system with circulators at the inputs.
The propagation direction control and delays between amplitudes are discussed in subsections.
We use a single multiport device to show the control effect and we introduce two multiport devices in the system for delayed effect.
\begin{figure}[htp!]
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator1.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator2.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator3.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator4.pdf}%
}
\caption{Higher dimensional HOM effect with directional control.
Correlated photons, $a_0b_0$, are sent in from the circulators into the first multiport.
After the first multiport interaction, the incoming photon pair splits into right-moving and left-moving two-photon amplitudes.
The separately-moving amplitudes are bunched at the beam splitters on right and left sides.
We can controllably switch between four different output sites, and where the clustered output photons appear depends on the location of the phase shifter $P$.
In (a), no phase plates are introduced, and the output biphoton amplitudes leave $f_0$ and $e_1$. The final state is $\frac{1}{\sqrt{2}}(\ket{0,2}_0+\ket{2,0}_1)$, meaning superposition of two photons in mode $f_0$ and two in mode $e_1$. In case (b), the phase shifter $P=\pi$ is to the left, changing the relative phase between upper and lower arms.
Similarly in (c) and (d), other locations for the phase shifters cause biphotons to leave in other spatial modes.} \label{fig:HOM_control_fig}
\end{figure}
\vspace{-15px}
\subsection{Control of propagation direction}
\vspace{-10px}
Given that one two-photon amplitude must exit left and one right, there are four possible combinations of outgoing HOM pairs as indicated in Fig. \ref{fig:HOM_control_fig}.
The combinations are, (a): $(f_0^2,e_1^2)$, (b): $(e_0^2,e_1^2)$, (c): $(e_0^2,f_1^2)$, and (d): $(f_0^2,e_1^2)$.
This means, in the case of (a) for example, the left-moving two-photon amplitude leaves in mode f, and the right-moving amplitude leaves in mode e.
Directional control of the four cases is readily demonstrated, as follows. In case (a) there is only a beam splitter transformation after the multiport, giving
\begin{eqnarray}
&-\frac{1}{4}(a_0-b_0)^2\xrightarrow{BS} -\frac{1}{4\sqrt{2}}(e_0-f_0-e_0-f_0)^2 = \frac{1}{2}f_0^2, \nonumber \\
&\frac{1}{4}(c_1+d_1)^2 \xrightarrow{BS} -\frac{1}{4\sqrt{2}}(e_1-f_1+e_1+f_1)^2 = \frac{1}{2}e_1^2.
\end{eqnarray}
The final output state is,
\begin{eqnarray}
\frac{1}{2}(f_0^2+e_1^2)=\frac{1}{\sqrt{2}}(\ket{0,2}_0+\ket{2,0}_1).
\end{eqnarray}
In case (b), a phase plate is inserted in the lower arm of the left side to switch the exit port from $d$ to $c$. All the phase shifters P are set to $\pi$, therefore transforming $b \rightarrow -b$.
\begin{eqnarray}
&-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2 \xrightarrow{P} -\frac{1}{4}(a+b)^2+\frac{1}{4}(c+d)^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(-e_0^2+e_1^2) = \frac{1}{\sqrt{2}}(-\ket{2,0}_0+\ket{2,0}_1).
\end{eqnarray}
Compared to case (a), the exit port is switched from f to e. In (c), phase plates are inserted in the lower arms of both right and left sides.
Photons in modes $b$ and $d$ are transformed to $-b$ and $-d$, respectively.
\begin{eqnarray}
&-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2 \xrightarrow{P} -\frac{1}{4}(a+b)^2+\frac{1}{4}(c-d)^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(-e_0^2-f_1^2) = -\frac{1}{\sqrt{2}}(\ket{2,0}_0+\ket{0,2}_1).
\end{eqnarray}
In (d), a phase plate is inserted in the lower arm of the right side. A photon in mode $d$ is transformed to $-d$.
\begin{eqnarray}
&-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2 \xrightarrow{P} -\frac{1}{4}(a-b)^2+\frac{1}{4}(c-d)^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(f_0^2-f_1^2) = \frac{1}{\sqrt{2}}(\ket{0,2}_0-\ket{0,2}_1).
\end{eqnarray}
This demonstrates complete directional control of biphoton propagation direction using only linear optical devices.
Directional control does not require changing splitting ratios at each linear optical device (BS and multiport), and occurs in a lossless manner since no post-selection is required.
\begin{figure*}[htp!]
\centering
\subfloat[][Delayed HOM effect without reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_circulator_arrays.pdf}\label{}}
\subfloat[][Delayed HOM effect with reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_circulator_arrays_phase_plate.pdf}\label{}}
\caption{Delayed HOM effect. The two-photon amplitude transformation progresses in time from top to bottom. The distance traveled in a single time step is indicated by vertical dashed lines. The original photons as well as photons in the target state are indicated using red circles. The green striped circles indicate intermediate transformed state. The total number of photons are always two through out the transformations. (a) Two multiports and beam splitters {\it without} phase shifters between the multiports. At the first step, the behavior is the same as for a single multiport with beam splitters. The right-moving amplitude propagates through the second multiport, and left-moving amplitude propagates through the beam splitter. The right moving amplitude is delayed by one additional multiport transformation before a two-photon observation probability will become available in spatial modes on the right. (b) Two multiports and beam splitters {\it with} a phase shifter P set at $\pi$ between multiports. When the P is present, the right-moving amplitude gains a relative phase between modes $
a_1$ and $b_1$. Reflection occurs at the multiport when the relative phase between the two is $\pi$. Therefore, the transformed amplitude reflects upon a second multiport encounter, going back to the original state with opposite propagation direction. Reflection does not occur on this transformed left-moving amplitude, therefore it continues to propagate leftward. The original left-moving amplitude becomes available for detection earlier than the transformed left-moving amplitude.}
\label{fig:delayed_HOM}
\end{figure*}
\subsection{Delayed HOM effect}
\vspace{-10px}
\subsubsection{Delayed HOM effect without reflection}
\vspace{-10px}
We introduce a phase shifter between two multiports as in Fig. \ref{fig:delayed_HOM} (b). Without the phase plate between two multiport devices, the photons behave exactly the same as in the previous subsection. However, the phase shifter can change propagation direction of right moving amplitude to the left. This reflection results in detecting HOM pairs only on the left side, but with some delay between the two exiting amplitudes. We start with the case without the phase shifter. The photon insertion is the same as the previous case, coming from the left side of the first multiport.
\vspace{-10px}
\begin{eqnarray}
&{a_{0}b_{0}}_R\xrightarrow{M}-\frac{1}{4}(a_{0}-b_{0})_L^2+ \frac{1}{4}(c_{0}+d_{0})_R^2 \nonumber \\
&\xrightarrow{T+BS,\mbox{ }T} -\frac{1}{2}f_{0L}^2 + \frac{1}{4}(a_{1}+b_{1})_R^2 \nonumber \\
&\xrightarrow{M} -\frac{1}{2}f_{0L}^2 + \frac{1}{4}(c_{1}+d_{1})_R^2 \xrightarrow{BS} -\frac{1}{2}f_{0L}^2 + \frac{1}{2}e_{1R}^2,
\end{eqnarray}
where M, T, BS represents multiport, translation and beam splitter transformation respectively. We use subscript $R$ and $L$ to illustrate amplitudes propagating to the right or left. $T$ translates a photon amplitude by a single time step (for example, $\frac{1}{4}(c_{0}+d_{0})^2 \rightarrow \frac{1}{4}(a_{1}+b_{1})^2$). The second transformation $T+BS$, $T$ is read as applying $T+BS$ on the first term and $T$ on the second term.
The final state is,
\begin{eqnarray}
-\frac{1}{2}f_{0L}^2 + \frac{1}{2}e_{1R}^2=-\frac{1}{\sqrt{2}}(\ket{0,2}_{0{T_0}L}-\ket{2,0}_{1{T_1}R}),
\end{eqnarray}
where $T_0$ is the time when the first biphoton amplitude leaves the system and $T_1$ is the exit time of the second.
The right moving amplitude stays in the system longer than the left moving amplitude because of the extra multiport device in the system, leading to time delay $\Delta T = T_1-T_0$.
\vspace{-10px}
\subsubsection{Delayed HOM effect with reflection}
\vspace{-10px}
When a $\pi$-phase shifter is inserted on one path between the multiports, the right-moving amplitude gets reflected upon the second multiport encounter.
Instead of having two-photon amplitudes on the right and left sides of the system, both photon amplitudes end up leaving from the left.
The HOM effect still occurs but now with some delay between the two amplitudes at the end of the BS.
This is indicated in Fig. \ref{fig:delayed_HOM} (b).
\vspace{-10px}
\begin{eqnarray}
&{a_{0}b_{0}}_R\xrightarrow{M}-\frac{1}{4}(a_{0}-b_{0})_L^2+ \frac{1}{4}(c_{0}+d_{0})_R^2 \nonumber \\
&\xrightarrow{T+BS,\mbox{ }T+P} -\frac{1}{2}f_{0L}^2 + \frac{1}{4}(a_{1}-b_{1})_R^2.
\end{eqnarray}
The second transformation $T+BS$, $T+P$ is read as applying $T+BS$ on the first term and $T+P$ on the second term.
Left-moving photons leave before right-moving photons.
\begin{eqnarray}
\xrightarrow{M}&\frac{1}{4}(a_{1}-b_{1})_L^2 \xrightarrow{P+T} \frac{1}{4}(c_{0}+d_{0})_L^2 \nonumber \\
&\xrightarrow{M} \frac{1}{4}(a_{0}+b_{0})_L^2 \xrightarrow{BS} \frac{1}{2}e_{0L}^2.
\end{eqnarray}
The final state,
\begin{eqnarray}
-\frac{1}{2}f_{0L}^2+\frac{1}{2}e_{0L}^2=-\frac{1}{\sqrt{2}}(\ket{0,2}_{0T_0L}-\ket{2,0}_{0T_2L}),
\end{eqnarray}
is now two HOM pair amplitudes, both on the left side of the system, at output ports $e_0$ and $f_0$, with some time delay $\Delta T = T_2 - T_0$ between them. The first amplitude leaves port $f_0$ at $T_0$, then the second leaves $e_0$ and the time labeled $T_2$.
\vspace{-10px}
\section{Transformation pattern II: state redistribution in higher-dimensional spatial and temporal modes} \label{sec:pattern_II}
\vspace{-10px}
\subsection{State transformation and propagation}
\vspace{-10px}
We have considered the case where the input photon state is transformed by the multiport device right after photon insertion in the previous section.
Instead of using circulators, we can transform the input state by the BS in advance and then transform the state by using the multiport device.
Even though the Grover matrix spreads the input state equally in four directions, the end result preserves the original form of the input state.
We demonstrate a state redistribution property using distinguishable and indistinguishable photons, meaning the input state gets redistributed between right and left side without changing amplitudes.
The propagation result is different from the previous case.
Consider sending two indistinguishable photons in the system.
The input two photons have the same polarization to make them indistinguishable.
The input photons are inserted from the left side of the beam splitter.
The beam splitter transforms the input state and propagates from the left side to right side of the device without any reflections.
The amplitudes are transformed by the multiport device after the beam splitter transformation.
This transformation splits input photons into coupled right- moving and left-moving amplitudes.
The coupled left moving amplitudes reflected from the first multiport counter propagates and transformed by the first beam splitter from the right to the left.
The right moving amplitude is transmitted without changes in amplitude.
This amplitude gets transmitted by the right side beam splitter at the end.
\vspace{-15px}
\subsubsection{Indistinguishable photons}
\vspace{-10px}
We examine the mathematical details on indistinguishable photons in the system without circulators first.
We consider three cases by sending photons in spatial modes e and f.
First, we consider indistinguishable a pair of single photons from spatial mode e and f.
\vspace{-5px}
\begin{eqnarray}
&e_{H0}f_{H0}\xrightarrow{BS}-\frac{1}{2}(a_{H0}^2-b_{H0}^2)\nonumber \\
&\xrightarrow{M} -\frac{1}{2}(a_{H0}-b_{H0})(c_{H0}+d_{H0}) \nonumber \\
&\xrightarrow{BS} -e_{H0}e_{H1}.
\end{eqnarray}
HOM state with relative phase between two amplitudes equal to +1 is considered here.
\begin{eqnarray}
&\frac{1}{2}(e_{H0}^2+f_{H0}^2)\xrightarrow{BS}\frac{1}{2}(a_{H0}^2+b_{H0}^2) \nonumber \\
&\xrightarrow{M} \frac{1}{4}(a_{H0}-b_{H0})^2+\frac{1}{4}(c_{H0}+d_{H0})^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(e_{H0}^2+f_{H1}^2).
\end{eqnarray}
The input state is redistributed in a sense that one amplitude is on the right side of the system and the other amplitude is on the left side while maintaining the original structure of the state.
HOM state with relative phase between two amplitudes equal to -1 is considered here.
\vspace{-5px}
\begin{eqnarray}
&\frac{1}{2}(e_{H0}^2-f_{H0}^2) \xrightarrow{BS} -a_{H0}b_{H0} \nonumber \\
&\xrightarrow{M} \frac{1}{4}(a_{H0}-b_{H0})^2-\frac{1}{4}(c_{H0}+d_{H0})^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(e_{H0}^2-e_{H1}^2).
\end{eqnarray}
In both cases, the output state is identical to the input state except for the spatial modes.
\vspace{-10px}
\subsubsection{Distinguishable photons}
\vspace{-10px}
Now, we examine the case of distinguishable two photon input.
The procedure is identical to the the previous case.
We begin with two distinguishable photons at each modes without superposition.
\begin{eqnarray}
&e_{H0}f_{V0} \xrightarrow{BS} \frac{1}{2}(a_{H0}-b_{H0})(a_{V0}+b_{V0})\nonumber \\
&\xrightarrow{M} -\frac{1}{2}(a_{H0}-b_{H0})(c_{V0}+d_{V0})\nonumber\\
&\xrightarrow{BS} - e_{H0} e_{V1}.
\end{eqnarray}
We examine the case of HOM states.
\begin{eqnarray}
&\frac{1}{2}(e_{H0}^2 \pm f_{V0}^2) \xrightarrow{BS}\frac{1}{4}(a_{H0}-b_{H0})^2\pm\frac{1}{4}(a_{V0}+b_{H0})^2\nonumber\\
&\xrightarrow{M}\frac{1}{4}(a_{H0}-b_{H0})^2\pm\frac{1}{4}(c_{V0}+d_{H0})^2\nonumber\\
&\xrightarrow{BS} \frac{1}{2}(e_{H0}^2 \pm e_{V1}^2).
\end{eqnarray}
The control of exit location can be performed as well in this scheme by introducing phase shifters in the system as indicated in Fig. \ref{fig:control_without_circ}.
This procedure does not destroy the redistribution property.
There are four potential spatial modes and by switching the phase shift before beam splitters, the direction of propagation switches.
The combinations are, (a): $(e_0,f_0)\rightarrow(e_0,e_1)$, (b): $(e_0,f_0)\rightarrow(e_0,f_1)$, (c): $(e_0,f_0)\rightarrow(f_0,e_1)$, and (d): $(e_0,f_0)\rightarrow(f_0,f_1)$.
The result from the system with circulators is summarized in Table. \ref{tab:table_1}, the system without them is in Table. \ref{tab:table_2}.
In the case of indistinguishable photons, the results are cyclic in a sense that all three states can be produced by using the other system.
However, there is a significant difference when distinguishable photons are considered.
\begin{figure}[htp!]
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator2.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator3.pdf}%
}
\hspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator4.pdf}%
}
\caption{Quantum state redistribution with control of propagation direction. We performed the same analysis as the higher dimensional HOM effect with direction control. By introducing phase shifters in the system before beam splitters, we can change the exit direction of the amplitudes. The starting state is $e_0f_0$. The first beam splitter transforms the input state, then they enter the multiport device. The multiport transformed state goes through beam splitters on the right and left side. The final outcome has the same form as the input state.}
\label{fig:control_without_circ}
\end{figure}
\begin{table}
\begin{tabular}{P{0.4\linewidth}P{0.6\linewidth}}
\hline \hline
\noalign{\vskip 1ex}
\multicolumn{2}{c}{State transformation with circulators}\\
\noalign{\vskip 1ex}
\hline \hline
\noalign{\vskip 1ex}
Indistinguishable photons & $a_{H0}b_{H0} \rightarrow -\frac{1}{2}(e_{H0}^2 - e_{H1}^2)$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with +1 relative phase & $\frac{1}{2}(a_{H0}^2 + b_{H0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 + e_{H1}^2)$\\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with $-1$ relative phase & $\frac{1}{2}(a_{H0}^2 - b_{H0}^2) \rightarrow -e_{H0}e_{H1}$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable photons & $a_{H0}b_{V0} \rightarrow -\frac{1}{2}(e_{H0}-e_{H1})(e_{V0}+e_{V1})$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable HOM pair & $\frac{1}{2}(a_{H0}^2 \pm b_{V0}^2) \rightarrow \frac{1}{4}\{(e_{H0}-e_{H1})^2\pm(e_{V0}-e_{V1})^2\}$ \\
\noalign{\vskip 1ex}
\hline \hline
\end{tabular}
\caption{State transformations in a system with circulators.
The first three states deal with indistinguishable photons by giving them the same polarization.
A state consisting of two single photons will become an HOM state.
We analyzed HOM states as an initial state, and they become either the HOM state or a two single-photon state.
Distinguishable photons are also analyzed by introducing orthogonal polarizations.
The output states become coupled states meaning the original states are not preserved.
} \label{tab:table_1}
\end{table}
\begin{table}
\begin{tabular}{P{0.4\linewidth}P{0.6\linewidth}}
\hline \hline
\noalign{\vskip 1ex}
\multicolumn{2}{c}{State transformation without circulators} \\
\noalign{\vskip 1ex}
\hline \hline
\noalign{\vskip 1ex}
Indistinguishable photons & $e_{H0}f_{H0} \rightarrow -e_{H0}e_{H1}$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with +1 relative phase & $\frac{1}{2}(e_{H0}^2 + f_{H0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 + e_{H1}^2)$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with $-1$ relative phase & $\frac{1}{2}(e_{H0}^2 - f_{H0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 - f_{H1}^2)$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable photons & $e_{H0}f_{V0} \rightarrow -e_{H0}e_{V1}$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable HOM pair & $\frac{1}{2}(e_{H0}^2 \pm f_{V0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 \pm f_{V1}^2)$ \\
\noalign{\vskip 1ex}
\hline \hline
\end{tabular}
\caption{State transformations in a system without circulators.
The structure of the table is the same as the Table. I.
The first three states deal with indistinguishable photons by giving them the same polarization.
The last two states handle distinguishable photons.
The output states preserve the same form as the input state.
We start the transformation from the system location 0, then the transformed states are redistributed between location 0 and location 1.
The result shows coherent transportation of input states.} \label{tab:table_2}
\end{table}
\vspace{-10px}
\subsection{Delayed state redistribution}
\vspace{-10px}
\begin{figure*}[htp!]
\centering
\subfloat[][State redistribution without reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_array.pdf}\label{}}
\subfloat[][State redistribution with reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_array_phase_plate.pdf}\label{}}
\caption{Delayed state redistribution.
The two-photon amplitude transformation progresses in time from top to bottom.
The distance traveled in a single time step is indicated by vertical dashed lines. The total photon numbers are two in the system through out the propagation. At the first step for both cases, the input two-photon state is transformed by the BS.
The transformed state becomes the HOM state, and it is indicated as red transparent overlapped circles occupying both modes.
The initial and the final transformed state are indicated using solid red circles, and intermediate states are indicated in striped yellow circles.
(a) Two multiports and beam splitters {\it without} phase shifters between the multiports.
The HOM state enters the multiport and transformed taking the form of $-\frac{1}{2}(a_{0}-b_{0})(c_{0}+d_{0})$.
The amplitudes are coupled, however, they propagate without changing its amplitude. After several steps, the amplitudes occupying two rails converges to a single mode state after transformation by beam splitters.
The final state has the same form as the input state.
(b) Two multiports and beam splitters {\it with} a phase shifter P set at $\pi$ between multiports.
When the P is present, the right-moving coupled amplitude gains a relative phase between modes $a_1$ and $b_1$.
Reflection occurs at the multiport when the relative phase between the two is $\pi$.
Therefore, the transformed amplitude reflects upon a second multiport encounter, going back to the original state with opposite propagation direction.
Reflection does not occur on this transformed coupled left-moving amplitude, therefore it continues to propagate leftward.
The original left-moving amplitude becomes available for detection earlier than the transformed left-moving amplitude.}
\label{fig:delayed_state_dist}
\end{figure*}
We introduce the temporal delay effect as the higher dimensional HOM case by introducing a phase shifter between two multiports.
\vspace{-10px}
\subsubsection{Without reflection}
\vspace{-10px}
When there is no phase shifter between the two multiports, the result is identical to the system with a single multiport from the previous section.
The state transformation and propagation is provided schematically in Fig. \ref{fig:delayed_state_dist} (a).
The photons are initially sent from the left side of the BS.
The correlated photons are transformed to HOM state through the BS.
\vspace{-10px}
\begin{eqnarray}
&{e_{0}f_{0}}_R \xrightarrow{BS} \frac{1}{2}(a_{0}^2-b_{0}^2)_R \nonumber\\
&\xrightarrow{M}-\frac{1}{2}(a_{0}-b_{0})_L(c_{0}+d_{0})_R.
\end{eqnarray}
The HOM state is transformed by the multiport device.
This state is in a coupled state because right moving and left moving amplitudes are not separated.
We propagate this state through the BS on the left and translate the amplitudes moving to the right.
\begin{eqnarray}
&\xrightarrow{BS,T}-\frac{1}{\sqrt{2}}f_{0L}(a_{1}+b_{1})_R \xrightarrow{M} -\frac{1}{\sqrt{2}}f_{0L}(c_{1}+d_{1})_R\nonumber\\
&\xrightarrow{BS} -f_{0T_0L}e_{1T_1R}
\end{eqnarray}
The left moving amplitude is transformed by the left BS while right moving amplitude propagates to the second multiport device.
We introduced temporal difference between the right moving and the left moving photons.
\vspace{-15px}
\subsubsection{With reflection}
\vspace{-10px}
Reflection of amplitudes are introduced when there is a phase shifter between two multiport devices as indicated in fig. \ref{fig:delayed_state_dist} (b).
\begin{eqnarray}
&{e_{0}f_{0}}_R \xrightarrow{BS} \frac{1}{2}(a_{0}^2-b_{0}^2)_R
\xrightarrow{M} -\frac{1}{2}(a_{0}-b_{0})_L(c_{0}+d_{0})_R \nonumber \\
&\xrightarrow{BS,T+P}-\frac{1}{\sqrt{2}}f_{0L}(a_{1}-b_{1})_R
\end{eqnarray}
The right moving amplitude gains relative phase between upper and lower rails, and this relative phase allows the amplitude to get reflected upon multiport encounter.
\begin{eqnarray}
&\xrightarrow{M} -\frac{1}{\sqrt{2}}f_{0L}(a_{1}-b_{1})_L \xrightarrow{T+P} -\frac{1}{\sqrt{2}}f_{0L}(c_{0}+d_{0})_L\nonumber\\
&\xrightarrow{M} -\frac{1}{\sqrt{2}}f_{0L}(a_{0}+b_{0})_L \xrightarrow{BS} -f_{0T_{0}L} e_{0T_{2}L}
\end{eqnarray}
The input photons do not have any delays between the two at the beginning. The delay $\Delta T = T_2 - T_0$ is introduced from the reflection in the system.
\vspace{-10pt}
\section{Conclusion}
\vspace{-10px}
We demonstrated higher dimensional quantum state manipulation such as the HOM effect and state redistribution by applying linear-optical four-ports realizing four-dimensional Grover matrix accompanied by beam splitters and phase shifters.
Identical photons are sent into two of the four input-output ports and split into right-moving and left-moving amplitudes, with no cross terms to observe the HOM effect.
This absolute separation of propagation direction without mixing of right-moving and left-moving amplitudes insures the photons remain clustered as they propagate through the system.
Variable phase shifts in the system allow the HOM photon pairs to switch between four spatial output destinations, which can increase information capacity.
Time delays between emerging parts of the clustered two-photon state illustrating “delayed” HOM effect can be engineered using two multiports.
In addition, depending on the phase shifter position, the propagation direction can be reversed so that the right moving amplitude can get reflected at the second multiport, resulting in HOM pairs always leaving only from the left side of the system and with a particular time-bin delay.
The same situations have been investigated in a system without circulators. This system allows to redistribute the input state between the right and the left side of the system without changing amplitudes.
The HOM effect and clustered photon pairs are widely used in quantum information science. The approach introduced here adds extra degrees of freedom, and paves the way for new applications that require control over the spatial and temporal modes of the HOM amplitudes as they move through one- and two-dimensional networks.
We have demonstrated two photon amplitude control in both spatial and temporal modes. This two photon system can be extended to multiphoton input states, and manipulation of more complex entangled states would be the next milestones to be achieved.
\vspace{-15pt}
\section*{Appendix}
\section{\label{sec:level1}Introduction}
\begin{figure}
\vspace{-5mm}
\centering
\includegraphics[width=3in]{HOM_cancelling_4.pdf}
\caption{Hong-Ou-Mandel effect. Two identical photons are sent into different beam splitter input ports. There are four possible outcomes, two photons leaving one port, two photons leaving the other port, each photon reflecting to give single photons at each exit, and both transmitting to give single photons at each port. The coincidence terms cancel out since they are identical but enter with opposite sign. The final state is a superposition of two outcomes, each with both photons clustered together at the same exit port.}
\label{fig:HOM_effect}
\end{figure}
\section{Introduction}
\vspace{-10px}
The Hong-Ou-Mandel (HOM) effect is one of the most recognized quantum two-photon interference effects \cite{hong1987measurement}.
When two indistinguishable photons arrive simultaneously at different inputs of a 50:50 beam splitter (BS), single-photon amplitudes at each output cancel, resulting in quantum superposition of two-photon states appearing at each output port, as in Fig. \ref{fig:HOM_effect}.
This traditional HOM method, observed on a BS having two input and two output ports, always has the two-photon state simultaneously occupying both output spatial modes, leaving no room to engineer control of propagation direction.
Various types of studies on quantum state transformations in multiport devices have been performed such as two photon propagation in a multimode system \cite{weihs1996two,zukowski1997realizable},
quantum interference effects using a few photons \cite{meany2012non,de2014coincidence,tichy2011four,campos2000three}, and propagation of multi-photons \cite{lim2005generalized,tillmann2015generalized,menssen2017distinguishability}. Internal degrees of freedom are also incorporated to enhance communication capacity \cite{walborn2003multimode,poem2012two,zhang2016engineering}.
Systems and procedures using multi-photon states, such as boson sampling, have been analyzed using multiport beam splitters both theoretically and experimentally \cite{aaronson2011computational,tillmann2013experimental,spring2013boson,bentivegna2015experimental,he2017time,wang2019boson}.
The HOM effect plays an important role in the field of quantum metrology when two-photon $|2002\rangle$-type states are extended to $N$-photon $N00N$ state \cite{dowling2008quantum,motes2015linear}.
Additionally, coherent transport of quantum states has been attracting attention, where single- and two-photon discrete-time quantum walk schemes are employed to transfer and process quantum states \cite{bose2003quantum,perez2013coherent,lovett2010universal,chapman2016experimental,nitsche2016quantum}.
A quantum routing approach has been proposed to transfer unknown states in 1D and 2D structures to assist quantum communication protocols \cite{zhan2014perfect,vstefavnak2016perfect,bartkiewicz2018implementation}.
Photon propagation control is especially crucial in a large optical network to distribute quantum states between two parties.
The network can be formed by combining multiple copies of four-port devices.
The state manipulation schemes we present can be integrated in quantum communication protocols since state retrieval timing can be chosen at will.
In this manuscript, we propose two-photon quantum state engineering and transportation methods with a linear-optical system which allows manipulation of photon amplitudes by using linear-optical devices such as optical multiports, beam splitters, and phase shifters.
Previously, such multiports have been introduced to demonstrate a two-photon clustering effect in quantum walks when multiple multiport devices are connected to form a chain \cite{simon2020quantum}. Clustering of two photons means that after encountering a multiport, the input two-photon amplitude separates into a superposition of a right-moving and a left-moving two-photon amplitude, with no amplitude for the photons to move in opposite directions.
By utilizing this separation, a higher-dimensional unitary transformation enables flexible quantum engineering designs of possible travel path combinations by switching relative phases within right moving and left moving amplitudes independently.
When two or more multiports are combined, this control of quantum amplitudes in a two-photon state allows demonstration of a “delayed" HOM effect engaging also time-bin modes in addition to spatial modes.
To perform this delayed effect, two or more multiports are required, and relative phase shifts between two rails can reflect the incoming amplitudes.
This controllable reflection without mirrors can also be seen as an additional state manipulation feature.
We introduce two distinct systems. The first system utilizes direct transformation of two-photons by the four-port device using circulators.
The second case does not have circulators in the system.
The photons are sent from the left side of the beam splitters, then the amplitudes encounter the multiport device.
This second system has not been analyzed in the past. Specific input states for both distinguishable and indistinguishable photons are redistributed between two parties coherently. Therefore, this system is particularly useful in quantum routing type applications.
This article is organized as follows. In Sec. II, we introduce the main optical components used in this manuscript to perform quantum state transformation. These basic linear optical devices are used to show HOM effect engaging in spatial modes and time-bin modes, and is addressed in Sec. III. In Sec. IV, we show redistribution of two photon states using the devices introduced that are presented in Sec. II. The summary of the results are given in Sec. V.
\begin{figure}[htp]
\subfloat[]{%
\includegraphics[clip,width=0.8\columnwidth]{multiport_systematic.pdf}%
}
\vspace{-10px}
\subfloat[]{%
\includegraphics[clip,width=0.8\columnwidth]{multiport_systematic_photons.pdf}%
}
\vspace{-10px}
\caption{(a) A possible experimental realization of a directionally-unbiased linear-optical four-port consists of four beam splitters, four mirrors, and four phase shifters. A photon can enter any of the four ports, and exit at any of the four ports (labeled as $a,b,c$, and $d$). With a specific choice of phase settings, a Grover matrix can be realized by coherently summing all possible paths to each output \cite{osawa2019directionally}. A schematic symbol for this device is shown on the right. (b) Single multiport transformation of a two-photon input state. The input state of two correlated photons entering from the left is depicted as $ab$ $(\ket{1,1})$. After scattering by a Grover multiport, the state transforms into $-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2$, which has clear separation of right- and left-moving two-photon amplitudes. No cross-terms with photons moving in opposite directions occur.}
\label{fig:combined_four_port}
\vspace{-10pt}
\end{figure}
\section{Photonic state transformations via linear optical devices}
\vspace{-10px}
In this section, we consider photonic state transformations in higher-dimensional spatial modes using a unitary four-dimensional Grover matrix \cite{grover1996fast} in place of the beam splitter.
In this section, we introduce the main systems that will be used for linear state transformations, followed by the basic photonic devices to implement them.
Beam splitters and the four-dimensional Grover matrix are the central system component.
We mainly use photon number representation to describe states through out the manuscript.
The general beam splitter transformation matrix is
\begin{eqnarray}
\label{eqn:BS}
&\begin{pmatrix}
\hat{c}
\\
\hat{d}
\end{pmatrix}=
\frac{1}{\sqrt{2}}
\begin{pmatrix}
1 & 1\\
-1 & 1
\end{pmatrix}
\begin{pmatrix}
\hat{a}\\
\hat{b}
\end{pmatrix}
\end{eqnarray}
where $\hat{a}$,$\hat{b}$,$\hat{c}$, and $\hat{d}$ are used to describe the input photon state transformation.
The labels are generalized here, therefore the specific location dependent beam splitter transformations are redefined in later sections.
We use photon number states to describe the system unless otherwise specified. The input state is denoted as $\hat{a}\hat{b}$ where $\hat{a}$ and $\hat{b}$ are respectively creation operators for the spatial modes $a$ and $b$.
The hat notation is dropped henceforth.
For a photon in spatial mode $a$ with horizontally polarized photon is denoted as $a_{H}$ and horizontally polarized photon in mode $b$ is denoted as $b_{H}$.
We omit polarization degrees of freedom when identical photons are used through out the system.
Photonic implementations of the Grover matrix can be readily realized \cite{carolan2015universal,crespi2013anderson,spagnolo2013three,fan1998channel,nikolopoulos2008directional}.
To be concrete, we use directionally-unbiased linear-optical four-ports (Fig. \ref{fig:combined_four_port} (a)) as an example.
Consider sending two indistinguishable photons into a four-dimensional multiport device realization of a Grover matrix.
This Grover operator, the multiport, described by the unitary matrix
\begin{equation}
\label{eqn:Grover}
Grover =
\frac{1}{2}
\begin{pmatrix}
-1&1&1&1\\
1&-1&1&1\\
1&1&-1&1\\
1&1&1&-1
\end{pmatrix},
\end{equation}
has equal splitting ratios between all input-output combinations and generalizes the BS transformation matrix given below in Eq.\eqref{eqn:BS}.
In general, photons in modes $a$ and $b$ are transformed in the following manner,
\begin{eqnarray}
&a \xrightarrow{M} \frac{1}{2}(-a+b+c+d)\; \mbox{ and }\;\\ \nonumber
&b \xrightarrow{M} \frac{1}{2}(a-b+c+d).
\end{eqnarray}
Theoretical analysis of the reversible Grover matrix has been performed by linear-optical directionally-unbiased four-ports \cite{simon2016group,simon2018joint,osawa2019directionally}, which consist of four beam splitters, four phase shifters, and four mirrors as indicated in Fig. \ref{fig:combined_four_port}(a).
They are represented schematically by the symbol in Fig. \ref{fig:combined_four_port}(b).
The three-port version of this device has been experimentally demonstrated using bulk optical devices \cite{osawa2018experimental}.
To have better and precise control of phases, miniaturization of the device is highly preferred to realize the four-port especially when several multiport devices are required to carry out an experiment.
In general, directional unitary devices such as those of the Reck and some other unitary matrix decomposition models \cite{reck1994experimental,su2019hybrid,clements2016optimal,de2018simple,motes2014scalable} can also realize a Grover matrix.
However, directionally-unbiased devices are advantageous when designing the delayed HOM effect, as well as requiring fewer optical resources.
Identical photons are sent into two of the four input-output ports from the left side (indicated in Fig. \ref{fig:combined_four_port}(b)).
We used multiport devices and beam splitters to form two systems for state propagation.
The photons are sent from the left side of the system through out the manuscript.
The first BS multiport composite system is denoted as subscript 0 and the other half is denoted as subscript 1.
The result differs depending on the input location of photons.
Consider a system consisting of two multiports and two beam splitters.
There are several ways to insert photons in the system, however we choose two specific ones in this manuscript.
To be able to send a photon into the middle of the system, the setup needs to be supplied with circulators, is shown in Fig. \ref{fig:multiport_circ}.
Another setup requires no circulators to propagate input photons.
The photons experience an extra transformation by a beam splitter upon photon entrance.
The system is graphically supplied in Fig. \ref{fig:multiport_no_circ}.
It needs to be noted that the number of multiports in the system does not change the final outcome.
We are using two multiports as an example, however, the result is the same when the system has a single multiport or more than two multiports as long as the devices are assumed to be lossless during the propagation.
Brief comments on the mathematical structure of the transformations carried out by the configurations given in Figs. \ref{fig:multiport_circ} and \ref{fig:multiport_no_circ} are given in the Appendix.
\vspace{-15px}
\subsection{Photon propagation using circulators}
\vspace{-10px}
\begin{figure}
\vspace{-5mm}
\centering
\includegraphics[width=3.5in]{multiport_two_circulator.pdf}
\caption{A system setup with input photons supplied by circulators. The system consists of two beam splitters, two multiport devices, and two circulators. These circulators allow us to send photons from the left side of the multiport device without experiencing a beam splitter transformation before entering the multiport device. The input state split into right moving and left moving amplitudes (shown as dotted arrows) upon multiport transformation.}
\label{fig:multiport_circ}
\end{figure}
This method is used to distribute HOM pair between the right and the left side of the system.
The original input state $a_0b_0$ transforms to:
\begin{eqnarray} \label{eq:trans}
a_0b_0 &\xrightarrow{M} \frac{1}{2}(-a_0+b_0+c_0+d_0)\frac{1}{2}(a_0-b_0+c_0+d_0) \nonumber \\
&=-\frac{1}{4}(a_0^2+b_0^2)+\frac{1}{2}a_0b_0+\frac{1}{4}(c_0^2+d_0^2)+\frac{1}{2}c_0d_0\nonumber \\
&= -\frac{1}{4}(a_0-b_0)^2+\frac{1}{4}(c_0+d_0)^2,
\end{eqnarray}
where we have used the commutation relation $ab = ba$ since the photons are identical and in different spatial locations.
Eq.\eqref{eq:trans} shows that correlated photons are split into right moving $\frac{1}{4}(c_0+d_0)^2$ and left moving $-\frac{1}{4}(a_0-b_0)^2$ amplitudes, with no cross terms.
This absolute separation of propagation direction without mixing of right moving and left moving amplitudes is important because the photon pairs remain distinctly localized and clustered at each step \cite{simon2020quantum}.
The right moving amplitude is translated to $\frac{1}{4}(a_1+b_1)^2$ and propagates without changing its form.
$\frac{1}{4}(a_1+b_1)^2 \xrightarrow{M} \frac{1}{4}(c_1+d_1)^2$.
The left moving amplitude $-\frac{1}{4}(a_0-b_0)^2$ stays the same until BS transformation.
The controlled HOM effect can be observed in higher-dimensional multiports assisted by extra beam splitters.
Imagine beam splitters inserted in the system as in Fig. \ref{fig:multiport_circ}.
Input state $ab$ is now transformed into $-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2$ as indicated above, then further transformed by beam splitters to obtain HOM pairs between the right side and left side of the system.
The right and left sides of the system each have two output ports, and the exit port of the photon pair can be controlled by varying phase shift settings before the beam splitters.
A phase shift on the left side of the system does not affect the result of the right side amplitude, and vice versa.
This system, having circulators at the beginning of the system, is denoted as transformation pattern I, and the detailed discussions of its transformation are in Sec. \ref{sec:pattern_I}.
\vspace{-15px}
\subsection{Photon propagation without circulators}
\vspace{-10px}
\begin{figure}
\vspace{-5mm}
\centering
\includegraphics[width=3.5in]{multiport_two.pdf}
\caption{System setup without circulators. The input photons are subjected to a beam splitter before they enter the multiport. The input state is transformed and propagated in one direction (shown as dotted arrows). The BS transformed input state is transformed again by the first multiport devices.}
\label{fig:multiport_no_circ}
\end{figure}
This method allows to redistribute input states between right and left side of the system without changing amplitudes.
Consider sending two photons from the left side of the beam splitter as indicated in fig. \ref{fig:multiport_no_circ} then transform the output state by the multiport device.
We only consider the first multiport transformation here.
The rest of the transformation is given in sec. \ref{sec:pattern_II}.
\begin{equation}
e_0f_0\xrightarrow{BS}-\frac{1}{2}(a_0^2-b_0^2)
\xrightarrow{M}-\frac{1}{2}(a_0-b_0)(c_0+d_0)
\end{equation}
The final state has cross-terms, and it is different from the case with circulators in a sense that the output state is \textit{coupled}.
The state does not provide clear separation between right moving and left moving amplitudes.
Even though, the state does not have clear distinction between right moving and left moving, we still refer the amplitudes right and left moving amplitudes unless special attention is required.
This system having no circulators is denoted as transformation pattern II, and the detailed discussions of its transformation are in Sec. \ref{sec:pattern_II}
\vspace{-10px}
\section{Transformation pattern I: directionally-controllable HOM effect in higher-dimensional spatial and temporal modes} \label{sec:pattern_I}
\vspace{-10px}
In this section we discuss the transformation pattern I.
The higher dimensional HOM effect is generated by the multiport-based linear optics system with circulators at the inputs.
The propagation direction control and delays between amplitudes are discussed in subsections.
We use a single multiport device to show the control effect and we introduce two multiport devices in the system for delayed effect.
\begin{figure}[htp!]
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator1.pdf}%
}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator2.pdf}%
}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator3.pdf}%
}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_with_circulator4.pdf}%
}
\caption{Higher dimensional HOM effect with directional control.
Correlated photons, $a_0b_0$, are sent in from the circulators into the first multiport.
After the first multiport interaction, the incoming photon pair splits into right-moving and left-moving two-photon amplitudes.
The separately-moving amplitudes are bunched at the beam splitters on right and left sides.
We can controllably switch between four different output sites, and where the clustered output photons appear depends on the location of the phase shifter $P$.
In (a), no phase plates are introduced, and the output biphoton amplitudes leave $f_0$ and $e_1$. The final state is $\frac{1}{\sqrt{2}}(\ket{0,2}_0+\ket{2,0}_1)$, meaning superposition of two photons in mode $f_0$ and two in mode $e_1$. In case (b), the phase shifter $P=\pi$ is to the left, changing the relative phase between upper and lower arms.
Similarly in (c) and (d), other locations for the phase shifters cause biphotons to leave in other spatial modes.} \label{fig:HOM_control_fig}
\end{figure}
\vspace{-15px}
\subsection{Control of propagation direction}
\vspace{-10px}
Given that one two-photon amplitude must exit left and one right, there are four possible combinations of outgoing HOM pairs as indicated in Fig. \ref{fig:HOM_control_fig}.
The combinations are, (a): $(f_0^2,e_1^2)$, (b): $(e_0^2,e_1^2)$, (c): $(e_0^2,f_1^2)$, and (d): $(f_0^2,e_1^2)$.
This means, in the case of (a) for example, the left-moving two-photon amplitude leaves in mode f, and the right-moving amplitude leaves in mode e.
Directional control of the four cases is readily demonstrated, as follows. In case (a) there is only a beam splitter transformation after the multiport, giving
\begin{eqnarray}
&-\frac{1}{4}(a_0-b_0)^2\xrightarrow{BS} -\frac{1}{4\sqrt{2}}(e_0-f_0-e_0-f_0)^2 = \frac{1}{2}f_0^2, \nonumber \\
&\frac{1}{4}(c_1+d_1)^2 \xrightarrow{BS} -\frac{1}{4\sqrt{2}}(e_1-f_1+e_1+f_1)^2 = \frac{1}{2}e_1^2.
\end{eqnarray}
The final output state is,
\begin{eqnarray}
\frac{1}{2}(f_0^2+e_1^2)=\frac{1}{\sqrt{2}}(\ket{0,2}_0+\ket{2,0}_1).
\end{eqnarray}
In case (b), a phase plate is inserted in the lower arm of the left side to switch the exit port from $d$ to $c$. All the phase shifters P are set to $\pi$, therefore transforming $b \rightarrow -b$.
\begin{eqnarray}
&-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2 \xrightarrow{P} -\frac{1}{4}(a+b)^2+\frac{1}{4}(c+d)^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(-e_0^2+e_1^2) = \frac{1}{\sqrt{2}}(-\ket{2,0}_0+\ket{2,0}_1).
\end{eqnarray}
Compared to case (a), the exit port is switched from f to e. In (c), phase plates are inserted in the lower arms of both right and left sides.
Photons in modes $b$ and $d$ are transformed to $-b$ and $-d$, respectively.
\begin{eqnarray}
&-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2 \xrightarrow{P} -\frac{1}{4}(a+b)^2+\frac{1}{4}(c-d)^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(-e_0^2-f_1^2) = -\frac{1}{\sqrt{2}}(\ket{2,0}_0+\ket{0,2}_1).
\end{eqnarray}
In (d), a phase plate is inserted in the lower arm of the right side. A photon in mode $d$ is transformed to $-d$.
\begin{eqnarray}
&-\frac{1}{4}(a-b)^2+\frac{1}{4}(c+d)^2 \xrightarrow{P} -\frac{1}{4}(a-b)^2+\frac{1}{4}(c-d)^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(f_0^2-f_1^2) = \frac{1}{\sqrt{2}}(\ket{0,2}_0-\ket{0,2}_1).
\end{eqnarray}
This demonstrates complete directional control of biphoton propagation direction using only linear optical devices.
Directional control does not require changing splitting ratios at each linear optical device (BS and multiport), and occurs in a lossless manner since no post-selection is required.
\begin{figure*}[htp!]
\centering
\subfloat[][Delayed HOM effect without reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_circulator_arrays.pdf}\label{}}
\subfloat[][Delayed HOM effect with reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_circulator_arrays_phase_plate.pdf}\label{}}
\caption{Delayed HOM effect. The two-photon amplitude transformation progresses in time from top to bottom. The distance traveled in a single time step is indicated by vertical dashed lines. The original photons as well as photons in the target state are indicated using red circles. The green striped circles indicate intermediate transformed state. The total number of photons are always two through out the transformations. (a) Two multiports and beam splitters {\it without} phase shifters between the multiports. At the first step, the behavior is the same as for a single multiport with beam splitters. The right-moving amplitude propagates through the second multiport, and left-moving amplitude propagates through the beam splitter. The right moving amplitude is delayed by one additional multiport transformation before a two-photon observation probability will become available in spatial modes on the right. (b) Two multiports and beam splitters {\it with} a phase shifter P set at $\pi$ between multiports. When the P is present, the right-moving amplitude gains a relative phase between modes $
a_1$ and $b_1$. Reflection occurs at the multiport when the relative phase between the two is $\pi$. Therefore, the transformed amplitude reflects upon a second multiport encounter, going back to the original state with opposite propagation direction. Reflection does not occur on this transformed left-moving amplitude, therefore it continues to propagate leftward. The original left-moving amplitude becomes available for detection earlier than the transformed left-moving amplitude.}
\label{fig:delayed_HOM}
\end{figure*}
\subsection{Delayed HOM effect}
\vspace{-10px}
\subsubsection{Delayed HOM effect without reflection}
\vspace{-10px}
We introduce a phase shifter between two multiports as in Fig. \ref{fig:delayed_HOM} (b). Without the phase plate between two multiport devices, the photons behave exactly the same as in the previous subsection. However, the phase shifter can change propagation direction of right moving amplitude to the left. This reflection results in detecting HOM pairs only on the left side, but with some delay between the two exiting amplitudes. We start with the case without the phase shifter. The photon insertion is the same as the previous case, coming from the left side of the first multiport.
\vspace{-10px}
\begin{eqnarray}
&{a_{0}b_{0}}_R\xrightarrow{M}-\frac{1}{4}(a_{0}-b_{0})_L^2+ \frac{1}{4}(c_{0}+d_{0})_R^2 \nonumber \\
&\xrightarrow{T+BS,\mbox{ }T} -\frac{1}{2}f_{0L}^2 + \frac{1}{4}(a_{1}+b_{1})_R^2 \nonumber \\
&\xrightarrow{M} -\frac{1}{2}f_{0L}^2 + \frac{1}{4}(c_{1}+d_{1})_R^2 \xrightarrow{BS} -\frac{1}{2}f_{0L}^2 + \frac{1}{2}e_{1R}^2,
\end{eqnarray}
where M, T, BS represents multiport, translation and beam splitter transformation respectively. We use subscript $R$ and $L$ to illustrate amplitudes propagating to the right or left. $T$ translates a photon amplitude by a single time step (for example, $\frac{1}{4}(c_{0}+d_{0})^2 \rightarrow \frac{1}{4}(a_{1}+b_{1})^2$). The second transformation $T+BS$, $T$ is read as applying $T+BS$ on the first term and $T$ on the second term.
The final state is,
\begin{eqnarray}
-\frac{1}{2}f_{0L}^2 + \frac{1}{2}e_{1R}^2=-\frac{1}{\sqrt{2}}(\ket{0,2}_{0{T_0}L}-\ket{2,0}_{1{T_1}R}),
\end{eqnarray}
where $T_0$ is the time when the first biphoton amplitude leaves the system and $T_1$ is the exit time of the second.
The right moving amplitude stays in the system longer than the left moving amplitude because of the extra multiport device in the system, leading to time delay $\Delta T = T_1-T_0$.
\vspace{-10px}
\subsubsection{Delayed HOM effect with reflection}
\vspace{-10px}
When a $\pi$-phase shifter is inserted on one path between the multiports, the right-moving amplitude gets reflected upon the second multiport encounter.
Instead of having two-photon amplitudes on the right and left sides of the system, both photon amplitudes end up leaving from the left.
The HOM effect still occurs but now with some delay between the two amplitudes at the end of the BS.
This is indicated in Fig. \ref{fig:delayed_HOM} (b).
\vspace{-10px}
\begin{eqnarray}
&{a_{0}b_{0}}_R\xrightarrow{M}-\frac{1}{4}(a_{0}-b_{0})_L^2+ \frac{1}{4}(c_{0}+d_{0})_R^2 \nonumber \\
&\xrightarrow{T+BS,\mbox{ }T+P} -\frac{1}{2}f_{0L}^2 + \frac{1}{4}(a_{1}-b_{1})_R^2.
\end{eqnarray}
The second transformation $T+BS$, $T+P$ is read as applying $T+BS$ on the first term and $T+P$ on the second term.
Left-moving photons leave before right-moving photons.
\begin{eqnarray}
\xrightarrow{M}&\frac{1}{4}(a_{1}-b_{1})_L^2 \xrightarrow{P+T} \frac{1}{4}(c_{0}+d_{0})_L^2 \nonumber \\
&\xrightarrow{M} \frac{1}{4}(a_{0}+b_{0})_L^2 \xrightarrow{BS} \frac{1}{2}e_{0L}^2.
\end{eqnarray}
The final state,
\begin{eqnarray}
-\frac{1}{2}f_{0L}^2+\frac{1}{2}e_{0L}^2=-\frac{1}{\sqrt{2}}(\ket{0,2}_{0T_0L}-\ket{2,0}_{0T_2L}),
\end{eqnarray}
is now two HOM pair amplitudes, both on the left side of the system, at output ports $e_0$ and $f_0$, with some time delay $\Delta T = T_2 - T_0$ between them. The first amplitude leaves port $f_0$ at $T_0$, then the second leaves $e_0$ and the time labeled $T_2$.
\vspace{-10px}
\section{Transformation pattern II: state redistribution in higher-dimensional spatial and temporal modes} \label{sec:pattern_II}
\vspace{-10px}
\subsection{State transformation and propagation}
\vspace{-10px}
We have considered the case where the input photon state is transformed by the multiport device right after photon insertion in the previous section.
Instead of using circulators, we can transform the input state by the BS in advance and then transform the state by using the multiport device.
Even though the Grover matrix spreads the input state equally in four directions, the end result preserves the original form of the input state.
We demonstrate a state redistribution property using distinguishable and indistinguishable photons, meaning the input state gets redistributed between right and left side without changing amplitudes.
The propagation result is different from the previous case.
Consider sending two indistinguishable photons in the system.
The input two photons have the same polarization to make them indistinguishable.
The input photons are inserted from the left side of the beam splitter.
The beam splitter transforms the input state and propagates from the left side to right side of the device without any reflections.
The amplitudes are transformed by the multiport device after the beam splitter transformation.
This transformation splits input photons into coupled right- moving and left-moving amplitudes.
The coupled left moving amplitudes reflected from the first multiport counter propagates and transformed by the first beam splitter from the right to the left.
The right moving amplitude is transmitted without changes in amplitude.
This amplitude gets transmitted by the right side beam splitter at the end.
\vspace{-15px}
\subsubsection{Indistinguishable photons}
\vspace{-10px}
We examine the mathematical details on indistinguishable photons in the system without circulators first.
We consider three cases by sending photons in spatial modes e and f.
First, we consider indistinguishable a pair of single photons from spatial mode e and f.
\vspace{-5px}
\begin{eqnarray}
&e_{H0}f_{H0}\xrightarrow{BS}-\frac{1}{2}(a_{H0}^2-b_{H0}^2)\nonumber \\
&\xrightarrow{M} -\frac{1}{2}(a_{H0}-b_{H0})(c_{H0}+d_{H0}) \nonumber \\
&\xrightarrow{BS} -e_{H0}e_{H1}.
\end{eqnarray}
HOM state with relative phase between two amplitudes equal to +1 is considered here.
\begin{eqnarray}
&\frac{1}{2}(e_{H0}^2+f_{H0}^2)\xrightarrow{BS}\frac{1}{2}(a_{H0}^2+b_{H0}^2) \nonumber \\
&\xrightarrow{M} \frac{1}{4}(a_{H0}-b_{H0})^2+\frac{1}{4}(c_{H0}+d_{H0})^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(e_{H0}^2+f_{H1}^2).
\end{eqnarray}
The input state is redistributed in a sense that one amplitude is on the right side of the system and the other amplitude is on the left side while maintaining the original structure of the state.
HOM state with relative phase between two amplitudes equal to -1 is considered here.
\vspace{-5px}
\begin{eqnarray}
&\frac{1}{2}(e_{H0}^2-f_{H0}^2) \xrightarrow{BS} -a_{H0}b_{H0} \nonumber \\
&\xrightarrow{M} \frac{1}{4}(a_{H0}-b_{H0})^2-\frac{1}{4}(c_{H0}+d_{H0})^2 \nonumber \\
&\xrightarrow{BS} \frac{1}{2}(e_{H0}^2-e_{H1}^2).
\end{eqnarray}
In both cases, the output state is identical to the input state except for the spatial modes.
\vspace{-10px}
\subsubsection{Distinguishable photons}
\vspace{-10px}
Now, we examine the case of distinguishable two photon input.
The procedure is identical to the the previous case.
We begin with two distinguishable photons at each modes without superposition.
\begin{eqnarray}
&e_{H0}f_{V0} \xrightarrow{BS} \frac{1}{2}(a_{H0}-b_{H0})(a_{V0}+b_{V0})\nonumber \\
&\xrightarrow{M} -\frac{1}{2}(a_{H0}-b_{H0})(c_{V0}+d_{V0})\nonumber\\
&\xrightarrow{BS} - e_{H0} e_{V1}.
\end{eqnarray}
We examine the case of HOM states.
\begin{eqnarray}
&\frac{1}{2}(e_{H0}^2 \pm f_{V0}^2) \xrightarrow{BS}\frac{1}{4}(a_{H0}-b_{H0})^2\pm\frac{1}{4}(a_{V0}+b_{H0})^2\nonumber\\
&\xrightarrow{M}\frac{1}{4}(a_{H0}-b_{H0})^2\pm\frac{1}{4}(c_{V0}+d_{H0})^2\nonumber\\
&\xrightarrow{BS} \frac{1}{2}(e_{H0}^2 \pm e_{V1}^2).
\end{eqnarray}
The control of exit location can be performed as well in this scheme by introducing phase shifters in the system as indicated in Fig. \ref{fig:control_without_circ}.
This procedure does not destroy the redistribution property.
There are four potential spatial modes and by switching the phase shift before beam splitters, the direction of propagation switches.
The combinations are, (a): $(e_0,f_0)\rightarrow(e_0,e_1)$, (b): $(e_0,f_0)\rightarrow(e_0,f_1)$, (c): $(e_0,f_0)\rightarrow(f_0,e_1)$, and (d): $(e_0,f_0)\rightarrow(f_0,f_1)$.
The result from the system with circulators is summarized in Table. \ref{tab:table_1}, the system without them is in Table. \ref{tab:table_2}.
In the case of indistinguishable photons, the results are cyclic in a sense that all three states can be produced by using the other system.
However, there is a significant difference when distinguishable photons are considered.
\begin{figure}[htp!]
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator.pdf}%
}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator2.pdf}%
}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator3.pdf}%
}
\subfloat[]{%
\includegraphics[clip,width=\columnwidth]{output_without_circulator4.pdf}%
}
\caption{Quantum state redistribution with control of propagation direction. We performed the same analysis as the higher dimensional HOM effect with direction control. By introducing phase shifters in the system before beam splitters, we can change the exit direction of the amplitudes. The starting state is $e_0f_0$. The first beam splitter transforms the input state, then they enter the multiport device. The multiport transformed state goes through beam splitters on the right and left side. The final outcome has the same form as the input state.}
\label{fig:control_without_circ}
\end{figure}
\begin{table}
\begin{tabular}{P{0.4\linewidth}P{0.6\linewidth}}
\hline \hline
\noalign{\vskip 1ex}
\multicolumn{2}{c}{State transformation with circulators}\\
\noalign{\vskip 1ex}
\hline \hline
\noalign{\vskip 1ex}
Indistinguishable photons & $a_{H0}b_{H0} \rightarrow -\frac{1}{2}(e_{H0}^2 - e_{H1}^2)$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with +1 relative phase & $\frac{1}{2}(a_{H0}^2 + b_{H0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 + e_{H1}^2)$\\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with $-1$ relative phase & $\frac{1}{2}(a_{H0}^2 - b_{H0}^2) \rightarrow -e_{H0}e_{H1}$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable photons & $a_{H0}b_{V0} \rightarrow -\frac{1}{2}(e_{H0}-e_{H1})(e_{V0}+e_{V1})$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable HOM pair & $\frac{1}{2}(a_{H0}^2 \pm b_{V0}^2) \rightarrow \frac{1}{4}\{(e_{H0}-e_{H1})^2\pm(e_{V0}-e_{V1})^2\}$ \\
\noalign{\vskip 1ex}
\hline \hline
\end{tabular}
\caption{State transformations in a system with circulators.
The first three states deal with indistinguishable photons by giving them the same polarization.
A state consisting of two single photons will become an HOM state.
We analyzed HOM states as an initial state, and they become either the HOM state or a two single-photon state.
Distinguishable photons are also analyzed by introducing orthogonal polarizations.
The output states become coupled states meaning the original states are not preserved.
} \label{tab:table_1}
\end{table}
\begin{table}
\begin{tabular}{P{0.4\linewidth}P{0.6\linewidth}}
\hline \hline
\noalign{\vskip 1ex}
\multicolumn{2}{c}{State transformation without circulators} \\
\noalign{\vskip 1ex}
\hline \hline
\noalign{\vskip 1ex}
Indistinguishable photons & $e_{H0}f_{H0} \rightarrow -e_{H0}e_{H1}$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with +1 relative phase & $\frac{1}{2}(e_{H0}^2 + f_{H0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 + e_{H1}^2)$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
HOM pair with $-1$ relative phase & $\frac{1}{2}(e_{H0}^2 - f_{H0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 - f_{H1}^2)$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable photons & $e_{H0}f_{V0} \rightarrow -e_{H0}e_{V1}$ \\
\noalign{\vskip 1ex}
\hline
\noalign{\vskip 1ex}
Distinguishable HOM pair & $\frac{1}{2}(e_{H0}^2 \pm f_{V0}^2) \rightarrow \frac{1}{2}(e_{H0}^2 \pm f_{V1}^2)$ \\
\noalign{\vskip 1ex}
\hline \hline
\end{tabular}
\caption{State transformations in a system without circulators.
The structure of the table is the same as the Table. I.
The first three states deal with indistinguishable photons by giving them the same polarization.
The last two states handle distinguishable photons.
The output states preserve the same form as the input state.
We start the transformation from the system location 0, then the transformed states are redistributed between location 0 and location 1.
The result shows coherent transportation of input states.} \label{tab:table_2}
\end{table}
\vspace{-10px}
\subsection{Delayed state redistribution}
\vspace{-10px}
\begin{figure*}[htp!]
\centering
\subfloat[][State redistribution without reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_array.pdf}\label{}}
\subfloat[][State redistribution with reflection]{\includegraphics[clip,width=\columnwidth]{multiport_two_array_phase_plate.pdf}\label{}}
\caption{Delayed state redistribution.
The two-photon amplitude transformation progresses in time from top to bottom.
The distance traveled in a single time step is indicated by vertical dashed lines. The total photon numbers are two in the system through out the propagation. At the first step for both cases, the input two-photon state is transformed by the BS.
The transformed state becomes the HOM state, and it is indicated as red transparent overlapped circles occupying both modes.
The initial and the final transformed state are indicated using solid red circles, and intermediate states are indicated in striped yellow circles.
(a) Two multiports and beam splitters {\it without} phase shifters between the multiports.
The HOM state enters the multiport and transformed taking the form of $-\frac{1}{2}(a_{0}-b_{0})(c_{0}+d_{0})$.
The amplitudes are coupled, however, they propagate without changing its amplitude. After several steps, the amplitudes occupying two rails converges to a single mode state after transformation by beam splitters.
The final state has the same form as the input state.
(b) Two multiports and beam splitters {\it with} a phase shifter P set at $\pi$ between multiports.
When the P is present, the right-moving coupled amplitude gains a relative phase between modes $a_1$ and $b_1$.
Reflection occurs at the multiport when the relative phase between the two is $\pi$.
Therefore, the transformed amplitude reflects upon a second multiport encounter, going back to the original state with opposite propagation direction.
Reflection does not occur on this transformed coupled left-moving amplitude, therefore it continues to propagate leftward.
The original left-moving amplitude becomes available for detection earlier than the transformed left-moving amplitude.}
\label{fig:delayed_state_dist}
\end{figure*}
We introduce the temporal delay effect as the higher dimensional HOM case by introducing a phase shifter between two multiports.
\vspace{-10px}
\subsubsection{Without reflection}
\vspace{-10px}
When there is no phase shifter between the two multiports, the result is identical to the system with a single multiport from the previous section.
The state transformation and propagation is provided schematically in Fig. \ref{fig:delayed_state_dist} (a).
The photons are initially sent from the left side of the BS.
The correlated photons are transformed to HOM state through the BS.
\vspace{-10px}
\begin{eqnarray}
&{e_{0}f_{0}}_R \xrightarrow{BS} \frac{1}{2}(a_{0}^2-b_{0}^2)_R \nonumber\\
&\xrightarrow{M}-\frac{1}{2}(a_{0}-b_{0})_L(c_{0}+d_{0})_R.
\end{eqnarray}
The HOM state is transformed by the multiport device.
This state is in a coupled state because right moving and left moving amplitudes are not separated.
We propagate this state through the BS on the left and translate the amplitudes moving to the right.
\begin{eqnarray}
&\xrightarrow{BS,T}-\frac{1}{\sqrt{2}}f_{0L}(a_{1}+b_{1})_R \xrightarrow{M} -\frac{1}{\sqrt{2}}f_{0L}(c_{1}+d_{1})_R\nonumber\\
&\xrightarrow{BS} -f_{0T_0L}e_{1T_1R}
\end{eqnarray}
The left moving amplitude is transformed by the left BS while right moving amplitude propagates to the second multiport device.
We introduced temporal difference between the right moving and the left moving photons.
\vspace{-15px}
\subsubsection{With reflection}
\vspace{-10px}
Reflection of amplitudes are introduced when there is a phase shifter between two multiport devices as indicated in fig. \ref{fig:delayed_state_dist} (b).
\begin{eqnarray}
&{e_{0}f_{0}}_R \xrightarrow{BS} \frac{1}{2}(a_{0}^2-b_{0}^2)_R
\xrightarrow{M} -\frac{1}{2}(a_{0}-b_{0})_L(c_{0}+d_{0})_R \nonumber \\
&\xrightarrow{BS,T+P}-\frac{1}{\sqrt{2}}f_{0L}(a_{1}-b_{1})_R
\end{eqnarray}
The right moving amplitude gains relative phase between upper and lower rails, and this relative phase allows the amplitude to get reflected upon multiport encounter.
\begin{eqnarray}
&\xrightarrow{M} -\frac{1}{\sqrt{2}}f_{0L}(a_{1}-b_{1})_L \xrightarrow{T+P} -\frac{1}{\sqrt{2}}f_{0L}(c_{0}+d_{0})_L\nonumber\\
&\xrightarrow{M} -\frac{1}{\sqrt{2}}f_{0L}(a_{0}+b_{0})_L \xrightarrow{BS} -f_{0T_{0}L} e_{0T_{2}L}
\end{eqnarray}
The input photons do not have any delays between the two at the beginning. The delay $\Delta T = T_2 - T_0$ is introduced from the reflection in the system.
\vspace{-10pt}
\section{Conclusion}
\vspace{-10px}
We demonstrated higher dimensional quantum state manipulation such as the HOM effect and state redistribution by applying linear-optical four-ports realizing four-dimensional Grover matrix accompanied by beam splitters and phase shifters.
Identical photons are sent into two of the four input-output ports and split into right-moving and left-moving amplitudes, with no cross terms to observe the HOM effect.
This absolute separation of propagation direction without mixing of right-moving and left-moving amplitudes insures the photons remain clustered as they propagate through the system.
Variable phase shifts in the system allow the HOM photon pairs to switch between four spatial output destinations, which can increase information capacity.
Time delays between emerging parts of the clustered two-photon state illustrating “delayed” HOM effect can be engineered using two multiports.
In addition, depending on the phase shifter position, the propagation direction can be reversed so that the right moving amplitude can get reflected at the second multiport, resulting in HOM pairs always leaving only from the left side of the system and with a particular time-bin delay.
The same situations have been investigated in a system without circulators. This system allows to redistribute the input state between the right and the left side of the system without changing amplitudes.
The HOM effect and clustered photon pairs are widely used in quantum information science. The approach introduced here adds extra degrees of freedom, and paves the way for new applications that require control over the spatial and temporal modes of the HOM amplitudes as they move through one- and two-dimensional networks.
We have demonstrated two photon amplitude control in both spatial and temporal modes. This two photon system can be extended to multiphoton input states, and manipulation of more complex entangled states would be the next milestones to be achieved.
\vspace{-15pt}
\section*{Appendix}
| 22a79c1050fb2f65ce3f6226fd3438a0725b2f57 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{section intro}
The lifting of representations between reductive algebraic groups plays an important role in representation theory.
The endoscopic transfer in Langlands functionality and the theta correspondence in Howe's reductive dual pairs are primary examples.
Dirac operators are employed for geometric construction of discrete series by Parthasarathy \cite{P}, Atiyah and Schmid \cite{AS}, and tempered representations by Wolf \cite{W}. In the late 1990's, Vogan made a conjecture on the algebraic property of the Dirac operators in Lie algebra setting. This conjecture was proved by Pandzic and myself in 2002 \cite{HP1}. This led us to study Dirac cohomology of Harish-Chandra modules \cite{HP2}.
Kostant extended the concept of Dirac cohomology and Vogan's conjecture to the more general setting of the cubic Dirac operator \cite{Ko2}.
In the formulation of central problems in Langlands program stable conjugacy plays an pivotal role. The theory of endoscopy investigates the difference between orbital integral over ordinary and stable conjugacy classes. As Dirac cohomology of a Harish-Chandra module determines its K-character \cite{HPZ}, it corresponds to the dual object of the orbital integral on elliptic elements. In this note we focus on using Dirac cohomology
for endoscopic transfer. The transfer factor
is difference of characters for the even and odd parts of the spin modules, or Dirac index of the trivial representation \cite{H}.
The aim of this note is to extend the calculation of lifting of characters to tempered representations with nonzero Dirac cohomology.
The ultimate goal is to understand lifting of characters for all unitary representations with nonzero Dirac cohomology (the Dirac series).
Jeff Adams \cite{A} defined a lifting on characters between orthogonal groups and nonlinear metaplectic groups over real numbers,
and it was extended to the p-adic case by Tatiana Howard \cite{Hd}.
This lifting of characters is closely related to both endoscopy and theta correspondence,
which also appears in the work of David Renard \cite{R}
and Wen-Wei Li \cite {Li}.
The Adams lifting of characters for orthogonal and
symplectic groups is given by the formal difference of the oscillator representations of
metaplectic groups and are related to the symplectic Dirac cohomology for Lie superalgebra \cite{HP3}.
Since the symplectic Dirac cohomology of modules for Lie superalgebras is very different from
the Dirac cohomology of Harish-Chandra modules, we deal with the transfer factor of the Adams lifting
in another paper \cite{H2}.
\section{Preliminaries on Dirac cohomology}
For a real reductive group $G$ with a Cartan involution $\theta$, denote by $\mathfrak{g}_0$ its Lie algebra
and assume that $K=G^\theta$ is a maximal compact subgroup of $G$.
Let $\mathfrak{g}=\mathfrak{k}\oplus\mathfrak{p}$ be the Cartan decomposition
for the complexified Lie algebra of $G$. Let $B$ be a non-degenerate invariant symmetric bilinear
form on $\mathfrak{g}$, which restricts to the Killing form on the semisimple part $[\mathfrak{g},\mathfrak{g}]$ of $\mathfrak{g}$.
Let $U(\mathfrak{g})$ be the universal enveloping
algebra of $\mathfrak{g}$ and $C(\mathfrak{p})$ the Clifford algebra of
$\mathfrak{p}$ with respect to $B$. Then one can consider the following version of the Dirac
operator:
$$
D=\sum_{i=1}^n Z_i\otimes Z_i \in U(\mathfrak{g})\otimes C(\mathfrak{p});
$$
here $Z_1,\dots,Z_n$ is an orthonormal basis of $\mathfrak{p}$ with respect to the symmetric bilinear
form $B$. It follows that $D$ is independent of the choice of the orthonomal basis $Z_1,\dots,Z_n$
and it is invariant under the diagonal adjoint action of $K$.
The Dirac operator $D$ is a square root of Laplace operator
associated to the symmetric pair $(\mathfrak{g},\mathfrak{k})$. To explain this, we
start with a Lie algebra map
\begin{equation*}
\alpha:\mathfrak{k}\rightarrow C(\mathfrak{p})
\end{equation*}
which is defined by the adjoint map $\mathop{\hbox {ad}}\nolimits:\mathfrak{k}\rightarrow\mathfrak{so}(\mathfrak{p})$
composed with the embedding of $\mathfrak{so}(\mathfrak{p})$ into $C(\mathfrak{p})$ using
the identification $\mathfrak{so}(\mathfrak{p})\simeq\bigwedge^2\mathfrak{p}$. The explicit
formula for $\alpha$ is (see \cite[\S2.3.3]{HP2})
\begin{equation}
\alpha(X)=-\frac{1}{4}\sum_{j}[X,Z_j]Z_j.
\end{equation}
Using $\alpha$ we can embed the Lie algebra $\mathfrak{k}$ diagonally into
$U(\mathfrak{g})\otimes C(\mathfrak{p})$, by
\begin{equation*}\label{Delta map}
X\mapsto X_\Delta=X\otimes1+1\otimes\alpha(X).
\end{equation*}
This embedding extends to $U(\mathfrak{k})$. We denote the image of $\mathfrak{k}$
by $\mathfrak{k}_\Delta$, and then the image of $U(\mathfrak{k})$ is the enveloping
algebra $U(\mathfrak{k}_\Delta)$ of $\mathfrak{k}_\Delta$.
Let $\Omega_\frak g$ be the Casimir operator
for $\frak g$, given by $\Omega_\frak g = \sum Z_i^2 -\sum W_j^2$,
where $W_j$ is an orthonormal basis for $\frak k_0$ with respect to
the inner product $-B$, where $B$ is the Killing form. Let
$\Omega_{\mathfrak{k}}=-\sum W_j^2$ be the Casimir operator for $\frak k$.
The image of $\Omega_\mathfrak{k}$ under $\Delta$ is denoted by
$\Omega_{\mathfrak{k}_\Delta}$.
Then
\begin{equation}\label{D^2}
D^2 = -\Omega_\frak g\otimes 1 + \Omega_{\frak k_\Delta} +
(||\rho_c||^2-||\rho||^2)1\otimes 1,
\end{equation}
where $\rho$ and $\rho_c$ are half sums of positive roots and compact
positive roots respectively.
The Vogan conjecture says that every element $z\otimes 1$ of
$Z(\mathfrak{g})\otimes 1 \subset U(\mathfrak{g})\otimes C(\mathfrak{p})$ can be written as
$$\zeta(z)+Da+bD$$
where $\zeta(z)$ is in $Z(\k_\Delta)$,
and $a,b\in U(\mathfrak{g})\otimes C(\mathfrak{p})$.
A main result in \cite{HP1} is introducing a differential $d$
on the $K$-invariants in $U(\mathfrak{g})\otimes C(\mathfrak{p})$ defined by a super bracket with $D$,
and determination of the cohomology of this
differential complex. As a consequence,
Pand\v{z}i\'c and I proved the following theorem.
In the following we denote by $\mathfrak{h}$ a Cartan subalgebra of $\mathfrak{g}$
containing a Cartan subalgebra $\mathfrak{t}$ of $\mathfrak{k}$ so that $\mathfrak{t}^*$ is embedded into $\mathfrak{h}^*$,
and by $W$ and $W_K$ the Weyl groups of $(\mathfrak{g},\mathfrak{h})$ and $(\mathfrak{k},\mathfrak{t})$
respectively.
\begin{thm}[\cite{HP1}]
Let $\zeta:
Z(\mathfrak{g})\rightarrow Z(\mathfrak{k})\cong Z(\mathfrak{k}_{\Delta})$ be the algebra homomorphism
that is determined
by the following commutative diagram:
\begin{equation*}\label{Vogan's diagram}
\CD
Z(\mathfrak{g}) @> \zeta >> Z(\mathfrak{k}) \\
@V \eta VV @V \eta_{\mathfrak{k}} VV \\
P(\mathfrak{h}^*)^{W} @>\mathrm{Res}>>
P(\mathfrak{t}^*)^{W_K},
\endCD
\end{equation*}
where $P$ denotes the polynomial algebra, and vertical maps $\eta$ and $\eta_\mathfrak{k}$ are Harish-Chandra
isomorphisms.
Then for each $z\in Z(\mathfrak{g})$ one has
\begin{equation*}\label{Vogan's equation}
z\otimes 1-\zeta(z)=Da+aD, \text{\ for some\ }a\in U(\mathfrak{g})\otimes C(\mathfrak{p}).
\end{equation*}
\end{thm}
For any admissible $(\mathfrak{g},K)$-module $X$,
Vogan (\cite{V}, \cite{HP1}) introduced the notion of \emph{Dirac cohomology} $H_D(X)$ of $X$.
Consider the action of the Dirac operator $D$ on $X\otimes S$,
with $S$ the spinor module for the Clifford algebra $C(\mathfrak{p})$. The Dirac cohomology is defined as follows:
$$H_D(X)\colon =\mathop{\hbox{Ker}}\nolimits D/ (\mathop{\hbox {Im}}\nolimits D \cap \mathop{\hbox{Ker}}\nolimits D).$$
It follows from the identity (\ref{D^2}) that $H_D(X)$ is a finite-dimensional module
for the spin double cover $\widetilde{K}$ of $K$. In case $X$ is unitary, $H_D(X)=\mathop{\hbox{Ker}}\nolimits D=\mathop{\hbox{Ker}}\nolimits D^2$
since $D$ is self-adjoint with respect to a natural Hermitian inner product on $X\otimes S$.
As a consequence of the above theorem, we have that
$H_D(X)$, if nonzero, determines the infinitesimal character of $X$.
\begin{thm}[\cite{HP1}]
Let $X$ be an admissible $(\mathfrak{g},K)$-module with standard infinitesimal character parameter
$\Lambda\in \mathfrak{h}^*$. Suppose that $H_D(X)$ contains a representation of
$\widetilde{K}$ with infinitesimal character $\lambda$. Then
$\Lambda$ and $\lambda\in\mathfrak{t}^*\subseteq \mathfrak{h}^*$ are conjugate under $W$.
\end{thm}
The above theorem is proved in \cite{HP1} for a connected semisimple
Lie group $G$. It is straightforward to extend the result to a
possibly disconnected reductive Lie group in Harish-Chandra's class \cite{DH}.
Vogan's conjecture implies a refinement of the celebrated
Parthasarathy's Dirac inequality, which is an extremely useful tool
for the classification of irreducible unitary representations of
reductive Lie groups.
\begin{thm}[Extended Dirac Inequality \cite{P}, \cite{HP1}]
Let $X$ be an irreducible unitary $(\mathfrak{g},K)$-module with infinitesimal
character $\Lambda$. Fix a representation of $K$ occurring in $X$
with a highest weight $\mu\in\t^*$, and a positive root system $\Delta^+(\mathfrak{g})$
for $\mathfrak{t}$ in $\mathfrak{g}$. Here $\t$ is a Cartan subalgebra of $\k$.
Write
$$\rho_c=\rho(\Delta^+(\mathfrak{k})),\ \rho_n=\rho(\Delta^+(\mathfrak{p})).$$
Fix an element $w\in W_K$ such that $w(\mu-\rho_n)$ is dominant
for $\Delta^+(\mathfrak{k})$. Then
$$\langle w(\mu-\rho_n)+\rho_c,w(\mu-\rho_n)+\rho_c\rangle
\geq \langle \Lambda,\Lambda\rangle.$$
The equality holds if and only if there exists a $w\in W$
such that
$$\Lambda = w(\mu-\rho_n)+\rho_c.$$
\end{thm}
\section{Dirac series and elliptic representations}
We say that an irreducible representation is in the {\it Dirac series} if it is a unitary representation with nonzero Dirac cohomology.
The Dirac series contains several families of very important unitary
representations including discrete series, unitary highest weight modules and unitary representations with nonzero
$(\mathfrak{g},K)$-cohomology. As shown in the previous section, a unitary $A_\mathfrak{q}(\lambda)$-module
with admissible $\lambda$ is a Dirac series. Any irreducible unitary highest
module is also a Dirac series. Thus, there exists Dirac series other than $A_\mathfrak{q}(\lambda)$-module.
We now discuss the relationship between Dirac series and elliptic representations.
Let $G$ be a connected semisimple algebraic group
over a local field $F$ of characteristic $0$. Arthur \cite{A1} studied a subset $\Pi_{temp, ell}(G(F))$
of tempered representations of $G(F)$, namely elliptic tempered representations.
The set of tempered representations $\Pi_{temp}(G(F))$ includes the discrete series
and in general the irreducible constituents of representations induced from discrete series.
These are exactly the representations which occur
in the Plancherel formula for $G(F)$.
In Harish-Chandra's theory\cite{HC1,HC2}, the character of an infinite dimensional representation $\pi$ is defined as a distribution
$$\Theta(\pi,f)=\mathop{\hbox {tr}}\nolimits \big{(} \int_{G(F)}f(x)\pi(x)dx\ \big{)}, \ \ \ \ f\in C^\infty_c(G(F) ),$$
which can be identified with a function on $G(F)$. In other words,
$$\Theta(\pi,f)=\int_{G(F)}f(x)\Theta(\pi,x)dx, \ \ \ \ f\in C^\infty_c(G(F)),$$
where $\Theta(\pi,x)$ is a locally integrable function on $G(F)$ that is smooth on the
open dense subset $G_{reg}(F)$ of regular elements.
A representation $\pi$ is called elliptic if $\Theta(\pi,x)$ does not vanish on the set of
elliptic elements in $G_{reg}(F)$.
The central objects in \cite{A1} are the normalized
characters $\Phi(\pi,\gamma)$, namely the functions defined by
$$\Phi(\pi,\gamma)=|D(\gamma)|^{1\over 2}\Theta(\pi,\gamma),\ \pi \in \Pi_{temp, ell}(G(F)), \
\gamma\in G_{reg}(F),$$
where
$$D(\gamma)=\mathop{\hbox {det}}\nolimits(1-\mathop{\hbox {Ad}}\nolimits(\gamma))_{\mathfrak{g}/\mathfrak{g}_\gamma},$$
is the Weyl discriminant. We will show how this normalized
character $\Phi(\pi,\gamma)$ is related to the Dirac cohomology of the
Harish-Chandra module of $\pi$
for a real group $G(\bbR)$.
From now on we are concerned only with the real group $G(\bbR)$.
Note that $G(\bbR)$ has elliptic elements if and only if
it is of equal rank with $K(\bbR)$. We also assume
this equal rank condition.
Induced representations from proper parabolic subgroups are not elliptic.
Consider the quotient of Grothendieck group of the category of finite length
Harish-Chandra modules by the subspace generated by induced
representations. Let us call this quotient group the elliptic Grothendieck
group. Authur \cite{A1} found an orthonormal basis of this elliptic
Grothendieck group in terms of elliptic tempered (possibly virtual) characters.
Those characters are the super tempered distributions defined by Harish-Chandra \cite{HC3}.
The tempered elliptic representations for
the real group $G(\bbR)$ are the representations
with non-zero Dirac index, which are studied in \cite{Lab1}.
Labesse shows that the tempered elliptic representations are precisely the fundamental series.
We now discuss the general elliptic representations
and and their Dirac index.
Recall that if $X$ is an
admissible $(\mathfrak{g},K)$-module with $K$-type decomposition
$X=\bigoplus_{\lambda}m_\lambda E_\lambda$, then the $K$-character of $X$ is the formal series
\[
\mathop{\hbox {ch}}\nolimits X=\sum_{\lambda}m_\lambda\mathop{\hbox {ch}}\nolimits E_\lambda,
\]
where $\mathop{\hbox {ch}}\nolimits E_\lambda$ is the character of the irreducible $K$-module $E_\lambda$.
Moreover, this definition makes sense also for virtual $(\mathfrak{g},K)$-modules $X$;
in that case, the integers $m_\lambda$ can be negative. In the following
we will often deal with representations of the spin double cover $\widetilde{K}$ of $K$,
and not $K$, but we will still denote the corresponding character by $\mathop{\hbox {ch}}\nolimits$.
Since $\mathfrak{p}$ is even-dimensional, the spin module $S$ decomposes as $S^+\oplus S^-$, with the $\mathfrak{k}$-submodules
$S^\pm$ being the even respectively odd part of $S\cong{\textstyle\bigwedge}\mathfrak{p}^+$.
Let $X=X_\pi$ be the Harish-Chandra module of an irreducible admissible
representation $\pi$ of $G(\mathbb{R})$. We consider the following difference
of $\widetilde{K}$-modules, the spinor index of $X$:
\[
I(X)=X\otimes S^+-X\otimes S^-.
\]
It is a virtual $\widetilde{K}$-module, an integer combination of
finitely many $\widetilde{K}$-modules.
The Dirac operator $D$ induces the action of the following $\widetilde{K}$-equivariant operators
\[
D^{\pm}:X\otimes S^{\pm}\rightarrow X\otimes S^{\mp}.
\]
Since $D^2$ acts by a scalar on each $\widetilde{K}$-type, most of $\widetilde{K}$-modules in
$X\otimes S^+$ are the same as in $X\otimes S^-$. It is straightforward to show the following identity.
\begin{lemma}\cite[Lemma 8.1]{H} The spinor index is equal to the Euler characteristic
of Dirac cohomology, i.e.,
\[
I(X)=H_D^+(X)-H_D^-(X).
\]
\end{lemma}
The spinor index $I(X)$ is also called the Dirac index of $X$,
since it is equal to the index of $D^+$, in the sense of index for
a Fredholm operator. It is also identical to the Euler characteristic of Dirac cohomology $H_D(X)$.
We denote by $\theta(X)$ the character of $I(X)$.
In terms of characters, this reads
$$
\theta(X)=\mathop{\hbox {ch}}\nolimits I(X)=\mathop{\hbox {ch}}\nolimits X(\mathop{\hbox {ch}}\nolimits S^+-\mathop{\hbox {ch}}\nolimits S^-)=\mathop{\hbox {ch}}\nolimits H_D^+(X)-\mathop{\hbox {ch}}\nolimits H_D^-(X).
$$
If we view $\mathop{\hbox {ch}}\nolimits E_\lambda$ as functions on $K$, then the series
$$\mathop{\hbox {ch}}\nolimits X=\sum_{\lambda}m_\lambda\mathop{\hbox {ch}}\nolimits E_\lambda$$
converges to a distribution on $K$ and it coincides with
$\Theta(X)$ on $K\cap G_{reg}$, according to Harish-Chandra \cite{HC1}.
Then the absolute value $|\theta_\pi|$ coincides with the absolute value
$
|\Phi(\pi,\gamma)|=|D(\gamma)|^{1\over 2}|\Theta(\pi,\gamma)|
$
on regular elliptic elements. We write this fact as the following lemma.
\begin{lemma} \label{character-theta} For any regular elliptic elements $\gamma$, we have
$$|\theta_\pi(\gamma)|=|\Phi(\pi,\gamma)|.$$
\end{lemma}
The following theorem follows immediately from the lemma.
\begin{thm}\cite[Theorem 8.3]{H}
Let $\pi$ be an irreducible admissible representation of $G(\bbR)$
with Harish-Chandra module $X_\pi$.
Then $\pi$ is elliptic if and only if the Dirac index $I(X_\pi)\neq 0$.
\end{thm}
We also recall a result from \cite{H}.
\begin{thm}\cite[Theorem 10.5]{H} Suppose $\pi$ is an irreducible unitary elliptic representation of
$G(\bbR)$ with a regular infinitesimal character.
Then $X_\pi\cong A_\mathfrak{q}(\lambda)$.
\end{thm}
As a consequence of the above two theorems, we have the following
\begin{cor} Suppose that rank of $G(\bbR)$ is equal to rank of $K(\bbR)$. Then any Dirac series of $G(\bbR)$ with regular infinitesimal character is
an $A_\mathfrak{q}(\lambda)$-module.
\end{cor}
The Dirac index of a representation determines its character on compact Cartan subgroups.
As shown by Harish-Chandra \cite{HC3}, the character of a discrete series is determined
completely on the set of regular elliptic elements.
It is a natural question whether the $A_\mathfrak{q}(\lambda)$ has the same property,
namely whether the Dirac index determines the representation.
It was shown in \cite{HPV} that it is indeed true for most of simple Lie groups
except for a few exception. In those exceptions, one needs the Dirac cohomology
together with the rank of $[\mathfrak{l},\mathfrak{l}]$ to determine the corresponding representation $A_\mathfrak{q}(\lambda)$.
\section{Cubic Dirac operators and associated cohomology}
We now recall the definition of Kostant's cubic Dirac operator
and the basic properties of the corresponding Dirac cohomology.
Let $\mathfrak{g}$ be a semisimple complex Lie algebra with Killing form $B$.
Let $\mathfrak{r}\subset\mathfrak{g}$ be a reductive Lie subalgebra such that
$B|_{\mathfrak{r}\times \mathfrak{r}}$ is non-degenerate. Let
$\mathfrak{g}=\mathfrak{r}\oplus\mathfrak{s}$ be the orthogonal decomposition with respect
to $B$. Then the restriction
$B|_\mathfrak{s}$ is also non-degenerate. Denote by $C(\mathfrak{s})$ the Clifford
algebra of $\mathfrak{s}$ with
\begin{equation*}
uu'+u'u=-2B(u, u')
\end{equation*}
for all $u, u'\in\mathfrak{s}$. The above choice of sign is the same as in \cite{HP2}, but
different from the definition in \cite{Ko1}, as well as in \cite{Ko2}. The two different
choices of signs have no essential difference
since the two bilinear forms are equivalent over $\mathbb{C}$. Now
fix an orthonormal basis $Z_1, \ldots, Z_m$ of $\mathfrak{s}$. Kostant
\cite{Ko1} defines the cubic Dirac operator $D$ by
\begin{equation*}
D=\sum_{i=1}^m{Z_i\otimes Z_i+1\otimes v}\in U(\mathfrak{g})\otimes C(\mathfrak{s}).
\end{equation*}
Here $v\in C(\mathfrak{s})$ is the image of the fundamental 3-form
$w\in\bigwedge^3(\mathfrak{s}^*)$,
\begin{equation*}
w(X,Y,Z)=\frac{1}{2}B(X,[Y,Z]),
\end{equation*}
under the Chevalley map $\bigwedge(\mathfrak{s}^*)\rightarrow C(\mathfrak{s})$ and
the identification of $\mathfrak{s}^*$ with $\mathfrak{s}$ by the Killing form $B$.
Explicitly,
\begin{equation*}
v=\frac{1}{2}\sum_{1\leq i<j<k \leq m}B([Z_i, Z_j],Z_k)Z_iZ_jZ_k.
\end{equation*}
The cubic Dirac operator has a good square in analogue with
the Dirac operator associated with the symmetric pair $(\mathfrak{g},\mathfrak{k})$ in Section 2. We
have a similar Lie algebra map
\begin{equation*}
\alpha:\mathfrak{r}\rightarrow C(\mathfrak{s})
\end{equation*}
which is defined by the adjoint map $\mathop{\hbox {ad}}\nolimits:\mathfrak{r}\rightarrow\mathfrak{so}(\mathfrak{s})$
composed with the embedding of $\mathfrak{so}(\mathfrak{s})$ into $C(\mathfrak{s})$ using
the identification $\mathfrak{so}(\mathfrak{s})\simeq\bigwedge^2\mathfrak{s}$. The explicit
formula for $\alpha$ is (see \cite[\S 2.3.3]{HP2})
\begin{equation}\label{mapalpha}\alpha(X)=-\frac{1}{4}\sum_{j}[X, Z_j]Z_j, \quad\
X\in\mathfrak{r}.
\end{equation}
Using $\alpha$ we can embed the Lie algebra $\mathfrak{r}$ diagonally into
$U(\mathfrak{g})\otimes C(\mathfrak{s})$, by
\begin{equation*}\label{Delta map}
X\mapsto X_\Delta=X\otimes1+1\otimes\alpha(X).
\end{equation*}
This embedding extends to $U(\mathfrak{r})$. We denote the image of $\mathfrak{r}$
by $\mathfrak{r}_\Delta$, and then the image of $U(\mathfrak{r})$ is the enveloping
algebra $U(\mathfrak{r}_\Delta)$ of $\mathfrak{r}_\Delta$. Let $\Omega_\mathfrak{g}$ (resp.
$\Omega_\mathfrak{r}$) be the Casimir elements for $\mathfrak{g}$ (resp. $\mathfrak{r}$).
The image of $\Omega_\mathfrak{r}$ under $\Delta$ is denoted by
$\Omega_{\mathfrak{r}_\Delta}$.
Let $\mathfrak{h}_\mathfrak{r}$ be a Cartan subalgebra of $\mathfrak{r}$ which is contained
in $\mathfrak{h}$. It follows from Kostant's calculation (\cite{Ko1}, Theorem 2.16) that
\begin{equation}\label{square}
D^2=-\Omega_\mathfrak{g}\otimes1+\Omega_{\mathfrak{r}_\Delta}-(\|\rho\|^2 - \|\rho_{\mathfrak{r}}\|^2)1\otimes1,
\end{equation}
where $\rho_\mathfrak{r}$ denote the half sum of positive roots for $(\mathfrak{r},
\mathfrak{h}_\mathfrak{r})$. We also note the sign difference with Kostant's formula
due to our choice of bilinear form for the definition of the Clifford algebra $C(\mathfrak{s})$.
We denote by $W$ the Weyl group associated to the root system
$\Delta(\mathfrak{g},\mathfrak{h})$ and $W_\mathfrak{r}$ the Weyl group associated to the root system
$\Delta(\mathfrak{r},\mathfrak{h}_\mathfrak{r})$. The following theorem due to Kostant is an
extension of Vogan's conjecture on the symmetric pair case which
is proved in \cite{HP1}.
(See \cite{Ko2} Theorems 4.1 and 4.2 or \cite{HP2} Theorem 4.1.4).
\begin{thm}\label{Vogan's conjecture1}
There is an algebra homomorphism
$$\zeta:
Z(\mathfrak{g})\rightarrow Z(\mathfrak{r})\cong Z(\mathfrak{r}_{\Delta})$$ such that for any $z\in Z(\mathfrak{g})$ one has
\begin{equation*}\label{Vogan's equation}
z\otimes 1-\zeta(z)=Da+aD \text{\ for some\ }a\in U(\mathfrak{g})\otimes C(\mathfrak{s}).
\end{equation*}
Moreover, $\zeta$ is determined
by the following commutative diagram:
\begin{equation*}\label{Vogan's diagram}
\CD
Z(\mathfrak{g}) @> \zeta >> Z(\mathfrak{r}) \\
@V \eta VV @V \eta_{\mathfrak{r}} VV \\
P(\mathfrak{h}^*)^{W} @>\mathrm{Res}>>
P(\mathfrak{h}_\mathfrak{r}^*)^{W_{\mathfrak{r}}}.
\endCD
\end{equation*}
Here the vertical maps $\eta$ and $\eta_\mathfrak{r}$ are Harish-Chandra
isomorphisms.
\end{thm}
\begin{defi} Let $S$ be a spin module of $C(\mathfrak{s})$.
Consider the action of $D$ on $V\otimes S$
\begin{equation}\label{Dirac map}
D:V\otimes S\rightarrow V\otimes S
\end{equation}
with $\mathfrak{g}$ acting on $V$ and $C(\mathfrak{s})$ on $S$. The {\it Dirac
cohomology} of $V$ is defined to be the $\mathfrak{r}$-module
\begin{equation*}
H_D(V):=\mathop{\hbox{Ker}}\nolimits D/(\mathop{\hbox{Ker}}\nolimits D\cap \mathop{\hbox {Im}}\nolimits D).
\end{equation*}
\end{defi}
The following theorem is a consequence of the above theorem.
\begin{thm}[\cite{Ko2},\cite{HP2}]
Let $V$ be a $\mathfrak{g}$-module with
$Z(\mathfrak{g})$ infinitesimal character $\chi_\Lambda$. Suppose that an
$\mathfrak{r}$-module $N$ is contained in the Dirac cohomology
$H_D(V)$ and has $Z(\mathfrak{r})$ infinitesimal
character $\chi_\lambda$ . Then $\lambda=w\Lambda$ for some $w\in W$.
\end{thm}
Suppose that $V_\lambda$ is a finite-dimensional representation
with highest weight $\lambda\in \mathfrak{h}^*$. Kostant \cite{Ko2} calculated
the Dirac cohomology of $V_\lambda$ with respect to any equal rank
quadratic subalgebra $\mathfrak{r}$ of $\mathfrak{g}$. Assume that $\mathfrak{h}\subset\mathfrak{r}\subset\mathfrak{g}$ is
the Cartan subalgebra for both $\mathfrak{r}$ and $\mathfrak{g}$.
Define $W(\mathfrak{g},\mathfrak{h})^1$ to be the subset of the Weyl group $W(\mathfrak{g},\mathfrak{h})$ by
$$
W(\mathfrak{g},\mathfrak{h})^1=\{w\in W(\mathfrak{g},\mathfrak{h})\ | w(\rho) \text{ is }\Delta^+(\mathfrak{r},\mathfrak{h})-\text{dominant}
\}.
$$
This is the same as the subset of elements $w\in W(\mathfrak{g},\mathfrak{h})$ that map the positive Weyl $\mathfrak{g}$-chamber
into the positive $\mathfrak{r}$-chamber. There is a bijection
$W(\mathfrak{r},\mathfrak{h})\times W(\mathfrak{g},\mathfrak{h})^1\rightarrow W(\mathfrak{g},\mathfrak{h})$ given by $(w,\tau)\mapsto w\tau$.
Kostant \cite{Ko2} proved the following result.
\begin{prop}[Kostant \cite{Ko2}] One has
$$
H_D(V_\lambda)=\bigoplus_{w\in W(\mathfrak{g},\mathfrak{h})^1}E_{w(\lambda+\rho)-\rho_\mathfrak{r}}.
$$
\end{prop}
\section{Endoscopic transfer factor}
Many important questions in harmonic analysis on Lie groups boil
down to the study of distributions on groups that are invariant under conjugacy.
The fundamental objects of invariant harmonic analysis are orbital integrals
as the geometric objects and characters of representations as the spectral objects.
The correspondence of these two kinds of objects reflects the core idea of harmonic analysis.
The orbital integrals are parameterized by the set of regular semisimple conjugacy classes in $G$.
Recall for such a $\gamma$, the orbital integral is defined as
$$\mathcal{O}_\gamma(f)=\int_{G/G_\gamma}f(x^{-1}\gamma x) dx, \ \ f\in C^\infty_c(G),$$
and the stable orbital integral is defined as
$$S\mathcal{O}_\gamma(f)=\sum_{\gamma'\in S(\gamma)}\mathcal{O}_{\gamma'}(f),$$
where $S(\gamma)$ is the stable conjugacy class.
Let $1\!\!1$ denote the trivial representation of $G$ and $\theta_{1\!\!1}$
the character of the Dirac index of the trivial representation. That is
$$\theta_{1\!\!1} = \mathop{\hbox {ch}}\nolimits H_D^+(1\!\!1)-\mathop{\hbox {ch}}\nolimits H_D^-(1\!\!1)=\mathop{\hbox {ch}}\nolimits S^+ -\mathop{\hbox {ch}}\nolimits S^-.$$
We note that
$$\overline{\theta_{1\!\!1}} =(-1)^q (\mathop{\hbox {ch}}\nolimits S^+ -\mathop{\hbox {ch}}\nolimits S^-)=(-1)^q \theta_{1\!\!1},$$
where $q={1\over 2} \mathop{\hbox {dim}}\nolimits G(\mathbb{R})/K(\mathbb{R})$.
Recall that $\theta_{\pi}$ denotes the character of the Dirac
index of $\pi$. If $\pi$ is the discrete series representation with Dirac cohomology $E_\mu$,
then $$\theta_{\pi}= (-1)^q \chi_\mu.$$
Labesse showed that there exists a function $f_\pi$ so that for any admissible
representations $\pi'$,
$$\mathop{\hbox {tr}}\nolimits \pi'(f_\pi)=\int_K \Theta_{\pi'}(k)\overline{\theta_{1\!\!1}\cdot \theta_{\pi}}dk.$$
Denote by $\theta_{\pi'}$ the character of its Dirac index for $\pi'$. Then one has
$$\mathop{\hbox {tr}}\nolimits \pi'(f_\pi)=(-1)^q\int_K \theta_{\pi'}\cdot \overline{\theta_{\pi}}dk.$$
Let $\pi'$ be a discrete series representation with Dirac cohomology $E_{\mu'}$.
It follows that
$$\mathop{\hbox {tr}}\nolimits \pi'(f_\pi)=\int_K \Theta_{\pi'}(k)\overline{\theta_{1\!\!1}\cdot \theta_{\pi}}dk(\chi_{\mu'},\chi_{\mu})=\mathop{\hbox {dim}}\nolimits \mathop{\hbox {Hom}}\nolimits_K(E_{\mu'},E_{\mu}).$$
Consequently we prove the following theorem due to Labesse.
\begin{thm}[Labesse \cite{Lab1}]
The function $f_\pi$ is a pseudo-coefficient for the discrete series $\pi$,
i.e., for any irreducible tempered representation $\pi'$,
$$\mathop{\hbox {tr}}\nolimits \pi'(f_\pi)=\begin{cases}1 \text{\ \ \ if \ }\pi\cong \pi'\\
0\text{\ \ \ \ otherwise}.\\ \end{cases} $$
\end{thm}
\begin{rmk}\label{character}
The orbital integrals of the pseudo-coefficient $f_\pi$ are easily computed for $\gamma$ regular semisimple:
$$\mathcal{O}_\gamma(f_\pi)=\begin{cases}\Theta_\pi(\gamma^{-1}) \text{\ \ \ if $\gamma$ is elliptic}\\
0\text{\ \ \ \ \ \ \ \ \ \ \ if $\gamma$ is not elliptic}.\\ \end{cases} $$
\end{rmk}
In the Langlands program a cruder form of conjugacy called stable conjugacy plays an important role.
The study of Langlands functoriality often leads to correspondence that is defined only up to stable conjugacy.
The endoscopy theory investigates the difference between ordinary and stable conjugacy
and how to understand ordinary conjugacy inside stable conjugacy.
The aim is to recover orbital integrals and characters from endoscopy groups.
The endoscopy theory for real groups is established by Shelstad in a series of papers
[Sh1-5].
Recasting Shelstad's work explicitly
in terms of the general transfer factors defined later by Langlands and
Shelstad [LS] is the first of the `Problems for Real Groups' proposed by Arthur \cite{A3}.
Recall that $G$ is a connected reductive algebraic group defined over $\mathbb{R}$.
Denote by $G^\vee$ the complex dual group and ${}^LG$ the $L$-group which is
the semidirect product of $G^\vee$ and the Weil group $W_\mathbb{R}$. A Langlands
parameter is an $L$-homomorphism
$$\phi\colon W_\mathbb{R}\rightarrow {}^LG.$$
Two Langlands parameters are equivalent if they are conjugated by an inner automorphism of $G^\vee$.
An equivalence class of Langlands parameters is associated
to a packet of irreducible admissible representations of
$G(\mathbb{R})$ [L2]. The $L$-packets of Langlands parameters with bounded image
consist of tempered representations. Temperedness is respected by
$L$-packets, but not unitarity.
The discrete series $L$-packets are in bijection with the irreducible finite-dimensional
representations of the same infinitesimal character.
One can construct all tempered
irreducible representations using unitary parabolic induction and by taking subrepresentations.
Two tempered irreducible representations $\pi$ and $\pi'$ are in the same
$L$-packet if up to equivalence, $\pi$ and $\pi'$ are subrepresentations of parabolically
induced representations from discrete series $\sigma$ and $\sigma'$ in the same
$L$-packets.
A stable distribution is any element of the closure of the space spanned by all distributions
of the form $\sum_{\pi\in \Pi}\Theta_\pi$ for $\Pi$ any tempered $L$-packet.
Such distributions can be transferred to inner forms of $G$ via the matching
of the stable orbital integrals, while unstable distributions cannot be.
In the setting of endoscopy embedding
$$\xi: {}^LH\rightarrow {}^LG,$$
one has a map from Langlands parameters for $H$ to that for $G$.
The Langlands functoriality principle asserts that there should be a
map from the Grothendieck group of virtual representations of $H(\mathbb{R})$ to
that of $G(\mathbb{R})$, compatible with $L$-packets.
We follow Labesse \S6.7 \cite{Lab2} for the description of the endoscopic transfer.
Let $T$ be an elliptic torus of $G$ and
$\kappa$ an endoscopic character. Let $H$ be the endoscopic
group defined by $(T,\kappa)$. Let $B_G$ be a Borel subgroup of $G$ containing $T$. Set
$$\Delta_B(\gamma)=\Pi_{\alpha>0}(1-\gamma^{-\alpha}),$$
where the product is over the positive roots defined by $B$. There is only one choice
of a Borel subgroup $B_H$ in $H$, containing $T_H$ and compatible with the isomorphism
$j\colon T_H\cong T$.
Assume $\eta\colon {}^LH\rightarrow {}^LG$ is an admissible embedding (see \S6.6 \cite{Lab2}).
Then for any pseudo-coefficent $f$ of a discrete series of $G$, there is a linear combination
$f^H$ of pseudo-coefficents of discrete series of $H$ such that for $\gamma=\gamma_G=j(\gamma_H)$ regular
in $T(\mathbb{R})$ (see Prop. 6.7.1 \cite{Lab2}), one has
\begin{equation}\label{endoscopy}
\mathcal{S}\mathcal{O}_{\gamma_H}(f^H)=\Delta(\gamma_H,\gamma)\mathcal{O}^\kappa_{\gamma}(f),
\end{equation}
where the transfer factor
\begin{equation*}\label{transfer}
\Delta(\gamma_H,\gamma)=(-1)^{q(G)-q(H)}\chi_{G,H}(\gamma)\Delta_B(\gamma^{-1})\Delta_{B_H}(\gamma_H^{-1})^{-1}.
\end{equation*}
The transfer $f\mapsto f^H$ of the pseudo-coefficents of discrete series can be extended to all of
functions in $C^\infty_c(G(\mathbb{R}))$ with extension of the correspondence $\gamma\mapsto \gamma_H$ (see
Theorem 6.7.2 \cite{Lab2}) so that the above identity (\ref{endoscopy}) holds for all $f$.
The geometric transfer $f\mapsto f^H$ is dual of a transfer for representations.
Given any admissible irreducible representation $\sigma$ of $H(\mathbb{R})$, it corresponds to
an element $\sigma_G$ in the Grothendieck group of virtual representations of $G(\mathbb{R})$ as follows.
Let $\phi$ be the Langlands parameter for $\sigma$.
Let $\Sigma$ be the $L$-packet of the admissible irreducible representations
of $H(\mathbb{R})$ corresponding to a Langlands parameter $\phi$ and $\Pi$ the L-packet of representations of
$G(\mathbb{R})$ corresponding to $\eta \circ \phi$ (that can be an empty set if this parameter
is not relevant for $G$).
\begin{thm} [Theorem 4.1.1 \cite{S}, Theorem 6.7.3 \cite{Lab2}] There is a function
$$\epsilon\colon \Pi \rightarrow \pm 1$$
such that, if we consider $\sigma_G$ in the Grothendieck group defined by
$$\sigma_G=\sum_{\pi\in\Pi}\epsilon(\pi)\pi$$
then the transfer $\sigma\mapsto \sigma_G$ satisfies
$$\mathop{\hbox {tr}}\nolimits \sigma_G(f)=\mathop{\hbox {tr}}\nolimits \sigma(f^H).$$
\end{thm}
In the following we suppose that $G(\mathbb{R})$ has a compact maximal torus $T(\mathbb{R})$,
and $\rho-\rho_H$ the difference of half sum of positive
roots for $G$ and $H$ respectively, defines a character of $T(\mathbb{R})$.
In \S 7.2 of [Lab2] Labesse shows that the canonical
transfer factor:
$$\Delta(\gamma^{-1})=(-1)^{q(G)-q(H)}
\frac{\sum_{w\in W(\mathfrak{g})}\epsilon (w)\gamma^{w\rho}}{\sum_{w\in W(\mathfrak{h})}\epsilon (w)\gamma^{w\rho_H}}$$
is well-defined function.
Then the transfer factor can be expressed more explicitly if $H$ is a subgroup of $G$.
Suppose that $\mathfrak{g}=\mathfrak{h}\oplus \mathfrak{s}$ is the orthogonal decomposition with
respect to a non-degenerate invariant bilinear form so that the form is non-degenerate on $\mathfrak{s}$.
We write $S(\mathfrak{g}/\mathfrak{h})$ for the spin-module of the Clifford algebra $C(\mathfrak{s})$. Then
$$\Delta(\gamma^{-1})=\mathop{\hbox {ch}}\nolimits S^+(\mathfrak{g}/\mathfrak{h}) -\mathop{\hbox {ch}}\nolimits S^-(\mathfrak{g}/\mathfrak{h}).$$
In other words, $\Delta(\gamma^{-1})$ is equal to the character of the Dirac index of
the trivial representation with respect to the Dirac operator $D(\mathfrak{g},\mathfrak{h})$.
If $\Theta_\pi$ is the character of a finite-dimensional representation $\pi$,
then
$$\Delta(\gamma^{-1})\Theta_\pi$$ is the character of the Dirac index of $\pi$.
This character can be calculated easily from the Kostant formula in Theorem 4.7.
We denote by $F_\lambda$ the irreducible finite-dimensional representation of $G(\mathbb{R})$
with highest weight $\lambda$ and by $E_\mu$ irreducible finite-dimensional representation
of $H(\mathbb{R})$ with highest weight $\mu$. Then
$$\Delta(\gamma^{-1})\Theta_{F_\lambda}= \sum_{w\in W^1}\Theta_{E_w(\lambda+\rho)-\rho_\mathfrak{h}}.$$
Here $W^1$ is a subset of elements in $W$ corresponding to $W_\mathfrak{h} \backslash W$ as before.
It is straightforward to use the transfer factor to calculate lifting of discrete series characters.
This lifting is closely related to the geometric transfer of the pseudo-coefficents of discrete series.
The Harish-Chandra module of a discrete series representation is isomorphic to $A_\mathfrak{b}(\lambda)$
for some $\theta$-stable Borel subalgebra and corresponding Harish-Chandra parameter is $\lambda+\rho$.
It follows from \cite[Theorem 7.5]{DH} that a tempered representations with nonzero Dirac cohomology is
$\pi_{\lambda+\rho}=A_\mathfrak{b}(\lambda)$ and it has Dirac cohomology equal to an irreducble $K$-module $E_{\lambda+\rho_n}$.
In the equal rank case, it is simply a limit of discrete series.
The calculation for discrete series extends to tempered representations with nonzero Dirac cohomology.
\begin{prop} Let $\pi_\lambda$ be a discrete series of $G$ with Harish-Chandra parameter $\lambda$.
Then we have
$$\Delta(\gamma^{-1})\Theta{\pi_\lambda}=\sum_{w\in W_{K}^1} \text{sign}(w)\Theta_{\tau_{w\lambda}}.$$
The above formula extends to limits of discrete series.
\end{prop}
\proof
In view of Remark \ref{character}, the right hand side of (\ref{endoscopy})
is the Dirac index of a combination of discrete series of $G(\mathbb{R})$
and the left hand side is a linear combination of discrete series of $H(\mathbb{R})$.
It follows from the Harish-Chandra formula for the character of discrete series
and supertempered distributions \cite{HC3} that
the Dirac index of a discrete series $\pi_\lambda$ with Harish-Chandra parameter $\lambda$ is
$$\Delta(\gamma^{-1})\Theta{\pi_\lambda}=\sum_{w\in W_{K}^1} \text{sign}(w)\Theta_{\tau_{w\lambda}}.$$
Here ${\tau_{w\lambda}}$ denotes the discrete series for $H(\mathbb{R})$ with Harish-Chandra parameter $w\lambda$, and
$W_{K}^1$ is a subset of elements in $W_{K}$ corresponding to $W_{H\cap K} \backslash W_{K}$.
This calculation is compatible with Labesse's calculation of the transfer of the pseudo-coefficients
of discrete series in \S 7.2 \cite{Lab2}.
It remians to show that this formula extends to tempered elliptic representations nonzero Dirac cohomology. By
Theorem 7.5 of \cite{DH} that a tempered elliptic representations nonzero Dirac cohomology is a limit of discrete sereis
$\pi_{\lambda+\rho}=A_\mathfrak{b}(\lambda)$ and it has Dirac cohomology equal to an irreducble $K$-module $E_{\lambda+\rho_n}$.
As the parmeter for a limit of discrete series, $\lambda+\rho$ is regular with respect to compact roots.
Thus, the same calculation for discrete series applies here.
\qed
\medskip
As a final remark, we note that we may use the Arthur packets to deal with non-tempered case.
The Arthur packets are parameterized by mappings
$$\psi\colon W_\mathbb{R} \times SL(2,\mathbb{C}) \rightarrow {}^LG$$
for which the projection onto the dual group $G^\vee$ of $\psi(W_\mathbb{R})$ is relatively compact.
Adams and Johnson \cite{AJ} have constructed some $A$-packets consisting of unitary $A_\mathfrak{q}(\lambda)$-modules.
As most of unitary $A_\mathfrak{q}(\lambda)$-modules can be classified
by their Dirac cohomology \cite{HPV}, the determination of Dirac cohomology of
$A_\mathfrak{q}(\lambda)$-modules may have some bearing on answering Arthur's questions (See \cite[\S9]{A2})
on Arthur packet $\Pi_\psi$.
| b25a88f166782dff168e7075840ccff42d739913 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sect:intro}
Generally, capturing high-quality images in dim light or back light conditions is challenging, since insufficient lighting can significantly degrade the visibility of images. Especially the lost details and low contrast would not only cause unpleasant subjective perceptions, but also hurt the performance of back-end computer vision systems which are designed for normal-light images. Though modern imaging sensors can automatically set high ISO, long exposure, and flash according to different circumstances to compensate for the low light, they suffer from different drawbacks.
One solution is to use high dynamic range (HDR) imaging techniques, which has been integrated into modern cameras to tackle with dark light environment in the image acquisition stage. However, when it comes to restore existing poor-quality low-light images, HDR needs bunch of images under different illumination conditions as inputs, which limits its practical application.
Thus the low-light image enhancement has been a long-standing problem in the community with a great progress made over the past years.
It can not only be used to increase the visual aesthetics of photos for people's daily use, but also to stable the performance of many computer vision algorithms such as object detection. Existing methods have two mainstream ideologies. One tries to consider the low-light enhancement problem as an image decomposition problem based on the Retinex theory, in which each image can be separated into independent components and the image reflectance part can be seen as a reasonable enhancement result \cite{fu2016weighted,guo2016lime,ren2018joint,zhang2018high}. However, these methods tend to introduce unexpected artifacts in the enhanced results. The other resorts to machine learning techniques based on large-scale image databases consisting of pairs of low-light image and corresponding enhanced image restored by image processing software like Photoshop \cite{wei2018deep,dale2009image,bychkovsky2011learning,shen2017msr}. These methods have achieved impressive performance but usually limited by the quality and quantity of the training database.
To get rid of the restriction of using training data, in this paper, we present a novel self-supervised low-light image enhancement framework called SID-NISM, which could restore the quality of any single low-light image only relying on the visual information of the image itself. Our major contributions are summarized as follows.
(1) A self-supervised image decomposition network, SID-Net, is proposed to decompose the input image into lighting-independent reflectance, structure-aware smooth illumination and reflectance-related noise straightforwardly according to the robust Retinex theory~\cite{li2018structure}. Taking as inputs any given input image and its corresponding histogram equalization image, SID-Net can converge to the optimal decomposed maps within limited iterations, in which two novel loss terms related to the formation of the reflectance and noise maps together with several common-used basic terms are adopted to guide the decomposition procedure preciously. As pointed out above, SID-Net is a image-specific network without depending on any prior training or reference image, which greatly distinguish it from existing supervised-learning methods like Retinex-Net~\cite{wei2018deep}.
(2) A nonlinear illumination saturation mapping function (NISM) is constructed to refine the decomposed low-light illumination map. Specifically, it could brighten up the whole image to a proper lighting level on the premise of preserving the contrast between foreground and background, which combats the weaknesses of Gamma correction in contrast preservation and bright regions enhancement. The final normal-light result can be restored by combining the denoised reflectance with the enhanced illumination.
Comprehensive experiments are conducted to illustrate the performance of the proposed method. Human perception user study suggests that people are more inclined to prefer the output of our method and find less unexpected artifacts in our results when the results of multiple competing methods are presented in front of them, which is consistent with the objective evaluations. Both subjective and objective experiments demonstrate the superiority of our method over the state-of-the-art methods in producing natural and attractive enhancement results.
\section{Related Work}\label{sect:related}
Researchers have devoted their efforts to solving the problem of low-light image enhancement in the past decades. Many techniques have been developed to improve the quality of low-light images, which can be classified into two major categories as mentioned in Sect. \ref{sect:intro}: Retinex-based methods and learning-based methods.
\textbf{Retinex-based Methods.} According to the classic Retinex theory~\cite{land1971lightness}, an observed image can be decomposed into two components, reflectance and illumination, in which the former one can be taken as a kind of compelling low-light image enhancement result. Jobson \textit{et al.}~\cite{jobson1997multiscale,jobson1997properties} made some early attempts based on this model, but their results are usually unrealistic. Wang \textit{et al.}~\cite{wang2013naturalness} presented a naturalness preserved method by utilizing a lightness-order-error measure. However, it tends to produce results with dim artifacts and requires expensive computational cost. Fu \textit{et al.}~\cite{fu2016weighted} and Ren \textit{et al.}~\cite{ren2018joint} succeed to simultaneously estimate reflectance and illumination maps through a weighted variational model and a sequential model respectively, but both suffered from preserving the property of Retinex model insufficiently. While Guo \textit{et al.}~\cite{guo2016lime} tired to only estimate the illumination map from the maximum values of three channels with constraints on preserving the main contour, then compute the reflectance by conducting element-wise division. And following the illumination estimation constraints in \cite{guo2016lime}, Zhang \textit{et al.}~\cite{zhang2018high} proposed two more constraints for estimating illumination based on perceptually bidirectional similarity. These two methods have pretty good performance but usually introduce unexpected artifacts in the enhanced images. Besides, with the development of neural network, some researchers~\cite{zhang2019kindling,wang2019underexposed,wei2018deep} sought to design image decomposition networks and illumination adjustment networks based on low-light and normal-light image pairs.
\textbf{Learning-based Methods.} Using machine learning tools to solve the problem of low-light image enhancement is a recent trend and also a promising direction. Dale \textit{et al.}~\cite{dale2009image} first established a database comprising 1 million images. Given an input image to be enhanced, their system executes a visual search to find the closest images in the database; these images define the input’s visual context, which can be further exploited to instantiate the restoration operations. Kang \textit{et al.}~\cite{kang2010personalization} constructed a database which stored the feature vectors describing training images along with vectors of enhancement parameters. Given a test image, the database was then searched for the best matching image, and the corresponding enhancement parameters were used to perform adjustment. Following the similar idea, Bychkovsky \textit{et al.}~\cite{bychkovsky2011learning} constructed a dataset containing 5,000 input-output image pairs that could be used to learn global tonal adjustments. To avoid collecting large-scale datasets, Shen \textit{et al.}~\cite{shen2017msr} trained their MSR-net designed upon the multi-scale Retinex theory on synthesized pairwise images.
With the power of neural network, these methods can get outstanding enhanced results, but at the same time it should be noticed that they are all based on supervised-learning frameworks and thus their performance highly depends on the quality and quantity of training datasets. Therefore Zhang \textit{et al.}~\cite{zhang2019zero} proposed the first unsupervised-learning back-lit image restoration network based on the S-curve theory~\cite{yuan2012automatic}, which estimated the image specific S-curve through region-level optimal exposure evaluation. However, it is limited by the mid-gray assumption \cite{guo2016lime} in S-curve, which leads to unrealistic enhanced results with halos and gray shadows.
\section{Retinex Theory}\label{sect:retinex}
The classic Retinex theory~\cite{land1971lightness} models the image compositions which could reflect the formation of low-light images to some extent. It assumes that the captured image can be decomposed into two components, reflectance and illumination. Let $S$ represents the source image, then the classic Retinex theory can be denoted by,
\begin{equation}\label{equ:retinex}
S = R \times L
\end{equation}
where $R$ is reflectance, $L$ is illumination and $\times$ is element-wise multiplication. Illumination $L$ refers to the various lightness on observed objects, while reflectance $R$ corresponds to the material RGB color that describes how objects reflect light, which is considered to be invariant to $L$ and other possible imaging conditions. In short, $R$ represents the intrinsic property of captured objects. Besides, in color images each channel of $R$ can be regarded as sharing the same grayscale illumination map $L$.
Since low-light images usually suffer from darkness and unbalanced illumination distributions, the low-light image enhancement problem can be regarded as a procedure of estimating the illumination-independent reflectance according to the Retinex theory, which aims to remove the illumination effect and recover the original appearance of the scene objects. However, directly decomposing an input image into reflectance and illumination yields the intrinsic image decomposition problem~\cite{grosse2009ground}, which is inherently ill-posed and may produce unrealistic results~\cite{guo2016lime}. Therefore, some researchers~\cite{guo2016lime,fu2016weighted,ren2018joint,wei2018deep} converted the decomposition problem into an optimization problem and then computed the optimal solution through conventional optimization solvers or machine learning techniques. Moreover, instead of taking image reflectance as the enhancement result directly, existing methods usually project the enhanced illumination back to the reflectance by $R \times f(L)$ at the end, where $f(\cdot)$ stands for a manipulation operator adopted to enhance the illumination such as Gamma correction.
Another noticeable issue in low-light image enhancement is the unexpected noise raised by enhancing dark regions. Although the low-light source image doesn't suffer from the noise problem, the noise hidden in the dark would be amplified along with stretching the contrast of dark regions. To address the intensive noise problem, researchers seek for help from denoising algorithms like BM3D~\cite{dabov2007image} as the post-processing method, which is straightforward but not designed to solve the specific denoising problem here. Therefore we decide to follow the robust Retinex model~\cite{li2018structure}, which integrate the noise term $N$ into the classic Retinex theory directly,
\begin{equation}\label{equ:retinex}
S = R \times L + N
\end{equation}
Once the source image could be decomposed into the three components successfully, the reflectance map would be not only independent with the illumination map but also get rid of unexpected noise.
\section{SID-NISM: Self-supervised Low-light Image Enhancement Framework} \label{sect:sid-nism}
In this section, the proposed low-light image enhancement framework SID-NISM is presented in details. Fig.~\ref{fig:Framework} illustrates the overall structure of SID-NISM, which consists of two stages, decomposition and enhancement, corresponding to the two main parts, SID-Net and NISM.
\begin{figure*}[t]
\centering
\includegraphics[scale=0.32]{framework-eps-converted-to.pdf}
\caption{The overall structure of SID-NISM. It consists of two stages, decomposition and enhancement, which are corresponding to the two main parts SID-Net and NISM.}
\label{fig:Framework}
\end{figure*}
\subsection{SID-Net: A Self-supervised Image Decomposition Network}
According to the definition of the Retinex theory, the reflectance component of an image should be invariant to the illumination component and other imaging conditions, which suggests that any pair of images with the same content should share the same reflectance regardless of their different illumination conditions.
Thus inspired by the observation, a self-supervised learning image decomposition network called SID-Net is designed to split given low-light image into its reflectance, illumination and noise directly by taking as inputs the original image and its histogram equalization version. Notice that since the histogram equalization image is served as one of the paired images under different illumination condition to share the same reflectance, not the ground-truth normal-light images as in the supervised-learning methods, it doesn't matter whether histogram equalization would produce images with unpleasant color distortion or unbalanced illumination. Besides, although theoretically capturing paired images or utilizing other image generation methods are acceptable as well, considering the practical application and the computational cost, histogram equalization image becomes the best option.
As illustrated in Fig.~\ref{fig:Framework}, SID-Net firstly takes as inputs the source image $S_{low}$ and its corresponding histogram equalization version $S_{he}$ to split both of them into three components according to the robust Retinex theory, then iteratively converges to the optimal decomposition maps by minimizing the designed loss function.
Since it is a self-supervised network without external information from training images, the key point of SID-Net is to devise reasonable and effective loss function to guide the network to generate expected decomposition maps. The loss function of SID-Net is formulated as follows,
\begin{equation}
\mathcal{L} = \lambda_{rc}\mathcal{L}_{rc} + \sum_{i=low,he} {(\mathcal{L}_{rec}^i + \lambda_{L}\mathcal{L}_{L}^i + \lambda_{R}\mathcal{L}_{R}^i + \lambda_{N}\mathcal{L}_{N}^i)}
\end{equation}
where $\lambda_{rc}$, $\lambda_{L}$, $\lambda_{R}$, $\lambda_{N}$ denote the coefficients to balance the reflectance similarities and the guidance of generating illumination, reflectance and noise maps. They are signi ficantly smaller than 1 to address the importance of the fidelity term $\mathcal{L}_{rec}$ in the optimization, which are 0.01, 0.1, 0.001, 0.01 respectively in experiments. Notice that the reconstruction loss $\mathcal{L}_{rec}$ and the three guidance loss terms $\mathcal{L}_{L}, \mathcal{L}_{R}, \mathcal{L}_{N}$ are both applied on the original low-light image $S_{low}$ and the histogram equalization image $S_{he}$ to separate them into reflectance $R_{low}, R_{he}$, illumination $L_{low}, L_{he}$ and noise $N_{low}, N_{he}$ in a self-supervised way.
\textbf{Retinex reconstruction loss.} First and foremost, according to the robust Retinex theory~\cite{li2018structure}, the three decomposed maps should be able to reconstruct the original image. Thus the reconstruction loss $\mathcal{L}_{rec}$ is formulated as,
\begin{equation}
\mathcal{L}_{rec} = \left\|R\times L+N-S\right\|_1
\end{equation}
It corresponds to the image formation, which could constrain the distance between estimated $R\times L+N$ and the source image $S$.
\textbf{Reflectance consistency loss.} As described in Sect.~\ref{sect:retinex}, the reflectance reflects the intrinsic property of captured objects, which should be invariant to the scene lighting and imaging conditions. That is to say, in the case of the same captured objects, i.e. image content, the reflectance maps of the low-light image and its histogram equalization version should be as close as possible. Therefore, based on the above assumption, the reflectance consistency loss $\mathcal{L}_{rc}$ is designed to constrain the reflectance similarities between $R_{low}$ and $R_{he}$.
\begin{equation}
\mathcal{L}_{rc} = \left\|R_{low}-R_{he}\right\|_1
\end{equation}
\textbf{Illumination smoothness and consistency loss.} The loss function designed for the illumination map $\mathcal{L}_{L}$ is consist of two parts, the smoothness term and the consistency term, which can be expressed as,
\begin{equation}
\begin{split}
\mathcal{L}_{L} &= {\left\|\nabla L \times \exp(-\alpha\nabla R)\right\|_1} \\
&+ \left\|\nabla L \times \exp(-\alpha\sum_{j=low,he}{\nabla L_j})\right\|_1
\end{split}
\end{equation}
where $\nabla$ denotes the gradient including $\nabla_h$ (horizontal) and $\nabla_v$ (vertical) and $\alpha$ denotes the coefficient balancing the strength of structure-awareness and edge-preservation which is set as 10 in experiments.
The first term is adopted to guide the illumination map to be piece-wise smooth in textural details while preserving the general structure of the original images. A common-used smoothness prior in various image restoration tasks, total variation minimization (TV)~\cite{ma2012tv,ng2011total}, which minimizes the gradient of the whole image, is selected here as the fundamental of the illumination smoothness term. The second one is designed to preserve strong mutual edges while depressing weak ones. Similar to the illumination smoothness loss, it utilizes the weighted version of TV loss. Notice that since TV loss is structure-blindness, the original TV function is weighted with the gradient of the reflectance~\cite{wei2018deep} and illumination maps~\cite{zhang2019kindling} respectively to make the constructed loss be aware of the image structure and the two illumination maps be consistent with each other in terms of the image edges.
\textbf{Reflectance contrast and color loss.} To generate accurate reflectance map, in addition to the basic reconstruction loss, a novel reflectance contrast and color loss $\mathcal{L}_{R}$ is introduced to improve the image contrast and restore the image color as much as possible, which can be expressed as,
\begin{equation}
\begin{split}
\mathcal{L}_{R} &= \frac{1}{3}\sum_{ch=R,G,B}{\left\|\nabla R^{ch} - \beta\nabla \hat{S}^{ch} \right\|_F} \\
&+ \left\|R^H-S^H\right\|_2
\end{split}
\end{equation}
where $\beta$ is the gradient amplification factor which is set as 10 in experiments. $R^H$ and $S^H$ are the hue channels of the reflectance map and the source image after converting them from the RGB to the HSV color space.
For the first term, as we know, low-light images always suffer from low contrast, which often indicates smaller gradient magnitudes~\cite{li2018structure}. Hence we attempt to manipulate the gradient magnitudes of the reflectance to boost the contrast of the enhanced results by amplifying the gradient of the input image with the factor $\beta$. Notice that $\nabla \hat{S}$ is a variant of the gradient of the input image $\nabla{S}$, in which small gradients, i.e. the noise, are suppressed before amplification.
\begin{equation}
\nabla \hat{S}=
\begin{cases}
0& \text{if } |\nabla{S}|<\epsilon\\
\nabla{S}& \text{otherwise}
\end{cases}`
\end{equation}
The second term is adopted to avoid color mismatch in the decomposed reflectance, which enforce the hue component of the reflectance map to be close enough to that of the source image.
\textbf{Noise estimation loss.} Although the reflectance contrast loss attempts to suppress noise by ignoring small gradients, noise may still appear in flat dark regions. It is necessary to guide the decomposition of the noise map by,
\begin{equation}
\mathcal{L}_{N} = \left\|S \times N\right\|_F
\end{equation}
which constrains the overall intensity of the noise~\cite{li2018structure}.
As for the detailed network configuration, SID-Net first adopts a commonly used feature extractor, which could be any classic network such as ResNet, U-Net, or even a simple CNN structure. Then, three separate $3\times3$ convolutional layers project the image features into branches, namely reflectance $R$, illumination $L$, and noise $N$ respectively. At the end, sigmoid function is used to constrain both $R$ and $L$ in the range of [0, 1], while tanh function is adopted to simulate additive noise to make $N$ fall in [-1, 1].
\subsection{NISM: Nonlinear Illumination Saturation Mapping Function}
After obtaining the low-light illumination map $L$ of the given image $S$ by SID-Net, a manipulation operator would be adopted to enhance $L$ as described in Sect.~\ref{sect:retinex}. In the previous methods~\cite{guo2016lime,ren2018joint,zhang2018high}, researchers got used to refine the illumination map through Gamma correction, say $\hat{L}=L^{1/\gamma}$. When $\gamma>1$, the illumination map will be brightened up. In Fig.~\ref{fig:NISM}(a), the function curve of Gamma transformation with $\gamma=2.2$ is shown by the red line. It can be seen that pixels whose intensity are less than 20\% of the maximum illumination value, namely 0.2 in the scale of [0, 1], will be brightened to 50\%, while the pixels whose original intensity is large enough ($>80\%$) will keep the themselves basically unchanged. This property of Gamma correction would definitely do favor to enhance the illumination map especially in those dark regions, but it also brings two major flaws: 1) the contrast of the original image is destroyed due to the over-brightening of dark areas; 2) the illumination level of the final enhanced image is still insufficient because the bright areas are barely changed. As shown in Fig.~\ref{fig:NISM}(c), the overall quality of the enhanced result is improved compared with the low-light input Fig.~\ref{fig:NISM}(b), but it haven't been unable to reach the standard of 'normal-light' in human perception.
\begin{figure}
\centering
\begin{minipage}[b]{0.3\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{curve-eps-converted-to.pdf}}
\centerline{(a) Curves}\medskip
\end{minipage}
\begin{minipage}[b]{0.35\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{547-low.png}}
\centerline{(b) Input}\medskip
\end{minipage}\\
\begin{minipage}[b]{0.35\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{547-gamma.png}}
\centerline{(c) Gamma}\medskip
\end{minipage}
\begin{minipage}[b]{0.35\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{547-NISM.png}}
\centerline{(d) NISM}\medskip
\end{minipage}
\caption{Comparison between the Gamma correction and the proposed manipulation operator NISM. (a) are the two function curves when $\gamma=2.2$ and $\eta=2.2$. (c) and (d) are the corresponding enhanced results when adopting Gamma correction and NISM on the illumination map of the low-light image (b) respectively.}
\label{fig:NISM}
\end{figure}
To combat the existing issues of Gamma correction, we consider depressing the brightening of dark pixels while increasing the illumination saturation of bright pixels. More concretely, for dark pixels, it is still necessary to be brightened up but should be carried on at a relatively lower level as compared with Gamma correction. In other words, the slope of the illumination operator at pixels whose intensity are less than 0.2 should be smaller than that of Gamma correction. On the contrary, the slope at bright pixels should be larger than that of Gamma correction, which could increase the illumination levels of bright pixels so as to render the final enhanced results bright enough.
According to the above considerations, a novel manipulation operator called Nonlinear Illumination Saturation Mapping (NISM) is proposed, which is formulated as,
\begin{equation}
\hat{L}=1-(1-L)^\eta
\end{equation}
where $\eta$ is the control parameter as $\gamma$ in Gamma correction. The function curve of NISM with $\eta=2.2$ is depicted in the blue line in Fig.~\ref{fig:NISM}(a). It can be seen that the curve shape of NISM is consistent with the previous analysis about the improved curve slope. Indeed, the constructed NISM and the Gamma correction are symmetrical about $y=1-x$.
The final enhanced image is generated by recombining the decomposed reflectance $R$ and the refined illumination $\hat{L}$ by $R\times\hat{L}$. The corresponding result is shown in Fig.~\ref{fig:NISM}(d), which reflects that NISM could transform the decomposed illumination map of the original low-light image into a normal-light level.
In implementation $\eta$ is not fixed. It is computed by first clustering pixels of the low-light illumination map into two clusters, bright pixels and dark pixels, by KMeans. Then taking the minimum illumination value of bright pixels as threshold $T$, $\eta$ is calculated by,
\begin{equation}
\eta=\log(1-0.8)/\log(1-T)
\end{equation}
which means to map the minimum illumination value of bright pixels to 0.8 under NISM.
\section{Experimental Results}
\subsection{Experiment Settings}
As a self-supervised framework, SID-NISM doesn't need any training dataset or prior information. For any given low-light image, SID-Net could split it into reflectance, illumination and noise components directly within hundred iterations, then NISM would enhance the decomposed illumination to combine with the reflectance. Besides, all the coefficients mentioned in Sect. \ref{sect:sid-nism} would remain the same during the experiments, which means it is not necessary to adjust the parameters for different inputs.
Evaluations are performed on real-scene images from four public datasets, DICM~\cite{lee2013contrast}, LIME~\cite{guo2016lime}, LOL~\cite{wei2018deep}, and MEF~\cite{ma2015perceptual}. And the proposed method SID-NISM is compared with 6 state-of-the-art low-light image enhancement algorithms, including 1) a weighted variational model for simultaneous reflectance and illumination estimation (SRIE)~\cite{fu2016weighted}, 2)illumination map estimation method (LIME)~\cite{guo2016lime}, 3) joint enhancement and denoising method via sequential decomposition (JED)~\cite{ren2018joint}, 4) perceptually bidirectional similarity based illumination estimation (PBS)~\cite{zhang2018high}, 5) deep learning based Retinex decomposition (Retinex-Net)~\cite{wei2018deep}, and 6) unsupervised-learning based back-lit image restoration network (ExCNet)~\cite{zhang2019zero}.
In the following subsections, both subjective and objective experimental results will be exhibited. More experiments including the intermediate decomposition results and the analysis of the effect on the object detection task can be found in the supplementary materials.
\subsection{Image Decomposition Results}
Since Retinex-based low-light image enhancement methods consider the problem as an image decomposition problem, it is necessary to show our intermediate decomposition results firstly. Fig.~\ref{fig:decomposition1} exhibits the decomposition results of two low-light images, \textit{Totoro} and \textit{toy} in LOL dataset. Their final enhanced results can be found in the main paper. In general, the proposed SID-Net can decompose reasonable reflectance, illumination and noise maps for given input images in a self-supervised way successfully. The decomposed reflectance maps contain sufficient color and structure details of the captured scene. The illumination maps reflect the lighting conditions as observed in human eyes. While the noise maps extract underlying noise component in dark regions.
\begin{figure*}
\centering
\begin{minipage}[b]{0.2\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22.png}}
\end{minipage}
\begin{minipage}[b]{0.2\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-R.png}}
\end{minipage}
\begin{minipage}[b]{0.2\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-L.png}}
\end{minipage}
\begin{minipage}[b]{0.2\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-N.png}}
\end{minipage}\\
\begin{minipage}[b]{0.2\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493.png}}
\centerline{(a) Source $S$}\medskip
\end{minipage}
\begin{minipage}[b]{0.2\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-R.png}}
\centerline{(b) Reflectance $R$}\medskip
\end{minipage}
\begin{minipage}[b]{0.2\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-L.png}}
\centerline{(c) Illumination $L$}\medskip
\end{minipage}
\begin{minipage}[b]{0.2\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-N.png}}
\centerline{(d) Noise $N$}\medskip
\end{minipage}
\caption{Intermediate decomposition results of two low-light images, \textit{Totoro} and \textit{toy} in LOL dataset.}
\label{fig:decomposition1}
\end{figure*}
Fig.~\ref{fig:decomposition2} compares the decomposition results of the proposed image decomposition network SID-Net with three state-of-the-art Retinex-based methods, LIME, JED and Retinex-Net, on \textit{Bookshelf} in LOL dataset and \textit{Balloon} in MEF dataset.
It can be seen that all of the four methods can decompose reasonable reflectance and illumination maps for given input images as expected. However, there also exists some differences among the four methods. In the aspect of reflectance, it is obvious that LIME can restore more color and structure details of the original objects compared with the other methods, because the reflectance of LIME is calculated by conducting element-wise division on the source image and the estimated illumination, which preserve the Retinex theory to a large extent. As for the illumination maps, SID-Net and Retinex-Net outperform other methods due to the specific structure-awareness illumination smoothness loss term, which makes it more consistent with the image structure. Besides, both as image decomposition networks, the biggest difference between SID-Net and Retinex-Net is that the proposed SID-Net is an unsupervised learning network without any prior training process, while the other one is a supervised learning network based on large-scale training dataset. As for the practical image decomposition performance, the two networks are evenly matched. Even in the reflectance maps, the results of SID-Net are clearer with more details than that of Retinex-Net.
In general, as shown in Fig.~\ref{fig:decomposition2}, SID-Net can extract underlying consistent reflectance from the given input image directly, and portray the lightness and shadow of the image in the illumination map. Although without prior training, the SID-Net can achieve the same or even better performance at the image decomposition stage.
\begin{figure*}[t]
\centering
\includegraphics[scale=0.4]{decom-eps-converted-to.pdf}
\caption{The decomposition results of LIME, JED, Retinex-Net, and SID-Net for \textit{Bookshelf} in LOL dataset and \textit{Balloon} in MEF dataset. For each of the example, the first row are the decomposed reflectance, while the second are the illumination maps}
\label{fig:decomposition2}
\end{figure*}
\subsection{Human Perception User Study}\label{sect:userstudy}
User study was conducted to perform comparisons between state-of-the-art methods in the dimensions of the overall image quality and some unpleasing artifacts through human perceptions. To evaluate the overall image quality, the subjects gave their personal feelings on the scale of 0 to 10, where 0 represents ``poor'', 5 means ``good'', and 10 is ``excellent''. As for the artifacts, there are five artifacts to be selected from (multiple-choice). Those are black edges, blur, overexposure, gray shadow and halo as shown in Fig.~\ref{fig:artifacts}. Once the subjects observe phenomenon in the images consisting with the description of the artifacts, the corresponding option would be checked. For all the enhanced images obtained by different methods, the subjects move the slider to give the image quality score and select specific artifacts according to their observations. The above procedure was completed on the Amazon Mechanical Turk platform by thirty workers.
\begin{figure}
\centering
\begin{minipage}[b]{0.29\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{blackedges.JPG}}
\centerline{(a) Black Edges}\medskip
\end{minipage}
\begin{minipage}[b]{0.29\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{blur.JPG}}
\centerline{(b) Blur}\medskip
\end{minipage}\\
\begin{minipage}[b]{0.29\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{overexposure.jpg}}
\centerline{(c) Overexposure}\medskip
\end{minipage}
\begin{minipage}[b]{0.29\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{grayshadow.png}}
\centerline{(d) Gray Shadow}\medskip
\end{minipage}
\begin{minipage}[b]{0.29\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{halo.jpg}}
\centerline{(e) Halo}\medskip
\end{minipage}
\caption{Five artifacts examples from (a) Retinex-Net \cite{wei2018deep} on \textit{hill} in DICM dataset; (b) JED \cite{ren2018joint} on \textit{tree} in DICM dataset; (c) LIME \cite{guo2016lime} on \textit{stormtrooper} in LIME dataset; (d) and (e) ExCNet \cite{zhang2019zero} on \textit{roof} in MEF dataset.}
\label{fig:artifacts}
\end{figure}
The results of the user study are summarized in graphs shown in Fig.~\ref{fig:qualityUser} and \ref{fig:artifactsUser}. In Fig.~\ref{fig:qualityUser}, a boxplot is used to exhibit the distributions of the image quality scores for different methods, in which the red solid line represents the median while the blue dotted line represents the mean score. Through observing the median, mean, and the first quartile, it is obvious that the participants showed a strong bias in preference towards our results (median: 6.0, mean: 5.51, quartile: 8.0) when compared to Retinex-Net~\cite{wei2018deep} (median: 3.0, mean: 3.05, quartile: 4.0), gave higher scores when compared to SRIE~\cite{fu2016weighted} (median: 5.0, mean: 5.04, quartile: 7.0), JED~\cite{ren2018joint} (median: 5.0, mean: 5.34, quartile: 7.0), PBS~\cite{zhang2018high} (median: 5.0, mean: 5.31, quartile: 7.0) and ExCNet~\cite{zhang2019zero} (median: 5.0, mean: 5.41, quartile: 7.0), and had no preference when compared with LIME~\cite{guo2016lime} (median: 6.0, mean: 5.67, quartile: 8.0). The results in Fig.~\ref{fig:qualityUser} indicate that SID-NISM outperforms most of the state-of-the-art methods in terms of the overall image quality.
\begin{figure}
\centering
\includegraphics[scale=0.32]{quality-eps-converted-to.pdf}
\caption{The image quality scores for different methods.}
\label{fig:qualityUser}
\end{figure}
In Fig.~\ref{fig:artifactsUser}, a stacked bar graph is adopted to illustrate the average counts of artifacts in the results of different methods, in which each color bar corresponds to the five artifacts in Fig.~\ref{fig:artifacts} respectively. Specifically, the average counts are computed by $M/(N*P)$, where $M$ is the total counts of artifacts in the results of each method, $N$ is the number of images, $P$ is the number of subjects. It can be seen that SID-NISM tends to introduce less artifacts (average counts: 0.6) than other methods (average counts: above 0.8) in total. Generally, SID-NISM doesn't lead too many unexpected artifacts. SRIE~\cite{fu2016weighted} would introduce gray shadow into their results, while the results of LIME~\cite{guo2016lime} may be over-enhanced to become overexposure. JED~\cite{ren2018joint} has more blurry cases, while Retinex-Net~\cite{wei2018deep} contains the most black edges in their results. These phenomenons can be also observed in the visual examples of the next subsection.
\begin{figure}
\centering
\includegraphics[scale=0.32]{artifacts-eps-converted-to.pdf}
\caption{The average counts of artifacts in different methods.}
\label{fig:artifactsUser}
\end{figure}
\subsection{Visual Quality Comparisons}\label{sect:visualQuality}
In order to facilitate readers to visually compare the results of different low-light image enhancement approaches, several examples are exhibited in Fig.~\ref{fig:vision1} and ~\ref{fig:vision2}.
\begin{figure*}
\centering
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-input.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-SRIE.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-LIME.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-JED.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-SID-new.png}}
\end{minipage}\\
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-input-local.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-SRIE-local.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-LIME-local.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-JED-local.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{22-SID-new-local.png}}
\end{minipage}\\
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-input.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-SRIE.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-LIME.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-JED.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-SID-new.png}}
\end{minipage}\\
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-input-local.png}}
\centerline{(a) Input}\medskip
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-SRIE-local.png}}
\centerline{(b) SRIE}\medskip
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-LIME-local.png}}
\centerline{(c) LIME}\medskip
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-JED-local.png}}
\centerline{(d) JED}\medskip
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{6079-SID-new-local.png}}
\centerline{(e) SID-NISM}\medskip
\end{minipage}
\caption{Compared with the results obtained by SRIE \cite{fu2016weighted}, LIME \cite{guo2016lime} and JED \cite{ren2018joint} on two natural images: \textit{Totoro} from LOL dataset and \textit{candle} from MEF dataset. Below are the enlarged detailed images in the red box. (Best viewed on screen)}
\label{fig:vision1}
\end{figure*}
\begin{figure*}
\centering
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-input.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-PBS.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-Retinex.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-ExCNet.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-SID-new.png}}
\end{minipage}\\
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-input-local.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-PBS-local.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-Retinex-local.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-ExCNet-local.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{1-SID-new-local.png}}
\end{minipage}\\
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-input.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-PBS.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-Retinex.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-ExCNet.png}}
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-SID-new.png}}
\end{minipage}\\
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-input-local.png}}
\centerline{(a) Input}\medskip
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-PBS-local.png}}
\centerline{(b) PBS}\medskip
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-Retinex-local.png}}
\centerline{(c) Retinex-Net}\medskip
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-ExCNet-local.png}}
\centerline{(d) ExCNet}\medskip
\end{minipage}
\begin{minipage}[b]{0.16\linewidth}
\centering
\centerline{\includegraphics[width=1\linewidth]{493-SID-new-local.png}}
\centerline{(e) SID-NISM}\medskip
\end{minipage}
\caption{Compared with the results obtained by PBS \cite{zhang2018high}, Retinex-Net \cite{wei2018deep} and ExCNet \cite{zhang2019zero} on two natural images: \textit{street} from LIME dataset and \textit{toy} from LOL dataset. Below are the enlarged detailed images in the red box. (Best viewed on screen)}
\label{fig:vision2}
\end{figure*}
In general, it is clear that SID-NISM could brighten up the dark regions of low-light images successfully. More importantly, it can also restore the strong contrast of scene content sufficiently. For example, taking \textit{candle} in Fig.~\ref{fig:vision1} as input, the contrast between the candle in the foreground and the glass cup in the background is clearer and stronger in the result of SID-NISM compared with other methods. Another example is the \textit{toy} in Fig.~\ref{fig:vision2}. It is obvious that the contrast between the toy's face and the background is stronger in our result. This property renders the results of our method seem more active and attractive, which is consistent with the user study about the image overall perceptions.
In addition, as discussed in Sect.~\ref{sect:userstudy}, SID-NISM could restore the overall quality of low-light images on the premise of introducing less unexpected artifacts, including preserving clear boundary (glass cup in \textit{candle}), avoiding gray shadow (the inside of the door in \textit{street}) and halo (boundary of the door in \textit{street}). Especially, our results contain less noise such as the sky in \textit{street} and the white background in \textit{toy}. Combining with the above intuitive observation and data statistics in the user study, it can be seen that our method is more robust with less artifacts, which makes our framework more competitive and practical in both people's daily use and commercial cases.
\subsection{Objective Evaluation Indexes}
\begin{table*}[]
\centering
\begin{tabular}{c|ccccccc}
\toprule
& GE & CE & GMI & GMG & NIQE & PSNR & SSIM \\
\hline
SRIE & 6.5584 & 18.4951 & 56.0559 & \textcolor{red}{\textbf{6.2808}} & 3.8777 & 12.8554 & 0.5298 \\
LIME & 7.2652 & 19.8994 & 93.2529 & 11.4664 & 4.2036 & \textcolor{blue}{\textbf{17.1717}} & 0.6355 \\
JED & 6.5296 & 19.0923 & 73.9586 & 4.6584 & \textcolor{red}{\textbf{3.3546}} & 14.9944 & 0.6325 \\
PBS & 7.0828 & 20.2996 & 84.7626 & 10.8144 & 4.4976 & 16.3765 & \textcolor{blue}{\textbf{0.6375}} \\
Retinex-Net & 6.9750 & \textcolor{red}{\textbf{21.0380}} & \textcolor{red}{\textbf{110.2008}} & 12.8826 & 5.7127 & 14.4061 & 0.4932 \\
ExCNet & 7.0548 & 18.7087 & 87.4292 & 10.2022 & 4.4485 & 14.9694 & 0.601 \\
SID-NISM w/o $\mathcal{L}_{R}$ and $\mathcal{L}_{N}$ & \textcolor{red}{\textbf{7.1978}} & 18.8489 & 99.9323 & 11.1822 & 4.4493 & 17.0313 & 0.5926 \\
SID-NISM & \textcolor{blue}{\textbf{7.2499}} & \textcolor{blue}{\textbf{20.7416}} & \textcolor{blue}{\textbf{103.3472}} & \textcolor{blue}{\textbf{9.9731}} & \textcolor{blue}{\textbf{3.8579}} & \textcolor{red}{\textbf{17.7576}} & \textcolor{red}{\textbf{0.6633}} \\
\hline
Reference & 7.1949 & 21.0804 & 110.8416 & 7.1368 & 3.5593 & / & 1 \\
\toprule
\end{tabular}
\caption{Quantitative comparison on LOL and MEF datasets in terms of different metrics. GE, CE, GMI, GMG should be close to the reference images. Lower NIQE, higher PSNR and SSIM indicate better image quality. Top2 results are highlighted in bold (\textcolor{red}{Top1}-\textcolor{blue}{Top2}).}
\label{tab:objEva}
\end{table*}
\begin{table*}[t]
\centering
\begin{tabular}{c|p{1.0cm}<{\centering}p{0.8cm}<{\centering}p{0.8cm}<{\centering}p{0.8cm}<{\centering}p{0.8cm}<{\centering}p{0.8cm}<{\centering}p{0.8cm}<{\centering}p{0.8cm}<{\centering}p{0.8cm}<{\centering}p{1.0cm}<{\centering}|c}
\toprule
& bicycle & boat & bottle & bus & car & cat & chair & dog & motor & person & mAP \\
\hline
Original & 42.36 & 8.34 & 21.38 & 42.31 & 30.74 & 20.44 & 34.77 & 25.53 & 33.12 & 36.66 & 31.54\\
SRIE & 44.75 & 10.76 & 27.26 & 54.01 & 30.80 & 21.28 & 36.96 & 26.27 & 34.02 & 38.33 & 33.92 \\
LIME & 47.56 & 9.25 & 24.26 & 53.24 & 32.72 & 20.40 & 32.25 & 28.29 & 32.50 & 39.98 & \textbf{33.98} \\
JED & 41.50 & 8.53 & 21.42 & 46.53 & 23.72 & 20.46 & 31.36 & 30.24 & 31.77 & 33.83 & 29.28 \\
PBS & 47.33 & 8.31 & 23.85 & 50.43 & 33.12 & 17.65 & 33.44 & 26.81 & 33.19 & 38.29 & 33.20 \\
Retinex-Net & 34.59 & 5.70 & 17.80 & 38.11 & 20.99 & 16.18 & 27.57 & 6.14 & 15.53 & 27.90 & 23.15 \\
ExCNet & 39.24 & 8.86 & 19.46 & 55.37 & 32.99 & 20.68 & 32.00 & 30.13 & 32.49 & 39.04 & 32.84 \\
SID-NISM & 47.64 & 9.69 & 18.26 & 51.36 & 32.02 & 17.19 & 36.48 & 28.53 & 32.75 & 39.58 & \textbf{33.39} \\
\toprule
\end{tabular}
\caption{The performance (mAP: \%) of object detection on real low-light images and their enhanced versions by low-light image enhancement methods.}
\label{tab:objectmap}
\end{table*}
In addition to the above subjective experiments, there are many indexes with or without reference images that can be used to evaluate the quality of the enhanced images objectively. In this section, we take use of gray entropy (GE), color entropy (CE, sum of the entropy of RGB channels), gray mean illumination (GMI), gray mean gradient (GMG), NIQE~\cite{mittal2012making}, PSNR, and SSIM~\cite{wang2004image} to compare different low-light image enhancement methods. Generally, the first four indexes could reflect the gap between the enhanced results and the reference high-light images from four aspects including the information in gray and color channels, the illumination degree and the image gradient magnitude. NIQE assesses the overall naturalness of the enhanced results and a lower value roughly corresponds to a higher overall naturalness. PSNR could verify the performance of denoising algorithms and a higher value usually indicates a better image quality. SSIM measures the structure similarity between the enhanced results and the reference images, which is in the range of [0, 1]. It should be noted that these indexes can only reflect the image quality in specific aspects, which may be not completely consistent with the evaluation results given by the human visual system.
The quantitative results over LOL and MEF datasets (the two datasets with reference normal-light images) are reported in Table \ref{tab:objEva}. Obviously, the proposed SID-NISM could achieve better performance compared with existing state-of-the-art methods in terms of the image information, the overall quality and the noise condition. It is worth mentioning that we also compare the performance of our framework with and without $\mathcal{L}_{R}$ and $\mathcal{L}_{N}$ to quantitatively evaluate the merit brought by the two novel reflectance and noise terms in the loss function of SID-Net. It can be seen that they not only help improve the image color and gradient information of the enhanced results greatly, but also decrease the noise influence and promote the overall image quality.
\section{Object Detection Test}
We also take a step further to look into the potential of low-light image enhancement in improving the performance of one popular computer vision task, object detection. To do so, tests were performed on YOLOv2 platform on real low-light images and their enhanced versions by our proposed framework and other state-of-the-art methods. Here we directly used the pre-trained VOC2007+2012 model without any re-training or fine-tuning as we would like to observe the outcome of the originally optimized model when given low-light and enhanced images. Experiments were conducted on Exclusively Dark Image Dataset (ExDark), which is consist of low-light images taken in 10 different low-light environments with corresponding 12 image classes and object level annotations.
Table \ref{tab:objectmap} exhibits the mean of Average Precision (mAP) over 10 classes when detecting different versions of images. Obviously, compared to detecting objects on the original low-light images directly, the performance of the object detection model would be improved by preprocessing the inputs through our proposed low-light image enhancement method from 31.54\% to 33.39\% in terms of mAP. Among existing methods, LIME reaches the best mAP (33.98\%) while SID-NISM shows its competitive power (within 1\%) in improving the performance of downstream tasks. This test illustrates the feasibility of incorporating image enhancement as a support for practical applications.
\section{Conclusion}
In this paper, a self-supervised low-light image enhancement method called SID-NISM is proposed, which can first decompose given input images in an unsupervised way without relying on any external examples or prior training, and then enhance decomposed illumination map by a well-designed nonlinear illumination mapping function. The proposed scheme is concise yet powerful. Experimental results show that, even though our method does not need any support data or prior knowledge, it can produce visually pleasing enhancement results with less unexpected artifacts as well as a good representation of image decomposition.
{\small
\bibliographystyle{ieee_fullname}
| 61df11a09c4aa62c3f98b3e4145b9bb7b0ca3eed | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Machine learning models are shown to be susceptible to various types of adversarial attacks targeted to degrade the performance of machine learning models. Research in adversarial machine learning has mostly focused on targeting accuracy as a measure \cite{chakraborty2018adversarial,li2018security}. However, with the advancement of research in fair machine learning, attention has geared toward fairness measures and their importance. A broad area of research has arisen by proposing different fairness measures and definitions \cite{dwork2012fairness,hardt2016equality,NIPS2017_6995,10.1145/3194770.3194776,mehrabi2019survey}. It is important to have systems that are away from discriminatory behavior as these systems can be used in sensitive environments, such as courts to make bail decisions \cite{Dresseleaao5580}; thus, crucial to have measures and standards to ensure that artificial intelligence (AI) systems are fair.
Similar to accuracy, fairness measures are important in machine learning research and can be targeted by adversaries for malicious intent. For instance, adversaries can attack models used in government and other sensitive agencies to make them appear unfair in order to depreciate their values and credibility. Some adversaries can even gain profit with such attacks off of biasing decisions for their benefit, in credit or loan applications for instance \cite{chen2019fairness}; therefore, thinking about ways these attacks could happen for designing more robust systems in the future is crucial. The goal is to have systems that are robust to fairness attacks; thus, it is important to consider attacks targeted toward fairness.
\textbf{Our contributions.} In this work, we propose data poisoning attacks that are able to target fairness. We propose two families of poisoning attacks. Those that target fairness only without having noticeable effects on accuracy to make them more subtle and undetectable, and those that affect both fairness and accuracy by injecting poisoned points during train time. Some adversaries may want to harm systems with regards to fairness and accuracy at the same time, while others might only consider fairness. In light of this, we introduce anchoring attack and influence attack on fairness techniques. In anchoring attack, we target fairness only without harming accuracy as much, while in influence attack on fairness, we target both accuracy and fairness measures by incorporating a loss function maximizing and attacking which would degrade fairness and accuracy.
Through experimentation on three different datasets with different fairness measures and definitions, we show the effectiveness of our attacks in achieving the desired goal of either affecting fairness only without significant harm on accuracy or accuracy and fairness at the same time. In addition, we incorporate different baseline models to evaluate different aspects of our attacks. We demonstrate that original data poisoning attacks designed to attack accuracy are not suitable for fairness attacks; thus, highlighting the importance of attacks designed for fairness specifically. We also compare our methods against concurrent work on adversarial attacks on fairness and show the effectiveness of our methods over them.
\section{Conclusion and Future Work}
In this work, we introduced two families of poisoning attacks that can target fairness. We showed the effectiveness of these attacks through experimentation on different real world datasets with different measures. Our influence attack on fairness (IAF) used the attack strategy as in influence attack \cite{koh2018stronger,koh2017understanding}. As an extension, we modified the loss function so that it can harm fairness as well as accuracy. Furthermore, we explore an attack strategy called the ``anchoring attack'' that harms fairness by placing poisoned points near target points in order to bias the outcome. Our paper also introduced two ways of sampling these target points. A direct extension of this approach is to explore other methods of sampling points to increase the effectiveness of this attack. The introduced attacks each have their own advantages and disadvantages. The goal was to design attacks that can complement each other. For instance, influence attack on fairness which is gradient-based can be slow. Anchoring attack, however, does not use gradients and is considerably faster. Further, while influence attack on fairness targets fairness harshly, anchoring attack is more subtle. And if anchoring attack can not explicitly control for accuracy fairness trade-off, influence attack on fairness can control this trade-off.
This work points out several important angles for future research. Some important extensions are as follows: what other ways can machine learning systems be harmed by data poisoning attacks? How can we design and adapt defenses that can be effective against malicious attacks targeting fairness? Another question worth pursuing is from the perspective of the defender. Can current defenses against accuracy attacks be useful against the types of attacks that target fairness. If not, how do we adapt defenders so that they can prevent fairness attacks? These questions can help us design more fair, and accurate models that are robust to poisoning attacks. By extension, one can also think about stronger attacks against fairness than ours. We anticipate that continuing to blend the fields of adversarial and fair machine learning can create interdisciplinary ideas that can help us develop more robust and fair machine learning models.
\section{Acknowledgments}
This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR0011890019. We thank the anonymous reviewers and Mozhdeh Gheini for their constructive feedback.
\section{Evaluation}
In our experiments, we evaluate our attacks with regards to different measures, such as accuracy and foundational fairness measures: statistical parity, and equality of opportunity differences.
We also utilize three real world datasets in our experiments, introduced below. We compare against a suite of baselines that test our attacks' performance with regards to accuracy and fairness. Our results indicate that our attacks, the anchoring attack and influence attack on fairness, are effective in terms of affecting fairness aspects of the model.
\subsection{Datasets}
We use three different real world datasets in our experiments with gender as the sensitive attribute. The data was split into an 80-20 train and test split. \\
\textbf{German Credit Dataset.} This dataset comes from UCI machine learning repository \cite{Dua:2019}. It contains the credit profile about individuals with 20 attributes associated to each data person. In our experiments, we utilized all the 20 attributes from this dataset. The classification goal is to predict whether an individual has good or bad credit. \\
\textbf{COMPAS Dataset.} Propublica's COMPAS dataset contains information about defendants from Broward County \cite{larson2016compas}. We utilized the features in Table \ref{compas_features} as our prediction features. The classification goal is to predict whether an individual will recommit a crime within two years. \\
\textbf{Drug Consumption Dataset.} This dataset comes from the UCI machine learning repository \cite{Dua:2019}. It contains information about individuals \cite{fehrman2017five}. We utilized the features listed in Table \ref{compas_features} as our prediction features. The classification goal is to predict whether an individual has consumed cocaine or not in their lifetime.
\begin{table}[h]
\centering
\begin{tabular}{ p{2.5cm} p{2.5cm}}
\toprule
COMPAS&\\
\midrule
sex&age\_cat \\
juv\_fel\_count&juv\_misd\_count\\
priors\_count&c\_charge\_degree\\
race & juv\_other\_count\\[0.5pt]
\bottomrule
\end{tabular}
\begin{tabular}{ p{1.7cm} p{1.7cm} p{1.7cm} p{0.5cm}}
\toprule
Drug&&\\
\midrule
ID&Age& Gender & SS\\
Education&Country&Ethnicity&\\
Nscore&Escore&Oscore&\\
Ascore&Cscore &Impulsive& \\ [0.5pt]
\bottomrule
\end{tabular}
\caption{Features used from the COMPAS and Drug Consumption datasets.}
\label{compas_features}
\end{table}
\subsection{Measures}
In addition to accuracy, we have utilized two well-known fairness measures to analyze the performance of different attacks with regard to fairness, detailed below. \\
\textbf{Statistical Parity Difference}
Statistical parity is a well-known measure (definition) introduced in \cite{dwork2012fairness}. We utilize this measure as one of our metrics for fairness. It captures the predictive outcome differences between different (advantaged and disadvantaged) demographic groups. The measure is defined below and is referred to as statistical parity throughout our paper.
\[ SPD = |p(\hat{Y}=+1|x \in \mathcal{D}_a)-p(\hat{Y}=+1|x \in \mathcal{D}_d)| \]
\textbf{Equality of Opportunity Difference}
Equality of opportunity is another well-known fairness definition introduced in \cite{hardt2016equality}. We utilized the equality of opportunity difference as another fairness metric. It captures differences in the true positive rate between different (advantaged and disadvantaged) demographic groups. The measure is defined below and is addressed as equality of opportunity throughout this paper.
\begin{align*}
EOD = |p(\hat{Y}=+1|x \in \mathcal{D}_a, Y=+1) \\ -p(\hat{Y}=+1|x \in \mathcal{D}_d, Y=+1)|
\end{align*}
\subsection{Methods}
\begin{figure*}[h]
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/German_acc.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/German_parity.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/German_equality.pdf}
\includegraphics[width=0.33\textwidth,,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/compas_acc.pdf}
\includegraphics[width=0.33\textwidth,,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/compas_parity.pdf}
\includegraphics[width=0.33\textwidth,,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/compas_equality.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/drug_acc.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/drug_parity.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/drug_equality.pdf}
\caption{Results obtained for different attacks with regards to different fairness (SPD and EOD) and accuracy (test error) measures on three different datasets (German Credit, COMPAS, and Drug Consumption) with different $\epsilon$ values.}
\label{attack_results}
\end{figure*}
\begin{figure*}[h]
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/German_lambda_acc.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/German_lambda_parity.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/German_lambda_equality.pdf}
\includegraphics[width=0.33\textwidth,,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/compas_lambda_acc.pdf}
\includegraphics[width=0.33\textwidth,,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/compas_lambda_parity.pdf}
\includegraphics[width=0.33\textwidth,,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/compas_lambda_equality.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/drug_lambda_acc.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/drug_lambda_parity.pdf}
\includegraphics[width=0.33\textwidth,trim=0.5cm 0cm 2cm 0cm,clip=true]{images/drug_lambda_equality.pdf}
\caption{Results obtained for different lambda values for the IAF attack with regards to different fairness (SPD and EOD) and accuracy (test error) measures on three different datasets (German Credit, COMPAS, and Drug Consumption) with different $\epsilon$.}
\label{lambda_attack_results}
\end{figure*}
To evaluate our attacks, we compared them against an attack that does not consider fairness and only considers accuracy to show that such attacks are not necessarily effective for fairness, motivating the need for fairness attacks. Also, we compared our attacks in terms of how they attack accuracy as a measure versus attacks that are specifically designed to target accuracy. We also compared our attacks to an attack that is optimized for fairness.
The evaluated methods are listed below. In our experiments, the poisoned points are inversely proportional to class balance as also suggested in \cite{koh2018stronger}, so we made $(|\mathcal{D}_{c}^+|\epsilon)$ copies from the negative poison instance ($\mathcal{I}_{-}$) and $(|\mathcal{D}_{c}^-|\epsilon)$ copies from the positive poison instance ($\mathcal{I}_{+}$) in which $|\mathcal{D}_{c}^+|$ and $|\mathcal{D}_{c}^-|$ denote the number of positive and negative points in the clean data respectively. Hinge loss was used to control for accuracy for all the methods in our experiments as in \cite{koh2018stronger}. \\
\textbf{Influence Attack on Fairness (IAF)} In this paper, our influence attack on fairness is where the attack tries to maximize the covariance between the signed distance of feature vectors from the decision boundary to the sensitive features, which would then cause the attack to target and degrade fairness. In our experiments we set $\lambda =1$.
\textbf{Random Anchoring Attack (RAA)} The anchoring attack where a target point is picked at random. In this new set of attacks, the goal is to place poisoned points in the vicinity of the target points in which the poisoned and target points have the same demographic group but different labels. In our experiments we set $\tau=0$ indicating the closest vicinity.
\textbf{Non-random Anchoring Attack (NRAA)} This attack builds upon the random anchoring attack; however, in this attack, the target point is not chosen randomly. In this attack, the point with the most neighbors similar to it (with the same demographic group and label) is chosen as the target point so that we can infect as many similar points to the target point as possible. This can be effective because we are infecting more targeted points; however, in some cases it might be less effective since more poisoned points may be needed in order to achieve the goal of infecting many points and shifting the decision boundary. In our experiments we set $\tau=0$.
\textbf{Influence Attack (Koh et al.)} This is a type of attack that is targeted only toward affecting accuracy \cite{koh2018stronger,koh2017understanding}. The reason we include this type of attack along with attacks targeted toward fairness is that it can help us understand how attacks targeting only accuracy affect fairness measures. Attacks of this nature can also serve as a good comparison because they show the effect of attacks on accuracy; because this attack is specifically designed to target accuracy, it can be a strong method to compare against.
\textbf{Poisoning Attack Against Algorithmic Fairness (Solans et al.)} In \cite{solans2020poisoning}, the authors propose a loss function that claims to target fairness measures. We utilized the loss introduced in this paper as depicted below in equation \eqref{baseline_loss} in the influence attack from \cite{koh2018stronger,koh2017understanding} and compared it to our proposed attacks. The goal of \cite{solans2020poisoning} was to incorporate the loss in \eqref{baseline_loss} into an attack strategy that would maximize the loss; thus, we incorporated this loss into the influence attack \cite{koh2018stronger,koh2017understanding}, which we found to be a strong attack strategy in maximizing the loss and also the same attack strategy used in our influence attack on fairness. In our experiments, we utilized the same $\lambda$ value as proposed in \cite{solans2020poisoning} to balance the class priors.
\begin{align*}
L_{adv}(\hat{\theta};\mathcal{D}_{test}) = &\underbrace{\sum_{k=1}^p \ell(\hat{\theta};x_k,y_k)}_\text{disadvantaged} + \lambda \underbrace{\sum_{j=1}^m \ell(\hat{\theta};x_j,y_j)}_\text{advantaged} \\
& where \;\; \lambda = \frac{p}{m}.
\numberthis \label{baseline_loss}
\end{align*}
\subsection{Results}
The results in Figure \ref{attack_results} demonstrate that the influence attack (Koh et al.), although performing remarkably well in attacking accuracy, does not attack fairness well. The results also confirm that our influence attack on fairness method outperforms (Solans et al.) \cite{solans2020poisoning} in affecting fairness measures, and anchoring attack outperforms (Solans et al.) \cite{solans2020poisoning} in affecting fairness measures in most of the cases. One can observe that influence attack on fairness is the most effective amongst all the attacks in attacking fairness measures.
Due to the nature of our influence attack on fairness loss function and its controlling parameter on accuracy and fairness, it can be utilized in scenarios where the adversary wants to maliciously harm the system in terms of accuracy, or fairness, or both. On the other hand, anchoring attacks can be utilized in places where the adversary wants to subtly harm accuracy with an effective harm on fairness. These types of attacks can be used by, e.g., adversaries who would want to gain profit off of biasing decisions for their benefit; thus, to remain less detectable they do not harm accuracy. Although it is possible that anchoring attack can harm accuracy to a higher degree, as shown empirically in our results, it is less likely that anchoring attack is able to degrade accuracy by a large amount in practice for real world datasets.
In addition, in Figure \ref{lambda_attack_results} we demonstrate the effect of our regularized loss in the influence attack on fairness. The results show that with the increase of lambda the attack affects fairness measures more as expected from the loss; however, for the lower lambda values the attack acts similar to the original influence attack targeted towards accuracy. The results also show that higher epsilon values highlight the behavior of the loss more as expected such that for high epsilon value of 1 the changes are more significant with modifications to the lambda value in the loss function, while less subtle for lower epsilon values such as 0.1.
\section{Related Work}
Here, we cover related work from both fair machine learning as well as adversarial machine learning research.
\subsection{Adversarial Machine Learning}
Research in adversarial machine learning is mostly focused on designing defenses and attacks against machine learning models \cite{NIPS2017_6943,chakraborty2018adversarial,li2018security}. Ultimately, the goal is for machine learning models to be robust toward malicious activities designed by adversaries. Thus, it is important to consider both sides of the spectrum in terms of designing the attacks and defenses that can overcome the attacks. In adversarial machine learning, different types of attacks, such as data poisoning and evasion attacks, exist. In evasion attacks, the goal is to come up with adversarial examples that are imperceptible to human eye but can deceive benign machine learning models during test time \cite{biggio2013evasion,moosavi2016deepfool,DBLP:journals/corr/GoodfellowSS14}. On the other hand, in data poisoning attacks, the goal is to manipulate the training data--via adding, removing, or changing instances--so that the learned model is malicious \cite{10.5555/3042573.3042761,shafahi2018poison}. Different algorithms and approaches have been proposed for poisoning attacks focusing on accuracy as the performance measure \cite{10.5555/3042573.3042761,shafahi2018poison}. In this paper, we also focused on data poisoning attack while considering fairness as a performance measure in addition to accuracy.
\subsection{Fair Machine Learning}
Research in fair machine learning has gained attention recently, with many active research areas. For instance, some work introduces new definitions and measures for fairness \cite{dwork2012fairness,hardt2016equality,NIPS2017_6995,mehrabi2020statistical}. \cite{10.1145/3194770.3194776} has a complete list of the definitions on fairness. Other work utilizes these definitions and tries to design and learn fair classification \cite{zafar2015learning,pmlr-v97-ustun19a}, regression \cite{agarwal2019fair}, and representations \cite{moyer2018invariant}. The battle to mitigate unfairness can happen in different phases. Some target making the data more fair \cite{zhang2017achieving}, while others target the algorithms \cite{zafar2015learning}. These mitigation techniques can also vary in when and how they are applied. For instance, some approaches are pre-processing techniques \cite{Kamiran2012} in which the focus is to remove discrimination from the data before the learning phase. Others try to impose fairness during training via incorporation of fair loss functions or other approaches during the training phase, known as in-processing \cite{kamishima2012fairness}, while some are post-processing approaches \cite{NIPS2017_7151} in which the model is treated as a black box system and discrimination removal is performed on the output of the model. \citet{mehrabi2019survey} performs a literature review of fair machine learning research in different subject domains, which can be referenced for more detail. In our work, we utilize some of the definitions and measures widely used in fair machine learning research \cite{dwork2012fairness,hardt2016equality} in measuring the performance of our attacks with regard to fairness. We were also inspired by some loss functions introduced in fair classification tasks in one of our attacks \cite{zafar2015learning}.
\subsection{Adversarial Fair Machine Learning}
The rapid and significant growth of research in algorithmic fairness highlights the importance of machine learning models being fair and robust toward any unfair behavior. To this end, it is important to think about attacks that can make models unfair in order to strengthen models against such attacks. This recent and interesting line of work combines the two fields of fair and adversarial machine learning. The only work we are aware of that proposes poisoning attacks against algorithmic fairness is \cite{solans2020poisoning}. In \cite{solans2020poisoning}, the authors propose an attack that targets fairness. We compared this attack with our two newly proposed attacks using three real world datasets. Our anchoring attack does not rely on any loss function making it different in nature with the previous work. In our influence attack on fairness we introduce a new loss function different than the previous work which is more in line with fairness literature and work done in fairness domain making our attack more intuitive. In addition, our influence attack on fairness is able to control a fairness-accuracy trade-off with the hyper-parameter involved in its loss function which is also shown in Figure \ref{lambda_attack_results} as an additional experimental result. This line of work can bring researchers from both fields closer and inspire new and interesting research problems. Another interdisciplinary research field combining concepts from fairness and privacy includes the differential privacy line of work \cite{dwork2008differential,bagdasaryan2019differential,pmlr-v97-jagielski19a,pujol2020fair}.
\section{Ethics Statement}
This paper furthers ethics in the machine learning community in two major ways:
\begin{itemize}
\item
Despite extensive research in adversarial machine learning, not much attention has been given to scenarios where fairness is a possible target of deliberate attacks. We suggest that fairness metrics are as important as accuracy, because they can be manipulated in sensitive environments to achieve malicious goals. Our work points out potential vulnerabilities of machine learning models against fairness-targeting attacks. This line of research can raise awareness, and motivate researchers to introduce methods to mitigate harmful effects of adversarial attacks on fairness. The attacks proposed in this paper are meant to ensure the robustness of fairness in machine learning applications. Nevertheless, we acknowledge that in the wrong hands these type of tools could enable an attacker to harm fairness in extant machine learning systems.
\item Fairness and adversarial machine learning are both very important research areas with major safety, security, and ethics implications both within and beyond on the AI/machine learning community. This work combines ideas from both adversarial and fair machine learning, and will hopefully facilitate collaboration among researchers from both communities, eventually leading to more robust and fair machine learning models.
\end{itemize}
\section{Introduction}
With proliferation of machine learning (ML) applications in everyday life, it is imperative that ML algorithms underlying those applications do not discriminate, especially when it comes to potentially sensitive and consequential decisions, such as bail decisions~\cite{Dresseleaao5580}. Thus, recent research has looked into possible biases present in ML algorithms, and proposed different measures and definitions for characterizing fairness \cite{dwork2012fairness,hardt2016equality,NIPS2017_6995,10.1145/3194770.3194776,mehrabi2019survey}.
Despite this interest, not much is known about the robustness of various fairness measures with respect to random, or perhaps malicious, perturbations. Indeed, it is known that machine learning models can be susceptible to various types of adversarial attacks targeted to degrade the performance of machine learning models. However, research in adversarial machine learning has mostly focused on targeting accuracy \cite{chakraborty2018adversarial,li2018security}. We argue that, like accuracy, fairness measures can be targeted by malicious adversaries as well. For instance, adversaries can attack models used by a government agency with the goal of making them appear unfair in order to depreciate their value and credibility. Some adversaries can even profit from such attacks by biasing decisions for their benefit, e.g., in credit or loan applications. Thus, one should consider fairness when assessing the robustness of ML systems.
\textbf{Our contributions.} In this work, we propose data poisoning attacks that target fairness. We propose two families of poisoning attacks: {\em anchoring} and {\em influence}\footnote{https://github.com/Ninarehm/attack}. In anchoring attacks the goal is to place poisoned points to affect fairness without modifying the attacker loss. On the other hand, our influence attack on fairness can affect both fairness and accuracy by injecting poisoned points during train time via a specific adversarial loss that regularizes between fairness and accuracy losses. Some adversaries may want to harm systems with regard to fairness and accuracy at the same time, while others might only consider one that can be achieved by this regularization. In the anchoring attack, we place poisoned points to bias the decision boundary; in the influence attack, we target fairness measures by incorporating a loss function maximizing and attacking which can degrade fairness by maximizing the covariance between the decision outcome and sensitive attributes.
Through experimentation on three different datasets with different fairness measures and definitions, we show the effectiveness of our attacks in achieving the desired goal of affecting fairness. In addition, we incorporate different baseline models to evaluate different aspects of our attacks. We demonstrate that original data poisoning attacks designed to attack accuracy are not suitable for fairness attacks, thus highlighting the importance of attacks designed for fairness. We also compare our methods against concurrent work on adversarial attacks on fairness and show the effectiveness of our methods in comparison.
\section{Background on Poisoning Attacks}
Consider a supervised learning problem characterized by a loss function $\mathcal{L}(\theta;\mathcal{D})$ and an adversarial loss $L_{adv}(\hat\theta;\mathcal{D})$, where $\hat\theta$ is the set of learnable parameters and $\mathcal{D}$ is a labeled dataset. Let $\mathcal{D}_{train}$ be the training dataset. We assume that the adversary can poison a fraction of those data points, so that $\mathcal{D}_{train}=\mathcal{D}_{c} \cup \mathcal{D}_{p}$, where $\mathcal{D}_{c}$ and $\mathcal{D}_{p}$ are the set of clean and poisoned data points, respectively. We assume that $|\mathcal{D}_{p}|= \epsilon |\mathcal{D}_{c}|$. Furthermore, $\mathcal{D}_{p} \subseteq \mathcal{F}_{\beta}$ where $\mathcal{F}_{\beta}$ is the feasible set, which is a set selected by a defense mechanism based on anomaly detection techniques, containing elements that the defender considers as sanitized data to train its model. The existence of the feasible set in the objective helps the poisoned points to blend with the natural data and make it more difficult for anomaly detector techniques to detect them \cite{koh2018stronger}.
A data poisoning attack can be written as the following optimization problem (over the set of poisoned data points):
\begin{align*}
\underset{\mathcal{D}_p}\max & \; L_{adv}(\hat\theta;\mathcal{D}_{test}) \\
s.t. \;\; & |\mathcal{D}_p| = \epsilon |\mathcal{D}_{c}| \\
& \mathcal{D}_{p} \subseteq \mathcal{F}_{\beta} \qquad \qquad \\
\text{where} \;\; & \hat\theta = \argmin_{\theta} \; \mathcal{L}(\theta;\mathcal{D}_{c} \cup \mathcal{D}_{p}).
\numberthis \label{objective}
\end{align*}
\ignore{
\begin{align*}
\underset{\mathcal{D}_p}\max & \; L(\hat\theta;\mathcal{D}_{test}) \qquad \;\;\;\;\;\;\; \text{($L$ being the loss function and $\hat\theta$ model parameter)}\\
s.t. \;\; & |\mathcal{D}_p| = \epsilon |\mathcal{D}_{c}| \qquad \;\;\;\;\;\; \text{($|\mathcal{D}_p|$ and $|\mathcal{D}_{c}|$ number of poisoned and clean instances)}\\
& \mathcal{D}_{p} \subseteq \mathcal{F}_{\beta} \qquad \qquad \;\;\;\; \text{($\mathcal{F}_{\beta}$ denotes the feasible set)}\\
& \mathcal{D}_{p} \subseteq (\mathcal{D}_{a} \cup \mathcal{D}_{d}) \qquad \text{($\mathcal{D}_{a}$ and $\mathcal{D}_{d}$ advantaged and disadvantaged instances)}\\
& \beta = B(\mathcal{D}_{c} \cup \mathcal{D}_{p}) \qquad \text{($\beta$ anomaly detector parameters)}\\
\text{where} \;\; & \hat\theta = \argmin_{\theta} \; L(\theta;\mathcal{D}_{c} \cup \mathcal{D}_{p}).
\numberthis \label{objective}
\end{align*}
}
In essence, the adversary attempts to maximize its test loss $L_{adv}$ by carefully selecting poisoned data points. These types of attacks are shown to be powerful against defenders that are trying to minimize their own loss $\mathcal{L}$, while the attacker is trying to harm the defense \cite{koh2018stronger}. In \cite{koh2018stronger}, authors propose to sample a positive $(\Tilde{x}_+,+1)$ and a negative $(\Tilde{x}_-,-1)$ instance and make $\epsilon |\mathcal{D}_c|$ copies from these sampled instances to serve as poisoned data points inversely proportional to the class balance such that there are $(|\mathcal{D}_{c}^+|\epsilon)$ copies from the negative poison instance $(\Tilde{x}_-,-1)$ and $(|\mathcal{D}_{c}^-|\epsilon)$ copies from the positive poison instance $(\Tilde{x}_+,+1)$ in which $|\mathcal{D}_{c}^+|$ and $|\mathcal{D}_{c}^-|$ represent the number of positive and negative points in the clean data respectively.
\section{Poisoning Attacks against Fairness}
\begin{algorithm}[h]
\SetAlgoLined
Input: clean data set $\mathcal{D}_{c}=\{(x_1,y_1),(x_2,y_2),...,(x_n,y_n)\}$, poison fraction $\epsilon$, and step size $\eta$. \\
Output: poisoned data set $\mathcal{D}_{p}=\{(\Tilde {x}_1,\Tilde{y}_1),(\Tilde{x}_2,\Tilde{y}_2),...,(\Tilde{x}_{\epsilon n},\Tilde{y}_{\epsilon n})\}$. \\
From $\mathcal{D}_{a}$ randomly sample the positive poisoned instance $\mathcal{I}_{+} \leftarrow (\Tilde{x}_1,\Tilde{y}_1)$. \\
From $\mathcal{D}_{d}$ randomly sample the negative poisoned instance $\mathcal{I}_{-} \leftarrow (\Tilde{x}_2,\Tilde{y}_2)$. \\
Make copies from $\mathcal{I}_{+}$ and $\mathcal{I}_{-}$ until having $\epsilon |\mathcal{D}_{c}|$ poisoned copies $\mathcal{C}_{p}$. \\
Load poisoned data set $\mathcal{D}_{p} \leftarrow \{\mathcal{C}_{p}\}$. \\
Load feasible set by applying anomaly detector $B$ $ \mathcal{F}_{\beta} \leftarrow B(\mathcal{D}_{c} \cup \mathcal{D}_{p})$. \\
\For{t= 1,2,...}{
$\hat{\theta} \leftarrow argmin_{\theta} \; \mathcal{L}(\theta;(\mathcal{D}_{c} \cup \mathcal{D}_{p})).$ \\
Pre-compute $g^{\top}_{\hat{\theta},\mathcal{D}_{test}} H^{-1}_{\hat{\theta}}$ from $L_{adv}$ for details refer to \cite{koh2018stronger}. \\
\For{i= 1,2}{
Set $\Tilde{x}_{i}^0 \leftarrow \Tilde{x}_{i} - \eta g^{\top}_{\hat{\theta},\mathcal{D}_{test}} H^{-1}_{\hat{\theta}} \frac{\partial^2 \ell(\hat{\theta};\Tilde{x}_i,\Tilde{y}_i) }{\partial \hat{\theta}\partial \Tilde{x}_i}. $ \\
Set $\Tilde{x}_{i} \leftarrow argmin_{x \in \mathcal{F}_{\beta}} ||x - \Tilde{x}_{i}^0 ||_2.\;\;\;$ (To project $\mathcal{D}_{p}$ back to $\mathcal{F}_{\beta}$).\\
}
Update copies $\mathcal{C}_{p}$ based on updates on $\mathcal{I}_{+}$ and $\mathcal{I}_{-}$. \\
Update feasible set $ \mathcal{F}_{\beta} \leftarrow B(\mathcal{D}_{c} \cup \mathcal{D}_{p})$. \\
}
\caption{Influence Attack on Fairness}
\label{hard_biasing_influence}
\end{algorithm}
\begin{figure*}[h]
\includegraphics[width=0.17\textwidth,trim=1cm 6cm 25.3cm 6.2cm,clip=true]{images/poisoned.pdf}
\begin{subfigure}[b]{0.37\textwidth}
\caption{Before Attack}
\includegraphics[width=\textwidth,trim=8.8cm 3cm 10.3cm 4cm,clip=true]{images/poison_pic_bold3.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.37\textwidth}
\caption{Anchoring Attack}
\includegraphics[width=\textwidth,trim=8.8cm 3cm 10.3cm 4cm,clip=true]{images/heuristic_attack.pdf}
\end{subfigure}
\caption{Anchoring attack representation. The figure on the left represents the before attack, while the right figure represents the anchoring attack in which poisoned points are located in close vicinity (depicted as the large solid circle) of target points.}
\label{attack_disc}
\end{figure*}
Now that we have discussed poisoning attacks, we will discuss how these attacks can be extended to fairness. We follow a common fairness setup where there are two groups: advantaged and disadvantaged. An example of advantaged and disadvantaged groups can be male and female in the job market where males could have advantage over females in getting hired in certain jobs. We assume that all poisoned points belong to either the advantaged or disadvantaged group, $\mathcal{D}_{p} \subseteq (\mathcal{D}_{a} \cup \mathcal{D}_{d})$, in which $\mathcal{D}_{a}$ represents data points from the advantaged demographic group and $\mathcal{D}_{d}$ represents data points from the disadvantaged demographic group.
\subsection{Influence Attack on Fairness}
For the influence attack on fairness, we use the influence attack introduced in \cite{koh2018stronger,koh2017understanding}, with a modification that includes the demographic information, in which the attack tries to maximize a given loss.
We then incorporate a loss function maximizing which using the influence attack can harm fairness. In \cite{zafar2015learning}, authors propose a loss function for fair classification with a constraint involving the covariance between the sensitive features ($z$) and the signed distance from feature vectors to the decision boundary ($d_{\theta}(x)$) formalized as:
\[ Cov(z,d_{\theta}(x)) \approx \frac{1}{N} \sum_{i=1}^N (z_i - \Bar{z})d_{\theta}(x_i). \]
By combining the above constraint with the original classification loss and maximizing it, the attacker can harm both fairness and accuracy at the same time via a regularization term, $\lambda$, that controls the trade-off between these two terms. Thus, the loss in our influence attack on fairness contains two parts: $\ell_{acc}$ and $\ell_{fairness}$ in which $\ell_{acc}$ controls for accuracy and $\ell_{fairness}$ controls for fairness constraints.
\begin{align*}
L_{adv}(\hat\theta;\mathcal{D}_{test}) & = \ell_{acc} +\lambda \ell_{fairness} \\
where \;\; \ell_{fairness} & = \frac{1}{N} \sum_{i=1}^N (z_i - \Bar{z})d_{\hat\theta}(x_i).
\numberthis \label{hard_biasing_loss}
\end{align*}
In other words, the influence attack on fairness would try to harm the fairness constraint and affect a model with respect to disparate impact~\cite{zafar2017fairness}. This loss can affect a model in terms of both fairness and accuracy with the regularization term $\lambda$ that controls the trade-off. In order to maximize the loss in
\eqref{hard_biasing_loss}, we use the influence attack strategy \cite{koh2018stronger,koh2017understanding} with changes that would incorporate demographic information as shown in Algorithm \ref{hard_biasing_influence}. Similar to the convention in \cite{koh2018stronger}, we sample one positive and one negative instance uniformly at random and make copies of the sampled instances that serve as our poisoned points. However, since we now have to take demographics into consideration for maximizing the bias and harming fairness, we sample the positive instance from $\mathcal{D}_{a}$ and the negative instance from $\mathcal{D}_{d}$. Notice that the opposite is also possible if an adversary wants to skew the disadvantaged group into being advantageous; however, for the goals of this paper and showing how our methods can increase the bias and harm fairness, we follow the aforementioned sampling procedure.
\subsection{Anchoring Attack}
We now describe a simple generic anchoring attack that can work with any loss function. Our results indicate that the proposed attack harms the model with regard to fairness.
The anchoring attack works as follows (details in Algorithm \ref{soft_biasing_alg}). First, the attacker samples a target $x_{target}$ that belongs to the clean data, $x_{target} \in \mathcal{D}_c$. Next, the attacker generates poisoned data point $\Tilde{x}$ in the vicinity of $x_{target}$, so that this new point has the same demographic but the opposite label, $demographic(x_{target})=demographic(\Tilde{x})$ and $y_{target}\neq \Tilde{y}$.
The general idea of the attack is to target some points ($x_{target}$) and cloud their labels through poisoned points that have opposite labels, which would lead to a skewed decision boundary, change in predictive labels of clean target points, and more biased outcomes. The right plot in Figure \ref{attack_disc} depicts an anchoring attack in which the poisoned points colored in black are placed to lie close to the target points that have the same demographic group but opposite label to bias the predictive outcome (black advantaged poisoned points with label +1 are targeting advantaged point with label -1, and black disadvantaged poisoned points with label -1 are targeting disadvantaged point with label +1). This placement of poisoned points in the space during the learning procedure will lead the decision boundary to change and, as a result, will cause more advantaged points to have a predictive outcome of +1 and more disadvantaged points to have a predictive outcome of -1, which is biasing the model's prediction.
$\mathbf{x_{target}}$ can be sampled in several ways. We introduce two ways, \textit{random} and \textit{non-random}, for sampling $\mathbf{x_{target}}$.
\textbf{Random Anchoring Attack.} In random anchoring attack, $\mathbf{x_{target}}$ is sampled uniformly at random for each demographic group.
\textbf{Non-random Anchoring Attack.} In the non-random anchoring attack, we choose popular $\mathbf{x_{target}}$
as our target for each demographic group.
\begin{algorithm}[h]
\SetAlgoLined
Input: clean data set $\mathcal{D}_{c}=\{(x_1,y_1),(x_2,y_2),...,(x_n,y_n)\}$, poison fraction $\epsilon$, and vicinity distance $\tau$. \\
Output: poisoned data set $\mathcal{D}_{p}=\{(\Tilde {x}_1,\Tilde{y}_1),(\Tilde{x}_2,\Tilde{y}_2),...,(\Tilde{x}_{\epsilon n},\Tilde{y}_{\epsilon n})\}$. \\
\For{t= 1,2,...}{
Sample negative $x_{target^{-}}$ from $\mathcal{D}_a$ and positive $x_{target^{+}}$ from $\mathcal{D}_d$ with random or non-random technique. \\
$\mathcal{G}_+$: Generate $(|\mathcal{D}_{c}^-|\epsilon)$ positive poisoned points $(\tilde{x}_{+},+1)$ with $\mathcal{D}_a$ in the close vicinity of $x_{target^{-}}$ s.t. $||\tilde{x}_{+} - x_{target^{-}}||_2 \leq \tau$. \\
$\mathcal{G}_-$: Generate $(|\mathcal{D}_{c}^+|\epsilon)$ negative poisoned points $(\tilde{x}_{-},-1)$ with $\mathcal{D}_d$ in the close vicinity of $x_{target^{+}}$ s.t. $||\tilde{x}_{-} - x_{target^{+}}||_2 \leq \tau$.\\
Load $\mathcal{D}_p$ from the generated data above $\mathcal{D}_p
\leftarrow \mathcal{G}_+ \cup \mathcal{G}_-$. \\
Load the feasible set $ \mathcal{F}_{\beta} \leftarrow B(\mathcal{D}_{c} \cup \mathcal{D}_{p})$. \\
\For{i=1,2,...,$\epsilon n$}{
Set $\Tilde{x}_{i} \leftarrow argmin_{x \in \mathcal{F}_{\beta}} ||x - \Tilde{x}_{i} ||_2.\;\;\;$ (To project $\mathcal{D}_{p}$ back to $\mathcal{F}_{\beta}$).\\
}
$argmin_{\theta} \; \mathcal{L}(\theta;(\mathcal{D}_{c} \cup \mathcal{D}_{p})).$ \\
}
\caption{Anchoring Attack}
\label{soft_biasing_alg}
\end{algorithm}
Here, popular $\mathbf{x_{target}}$ means the point that is close to more similar instances $x_i$, eligible to serve as targets, such that $demographic(x_i) = demographic(x_{target})$ and $y_i = y_{target}$. By doing this, we can ensure to affect as much as points similar to $\mathbf{x_{target}}$ as possible to maximize our biasing goal. Pick $x$ with $max(c)$ as $x_{target}$ where $c$ is calculated for each $x$ as follows: $\forall x_i$ if $demographic(x_i) = demographic(x)$ and $y_i = y$ and $||x_i - x || < \sigma$ then increase $c$ for $x$.
| 7a95750157392e74cba91b8c2998ccfb35b4548b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec-introduction}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig01.pdf}
\caption{Mean rank of {\textsc{MiniRocket}} in terms of accuracy versus other SOTA methods over 30 resamples of 109 datasets from the UCR archive.}
\Description[In terms of accuracy, Apricot ranks just ahead of Rocket, but behind both TS-CHIEF and HIVE-COTE/TDE]{In terms of accuracy, Apricot ranks just ahead of Rocket, but behind both TS-CHIEF and HIVE-COTE/TDE. Apricot is in the same clique as InceptionTime, Rocket, TS-CHIEF, and HIVE-COTE, that is, the pairwise differences between these classifiers are not statistically significant.}
\label{fig-rank-ucr109}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig02.pdf}
\caption{Transform time for {\textsc{MiniRocket}} versus {\textsc{Rocket}} for the same 109 datasets from the UCR archive.}
\Description[In terms of transform time, Apricot is between approximately 5 and 70 times faster than Rocket]{Scatter plot showing the total transform time (training and test) for Apricot against the total transform time for Rocket for 109 datasets from the UCR archive, on a log scale. In terms of total transform time, Apricot is between approximately 5 and 70 times faster than Rocket. The difference in total transform time increases, that is, Apricot is relatively faster, as total transform time increases. For shorter transform times, Apricot is closer to 10 times faster than Rocket, for longer transform times, Apricot is approaching 70 times faster than Rocket.}
\label{fig-transform-time-minirocket-vs-rocket}
\end{figure}
Until recently, the most accurate methods for time series classification were limited by high computational complexity. While there have been considerable advances in recent years, computational complexity and a lack of scalability remain persistent problems.
{\textsc{Rocket}} \citep{dempster_etal_2020} achieves state-of-the-art accuracy with a fraction of the computational expense of any method of comparable accuracy by transforming input time series using random convolutional kernels, and using the transformed features to train a linear classifier. We show that it is possible to reformulate {\textsc{Rocket}}, making it up to $75$ times faster on larger datasets, and making it almost entirely deterministic (and optionally, with additional computational expense, fully deterministic), while maintaining essentially the same accuracy. We call this method {\textsc{MiniRocket}} (for \textbf{MINI}mally \textbf{R}and\textbf{O}m \textbf{C}onvolutional \textbf{KE}rnel \textbf{T}ransform).
Like {\textsc{Rocket}}, {\textsc{MiniRocket}} transforms input time series using convolutional kernels, and uses the transformed features to train a linear classifier. However, unlike {\textsc{Rocket}}, {\textsc{MiniRocket}} uses a small, fixed set of kernels, and is almost entirely deterministic. {\textsc{MiniRocket}} maintains the two most important aspects of {\textsc{Rocket}}: dilation and PPV, i.e., `proportion of positive values' pooling \citep{dempster_etal_2020}. {\textsc{MiniRocket}} exploits various properties of the kernels, and of PPV, in order to massively reduce the time required for the transform. {\textsc{MiniRocket}} demonstrates that, while random convolutional kernels are highly effective, it is possible to achieve essentially the same accuracy using a mostly-deterministic and much faster procedure.
Figure \ref{fig-rank-ucr109} shows the mean rank of {\textsc{MiniRocket}} in terms of accuracy versus other state-of-the-art methods over 30 resamples of 109 datasets from the UCR archive of benchmark time series \citep{dau_etal_2019}. On average, {\textsc{MiniRocket}} is marginally more accurate than {\textsc{Rocket}}, and slightly less accurate than the most accurate current methods.
\begin{sloppypar}
Figure \ref{fig-transform-time-minirocket-vs-rocket} shows total transform time (training and test) for {\textsc{MiniRocket}} versus {\textsc{Rocket}}, for the same 109 datasets. On average, {\textsc{MiniRocket}} is more than $30$ times faster than {\textsc{Rocket}} (the advantage is even greater for larger datasets: see Section \ref{subsec-scalability}). Restricted to a single CPU core, total compute time for {\textsc{Rocket}} over all 109 datasets is 2 hours 2 minutes (1h 55m transform time), versus just 8 minutes for {\textsc{MiniRocket}} (2m 30s transform time). To put this in context, total compute time for {\textsc{MiniRocket}} for all 109 datasets is less than the compute time for {\textsc{Rocket}} for just one of those datasets. (Compute times are averages over 30 resamples, run on a cluster using Intel Xeon E5-2680 v3/4 and Xeon Gold 6150 CPUs, restricted to a single CPU core per dataset per resample.)
\end{sloppypar}
While only broadly comparable due to hardware and software differences, total compute time for the same 109 datasets using a single CPU thread is approximately 13 hours for cBOSS, more than a day for CIF, more than two days for TDE, approximately a week for Proximity Forest, more than two weeks for HIVE-COTE, and several weeks for TS-CHIEF \citep{bagnall_etal_2020,middlehurst_etal_2020a,middlehurst_etal_2020b}. Total compute time for InceptionTime (using GPUs, and for the original training/test splits rather than resamples) is more than 4 days \citep{ismailfawaz_etal_2020}.
{\textsc{MiniRocket}} represents a significant advance in accuracy relative to computational cost. {\textsc{MiniRocket}} is significantly faster than any other method of comparable accuracy (including {\textsc{Rocket}}), and significantly more accurate than any other method of even roughly-similar computational expense.
The rest of this paper is structured as follows. In Section \ref{sec-related-work}, we review relevant related work. In Section \ref{sec-method}, we detail the changes from {\textsc{Rocket}} to {\textsc{MiniRocket}}. In Section \ref{sec-experiments}, we present experimental results for {\textsc{MiniRocket}} in terms of accuracy and scalability, as well as a sensitivity analysis in relation to key parameter choices.
\section{Related Work} \label{sec-related-work}
\subsection{Current State of the Art}
Recent advances in accuracy have largely superseded the most accurate methods originally identified in \citep{bagnall_etal_2017}. According to \citep{bagnall_etal_2020,middlehurst_etal_2020a,middlehurst_etal_2020b}, the most accurate current methods for time series classification are HIVE-COTE and its variants \citep{lines_etal_2018}, TS-CHIEF \citep{shifaz_etal_2020}, InceptionTime \citep{ismailfawaz_etal_2020}, and {\textsc{Rocket}} \citep{dempster_etal_2020}. However, while accuracy has improved, with some exceptions computational complexity and a lack of scalability remain persistent problems.
TS-CHIEF builds on Proximity Forest, an ensemble of decision trees using distance measures as splitting criteria \citep{lucas_etal_2019}. In addition to distance measures, TS-CHIEF uses interval-based and spectral-based splitting criteria \citep{shifaz_etal_2020}.
InceptionTime is an ensemble of convolutional neural networks based on the Inception architecture, and is the most accurate convolutional neural network model for time series classification \citep{ismailfawaz_etal_2020}.
The Temporal Dictionary Ensemble (TDE) is a recent dictionary method based on the frequency of occurrence of patterns in time series \citep{middlehurst_etal_2020a}. TDE combines aspects of earlier dictionary methods including cBOSS \citep{middlehurst_etal_2019}, a more scalable variant of BOSS \citep{schafer_2015}.
Catch22 is a transform based on 22 predefined time series features, used in combination with a decision tree or random forest \citep{lubba_etal_2019}. On its own, catch22 is fast, but highly inaccurate: see \citep{dempster_etal_2020,middlehurst_etal_2020b}. The Canonical Interval Forest (CIF) is a recent method which adapts the Time Series Forest (TSF) to use catch22 features \citep{middlehurst_etal_2020b}. CIF is significantly more accurate than either catch22 or TSF.
\begin{sloppypar}
HIVE-COTE is an ensemble of other methods including BOSS and TSF. Two recent variants of HIVE-COTE, namely HIVE-COTE/TDE (using TDE in place of BOSS) and HIVE-COTE/CIF (using CIF in place of TSF) have been shown to be significantly more accurate than HIVE-COTE, or any other existing method for time series classification \citep{middlehurst_etal_2020a,middlehurst_etal_2020b}. These variants are, in turn, based on an updated `base' version of HIVE-COTE \citep{bagnall_etal_2020}.
\end{sloppypar}
While state of the art in terms of accuracy, with the exception of cBOSS these methods are limited by high computational complexity, requiring days or even weeks to train on the datasets in the UCR archive. While more scalable, cBOSS is significantly less accurate than most of the other methods.
\subsection{{\textsc{Rocket}}}
{\textsc{Rocket}} achieves state-of-the-art accuracy, matching the most accurate methods for time series classification (with the exception of the most recent variants of HIVE-COTE), but is considerably faster and more scalable than other methods of comparable accuracy \citep{dempster_etal_2020}.
{\textsc{Rocket}} transforms input time series using random convolutional kernels, and uses the transformed features to train a linear classifier. Each input time series is convolved with $10{,}000$ random convolutional kernels. {\textsc{Rocket}} applies global max pooling and PPV (for `proportion of positive values') pooling to the resulting convolution output to produce two features per kernel per input time series, for a total of $20{,}000$ features per input time series. The transformed features are then used to train a linear classifier: a ridge regression classifier, or logistic regression trained using stochastic gradient descent (for larger datasets).
The kernels are random in terms of their length, weights, bias, dilation, and padding: see Section \ref{subsec-removing-randomness}. The two most important aspects of {\textsc{Rocket}} in terms of achieving state-of-the-art accuracy are the use of dilation, sampled on an exponential scale, and the use of PPV. {\textsc{Rocket}} forms the basis for {\textsc{MiniRocket}}. The differences between {\textsc{Rocket}} and {\textsc{MiniRocket}} are detailed in Section \ref{sec-method}.
\subsection{Other Methods}
The use of a small, fixed set of kernels differentiates {\textsc{MiniRocket}} from both {\textsc{Rocket}}, which uses random kernels, and convolutional neural networks such as InceptionTime, which use learned kernels. It also differentiates {\textsc{MiniRocket}} from other methods with at least superficial similarities to {\textsc{Rocket}}, such as random shapelet methods as in \citep{karlsson_etal_2016}, and other random methods such as those based on \citep{rahimi_and_recht_2008}.
In using kernels with weights constrained to two values (see Section \ref{subsubsec-weights}), there are obvious similarities with binary and quantised convolutional neural networks \citep{rastegari_etal_2016,hubara_etal_2018}. {\textsc{MiniRocket}} makes use of at least two advantages of binary/quantised kernels, namely, the ability to perform the convolution operation via addition, as well as efficiencies arising from the relatively small number of possible binary kernels of a given size, e.g., \citep{rastegari_etal_2016,juefeixu_etal_2017,hubara_etal_2018}. However, while the kernels used in {\textsc{MiniRocket}} are binary in the sense of having only two values, these values are \textit{not} $0$ and $1$ (or $-1$ and $1$). In fact, the actual values of the weights are not important: see Section \ref{subsubsec-weights}. {\textsc{MiniRocket}} does not use bitwise operations, and the input and convolution output are used at full precision.
The optimisations used in {\textsc{MiniRocket}} are similar in motivation to several optimisations developed for convolutional neural networks, i.e., broadly speaking, to reduce the number of operations (especially multiplications) required to perform the convolution operation, e.g., \citep{liu_etal_2015,lavin_and_gray_2016,chollet_2017,mehta_etal_2018}. In precomputing the product of the kernel weights and the input, and using those precomputed values to construct the convolution output (see Sections \ref{subsubsec-factoring-out} and \ref{subsubsec-all-kernels-at-once}), the optimisations used in {\textsc{MiniRocket}} bear some resemblance to highly simplified versions of shift-based methods \citep{wu_etal_2018}, where conventional convolutional kernels are replaced by a combination of $1 \times 1$ convolutions and spatial shifts in the input, and lookup-based methods \citep{bagherinezhad_etal_2017}, where the convolution operation is performed via linear combinations of the precomputed convolution output for a small `dictionary' of kernels.
However, the optimisations used in {\textsc{MiniRocket}} are much simpler than these methods. {\textsc{MiniRocket}} uses a fixed set of kernels, and uses the convolution output for these kernels directly, rather than through a learned linear combination, c.f., e.g., \citep{bagherinezhad_etal_2017,juefeixu_etal_2017}. The optimisations arise as a natural result of using this fixed set of kernels, rather than being general-purpose optimisations.
Several things further distinguish {\textsc{MiniRocket}} (and {\textsc{Rocket}}) from most approaches involving convolutional neural networks. The features produced by the transform are all independent of each other (there is no hidden layer). Neither the convolution output nor the pooled features are transformed through, e.g., a sigmoid function or rectified linear unit (ReLU). As such, the classifier learns a direct linear function of the features produced by the transform. {\textsc{MiniRocket}} is also distinguished by its use of dilation (similar to using many different dilations in a single convolutional layer, with dilations taking any integer value not just powers of two), and PPV.
\section{Method} \label{sec-method}
{\textsc{MiniRocket}} involves making certain key changes in order to remove almost all randomness from {\textsc{Rocket}} (Section \ref{subsec-removing-randomness}), and exploiting these changes in order to dramatically speed up the transform (Section \ref{subsec-optimising-the-transform}). In tuning kernel length, weights, bias, etc., we have restricted ourselves to the same 40 `development' datasets as used in \citep{dempster_etal_2020}, with the same aim of avoiding overfitting the entire UCR archive. (Note, however, that it is not necessarily the aim of {\textsc{MiniRocket}} to maximise accuracy \textit{per se}, but rather to balance accuracy with parameter choices which remove randomness and are conducive to optimising the transform.) The procedures for setting the parameter values and performing the transform are set out in \texttt{\ref{pseudo-fit}} and \texttt{\ref{pseudo-transform}} in Appendix \ref{sec-appendix-pseudocode}.
As for {\textsc{Rocket}}, we implement {\textsc{MiniRocket}} in Python, compiled via Numba \citep{lam_etal_2015}. We use a ridge regression classifier from scikit-learn \citep{pedregosa_etal_2011}, and logistic regression implemented using PyTorch \citep{paszke_etal_2019}. Our code is available at: \url{https://github.com/angus924/minirocket}.
\subsection{Removing Randomness} \label{subsec-removing-randomness}
\begin{table}
\centering
\caption{Summary of changes from {\textsc{Rocket}} to {\textsc{MiniRocket}}.}
\begin{tabular}{ccc}
\toprule
& {\textsc{Rocket}} & {\textsc{MiniRocket}} \\
\midrule
length & $\{7, 9, 11\}$ & 9 \\
weights & $\mathcal{N}(0, 1)$ & $\{-1, 2\}$ \\
bias & $\mathcal{U}(-1, 1)$ & from convolution output \\
dilation & random & fixed (rel. to input length) \\
padding & random & fixed \\
features & PPV + max & PPV \\
num. features & 20K & 10K \\
\bottomrule
\end{tabular}
\label{table-changes-rocket-to-minirocket}
\end{table}
{\textsc{Rocket}} uses kernels with lengths selected randomly from $\{7,9,11\}$, weights drawn from $\mathcal{N}(0, 1)$, bias terms drawn from $\mathcal{U}(-1, 1)$, random dilations, and random paddings. Two features, PPV and max, are computed per kernel, for a total of $20{,}000$ features. {\textsc{MiniRocket}} is characterised by a number of key changes to the kernels in terms of length, weights, bias, dilation, and padding, as well as resulting changes to the features, as summarised in Table \ref{table-changes-rocket-to-minirocket}.
\subsubsection{Length}
{\textsc{MiniRocket}} uses kernels of length 9, with weights restricted to two values, building on the observation in \citep{dempster_etal_2020} that weights drawn from $\{-1, 0, 1\}$ produce similar accuracy to weights drawn from $\mathcal{N}(0, 1)$.
In order to maximise computational efficiency, the set of kernels should be as small as possible: see Section \ref{subsubsec-reusing-output}. The set of possible two-valued kernels grows exponentially with length. There are $2^{3} = 8$ possible kernels of length 3, but $2^{15} = 32{,}768$ possible kernels of length 15. With more than two values, the set of possible kernels grows even faster with length. For example, there are $3^{15} \approx 14\text{ million}$ possible three-valued kernels of length 15.
There are $2^{9} = 512$ possible two-valued kernels of length 9. {\textsc{MiniRocket}} uses a subset of 84 of these kernels, a subset which balances accuracy with the computational advantages of using a small number of kernels: see Section \ref{subsubsec-sensitivity-kernels}. (A length of 9 is also consistent with the average length used in {\textsc{Rocket}}.)
\subsubsection{Weights} \label{subsubsec-weights}
Kernels with weights restricted to two values, $\alpha$ and $\beta$, can be characterised in terms of the number of weights with the value $\beta$ (or, equivalently, the number of weights with the value $\alpha$). In this sense, the full set of two-valued kernels of length 9 includes the subset of kernels with 1 value of $\beta$ (e.g., $[\alpha,\alpha,\alpha,\alpha,\alpha,\alpha,\alpha,\alpha,\beta]$), the subset of kernels with 2 values of $\beta$ (e.g., $[\alpha,\alpha,\alpha,\alpha,\alpha,\alpha,\alpha,\beta,\beta]$), and so on. {\textsc{MiniRocket}} uses the subset kernels with 3 values of $\beta$:
\begin{gather*}
[\alpha,\alpha,\alpha,\alpha,\alpha,\alpha,\beta,\beta,\beta] \\
[\alpha,\alpha,\alpha,\alpha,\alpha,\beta,\alpha,\beta,\beta] \\
[\alpha,\alpha,\alpha,\alpha,\beta,\alpha,\alpha,\beta,\beta] \\
...
\end{gather*}
For {\textsc{MiniRocket}}, we set $\alpha = -1$ and $\beta = 2$. However, the choice of $\alpha$ and $\beta$ is arbitrary, in the sense that the scale of these values is unimportant. For an input time series, $X$, kernel, $W$, and bias, $b$, PPV is given by $\text{PPV}(X * W - b) = \frac{1}{n} \sum [X * W - b > 0]$ or, equivalently, $\text{PPV}(X * W) = \frac{1}{n} \sum [X * W > b]$, where `$*$' denotes convolution, and $[X \in a]$ denotes the indicator function. As such, computing PPV is essentially equivalent to computing the empirical cumulative distribution function. Accordingly, the scale of the weights is unimportant, because bias values are drawn from the convolution output, $X * W$ (see Section \ref{subsubsec-bias}), and so by definition match the scale of the weights and the scale of the input. (Hence, in contrast to {\textsc{Rocket}}, it is not necessary to normalise the input.)
It is only important that the sum of the weights should be zero or, equivalently, that $\beta = -2 \alpha$. Otherwise, the values of $\alpha$ and $\beta$ are not important. This constraint---that the weights sum to zero---ensures that the kernels are only sensitive to the relative magnitude of the values in the input, i.e., that the convolution output is invariant to the addition or subtraction of any constant value to the input, i.e., $X * W = (X \pm c) * W$.
As PPV is bounded between 0 and 1, in computing PPV for a given kernel, $W$, we get an equivalent feature for the inverted kernel, $-W$, `for free': see Section \ref{subsubsec-computing-ppv}. Accordingly, there is no need to use both the set of kernels with weights $\alpha = -1$ and $\beta = 2$, and the corresponding inverted set of kernels with weights $\alpha = 1$ and $\beta = -2$, as we get these inverted kernels `for free'.
The set of 84 kernels of length 9 with three weights with the value $\beta = 2$, and six weights with the value $\alpha = -1$, has the desirable properties of being a relatively small, fixed set of kernels---conducive to the optimisations pursued in Section \ref{subsec-optimising-the-transform}---and producing high classification accuracy. However, we stress that there is not necessarily anything `special' about this set of kernels. Other subsets of kernels of length 9, and kernels of other lengths, produce similar accuracy: see Section \ref{subsubsec-sensitivity-kernels}. This is in addition to the observations in \citep{dempster_etal_2020}, i.e., that kernels (of various lengths) with weights drawn from $\mathcal{N}(0, 1)$, or from $\{-1, 0, 1\}$, are also effective.
\subsubsection{Bias} \label{subsubsec-bias}
Bias values are drawn from the convolution output, and are used to compute PPV as set out above in Section \ref{subsubsec-weights}. By default, for a given kernel/dilation combination, bias values are drawn from the quantiles of the convolution output for a single, randomly-selected training example. For a given kernel, $W$, and dilation, $d$, we compute the convolution output for a randomly-selected training example, $X$, i.e., $W_{d} * X$. We take, e.g., the $[0.25, 0.5, 0.75]$ quantiles from $W_{d} * X$ as bias values, to be used in computing PPV. We use a low-discrepancy sequence to assign quantiles to different kernel/dilation combinations \citep{schretter_etal_2016}.
The selection of training examples for the purpose of sampling bias values is the only stochastic element of {\textsc{MiniRocket}}. Further, while the choice of training example is random, in drawing bias values from the convolution output, we are selecting values produced by an otherwise entirely deterministic procedure. This is why we characterise {\textsc{MiniRocket}} as `minimally random'.
For the deterministic variant of {\textsc{MiniRocket}}, bias values are drawn from the convolution output for the entire training set, rather than a single, randomly-selected training example. This is the only substantive difference between the default and deterministic variants, and the difference in accuracy between the two variants is negligible: see Section \ref{subsec-ucr-archive}.
The advantage of using the entire training set is an entirely deterministic transform, for applications where this is desirable. However, this comes at additional computational cost, which is unlikely to be practical for larger datasets. Crucially, however, it demonstrates that the accuracy of {\textsc{Rocket}} is achievable using an entirely deterministic transform. In practice, using a single, randomly-selected training example has little impact in terms of accuracy.
A variant of {\textsc{Rocket}} using the same method for sampling bias values as {\textsc{MiniRocket}} is slightly more accurate than default {\textsc{Rocket}} but, overall, the difference is relatively minor: see Section \ref{subsec-ucr-archive}.
\subsubsection{Dilation} \label{subsubsec-dilation}
Dilation is used to `spread' a kernel over the input. For dilation, $d$, a given kernel is convolved with every $d^{\text{th}}$ element of the input \citep{yu_and_koltun_2016,dempster_etal_2020}. Each kernel is assigned the same fixed set of dilations, adjusted to the length of the input time series. We specify dilations in the range $D = \{\lfloor 2^{0} \rfloor, ..., \lfloor 2^{\text{max}} \rfloor\}$, where the exponents are uniformly spaced between 0 and $\text{max} = \log_2 ( l_{\text{input}} - 1 ) / ( l_{\text{kernel}} - 1 )$, where $l_{\text{input}}$ is input length and $l_{\text{kernel}}$ is kernel length (i.e., 9), such that the maximum effective length of a kernel, including dilation, is the length of the input time series. The count of each unique integer dilation value in $D$ determines the number of features to be computed per dilation (scaled according to the total number of features), ensuring that, as in {\textsc{Rocket}}, exponentially more features are computed for smaller dilations.
As time series length increases, the number of possible dilation values increases. This means that, for a fixed number of features, the number of features computed per dilation decreases (unless constrained in some way), making the transform less efficient: see Section \ref{subsubsec-reusing-output}. Hence, by default, we limit the maximum number of dilations per kernel to 32. While technically an additional hyperparameter, this has little effect on accuracy (see Section \ref{subsubsec-sensitivity-dilation}), and is intended to be kept at its default value.
\subsubsection{Padding} \label{subsubsec-padding}
Padding is alternated for each kernel/dilation combination such that, overall, half the kernel/dilation combinations use padding, and half do not. As for {\textsc{Rocket}}, {\textsc{MiniRocket}} uses standard zero padding. In effect, zeros are added to the start and end of each input time series such that the convolution operation begins with the `middle' element of the kernel centered on the first element of the time series, and ends with the `middle' element of the kernel centered on the last element of the time series \citep{goodfellow_etal_2016}.
\subsubsection{Features} \label{subsubsec-features}
Given the other changes, there is no longer any benefit in terms of accuracy in using global max pooling in addition to PPV: see Section \ref{subsubsec-sensitivity-features}. Accordingly, {\textsc{MiniRocket}} `drops' global max pooling and uses only PPV.
We do not replace global max pooling with additional PPV features. As for {\textsc{Rocket}}, the number of features represents a tradeoff between accuracy and computational expense. {\textsc{MiniRocket}} with $10{,}000$ features already matches {\textsc{Rocket}} in terms of accuracy, and there is little or no benefit in terms of accuracy to increasing the number of features beyond $10{,}000$: see Section \ref{subsubsec-sensitivity-num-features}.
Accordingly, by default, {\textsc{MiniRocket}} uses $10{,}000$ features (or, more precisely, the nearest multiple of 84---the number of kernels---less than $10{,}000$, i.e., $9{,}996$). While technically a hyperparameter, this is intended to be kept at its default value.
\subsection{Optimising the Transform} \label{subsec-optimising-the-transform}
{\textsc{MiniRocket}} takes advantage of the properties of the small, fixed set of two-valued kernels, and of PPV, to significantly speed up the transform through four key optimisations:
\begin{enumerate}
\item computing PPV for $W$ and $-W$ at the same time;
\item reusing the convolution output to compute multiple features;
\item avoiding multiplications in the convolution operation; and
\item for each dilation, computing all kernels (almost) `at once'.
\end{enumerate}
\subsubsection{Computing PPV for $W$ and $-W$ at the Same Time} \label{subsubsec-computing-ppv}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig03.pdf}
\caption{Illustration of $\text{{\normalfont PPV}}(X * W - b) = 1 - \text{{\normalfont PPV}}(b - (X * W))$.}
\Description[PPV for a given kernel, W, is equivalent to a very similar feature for the inverted kernel, negative W]{A toy example showing that PPV for a given kernel, W, is equivalent to a very similar feature for the inverted kernel, negative W. Two plots: the plot on the left shows the convolution output for input X and kernel W minus bias b, and the resulting proportion of positive values, or PPV; the plot on the right shows the convolution output for the same input X and the inverted kernel negative W plus bias b, and the proportion of negative values, or PNV. PNV for the inverted kernel negative W is shown to be the same as PPV for the kernel W.}
\label{fig-diagram-ppv-inverse}
\end{figure}
For $C = X * W - b$, PPV is given by $\text{PPV}(C) = \frac{1}{n} \sum [c > 0].$ PPV is bounded between 0 and 1. By definition, the proportion of negative values (or PNV) is the complement of PPV, i.e., $1 - \text{PPV}(X * W - b) = \text{PNV}(X * W - b).$ That is, by computing PPV, we also implicitly compute PNV and vice versa. In this sense, PPV and PNV are equivalent.
Further, the convolution operation is associative, such that $X * -W = -(X * W)$. Accordingly, by computing PPV for a given kernel, $W$, we unavoidably also compute an equivalent feature (i.e., PNV) for $-W$, that is, $\text{PPV}(X * W - b) = 1 - \text{PPV}(b - (X * W)).$ This relationship is illustrated in Figure \ref{fig-diagram-ppv-inverse}. This means that, for the purposes of PPV, it is unnecessary to compute both $X * W$ and $X * -W$. In fact, it would be redundant to do so.
This means that, in practice, for a given set of kernels where each kernel, $W$, is matched by a corresponding inverted kernel, $-W$, we only need to perform the convolution operation for $W$, i.e., for half of the kernels. We get $-W$ `for free'. Accordingly, as set out in Section \ref{subsubsec-weights}, {\textsc{MiniRocket}} only uses a set of kernels with weights $\alpha = -1$ and $\beta = 2$, as it is unnecessary to also use the corresponding set of inverted kernels with weights $\alpha = 1$ and $\beta = -2$.
\subsubsection{Reusing the Convolution Output} \label{subsubsec-reusing-output}
For {\textsc{MiniRocket}}, the same kernel/dilation combination is used to compute multiple features, at least for smaller dilations (exponentially fewer features are computed for larger dilations: see Section \ref{subsubsec-dilation}).
For a given kernel, $W$, and dilation, $d$, we compute $C = X * W_{d}$ and then reuse the convolution output, $C$, to compute multiple features, i.e., for multiple different bias values. This has the effect that multiple features are computed with the computational cost of a single convolution operation, plus the much lower cost of computing PPV for each bias value.
\subsubsection{Avoiding Multiplications} \label{subsubsec-factoring-out}
Restricting the kernel weights to two values allows us to, in effect, `factor out' the multiplications from the convolution operation, and to perform the convolution operation using only addition.
For input time series $X = [x_{0}, x_{1}, ..., x_{n - 1}]$, and kernel $W = [w_{0}, w_{1}, ..., w_{m - 1}]$, with dilation, $d$, the convolution operation can be formulated as:
$$
X * W_{d} = \sum_{j=0}^{m - 1} x_{i - (\lfloor \frac{m}{2} \rfloor \cdot d) + (j \cdot d)} \cdot w_{j}, \forall i \in \{0, 1, ..., n - 1\}.
$$
Equivalently, the convolution operation can be thought of as the column sums of a matrix, $\boldsymbol{\hat{C}}$, where each row corresponds to the input time series multiplied by the appropriate kernel weight, and the alignment of the rows corresponds to dilation (values of 0 in $\boldsymbol{\hat{C}}$ represent zero padding), e.g.:
$$
\hat{\boldsymbol{C}} =
\begin{bmatrix}
0 & 0 & 0 & 0 & w_{0} x_{0} & \cdots \\
0 & 0 & 0 & w_{1} x_{0} & w_{1} x_{1} & \cdots \\
0 & 0 & w_{2} x_{0} & w_{2} x_{1} & w_{2} x_{2} & \cdots \\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots \\
w_{m - 1} x_{4} & w_{m - 1} x_{5} & w_{m - 1} x_{6} & w_{m - 1} x_{7} & w_{m - 1} x_{8} & \cdots
\end{bmatrix}
$$
The result of the convolution operation is given by the column sums of $\boldsymbol{\hat{C}}$, i.e., $C = X * W = \boldsymbol{1}^{\top}\boldsymbol{\hat{C}}$, where $\boldsymbol{1}$ is a vector, $[1,1,...,1]^{\top}$, of length $n$.
Where the weights of the kernels are restricted to two values, $\alpha$ and $\beta$, we can `factor out' the multiplications by precomputing $A = \alpha X$ and $B = \beta X$ and then, for a given kernel, e.g., $W = [\alpha, \beta, \alpha, ..., \alpha]$, completing the convolution operation by summation using $A = [a_{0}, a_{1}, ..., a_{n - 1}]$ and $B = [b_{0}, b_{1}, ..., b_{n - 1}]$:
$$
\hat{\boldsymbol{C}} =
\begin{bmatrix}
0 & 0 & 0 & 0 & a_{0} & \cdots & a_{n-5} \\
0 & 0 & 0 & b_{0} & b_{1} & \cdots & b_{n-4} \\
0 & 0 & a_{0} & a_{1} & a_{2} & \cdots & b_{n-3} \\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots & \vdots \\
a_{4} & a_{5} & a_{6} & a_{7} & a_{8} & \cdots & 0
\end{bmatrix}
$$
In other words, it is only necessary to compute $\alpha X$ and $\beta X$ once for each input time series, and then reuse the results to complete the convolution operation for each kernel by addition.
\subsubsection{Computing All the Kernels (Almost) `At Once'} \label{subsubsec-all-kernels-at-once}
We can take further advantage of using only two values for the kernel weights in order to perform most of the computation required for all 84 kernels `at once' for each dilation value. More precisely, as {\textsc{MiniRocket}} uses kernels with six weights of one value, and three weights of another value, we can perform $\frac{6}{9} = \frac{2}{3}$ of the computation for all 84 kernels `at once' for a given dilation.
This is possible by treating all kernel weights as $\alpha = -1$, precomputing convolution output, $C_{\alpha}$, and later adjusting $C_{\alpha}$ for each kernel. Per Section \ref{subsubsec-factoring-out}, $C_{\alpha}$ can be thought of as the column sums of a matrix with 9 rows, where each row corresponds to $\alpha X = -X$, aligned according to dilation. For example, for a dilation of 1:
$$
\hat{\boldsymbol{C}}_{\alpha} =
\begin{bmatrix}
0 & 0 & 0 & 0 & -x_{0} & \cdots & -x_{n-5} \\
0 & 0 & 0 & -x_{0} & -x_{1} & \cdots & -x_{n-4} \\
0 & 0 & -x_{0} & -x_{1} & -x_{2} & \cdots & -x_{n-3} \\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots & \vdots \\
-x_{4} & -x_{5} & -x_{6} & -x_{7} & -x_{8} & \cdots & 0
\end{bmatrix}
$$
For {\textsc{MiniRocket}}, the kernel weights are $\alpha = -1$ and $\beta = 2$. Let $\gamma = 3$, noting that $2 = -1 + 3$. As for $\hat{\boldsymbol{C}}_{\alpha}$, we then form $\hat{\boldsymbol{C}}_{\gamma}$, where each row corresponds to $\gamma X = 3X$, aligned according to dilation. For each kernel, $C_{\gamma}$ is equivalent to the column sums of those rows in $\hat{\boldsymbol{C}}_{\gamma}$ corresponding to the position of the $\beta$ weights in the given kernel. For example, for kernel $W = [\beta, \alpha, \beta, \alpha, \beta, \alpha, \alpha, \alpha, \alpha]$:
$$
\hat{\boldsymbol{C}}_{\gamma}^{(W)} =
\begin{bmatrix}
0 & 0 & 0 & 0 & 3x_{0} & \cdots & 3x_{n-5} \\
0 & 0 & 3x_{0} & 3x_{1} & 3x_{2} & \cdots & 3x_{n-3} \\
3x_{0} & 3x_{1} & 3x_{2} & 3x_{3} & 3x_{4} & \cdots & 3x_{n-1}
\end{bmatrix}
$$
The final convolution output for a given kernel is then given by $C = C_{\alpha} + C_{\gamma}$. In other words, we can reuse $C_{\alpha}$, computed once for a given dilation, to compute the convolution output for all 84 kernels for that dilation. For each kernel, computing $C$ only involves adding $C_{\gamma}$ to $C_{\alpha}$. In performing the convolution operation in this way, we only have to compute $C_{\gamma}$ for each kernel, i.e., $\frac{1}{3}$ of the computation otherwise required.
\subsection{Classifiers}
Like {\textsc{Rocket}}, {\textsc{MiniRocket}} is a transform, producing features which are then used to train a linear classifier. We use the same classifiers as {\textsc{Rocket}} to learn the mapping from the features to the classes, i.e., a ridge regression classifier or, for larger datasets, logistic regression trained using Adam \citep{kingma_and_ba_2015}. As for {\textsc{Rocket}}, we suggest switching from the ridge regression classifier to logistic regression when there are more training examples than features, i.e., when there are more than approximately $10{,}000$ training examples.
\subsection{Complexity}
Fundamentally, the scalability of {\textsc{MiniRocket}} remains the same as for {\textsc{Rocket}}: linear in the number of kernels/features ($k$), the number of examples ($n$), and time series length ($l_{\text{input}}$) or, formally, $O(k \cdot n \cdot l_{\text{input}})$. While {\textsc{MiniRocket}} uses a smaller number of kernel/dilation combinations, and computes multiple features for each kernel/dilation combination, complexity is still proportional to the number of kernels/features. Similarly, while {\textsc{MiniRocket}} `factors out' multiplications from the convolution operation, the number of addition operations is still proportional to the number of kernels and time series length, and while {\textsc{MiniRocket}} performs the majority of the computation required for all 84 kernels `at once', the remaining computation is still proportional to the number of kernels/features. However, within this broad class of complexity, the various optimisations pursued in Section \ref{subsec-optimising-the-transform} make {\textsc{MiniRocket}} significantly faster in practice.
\subsection{Memory}
Compared to {\textsc{Rocket}} (which does not store any intermediate values), {\textsc{MiniRocket}} temporarily stores up to 13 additional vectors, namely, $A = -X$, $G = \gamma X = 3X$ (plus 9 variants of $G$ pre-aligned for the given dilation), $C_{\alpha}$, and $C$: see Sections \ref{subsubsec-factoring-out} and \ref{subsubsec-all-kernels-at-once}. This is equivalent to storing 13 additional copies of a single input time series (approx. $1{,}000{,}000 \times 4 \times 13 = 52 \text{MB}$ for time series of length 1 million), which should be negligible in almost all cases.
When transforming the training set, the deterministic variant stores the convolution output for a given kernel/dilation combination for the entire training set, which is equivalent to storing one additional copy of the entire training set. This is impractical for larger datasets, which is why it is avoided by default.
\section{Experiments} \label{sec-experiments}
We evaluate {\textsc{MiniRocket}} on the datasets in the UCR archive (Section \ref{subsec-ucr-archive}), showing that, on average, {\textsc{MiniRocket}} is marginally more accurate than {\textsc{Rocket}}, and not significantly less accurate than the most accurate current methods for time series classification. We demonstrate the speed and scalability of {\textsc{MiniRocket}} in terms of both training set size and time series length (Section \ref{subsec-scalability}), showing that {\textsc{MiniRocket}} is up to $75$ times faster than {\textsc{Rocket}} on larger datasets. We also explore the effect of key parameters in relation to kernel length, bias, output features, and dilation (Section \ref{subsec-sensitivity-analysis}).
\subsection{UCR Archive} \label{subsec-ucr-archive}
We evaluate {\textsc{MiniRocket}} on the datasets in the UCR archive \citep{dau_etal_2019}. We compare {\textsc{MiniRocket}} against the most accurate current methods for time series classification, namely, HIVE-COTE/TDE (representative of HIVE-COTE and its variants), TS-CHIEF, InceptionTime, and {\textsc{Rocket}}, as well as TDE, CIF, cBOSS and Proximity Forest.
For consistency and direct comparability with the most recent published results for other state-of-the-art methods \citep{bagnall_etal_2020,middlehurst_etal_2020a,middlehurst_etal_2020b}, we evaluate {\textsc{MiniRocket}} on 30 resamples of 109 datasets from the archive. We use the same 30 resamples (including the default training/test split) as in \citep{bagnall_etal_2020,middlehurst_etal_2020a,middlehurst_etal_2020b}. (Full results are available in the accompanying repository.)
Figure \ref{fig-rank-ucr109} on page \pageref{fig-rank-ucr109} shows the mean rank of {\textsc{MiniRocket}} versus the other state-of-the-art methods. Methods for which the pairwise difference in accuracy is not statistically significant, per a Wilcoxon signed-rank test with Holm correction (as a post hoc test to the Friedman test), are connected with a black line \citep{demsar_2006,garcia_and_herrera_2008,benavoli_etal_2016}.
{\textsc{MiniRocket}} is, on average, marginally more accurate than {\textsc{Rocket}}, and somewhat less accurate than the most accurate current methods, namely TS-CHIEF and HIVE-COTE/TDE, although the differences in accuracy are not statistically significant. However, as noted in Section \ref{sec-introduction}, the total compute time for {\textsc{MiniRocket}} on these datasets is a tiny fraction of the total compute time required by the other methods (even {\textsc{Rocket}}, which is already considerably faster than even the fastest of the other methods).
\paragraph{{\textsc{MiniRocket}} versus {\textsc{Rocket}}.}
\begin{figure}
\centering
\includegraphics[width=0.65\linewidth]{./fig04.pdf}
\caption{Pairwise accuracy of {\textsc{MiniRocket}} versus {\textsc{Rocket}}.}
\Description[Apricot is more accurate than Rocket on 61 of 109 datasets]{Scatter plot showing the accuracy of Apricot against the accuracy of Rocket for 109 datasets from the UCR archive. Apricot is more accurate on 61 datasets, as accurate on 3 datasets, and less accurate on 45 datasets. The accuracy of Apricot and Rocket is similar for most datasets. For one dataset, PigAirWayPressure, Apricot is considerably more accurate than Rocket.}
\label{fig-pairwise-minirocket-vs-rocket}
\end{figure}
Figure \ref{fig-pairwise-minirocket-vs-rocket} shows the pairwise accuracy of {\textsc{MiniRocket}} versus {\textsc{Rocket}} for the same 109 datasets. Overall, {\textsc{MiniRocket}} and {\textsc{Rocket}} achieve very similar accuracy. {\textsc{MiniRocket}} is more accurate than {\textsc{Rocket}} on 61 datasets, and less accurate on 45 datasets, but the differences in accuracy are mostly small. The large difference in accuracy between {\textsc{MiniRocket}} and {\textsc{Rocket}} on one dataset, \textit{PigAirwayPressure}, appears to be due to the way the bias values are sampled. We also evaluated a variant of {\textsc{Rocket}} which uses the same method of sampling bias values as {\textsc{MiniRocket}}. Overall, this variant is slightly more accurate than default {\textsc{Rocket}}, but the difference is relatively minor, with a win/draw/loss of 50/6/53 against {\textsc{MiniRocket}}.
\paragraph{Deterministic variant.}
\begin{figure}
\centering
\includegraphics[width=0.65\linewidth]{./fig05.pdf}
\caption{Pairwise accuracy of default {\textsc{MiniRocket}} versus the deterministic variant.}
\Description[Apricot is more accurate than the deterministic variant of Apricot on 42 of 109 datasets]{Scatter plot showing the accuracy of Apricot against the accuracy of the deterministic variant of Apricot for 109 datasets from the UCR archive. Apricot is more accurate on 42 datasets, as accurate on 11 datasets, and less accurate on 56 datasets. The accuracy of Apricot and the deterministic variant is extremely similar on all datasets, with only very minor, almost imperceptible, differences.}
\label{fig-pairwise-minirocket-vs-minirocket-dv}
\end{figure}
Figure \ref{fig-pairwise-minirocket-vs-minirocket-dv} shows the pairwise accuracy of default {\textsc{MiniRocket}} vs the deterministic variant (or {\textsc{MiniRocket}}$_{\text{DV}}$) for the same 109 datasets. Overall, the deterministic variant produces essentially the same accuracy as the default variant.
\subsection{Scalability} \label{subsec-scalability}
\subsubsection{Training Set Size}
\begin{figure*}
\centering
{\includegraphics[width=0.70\linewidth]{./fig06a.pdf}\phantomsubcaption}
\hfill
{\includegraphics[width=0.230\linewidth]{./fig06b.pdf}\phantomsubcaption}
\caption{Training time versus (left) training set size and (right) time series length.}
\Description[Apricot is much faster than Rocket on large datasets in terms of both training set size and time series length]{Three line plots showing training set size against total training time (transform plus classifier training), and one line plot showing time series length against total training time, for both Apricot and Rocket. In terms of training set size, Apricot is 66 times faster for the FruitFlies dataset, 43 times faster for the InsectSound dataset, and 75 times faster for the MosquitoSound dataset. In terms of time series length, Apricot is 19 times faster for the DucksAndGeese dataset.}
\label{fig-scalability}
\end{figure*}
\begin{table}
\centering
\caption{Accuracy and total training time.}
\begin{tabular}{cccrr}
\toprule
& \multicolumn{2}{c}{Accuracy} & \multicolumn{2}{c}{Training Time} \\
\cmidrule(lr){2-3} \cmidrule(lr){4-5}
& {\textsc{Rocket}} & {\textsc{MiniRocket}} & {\textsc{Rocket}} & {\textsc{MiniRocket}} \\
\midrule
\textit{Fruit} & 0.9491 & 0.9568 & $2\text{h }36\text{m }40\text{s }$ & $2\text{m }22\text{s }$ \\
\textit{Insect} & 0.7796 & 0.7639 & $26\text{m }44\text{s }$ & $37\text{s }$ \\
\textit{Mosquito} & 0.8271 & 0.8165 & $15\text{h }34\text{m }58\text{s }$ & $12\text{m }32\text{s }$ \\
\bottomrule
\end{tabular}
\label{table-scalability-training-set-size}
\end{table}
We demonstrate the speed and scalability of {\textsc{MiniRocket}} in terms of training set size on the three largest datasets in the UCR archive, namely, \textit{MosquitoSound} ($139{,}780$ training examples, each of length $3{,}750$), \textit{InsectSound} ($25{,}000$ training examples, each of length $600$), and \textit{FruitFlies} ($17{,}259$ training examples, each of length $5{,}000$). These recent additions are significantly larger than other datasets in the archive.
For this purpose, following \citep{dempster_etal_2020}, we integrate {\textsc{MiniRocket}} (and {\textsc{Rocket}}) with logistic regression, trained using Adam. Training details are provided in Appendix \ref{sec-appendix-training-details}. The experiments were performed on the same cluster as noted in Section \ref{sec-introduction} and, again, both {\textsc{Rocket}} and {\textsc{MiniRocket}} are restricted to a single CPU core.
Figure \ref{fig-scalability} shows training time vs training set size for {\textsc{MiniRocket}} and {\textsc{Rocket}}. Training time includes the transform for both validation and training sets, and classifier training. Table \ref{table-scalability-training-set-size} shows test accuracy and total training time (for the full training set).
{\textsc{MiniRocket}} is slightly more accurate on one of the datasets, and slightly less accurate on two of the datasets. This is consistent with the small differences in accuracy observed on the other datasets in the UCR archive: see Section \ref{subsec-ucr-archive}. However, {\textsc{MiniRocket}} is considerably faster than {\textsc{Rocket}}: $43$ times faster on \textit{InsectSound}, $66$ times faster on \textit{FruitFlies}, and $75$ times faster on \textit{MosquitoSound}.
The accuracy of {\textsc{Rocket}} and {\textsc{MiniRocket}} on the \textit{InsectSound} and \textit{MosquitoSound} datasets appears to be broadly comparable to reported results for other methods for these datasets or versions of these datasets \citep{chen_etal_2014,zhang_etal_2017,fanioudakis_etal_2018,flynn_and_bagnall_2019}, although some deep learning approaches are significantly more accurate on \textit{MosquitoSound} \citep{fanioudakis_etal_2018}.
\subsubsection{Time Series Length}
\begin{sloppypar}
We demonstrate the scalability of {\textsc{MiniRocket}} in terms of time series length on the dataset in the UCR archive with the longest time series, \textit{DucksAndGeese} ($50$ training examples, each of length $236{,}784$). This recent addition has significantly longer time series than other datasets in the archive.
\end{sloppypar}
Figure \ref{fig-scalability} shows training time versus time series length for both {\textsc{Rocket}} and {\textsc{MiniRocket}}. Training time includes the transform and classifier training. (With only 50 training examples, we use the ridge regression classifier.)
While both {\textsc{Rocket}} and {\textsc{MiniRocket}} are linear in time series length, {\textsc{MiniRocket}} is considerably faster for a given length. With more training examples, we would expect the difference in training time to be considerably larger. With only 50 training examples, the overhead of sampling bias values (which is unrelated to training set size) constitutes a significant proportion of the total training time for {\textsc{MiniRocket}}.
\subsection{Sensitivity Analysis} \label{subsec-sensitivity-analysis}
We explore the effect of key parameter choices on accuracy:
\begin{itemize}
\item kernel length;
\item sampling bias from the convolution output versus $\mathcal{U}(-1, 1)$;
\item using only PPV versus both PPV and global max pooling;
\item the number of features; and
\item limiting the maximum number of dilations per kernel.
\end{itemize}
We perform the analysis using the 40 `development' datasets (default training/test splits). Results are mean results over 10 runs.
\subsubsection{Kernels} \label{subsubsec-sensitivity-kernels}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig07.pdf}
\caption{Mean rank for different kernel lengths.}
\Description[The subset of kernels of length 9 having three weights of one value ranks just behind the full set of length 9]{The subset of kernels of length 9 having three weights of one value and six weights of another value ranks just behind the full set of kernels of length 9, and ahead of any other subset of length 9 and kernels of length 7 or 11. The pairwise differences are (with one exception) not statistically significant.}
\label{fig-sensitivity-kal}
\end{figure}
Figure \ref{fig-sensitivity-kal} shows the effect of kernel length on accuracy. For kernels of length 9, a subscript refers to a particular subset of kernels in the sense discussed in Section \ref{subsubsec-weights}. (E.g., $9_{\{3\}}$ refers to kernels with three weights of one value, and six weights of another value.) The total number of features is kept constant (to the nearest multiple of the number of kernels less than $10{,}000$: see Section \ref{subsubsec-features}), such that more features are computed per kernel for smaller sets of kernels and vice versa.
Kernels of length 9 are most accurate, but kernels of length 7 or 11 are not significantly less accurate. This is consistent with the findings in \citep{dempster_etal_2020} in relation to {\textsc{Rocket}}. The actual differences in accuracy between kernels of different lengths is very small.
Crucially, however, as noted in Section \ref{subsubsec-weights}, the $9_{\{3\}}$ subset is nearly as accurate as the full set of kernels of length 9. This is a relatively small subset of kernels, and is particularly well suited to the optimisations pursued in Section \ref{subsec-optimising-the-transform}.
\subsubsection{Bias} \label{subsubsec-sensitivity-bias}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig08.pdf}
\caption{Mean rank for bias sampled from the convolution output versus bias sampled from $\mathcal{U}(-1, 1)$.}
\Description[Bias sampled from the convolution output ranks well ahead of bias sampled uniformly]{Bias sampled from the convolution output ranks well ahead of bias sampled uniformly from negative 1 to 1, and the difference is statistically significant.}
\label{fig-sensitivity-bias}
\end{figure}
Figure \ref{fig-sensitivity-bias} shows the effect in terms of accuracy of sampling bias from the convolution output versus from $\mathcal{U}(-1, 1)$ as in {\textsc{Rocket}}. {\textsc{MiniRocket}} is significantly less accurate when sampling bias from $\mathcal{U}(-1,1)$. The change to sampling bias from the convolution output is critical to matching the accuracy of {\textsc{Rocket}}.
\subsubsection{Features} \label{subsubsec-sensitivity-features}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig09.pdf}
\caption{Mean rank for PPV vs PPV and global max pooling.}
\Description[PPV by itself ranks ahead of the combination of PPV and global max pooling]{PPV by itself ranks ahead of the combination of PPV and global max pooling, but the difference is not statistically significant.}
\label{fig-sensitivity-ppv}
\end{figure}
Figure \ref{fig-sensitivity-ppv} shows the effect of using only PPV versus both PPV and global max pooling. With the other changes to {\textsc{MiniRocket}}---in particular, with the change to sampling bias from the convolution output---there is no advantage to using global max pooling in addition to PPV. In fact, using global max pooling in addition to PPV is less accurate than just using PPV, although the difference is not statistically significant.
\subsubsection{Number of Features} \label{subsubsec-sensitivity-num-features}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig10.pdf}
\caption{Mean rank for different numbers of features.}
\Description[9,996 features, the default, ranks just behind 49,980 features, but ahead of any other number of features]{9,996 features, the default, ranks just behind 49,980 features, but ahead of 99,960 features and any smaller number of features. 9,996 features is in the same clique as 4,956, 99,960, and 49,980 features, that is, the pairwise differences are not statistically significant.}
\label{fig-sensitivity-num-features}
\end{figure}
Figure \ref{fig-sensitivity-num-features} shows the effect of different numbers of features between $84$ and $99{,}960$ (the nearest multiple of 84 less than 100, 500, $1{,}000$, ...). Increasing the number of features noticeably increases accuracy up to approximately $10{,}000$ features. There is little or no benefit to increasing the number of features beyond $10{,}000$, at least for shorter time series, because there is little benefit in computing PPV for many more than $l_{\text{input}}$ bias values for time series of length $l_{\text{input}}$ (more and more features will be the same). For $49{,}980$ and $99{,}960$ features, we have endeavoured to avoid this limitation as much as possible by setting the maximum number of dilations per kernel to 119 (see Section \ref{subsubsec-dilation}) and, where necessary, sampling bias values from multiple training examples.
\subsubsection{Dilation} \label{subsubsec-sensitivity-dilation}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{./fig11.pdf}
\caption{Mean rank of different values for the maximum number of dilations per kernel.}
\Description[A maximum of 32 dilations per kernel, the default, ranks ahead of all other values]{A maximum of 32 dilations per kernel, the default, ranks ahead of all other values for the maximum number of dilations per kernel. The pairwise differences are (with one exception) not statistically significant.}
\label{fig-sensitivity-dilation}
\end{figure}
Figure \ref{fig-sensitivity-dilation} shows the effect in terms of accuracy of different values for the maximum number of dilations per kernel. The total number of features is kept constant, such that more features are computed per dilation for a smaller number of maximum dilations per kernel and vice versa: see Section \ref{subsubsec-dilation}. There is little difference in accuracy between values of 16 and 119 (119 being the largest possible number of dilations per kernel for the default number of features, i.e., $\lfloor 10{,}000 / {84} \rfloor = 119$). A value of 32 balances accuracy with the computational advantage of limiting the number of dilations per kernel, as discussed in Section \ref{subsubsec-dilation}.
\section{Conclusion}
We reformulate {\textsc{Rocket}} into a new method, {\textsc{MiniRocket}}, making it up to $75$ times faster on larger datasets. {\textsc{MiniRocket}} shows that it is possible to achieve essentially the same accuracy as {\textsc{Rocket}} using a mostly-deterministic and much faster procedure.
{\textsc{MiniRocket}} represents a significant advance in accuracy relative to computational cost. {\textsc{MiniRocket}} is much faster than any other method of comparable accuracy (including {\textsc{Rocket}}), and far more accurate than any other method of even roughly-similar computational expense. Accordingly, we suggest that {\textsc{MiniRocket}} should be considered and used as the default variant of {\textsc{Rocket}}. We provide a na{\"i}ve facility for applying {\textsc{MiniRocket}} to multivariate time series (available through the accompanying repository). In future work, we propose to investigate more sophisticated approaches to multivariate time series, to explore the integration of {\textsc{MiniRocket}} with nonlinear classifiers, and the use of {\textsc{MiniRocket}} beyond time series data.
\begin{acks}
This material is based on work supported by an Australian Government Research Training Program Scholarship, and the Australian Research Council under award DP190100017. The authors would like to thank Professor Eamonn Keogh and all the people who have contributed to the UCR time series classification archive. Figures showing mean ranks were produced using code from \citep{ismailfawaz_etal_2019}.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
| 65d2824f5406e9e8cd3c3daf5a62881f16a4ec3f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Criminal activities have become a major social problem due to their adverse effect on human life, economy and safety. The availability of crime data in recent years has enabled researchers to develop models for crime prediction. The government and responsible authorities can take preventive measures if they know about a crime event in advance. Knowing the insight behind the prediction of a crime occurrence would allow them to plan preventive measures appropriately and keep the society safe from the happening of the crime. Interpretable predictions ensure the transparency and accountability of the model. Thus, both accuracy and interpretability are two essential and desired properties for a crime prediction model. We propose an \emph{\underline{A}ttention-based \underline{I}nterpretable \underline{S}patio \underline{T}emporal Network (AIST)}, an interpretable deep learning model for crime prediction.
\begin{figure}[htbp]
\begin{subfigure}{0.32\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/map.pdf_tex}
\caption{Chicago Communities}
\label{Fig:chicago_map}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/sp_cor.pdf_tex}
\caption{Spatial correlation}
\label{Fig:sp_cor}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/tem_cor.pdf_tex}
\caption{Temporal correlation}
\label{Fig:tem_cor}
\end{subfigure}
\caption{Spatio-temporal dependencies of crime distribution}
\label{Fig:map_sp_tem_cor}
\vspace{-1em}
\end{figure}
Crime events exhibit spatial and temporal correlations and external features (e.g., taxi flow) often have influence on the crime occurrence.
\textbf{Spatial Correlation.} Spatially nearby regions show a similar crime distribution and the extent of this similarity varies across regions and time. Figure~\ref{Fig:chicago_map} shows the communities (i.e., regions) of Chicago and Figure~\ref{Fig:sp_cor} shows an example on January, 2019 Chicago crime data. Regions 8 and 32 show strong spatial correlation while Regions 8 and 7 do not, though both of them are spatially nearby. Also, the spatial correlation between Regions 8 and 32 changes with time.
\textbf{Temporal Correlation.} Crime occurrences of a region show both short and long term temporal correlations and these correlations vary with crime categories. Fig~\ref{Fig:tem_cor} shows an example for Region 8: deceptive practice (C0) and theft (C1) peak during mid night, whereas robbery (C5) peak during late night or early morning. There is also a significant difference of crime distributions across different crime categories: deceptive practice (C0) and theft (C1) occur at regular intervals whereas robbery is not so common for Region $8$. Besides, the crime distributions of the same category differ throughout the week.
\vspace{-1em}
\begin{figure}[htbp]
\begin{subfigure}{0.32\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/feat_cor1.pdf_tex}
\caption{}
\label{Fig:ext_cor_1}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/feat_cor2.pdf_tex}
\caption{}
\label{Fig:ext_cor_2}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/feat_cor3.pdf_tex}
\caption{}
\label{Fig:ext_cor_3}
\end{subfigure}
\caption{Influence of Taxi Flows on the crime distribution in Chicago Communities}
\label{Fig:ext_cor}
\vspace{-1em}
\end{figure}
\textbf{External Features.} Functionalities and urban characteristics of a region like points of interests (POIs), traffic flow have direct influence on its crime occurrences. The influence of these external features on the crime occurrences tend to vary from time to time and region to region. Figure~\ref{Fig:ext_cor} shows such an example in Region $8$, where the distribution of deceptive practice (C0) (Figure~\ref{Fig:ext_cor_1}) and theft (C1) (Figure~\ref{Fig:ext_cor_2}) have a strong correlation with taxi flows, whereas robbery (C5) (Figure~\ref{Fig:ext_cor_3}) comparatively shows a weaker correlation with taxi flows.
Modeling these diverse spatio-temporal correlations and learning meaningful external features and their probable influence on crime are challenging tasks. Traditional interpretable machine learning and data mining methods~\cite{DBLP:conf/kdd/WangKGL16, 4666600, DBLP:conf/pakdd/Yu0CM14} cannot model these non-linear spatio-temporal correlations and thus fail to predict the crime occurrences accurately. Recent deep learning models~\cite{DBLP:conf/cikm/HuangZZC18, DBLP:conf/www/HuangZZWCY19} capture this non-linear spatial and temporal dependencies to some extent and improve the accuracy of traditional models. They still have major limitations:
\begin{compactitem}
\item The models only learn static spatial correlations. However, the correlations for two regions may vary with time.
\item The models do not address long term (e.g. daily, weekly) temporal correlations.
\item The models do not consider the external features and hence the learned region embedding is incomplete.
\item The models lack interpretability. Both these models use LSTM based attention weights which are difficult to interpret because of the recurrence on the hidden states generated by LSTMs~\cite{DBLP:conf/nips/ChoiBSKSS16}. They are also not sparse enough to be meaningful for long sequence.
\end{compactitem}
To overcome the limitations, we develop AIST that captures dynamic spatio-temporal correlation for crime prediction and provides quantitative insights based on external features behind a prediction. We develop two novel variants of graph attention networks (GAT)~\cite{DBLP:conf/iclr/VelickovicCCRLB18}, $hGAT$ and $fGAT$ to learn the crime and feature embedding of the nodes (regions), respectively at each time step. These embedding are then fed to three sparse attention based-LSTMs (SAB-LSTMs)~\cite{DBLP:conf/nips/KeGBBMPB18} for modeling recent, daily and weekly crime trends. Finally, AIST applies a location-based attention mechanism to identify the significance of different trends to make a prediction.
GAT does not consider the hierarchical information of nodes. However, in real-world scenarios nodes tend to form clusters and belong to different hierarchies based on similar characteristics. In urban context, nodes (regions) that belong to a same hierarchy shares similar functionalities and crime distributions. We propose $hGAT$ that incorporates this prior knowledge of hierarchical information into GAT's architecture to produce a better crime embedding of nodes.
Concatenating the feature vectors with spatial~\cite{DBLP:conf/aaai/Yao0KTJLGYL18, DBLP:conf/aaai/YaoTWZL19} or temporal view~\cite{DBLP:conf/aaai/LiZKXZ19} either directly or after a linear transformation is a common practice for incorporating the external features into the model. However, it fails to fully utilize the features and generate insights for a model's prediction. We propose $fGAT$ that replaces the additive self-attention mechanism of GAT with a novel scaled dot product self-attention mechanism~\cite{DBLP:conf/nips/VaswaniSPUJGKP17} to learn crime and region specific relevant feature embedding.
The unique challenge of a crime prediction problem that does not apply to other spatio-temporal prediction problems (e.g., traffic flow prediction, crowd flow prediction, passenger demand prediction) is the fact that crime data is spatially, temporally and categorically extremely sparse. AIST utilizes the feature embedding learned from fGAT to tackle the sparseness of crime data. On top of that, it is also necessary to keep the crime prediction model's architecture reasonably interpretable, which makes the tasks even harder than building a spatio-temporal model that does not consider interpretability~\cite{10.1038/s42256-019-0048-x}.
AIST is interpretable because it takes transparent decisions at each prediction step based on the different attention modules used in the model architecture. To explain a prediction, we first find whether the prediction is based on recent occurrences or any recurring trend and then identify the previous time steps that are given the most importance. Our model knows why a time step is given importance as the input at each time step is an interpretable spatial embedding. Hence, if we backtrack we can find the most important regions and features for a specific time step. Even though attention as a form of explanation is not new in spatio-temporal literature~\cite{DBLP:conf/cikm/HuangZZC18, DBLP:conf/www/HuangZZWCY19, DBLP:conf/aaai/GuoLFSW19, DBLP:conf/cikm/ZhangHXX20}, simply using an attention module does not make a model interpretable~\cite{DBLP:conf/acl/SerranoS19,DBLP:conf/naacl/JainW19}. Keeping this in mind, unlike existing spatio-temporal literature the model architecture of AIST is designed so that it complies with the conditions presented in~\cite{DBLP:conf/emnlp/WiegreffeP19} under which attentions can be regarded as faithful explanations.
Besides inherent interpretable architectures~\cite{DBLP:conf/nips/ChenLTBRS19}, recently post-hoc local explanation techniques~\cite{DBLP:conf/nips/LundbergL17} that provide approximate explanations to a model's decision making have been explored to imitate the behavior of deep learning black box models. However, they are not well received considering the fact that if these explanations had been adequate enough, there would be no need for the original model~\cite{10.1038/s42256-019-0048-x}. Hence, we keep AIST architecture inherently interpretable while ensuring its accuracy.
In summary, the contributions of this paper are as follows:
\begin{compactitem}
\item We propose a novel interpretable spatio-temporal deep learning model, AIST which is able to capture diverse spatio-temporal correlations based on past crime occurrences, external features and recurring trends.
\item We propose $hGAT$, a novel GAT variant that allows AIST to learn more faithful node embedding.
\item We propose $fGAT$, another novel GAT variant that provide insights behind the predictions of AIST.
\item We conduct experiments on Chicago crime data. AIST achieves a higher accuracy than the state-of-the-art methods and provides useful insights for its predictions. Experiment results also validate that the explanations provided by different attention modules in hGAT, fGAT and SAB-LSTMs are faithful.
\end{compactitem}
The remaining of the paper is organized as follows. We discuss the related work in Section~\ref{Sec:relatedWork} and formulate the crime prediction problem in Section~\ref{Sec:problemFormulation}. We present our model, $AIST$ in Section~\ref{sec:model}. Section~\ref{Sec:experimentalResult} presents the experimental results and evaluates the accuracy and the interpretability of AIST. Section~\ref{sec:conclusion} concludes the paper.
\section{Related Work}
\label{Sec:relatedWork}
Data-driven crime prediction problems have received wide attention from the researchers for decades. Existing studies on crime prediction can be divided into following categories: (i) \emph{crime rate inference} that predicts the crime rate of a region, (ii) \emph{crime hotspot detection} that finds the locations where crimes are clustered, and (iii) \emph{crime occurrence prediction} that forecasts the occurrence of a crime category for a location at a future timestamp. Our work falls in the third category. In Sections~\ref{sec:crimeModels} and~\ref{sec:interpretableModels}, we elaborate existing crime prediction models and interpretable models, respectively. In Section~\ref{spatio-temporal}, we discuss the deep learning methods used for spatial-temporal prediction.
\subsection{Crime Prediction Models}
\label{sec:crimeModels}
\textbf{Statistical and Classic Machine Learning Methods.} Recent studies~\cite{DBLP:conf/kdd/WangKGL16, DBLP:journals/tbd/WangYKGL19, DBLP:journals/corr/abs-1908-02570} used statistical and classic machine learning methods (e.g., linear regression, negative binomial regression, geographically weighted regression, random forest) for crime rate inference problem. In~\cite{DBLP:conf/kdd/WangKGL16, DBLP:journals/tbd/WangYKGL19}, the authors studied the effect of point of interest (POI) (e.g., a restaurant or a shopping mall) and taxi flow information along with the traditional demographics features of a region while in~\cite{DBLP:journals/corr/abs-1908-02570}, the authors utilized FourSquare check-in data for estimating the crime rate of a particular region. Researchers have also employed kernel density estimation (KDE)~\cite{article, DBLP:journals/Hart, eck2005mapping, DBLP:conf/sibgrapi/NetoSV16} for predicting hot-spot maps. However, these works only take spatial features and dependencies into account ignoring the temporal dynamics of crime.
To address the temporal dynamics, time-series models such as autoregressive integrated moving average (ARIMA)~\cite{4666600} have been proposed for one-week ahead crime occurrence prediction. In~\cite{doi:10.1198/jasa.2011.ap09546}, the authors implemented a self-exciting point process similar to one used by the seismologists in the context of urban crime to understand the temporal trends of burglary. Even though these models acknowledge the temporal dynamics, they do not incorporate the spatial context of crimes.
Both spatial and temporal information have been also explicitly modeled in the literature. In~\cite{DBLP:conf/pakdd/Yu0CM14}, the authors proposed an algorithm that constructs a global crime pattern from local crime cluster distributions, and employed it for predicting residential burglary. In~\cite{DBLP:journals/tgis/NakayaY10}, the authors employed STKDE, a variant of KDE for mapping transient and stable crime clusters. The work in~\cite{DBLP:journals/tist/TooleEP11} used analytic and statistical techniques to identify the spatio-temporal crime patterns. In~\cite{DBLP:conf/cikm/ZhaoT17}, spatio-temporal correlations like intra-region temporal correlation and inter-region spatial correlation have been considered for crime occurrence prediction. However, all of these methods cannot fully model the complex non-linear relation of space and time and the dynamicity of spatial-temporal correlation.
Besides spatio-temporal features, incorporating additional data (e.g. Twitter, demographics data) improve the accuracy of existing crime prediction models. The authors in ~\cite{DBLP:journals/dss/Gerber14} added Twitter-based features extracted from topic based modeling for improving the prediction of models. In~\cite{10.1145/1938606.1938608}, the authors used fuzzy association rule mining to find consistent crime patterns using population demographics information of communities. Another line of work~\cite{DBLP:conf/gis/XiongSKDPS19, DBLP:conf/www/WangJWWL19} explores the heterogeneous and task-specific division of spatial regions over traditional grid and community based division which helps improve the accuracy of the crime prediction.
\textbf{Deep Learning Methods.} Deep learning models have recently been shown to be very effective in domains like computer vision, speech recognition and natural language processing. Recent deep learning models have also attempted to capture the non-linear spatio-temporal dependencies of crime. DeepCrime~\cite{DBLP:conf/cikm/HuangZZC18}, a hierarchical recurrent framework with attention mechanism, considers temporal correlation, its inter-relation with ubiquitous data and category dependencies for future crime prediction. However, DeepCrime does not consider spatial correlations of crimes. In~\cite{DBLP:journals/corr/WangZZBB17}, the authors applied ST-ResNet architecture~\cite{DBLP:conf/aaai/ZhangZQ17} for crime intensity prediction while in~\cite{DBLP:conf/www/HuangZZWCY19}, the authors developed MiST, a LSTM based neural network architecture with attention mechanism to model spatio-temporal and cross-categorical correlation for crime prediction. None of these models can capture dynamic spatial correlation and identify the impact of external features on crime predictions. Besides, these models are not interpretable. DeepCrime and MiST employ attention based RNNs which lack interpretability because of the recurrence on the hidden states generated by RNNs and their non-sparse attention weights for longer sequences. ST-ResNet uses deep residual units with hundreds and thousands of CNNs stacked altogether which makes it harder to interpret the model's prediction.
\subsection{Interpretable Models}
\label{sec:interpretableModels}
The statistical and classic machine learning models have an advantage over deep learning models in terms of interpretability. However, they cannot model the complex non-linearity of space and time and thus lacks accuracy. On the other hand, though neural networks can capture the spatial-temporal non-linear relationship, they are not interpretable.
Attention-based models focus on the most relevant information while performing a certain task. These models have become very popular in image processing~\cite{DBLP:conf/icml/XuBKCCSZB15, DBLP:conf/nips/MnihHGK14, DBLP:journals/corr/BaMK14, DBLP:conf/cvpr/FuZM17} and natural language processing~\cite{DBLP:journals/corr/BahdanauCB14, DBLP:conf/nips/VaswaniSPUJGKP17}, and health-care predictions~\cite{DBLP:conf/nips/ChoiBSKSS16, DBLP:conf/kdd/ChoiBSSS17, DBLP:conf/kdd/BaiZEV18, DBLP:conf/kdd/MaCZYSG17} for ensuring interpretability. Similar to the statistical and classic machine learning models, despite of having a self-explanatory structure, simply using an attention based model does not make an explanation of a prediction faithful. The model architecture of AIST provides faithful explanations, which is also validated by our experiment results.
The other category of interpretable models is post-hoc models, where a separate model is used for explanation. Examples of post-hoc models include LIME~\cite{DBLP:conf/kdd/Ribeiro0G16}, SHAP~\cite{DBLP:conf/nips/LundbergL17}, rule-based learning~\cite{DBLP:journals/corr/SuWVM16} and saliency visualizations~\cite{DBLP:conf/nips/DabkowskiG17}.
\subsection{Deep Learning for Spatio-temporal Prediction}
\label{spatio-temporal}
Deep learning methods have become popular in recent years in the domain of spatial temporal prediction. A common approach is to use the convolution based architecture (CNN)~\cite{DBLP:conf/gis/ZhangZQLY16, DBLP:journals/tits/ChenYL18, DBLP:conf/icdm/ChenLTCZYVFZ18} for finding the spatial correlation and the recurrent based architecture~\cite{DBLP:conf/kdd/RongXYM18, DBLP:journals/corr/abs-1801-02143} for finding the temporal correlation. In~\cite{DBLP:conf/aaai/YaoTWZL19}, the authors used both CNN and attention-based LSTM to capture the dynamic spatio-temporal dependencies for traffic prediction. In~\cite{DBLP:conf/ijcai/LiangKZYZ18}, the authors proposed a multi-level attention mechanism along with a recurrent layer and a fusion module to incorporate the external features for geo-sensory time prediction.
Recent literature~\cite{DBLP:conf/aaai/GuoLFSW19, DBLP:conf/ijcai/YuYZ18, DBLP:journals/TKDE/9139357, DBLP:conf/cikm/ZhangHXX20, DBLP:conf/cikm/XieG0X0Z20, DBLP:conf/www/Wang0WJWTJY20, DBLP:conf/kdd/HongLYLFWQY20} has also started exploring the graph neural networks for such prediction. In~\cite{DBLP:conf/aaai/GuoLFSW19, DBLP:conf/ijcai/YuYZ18, DBLP:conf/kdd/HongLYLFWQY20}, the authors proposed a pure convolutional structure in the form of a graph convolution in the spatial dimension and a general convolution in the temporal dimension to model the traffic flows. In~\cite{DBLP:journals/TKDE/9139357}, several temporal views are fed to their respective graph convolution layers and then fused altogether along with semantic views to model the crowd flows. In~\cite{DBLP:conf/www/Wang0WJWTJY20}, the authors modeled traffic flows with a graph convolutional network (GCN) followed by a recurrent layer and a transformer to capture the local and global temporal correlation, respectively. Both these models~\cite{DBLP:journals/TKDE/9139357, DBLP:conf/www/Wang0WJWTJY20} incorporate geospatial position of nodes into the GCN to better model the spatial dependencies. In~\cite{DBLP:conf/cikm/ZhangHXX20}, the authors proposed STCGA that combines multiple self-attention, graph attention, and convolutional residual networks to predict the traffic flow.~\cite{DBLP:conf/cikm/XieG0X0Z20} proposes DIGC and a pre-trained binary classifier, both of which consists of a GCN followed by an LSTM to extract the spatio-temporal and latent incident crime features, respectively for traffic speed prediction.
None of these spatio-temporal prediction models are designed to handle the sparseness of crime data. The finer the spatial, temporal or categorical resolution gets, the sparser the crime data becomes; which makes it even harder to model the crime. Hence, the spatio-temporal prediction models fail to perform well for crime prediction tasks. Unlike existing spatio-temporal literature, AIST chooses a handful of region and crime category specific external features, and applies fGAT to learn a more stable and faithful feature embedding of the target region as a substitute of the sparse crime data. This learned feature embedding along with the crime embedding learned by hGAT allows AIST to capture the slightest of changes in the feature or crime embedding of the target region over time and make predictions accordingly. Our experiment results also show that AIST outperforms the high-performance spatio-temporal prediction models.
\begin{small}
\begin{table}[htbp]
\caption{Notations and their meanings}
\centering
\begin{tabular}{|c|p{9cm}|
\hline \hline
Notation & Symbol \\
\hline
$N, T, K, J$ & Number of regions, time steps, crime categories, external features\\
\hline
$\mathcal{N}_i$ & First-order neighbors of region $r_i$ (including itself)\\
\hline
$\tau$ & Length of a time step\\
\hline
$x_{i, t}^k$ & Crime occurrences of $k$-th category at region $r_i$ during $t$-th time step\\
\hline
$\textbf{x}_{i, t}$ & Crime occurrences of all categories at region $r_i$ during $t$-th time step\\
\hline
$\textbf{X}_{t}$ & Crime occurrences of all categories at all regions during $t$-th time step\\
\hline
$f_{i, t}^j$ & $j$-th external feature of region $r_i$ during $t$-th time interval\\
\hline
$\textbf{f}_{i, t}$ & All external features of region $r_i$ during $t$-th time step\\
\hline
$\textbf{F}_{t}$ & All external features of all regions during $t$-th time step\\
\hline
$\hat{\textbf{Y}}_{T+1}$ & Predicted crime occurrences of all regions and categories of the city at $(T+1)$-th time step\\
\hline
\end{tabular}
\label{symbols}
\end{table}
\end{small}
\section{Problem Formulation}
\label{Sec:problemFormulation}
In this section, we introduce some notations\footnote{Bold letters, e.g. \textbf{$A, a$} denote matrices and vectors respectively and small letters, e.g. $a$ denote scalars} and formulate crime prediction problem as a regression task. Table~\ref{symbols} summarizes the notations used in the paper.
\textbf{Region.} We model a city with an undirected graph $G = (V, E)$, where $V$ represents a set of $N$ regions $\{r_1, r_2, r_3,\ldots, r_N\}$ and $E$ represents a set of edges connecting them. In this study, a region denotes a community area: a pre-defined administrative boundary that serves various planning and statistical purposes. For a region $r_{i}$ to be connected to region $r_{i'}$ they must share a common boundary.
\textbf{Crime Occurrence.} Let $x_{i, t}^k \in \mathbb{R}$ represent the number of crimes reported of category $k$ (e.g. theft) at region $r_i$ during $t$-th time step\footnote{We use time step and time interval synonymously.}. If $\mathbf{x}_{i, t} = [x_{i, t}^1, x_{i, t}^2, x_{i, t}^3, \ldots, x_{i, t}^K] \in \mathbb{R}^{K}$ denotes the reported crimes of all $K$ categories at region $r_i$ during $t$-th time step and $\mathbf{X}_{t} = (\mathbf{x}_{1, t}, \mathbf{x}_{2, t}, \mathbf{x}_{3, t}, \ldots, \mathbf{x}_{N, t}) \in \mathbb{R}^{K\times N}$ denotes the reported crimes of all categories at all $N$ regions during $t$-th time step, then the crime occurrences of the whole city for $T$ time steps can be denoted as $\mathcal{X} = (\mathbf{X}_1, \mathbf{X}_2, \ldots, \mathbf{X}_T) \in \mathbb{R}^{K\times N\times T}$.
\textbf{External Feature.} We use POI information, traffic inflow and traffic outflow as external features for improving the model's prediction accuracy. The external features of the city during $T$ time steps are denoted as $\mathcal{F} = (\mathbf{F}_1, \mathbf{F}_2, \ldots, \mathbf{F}_T) \in \mathbb{R}^{J\times N\times T}$ for $J$ external features, where $\mathbf{F}_{t} = (\mathbf{f}_{1, t}, \mathbf{f}_{2, t}, \mathbf{f}_{3, t}, \ldots, \mathbf{f}_{N, t}) \in \mathbb{R}^{J\times N}$ denotes the external features of the city during $t$-th time step, $\mathbf{f}_{i, t} = [f_{i, t}^1, f_{i, t}^2, f_{i, t}^3, \ldots, f_{i, t}^J] \in \mathbb{R}^{J}$ denotes the external features of a region $r_i$ during time step $t$ and $f_{i, t}^j \in \mathbb{R}$ denotes the $j$-th external feature of a region $r_i$ during time step $t$.
\textbf{Problem Definition.} Given past crime occurrences $\mathcal{X}$ and external features $\mathcal{F}$ for last $T$ time steps, predict $\hat{\textbf{Y}}_{T+1}$, the crime occurrences of the city during $(T+1)$-th time step.
\begin{figure*}[tbp]
\centering
\def0.9\textwidth{\textwidth}
\input{Figures/Model_Overview.pdf_tex}
\caption{Model Overview of AIST: (a) hGAT is applied to calculate the crime embedding $c_1$ of target region $r_1$, (b) fGAT is applied to calculate the feature embedding $e_1$ of $r_1$, (c) both $c_1$ and $e_1$ are concatenated to produce spatial representation $s_{1}$ of $r_1$ at time step $t$, (d) the spatial representations generated at different time steps are then fed to three SAB-LSTMs to capture recent, daily and weekly crime trends at $r_1$ and a location-based attention is applied to predict the crime occurrence of $r_1$ at $(T+1)$-th time step, $\hat{y}_{T+1}$.}
\label{Fig:overview}
\vspace{0em}
\end{figure*}
\section{Model Description}
\label{sec:model}
The key idea behind our model's high prediction accuracy is that we exploit (i) hierarchical information of regions, (ii) external features, and (iii) short, long term crime patterns to capture the dynamic spatio-temporal dependencies while keeping the model's architecture reasonably interpretable. Given the crime occurrences of category $k$ at region $r_i$ during time steps $[1..T]$, we find the crime embedding $\mathbf{c_{i, t}^k}$ and feature embedding $\mathbf{e_{i, t}^k}$ of $r_i$ using our proposed hGAT and fGAT, respectively, and concatenate them to produce spatial representation $\mathbf{s_{i, t}^k}$ for each of these time step. These embedding are then fed to three SAB-LSTMs for capturing recent, daily and weekly trends which outputs the hidden states $\mathbf{h_{T+1}^r}, \mathbf{h_{T+1}^d}, \mathbf{h_{T+1}^w}$, respectively. After applying an attention mechanism on these hidden states, a context vector, $\mathbf{c_{T+1}}$ is generated to predict the crime occurrence at $(T+1)$-th time step, $\hat{y}_{i, T+1}^k$. Figure~\ref{Fig:overview} gives an overview of the model. In Section~\ref{spatial}, we elaborate on how we generate crime embedding using hGAT and feature embedding using fGAT to produce spatial representation (Figures~\ref{Fig:overview}a--~\ref{Fig:overview}c). In Sections~\ref{temporal} and~\ref{prediction}, we discuss our crime trend generation and prediction steps, respectively (Figure~\ref{Fig:overview}d).
\subsection{Spatial View}
\label{spatial}
Convolutional neural networks (CNNs)~\cite{DBLP:conf/gis/ZhangZQLY16} and its variants~\cite{DBLP:conf/aaai/YaoTWZL19} have been applied to model spatial correlation between regions in spatio-temporal prediction. Though CNNs learn meaningful features on regular grid structured data, they do not perform well on irregular graph data because the number of nodes in a graph and their neighbor counts are variables. Urban crime data exhibit a clear graph structure considering the correlation between regions and other external features. Modeling them as grid structured data results in incomplete information and makes it hard to learn meaningful information. To address this issue, graph convolutional networks (GCNs)~\cite{DBLP:conf/iclr/KipfW17} have gained popularity in recent times. GCNs learn a node's embedding as an aggregation of its neighbor's features and calculate their contribution with predefined Laplacian Matrix, which is the difference of the degree matrix and the adjacency matrix of the graph. Since the contributions of neighbor nodes are static, GCNs can not capture the dynamic spatial correlation between regions.
Based on these observations, we choose GAT as the base architecture to capture the spatial dependencies for crime prediction. Similar to GCN, GAT learns a node's embedding as an aggregation of its neighbor's features but uses a self-attention mechanism to learn their contributions instead. GAT does not require any costly matrix operation and knowledge about the graph structure upfront, which allows GAT to learn dynamic spatial correlation between regions.
We use two GAT variants: hGAT and fGAT to learn the crime and feature embedding of a target region as follows.
\textbf{\emph{Crime Embedding.}}
The city of Chicago is divided into $77$ communities (regions) and the communities are grouped into $9$ districts or sides forming a containment hierarchy. Communities under the same side tend to share similar socio-economic, demographic and urban features, which result in similar crime distribution than those under different sides. Hence, while aggregating the node (community) information in GAT, prioritizing the nodes that fall under the same side with the target node over others may help to learn better crime representation of a target node. From this intuition, we propose $hGAT$ to amplify the signals of the nodes that fall under the same side with the target node than those which do not. Since almost every city can be divided into multiple partitions at different spatial resolutions, this idea can be generalized to other cities as well.
\begin{figure}[tbp]
\begin{subfigure}{3.5 cm}
\def0.9\textwidth{\columnwidth}
\input {Figures/Community_Structure_v2.pdf_tex}
\caption{First-hop neighbors}
\end{subfigure}
\hspace{4 mm }
\begin{subfigure}{4.5 cm}
\def0.9\textwidth{\columnwidth}
\input {Figures/Community_Structure_v1.pdf_tex}
\caption{Hierarchical Structure}
\end{subfigure}
\caption{Complex spatial interaction between regions}
\label{Fig:region_structure}
\vspace{-1em}
\end{figure}
Figure~\ref{Fig:region_structure} represents a scenario where the first-order neighbors of target region $8$ are $\{7, 8, 24, 28, 32\}$ and target region $8$ along with region $32$ and $33$ fall under the same side (represented as grey circles in Figure~\ref{Fig:region_structure}). It is evident from Figure~\ref{Fig:hgat_cor_total} that R8 is strongly correlated to R32 than other nearby regions (R7, R24) in terms of crime distribution and external features (POI and taxi flows). Similarly, R24 shows a stronger correlation with R28 than R8. Thus, it is expected that the target region 8 is influenced more by region $32$ which is not only a first-hop neighbor but also falls under the same side, whereas regions $7, 24$ or $28$ do not. To be clear, we do not consider the influence of regions such as $33$ that falls under the same side with the target region 8, but is not a first-hop neighbor. We only want to amplify the signal from those regions which satisfy both conditions: falls under the same side and is a first-hop neighbor.
\begin{figure}[htbp]
\centering
\def0.9\textwidth{0.75\textwidth}
\input{Figures/hgat_cor.pdf_tex}
\caption{Pearson correlation coefficient among regions of Chicago based on 2019 crime, POI and 2019 taxi flows distribution (left to right)}
\label{Fig:hgat_cor_total}
\vspace{-1em}
\end{figure}
hGAT considers two sets of features for each node ($r_i$): (i) node level features: crime occurrences at community level during time step $t$: $x_{i, t}^k$, (ii) parent level features: crime occurrences at district/side level during time step $t$: $z_{i, t}^k$ as input to GAT and an additional attention layer to capture the similarity between nodes based on parent level features. Let the parent node of a region $r_i$ be $p_j = Parent(r_i)$. Then the parent feature of node $r_i$ is calculated as $z_{i, t}^k = \sum_{\forall r_{n} Parent(r_{n})=p_j}x_{n, t}^k$. Basically, we sum the crime occurrences of category $k$ across the nodes whose parent node is $p_j$ to create the parent feature of target node $r_i$. Traditional GAT considers only node level features. Hence, it can not model hierarchical information into a node's embedding.
For hGAT, we use two transformation matrix, (i) $\mathbf{w_x}\in \mathbb{R}^{F}$ to learn the similarities between a target region and its neighbor's node level features, and (ii) $\mathbf{w_z}\in \mathbb{R}^{F'}$ to learn similarities between their parent-level features. Based on these information two separate feed-forward attention layer computes two sets of pair-wise unnormalized attention scores between the target region and its first-hop neighbors: $e_{ii'}^c$ and $e_{ii'}^p$, respectively. For clarity, we omit the indices of crime categories ($k$) and time step ($t$).
\begin{align*}
e_{ii'}^c &= \text{LeakyReLU}(\mathbf{a_x}^T[\mathbf{w_x}x_{i} \mathbin\Vert \mathbf{w_x}x_{i'}]) \\
e_{ii'}^p &= \text{LeakyReLU}(\mathbf{a_z}^T[\mathbf{w_z}z_{i} \mathbin\Vert \mathbf{w_z}z_{i'}])
\end{align*}
We perform an element-wise addition to combine these two sets of unnormalized attention scores and apply softmax over them to generate final attention scores, where $\mathcal{N}_i$ denotes first order neighbors including itself.
$$e_{ii'} = e_{ii'}^c + e_{ii'}^p$$
$$\alpha_{ii'} = \text{softmax}_{i'}(e_{ii'}) = \frac{exp(e_{ii'})}{\sum_{i'' \in \mathcal{N}_i}exp(e_{ii''})}$$
Finally, we use this combined attention score to update the crime embedding ${c}_{i, t}^k$ of target region $r_i$.
\begin{equation}
\label{eq:1}
\mathbf{{c}}_{i, t}^k = \sigma\left(\sum_{i' \in \mathcal{N}_i}\alpha_{ii'}\mathbf{w_x}x_{i', t}^k\right)
\end{equation}
Figure~\ref{Fig:overview}a gives an overview of generating the crime embedding.
\textbf{\emph{Feature Embedding.}} Besides historical crime observations, external features have been shown to be useful in crime prediction problems~\cite{DBLP:conf/cikm/HuangZZC18, DBLP:conf/cikm/ZhaoT17, DBLP:conf/kdd/WangKGL16}. We propose $fGAT$ that replaces additive self-attention mechanism with a novel scaled dot product self-attention mechanism~\cite{DBLP:conf/nips/VaswaniSPUJGKP17} to learn category specific feature embedding of regions.
The feature embedding of a target region is formulated as an aggregation of its neighbors' features based on their possible influence on the crimes of the target region. The intuition behind finding a possible influential feature is - if two regions having similar features experience similar crime occurrences at a specific time step then the features might influence crimes or serve as proxies for crime prediction in addition to the crime occurrences.
We compute the query vector $\mathbf{q}_{ii'}$ by multiplying the concatenated crime occurrences of a target region ($r_i$) and its neighbor region ($r_{i'}$) with weight matrix $\mathbf{W_q} \in \mathbb{R}^{d_q\times 2}$ to learn their crime distribution similarities. For preparing the key vector $\mathbf{k}_{ii'}^j$ for feature $j$, we multiply the concatenated features of $r_i$ and $r_{i'}$ with weight matrix $\mathbf{W_k} \in \mathbb{R}^{d_k\times 2}$ to learn their feature similarities. Here, $d_q$ and $d_k$ represents the dimension of the query and key vector, respectively.
\begin{align*}
\mathbf{q}_{ii'} = \mathbf{W_q}([x_{i, t}^k \mathbin\Vert x_{i', t}^k])\\
\mathbf{k}_{ii'}^j = \mathbf{W_k}([f_{i, t}^j \mathbin\Vert f_{i', t}^j])
\end{align*}
Then, the attention weight of $j$-th feature of $r_{i'}$ is calculated using the dot-product attention mechanism.
$$\beta_{ii'}^{j} =\text{softmax}_{j}(\frac{\mathbf{q}_{ii'}{\mathbf{k}_{ii'}^j}^T}{\sqrt{d_k}}) $$
Once the attention weights of individual features are found, the feature embedding ${e}_{i, t}^k$ of $r_i$ is formulated as follows.
\begin{equation}
\label{eq:2}
\mathbf{{e}}_{i, t}^k = \sigma\left(\sum_{i' \in \mathcal{N}_i}\left( \alpha_{ii'} \sum_{j=1}^{J}\beta_{ii'}^{j} \mathbf{w_v}f_{i', t}^j\right)\right)
\end{equation}
Here $ \beta_{ii'}^{j}\mathbf{w_v} f_{i', t}^j$ represents the contribution of $j$-th feature of $r_i'$ on the feature embedding of $r_i$ and $\mathbf{w_v} \in \mathbb{R}^{F}$. Figure~\ref{Fig:overview}b gives an overview of generating the feature embedding.
Finally, We concatenate the crime embedding, $\mathbf{{c}}_{i, t}^k$ and feature embedding, $\mathbf{{e}}_{i, t}^k$ to find spatial embedding $\mathbf{s}_{i, t}^k$ of target region $r_i$ at $t$-th time step for crime category $k$. This spatial embedding $\mathbf{s}_{i, t}^k$ is fed as input to a SAB-LSTM cell for time step $t$ as shown in Figure~\ref{Fig:overview}c.
$$\mathbf{s}_{i, t}^k = [\mathbf{{c}}_{i, t}^k \mathbin\Vert \mathbf{{e}}_{i, t}^k]$$
\subsection{Temporal View}
\label{temporal}
LSTM and GRU are two popular recurrent neural networks (RNNs) that capture temporal correlations. However, besides being non-interpretable they suffer from vanishing gradient problem for long sequences. To address these issues, attention-based RNNs are proposed that use attention mechanism to focus on relevant hidden states. These attention weights are difficult to interpret because of the recurrence on the hidden states generated by LSTMs~\cite{DBLP:conf/nips/ChoiBSKSS16}. They are also not sparse enough to be meaningful for long sequence.
SAB-LSTM~\cite{DBLP:conf/nips/KeGBBMPB18} back-propagates across only a selected small subset instead of all hidden states, which are selected using a sparse and hard attention mechanism. Thus it mitigates the gradient vanishing problem and is also interpretable. In this section, first we give a brief overview of a SAB-LSTM cell and then discuss how we apply them in predicting crimes.
At each time step $t$ the underlying LSTM of SAB-LSTM takes the spatial embedding of region $r_i$, $\mathbf{s}_{i, t}^k$ and the previous hidden state $\mathbf{h_{t-1}}$ as inputs for crime category $k$. It produces a new cell state $\mathbf{c_t}$ along with a provisional hidden state $\mathbf{\hat{h}_t}$.
$$\mathbf{\hat{h}}_{t}, \mathbf{c_{t}} = \text{LSTM}(\mathbf{s}_{i, t}^k, \mathbf{h_{t-1}})$$
The provisional hidden state, $\mathbf{\hat{h}_t}$ is concatenated with all the vectors stored in memory $\mathcal{M} = [\mathbf{h}_1^{mem}, \mathbf{h}_2^{mem}, \ldots, \mathbf{h}_{|\mathcal{M}|}^{mem}]$ and passed through a feed-forward neural network to generate unnormalized attention weights $(e_m)$ for each vector stored in the memory. Memory $\mathcal{M}$ contains a set of hidden states selected arbitrarily (after each $k_{att}$ time step) for comparison with the generated provisional hidden state.
$$e_m = \mathbf{W_m}tanh(\mathbf{\hat{h}}_{t} \mathbin\Vert \mathbf{h}_m^{mem})$$
Then, SAB-LSTM subtracts the $(k_{top}+1)$-th highest attention score from all the attention scores and use normalization to generate $k_{top}$ sparse attention weights to select only $k_{top}$ memory cells.
$$\alpha_{m} =\frac{e_{m}-e_{k_{top}+1}}{\sum_{m^{''} \in \mathcal{M}} (e_{m^{''}}-e_{k_{top}+1})}$$
Once the attention weights ($\alpha_{m}$) are obtained, it calculates a summary vector $\mathbf{sum_t}$ by summing over the $k_{top}$ memories. This summary vector is then concatenated to the previously generated hidden provisional state $\mathbf{\hat{h}_t}$ to get the final hidden state, $\mathbf{h_t}$.
\begin{align*}
\mathbf{sum_t} &= \sum_{m \in \mathcal{M}}\alpha_{m}\mathbf{h}_m^{mem}\\
\mathbf{h_t} &= \mathbf{\hat{h}}_t + \mathbf{sum_t}
\end{align*}
The hidden state generated at time step $t$ has two contributing factors. First, the provisional hidden vector $(\mathbf{\hat{h}_t})$ which is the output of a traditional LSTM at time $t$ and non-interpretable. Second, the summary vector $\mathbf{sum_t}$ which is the summation of dynamic, sparse hidden states aligned with current state and interpretable. For crime prediction task we omit the first contributing factor and only use the summary vector $\mathbf{sum_t}$ as our output hidden state $\mathbf{h_t}$. Even though the accuracy is slightly compromised but this makes SAB-LSTM more interpretable.
\begin{equation}
\label{eq:3}
\mathbf{h_t} = \mathbf{sum_t} = \sum_{m \in \mathcal{M}}\alpha_{m}\mathbf{h}_m^{mem}
\end{equation}
We use three SAB-LSTMs for our crime prediction task (Figure~\ref{Fig:overview}d). $\text{SAB-LSTM}_r$ captures recent crime trends based on a target region's spatial embedding during past $T$ time steps. $\text{SAB-LSTM}_d$ captures daily trends based on spatial embedding at the same time step as the predicted time step but on previous days . Finally, $\text{SAB-LSTM}_w$ captures weekly trends based on the the spatial embedding at the same time step as the predicted time step but on previous weeks. We formulate them as follows. For simplicity of representation, we omit region-index $i$ and crime category-index $k$.
\begin{align*}
\mathbf{h}_{T+1}^r &= \text{SAB-LSTM}_r(\mathbf{s}_{t}) \\
\mathbf{h}_{T+1}^d &= \text{SAB-LSTM}_d(\mathbf{s}_{(T+1)-t_{d}*m}) \\
\mathbf{h}_{T+1}^w &= \text{SAB-LSTM}_w(\mathbf{s}_{(T+1)-t_{w}*7*m})
\end{align*}
Here, $t = [1..T], t_d = [1..T_{d}], t_w = [1..T_{w}], m = 24/\tau$, $T_d = T/m$, $T_w = T/(m*7), \tau = \text{length of each time step}$. $\mathbf{h}_{T+1}^r, \mathbf{h}_{T+1}^d, \mathbf{h}_{T+1}^w \in \mathbb{R}^H$, $H = \text{hidden state dimension}$.
\vspace{2mm}
\subsection{Prediction}
\label{prediction}
After calculating the final hidden states of all three SAB-LSTMs we use location-based attention mechanism~\cite{DBLP:conf/emnlp/LuongPM15} to capture the contribution ($\alpha_a$) of the recent, daily and weekly trends. Then, a context vector is calculated using the generated attention weights. Here, $\mathbf{W_h} \in \mathbb{R}^{H\times A}, \mathbf{b_h} \in \mathbb{R}^A $ are learnable parameters, $A$ = attention dimension and $a=\{r, d, w\}$.
$$\alpha_a = \text{softmax}_{a} (\text{tanh}(\mathbf{W_h}^{T}\mathbf{h}_{T+1}^a + \mathbf{b_h})$$
\begin{equation}
\label{eq:4}
\mathbf{c}_{i, T+1}^k = \sum_a \alpha_a \mathbf{h}_{T+1}^a
\end{equation}
Finally, the context vector is fed to a fully connected layer for predicting the crime occurrence at time step ($T+1$) for region $r_i$ and crime category $k$ where, $\mathbf{w} \in \mathbb{R}^{H}, \mathbf{b} \in \mathbb{R}$ are learnable parameters. We add the previously omitted region-index $i$ and crime category-index $k$ below.
\begin{equation}
\label{eq:5}
\hat{y}_{i, T+1}^k = \text{tanh}(\mathbf{w}\mathbf{c}_{i, T+1}^k + b)
\end{equation}
Figure~\ref{Fig:overview}d gives an overview of generating the context vector and prediction.
\section{Experiment}
\label{Sec:experimentalResult}
\subsection{Experimental Settings}
\subsubsection{Data-sets}
We evaluate our model on publicly available 2019 Chicago crime data~\cite{chicago_crime2019}, following the state-of the-art~\cite{DBLP:conf/kdd/WangKGL16, DBLP:journals/tbd/WangYKGL19}. Chicago is one of the most violent cities of United States and the crime concentration of Chicago is very diverse; it has both some of the safest and some of the most crime prone neighborhoods. We use 2019 Chicago taxi trip data~\cite{chicago_taxi2019} and POI information as external features. We collect POI information from FourSquare API while Chicago crime and taxi data are publicly available.
\begin{itemize}
\item \textbf{Chicago-Crime} (2019). We select \num[group-separator={,}]{152720} crime records of four crime categories: theft, criminal damage, battery, narcotics from $1/1/2019$ to $31/12/2019$ and extract these information of each record: timestamp, primary category of crime, community area where it occurred.
\item \textbf{Chicago-POI} We select \num[group-separator={,}]{89324} POIs of $10$ categories: food, residence, travel, arts \& entertainment, outdoors \& recreation, education, nightlife, professional, shops and event.
\item \textbf{Chicago-Taxi} (2019). We select \num[group-separator={,}]{29110097} taxi trips from $1/1/2019$ to $31/12/2019$ and extract these information of each record: pickup timestamp, drop-off timestamp, pickup community area, drop-off community area.
\end{itemize}
\subsubsection{Data Preprocessing} We consider Chicago crime data of first 8 months as training set and 10\% and 90\% of the remaining last 4 months as validation set and test set, respectively. We use taxi inflow (F1), outflow (F2) and POI category: food (F3), residence (F4), travel (F5), arts \& entertainment (F6), outdoors \& recreation (F7), education (F8), nightlife (F9), professional (F10), shops (F11) and event (F12) as external features. We use Min-Max normalization to scale the crime events to [-1, 1] and later denormalize the prediction to get the actual number of crime events. Following~\cite{DBLP:journals/tbd/WangYKGL19}, we do not scale the external features.
\subsubsection{Parameter Settings} We optimize the hyperparameters of AIST using a grid search strategy. The search space for every hyperparameter is presented in Table~\ref{table:exp_hyperparameter} and the selected value in the search space is shown in bold. For simplicity, we use the same parameter settings across all crime categories and regions. AIST is trained using the Adam optimizer with batch size = 42 and initial learning rate = 0.001. We set the duration of each time step, $\tau = 4$ hours. We set the number of recent ($T$), daily ($T_d$) and weekly ($T_w$) time steps to 20, 20 and 3, respectively.
\emph{hGAT \& fGAT settings.} Both hGAT and fGAT are single layer GATs consisting of single attention head for computational efficiency.
We set the output size ($F$) of both hGAT and fGAT to 8. For fGAT, we set $d_q, d_k=40$. Dropout with $p=0.5$ is applied to unnormalized node-level ($e_{ii'}^c$), unnormalized parent-level ($e_{ii'}^p$), normalized combined attention weights ($\alpha_{ii'}$) in hGAT, and dropout with $p=0.5$ is applied to normalized dot-product attention weights ($\beta_{ii'}^j$) in fGAT.
\emph{SAB-LSTM settings.} All 3 SAB-LSTMs are single layered with hidden dimension $H=40$. For $SAB-LSTM_{r}$ and $SAB-LSTM_{d}$, we set $k_{att}=5, k_{top}=5, trunc_{length} = 5$. For $SAB-LSTM_w$ we set $k_{att}=1, k_{top}=5, trunc_{length} = 1$. Dropout with $p=0.2$ is applied to each output of 3 SAB-LSTMs. Finally, we set the attention dimension of location based attention as $A=30$.
\begin{small}
\begin{table*}[htbp]
\caption{Hyperparameter Settings of AIST}
\centering
\begin{tabular}[t]{|l|l|c|}
\hline
Hyperparameters & Search Space\\
\hline
Number of recent time step ($T$) & [16, \textbf{20}, 24, 28]\\
\hline
Number of daily time step ($T_d$) & [12, 16, \textbf{20}, 24] \\
\hline
Number of weekly time step ($T_w$) & [2, \textbf{3}, 4, 5]\\
\hline
Output size of both hGAT and fGAT ($F$) & [6, \textbf{8}, 10, 12]\\
\hline
Dimension of query and key vector of fGAT ($d_q, d_v$) & [36, \textbf{40}, 44, 48]\\
\hline
Dimension of hidden states of SAB-LSTM ($H$) & [24, 32, \textbf{40}, 48] \\
\hline
Dimension of Location Attention ($H$) & [22, \textbf{30}, 38, 46]\\
\hline
\end{tabular}
\label{table:exp_hyperparameter}
\end{table*}
\end{small}
\subsubsection{Evaluation Criteria}
We use mean average error (MAE) and mean square error (MSE) to evaluate AIST predictions. Here, $n$ represents the number of predictions, $y_i$ represents the predicted result and $\hat{y}_i$ represents the ground truth.
\begin{align*}
\text{MAE} = \frac{1}{n}\sum_{i=1}^{n}|y_i - \hat{y}_i| \qquad
\text{MSE} = \frac{1}{n}\sum_{i=1}^{n}(y_i - \hat{y}_i)^2 \end{align*}
We also use Total Variation Distance (TVD) for comparing prediction scores and Jensen-Shannon Divergence (JSD) for comparing attention weight distributions to evaluate the interpretations of AIST, where $\alpha = \frac{\alpha_1 + \alpha_1}{2}$ and $\text{KL}(p || q)$ calculates the Kullback–Leibler divergence between probability distributions $p$ and $q$.
$$\text{TVD} = \frac{1}{2}\sum_{i=1}^{|\textbf{y}|}(|\hat{y}_{1i} - \hat{y}_{2i}|)$$
$$\text{JSD}(\alpha_1, \alpha_2) = \frac{1}{2} \text{KL}[\alpha_1 || \alpha] + \frac{1}{2} \text{KL}[\alpha_2 || \alpha]$$
\subsubsection{Baselines}
We compare AIST with the following baselines.
\begin{compactitem}
\item ARIMA~\cite{4666600}. The most general case of models for predicting time series combining moving average and auto-regression.
\item DTR~\cite{DBLP:books/wa/BreimanFOS84}. A decision tree algorithm for regression that chooses the best random split while partitioning samples in multiple subsets.
\item Att-RNN~\cite{DBLP:journals/corr/BahdanauCB14}. It uses attention mechanism with RNN to capture the temporal correlation.
\item DeepCrime~\cite{DBLP:conf/cikm/HuangZZC18}. A hierarchical recurrent framework that encodes the temporal correlation and inter-dependencies between crimes and urban anomalies.
\item MiST~\cite{DBLP:conf/www/HuangZZWCY19}. It uses multiple LSTMs to encode the spatial, temporal and categorical views of crime.
\item GeoMAN*. A multi-level attention network with a sequential encoder-decoder architecture that models both spatial and temporal correlation, customized to predict crimes. Unlike GeoMAN~\cite{DBLP:conf/ijcai/LiangKZYZ18}, while calculating the spatial dependencies we only consider those who share a common boundary rather than considering all the regions in the network.
\item STGCN~\cite{DBLP:conf/ijcai/YuYZ18}. A graph convolutional layer is placed between two gated temporal convolution layers to model the spatio-temporal correlation.
\item MVGCN*. A GNN architecture with multiple graph convolution and fully connected layers to process different temporal and semantic views, respectively. MVGCN~\cite{DBLP:journals/TKDE/9139357} is customized by only considering those regions that share a common boundary for predicting crimes.
\end{compactitem}
\begin{small}
\begin{table*}[t]
\caption{Comparison of AIST with baselines on Chicago Crime Data (2019)}
\centering
\begin{tabular}[t]{|c|c|c|c|c|c|}
\hline
Model & Criteria & Theft (C1) &
\begin{tabular}{@{}c@{}}Criminal \\ Damage (C2)\end{tabular}
& Battery (C3) & Narcotics (C4) \\
\hline
\multirow{2}{*}{ARIMA~\cite{4666600}}
& MAE & 1.2010 & 0.5863 & 0.8840 & 0.5705\\
& MSE & 2.8492 & 0.7238 & 1.4242 & 0.7928 \\
\hline
\multirow{2}{*}{DTR~\cite{DBLP:books/wa/BreimanFOS84}}
& MAE & 1.1943 & 0.5590 & 0.8983 & 0.4901\\
& MSE & 3.3275 & 0.8123 & 1.9336 & 0.8522\\
\hline
\multirow{2}{*}{Att-RNN~\cite{DBLP:journals/corr/BahdanauCB14}}
& MAE & 1.0419 & 0.4096 & 0.7377 & 0.4128\\
& MSE & 2.5443 & 0.4427 & 1.0665 & 0.6380\\
\hline
\multirow{2}{*}{DeepCrime~\cite{DBLP:conf/cikm/HuangZZC18}}
& MAE & 1.0022 & 0.3727 & 0.7271 & 0.3702 \\
& MSE & 2.6279 & 0.4751 & 1.0567 & 0.6394\\
\hline
\multirow{2}{*}{MiST~\cite{DBLP:conf/www/HuangZZWCY19}}
& MAE & 1.0241 & 0.3727 & 0.7365 & 0.3701\\
& MSE & 2.5153 & 0.4836 & 1.0345 & 0.6495\\
\hline
\multirow{2}{*}
{\begin{tabular}{@{}c@{}}customized \\GeoMAN~\cite{DBLP:conf/ijcai/LiangKZYZ18}\end{tabular}}
& MAE & 0.9092 & 0.3876 & 0.7226 & 0.3450\\
& MSE & 1.9930 & \textbf{0.4385} & 0.9871 & \textbf{0.5595}\\
\hline
\multirow{2}{*}{STGCN~\cite{DBLP:conf/ijcai/YuYZ18}}
& MAE & 1.0416 & 0.5130 & 1.0869 & 0.3886\\
& MSE & 1.8121 & 0.4860 & 1.0595 & 0.6342\\
\hline
\multirow{2}{*}
{\begin{tabular}{@{}c@{}}customized \\MVGCN~\cite{DBLP:journals/TKDE/9139357}\end{tabular}}
& MAE & 1.5244 & 0.4641 & 0.7928 & 0.4093 \\
& MSE & 4.2593 & 0.7019 & 1.1949 & 0.8337\\
\hline
\multirow{2}{*}{AIST}
& MAE & \textbf{0.8747} & \textbf{0.3615} & \textbf{0.6910} & \textbf{0.3399}\\
& MSE & \textbf{1.6986} & 0.4837 & \textbf{0.9568} & 0.5609\\
\hline
\end{tabular}
\label{table:exp_result}
\vspace{-1em}
\end{table*}
\end{small}
\subsection{Prediction Performance}
\label{SubSec:pred}
\subsubsection{Comparison with baselines. }
The performance of the baselines and AIST is shown in Table~\ref{table:exp_result}. The baselines include both high-performance crime prediction models (e.g., DeepCrime, Mist) and high-performance spatio-temporal prediction models (e.g., customized GeoMAN, STGCN, customized MVGCN). All the baselines have been tuned optimally to produce the best prediction result. In general, AIST outperforms all baselines by achieving the lowest MAE and MSE scores across all crime categories (except the MSE scores for crime category Criminal Damage (C2) and Narcotics (C4)).
\begin{itemize}[leftmargin=*]
\item AIST learns the time varying spatial dependencies, diverse temporal correlation and crime relevant dynamic context to perform crime prediction tasks. Other competing deep learning models such as MVGCN*, STGCN, GeoMAN*, MiST, Att-RNN do not learn the crime and region specific relevant context; DeepCrime, Att-RNN do not consider the spatial correlation and MVGCN* do not consider the temporal correlation. As a result, in general these models fail to perform better than AIST for crime prediction tasks.
\item Aside from AIST, GeoMAN* has the second best MAE and MSE scores across all crime categories. On top of that, it has better MSE scores than AIST for category Criminal Damage (C2) and Narcotics (C4). C2 lacks periodical temporal properties and the surrounding context has less influence on C4. Since, GeoMAN* does not consider daily or weekly temporal properties and the influence of region specific context on a crime category, it performs better than AIST for these crime categories in terms of MSE scores.
\item Other attention-based neural network architectures: DeepCrime and MiST perform considerably worse than GeoMAN* because they do not consider the spatial correlation and external features, respectively. Though DeepCrime and MiST have similar MAE scores, MiST is better than DeepCrime in terms of the MSE score. This is because DeepCrime only captures the temporal correlation and region-category dependencies, whereas MiST captures both spatio-temporal and cross-categorical correlation of crimes.
\item STGCN has the third best MSE score across all crime categories. Only AIST and GeoMAN* have better MSE score than STGCN. Having a lower MSE score than those of other attention-based deep learning models such as Att-RNN, DeepCrime and MiST, STGCN is more likely to capture the sudden changes in crime distribution. However, in terms of the MAE score STGCN shows a poor performance in comparison with AIST and others such as GeoMAN*, MiST, DeepCrime, and Att-RNN.
\item Att-RNN, a plain recurrent neural network architecture that only considers the recent temporal correlation is behind DeepCrime and MiST, but above STGCN in terms of the performance based on the MAE score. However, in terms of the MSE score its performance is quite similar to DeepCrime. For category Criminal Damage (C2), Att-RNN performs better than AIST supporting our claim that periodical information is not helpful for predicting this category of crime.
\item MVGCN*, despite being a top performing architecture in crowd flows prediction, performs worst among the competing deep learning models in the crime prediction task based on MSE scores. Because of the sparsity of crime distributions, careful exploration of available crime data and context are an absolute necessity for capturing the sudden change in the crime distribution. However, the large MSE scores of MVGCN* mean that the graph convolution and fully connected layers in MVGCN* used to model the spatial correlation and the influence of the external features lack the ability to do so. Hence, it fails to compete with others.
\item Traditional time series analysis and machine learning methods such as ARIMA and DTR though interpretable, lack the ability to model the non-linear and complex crime patterns. This is because ARIMA only considers a fixed temporal pattern, whereas DTR does not consider the temporal properties of crime at all. Hence, in general they show poor performance than the deep learning models across all crime categories.
\end{itemize}
\subsubsection{Effectiveness of different spatial components of AIST} We consider the following variants of AIST to understand the influence of different spatial components on its prediction performance. Figure~\ref{Fig:eval_self} shows a comparative performance analysis of these spatial variants of AIST.
\begin{itemize}
\item $\text{AIST}_{g}.$ hGAT is replaced with traditional graph attention networks to learn crime embedding; fGAT is omitted from the spatial module.
\item $\text{AIST}_{h}.$ hGAT is used to learn crime embedding; fGAT is omitted from the spatial module.
\item $\text{AIST}_f.$ hGAT is used to learn crime embedding and the crime embedding is concatenated with the external features to produce the final spatial embedding; fGAT is omitted from the spatial module.
\item $\text{AIST}_{f'}.$ GAT (in place of hGAT) and fGAT constitute the spatial module.
\end{itemize}
$\text{AIST}_h$ has a better MAE score across all crime categories than $\text{AIST}_g$ which justifies the selection of hGAT over GAT for learning the crime embedding. Specifically, from Fig~\ref{Fig:eval_model7} it is evident that hGAT learns a better crime embedding of category Narcotics compared to other categories, which indicates the existence of strong narcotics networks in certain parts of Chicago. Besides, the fact that AIST consistently performs better than $\text{AIST}_{f'}$ shows the superiority of the spatial embedding learned by hGAT alongside fGAT over $\text{AIST}_{f'}$.
For categories Theft (C1) and Battery (C3), the concatenation of external features improves the prediction performance of $\text{AIST}_h$, which indicates the impact of contextual information for the crime prediction task, specially for these categories. However, the prediction performance of $\text{AIST}_f$ deteriorates significantly for category Narcotics (C4) and remains almost same as $\text{AIST}_h$ for category Criminal Damage (C2), which suggest its inability to constantly differentiate the influential features from noise across all crime categories. The significant
performance improvement of AIST over $\text{AIST}_f$ in predicting crime events across all categories suggests that careful extraction and learning of crime relevant feature embedding by fGAT is a necessity while performing crime prediction tasks.
\begin{figure}[tbp]
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/smodule1.pdf_tex}
\caption{C1: Theft}
\label{Fig:eval_model1}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/smodule2.pdf_tex}
\caption{C2: Criminal Damage}
\label{Fig:eval_model2}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/smodule3.pdf_tex}
\caption{C3: Battery}
\label{Fig:eval_model3}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/smodule7.pdf_tex}
\caption{C4: Narcotics}
\label{Fig:eval_model7}
\end{subfigure}
\caption{Effect of different spatial components}
\label{Fig:eval_self}
\vspace{-1em}
\end{figure}
\subsubsection{Effectiveness of different temporal components of AIST}
Figure~\ref{Fig:tmodel} shows a comparative performance analysis of different temporal components of AIST.
\begin{itemize}
\item $\text{AIST}_{r}.$ $\text{SAB-LSTM}_{d}$ and $\text{SAB-LSTM}_{w}$ are omitted from the temporal module.
\item $\text{AIST}_{d}.$ $\text{SAB-LSTM}_{w}$ is omitted from the temporal module.
\item $\text{AIST}_{w}.$ $\text{SAB-LSTM}_{d}$ is omitted from the temporal module.
\item $\text{AIST}_{l}.$ All three $\text{SAB-LSTMs}$ are replaced by traditional LSTMs.
\end{itemize}
The significant decrease in the MAE scores of $\text{AIST}_{d}$ and $\text{AIST}_{w}$ over $\text{AIST}_{r}$ for Theft (C1), Battery (C2) and Narcotics (C4) suggest that both daily and weekly trends are instrumental in crime prediction tasks. Between these two, $\text{AIST}_{w}$ has better MAE scores over $\text{AIST}_{d}$ across all crime categories indicating the dominance of weekly trends over daily trends. Above all, the better MAE scores of AIST over $\text{AIST}_{l}$ for all crime categories justify the selection of SAB-LSTMs over traditional LSTMs for the crime prediction tasks.
Contrary to the general observation discussed above, $\text{AIST}_{r}, \text{AIST}_{d}, \text{and AIST}_{w}$ show similar performance for category Criminal Damage (C2) (Figure~\ref{Fig:tmodel2}), which suggests that it does not follow any daily or weekly trend.
\begin{figure}[htbp]
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/tmodule1.pdf_tex}
\caption{C1: Theft}
\label{Fig:tmodel1}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/tmodule2.pdf_tex}
\caption{C2: Criminal Damage}
\label{Fig:tmodel2}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/tmodule3.pdf_tex}
\caption{C3: Battery}
\label{Fig:tmodel3}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/tmodule7.pdf_tex}
\caption{C4: Narcotics}
\label{Fig:tmodel7}
\end{subfigure}
\caption{Effect of different temporal components}
\label{Fig:tmodel}
\vspace{-1em}
\end{figure}
\subsubsection{Effect of Parameters}
Different parameters like number of recent ($T$), daily ($T_d$) and weekly ($T_w$) time steps, dimension of SAB-LSTM hidden states ($H$), output size of hGAT and fGAT ($F$), dimension of query ($d_q$), key ($d_k$) vectors and location attention ($A$) have impact on the performance of AIST. To better understand the crucial parameters of AIST and their effect on its prediction performance, we run several experiments and present the results in ~\Cref{Fig:param_T,Fig:param_TD,Fig:param_TW,Fig:param_H,Fig:param_F,Fig:param_Q,Fig:param_A}. To observe the effect of a parameter, the value of the parameter is varied within its range, and other parameters are set to their default values as presented in Table~\ref{table:exp_hyperparameter}.
Figures~\ref{Fig:param_T},~\ref{Fig:param_TD}, and~\ref{Fig:param_TW} suggest that AIST is sensitive to the number of recent ($T$), daily ($T_d$) and weekly ($T_w$) time steps, which are being fed to the three SAB-LSTMs as input. AIST shows poor performance for both smaller and larger $T, T_d$ due to the lack of data for learning temporal dependencies and the absence of long temporal correlation, respectively. Somewhere in between, when $T, T_d = 20$, AIST in general performs best by capturing the recent and periodic properties of the crime. On the contrary, AIST performs well when the number of weekly time steps $(T_w)$ is relatively small (Figure~\ref{Fig:param_TW}). However, unlike other categories, AIST performs best for category Theft (C1), when the number of recent and weekly time steps are comparatively larger (Figure~\ref{Fig:param_T1},~\ref{Fig:param_TW1}). This signifies the existence of long term temporal dependencies for category Theft.
AIST is also sensitive to the dimension $(H)$ of the hidden states of SAB-LSTMs (Figure~\ref{Fig:param_H}) and output size $(F)$ of hGAT and fGAT (Figure~\ref{Fig:param_F}). Limited spatial and temporal information make the training hard for AIST. As a result, the performance of AIST deteriorates. On the other hand, a larger output size and dimension of the hidden state make it easier for AIST to overfit the data. Hence, we set $F=8$ and $H=40$ so that AIST can generalize well by learning sufficient spatial and temporal information.
It is evident from Figures~\ref{Fig:param_Q} and~\ref{Fig:param_A} that the query and key dimensions $(d_q, d_k)$ and location attention dimension $(A)$ follow the same trend as the other hyperparameters discussed above. Based on the performance of AIST across different crime categories, we set $d_q, d_k = 40$ and $A = 30$.
\begin{figure}[htbp]
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_T1.pdf_tex}
\caption{C1: Theft}
\label{Fig:param_T1}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_T2.pdf_tex}
\caption{C2: Criminal Damage}
\label{Fig:param_T2}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_T3.pdf_tex}
\caption{C3: Battery}
\label{Fig:param_T3}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_T7.pdf_tex}
\caption{C4: Narcotics}
\label{Fig:param_T7}
\end{subfigure}
\caption{Effect of the Number of Recent Time Steps, $T$}
\vspace{-1em}
\label{Fig:param_T}
\end{figure}
\begin{figure}[htbp]
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_td1.pdf_tex}
\caption{C1: Theft}
\label{Fig:param_TD1}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_td2.pdf_tex}
\caption{C2: Criminal Damage}
\label{Fig:param_TD2}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_td3.pdf_tex}
\caption{C3: Battery}
\label{Fig:param_TD3}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_td7.pdf_tex}
\caption{C4: Narcotics}
\label{Fig:param_TD7}
\end{subfigure}
\caption{Effect of the Number of Daily Time Steps, $T_d$}
\label{Fig:param_TD}
\vspace{-1em}
\end{figure}
\begin{figure}[htbp]
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_tw1.pdf_tex}
\caption{C1: Theft}
\label{Fig:param_TW1}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_tw2.pdf_tex}
\caption{C2: Criminal Damage}
\label{Fig:param_TW2}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_tw3.pdf_tex}
\caption{C3: Battery}
\label{Fig:param_TW3}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_tw7.pdf_tex}
\caption{C4: Narcotics}
\label{Fig:param_TW4}
\end{subfigure}
\caption{Effect of the Number of Weekly Time Steps, $T_w$}
\label{Fig:param_TW}
\vspace{-1em}
\end{figure}
\begin{figure}[htbp]
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_h1.pdf_tex}
\caption{C1: Theft}
\label{Fig:param_H1}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_h2.pdf_tex}
\caption{C2: Criminal Damage}
\label{Fig:param_H2}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_h3.pdf_tex}
\caption{C3: Battery}
\label{Fig:param_H3}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_h7.pdf_tex}
\caption{C4: Narcotics}
\label{Fig:param_H7}
\end{subfigure}
\caption{Effect of the Dimension of Hidden State, $H$}
\label{Fig:param_H}
\vspace{-1em}
\end{figure}
\begin{figure}[htbp]
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_F1.pdf_tex}
\caption{C1: Theft}
\label{Fig:param_F1}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_F2.pdf_tex}
\caption{C2: Criminal Damage}
\label{Fig:param_F2}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_F3.pdf_tex}
\caption{C3: Battery}
\label{Fig:param_F3}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_F7.pdf_tex}
\caption{C4: Narcotics}
\label{Fig:param_F7}
\end{subfigure}
\caption{Effect of the Output Size of hGAT and fGAT, $F$}
\label{Fig:param_F}
\end{figure}
\begin{figure}[htbp]
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_Q1.pdf_tex}
\caption{C1: Theft}
\label{Fig:param_Q1}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_Q2.pdf_tex}
\caption{C2: Criminal Damage}
\label{Fig:param_Q2}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_Q3.pdf_tex}
\caption{C3: Battery}
\label{Fig:param_Q3}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_Q7.pdf_tex}
\caption{C4: Narcotics}
\label{Fig:param_Q7}
\end{subfigure}
\caption{Effect of the Dimension of Query and Key, $d_q, d_k$}
\label{Fig:param_Q}
\end{figure}
\begin{figure}[htbp]
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_a1.pdf_tex}
\caption{C1: Theft}
\label{Fig:param_A1}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_a2.pdf_tex}
\caption{C2: Criminal Damage}
\label{Fig:param_A2}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_a3.pdf_tex}
\caption{C3: Battery}
\label{Fig:param_A3}
\end{subfigure}
\begin{subfigure}{0.245\columnwidth}
\def0.9\textwidth{\columnwidth}
\input{Figures/param_a7.pdf_tex}
\caption{C4: Narcotics}
\label{Fig:param_A7}
\end{subfigure}
\caption{Effect of the Dimension of Location Attention, $A$}
\label{Fig:param_A}
\end{figure}
\subsubsection{Effect of Train/Test Ratio}
We run several experiments to learn the effect of the train-test ratio on the prediction performance of AIST. We consider the first $n \in \{6, 7, 8, 9, 10\}$ months of Chicago crime data (2019) as training set. We take 10\% of the remaining data of last $(12-n)\in \{6, 5, 4, 3, 12\}$ months for the validation set and use the rest as the test set. Table~\ref{table:tt_ratio} shows the prediction performance of AIST across all crime categories against different train-test ratio.
Similar to the most deep learning models, fewer training samples cause AIST to overfit the data and as a result AIST shows poor prediction performance. This is evident from the reported MAE and MSE scores of AIST when only 6 months of data is used for training. Once we gradually increase the size of the training data, the prediction performance of AIST improves significantly and reaches its peak when the size of the training data is 8 months. Adding additional training data beyond 8 months again deteriorates the prediction performance of AIST.
\begin{small}
\begin{table*}[htbp]
\caption{Effect of Train-Test Ratio on Crime Prediction Performance of AIST}
\centering
\begin{tabular}[t]{|E|D|D|D|D|D|D|D|D|}
\hline
Train / Test Ratio
& \multicolumn{2}{c|}{\centering Theft (C1)}
& \multicolumn{2}{c|}{\begin{tabular}{@{}c@{}} Criminal \\ Damage (C2)\end{tabular}}
& \multicolumn{2}{c|}{Battery (C3)}
& \multicolumn{2}{c|}{Narcotics (C4)}
\\
\hline
In Percent & MAE & MSE & MAE & MSE & MAE & MSE & MAE & MSE\\
\hline
0.50 / 0.45 & 1.2576 & 4.4737 & 0.3869 & 0.5121 & 0.7745 & 1.2920 & 0.4179 & 0.7457\\
\hline
0.58 / 0.38 & 0.9445 & 2.0771 & 0.3765 & 0.4972 & 0.7229 & 1.0653 & 0.3518 & 0.5619\\
\hline
0.67 / 0.30 & \textbf{0.8747} & \textbf{1.6986} & \textbf{0.3615} & 0.4837 & \textbf{0.6910} & \textbf{0.9568} & \textbf{0.3399} & \textbf{0.5609}\\
\hline
0.75 / 0.23 & 1.3403 & 4.1698 & 0.3709 & 0.4748 & 0.7915 & 1.2591 & 0.4208 & 0.7391\\
\hline
0.83 / 0.15 & 1.2807 & 3.8667 & 0.3646 & 0.\textbf{4650} & 0.7988 & 1.1857 & 0.4240 & 0.7518\\
\hline
\end{tabular}
\label{table:tt_ratio}
\end{table*}
\end{small}
\subsection{Evaluation of Interpretability}
The notion of interpretability mainly comes down to two points: i) plausibility: how understandable it is to humans, and ii) faithfulness: how accurately it refers to the true reasoning process of a model.
Besides human evaluations~\cite{DBLP:conf/chi/KaurNJCWV20,
DBLP:conf/iui/EhsanTCHR19, DBLP:conf/naacl/MullenbachWDSE18}, explanations that align directly with the input have been considered as plausible explanations~\cite{DBLP:conf/emnlp/LeiBJ16}. Attentions are plausible explanations because they assign importance weights to the inputs while making a prediction ~\cite{DBLP:conf/emnlp/WiegreffeP19}. Since AIST interprets the importance of different regions, features, time steps and trends on the crime prediction based on four attention modules, the interpretation of AIST is \textit{plausible}.
A recent study~\cite{DBLP:conf/emnlp/WiegreffeP19} shows the conditions under which attentions can be regarded as faithful explanations, and nullifies the claim~\cite{DBLP:conf/acl/SerranoS19,DBLP:conf/naacl/JainW19} that criticizes attention as a form of faithful explanation due to its weak correlation with other feature importance metrics and the existence of alternate adversarial attention weights. Specifically,~\cite{DBLP:conf/emnlp/WiegreffeP19} proposes a series of extensive experiments based on dataset and model properties: i) train on uniform attention weights: the attention distribution is frozen to uniform weights to validate whether the attention is actually necessary for a better performance, ii) calibration of variance: the model is trained with different initializing seeds to generate base variance for attention distributions, iii) train an MLP (multilayer perceptron): the LSTM cells are replaced by MLP and are trained separately and iv) train an adversary: the model is trained to provide similar predictions as the base model while keeping the attention distributions distant from the actual ones for ascertaining exclusivity. We evaluate the attention weights generated by AIST by performing these experiments (except iii since the attention modules used in AIST are either feed-forward neural networks or sparse which do not comply with the experimental settings) to validate their \textit{faithfulness}.
Since the faithfulness varies across model, tasks and input space, both~\cite{DBLP:conf/emnlp/WiegreffeP19, DBLP:conf/acl/JacoviG20} emphasize that the faithfulness should be evaluated in grayscale instead of a binary term, i.e., faithful or not faithful. Following~\cite{DBLP:conf/emnlp/WiegreffeP19, DBLP:conf/acl/JacoviG20}, we consider the degree of faithfulness as it allows to identify the interpretation that is sufficiently faithful to be useful in
practice.
Our process to generate the adversarial attention weights to establish the exclusivity, hence the faithfulness of the model is as follows. We train an adversarial model ($\mathcal{M}_{adv}$) with the objective of minimizing the prediction differences from our AIST model ($\mathcal{M}_{AIST}$) along with a divergent attention distribution for an instance $i$.
$$\mathcal{L}(\mathcal{M}_{AIST}, \mathcal{M}_{adv}) = \text{TVD}(\hat{y}_{AIST}^{(i)}, \hat{y}_{adv}^{(i)}) - \lambda \; \text{KL}(\alpha_{AIST}^{(i)}, \alpha_{adv}^{(i)})$$
Here, $\lambda$ is a hyperparameter that controls the tradeoff between TVD and JSD, where TVD is the levels of prediction variance and JSD (Jensen-Shannon Divergence) quantifies the difference between two attention distributions. In Figure~\ref{Fig:eval_int_adv}, we present the TVD between the predictions of the adversarial and AIST model against the increasing JSD between their attention distributions for a selected number of regions on specific crime categories. We believe the graphs in Figure~\ref{Fig:eval_int_adv} to be representative of all $4$ crime categories across $77$ regions as they show all of the possible three cases: not faithful, moderately faithful and concretely faithful. Fast increase in the prediction difference concurs that the attention scores are not easily manipulable and exclusive. Hence, they can be used as faithful explanations. We also include the scores of uniform model variant (\mysquare{blue}) and random seed initialization (\mytriangle{black!60!green}) in these TVD vs JSD graphs. Figure~\ref{Fig:eval_int_adv_24_1}, ~\ref{Fig:eval_int_adv_27_1}, ~\ref{Fig:eval_int_adv_31_1}, ~\ref{Fig:eval_int_adv_70_3}, ~\ref{Fig:eval_int_adv_43_3} establish attentions as faithful explanations for the specified regions and crime categories as the increase in JSD comes at a high price of the increased TVD (at different rates). However, Figure~\ref{Fig:eval_int_adv_0_1} shows an example where the attention distributions generated by AIST can not be deemed faithful as it is easy to manipulate the attentions without losing much of the prediction performance. We include the predictions of the best adversarial models with instance-average JSD > 1 in Table~\ref{table:exp_result_int}.
Table~\ref{table:exp_result_int} shows the superiority of the base model (AIST) across all four crime categories over its uniform variant and adversarial models. Substantial increase of MAE scores of the uniform model suggests that attention is indeed a necessary component for better performance in the crime prediction. Similarly, a higher MAE score of the adversarial models ascertain the exclusivity of the predictions generated by AIST model.
\begin{figure}[t]
\begin{subfigure}{0.32\columnwidth}
\includegraphics[width=\columnwidth]{Figures/24_1_adv.png}
\caption{R24 C1}
\label{Fig:eval_int_adv_24_1}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\includegraphics[width=\columnwidth]{Figures/27_1_adv.png}
\caption{R27 C1}
\label{Fig:eval_int_adv_27_1}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\includegraphics[width=\columnwidth]{Figures/0_1_adv.png}
\caption{R0 C1}
\label{Fig:eval_int_adv_0_1}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\includegraphics[width=\columnwidth]{Figures/31_1_adv.png}
\caption{R31 C1}
\label{Fig:eval_int_adv_31_1}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\includegraphics[width=\columnwidth]{Figures/70_3_adv.png}
\caption{R70 C3}
\label{Fig:eval_int_adv_70_3}
\end{subfigure}
\begin{subfigure}{0.32\columnwidth}
\includegraphics[width=\columnwidth]{Figures/43_3_adv.png}
\caption{R43 C3}
\label{Fig:eval_int_adv_43_3}
\end{subfigure}
\caption{Evaluation of interpretability (Averaged per-instance test set JSD and TVD from base model for each model variant. JSD is bounded at $\sim2.07$; \mytriangle{black!60!green}: random seed; \mysquare{blue} uniform weights; dotted line: our adversarial setup as $\lambda$ is varied)}
\vspace{-1em}
\label{Fig:eval_int_adv}
\end{figure}
\begin{small}
\begin{table*}[htbp]
\caption{Comparison of AIST with its uniform variant and adversarial models on MAE}
\centering
\begin{tabular}[t]{|c|c|c|c|}
\hline
Crime Category & Uniform & Base Model & Adversarial\\
\hline
Theft (C1) & 0.9776 & \textbf{0.8747} & 1.1807\\
\hline
Criminal Damage (C2) & 0.3738 & \textbf{0.3615} & 0.3734 \\
\hline
Battery (C3) & 0.7206 & \textbf{0.6910} & 0.8029 \\
\hline
Narcotics (C4) & 0.3634 & \textbf{0.3399} & 0.5537 \\
\hline
\end{tabular}
\label{table:exp_result_int}
\end{table*}
\end{small}
\subsection{Case Study}
We select three communities for exploration: (i) R8 (Near North Side): situated in downtown Central Chicago and experiences high crime distribution, (ii) R25 (Austin): situated on the Western side of Chicago and is not as busy as R8, but has a high crime distribution and (iii) R72 (Beverly): located in Southern Chicago and is a quiet residential community with low crime rate. For each community, we randomly select 200 samples from test set and present the contribution of neighbor regions, POI and taxi flow features, trends and important time steps as a heat map in Figure~\ref{Fig:case_8_25}. We denote crimes category Theft, Criminal Damage, Battery and Narcotics with C1, C2, C3 and C4, respectively.
From Equation~\ref{eq:1}, the contribution coefficient of the crime occurrences of region $r_i'\in\mathcal{N}_i$ to the crime embedding of target region $r_i$ during time step $t$ can be calculated as, $\phi(\mathbf{{c}}_{i, t}^k, x_{i', t}^k) = \alpha_{ii'}\mathbf{w_x}{x_{i', t}^k}$. From Equation~\ref{eq:2}, the contribution coefficient of feature $j$ on target region can be calculated as $\phi(\mathbf{{e}}_{i, t}^k, f_{t}^j) = \beta_{ii'}^{j} \sum_{i' \in \mathcal{N}_i}\alpha_{ii'} \mathbf{w_v}f_{i', t}^j$. Similarly, $\phi(\hat{y}_{i, T+1}^k, {h}_{T+1}^a) =\alpha_a \mathbf{w}\mathbf{h}_{T+1}^a$ denotes the contribution coefficient of recent, daily and periodic trends (Equations~\ref{eq:4},~\ref{eq:5}).
For R8, professional POIs (F10) made the highest contribution. R8 is a business region with thousands of jobs and has a large number of professional POIs. Thus it is expected that those POIs have large impact on it's crime embedding. On the other hand, R25 and R72 are residential regions. POI Shop (F11) contributed most for R25 and R72. Besides F11, Residence (F4) POIs also contributed for R72 for all categories except C4. Hence, our model learns both region and category specific influential features. An interesting observation for R8 is that though R8 has a large number of Food and Shop POIs, their contribution is almost none which signifies the quality of our model's prediction.
C1 shows strong long term temporal correlation for R8, whereas none of the crime category shows long term temporal correlation for R72 as the crime number for R72 is low. C3 in R25 shows a strong periodic correlation and unlike R8 and R72, it does not depend on recent crimes. C4 hardly present any long term correlation, which is intuitive. C2 shows daily periodicity in R8 and R25.
R7 and R24 have the most similar crime distribution as R8. However, these similarities vary with crime categories, e.g. for C1 and C3, R7 is given more attention whereas for C4, the attention shifts to R24 and R28. This is because both R24 and R28 experience large number of C4 crimes and have greater influence than R7. For R25, R18 is the most influential region across all crime categories. Unlike R8, the contribution of its neighboring regions are almost same except for R23 which is given less importance compared to other neighbors. The fact that R23 shares its boundary with different regions of different districts/sides makes their crime distribution less similar. R72 gives equal importance to each of its neighbors except R71. R71 has a higher number of crime occurrences than R72 and their crime distribution is quite different for all crime categories except C4. Thus, our model is able to capture diverse spatial correlation.
\begin{figure*}[t]
\captionsetup[subfigure]{aboveskip=-5pt, belowskip=1pt}
\begin{subfigure}{\textwidth}
\def0.9\textwidth{0.9\textwidth}
\centering
\input{Figures/r7.pdf_tex}
\caption{R8 (Near North Side)}
\label{Fig:case_8}
\end{subfigure}
\begin{subfigure}{\textwidth}
\def0.9\textwidth{0.9\textwidth}
\centering
\input{Figures/r24.pdf_tex}
\caption{R25 (Austin)}
\label{Fig:case_25}
\end{subfigure}
\begin{subfigure}{\textwidth}
\def0.9\textwidth{0.9\textwidth}
\centering
\input{Figures/r71.pdf_tex}
\caption{R72 (Beverly)}
\label{Fig:case_72}
\end{subfigure}
\caption {Case analysis of Region 8, 25 and 72}
\label{Fig:case_8_25}
\vspace{-3mm}
\end{figure*}
\section{Conclusion}
\label{sec:conclusion}
We propose AIST, a novel interpretable deep learning framework for crime prediction. AIST captures the dynamic spatio-temporal correlations based on the past crime occurrences, external features (e.g., traffic flow and POI information) and the recent and periodic crime trends. We develop two novel variants of GAT, $hGAT$ and $fGAT$ that allows AIST to improve prediction accuracy and provide the insights behind a prediction. Experiments and case studies on real-world Chicago crime data show that AIST outperforms the baseline models in terms of prediction accuracy and we can exploit attention weights associated with different parts of the model to interpret its prediction. On average, AIST shows a decrease of $8.3$\% on MAE and $20.98$\% on MSE over the state-of-the-art for crime prediction tasks.
AIST also outperforms the high-performance spatio-temporal models~\cite{DBLP:conf/ijcai/LiangKZYZ18, DBLP:conf/ijcai/YuYZ18, DBLP:journals/TKDE/9139357} developed for solving different domain of tasks (e.g., geo-sensory time series, traffic or crowd flow prediction). On average, AIST shows a decrease of $4.1\%$ on MAE and $7.45\%$ on MSE, when we customize these models for the crime prediction task.
Though we evaluate AIST for the crime prediction problem, AIST has the ability to learn an arbitrary function over the spatio-temporal-semantic space and can be adapted for any other spatio-temporal problem (e.g. traffic, citywide passenger demand, taxi demand prediction) that can benefit from incorporating semantically relevant information and knowing the interpretation of the prediction.
\bibliographystyle{ACM-Reference-Format}
| 2f13139a843897312b6d0c8a1f71abdd8bd6abb1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec.Introduction}
Age-of-Information (AoI) is a performance metric that captures the freshness of the information from the perspective of the destination. AoI measures the time that elapsed since the generation of the packet that was most recently delivered to the destination. This performance metric has been receiving attention in the literature \cite{Book_AoI_17,Book19,yates2020age} for its application in communication systems that carry time-sensitive data. In this paper, we consider a network with $M$ sources transmitting time-sensitive information to the destination over $N$ unreliable wireless channels, as illustrated in Fig.~\ref{fig.Network}.
Packets from each source are generated according to an i.i.d.\ stochastic process with known statistics and the state of each wireless channel (ON/OFF) varies according to an i.i.d.\ stochastic process with \emph{unknown statistics}. At every time slot, the learning algorithm schedules a single pair (source, channel) and the selected source attempts to transmit its packet via the selected wireless channel. When a packet with fresh information is successfully transmitted to the destination, the AoI associated with the selected source is reduced. The goal of the scheduler is to keep the information associated with every source in the network as fresh as possible, i.e.\ to minimize the AoI in the network. To decide which pair to select in a time slot, the scheduler takes into account: i) the packet generation processes at the $M$ sources; ii) the current values of AoI at the destination; and iii) the estimated reliability of the $N$ wireless channels.
In this sequential decision problem, the outcomes of previous transmission attempts are used to estimate the reliability of the wireless channels. This statistical learning problem is closely related to the stochastic multi-armed bandit (MAB) problem in which the wireless channels are the bandits that give i.i.d.\ rewards and the scheduler is the player that attempts to learn the statistics of the bandits in order to maximize the reward accumulated over time. The main challenge in the stochastic MAB problem is to strike a balance between exploiting the bandit that gave the highest rewards in the past and exploring other bandits that may give high rewards in the future. To evaluate the performance of different learning algorithms, we define regret. Regret is the difference between the expected cumulative reward of a \emph{genie algorithm} (that knows the statistics of the bandits a priori) and the expected cumulative reward of the \emph{learning algorithm} under consideration. The regret captures the penalty incurred by having to learn the statistics of the bandits over time. Some well-known order-optimal learning algorithms in terms of regret are: $\epsilon$-Greedy, Upper Confidence Bound (UCB), and Thompson Sampling (TS). The regret of these policies was shown to increase no more than logarithmically in time \cite{FiniteMAB,KL-UCB,TS_MAB}, $O(\log T)$, and this bound was shown to be tight \cite{Assympt}.
We refer to our problem as the \emph{Aging Bandit problem}. An important distinction between the stochastic MAB problem and the Aging Bandit problem is the reward structure. In the stochastic MAB problem, the player selects a bandit in each time slot and receives a reward that is i.i.d.\ over time and depends only on the probability distribution associated with the selected bandit. In the Aging Bandit problem, the scheduler selects a pair (source, channel) and the reward is the AoI reduction that results from a packet transmission to the destination. This reward depends on the state of the selected channel (which is i.i.d.\ over time), since a failed transmission gives zero reward, and it also depends on the history of previous packet deliveries and packet generations. In particular, if the selected source has recently delivered a fresh information update to the destination, then the reduction in AoI may be small. In contrast, if the selected source has not updated the destination for a long period, then the AoI reduction may be large. The reward structure of Aging Bandits is closely related to the AoI evolution (formally defined in Sec.~\ref{sec.Model}) which is history-dependent. This intricate reward structure has significant impact on the analysis of regret and on the development of learning algorithms when compared to the analysis of the traditional stochastic MAB.
The literature on MAB problems is vast, dating more than eight decades \cite{Thompson33}. For surveys on different types of MAB problems, we refer the readers to \cite{surveyMAB1,RMAB_book,surveyMAB2,surveyMAB3}. Most relevant to this work are \cite{RegretQueueing,RegretThomas,ThomasPHD,igorTON18,igorINFOCOM,YuPin18,YuPinTMC,WhittleCSMA,WhittleVishrant,RegretAoI,CorrelatedRegretAoI,DecentralizedRegretAoI}. The authors in \cite{RegretQueueing,RegretThomas,ThomasPHD} considered the problem of minimizing the expected queue-length in a system with a single queue and multiple servers with unknown service rates. In \cite{RegretQueueing}, the authors introduced the concept of queue-length regret, developed a learning algorithm inspired by Thompson Sampling, and analyzed its regret. In \cite{RegretThomas,ThomasPHD}, the authors used information particular to the queue evolution to develop a learning algorithm with $O(1)$ queue-length regret.
The authors in \cite{igorTON18,igorINFOCOM,YuPin18,YuPinTMC,WhittleCSMA,WhittleVishrant,RegretAoI,CorrelatedRegretAoI,DecentralizedRegretAoI} considered the problem of minimizing the average AoI in a single-hop wireless network with unreliable channels. In \cite{igorTON18,igorINFOCOM,YuPin18,YuPinTMC,WhittleCSMA,WhittleVishrant}, the authors posed the AoI minimization problem in a network with multiple sources and \emph{known channel statistics} as a restless MAB problem, developed the associated Whittle’s Index scheduling policy, and evaluated its performance in terms of the average AoI. In \cite{RegretAoI}, the authors considered the AoI minimization problem in a network with a single source-destination pair and unknown channel statistics, introduced the concept of AoI regret, and showed that the AoI regret of UCB and TS scale as $O(\log T)$. In \cite{CorrelatedRegretAoI}, the authors obtained similar results as in \cite{RegretAoI} for the more challenging case of correlated wireless channels. In \cite{DecentralizedRegretAoI}, the authors considered the AoI minimization problem in a network with multiple sources that generate and transmit fresh packets at every time slot through (possibly) different channels with unknown statistics. The authors in \cite{DecentralizedRegretAoI} showed that the AoI regret of a UCB-based distributed learning algorithm scales as $O(\log^2 T)$. An important modelling assumption common to \cite{RegretAoI,CorrelatedRegretAoI,DecentralizedRegretAoI} is that sources generate and transmit fresh packets at every time slot. The more realistic assumptions of random packet generation and scheduled transmissions have significant impact on the AoI evolution, on the analysis of AoI regret, and on the development of learning algorithms. For example, in Sec.~\ref{sec.Policy}, we leverage the random packet generation to develop a learning algorithm with $O(1)$ AoI regret.
In this paper, we study learning algorithms that attempt to minimize AoI in a network with multiple sources generating packets according to stochastic processes and transmitting these packets to the destination over wireless channels with initially unknown statistics. At every time slot, the learning algorithm schedules a single pair (source, channel) and the selected source attempts to transmit a packet through the selected channel. Note that the source policy, which selects a source at each time slot, and the channel policy, which selects the channel to be used in each time slot, can be naturally decoupled, as the optimal channel is independent of the source selected.
In this paper, we focus on the exploration-exploitation dilemma faced by the channel policy. In particular, we consider learning algorithms employing the \emph{optimal source policy} and \emph{different channel policies}. Our main contributions include:
\begin{itemize}
\item we analyze the performance of channel policies based on traditional MAB algorithms including $\epsilon$-Greedy, UCB, and TS, and show that their AoI regret scales as $\Theta(\log T)$. These results generalize the analysis in \cite{RegretAoI} to networks with multiple sources generating packets randomly. The analysis of the AoI regret is more challenging in this network setting since the AoI evolution depends on both the source policy and the stochastic packet generation process. These challenges are discussed in Sec.~\ref{sec.Regret};
\item we develop a novel learning algorithm and establish that it has $O(1)$ AoI regret. The key insight is that when packets are generated randomly, the learning algorithm can utilize times when the network has no packets to transmit, in order to learn the statistics of the channel. To the best of our knowledge, this is the first learning algorithm with bounded AoI regret.
\end{itemize}
The remainder of this paper is outlined as follows. In Sec.~\ref{sec.Model}, the network model and performance metrics are formally presented. In Sec.~\ref{sec.Regret}, we analyze the AoI regret of traditional learning algorithms. In Sec.~\ref{sec.Policy}, we develop an order-optimal learning algorithm and analyze its AoI regret. In Sec.~\ref{sec.Simulations}, we compare the AoI regret of different learning algorithms using simulations. The paper is concluded in Sec.~\ref{sec.Conclusion}. Some of the technical proofs have been omitted due to the space constraint, and will be made available in a technical report.
\section{System Model}\label{sec.Model}
Consider a single-hop wireless network with $M$ sources, $N$ channels and a single destination, as illustrated in Fig.~\ref{fig.Network}.
Each source generates packets containing time-sensitive information and these packets are to be transmitted to the destination through one of the wireless channels. Let the time be slotted, with slot index $t\in\{1,2,\cdots,T\}$, where $T$ is the time horizon of this discrete-time system. The slot duration allows for a single packet transmission. We normalize the slot duration to unity.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\columnwidth]{Network.png}
\end{center}
\vspace{-0.4cm}
\caption{Illustration of the wireless network with $M$ sources, $N$ channels, and a destination.}\label{fig.Network}
\vspace{-0.4cm}
\end{figure}
At the beginning of every slot $t$, each source generates a packet with probability $\lambda\in(0,1)$. Let $a_m(t)\in\{0,1\}$ be the indicator function that is equal to $1$ when source $m\in\{1,2,\cdots,M\}$ generates a packet in slot $t$, and $a_m(t)=0$ otherwise. This Bernoulli process with parameter $\lambda$ is i.i.d.\ over time and independent across different sources, with $\mathrm{P}\left(a_m(t)=1\right)=\lambda,\forall m,t$. A packet that is generated in slot $t$ can be transmitted during the same slot $t$. We denote the vector of packet generations in slot $t$ by $\vec{a}(t) = \left[ a_1(t)\ \cdots\ a_M(t) \right]^\mathsf{T}$.
Each source has a transmission queue to store its packets. Sources keep \emph{only} the most recently generated packet, i.e.\ the freshest packet, in their queue. When source $m$ generates a new packet at the beginning of slot $t$, older packets (if any) are discarded from its queue. Notice that delivering the most recently generated packet provides the freshest information to the destination. This queueing discipline is known to optimize the AoI in a variety of contexts \cite{AoI_management,AoI_LIFO,AoI_LGFS19_2}. After a packet delivery from source $m$, the queue remains empty until the next packet generation from the same source. However, while the queue is empty, a \emph{dummy packet} can be transmitted for the purpose of probing the channels.
The networked system is empty during slot $t$ if there are no data packets available for transmission, i.e.\ if the $M$ queues are empty. Let $E(t)\in\{0,1\}$ be the indicator function that is equal to $1$ if the system is empty during slot $t$, and $E(t)=0$ otherwise. Notice that if there is a packet generation at the beginning of slot $t$, then the system is nonempty during slot $t$ and $E(t)=0$. Recall that when the system is empty, sources can still transmit dummy packets.
In a slot, the learning algorithm selects a single pair $(m,n)$, where $m\in\{1,2,\cdots,M\}$ is the index of the source and $n\in\{1,2,\cdots,N\}$ is the index of the wireless channel. Then, during this slot, source $m$ transmits a packet to the destination through channel $n$. If channel $n$ is ON, then the packet is successfully transmitted to the destination, and if channel $n$ is OFF, then the transmission fails. The learning algorithm does not know the channel states while making scheduling decisions, and the outcome of a transmission attempt during slot $t$ is known at the beginning of slot $t+1$. Let $b_n(t)\in\{0,1\}$ be the indicator function that represents the state of channel $n$ during slot $t$. The channel is ON, $b_n(t)=1$, with probability $\mu_n\in(0,1]$, and the channel is OFF, $b_n(t)=0$, with probability $1-\mu_n$. The channel state process is i.i.d.\ over time and independent across different channels.
The \emph{reliability of channel} $n$ is represented by the probability of this channel being ON, $\mu_n$. Let $\vec\mu = \left[ \mu_1\ \cdots\ \mu_N \right]^\mathsf{T}$ be the vector of channel reliabilities.
Let $\mu^*$ be the maximum channel reliability and let $n^*$ be the index of the corresponding channel, i.e.\,
$\mu^* = \max_{n} \mu_n = \mu_{n^*}$.
For simplicity, we assume that the optimal channel $n^*$ is unique. Naturally, if the channel reliabilities were known by the learning algorithm in advance, then the algorithm would select channel $n^*$ in every slot $t$. However, since the channel reliabilities $\vec\mu$ are initially unknown, the learning algorithm has to estimate $\mu_n$ using observations from previous transmission attempts, while at the same time attempting to minimize the AoI in the network.
Next, we formulate the AoI minimization problem.
\subsection{Age of Information}
The AoI captures how old the information is from the perspective of the destination. Let $h_m(t)$ be a positive integer that represents the AoI associated with source $m$ at the beginning of slot $t$. By definition, we have $h_m(t):=t-\tau_m(t)$, where $\tau_m(t)$ is the generation time of the latest packet successfully transmitted from source $m$ to the destination\footnote{We define $\tau_m(t)=0$ prior to the first packet delivery from source $m$.}. If the destination does not receive a fresh packet from source $m$ during slot $t$, then in the next slot we have $h_m(t+1)=h_m(t)+1$, since the information at the destination is one slot older. In contrast, if the destination receives a fresh packet from source $m$ during slot $t$, then in the next slot the value of $\tau_m(t+1)$ is updated to the generation time of the received packet and the AoI is reduced by $\tau_m(t+1)-\tau_m(t)$. This difference is the ``freshness gain'' associated with the received packet. The evolution of $h_m(t)$ over time is illustrated in Fig.~\ref{fig.AoIevolution}. We define the vector of AoI in slot $t$ as $\vec{h}(t) = \left[ h_1(t)\ \cdots\ h_M(t) \right]^\mathsf{T}$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\columnwidth]{AoI_evolution.png}
\end{center}
\vspace{-0.3cm}
\caption{The blue and orange rectangles at the bottom represent packets generated at source $m$ and successful packet transmissions from source $m$, respectively. The orange curve shows the AoI evolution $h_m(t)$ associated with source $m$.} \label{fig.AoIevolution}
\vspace{-0.4cm}
\end{figure}
For capturing the information freshness of the entire network, we consider the \emph{expected total AoI} $\bar h(T)$, which is defined as the expected sum of the AoI over all sources and over time, namely
\begin{equation}\label{eq.totalAge}
\bar h(T) = \mathrm{E}\left[\sum_{m=1}^{M} \sum_{t=1}^{T} h_m(t)\right] \; ,
\end{equation}
where the expectation is with respect to the randomness in the channel states $b_n(t)$, packet generation process $\vec a(t)$, and scheduling decisions $(m,n)$.
The learning algorithm schedules pairs $(m,n)$ over time so as to minimize the expected total AoI $\bar h(T)$. Recall that in this sequential decision problem, the channel reliabilities $\mu_n$ are initially unknown by the learning algorithm and should be estimated over time. Next, we discuss the class of learning algorithms considered in this paper.
\subsection{Learning Algorithm}
In this section, we present three important concepts associated with the learning algorithm: the channel policy, the source policy, and the AoI regret. Prior to discussing these concepts, we introduce some notation.
In each slot $t$, the learning algorithm selects a single source and a single channel. Let $m(t)$ be the index of the source selected during slot $t$ and let $n(t)$ be the index of the channel selected during slot $t$. Then, the pair selected in each slot can be denoted as $(m(t),n(t))$. Notice that the learning algorithm can be divided into two components: the source policy, which selects $m(t)$, and the channel policy, which selects $n(t)$. Let $b(t)=b_{n(t)}(t)$ be the state of the channel selected during slot $t$, and recall that $\vec{a}(t)$ is the vector of packet generations and $\vec{h}(t)$ is the vector of AoI in slot $t$. Using this notation, we define the \emph{channel policy} and the \emph{source policy}.
The \emph{channel policy} may (or may not) take into account the status of the transmission queues at the sources (in particular $E(t)$) in making scheduling decisions $n(t)$. Hence, we define two types of channel policies: queue-independent channel policies and queue-dependent channel policies. Let $\Pi_B$ be the class of admissible \emph{queue-independent channel policies} $\pi_b$. In slot $t$, an arbitrary policy $\pi_b\in\Pi_B$ selects $n(t)$ using information about the outcome of previous transmission attempts. In particular, the queue-independent channel history in slot $t$ is given by $H_B(t) = \{ n(1),b(1),\cdots,n(t-1),b(t-1) \}$. Let $\bar\Pi_B$ be the class of admissible \emph{queue-dependent channel policies} $\bar\pi_b$. In slot $t$, an arbitrary policy $\bar\pi_b\in\bar\Pi_B$ selects $n(t)$ using information about the outcome of previous transmission attempts and about the current status of the transmission queues. In particular, the queue-dependent channel history in slot $t$ is given by $\bar{H}_B(t) = H_B(t)\cup\{E(t)\}$. In Sec.~\ref{sec.Policy}, we show that this small amount of information, namely $E(t)$, can have a significant impact on the performance of the channel policy.
It is easy to see that both the optimal queue-independent channel policy $\pi^*_b$ and the optimal queue-dependent channel policy $\bar \pi^*_b$ select the channel with highest reliability $\mu^*$ at every slot $t$. However, since the reliabilities $\vec\mu$ are not known a priori, the channel policies have to estimate $\vec\mu$ over time. In Sec.~\ref{sec.Regret}, we consider queue-independent channel policies and in Sec.~\ref{sec.Policy}, we consider queue-dependent channel policies.
The \emph{source policies} considered in this paper are work-conserving, i.e.\, policies that never transmit dummy packets when there are undelivered data packets in the system. Let $\Pi_A$ be the class of admissible work-conserving source policies $\pi_a$. In slot $t$, an arbitrary source policy $\pi_a\in\Pi_A$ selects $m(t)$ using information about the current AoI and the generation times of the packets waiting to be transmitted at the sources' queues. In particular, the source history in slot $t$ is given by $H_A(t) = \{ \vec{a}(1),\vec{h}(1),\cdots,\vec{a}(t),\vec{h}(t)\}$.
The optimal source policy $\pi^*_a\in\Pi_A$ is the transmission scheduling policy that minimizes the expected total AoI in \eqref{eq.totalAge}. A few works in the literature \cite{YuPin18,YuPinTMC,igorMobiHoc,igorTMC19} have addressed the problem of finding the transmission scheduling policy that minimizes AoI in wireless networks with stochastic packet generation and unreliable channels with \emph{known statistics}. Despite those efforts, a full characterization of the optimal source policy is still an open problem.
In this paper, we consider learning algorithms $\pi$ that are a composition of a source policy and a channel policy $\pi=(\pi_a,\pi_b)$. Our goal is to study the exploration-exploitation dilemma faced by the channel policy. To that end, we analyze the AoI regret of learning algorithms employing the optimal source policy and different channel policies. To analyze the AoI regret of learning algorithms without the full characterization of the optimal source policy $\pi_a^*$, we derive lower and upper bounds on the regret.
These bounds are discussed in Proposition~\ref{p1}, Proposition~\ref{t2}, and Theorem~\ref{t3}, where
we assumed that the optimal source policy $\pi_a^*$
is the same irrespective of the queue-independent channel policy $\pi_b$ under consideration, namely
\begin{equation}\label{eq.independenceAssumption}
\pi^*_a = \argmin_{\pi_a\in\Pi_A} \mathrm{E}\left[ \sum_{m=1}^{M} \sum_{t=1}^{T} h^{(\pi_a,\pi_b)}_m(t) \right] \; , \; \forall\pi_b\in\Pi_B \; ,
\end{equation}
where $h^{(\pi_a,\pi_b)}_m(t)$ denotes the AoI associated with source $m$ in slot $t$ when the learning algorithm $\pi=(\pi_a,\pi_b)$ is employed. An analogous assumption is utilized for the case of queue-dependent channel policies $\bar\pi_b\in\bar\Pi_B$.
The \emph{AoI regret} of a learning algorithm $\pi$ with queue-independent channel policy $\pi_b$ is defined as the difference between the expected total AoI $\bar h^\pi (T)$ when $\pi=(\pi_a,\pi_b)$ is employed and the expected total AoI $\bar h^* (T)$ when the optimal algorithm $\pi^*=(\pi_a^*,\pi_b^*)$ is employed, namely
\begin{equation}\label{eq.AoIregret}
R^\pi(T) = \mathrm{E}\left[ \sum_{m=1}^{M} \sum_{t=1}^{T} h^\pi_m(t) - \sum_{m=1}^{M} \sum_{t=1}^{T} h^*_m(t) \right] \; ,
\end{equation}
where the expectation is with respect to the randomness in the channel states $b(t)$, packet generation process $\vec a(t)$, and scheduling decisions $(m(t),n(t))$. The definition of AoI regret for a learning algorithm $\bar\pi$ with queue-dependent channel policy $\bar\pi_b$ is analogous to \eqref{eq.AoIregret}.
Next, we analyze the AoI regret of learning algorithms with \emph{queue-independent channel policies}.
\section{Regret Analysis}\label{sec.Regret}
The problem of learning channel reliabilities over time is closely related to the stochastic MAB problem. A natural class of channel policies to consider are traditional MAB algorithms such as $\epsilon$-Greedy, UCB, and TS.
In this section, we derive bounds on the AoI regret of learning algorithms that employ \emph{queue-independent channel policies}. Notice that the class of queue-independent channel policies $\Pi_B$ includes traditional MAB algorithms. We describe a learning algorithm employing TS as its channel policy in Algorithm~\ref{alg.TS}.
\begin{algorithm}[t]
\SetAlgoLined
Initialization: time $t=1$, estimates $\hat\mu_n=0$, counters $T_n=0$, parameters $\alpha_n=\beta_n=1$, $\forall n\in\{1,\cdots,N\}$\;
\While{$1\le t\le T$}{
Optimal source policy selects $m\in\{1,2,\cdots,M\}$\;
$\theta_n \sim \text{Beta}(\alpha_n,\beta_n)$\;
$n=\argmax_{n'\in\{1,\cdots,N\}}\theta_{n'}$\;
Source $m$ transmits packet through channel $n$ and observes channel state $b$\;
\eIf{$b=1$}{
$\alpha_n = \alpha_n + 1$\;
}{
$\beta_n = \beta_n + 1$\;
}
Compute new estimate $\hat\mu_n = \dfrac{\hat\mu_n T_n + b}{T_n+1}$\;
$T_n=T_n+1$\;
$t=t+1$\;
}
\caption{Learning Algorithm employing TS as its channel policy}\label{alg.TS}
\end{algorithm}
Scheduling decisions of a learning algorithm $\pi$ might differ from those of $\pi^*$ both in the source and in the channel, which makes the analysis of the AoI regret $\sum_{t=1}^T \sum_{m=1}^M \mathrm{E}[ h_m^\pi(t)-h_m^{*}(t) ]$ challenging. To alleviate this challenge, we use stochastic coupling to create \emph{equivalent coupled channel state processes} that are simpler to analyze. Similar coupling arguments were employed in \cite{RegretQueueing,RegretAoI}.
\begin{remark}[Coupled Channel States]\label{r1}
Let $\{U(t)\}_{t=1}^T$ be a sequence of i.i.d.\ random variables uniformly distributed in the interval $[0,1]$. In each slot $t$, the channel states $b_n(t)$ are determined as follows
\begin{equation}\label{coupling}
b_n(t) = 1 \iff 0\le U(t) \le \mu_n \; , \; \forall n \; .
\end{equation}
\end{remark}
By construction, the coupled channel states are no longer independent. In particular, if a channel is ON during slot $t$, then all channels with higher reliability $\mu_n$ are also ON during that slot. Notice that, in each slot $t$, each coupled channel $n$ has the same probability distribution as the associated original channel~$n$, namely $\mathrm{P}\left(b_n(t)=1\right)=\mu_n,\forall n,t$.
Hence, given the scheduling decision $(m(t),n(t))$ of $\pi$ during any slot $t$, the probability of a successful transmission attempt from source $m(t)$ through channel $n(t)$ is the same for both the coupled and original channel states. It follows that
the probability distribution of $h_m^\pi(t)$ also remains the same for all slots $t$ and for all sources $m$ and, thus, the AoI regret $R^\pi(T)$ in \eqref{eq.AoIregret} also remains the same for both the coupled and original channel state processes.
\emph{For simplicity of analysis, henceforth in this paper, we assume that the channel state processes are coupled as described in Remark~\ref{r1}}.
In Proposition~\ref{p1}, Proposition~\ref{t2}, and Corollary~\ref{c3}, we derive bounds on the AoI regret of a learning algorithm $\pi$ with respect to its \emph{expected number of suboptimal channel choices}, namely
\begin{equation}\label{eq.suboptimalChoices}
\mathrm{E}[K^\pi(T)] = \mathrm{E}\left[\sum_{t=1}^T\mathbbm{1}\left\{n^\pi(t) \neq n^*\right\}\right] \; ,
\end{equation}
where $\mathbbm{1}\left\{n^\pi(t) \neq n^*\right\}=1$ if $n^\pi(t) \neq n^*$, and $\mathbbm{1}\left\{n^\pi(t) \neq n^*\right\}=0$ otherwise.
We consider two classes of admissible learning algorithms
\begin{align}
\Pi = \left\{ \pi=(\pi_a,\pi_b) : \pi_a\in\Pi_A , \pi_b\in\Pi_B \right\} \; ; \label{eq.Pi}\\
\Pi^* = \left\{ \pi=(\pi_a,\pi_b) : \pi_a = \pi^*_a , \pi_b\in\Pi_B \right\} \; . \label{eq.Pi_star}
\end{align}
Both classes employ queue-independent channel policies. The difference is that $\Pi$ employs any admissible source policy $\pi_a\in\Pi_A$, while $\Pi^*$ employs the optimal source policy $\pi_a^*$. Naturally, we have $\Pi^*\subset\Pi$.
\begin{proposition}[Lower Bound]\label{p1}
For any given network configuration $(\lambda,\vec\mu)$, the AoI regret of any learning algorithm $\pi\in\Pi$ scales at least on the order of its expected number of suboptimal channel choices, namely\footnote{$f(t) = \Omega(g(t)) \iff \exists C>0\ \exists t_0\ \forall t>t_0 : f(t)\ge C\cdot g(n)$}
\begin{equation}\label{eq.p1}
R^\pi(T) = \Omega\left( \mathrm{E}\left[ K^{\pi}(T) \right] \right) \; .
\end{equation}
\end{proposition}
\noindent\emph{Proof outline.} In addition to the suboptimal channel choices, source choices $m^\pi(t)$ of algorithm $\pi\in\Pi$ can also differ from the source choices $m^{*}(t)$ of $\pi^*$. To overcome this challenge, we construct an auxiliary algorithm $\hat\pi^*$ with optimal channel policy
and a source policy that selects the same source\footnote{Notice that if the selected source $m^{\pi}(t)$ has no packet in its transmission queue, then the auxiliary algorithm attempts to transmit a dummy packet.} $m^{\pi}(t)$ as $\pi$ in every slot $t$. Then, we focus on the \emph{auxiliary AoI regret} $\sum_{t=1}^T \sum_{m=1}^M \mathrm{E}[ h_m^\pi(t)-h_m^{\hat\pi^*}(t)]$ associated with the auxiliary algorithm $\hat\pi^*$, which we show to be not greater than the original AoI regret $\sum_{t=1}^T \sum_{m=1}^M \mathrm{E}[ h_m^\pi(t)-h_m^{*}(t) ]$. We then observe that each suboptimal channel choice of $\pi$ results in a penalty to the auxiliary AoI regret, and we show that this penalty is lower bounded by a constant. Using this constant, we obtain the desired lower bound on the original AoI regret in \eqref{eq.p1}.
The details are omitted due to the space constraint.
%
%
%
\begin{proposition}[Upper Bound]\label{t2}
For any given network configuration $(\lambda,\vec\mu)$, the AoI regret of any learning algorithm $\pi\in\Pi^*$ scales at most on the order of its expected number of suboptimal channel choices, namely\footnote{$f(t) = O(g(t)) \iff \exists C>0\ \exists t_0\ \forall t>t_0 : f(t)\le C\cdot g(n)$}
\begin{equation}\label{eq.t2}
R^\pi(T) = O\left( \mathrm{E}\left[ K^{\pi}(T) \right] \right) \; .
\end{equation}
\end{proposition}
\noindent\emph{Proof outline}.
Despite the fact that both learning algorithms $\pi\in\Pi^*$ and $\pi^*$ employ the same optimal source policy $\pi_a^*$, they might select different sources $m^{\pi}(t) \neq m^{*}(t)$ over time, due to their different channel policies.
To address this challenge, we use an approach similar to the proof of Proposition~\ref{p1}. We construct an auxiliary algorithm $\hat\pi\in\Pi^*$
with a source policy that selects the same source $m^{*}(t)$ as $\pi^*$ in every slot $t$, and with a channel policy that selects the same channel $n^{\pi}(t)$ as $\pi$ in every slot $t$. Then, we show that the auxiliary AoI regret $\sum_{t=1}^T \sum_{m=1}^M \mathrm{E}[ h_m^{\hat\pi}(t) - h_m^{*}(t) ]$ associated with the auxiliary algorithm $\hat\pi$ is not lower than the original AoI regret $\sum_{t=1}^T \sum_{m=1}^M \mathrm{E}[ h_m^\pi(t)-h_m^{*}(t) ]$. To derive an upper bound on the auxiliary AoI regret, we analyze the penalty that results from each suboptimal channel choice of $\hat\pi$.
During a slot $t$ where $\hat\pi$ makes a suboptimal channel choice, if channel $n^{\hat\pi}(t)$ is OFF and channel $n^{*}$ is ON, then a \emph{discrepancy} is added to the difference between the AoI of $\hat\pi$ and the AoI of $\pi^*$, i.e.\, $h_m^{\hat\pi}(t+1) - h_m^{*}(t+1)>h_m^{\hat\pi}(t) - h_m^{*}(t)$.
This discrepancy lasts until the next successful transmission of a packet from source $m$ by the auxiliary algorithm $\hat\pi$, after which the values of $h_m^{\hat\pi}(\cdot)$ and $h_m^{*}(\cdot)$ become equal\footnote{Recall from Remark~\ref{r1} that channel states are coupled. Hence, if channel $n^{\hat\pi}(t)$ is ON, then channel $n^{*}$ is also ON.}. We refer to the duration of the discrepancy as its \emph{length}. The penalty that results from a suboptimal channel choice is the product of the discrepancy and its length. We characterize the auxiliary AoI regret by expressing it as the sum of the penalties arising from suboptimal channel choices. Then, using discrete phase-type distributions, we upper bound the discrepancies and the lengths by constants (in the expected sense) to obtain the result in \eqref{eq.t2}.
The details are omitted due to the space constraint.
\begin{corollary}\label{c3}
For any given network configuration $(\lambda,\vec\mu)$, the AoI regret of any learning algorithm $\pi\in\Pi^*$ scales with its expected number of suboptimal channel choices, namely\footnote{$f(t) = \Theta(g(t)) \iff f(t) = O(g(t)) \land f(t) = \Omega(g(t)) \iff \exists C_1,C_2>0\ \exists t_0\ \forall t>t_0 : C_1\cdot g(n) \le f(t)\le C_2\cdot g(n)$}
\begin{equation}
R^\pi(T) = \Theta\left( \mathrm{E}\left[ K^{\pi}(T) \right] \right) \; .
\end{equation}
\end{corollary}
Corollary~\ref{c3} follows directly from Propositions~\ref{p1} and ~\ref{t2}. Notice that the bounds in Proposition~\ref{t2} and Corollary~\ref{c3} are not valid for the broader class of learning algorithms $\Pi$ which includes suboptimal source policies. This is because suboptimal source choices may add to the AoI regret, possibly making it grow faster than $\mathrm{E}\left[ K^{\pi}(T) \right]$.
Prior to analyzing the AoI regret of learning algorithms that employ $\epsilon$-Greedy, UCB, and TS as their channel policy, we define $\alpha$\emph{-consistent learning algorithms} \cite{surveyMAB3,RegretQueueing} and discuss a few of their properties. Let $\mathrm{E}[T^\pi_n(T)]$ be the expected number of times channel $n$ is selected by $\pi\in\Pi$ in the first $T$ slots, namely
\begin{equation}\label{eq.channelChoices}
\mathrm{E}[T_n^\pi(T)] = \mathrm{E}\left[\sum_{t=1}^T\mathbbm{1}\left\{n^\pi(t) = n\right\}\right] \; .
\end{equation}
\begin{definition}[$\alpha$-consistency]\label{def.alphaConsistency}
For a given $\alpha\in(0,1)$, a learning algorithm $\pi\in\Pi$ is classified as \emph{$\alpha$-consistent} if, for any network configuration $(\lambda,\vec\mu)$, we have
$E\left[ T^\pi_n(T) \right] = O(T^\alpha)$
for \emph{all suboptimal channels} $n\neq n^*$.
\end{definition}
Intuitively, a learning algorithm $\pi\in\Pi$ is $\alpha$-consistent if its channel policy has good performance in \emph{every network configuration}. Consider a learning algorithm with a trivial channel policy that selects $n(t)=1$ in every slot $t$. In network configurations with $n^*=1$, this channel policy never selects suboptimal channels, i.e. $\mathrm{E}\left[ T^\pi_n(T) \right] = O(T^\alpha),\forall n\neq n^*$. However, in network settings with $n^*\neq 1$, this channel policy is such that $\mathrm{E}\left[ T^\pi_1(T) \right] = T$, which violates the definition of $\alpha$-consistency. In the remainder of this section, we focus on channel policies that have good performance in every network configuration. In particular, we analyze the AoI regret of $\alpha$-consistent learning algorithms with queue-independent channel policies.
\begin{remark}[AoI regret of $\alpha$-consistent algorithms]\label{r2}
In \cite[Corollary~$20$]{RegretQueueing}, the authors show that any learning algorithm $\pi\in\Pi$ that is $\alpha$-consistent has an expected number of suboptimal channel choices that scales as
$E\left[ K^{\pi}(T) \right] = \Omega(\log T)$,
for any network configuration $(\lambda,\vec\mu)$. Hence, it follows from the lower bound in Proposition~\ref{p1} that the associated AoI regret scales as
\begin{equation}\label{eq.alphaOmega}
R^\pi(T) = \Omega(\log T) \; ,
\end{equation}
for any network configuration $(\lambda,\vec\mu)$.
\end{remark}
Notice that the lower bound in Remark~\ref{r2} applies to $\alpha$-consistent learning algorithms with queue-independent channel policies that do not know the statistics of the channels in advance.
Learning algorithms that employ $\epsilon$-Greedy, UCB, and TS as their channel policy are known to have suboptimal channel choices scaling as $\mathrm{E}[K^\pi(T)] = O(\log T)$ for any network configuration $(\lambda,\vec\mu)$ \cite{FiniteMAB,furtherTS}, which implies that they are $\alpha$-consistent. Hence, it follows from the upper bound in Proposition~\ref{t2} and from \eqref{eq.alphaOmega} that the AoI regret of these learning algorithms scale as
\begin{equation}
R^\pi(T) = \Theta(\log T) \; .
\end{equation}
In \cite{RegretAoI}, the authors derived lower and upper bounds on the AoI regret of learning algorithms employing queue-independent channel policies, including UCB and TS, in networks with a single source generating and transmitting fresh packets in every slot $t$. Propositions~\ref{p1} and \ref{t2} generalize the results in \cite{RegretAoI} to networks with multiple sources generating packets according to stochastic processes. The analysis of the AoI regret is more challenging in this network setting for the following reasons:
i) the optimal source policy $\pi^*_a$ is unknown and there is no closed-form expression for the expected total AoI \eqref{eq.totalAge} of the optimal algorithm $\pi^*=(\pi_a^*,\pi_b^*)$; and
ii) the learning algorithm under consideration $\pi=(\pi_a,\pi_b)$ can make suboptimal choices both in terms of sources $m(t)$ and channels $n(t)$, and these two types of suboptimal choices affect the AoI regret $R^\pi(T)$ differently.
Next, we develop a learning algorithm that leverages information about the status of the transmission queues in making scheduling decisions $n(t)$, and show that this new learning algorithm has $O(1)$ AoI regret.
\section{Order-Optimal Learning Algorithm}\label{sec.Policy}
In this section, we develop a learning algorithm $\bar\eta\in\bar\Pi$ with a \emph{queue-dependent channel policy} that selects $n(t)$ using information about the outcome of previous transmission attempts, namely
$H_B(t) = \{ n(1),b(1),\cdots,n(t-1),b(t-1) \}$,
and about the current status of the transmission queues, $E(t)$. Then, we derive an upper bound on its AoI regret. In particular, we show that the AoI regret of $\bar\eta$ is such that $R^{\bar\eta}(T)=O(1)$. Notice that the only difference between the learning algorithms $\pi\in\Pi$ in Sec.~\ref{sec.Regret} and the order-optimal learning algorithm $\bar\eta$ is the knowledge of $E(t)$. This seemingly modest addition led to the reduction of the AoI regret from $R^{\pi}(T)=\Omega(\log T)$ to $R^{\bar\eta}(T)=O(1)$. To the best of our knowledge, this is the first learning algorithm with bounded AoI regret.
The key insight is that when packets are generated randomly, the learning algorithm $\bar\eta$ can utilize times when the network has no data packets to transmit, i.e.\ when $E(t)=1$, to transmit dummy packets and learn the statistics of the channels without incurring an opportunity cost. The order-optimal learning algorithm $\bar\eta=(\eta_a,\bar\eta_b)$ has optimal source policy $\eta_a=\pi^*_a$ and a channel policy $\bar\eta_b\in\bar\Pi_B$ that operates as follows: when the system is empty, $E(t)=1$, the policy chooses a channel uniformly at random and uses the outcome of the transmission attempt to update its estimates of the channel reliabilities and, when the system is nonempty, $E(t)=0$, the policy chooses the channel with the current highest estimated reliability. Notice that the channel policy only updates its estimates of the channel reliabilities when the system is empty. A similar channel policy was used in \cite{RegretThomas,ThomasPHD} to develop a learning algorithm with bounded queue-length regret. The order-optimal learning algorithm $\bar\eta$ is described in Algorithm~\ref{alg.Opt}. The upper bound on the AoI regret is established in the theorem that follows.
\begin{algorithm}[t]
\SetAlgoLined
Initialization: time $t=1$, estimates $\hat\mu_n=0$, counters $T_n=0$, $\forall n\in\{1,\cdots,N\}$\;
\While{$1\le t\le T$}{
Optimal source policy selects $m\in\{1,2,\cdots,M\}$\;
\eIf{system is empty}{
$n = \text{Unif}\{1,\cdots,N\}$\;
Source $m$ transmits dummy packet through channel $n$ and observes channel state $b$\;
$\hat\mu_{n} = \dfrac{\hat\mu_{n} T_{n} + b}{T_{n}+1}$\;
$T_{n}=T_{n}+1$\;
}{
$n=\argmax_{n'\in\{1,\cdots,N\}}\hat\mu_{n'}$\;
Source $m$ transmits data packet through channel $n$ and observes channel state $b$\;
}
$t=t+1$\;
}
\caption{Order-Optimal Learning Algorithm}\label{alg.Opt}
\end{algorithm}
\begin{theorem}\label{t3}
For any given network configuration $(\lambda,\vec\mu)$, the AoI regret of the order-optimal learning algorithm $\bar\eta$ is bounded, namely
\begin{equation}\label{eq.t3}
R^{\bar\eta}(T)=O(1) \; .
\end{equation}
\end{theorem}
\begin{proof}
Recall that the two components of the order-optimal learning algorithm are $\bar\eta = (\eta_a,\bar\eta_b)$. Similarly to the proof of Proposition~\ref{t2}, we start by constructing an auxiliary algorithm $\hat\eta = (\hat\eta_a,\bar\eta_b)$ which has a source policy $\hat\eta_a$ that selects the same source $m^{*}(t)$ as $\pi^*$ in every slot $t$. Since $\eta_a = \pi^*_a$ is optimal, it follows that $\hat\eta_a$ is suboptimal, which implies that
\begin{align}
\sum_{m=1}^M \sum_{t=1}^T \mathrm{E}[ h_m^{\bar\eta}(t)-h_m^{*}(t) ] \le \sum_{m=1}^M \sum_{t=1}^T \mathrm{E}[ h_m^{\hat\eta}(t)-h_m^{*}(t) ] \; . \label{aux_regret}
\end{align}
We denote the RHS of \eqref{aux_regret} as the \emph{auxiliary AoI regret}. Prior to deriving the upper bound on the auxiliary AoI regret, we introduce some definitions that are particular to the channel policy $\bar\eta_b$.
Consider the time slots when the system becomes empty, i.e.\ time slots $t$ such that $E(t-1)=0$ and $E(t)=1$. We denote the time interval between two such slots as a \emph{period}
and we divide time $t\in\{1,2,\cdots,T\}$ into successive periods, with period index $p\in\{1,2,\cdots,P\}$. By definition, the system is empty, $E(t)=1$, in the beginning of each period $p$ and it remains empty until the first packet generation. Once the first packet is generated, the system becomes nonempty, $E(t)=0$, and it remains nonempty until the end of the period. Hence, each period $p$ has two phases: an \emph{empty phase} and a \emph{nonempty phase}, with each phase having at least one slot.
Let $s_p$ and $f_p$ be the first and the last slots of period $p$, respectively, with $s_1=1$ and $s_{p+1} = f_p + 1,\forall p$.
Then, the cumulative AoI of source $m$ during period $p$ can be written as
\begin{equation}\label{eq.cumy}
y^{\hat\eta}_m(p) = \sum_{t=s_p}^{f_p} h^{\hat\eta}_m(t) \; .
\end{equation}
Recall from Algorithm~\ref{alg.Opt} that estimates of the channel reliabilities are only updated during empty phases. Within a nonempty phase, the estimates do not change and, thus, the selected channel also does not change. Let $\bar n(p)$ be the channel selected by policy $\bar\eta_b$ during the entire \emph{nonempty phase} of period $p$.
If $\bar n(p)=n^*$, we refer to period $p$ as an \emph{optimal period}. Otherwise, we refer to period $p$ as a \emph{suboptimal period}.
Next, we derive an upper bound on the auxiliary AoI regret in terms of the expected AoI contributions of the suboptimal periods.
\begin{lemma}\label{step1}
The auxiliary AoI regret is upper bounded by
\begin{align}
&\sum_{m=1}^M \sum_{t=1}^T \mathrm{E}[ h_m^{\hat\eta}(t)-h_m^{*}(t)] \nonumber\\
&\quad \le \sum_{m=1}^{M} \sum_{p=1}^{T} \mathrm{E}\left[ y^{\hat\eta}_m(p) \;\middle\vert\; \bar n(p)\neq n^* \right] \mathrm{P}\left(\bar n(p)\neq n^*\right) \; .\label{eqstep1}
\end{align}
\end{lemma}
To establish Lemma~\ref{step1}, we first show that if period $p$ is an optimal period, then $h^{\hat\eta}_m(t) = h^{*}_m(t), \forall m, \forall t \in \{s_p,\cdots,f_p\}$, which implies that optimal periods do not contribute to the auxiliary AoI regret. Then, we obtain the upper bound in \eqref{eqstep1} by manipulating the expression of the auxiliary AoI regret.
The complete proof of Lemma~\ref{step1} can be found in Appendix \ref{newA}.
In Lemmas~\ref{step2} and \ref{step3}, we derive upper bounds on the first and second terms on the RHS of \eqref{eqstep1}, respectively.
\begin{lemma}\label{step2}
There exists a constant $C_y$ such that
\begin{equation}\label{eqstep2}
\mathrm{E}\left[ y^{\hat\eta}_m(p) \;\middle\vert\; \bar n(p)\neq n^* \right] \le C_y \; .
\end{equation}
\end{lemma}
To establish Lemma~\ref{step2}, we first show that the cumulative AoI $y^{\hat\eta}_m(p)$ of source $m$ in period $p$ can be upper bounded by
\begin{align}
y^{\hat\eta}_m(p) &= \sum_{t=s_p}^{f_p} h^{\hat\eta}_m(t) \le \sum_{i=0}^{f_p-s_p} \left( h_m^{\hat\eta}(s_p) + i \right)\nonumber\\
&= h_m^{\hat\eta}(s_p) [f_p - s_p + 1] + \dfrac12 [(f_p - s_p)^2 + f_p - s_p] \; .\label{cumsum}
\end{align}
Then, we derive an upper bound on the conditional expectation of \eqref{cumsum}. In particular, we
show that $h_m^{\hat\eta}(s_p)$ can be upper bounded by a geometric random variable. Then, we show that the random variable $f_p - s_p$, which represents the length of period $p$, follows a discrete phase-type distribution. The upper bound on the conditional expectation of \eqref{cumsum} follows from the fact that the geometric random variable has finite second moment and the phase-type random variable has finite first and second moments.
The details are omitted due to the space constraint.
%
\begin{lemma}\label{step3}
There exists a constant $C_p$ such that
\begin{equation}
\sum_{p=1}^{T} \mathrm{P}\left(\bar n(p)\neq n^*\right) \le C_p \; .\label{eqstep3}
\end{equation}
\end{lemma}
To establish Lemma~\ref{step3}, we use Hoeffding's inequality to upper bound $\mathrm{P}\left(\bar n(p) = n\right)$ by an exponential function of $-p$, for every suboptimal channel $n$. The result in \eqref{eqstep3} follows directly from this upper bound. The complete proof of Lemma~\ref{step3} can be found in Appendix \ref{newC}.
From the upper bound on the AoI regret in \eqref{aux_regret} and the results in Lemmas~\ref{step1}, \ref{step2} and \ref{step3}, we have
\begin{align}
R^{\bar\eta}(T) &\le \sum_{m=1}^M \sum_{t=1}^T \mathrm{E}[ h_m^{\hat\eta}(t)-h_m^{*}(t) ]\nonumber\\
&\le \sum_{m=1}^{M} \sum_{p=1}^{T} \mathrm{E}\left[ y^{\hat\eta}_m(p) \;\middle\vert\; \bar n(p)\neq n^* \right] \mathrm{P}\left(\bar n(p)\neq n^*\right)\nonumber\\
&\le C_y M C_p
\end{align}
which establishes the bound in \eqref{eq.t3}.
\end{proof}
In the particular case of a network with sources generating fresh packets at every slot $t$, i.e.\ $\lambda=1$, the algorithm $\bar\eta$ cannot utilize slots in which the system is empty to learn the channel reliabilities without incurring a cost in terms of AoI regret, which results in a $R^{\bar\eta}(T)$ that grows over time.
The upper bound in Theorem~\ref{t3} is only valid for the network models described in Sec.~\ref{sec.Model}, in which $\lambda\in(0,1)$.
Next, we evaluate the AoI regret of the different learning algorithms discussed in this paper using MATLAB simulations and we propose a heuristic algorithm that leverages the fast learning rates of TS and the bounded regret of the order-optimal algorithm.
\section{Simulations}\label{sec.Simulations}
In this section, we evaluate the performance of learning algorithms in terms of the AoI regret in \eqref{eq.AoIregret}. We compare learning algorithms employing the Age-Based Max-Weight source policy \cite[Sec.~5]{igorTMC19} and different channel policies, namely: i) $\epsilon$-Greedy; ii) UCB; iii) TS; iv) Optimal; and v) Hybrid. The Age-Based Max-Weight source policy selects, in each slot $t$, the source $m$ associated with the packet that gives the largest AoI reduction, $\tau_m(t+1)-\tau_m(t)$, if the transmission in slot $t$ is successful. Intuitively, this policy is selecting the source with highest potential reward in terms of AoI. In \cite{igorTMC19}, the authors evaluate the performance of the Age-Based Max-Weight source policy both analytically and using simulations, and show that it achieves near optimal AoI. The first three channel policies, namely $\epsilon$-Greedy, UCB, and TS, were discussed in Sec.~\ref{sec.Regret}. The Optimal policy is the order-optimal channel policy $\bar\eta_b$ developed in Sec.~\ref{sec.Policy}. The Hybrid policy employs TS for a fixed period in the beginning of the simulation and then employs the Optimal policy in the remaining slots.
We simulate a network with a time horizon of $T=10^5$ slots, $M=3$ sources, each generating packets according to a Bernoulli process with rate $\lambda$, and $N=5$ channels with reliabilities $\vec\mu = [0.4\ 0.45\ 0.5\ 0.55\ 0.6]^\mathsf{T}$. Figures~\ref{fig.Regret010} and \ref{fig.Regret075} show simulation results of the evolution of the AoI regret over time for $\lambda=0.1$ and $\lambda=0.75$, respectively.
Figure~\ref{fig.Estimates} shows simulation results of the evolution of the reliability estimates associated with the channels with $\mu_4=0.55$ and $\mu_5=0.6$ over time for $\lambda=0.75$.
Each data point in Figs.~\ref{fig.Regret010}, \ref{fig.Regret075}, and \ref{fig.Estimates} is an average over the results of $10^3$ simulations.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\columnwidth]{Regret010.png}
\end{center}
\vspace{-0.4cm}
\caption{Simulation of a network with $\lambda=0.1$.} \label{fig.Regret010}
\vspace{-0.2cm}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\columnwidth]{Regret075.png}
\end{center}
\vspace{-0.4cm}
\caption{Simulation of a network with $\lambda=0.75$.} \label{fig.Regret075}
\vspace{-0.2cm}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\columnwidth]{Statistics.png}
\end{center}
\vspace{-0.4cm}
\caption{Simulation of a network with $\lambda=0.75$.} \label{fig.Estimates}
\vspace{-0.2cm}
\end{figure}
The results in Figs.~\ref{fig.Regret010} and \ref{fig.Regret075} suggest that, as expected, the AoI regret associated with Optimal and Hybrid is bounded, while the AoI regrets associated with $\epsilon$-Greedy, UCB and TS grow over time. By comparing the AoI regret of Optimal and TS in Figs.~\ref{fig.Regret010} and \ref{fig.Regret075}, it is clear that the AoI regret of the Optimal channel policy varies significantly with $\lambda$. In particular, for $T=10^5$, when $\lambda$ increases from $0.1$ to $0.75$, the AoI regret of TS increases by a factor of $1.5$ (from $1,318$ to $1,963$), while the AoI regret of Optimal increases by a factor of $491.0$ (from $1,068$ to $481,700$).
A main reason for this performance degradation is that when $\lambda$ increases, empty systems with $E(t)=1$ occur less often and, as a result, the Optimal channel policy takes longer to learn the reliability of the channels, as can be seen in Fig.~\ref{fig.Estimates}.
To improve the performance of the Optimal policy for networks with large $\lambda$, we propose a heuristic policy called Hybrid channel policy, which employs TS in the first $10^4$ slots to quickly learn the reliability of the channels, and then shifts to the Optimal policy which has bounded AoI regret in the long term. Figure~\ref{fig.Estimates} illustrates the difference in the learning rates between Optimal and Hybrid.
Notice in Fig.~\ref{fig.Estimates} that there are extended periods of time in which the Optimal channel policy assigns a larger estimated reliability to a suboptimal channel, which leads to the large AoI regret shown in Fig.~\ref{fig.Regret075}. However, as established in Theorem~\ref{t3}, for a long enough time-horizon $T$, the Optimal policy will eventually converge to the true reliabilities, at which point the AoI regret will stop increasing.
\section{Conclusion}\label{sec.Conclusion}
This paper considers a single-hop wireless network with $M$ sources transmitting time-sensitive information to the destination over $N$ unreliable channels. Packets from each source are generated according to a Bernoulli process with known rate $\lambda$ and the state of channel $n$ (ON/OFF) varies according to a Bernoulli process with unknown rate $\mu_n$. The reliabilities $\vec\mu$ of the wireless channels is to be learned through observation. At every slot $t$, the learning algorithm selects a single pair $(m(t), n(t))$ and the selected source $m(t)$ attempts to transmit its packet via the selected channel $n(t)$. The goal of the learning algorithm is to minimize the expected total AoI $\bar h(T)$. To analyze the performance of the learning algorithm, we derive bounds on the AoI regret $R^\pi(T)$ associated with different learning algorithms.
Our main contributions include: i) analyzing the performance of learning algorithms that employ channel policies based on traditional MAB algorithms ($\epsilon$-Greedy, UCB, and TS) and showing that their AoI regret scales as $\Theta(\log T)$; and ii) developing a novel learning algorithm and establishing that it has $O(1)$ AoI regret. To the best of our knowledge, this is the first learning algorithm with bounded AoI regret. Interesting extensions of this work include consideration of sources with unknown packet generation rates and channels with time-varying statistics.
\bibliographystyle{IEEEtran}
| 2dd4f3a645886e833469f037c7e8c3592b6eda36 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{section:introduction}
Transformer models~\citep{vaswani2017attention} pre-trained by language modeling tasks are very powerful on a wide variety of NLP tasks~\citep{devlin2018bert, liu2019roberta}.
However, the convergence efficiency of language pre-training methods hasn't been improved much~\citep{li2020train, kaplan2020scaling}.
Given finite amount of time and computing resources, language pre-training methods can be very under-trained~\citep{liu2019roberta}. Therefore unlike most of classic deep learning tasks in which overfitting is a concern, the challenge of language pre-training is actually underfitting the training data~\citep{liu2019roberta,carlini2020extracting,li2020train,shoeybi2019training}.
What makes data fitting so hard for language pre-training? Recent research shed light on this issue by interpreting the attention weights of pre-trained Transformer models\citep{clark2019does, rogers2020primer} or running probing tasks on pre-trained models with different situations\citep{liu2019linguistic,zhang2020you, talmor2019olmpics, tenney2019you, liu2021probing}.
One common observation from such work is that pre-trained model's attention modules can be good at capturing linguistic information such as syntactic patterns, while having a hard time at fitting more complex patterns such as human commonsense and reasoning. This is probably due to that models pre-trained with language modeling tasks rely too much on word co-occurrence information \citep{tenney2019you,talmor2019olmpics}.
We observed similar phenomenon by manually examining the word predictions of Masked Language Model (MLM) at masked positions during pre-training.
We found that even at the early stage of pre-training, the pre-trained model can surprisingly produce very fluent sentences. However, the model can have mis-predictions that semantically or logically contradict with the context through the entire pre-training. One example is,
\begin{center}
\textbf{\texttt{He went back to his \underline{bedroom} to continue his draft design of a new bed.}}
\end{center}
In the masked position of the sentence above, the ground-truth word is ``study'', which is mis-predicted as ``bedroom''. It is probably because the frequently co-occurring pattern between ''bedroom'' and ''bed'' that is easy to fit for the model, dominates pre-training and outruns the hard-to-fit semantics in the context.
Such easy-to-fit patterns can therefore prevent the model from fitting the more sophisticated patterns, such as reasoning and rare facts, and harm the model's performance on downstream tasks.
Fortunately, we believe that mis-predictions can help locate such dominating patterns the model has fitted that harm language understanding. When a mis-prediction occurs, there are likely to be some dominating patterns related to the mis-prediction in the context fitted by the model that cause this mis-prediction, for example, the frequently co-occurring word ''bed'' with the mis-prediction ''bedroom'' in the highlighted example. If we can add regularization to train the model to rely less on these dominating patterns such as word co-occurrences when a mis-prediction occurs, thus focusing more on the rest more subtle patterns, more information can be efficiently fitted at pre-training. Following this motivation, we propose a new language pre-training method, Using Mis-predictions as Harm Alerts (MPA).
Specifically, in MPA, when a mis-prediction occurs during pre-training, we use its co-occurrence information to guide several heads of the self-attention modules. These self-attention heads in the Transformer modules are optimized to assign higher attention weights to the words in the input sentence that rarely co-occur with the mis-prediction while assigning lower weights to the other words.
By doing so, the Transformer model is trained to rely less on the dominating co-occurring patterns of mis-predictions while focusing more on the rest of the contextual information.
Empirically we find that such a simple regularization can help pre-training methods quite effectively. We conduct experiments using BERT and ELECTRA~\citep{clark2019electra} as MPA's backbone methods. Our results show that MPA expedites the training processes of BERT and ELECTRA and improves their performances on downstream tasks.
Sensitivity analysis on hyper-parameter configurations shows that MPA improves the backbone methods in a wide range of hyper-parameter settings. The ablation study also illustrates that MPA's effectiveness largely comes from the assigning appropriate attention weights to rare context of mis-predictions.
\section{Using Mis-Predictions as Harm Alerts}
\label{section:main}
\vspace{-2mm}
In the section, we first briefly introduce BERT (in Appendix \ref{appx:preliminaries}) and ELECTRA as preliminaries.
Then we describe the proposed method, Mis-Predictions as Harm Alerts (MPA).
We use ELECTRA as the backbone method to describe MPA for ease of illustration, but it is important to note that MPA can potentially be applied to other Transformer-based language pre-training methods as well.
\subsection{Preliminaries}
\label{subsection:preliminaries}
\paragraph{ELECTRA} ELECTRA trains two Transformer models in parallel, one smaller BERT as generator (G) and the other normal-sized Transformer as discriminator (D). The training objective of the discriminator is to detect the mis-predicted tokens from the generator.
The ELECTRA generator's training objective and loss function are the same as those of BERT, as described in the previous section and Equation~\ref{eq:bert-loss}.
Specifically, the input of the generator is a token sequence masked similarly as in BERT pre-training. At every iteration during training, the generator makes predictions at the masked positions of each sequence. Then the generator outputs a new token sequence with tokens at the masked positions replaced by the generator's predictions. The generator's output sequence then would be used as the input of the discriminator. If we denote the generator's output sequence as $\mathbf{x}^r$, then the discriminator's loss function is,
\begin{equation}
\begin{aligned}
\mathcal{L}_{D}(\mathbf{x}, \mathbf{x}^{r}) &= \frac{1}{N} (\sum_{t=1}^N-\mathbf{1}(\mathbf{x}_t^{r}=\mathbf{x}_t)\cdot \log D(\mathbf{x}^{r},t) \\
&-\mathbf{1}(\mathbf{x}_t^{r}\neq \mathbf{x}_t)\cdot\log (1-D(\mathbf{x}^{r},t))).
\label{eq:electra-loss}
\end{aligned}
\end{equation}
$N$ is the length of the input sequence $x$. Equation~\ref{eq:electra-loss} is a cross-entropy loss for the binary classification task of ELECTRA's discriminator. With this loss, the discriminator of ELECTRA is trained to detect if the token at each position of the input is a mis-prediction. The final combined loss of ELECTRA with respect to $x$ is,
\vspace{-3mm}
\begin{align}
\mathcal{L}_(\mathbf{x},\mathbf{x}^{m}, \mathbf{x}^{r}) &= \mathcal{L}_G(\mathbf{x}, \mathbf{x}^{m})+\lambda \mathcal{L}_D(\mathbf{x}, \mathbf{x}^{r}),
\label{eq:electra-entire-loss}
\end{align}
$\lambda$ is the hyper-parameter controlling the ratio of the two losses. With this modified task, ELECTRA improves the pre-training efficiency by providing a better sampling efficiency.
\subsection{The Proposed Method}
\label{subsection:main}
\vspace{-2mm}
A mis-prediction usually occurs when there are dominating co-occurring patterns with the mis-prediction in the context that lead to it.
In this section, we firstly describe how MPA locates such dominating and harmful patterns with the help of mis-predictions. MPA achieves this goal by preparing beforehand a context matrix $\mathbf{S}$ with word oc-occurrence information. In context matrix $\mathbf{S}$, for a token $i$ and token $j$, if $j$ co-occurs frequently with $i$, $\mathbf{S}_{i,j}$ is close to $1$ and vice versa. Details of constructing $\mathbf{S}$ are in Appendix\ref{appx:context}.
Then we describe how MPA trains the Transformer model to rely less on such patterns and rely more on the rest information with the help of this context matrix.
\paragraph{Pre-training.} After recording the context information of tokens in the context matrix $S$, we start pre-training with MPA. During pre-training, given an partially-masked input sequence $\mathbf{x}^m$, we firstly forward it to the generator and get $\mathbf{x}^r$, same as ELECTRA. $\mathbf{x}^r$ is the output sequence from the generator, with tokens at the masked positions replaced by the generator's predictions. Then we collect the mis-predicted positions in $\mathbf{x}^r$ as,
\begin{align}
\mathcal{M}_{x} =\{1(x_t^{r} \neq x_t)\cdot t\}_{t=1}^{N_x}.
\label{eq:multi-head}
\end{align}
$N_x$ is the length of the input sequence $\mathbf{x}$. As described in Equation~\ref{eq:attention}, each query-key attention co-efficient $a(i,j)$ is calculated by a scaled dot-product of the query $i$ and the key $j$. We maintain the self-attention modules the same for all the other positions in the input sequence, except the mis-predicted positions in $\mathcal{M}_{x}$.
For each mis-predicted position $t$ in the input sequence $\mathcal{M}_{x}$, we firstly fetch the pre-calculated context vector of the mis-predicted token $x_t^{r}$ from $\mathbf{S}$. We denote it as $\mathbf{S}_{x_t^r}$.
The vector $\mathbf{S}_{x_t^r}$ consists of the context co-efficients of the mis-prediction $x_t^{r}$ with all tokens in the vocabulary.
Then from $\mathbf{S}_{x_t^r}$, we select context coefficients of mis-prediction $x_t^{r}$ with tokens in the input sentence $\mathbf{x}^r$ and denote the vector as $\mathbf{S}(t, \mathbf{x}^r)$. $\mathbf{S}(t, \mathbf{x}^r)$ is a vector with its dimension equal to the length of $\mathbf{x}^r$. Each element $\mathbf{S}(t, \mathbf{x}^r)_i$ in $\mathbf{S}(t, \mathbf{x}^r)$ is the context coefficient of token pair $(x_t^r, x_i^r)$. Figure~\ref{fig:fetch} describes this process in detail with an example.
With $\mathbf{S}(t, \mathbf{x}^r)$ prepared for the input sequence $\mathbf{x}^r$ and the mis-predicted position $t$, MPA guides several self-attention heads in the discriminator to focus less on the frequent context and more on the other context of each mis-prediction $x_t^{r}$.
Specifically, at the mis-predicted position $t$, for query $\mathbf{q}_t$, the context-guided self-attention calculates the attention coefficients as,
\vspace{-2mm}
\begin{align}
\mathbf{g}(\mathbf{q}_t, \mathbf{K}) = \frac{\mathbf{q}_t\mathbf{K}^T}{\sqrt{d_K}}\cdot (1-\mathbf{S}(t, \mathbf{x}^r)).
\label{eq:context-guilded}
\end{align}
\vspace{-1mm}
In this equation, we multiply the original pre-softmax attention co-efficients $\frac{\mathbf{q}_t\mathbf{K}^T}{\sqrt{d_K}}$ at position $t$ with $(1-\mathbf{S}(t, \mathbf{x}^r))$.
Through this way, keys ignored by the attention module could be set a larger weight in $\mathbf{g}(\mathbf{q}_t, \mathbf{K})$ if their context coefficients in $\mathbf{S}(t, \mathbf{x}^r)$ are smaller compared with other tokens in the sentence. Keys at positions of frequent context of the mis-prediction would be set a smaller weight in $\mathbf{g}(\mathbf{q}_t, \mathbf{K})$ since their context coefficients in $\mathbf{S}(t, \mathbf{x}^r)$ is relatively large.
We then use $\mathbf{g}(\mathbf{q}_t, \mathbf{K})$ as the supervised information to train these self-attention heads at the mis-predicted position $t$. Specifically, we minimize the L-2 loss between $\mathbf{g}(\mathbf{q}_t, \mathbf{K})$ and the original pre-softmax attention weights from the attention module,
\vspace{-3mm}
\begin{align}
\small
\mathcal{L}_A = \frac{1}{N_M}\sum_{t=0}^{N_M}(\frac{\mathbf{q}_t\mathbf{K}^T}{\sqrt{d}}-\mathbf{g}(\mathbf{q}_t, \mathbf{K}))^2
\label{eq:attention_loss_2}
\end{align}
\vspace{-3mm}
$N_M$ is the total number of mis-predictions. Note that we don't back-propergate through $\mathbf{g}(\mathbf{q}_t, \mathbf{K})$. $\mathbf{g}(\mathbf{q}_t, \mathbf{K})$ is only used as supervised information to guide the model to focus more on conflicting context of mis-predictions. Figure~\ref{fig:arch} in Appendix shows with an example the training of self-attention heads with this additional loss at the mis-predicted positions.
The final loss function of the proposed pre-training method is,
\vspace{-3mm}
\begin{align}
\small
\mathcal{L}_(\mathbf{x},\mathbf{x}^{m}, \mathbf{x}^{r}) &= \mathcal{L}_G(\mathbf{x}, \mathbf{x}^{m})+\lambda \mathcal{L}_D(\mathbf{x}, \mathbf{x}^{r}) + \gamma \mathcal{L}_A (\mathbf{x}, \mathbf{x}^{r}).
\label{eq:total-loss}
\end{align}
\vspace{-1mm}
$\lambda$ and $\gamma$ and hyper-parameters controlling the ratio of the 3 losses. We present the implementation details in Appendix \ref{appx:implementation}.
\section{Experiments}
\label{section:experiments}
\vspace{-2mm}
To verify the efficiency and effectiveness of MPA, we conduct pre-training experiments and evaluate pre-trained models on fine-tuning downstream tasks.
All codes are implemented with \emph{fairseq} \citep{ott2019fairseq} in \emph{PyTorch} \citep{paszke2017automatic}. All models are run on 8 NVIDIA Tesla V100 GPUs with mixed-precision \citep{micikevicius2017mixed}. We describe our experiment setup in Appendix \ref{appx:exp_setup}
\begin{table}
\small
\caption{Performances of all methods on all downstream tasks. Results of GPT, BERT and ELECTRA are from \citep{clark2019electra} and \citep{levine2020pmi}. The result of SpanBERT is obtained by fine-tuning the released checkpoint from \citep{joshi2019spanbert}. We also reproduce BERT and ELECTRA in our system for fair comparison. We report their results as BERT(ours) and ELECTRA(ours).
MPA outperforms its corresponding backbone methods on every downstream task.
}
\label{tab:general}
\small
\begin{tabular}{lccccc}
\toprule
& Params & Avg. GLUE & Avg. SuperGLUE & SQuAD2.0 F1 & SQuAD2.0 EM\\
\hline
GPT-2 & 117 M & 78.8 & -&- &-\\
BERT & 110 M & 82.2 & 66.1 &76.4 &79.6\\
SpanBERT & 110 M & 83.9 &- &77.1 &80.3\\
ELECTRA & 110 M & 85.1 & -& 80.5&83.3\\
\hline
BERT (Ours) & 110 M & 83.0 & 66.3& 76.9& 80.1\\
BERT-MPA & 110 M & $\mathbf{83.7}$ & $\mathbf{67.4}$& $\mathbf{77.5}$& $\mathbf{80.7}$\\
\hline
ELECTRA (Ours) & 110 M & 85.2 & 70.1& 80.2 &83.1\\
\textbf{ELECTRA-MPA} & 110 M & $\mathbf{86.0}$ & $\mathbf{72.2}$ & $\mathbf{83.1}$ & $\mathbf{86.1}$\\
\bottomrule
\end{tabular}
\end{table}
Table~\ref{tab:general} shows the performances of all pre-trained models on every task. We can see that MPA improves the performance of both BERT and ELECTRA on each of the task tested. For fair comparison, we pre-train the two baselines, BERT and ELECTRA, with our system and data and present the results as BERT(ours) and ELECTRA(ours).
We also present pre-trained models' performances on every individual sub-task of GLUE in Table~\ref{tab:glue} and those of SuperGLUE in Table~\ref{tab:superglue} in Appendix. Results show that MPA can outperform both of the backbone methods in the majority of sub-tasks. MPA also has considerable performance improvements on tasks requiring more complex semantic understanding and reasoning, such as MNLI, RTE, WiC and WSC.
We also present the average GLUE score curves during the pre-training in Figure~\ref{fig:curve} in Appendix. It shows that MPA outperforms its backbone methods throughout the entire pre-training. It also shows that MPA can reach the same average GLUE scores with fewer numbers of iterations compared with both baselines. The number of iterations required for MPA-ELECTRA is $40\%$ less than that of ELECTRA when reaching ELECTRA's final performance. Since MPA-ELECTRA's running time per iteration is identical with ELECTRA, MPA-ELECTRA's training time is also $40\%$ less than that of ELECTRA.
Finally we present our hyper-parameter analysis and ablation study in Appendix \ref{appx:ablation_study}
\section{Conclusion and Future Work}
\label{section:conclusion}
\vspace{-2mm}
In this work, we propose MPA (Mis-Predictions as Harm Alerts) to improve language pre-training methods. In MPA, we train several heads of Transformer's attention modules to rely less on the dominating co-occurring patterns with mis-predictions and to focus more on the rare context of them. Experimental results show that MPA can improve both BERT and ELECTRA on the majority of GLUE downstream tasks. MPA can also make the pre-training of BERT and ELECTRA more efficient.
The ablation study also confirms that MPA's effectiveness largely comes from the rare context of mis-predictions.
| 285d186be92ac2358764ab105aa8925c07f2b4a5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{sec:int}
As the reconfiguration of magnetic field geometry, magnetic reconnection plays an elemental role in magnetized plasma systems, e.g., the solar and stellar coronae and planetary magnetospheres, throughout the university \citep{2000mrmt.conf.....P}. It is used to explain the release of magnetic energy and its conversion to other forms, such as thermal and kinetic energy \citep{2010RvMP...82..603Y}. In solar physics, numerous theoretical studies of magnetic reconnection have been undertaken to explain various solar activities, such as flares, filament eruptions, coronal mass ejections, and jets \citep{1999Ap&SS.264..129S, 2000JGR...105.2375L, 2011LRSP....8....1C}. In the two-dimensional (2D) models, magnetic reconnection takes place at an X-point where anti-parallel magnetic field lines converge and reconnect \citep{2000mrmt.conf.....P, 2010RvMP...82..603Y, 2020PRSA...476.20190867N}. However, the process of magnetic reconnection is difficult to observe directly.
In the solar corona, magnetic flux is frozen into the coronal plasma \citep{2014masu.book.....P}. The coronal structures, e.g., loops and filament threads, and their structural changes thus outline the magnetic field topology and its evolution. Using the remote-sensing observations, many signatures of magnetic reconnection have been reported. These include reconnection inflows \citep{2001ApJ...546L..69Y, 2005ApJ...622.1251L, 2009ApJ...703..877L}, current sheets \citep{2008ApJ...686.1372C, 2010ApJ...723L..28L, 2012SoPh..276..261S, 2016NatPh..12..847L, 2016ApJ...829L..33L, 2018ApJ...866...64C, 2019ApJ...874..146H}, reconnection outflows \citep{2012ApJ...745L...6T, 2014ApJ...797L..14T, 2016ApJ...818L..27C, 2016Ap&SS.361..301L}, plasmoid ejections \citep{2012PhRvX...2b1015S, 2013A&A...557A.115K, 2019A&A...628A...8P, 2020A&A...633A.121X}, loop-top hard X-ray sources \citep{1994Natur.371..495M, 2013NatPh...9..489S}, supra-arcade downflows \citep{2000SoPh..195..381M, 2003SoPh..217..267I, 2016ApJ...829L..33L}, cusp-shaped post-flare loops \citep{1992PASJ...44L..63T, 2018ApJ...853L..18Y}, and coronal structural reconfigurations \citep{2013ApJ...776...57Z, 2014A&A...570A..93L, 2018ApJ...864L...4L, 2018ApJ...868L..33L, 2019ApJ...884...34L, 2018ApJ...863L..22K}.
In the solar chromosphere, observational evidence of magnetic reconnection has also been presented \citep{2020scts1143101914636Y}. Using the H$\alpha$ images from the New Vacuum Solar Telescope \citep[NVST;][]{2014RAA....14..705L} with high spatial and temporal resolution, signatures of magnetic reconnection are observed at an X-type configuration between a pair of interacting fibrils \citep{2015ApJ...798L..11Y, 2016ApJ...819L..24Y}. During the process of magnetic reconnection, inflows and outflows of fibrils are clearly detected \citep{2015ApJ...798L..11Y}. In another reconnection event, oscillation of newly reconnected fibrils after reconnection in the chromosphere is reported \citep{2016ApJ...819L..24Y}. Magnetic reconnection between the chromospheric fibrils and filament threads is studied, and suggested to play an important role in solar eruptions by releasing the magnetic twist \citep{2016NatCo...711837X, 2018ApJ...853L..26H}. Between two neighboring filaments, magnetic reconnection occurs, and forms two sets of new filaments \citep{2017ApJ...838..131Y} and two-sided-loop jets \citep{2019ApJ...883..104S, 2019ApJ...887..220Y}. Magnetic reconnection between the emerging and pre-existing fibrils is investigated \citep{2018ApJ...861..108Z, 2019ApJ...876...51Z}. An example of two-sided-loop jets simultaneously observed in the chromosphere, transition region, and corona is then described \citep{2018ApJ...861..108Z}. Recently, a small-scale oscillatory reconnection event is presented, that leads to the formation and disappearance of a flux rope \citep{2019ApJ...874L..27X}. Moreover, the disconnection of a filament caused by reconnection is revealed \citep{2020A&A...633A.121X}.
Magnetic reconnection modulated by non-local solar activities has been studied theoretically \citep[see a review by][]{2018SSRv..214...45M}. \citet{2006A&A...452..343N} numerically simulated the interaction of fast magnetoacoustic oscillations of a non-flaring loop with a nearby magnetic null point. They found that the fast magnetoacoustic wave coming into the null point from the outside oscillating loop can trigger the magnetic reconnection. \citet{2006SoPh..238..313C} performed magnetohydrodynamic (MHD) simulations of magnetic reconnection driven by five-minute solar p-mode oscillations. They pointed out that several typical and puzzling features of the transition-region explosive events can only be explained if there exist p-mode oscillations and the reconnection site is located in the upper chromosphere. \citet{2009A&A...493..227M} investigated the nature of nonlinear fast magnetoacoustic waves propagating in the neighborhood of a 2D magnetic X-point. They demonstrated that magnetic reconnection is naturally driven by the MHD wave propagation. However, magnetic reconnection affected by the external solar activities is rarely observed directly. In the wake of an erupting flux rope, oscillation of the current sheets caused by the neighboring filament eruption has been reported \citep{2016ApJ...829L..33L}. But the evolution of magnetic reconnection is not investigated. Recently, \citet{2017ApJ...851L...1Z} presented that the rising flux rope pushes the overlying loops, and forms an external current sheet where magnetic reconnection takes place. In this paper, we report a reconnection event accelerated by an adjacent filament eruption. The observations and results are shown separately in Sections\,\ref{sec:obs} and \ref{sec:res}. A summary and discussion is presented in Section\,\ref{sec:sum}.
\section{Observations}\label{sec:obs}
The NVST is a 1-meter ground-based solar telescope, located in the Fuxian Solar Observatory of the Yunnan Observatories, Chinese Academy of Sciences. It provides observations of the solar fine structures and their evolution in the solar lower atmosphere. On 2013 March 15, the NVST observed the active region (AR) 11696 with a field of view (FOV) of 200\arcsec$\times$186\arcsec~in the H$\alpha$ channel, centered at 6562.8\,\AA~with a bandwidth of 0.25 \AA, from 01:20\,UT to 06:40\,UT. The H$\alpha$ images have a time cadence of 12\,s and spatial sampling of 0.164\arcsec\,pixel$^{-1}$. They are processed first by flat field correction and dark current subtraction, and then reconstructed by speckle masking \citep[][and references therein]{2016NewA...49....8X}. The co-alignment of H$\alpha$ images is carried out by a fast sub-pixel image registration algorithm \citep{2012JKAS...45..167F, 2015RAA....15..569Y}.
The Atmospheric Imaging Assembly \citep[AIA;][]{2012SoPh..275...17L} onboard the Solar Dynamic Observatory \citep[SDO;][]{2012SoPh..275....3P} is a set of normal-incidence imaging telescopes, acquiring solar atmospheric images in ten wavelength bands. Different AIA channels show plasma at different temperatures, e.g., 131\,\AA~peaks at $\sim$10\,MK (Fe\,XXI) and $\sim$0.6\,MK (Fe\,VIII), 94\,\AA~peaks at $\sim$7.2\,MK (Fe\,XVIII), 335\,\AA~peaks at $\sim$2.5\,MK (Fe\,XVI), 211\,\AA~peaks at $\sim$1.9\,MK (Fe\,XIV), 193\,\AA~peaks at $\sim$1.5\,MK (Fe\,XII), 171\,\AA~peaks at $\sim$0.9\,MK (Fe\,IX), and 304\,\AA~peaks at $\sim$0.05\,MK (He\,II). In this study, we employ the AIA images in one ultraviolet (UV) channel (1600\,\AA) and seven extreme UV (EUV) channels (131, 94, 335, 211, 193, 171, and 304\,\AA) to investigate the evolution of magnetic reconnection and filament eruption. Here, the AIA images are processed to 1.5-level using ``aia\_prep.pro". Then the spatial sampling of AIA images is 0.6\arcsec\,pixel$^{-1}$, and the time cadences of AIA EUV and UV images are 12\,s and 24\,s, respectively. The Helioseismic and Magnetic Imager \citep[HMI;][]{2012SoPh..275..229S} onboard the SDO provides line of sight (LOS) magnetograms, with a time cadence of 45\,s and spatial sampling of 0.5\arcsec\,pixel$^{-1}$. We use the HMI LOS magnetograms to study the evolution of surface magnetic fields underlying the reconnection region and the erupting filament.
The NVST H$\alpha$ images have been rotated to match the orientation of SDO observations. All the data from different instruments, i.e., the SDO and NVST, and passbands have been aligned with a principle of best cross-correlation between images of two passbands with the closest characteristic temperatures.
\section{Results}\label{sec:res}
We analyze the observations that provide signatures of how a filament eruption causes magnetic reconnection to speed-up at a different location. On 2013 March 15, a saddle-like structure, located to the northeast of AR 11696, was observed by the NVST, see Figure\,\ref{f:general_information}(a). It is created by a set of four fibrils seen in the H$\alpha$ images, marked separately by L1, L2, L3, and L4, see the red, green, cyan, and blue dashed lines in Figure\,\ref{f:general_information}(a). A curved filament is located to the southeast of the saddle-like structure, outlined by the pink dotted line in Figure\,\ref{f:general_information}(a). We overlay the H$\alpha$ fibrils L1-L4 and their nearby filament on an HMI LOS magnetogram in Figure\,\ref{f:general_information}(b). For better descriptions, we label the positive and negative magnetic fields as P1 and P2, and N1, N2, and N3, based on the fibril connectivity, see Figure\,\ref{f:general_information}(b). The fibrils L1, L2, L3, and L4 connect the positive and negative magnetic fields P1 and N1, P1 and N2, P2 and N2, and P2 and N1, respectively, see the red, green, cyan, and blue dashed lines in Figure\,\ref{f:general_information}(b). The central positive magnetic fields P2 and their surrounding negative magnetic fields N1, N2, and N3 constitute a fan-spine magnetic field configuration, see Figures\,\ref{f:general_information}(b) and \ref{f:cartoon}. The filament is located upon the polarity inversion line between the central positive magnetic fields P2 and the surrounding negative magnetic fields N1 and N3, see the pink dotted line in Figure\,\ref{f:general_information}(b).
\subsection{Magnetic reconnection before the filament eruption}\label{sec:mr1}
The H$\alpha$ fibrils L2 and L4 constitute a saddle-type structure, see Figure\,\ref{f:mr_ha}(a). They move toward each other, and reconnect. Two sets of newly reconnected fibrils L1 and L3 then form, and retract away from the reconnection region, see the online animated version of Figure\,\ref{f:mr_ha}. No current sheet is observed in the H$\alpha$ diagnostics in the reconnection region between fibrils L2 and L4. Along the red line AB in Figure\,\ref{f:mr_ha}(c), a time slice of H$\alpha$ images is made, and displayed in Figure\,\ref{f:measurements_ha}(a). Inward motions of fibrils L2 and L4, with mean speeds of $\sim$16-18 km\,s$^{-1}$, see the green dotted lines in Figure\,\ref{f:measurements_ha}(a), toward the reconnection region, denoted by the red dashed line in Figure\,\ref{f:measurements_ha}(a), are clearly identified. Along the green line CD in Figure\,\ref{f:mr_ha}(d), another time slice of H$\alpha$ images is obtained, and shown in Figure\,\ref{f:measurements_ha}(b). At the same time, outward motions of the newly reconnected fibrils L1 and L3, with mean speeds of $\sim$24-26 km\,s$^{-1}$, see the green dotted lines in Figure\,\ref{f:measurements_ha}(b), away from the reconnection region, denoted by the red dashed line in Figure\,\ref{f:measurements_ha}(b), are evidently detected. Moreover, topological reconfiguration of the fibrils by reconnection, e.g., from fibrils L4 to L3 and L1, is observed, see Figures\,\ref{f:mr_ha}(a)-(b) and (e)-(f) and the online animated version of Figure\,\ref{f:mr_ha}. Along the pink and cyan lines EF and GH in Figures\,\ref{f:mr_ha}(b) and (e), time slices of H$\alpha$ images are made, and illustrated in Figures\,\ref{f:measurements_ha}(c) and (d), respectively. The fibrils L4 reconnect with L2, and then separately turn into L3 and L1, with mean moving speeds of $\sim$7 km\,s$^{-1}$ and $\sim$13 km\,s$^{-1}$, see the blue dotted lines in Figures\,\ref{f:measurements_ha}(c)-(d).
The process of magnetic reconnection is also observed by AIA, see Figure\,\ref{f:mr_aia1}. Four sets of coronal loops L1-L4 are recorded in AIA EUV images, forming a saddle-like structure, consistent with the H$\alpha$ fibrils L1-L4, see Figure\,\ref{f:mr_aia1}(a). In the same manner, the loops L2 and L4 constitute an X-type structure, see the red dotted lines in Figure\,\ref{f:mr_aia1}(c). At the interface of these two loops, magnetic reconnection takes place, see the online animated version of Figure\,\ref{f:mr_aia2}. Along the cyan line IJ in Figure\,\ref{f:mr_aia1}(a), a time slice of AIA 304\,\AA~images is made, and displayed in Figure\,\ref{f:measurements_aia1}(a). Different from the H$\alpha$ observations, motions of the EUV loops toward the reconnection region are hard to observe, see the online animated version of Figure\,\ref{f:mr_aia2}. Nevertheless, the inward motion of loops L2, similar to the H$\alpha$ fibrils L2, with a mean speed of $\sim$33 km\,s$^{-1}$, see the blue dotted line in Figure\,\ref{f:measurements_aia1}(a), toward the reconnection region, marked by the purple dashed line in Figure\,\ref{f:measurements_aia1}(a), is still identified.
In the reconnection region, a current sheet forms, denoted by green solid arrows in Figure\,\ref{f:mr_aia1}. In the particular snapshot, i.e., at 01:33:11 UT, it has a width of $\sim$2\,Mm, and a length of $\sim$7.2\,Mm in AIA 171\,\AA~images, see Figure\,\ref{f:mr_aia1}(c). Here, the length of current sheet is measured between two cusp-shaped structures at the ends of current sheet, marked by red pluses in Figure\,\ref{f:mr_aia1}(c). For the width of current sheet, first we get the intensity profile in the AIA 171 \AA~channel perpendicular to the current sheet, e.g., along the purple line in Figure\,\ref{f:mr_aia1}(c). Employing the intensity surrounding the current sheet, we calculate the background emission, and subtract it from the intensity profile. We fit the residual intensity profile using a single Gaussian, and obtain the full width at half maximum (FWHM) of the single Gaussian fit as the current sheet width. The magnetic reconnection rate, the ratio of the width and the length of the current sheet \citep{1957JGR....62..509P}, is thus $\sim$0.28. The current sheet appears in most AIA EUV channels, except the higher-temperature channels, e.g., 335 and 94\,\AA, see Figure\,\ref{f:mr_aia1} and the online animated version of Figure\,\ref{f:mr_aia2}. In the blue rectangle enclosing the current sheet in Figure\,\ref{f:mr_aia1}(e), the light curves of the AIA 304, 131, 171, 193, and 211\,\AA~channels are calculated, and shown in Figure\,\ref{f:measurements_aia1}(c). Considering the influence of background emission, similar temporal evolution of the AIA EUV light curves are identified. They reach the peaks at almost the same time, see the blue vertical dashed line in Figure\,\ref{f:measurements_aia1}(c). The current sheet repeatedly appears and disappears, see the online animated version of Figure\,\ref{f:mr_aia2}, with a mean period of $\sim$8\,minutes, obtained from the wavelet analysis of the AIA 304\,\AA~light curve in Figure\,\ref{f:measurements_aia1}(c). Same as before, the well-developed current sheets are measured using AIA 171\,\AA~images. They have the width of 1-2\,Mm with a mean value of 1.5\,Mm, the length of 5.3-7.2\,Mm with a mean value of 6.4\,Mm, and the reconnection rate of 0.18-0.3 with a mean value of 0.24.
As the current sheet is not observed in the AIA higher-temperature channels, e.g., 94 and 335\,\AA, plasma at that location could be cooler than $\sim$2.5\,MK, the characteristic temperature of AIA 335\,\AA~channel. The current sheet in AIA 131\,\AA~images, see Figure \ref{f:mr_aia1}(b), hence shows plasma with the lower characteristic temperature ($\sim$0.6\,MK) of AIA 131\,\AA~channel. Using six AIA EUV channels, including 94, 335, 211, 193, 171, and 131\,\AA, we analyze the temperature and emission measure (EM) of the current sheet. Here, we employ the differential EM (DEM) analysis using ``xrt\_dem\_iterative2.pro" \citep{2012ApJ...761...62C}. The current sheet region, enclosed by the red rectangle in Figure\,\ref{f:mr_aia1}(b), is chosen to compute the DEM. The region out of the current sheet, enclosed by the purple rectangle in Figure\,\ref{f:mr_aia1}(b), is chosen for the background emission that is subtracted from the current sheet region. In each region, the DN counts in each of the six AIA channels are temporally normalized by the exposure time and spatially averaged over all pixels. The DEM curve of the current sheet region is displayed in Figure\,\ref{f:measurements_aia1}(e). Consistent with the AIA imaging observations, the DEM shows a lack of hot plasma component in the current sheet, see the black curve in Figure\,\ref{f:measurements_aia1}(e). The average DEM-weighed temperature and EM are 2\,MK and 2.5$\times$10$^{27}$\,cm$^{-5}$, respectively.
Using the EM, the number density ($n_{p}$) of current sheet is estimated using $n_{p}=\sqrt{\frac{{\textrm{EM}}}{D}}$, where $D$ is the LOS depth of current sheet. Assuming that the depth $D$ equals the width ($W$) of current sheet, then the density is $n_{p}=\sqrt{\frac{{\textrm{EM}}}{W}}$. Employing EM=2.5$\times$10$^{27}$\,cm$^{-5}$ and W=2\,Mm, we obtain the density to be 3.5$\times$10$^{9}$\,cm$^{-3}$. The thermal energy (TE) of current sheet is also calculated using TE=$\frac{3}{2}$n$_{p}$$\cdot$k$_{B}$$\cdot$V$\cdot$$\delta$T. Here k$_{B}$ is Boltzmann's constant, V volume, and $\delta$T temperature increase from the temperature (T$_{1}$) of reconnection inflowing structure to that (T$_{2}$) of current sheet. Assuming that the current sheet is a cylinder, its volume is then V=$\pi$($\frac{W}{2}$)$^{2}$$\cdot$L, where L is the length of current sheet. Therefore the thermal energy is TE=$\frac{3}{2}$n$_{p}$$\cdot$k$_{B}$$\cdot$$\pi$($\frac{W}{2}$)$^{2}$$\cdot$L$\cdot$(T$_{2}$-T$_{1}$). As the reconnection inflow is observed mainly in H$\alpha$ images and the current sheet is identified in EUV images, we obtain T$_{1}$=10$^{4}$ K, the temperature of H$\alpha$ fibrils \citep{2012ApJ...749..136L}, and T$_{2}$ to be the average DEM-weighed temperature of current sheet. Employing n$_{p}$=3.5$\times$10$^{9}$\,cm$^{-3}$, W=1-2\,Mm, L=5.3-7.2\,Mm, T$_{2}$=2\,MK, and T$_{1}$=10$^{4}$ K, we calculate the thermal energy of current sheet, and get TE=(1.9$\pm$1.3)$\times$10$^{25}$\,erg.
Plasmoids appear in the current sheet, and propagate along it bi-directionally, and then further along the reconnection loops L2 and L4, see Figure \ref{f:mr_aia1}(d) and the online animated version of Figure\,\ref{f:mr_aia2}. Same as the current sheet, they are not observed in the AIA higher-temperature channels, e.g., 335 and 94\,\AA. Along the current sheet, see the green line KL in Figure\,\ref{f:mr_aia1}(c), a time slice of AIA 171\,\AA~images is made, and shown in Figure\,\ref{f:measurements_aia1}(b). It indicates that the plasmoids always form in the middle of the current sheet, marked by the red dashed line in Figure\,\ref{f:measurements_aia1}(b), and then move bi-directionally with mean speeds of $\sim$36-48\,km\,s$^{-1}$, see the green dotted lines in Figure\,\ref{f:measurements_aia1}(b). Along the blue line MN in Figure\,\ref{f:mr_aia1}(c), another time slice of AIA 171\,\AA~images is made, and illustrated in Figure\,\ref{f:measurements_aia1}(d). Several motions of plasmoids along the north leg of loops L2 are evidently identified, with a mean speed of $\sim$71\,km\,s$^{-1}$, see the green dotted line in Figure\,\ref{f:measurements_aia1}(d).
From $\sim$02:30 UT, a set of higher-lying loops L6, connecting the surrounding negative magnetic fields N1-N3 and the remote positive magnetic fields, appears in most AIA EUV channels, except 304\,\AA, see Figure\,\ref{f:filament_eruption}(a). They are likely to be heated by nanoflares \citep{2015A&A...583A.109L}, as no significant activity is detected associated with the brightening of loops. The current sheet, i.e., the reconnection region, is then covered by the loops L6 in these EUV channels. It, however, still appears in AIA 304\,\AA~images, see the online animated version of Figure\,\ref{f:mr_aia2}. This indicates that the reconnection between loops L2 and L4 continues to take place as before, consistent with the H$\alpha$ observations. A set of lower-lying loops L5 overlying the filament is also detected. It connects the positive and negative magnetic fields P2 and N3, see Figures\,\ref{f:filament_eruption}(a) and \ref{f:cartoon}. In the AIA higher-temperature channels, e.g., 94 and 335\,\AA, the loops L5 and L6 gradually disappear, indicating the cooling process of heated loops \citep{2015A&A...583A.109L}. Consistent with the previous observations, the current sheet forms in the lower-temperature, e.g., 304\,\AA, rather than the higher-temperature channels, e.g., 94 and 335\,\AA, see the online animated version of Figure\,\ref{f:mr_aia2}.
\subsection{Filament eruption}\label{sec:filament_eruption}
From $\sim$05:52 UT, the north, rather than the south, part of the filament, located to the southeast of the reconnection region, brightens, see Figure\,\ref{f:filament_eruption}(b), and then erupts, see Figure\,\ref{f:filament_eruption}(d). A partial eruption of the filament is thus observed \citep{2016NatPh..12..847L}. Along the erupting direction, see the green line PQ in Figure\,\ref{f:filament_eruption}(b), a time slice of AIA 304\,\AA~images is made, and displayed in Figure\,\ref{f:filament_eruption}(g). The filament erupts with a mean projection speed of $\sim$55 km\,s$^{-1}$, see the blue dotted line in Figure\,\ref{f:filament_eruption}(g). Assuming that the filament erupts outward along the radial direction, we obtain the corrected erupting speed to be 163 km\,s$^{-1}$, using the heliographic position N13\,W15 of the erupting filament. The erupting filament is prevented eventually by the higher-lying loops L5 and L6, showing a failed filament eruption, and makes the higher-lying loops bright, see Figure\,\ref{f:filament_eruption}(e) and the online animated version of Figure\,\ref{f:mr_aia2}. Moreover, the failed filament eruption may also be caused by the reconnection between the erupting filament and its overlying loops \citep{2016NatPh..12..847L, 2020ApJ...889..106Y}. Two flare ribbons and post-flare loops, associated with the filament eruption, appear, see Figures\,\ref{f:filament_eruption}(c) and (e). The south flare ribbon moves away from the polarity inversion line of the positive and negative magnetic fields P2 and N3, with a mean speed of $\sim$42 km\,s$^{-1}$, see the green dotted line in Figure\,\ref{f:filament_eruption}(g).
The filament eruption pushes away the higher-lying loops L6 covering the reconnection region, see Figure\,\ref{f:filament_eruption}(f) and the online animated version of Figure\,\ref{f:mr_aia2}. It thus leads to a disturbance propagating outward across the reconnection region. Along the propagating direction VW in the blue rectangle in Figure\,\ref{f:filament_eruption}(f), a time slice of AIA 211\,\AA~images is measured, and shown in Figure\,\ref{f:filament_eruption}(h). It indicates that the loops L6 is pushed away with a mean moving speed of $\sim$290\,km\,s$^{-1}$ by the propagating disturbance caused by the filament eruption, see the blue dotted line in Figure\,\ref{f:filament_eruption}(h). A dimming region then forms, and the reconnection region, e.g., the current sheet, reappears in these AIA EUV channels, e.g., 171, 193, and 211\,\AA, see the online animated version of Figure\,\ref{f:mr_aia2}.
\subsection{Magnetic reconnection after the filament eruption}\label{sec:mr2}
After the filament eruption, the current sheet between loops L2 and L4 has the width and length of $\sim$2\,Mm and $\sim$3.5\,Mm at 06:18:35 UT, and thus a reconnection rate of $\sim$0.57, see Figure\,\ref{f:mr_aia2}(c). Here, the width is similar to the current sheet widths (1-2\,Mm), the length is, however, smaller than the current sheet lengths (5.3-7.2\,Mm), and the reconnection rate is much larger than the reconnection rates (0.18-0.3), before the filament eruption. The current sheet, different from those occurring in the AIA low-temperature EUV channels before the filament eruption, appears in all AIA EUV channels, marked by the green solid arrows in Figure\,\ref{f:mr_aia2}. Same as in Section\,\ref{sec:mr1}, in the blue rectangle enclosing the reconnection region in Figure\,\ref{f:mr_aia2}(e), the light curves of the AIA 94, 335, 211, 193, 171, 131, and 304\,\AA~channels are calculated, and displayed in Figure\,\ref{f:measurements_aia2}(a). Here, the AIA 211, 193, and 171\,\AA~light curves before the filament eruption, denoted by the red vertical dotted line in Figure\,\ref{f:measurements_aia2}(a), are not measured, because before the filament eruption the reconnection region is covered by the higher-lying loops L6 in these channels, see Section\,\ref{sec:mr1}. Before the filament eruption, the AIA 94\,\AA~and 335\,\AA~light curves keep constant, indicating that no current sheet appears in these AIA higher-temperature channels. The AIA 304\,\AA~and 131\,\AA~light curves, however, evolve due to the appearance of current sheet in these lower-temperature channels. Right after the filament eruption, the AIA 335, 211, 193, and 171\,\AA~light curves decrease evidently, showing a dimming, as the higher-lying loops L6, covering the reconnection region, are pushed away by the propagating disturbance caused by the filament eruption, see also Figures\,\ref{f:filament_eruption}(f) and (h). All the light curves then exhibit rapid rise and reach the peaks. Among them, the AIA 94\,\AA~light curve reaches the peak at 06:09:37 UT, see the green vertical dashed line in Figure\,\ref{f:measurements_aia2}(a), $\sim$2.3 minutes later than the AIA 304\,\AA~light curve that peaks at 06:07:19 UT, see the blue vertical dashed line in Figure\,\ref{f:measurements_aia2}(a). Affected by the brightening of higher-lying loops caused by the filament eruption, see Section \ref{sec:filament_eruption}, the AIA 335 \AA~light curve reaches the peak several minutes later than the other light curves. It, however, has a small peak at 06:07:26 UT, identical to the peak of the AIA 304 \AA~light curve, see Figure\,\ref{f:measurements_aia2}(a). In addition, a much bright reconnection region is detected simultaneously in H$\alpha$ observations, see the online animated version of Figure\,\ref{f:mr_ha}.
As the current sheet appears in all AIA EUV channels, it, therefore, may contain plasma with both of the high and low temperature. The current sheet region, enclosed by the red rectangle in Figure\,\ref{f:mr_aia2}(d), is chosen to compute the DEM. The region out of the current sheet, enclosed by the purple rectangle in Figure\,\ref{f:mr_aia2}(d), is selected for the background emission that is subtracted from the current sheet region. The DEM curve of the current sheet region is shown in Figure\,\ref{f:measurements_aia2}(c). It indicates that more plasma with higher temperature and less plasma with lower temperature is detected in the current sheet, comparing to that in Figure\,\ref{f:measurements_aia1}(e). The average DEM-weighed temperature and EM are 5.5\,MK and 1.7$\times$10$^{28}$\,cm$^{-5}$, respectively. Both these quantities are significantly higher than those of current sheet before the filament eruption (2\,MK and 2.5$\times$10$^{27}$\,cm$^{-5}$). Similar to Section\,\ref{sec:mr1}, the density of current sheet is also estimated to be 9.2$\times$10$^{9}$\,cm$^{-3}$ by using the EM, under the assumption that the LOS depth (D) of the current sheet equals its width (W=2\,Mm). It is larger than that of current sheet before the filament eruption (3.5$\times$10$^{9}$\,cm$^{-3}$). Using n$_{p}$=9.2$\times$10$^{9}$\,cm$^{-3}$, W=2\,Mm, L=3.5\,Mm, T$_{2}$=5.5\,MK, and T$_{1}$=10$^{4}$ K, we also calculate the thermal energy of current sheet, and obtain the value to be TE=1.1$\times$10$^{26}$\,erg. It is also larger than that of current sheet before the filament eruption (1.9$\pm$1.3$\times$10$^{25}$\,erg).
In the current sheet, plasmoids appear and move along it and the reconnection loops, see Figure \ref{f:filament_eruption}(d) and the online animated version of Figure\,\ref{f:mr_aia2}. The north endpoint of loops L2, enclosed by the pink circles, then brightens, see Figure\,\ref{f:mr_aia2}(h). Along the green line RS in Figure\,\ref{f:mr_aia2}(c), a time slice of AIA 171\,\AA~images is made, and displayed in Figure\,\ref{f:measurements_aia2}(b). It shows that more plasmoids moving along the north leg of loops L2 are detected, comparing with that in Figure\,\ref{f:measurements_aia1}(d), with a similar mean speed of $\sim$70\,km\,s$^{-1}$, see the green dotted line in Figure\,\ref{f:measurements_aia2}(b). Moreover, the plasmoids, different from those appearing in the AIA lower-temperature EUV channels before the filament eruption, see Section\,\ref{sec:mr1}, appear in all AIA EUV channels. More hotter plasmoids are thus generated in the current sheet after the filament eruption. Comparing all the results before and after the filament eruption, see Sections\,\ref{sec:mr1} and \ref{sec:mr2}, we conclude that the reconnection between loops L2 and L4 is significantly accelerated by the filament eruption occurring to the southeast of the reconnection region.
\section{Summary and discussion}\label{sec:sum}
Employing the H$\alpha$ images from NVST, and the AIA images and HMI LOS magnetograms from SDO, we study the reconnection between fibrils (loops) L2 and L4, and its nearby filament eruption. The reconnection accelerated by the filament eruption is then reported. In H$\alpha$ images, a saddle-like structure, consisting of four sets of fibrils L1-L4, is observed. The fibrils L2 and L4 from opposite sides of the saddle region move together, and reconnect. The newly reconnected fibrils L1 and L3 then form, and retract away from the reconnection region. In AIA EUV images, similar loops L1-L4 and their evolution are identified. At the interface of loops L2 and L4, the current sheet repeatedly forms and disappears. Magnetic reconnection takes place in the current sheet. Plasmoids appear in the current sheet, and propagate along it, and then further along the reconnection loops. A filament, located to the southeast of the reconnection region, partially erupts, and leads to a flare. It is then prevented by the overlying loops as a failed filament eruption. After the filament eruption, a hotter, shorter current sheet forms with a much larger reconnection rate, where more hotter plasmoids appear. Based on the NVST H$\alpha$ images, and the SDO AIA EUV images and HMI LOS magnetograms, a schematic diagram of the magnetic reconnection between fibrils (loops) and its nearby filament eruption is demonstrated in Figure\,\ref{f:cartoon}. Here, the red star represents the reconnection point between magnetic field lines of loops L2 and L4.
A small-scale reconnection event among a saddle-like structure is observed by NVST. Similar to the small-scale reconnection events previously reported \citep{2015ApJ...798L..11Y, 2016ApJ...819L..24Y}, inward and outward motions of H$\alpha$ fibrils toward and away from the reconnection region are evidently detected, see Section\,\ref{sec:mr1}. The reconnection inflowing and outflowing speeds of $\sim$17 km\,s$^{-1}$ and $\sim$25 km\,s$^{-1}$ are consistent with those of the fast reconnection event in \citet{2015ApJ...798L..11Y}. Different from \citet{2015ApJ...798L..11Y}, in this study the current sheet appears only in the AIA EUV channels, rather than the H$\alpha$ channel, see Section\,\ref{sec:mr1}. This indicates that the current sheet is significantly heated during the reconnection process \citep{2016NatPh..12..847L, 2016ApJ...829L..33L, 2020A&A...633A.121X}. The width (1-2\,Mm) and length (3.5-7.2\,Mm) of current sheets are identical to those in \citet{2016NatCo...711837X, 2020A&A...633A.121X}, but larger than those in \citet{2015ApJ...798L..11Y} and \citet{2018ApJ...858L...4X}. Moreover, the reconnection rate (0.18-0.57) is similar to those in \citet{2016NatCo...711837X, 2018ApJ...858L...4X}, but larger than those in \citet{2020A&A...633A.121X}. Many plasmoids form in the current sheet, suggesting the presence of plasmoid instabilities during the process of magnetic reconnection \citep{2013A&A...557A.115K, 2016NatPh..12..847L, 2019A&A...628A...8P}. They propagate along the current sheet bi-directionally, and then further along the reconnection loops with a mean speed of $\sim$70 km\,s$^{-1}$. The moving speed here is consistent with those in \citet{2015ApJ...798L..11Y} and \citet{2020A&A...633A.121X}, but smaller than those in \citet{2016NatPh..12..847L}.
Magnetic reconnection accelerated by nearby filament eruption is observed. After the filament eruption, the length of current sheet decreases significantly from 5.3-7.2\,Mm to 3.5\,Mm. The reconnection rate, however, increases largely from 0.18-0.3 to 0.57, see Sections\,\ref{sec:mr1} and \ref{sec:mr2}. The enhancements of the AIA EUV light curves in the reconnection region after the filament eruption, see Figures\,\ref{f:mr_aia1}(c) and \ref{f:mr_aia2}(a), suggest the increase of temperature and/or density of plasma in the current sheet. The current sheet appears in the AIA higher-temperature channels, e.g., 335 and 94\,\AA, after rather than before the filament eruption. It is thus heated to much higher temperature after the filament eruption. This is also supported by the DEM curves of current sheet before and after the filament eruption, e.g., the average DEM-weighed temperature of current sheet increases from 2\,MK to 5.5\,MK. The increase of current sheet density from 3.5$\times$10$^{9}$ cm$^{-3}$ to 9.2$\times$10$^{9}$ cm$^{-3}$ after the filament eruption shows that more plasma (n$_{p}$$\cdot$$\pi$$\cdot$($\frac{W}{2}$)$^{2}$$\cdot$L) from (4.7$\pm$3.2)$\times$10$^{34}$ to 1$\times$10$^{35}$ is heated to higher temperature. More thermal energy of current sheet converted by reconnection after (1.1$\times$10$^{26}$\,erg) than before ((1.9$\pm$1.3)$\times$10$^{25}$\,erg) the filament eruption is then achieved. In addition, more hotter plasmoids form in the current sheet after the filament eruption during the same time intervals, see Figures\,\ref{f:measurements_aia1}(d) and \ref{f:measurements_aia2}(b). This indicates that more plasma is accelerated during the reconnection process. More kinetic energy is hence converted by reconnection after the filament eruption.
Magnetic reconnection may be accelerated by the fast MHD wave caused by filament eruption. \citet{2006A&A...452..343N} suggested that the fast wave coming into the magnetic null point from the outside leads to the increase of electric current density. The increasing electric current then efficiently induces plasma micro-instabilities of various kinds, and hence produces anomalous resistivity, which efficiently triggers the reconnection. Fast mode MHD waves generated by filament eruptions are indeed widely reported \citep[e.g.,][]{2010ApJ...723L..53L, 2012A&A...539A...7L, 2018ApJ...853....1S}. In this study, the higher-lying loops covering the reconnection region in the AIA EUV channels are pushed away right after the filament eruption, see Section\,\ref{sec:filament_eruption}. This suggests that the filament eruption leads to a disturbance propagating outward across the reconnection region, that could be related to fast mode MHD wave. Using v$_{A}$=$\frac{B}{\sqrt{4 \pi n_{p} \cdot m_{p}}}$, we calculate the Alfv$\acute{e}$n speed near the reconnection region, where B is the magnetic field strength, and m$_{p}$ is the proton mass. Employing the magnetic field strength B=(3$\pm$2) G in the corona \citep{2020Sci...369..694Y}, and the coronal density n$_{p}$=(1$\pm$0.5)$\times$10$^{9}$ cm$^{-3}$, less than the current sheet density (3.5$\times$10$^{9}$ cm$^{-3}$) before the filament eruption, the Alfv$\acute{e}$n speed is obtained to be v$_{A}$=(272$\pm$216) km\,s$^{-1}$. It is consistent with the propagating speed ($\sim$290 km\,s$^{-1}$) of the disturbance. This propagating disturbance is hence likely to represent the fast-mode MHD wave driven by the filament eruption. The fast wave comes into the current sheet, increases its electric current density \citep{2006A&A...452..343N}, and accelerates the magnetic reconnection. It could also enhance turbulent plasma motions at the current sheet, leading to a turbulent reconnection \citep[][]{2020ApJ...890L...2C}. Additionally, the propagating disturbance may also push more magnetic flux of loops (fibrils) L4, and thus more magnetic energy, into the reconnection region \citep{2017ApJ...851L...1Z}. All these effects will play a role in liberating magnetic energy with a larger reconnection rate. The released magnetic energy will then be converted to other forms of energy, e.g., thermal and kinetic.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.7\textwidth]{general_information.eps}
\caption{General information of H$\alpha$ fibrils and their nearby filament. (a) NVST H$\alpha$ image and (b) SDO/HMI LOS magnetogram. The red, green, cyan, and blue dashed lines outline the fibrils L1, L2, L3, and L4, respectively. The pink dotted lines represent the filament. The red rectangle in (a) shows the FOVs of Figures\,\ref{f:mr_ha}, \ref{f:mr_aia1}, \ref{f:filament_eruption}(a)-(f), and \ref{f:mr_aia2}. The N1, N2, and N3, and P1 and P2 in (b) separately denote the negative and positive magnetic fields. See Section\,\ref{sec:res} for details.
\label{f:general_information}}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\textwidth]{mr_ha.eps}
\caption{Magnetic reconnection between H$\alpha$ fibrils observed by NVST. (a)-(f) NVST H$\alpha$ images. The red, green, cyan, and blue arrows separately denote the fibrils L1, L2, L3, and L4. The red, green, pink, and cyan lines AB, CD, EF, and GH in (c), (d), (b), and (e) show the positions for time slices of NVST H$\alpha$ images displayed in Figures\,\ref{f:measurements_ha}(a), (b), (c), and (d), respectively. The FOV is denoted by the red rectangle in Figure\,\ref{f:general_information}(a). An animation of the unannotated NVST H$\alpha$ images is available. It covers $\sim$5.3\,hr starting at 01:20:44 UT, and the video cadence is 12\,s. See Section\,\ref{sec:mr1} for details. (An animation of this figure is available.)
\label{f:mr_ha}}
\end{figure}
\begin{figure}[ht!]
\includegraphics[width=\textwidth]{measurements_ha.eps}
\centering
\caption{Temporal evolution of the magnetic reconnection between H$\alpha$ fibrils observed by NVST. (a)-(d) Time slices of NVST H$\alpha$ images along the red, green, pink, and cyan lines AB, CD, EF, and GH in Figures\,\ref{f:mr_ha}(c), (d), (b), and (e), respectively. The green and blue dotted lines outline the motions of fibrils, with the moving speeds denoted by the numbers. The red dashed lines in (a)-(b) mark the reconnection region. The cyan dash-dotted lines in (a)-(d) separately show the times of H$\alpha$ images shown in Figures\,\ref{f:mr_ha}(c), (d), (b), and (e). See Section \ref{sec:mr1} for details.
\label{f:measurements_ha}}
\end{figure}
\begin{figure}[ht!]
\includegraphics[width=0.8\textwidth]{mr_aia1.eps}
\centering
\caption{Magnetic reconnection between loops before the filament eruption observed by SDO/AIA. (a) AIA 304 \AA, (b) 131 \AA, (c) 171 \AA, (d) 193 \AA, (e) 211 \AA, and (f) 335 \AA~images. The green solid arrows in (a)-(e) denote the current sheet. The cyan, green, and blue lines IJ, KL, and MN in (a) and (c) show the positions for time slices of AIA 304 \AA~and 171 \AA~images displayed in Figures\,\ref{f:measurements_aia1}(a), (b), and (d), respectively. The red and purple rectangles in (b) separately enclose the region for the DEM curve in Figure\,\ref{f:measurements_aia1}(e), and the location where the background emission is measured. In (c), the red pluses mark the positions between which the length of current sheet is measured, the purple line denotes the position along which the width of current sheet is calculated, and the red dotted lines outline the loops L2 and L4. The blue rectangle in (e) marks the region for the light curves of the AIA EUV channels as shown in Figure\,\ref{f:measurements_aia1}(c). The FOV is denoted by the red rectangle in Figure\,\ref{f:general_information}(a). See Section\,\ref{sec:mr1} for details.
\label{f:mr_aia1}}
\end{figure}
\begin{figure}[ht!]
\includegraphics[width=0.9\textwidth]{measurements_aia1.eps}
\centering
\caption{Temporal evolution of the magnetic reconnection between loops before the filament eruption observed by AIA. Time slices of (a) AIA 304\,\AA, and (b) and (d) 171\,\AA~images along the cyan, green, and blue lines IJ, KL, and MN in Figures\,\ref{f:mr_aia1}(a) and (c), respectively. (c) Light curves of the AIA EUV channels in the blue rectangle in Figure\,\ref{f:mr_aia1}(e). (e) DEM curve for a current sheet region enclosed by the red rectangle in Figure\,\ref{f:mr_aia1}(b). The blue and green dotted lines in (a), (b), and (d) separately outline the motions of loops and plasmoids, with moving speeds denoted by the numbers. The purple dashed line in (a) denotes the reconnection region. The red dashed line in (b) marks the middle of the current sheet. The blue vertical dashed line in (c) marks a peak of the EUV light curves. In (e), the black curve is the best-fit DEM distribution, and the red, green, and blue rectangles separately represent the regions containing 50\%, 51-80\%, and 81-95\% of the Monte Carlo solutions. See Section \ref{sec:mr1} for details.
\label{f:measurements_aia1}}
\end{figure}
\begin{figure}[ht!]
\includegraphics[width=0.66\textwidth]{filament_eruption.eps}
\centering
\caption{Filament eruption observed by AIA. (a) AIA 171 \AA, (b) and (d) 304 \AA, (c) 1600 \AA, (e) 131 \AA, and (f) 211 \AA~images. (g)-(h) Time slices of AIA 304\,\AA~and 211\,\AA~images along the green line PQ in (b) and along the VW direction in the blue rectangle in (f), respectively. The blue and green dotted lines in (g) separately outline the filament eruption and flare ribbon separation. The blue dotted line in (h) outlines the motion of loops. The moving speeds are denoted by the numbers in (g)-(h). The FOVs of (a)-(f) are denoted by the red rectangle in Figure\,\ref{f:general_information}(a). See Sections\,\ref{sec:mr1} and \ref{sec:filament_eruption} for details.
\label{f:filament_eruption}}
\end{figure}
\begin{figure}[ht!]
\includegraphics[width=0.66\textwidth]{mr_aia2.eps}
\centering
\caption{Magnetic reconnection between loops after the filament eruption observed by AIA. (a) AIA 304 \AA, (b) 131 \AA, (c) 171 \AA, (d) 193 \AA, (e) 211 \AA, (f) 335 \AA, (g) 94 \AA, and (h) 1600 \AA~images. The pink circles enclose the north endpoint of loops L2. The green solid arrows in (a)-(g) denote the current sheet. The green line RS in (c) shows the position for time slice of AIA 171 \AA~images as displayed in Figure\,\ref{f:measurements_aia2}(b). The red and purple rectangles in (d) separately enclose the region for the DEM curve in Figure\,\ref{f:measurements_aia2}(c), and the location where the background emission is measured. Same as in Figure\,\ref{f:mr_aia1}(e), the blue rectangle in (e) marks the region for the light curves of the AIA EUV channels as shown in Figure\,\ref{f:measurements_aia2}(a). The FOV is denoted by the red rectangle in Figure\,\ref{f:general_information}(a). An animation of the unannotated AIA images is available. It covers $\sim$5.4 hr starting at 01:00 UT, and the video duration is 12\,s. See Section \ref{sec:mr2} for details. (An animation of this figure is available.)
\label{f:mr_aia2}}
\end{figure}
\begin{figure}[ht!]
\includegraphics[width=0.8\textwidth]{measurements_aia2.eps}
\centering
\caption{Temporal evolution of the magnetic reconnection between loops after the filament eruption observed by AIA. (a) Light curves of the AIA EUV channels in the blue rectangle in Figure\,\ref{f:mr_aia2}(e). (b) Time slice of AIA 171 \AA~images along the green line RS in Figure\,\ref{f:mr_aia2}(c). (c) DEM curve for a current sheet region enclosed by the red rectangle in Figure\,\ref{f:mr_aia2}(d). In (a), the red vertical dotted line denotes the filament eruption, and the blue and green vertical dashed lines separately mark the peaks of the AIA 304 \AA~and 94 \AA~light curves. The green dotted line in (b) outlines the motion of plasmoids, with moving speed denoted by the number. In (c), the black curve is the best-fit DEM distribution, and the red, green, and blue rectangles represent the regions containing 50\%, 51-80\%, and 81-95\% of the Monte Carlo solutions, respectively. See Section \ref{sec:mr2} for details.
\label{f:measurements_aia2}}
\end{figure}
\begin{figure}[ht!]
\includegraphics[width=\textwidth]{cartoon.eps}
\centering
\caption{Schematic diagram of the magnetic reconnection between loops (fibrils) and its nearby filament eruption. The red, green, cyan, blue, yellow, and brown lines with arrows represent the directional magnetic field lines of loops (fibrils) L1, L2, L3, L4, L5, and L6, respectively. The pink line shows the filament. The red star denotes the reconnection point between field lines L2 and L4. The gray parallelogram represents the photosphere, and the black and white patches separately indicate the negative and positive magnetic fields N1, N2, and N3, and P1 and P2. See Section \ref{sec:sum} for details.
\label{f:cartoon}}
\end{figure}
\acknowledgments
The authors thank the referee for helpful comments. We are indebted to the NVST and SDO teams for providing the data. This work is supported by the Strategic Priority Research Program of Chinese Academy of Sciences, Grant No. XDB 41000000, the National Natural Science Foundations of China (12073042, 11673034, 11533008, 11790304, 11873059, 1111903050, and 11773039), and the Key Research Program of Frontier Sciences (ZDBS-LY-SLH013) and the Key Programs (QYZDJ-SSW-SLH050) of Chinese Academy of Sciences. We acknowledge the usage of JHelioviewer software \cite[][]{2017A&A...606A..10M} and NASA's Astrophysics Data System.
| 31383307c3c91c1fe912d0344b53ba66becce2c8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Star-forming regions in the Galaxy are distributed in a complex web of filaments that resemble a highly hierarchical network \cite[e.g.][]{2018A&A...610A..77H, 2010A&A...518L.100M, 2010A&A...518L.102A}. While open clusters are typically found in the densest parts of the structure, nearly 90\% of newborn stars become gravitationally unbound soon after the birth due to their dynamic interactions. Such loose ensembles of dispersing coeval stars are observed as stellar associations that keep the kinematic imprint of their local birth site up to $\sim$30~Myr before they become a part of the Galactic disk \citep{2019ARA&A..57..227K}. Because such groups of hundreds to thousands of stellar siblings were born from the same molecular cloud, they all have similar surface abundances \citep{2007AJ....133..694D}. These moving groups are thus the fossil records of the Galaxy that have a potential to link together star formation sites with the larger structures of the disk. They resemble an ideal laboratory to study a wide variety of important topics, from star- and planetary formation environments, the initial mass function and sequentially triggered star formation to dynamical processes that lead to the evaporation and finally the dispersal of an association.
A reliable reconstruction of stellar associations is thus of critical importance. While observations from the Hipparcos space astrometry mission allowed a major improvement in the search of overdensities in the kinematic phase space using stellar positions, parallaxes and proper motions \citep{1999AJ....117..354D}, it is high precision measurements from the Gaia space telescope -- including radial velocities for a subset of 7,000,000 stars -- that is revolutionising Galactic astrophysics \citep{2018A&A...616A...1G}. It has facilitated numerous attempts to study young stars above the main sequence and identify new members of the known moving groups in the Solar neighbourhood
\cite[e.g.][]{2018ApJ...862..138G, 2020MNRAS.491..215B}. Additionally
\citet{2018A&A...618A..93C} studied young populations on much larger Galactic scales and reported on the discovery of $\sim$1500 clusters.
Although a selection of the candidate members of a particular moving group is often based on the cuts in the kinematic space \cite[e.g.][]{2020arXiv200204801U},
the true nature of these groups appear to be diffuse due to their gradual dispersal. \citet{2019A&A...621L...3M} recently described extended structures emerging as the tidal streams of the nearby Hyades cluster, while \citet{2019A&A...623A.112D} found 11,000 pre-main sequence members of the Scorpius-Centaurus OB2 association residing in both compact and diffuse populations. Kinematic cuts in such cases are prone to be biased against the low-mass stars that are most likely to evaporate first.
Numerous works on young associations rely on multi-dimensional clustering algorithms. For example, \cite[e.g.][]{2019AJ....158..122K} report on the discovery of 1,900 clusters and comoving groups within 1~kpc with HDBSCAN (Hierarchival Density-Based Spatial Clustering of Applications with Noise described by \citealp{campello_density-based_2013}). However, the arrival of the Gaia's high precision parallaxes and proper motions enables reliable orbital simulations for the first time. For instance, \citet{2019MNRAS.489.3625C} were able to model an association at its birth time using Chronostar, perform its orbital trace-forward and blindly reconstruct the known Beta Pictoris association, reliably determine its members and, importantly, its kinematic age.
Stellar age is, besides the kinematics, one of the decisive parameters in the characterization of the young moving groups. Parallaxes of nearby stars with uncertainties better than 10\% enable the placement of stellar populations on the color-magnitude diagram. However, due to the numerous effects including the evolutionary model uncertainties and inflated radii on low-mass end of the population, the presence of binaries, background contamination and spread due to metallicity effects, and the variability of young stars, isochronal dating techniques remain non-trivial.
While gyrochronology relies on the multiple photometric measurements to determine the rotation period of a star, it is spectroscopic youth indicators that require only one observation for the estimation of stellar age.
Spectroscopic features of solar-like and cooler young stars up to the solar age are straightforward to observe. They emerge from the processes related to the magnetic activity of a star and manifest themselves in the excess emission in calcium H\&K lines (Ca~II~H\&K, 3969 and 3934~\AA; \citealt{2008ApJ...687.1264M}), H$\alpha$ line (6563~\AA; \citealt{2005A&A...431..329L}) and infrared calcium triplet (Ca~II~IRT; 8498, 8542 and 8662~\AA; \citealt{2013ApJ...776..127Z}). \citealp{2008ApJ...687.1264M} describe an age--activity relation that estimates age from the Ca~II~H\&K emission in the range from $\sim$10~Myr up to 10~Gyr, although \citealp{2013A&A...551L...8P} has shown later that there is no decay in chromospheric activity beyond 2~Gyr. The decline of the emission rate is the fastest in the youngest stars. Despite the variable nature of magnetic activity, especially in the pre-main sequence stars, it is easy to differentiate between stars of a few 10 and a few 100~Myr.
On the other hand, the presence of the lithium 6708~\AA~ line in GKM dwarfs directly indicates their youth and is a good age estimator for stars between 10-30~Myr -- which is a typical age of a stellar association.
Follow-up observations with the goal to detect the lithium line in young candidates have been performed by \citet{2019ApJ...877...60B} (who found lithium in 58 stars)
while \citet{2009A&A...508..833D}
report on the lithium measurements for $\sim$400 stars.
Over 3000 young K and M stars with a detectable lithium 6708~\AA~ line have recently been identified in the GALAH dataset \citep{2019MNRAS.484.4591Z}. While the majority of young early K dwarfs in the GALAH sample have practically settled on the main sequence, young late K and M stars with a detectable lithium line still reside 1~magnitude or more above the main sequence.
\citealp{2015MNRAS.448.2737R} have kinematically and photometrically selected candidate members of the Upper-Scorpius association and discovered 237 new members by the presence of lithium absorption.
In the Gaia era, the majority of stars in the Solar neighbourhood have parallaxes and proper motions precisely determined while spectroscopic age indicators and precise radial velocities are missing for a large fraction of low-mass young stars. Large spectroscopic surveys, such as GALAH \citep{2020arXiv201102505B}, typically avoid the crowded Galactic plane where most of the young stars reside.
This work aims to fill the gap and presents spectroscopic observations, their age indicators and radial velocities of 799 young star candidates within 200~pc with no pre-existing lithium measurements.
Section \ref{sec.data} describes the kinematically unbiased selection of all overluminous late K and early M stars within 200~pc.
We measure equivalent widths of the lithium absorption lines and the excess flux in Ca~II~H\&K and H$\alpha$ lines, as described in Section \ref{sec.youth_indicators}. Section \ref{sec.discussion} discusses age estimation and strategy success. The dataset is accompanied with radial velocities. Concluding remarks are given in Section \ref{sec.conclusions}.
\section{Data} \label{sec.data}
\subsection{Selection function}
Candidate young stars with Gaia magnitudes $10<G<14.5$ were selected from the \textit{Gaia}~DR2 catalogue \citep{2018A&A...616A...1G}. We focused only on the low-mass end of the distribution. The selection was based on their overluminosities in the colour-magnitude diagram. The colour index was chosen to be BP-W1 because it gives the narrowest main sequence with overluminous stars clearly standing out. BP is taken from \citet{2018A&A...616A...1G} and is described in more detail by \citet{2018A&A...616A...4E} while W1 is from \citet{2014yCat.2328....0C}. The relation used as a lower main sequence parametrisation $G(c)$
\begin{dmath}
G(c) = 4.717 \times 10^{-3} \; c^5 -0.149 \; c^4 + 1.662 \; c^3 - 8.374 \; c^2 + 20.728 \; c - 14.129
\end{dmath}
where G is absolute Gaia G magnitude and $c$=BP-W1 is described in more detail in \citet{2019MNRAS.484.4591Z} together with the arguments leading to the choice of BP-W1 being the best colour index for this purpose.
The colour-- temperature relation is determined from synthetic spectra while the temperature-spectral type relation is based on \citet{2013ApJS..208....9P}\footnote{In the version from 2018.08.02, available online: \url{http://www.pas.rochester.edu/~emamajek/EEM_dwarf_UBVIJHK_colors_Teff.txt}}.
Our criteria further exclude older stars and keep only objects that are found 1~magnitude or more above the main sequence. This approach largely avoids main sequence binaries (at most 0.75~mag above the main sequence).
The sample was color cut to include only stars between 3$<$BP-W1$<$5.6. This limit corresponds to K5-M3 dwarfs with $T_{\rm eff}=3400$--4400\,K and allows the optimisation of the observation strategy and a focus on the cool pre-main sequence objects with the fastest lithium depletion rate. The blue limit is chosen so that it minimises the contamination with subgiants but keeps most of the late K dwarfs in the sample. The red limit is set on the steep region of the lithium isochrones that divides early M dwarfs with the fast depletion processes from those cooler ones that need more than 100~Myr to show a significant change in lithium.
The upper luminosity boundary
\begin{equation}
G > G(c) - (1.33 c -3)
\end{equation}
rejects giants from the sample.
Since all stars disperse with time in the kinematic parameter space, young objects are found only in regions with low velocities. To avoid the kinematic bias towards the pre-selected clumps of young stars in the velocity parameter space that disfavors the low-mass stars, and to remove old stars, we compute the mean $UVW$ value of the sample and keep all objects within ($\pm$15, $\pm$15, $\pm$10)~$\mathrm{km\,s^{-1}}$ of the median $UVW$ = (-11.90, 215.77, 0.19)~$\mathrm{km\,s^{-1}}$.
No kinematic cut was performed on stars that have no radial velocities available in the Gaia catalogue \citep{2018A&A...616A...6S}.
A declination cut with $\delta<30\,\mathrm{deg}$ eliminated objects not visible from the Siding Spring Observatory, Australia, where the observations took place.
Known young stars
from the Simbad database and stars observed with the GALAH survey \citep{2018MNRAS.478.4513B} were removed from the list to maximise survey efficiency at detecting new young stars. This selection results in 799 candidate stars.
Finally, our sample of stars described in this work includes observations of 756 candidate objects from this list.
A color-magnitude diagram with all the candidates is shown in Figure \ref{fig.cmd}. Parallaxes are taken from Gaia~DR2 \citep{2018A&A...616A...1G}.
\begin{figure}
\includegraphics[width=\columnwidth]{cmd.pdf}
\caption{Colour-magnitude diagram with candidate young stars and their reddening estimated in this work.
Details on the reddening estimation are described in Sec. \ref{sec.reddening}. The most crowded region ($\sim$ K5 dwarfs) is contaminated with reddened hotter stars while M dwarfs show less contamination due to their proximity.
Red lines denote the main sequence (dashed line) and the selection function 1 magnitude above (solid line).
Contours show the density of stars in the Gaia catalog.
}
\label{fig.cmd}
\end{figure}
\subsection{Observations}
Observations were carried out between November 2018 and October 2019 over 64 nights with the ANU 2.3m telescope at Siding Spring Observatory.
In order to achieve better radial velocity precision, 349 stars brighter than G=12.5 were observed with the slit-fed Echelle spectrograph in the Nasmyth focus, covering wavelengths between $\sim$3900 and $\sim$6750~\AA~ at R=24,000. Exposure times were between 600~sec for the brightest and 1800~sec for the faintest objects, resulting in a typical S/N of 20 in the order containing the H$\alpha$ line. Blue wavelengths with the calcium H\&K lines have poor S/N but clearly show strong emission above the continuum when present (Fig. \ref{fig.echelle_calcium}). The spectra were reduced as per \citet{2014MNRAS.437.2831Z}. Wavelength calibration was provided by bracketing Thorium-Argon lamp exposures.
Fainter stars (449) between $12.5<G<14.5$ were observed with the Wide Field Spectrograph (WiFeS; \citealt{2007Ap&SS.310..255D}), namely with resolving power of 3000 in the blue and 7000 in the red, covering 3500-7000~\AA.
We typically used a RT480 beam splitter.
Typical exposure times were 5~minutes per star that resulted in the median S/N of 13 and 31 for the blue and the red band, respectively. Thorium-Argon lamp frames were taken every hour to enable wavelength calibration. WiFeS spectra were reduced with a standard PyWiFeS package \citep{2014Ap&SS.349..617C}, updated to be better suited for stellar reductions of a large number of nights.
\subsection{Synthetic Spectra} \label{sec.synthetic}
For computation of radial velocities and parameter estimation, we use a template grid of 1D LTE spectra that was previously described by \citet{2019MNRAS.488L.109N}.
Briefly, spectra were computed using the TURBOSPECTRUM code (v15.1; \citealt{1998A&A...330.1109A, 2012ascl.soft05004P}) and MARCS model atmospheres \citep{2008A&A...486..951G}.
For models with $\log\,g > 3.5$, we use $v_{\rm mic} = 1\,{\rm km\,s^{-1}}$; for models with $\log\,g \le 3.5$, we use $v_{\rm mic} = 2\,{\rm km\,s^{-1}}$ and perform the radiative transfer calculations under spherical symmetry taking into account continuum scattering.
The spectra are computed with a sampling step of $1\,$km$\,$s$^{-1}$, corresponding to a resolving power $R\sim300\,000$. We adopt the solar chemical composition and isotopic ratios from \citet{2009ARA&A..47..481A}, except for an alpha enhancement that varies linearly from $\text[\alpha / \text{Fe}] = 0$ when $\rm [Fe/H] \ge 0$ to $\text[\alpha/\text{Fe}] = +0.4$ when $\rm [Fe/H] \le -1$.
We use a selection of atomic lines from VALD3 \citep{2015PhyS...90e4005R} together with roughly 15 million molecular lines representing 18 different molecules, the most important of which for this work are CaH (Plez, priv. comm.), MgH \citep{2003ApJS..148..599S,1995ASPC...78..205K}, and TiO \citep[with updates via VALD3]{1998A&A...337..495P}.
We use this grid to generate two synthetic libraries for radial velocity determination and parameter estimation.
For the WiFeS spectra, we use a coarsely sampled version of this grid, broadened to $R\sim7000$ with $5400 \leq \lambda \leq 7000$, $3000 \leq T_{\rm eff} \leq 8000\,$K, $3.0 \leq \log g \leq 5.5$, and $-1.0 \leq $[Fe/H]$ \leq 0.5$, in steps of $100\,$K, $0.25\,$dex, and $0.25\,$dex respectively.
For the Echelle spectra, we adopted R=24,000 for $3000 \leq T_{\rm eff} \leq 6000\,$K, $4 \leq \log g \leq 5$, and [Fe/H]=0, in steps of $250\,$K and $0.5\,$dex, respectively. Additionally, $\log g$ for $T_{\rm eff}<4000\,$K was extended to 5.5. Spectra cover wavelengths from 4800 to 6700~\AA.
\subsection{Radial velocities}
Radial velocities for datasets from both instruments were determined with the same algorithms using synthetic spectra described in the previous section.
\subsubsection{WiFeS}
Radial velocities of the WiFeS R7000 spectra were determined from a least squares minimisation of a set of synthetic template spectra varying in temperature (see Section \ref{sec.synthetic} for details of model grid). We use a coarsely sampled version of this grid, computed at R$\sim7000$ over $5400 \leq \lambda \leq 7000$ for $3000 \leq T_{\rm eff} \leq 5500\,$K, $\log g = 4.5$, and [Fe/H]$= 0.0$, with $T_{\rm eff}$ steps of $100\,$K for radial velocity determination.
Prior to computing radial velocities, we normalise both our observed and synthetic template spectra. For warmer stars without the extensive molecular bands and opacities present in cool stars, continuum regions are typically used to continuum normalise the spectrum. For observed cool star spectra however, such regions are unavailable in the optical, so we must opt for another normalisation formalism, which we term here \textit{internally consistent normalisation}:
\begin{equation}
f_{\rm norm} = f_{*} \times e^{\big(a_0 + \frac{a_1}{\lambda}+\frac{a_2}{\lambda^2}\big)}
\end{equation}
where $f_{\rm norm}$ is the internally consistent normalised flux vector, $\lambda$ is the corresponding wavelength vector, and $a_0$, $a_1$, and $a_2$ are coefficients of a second order polynomial fitted to the logarithm of $f_{*}$, which is either an observed flux corrected spectrum, or a synthetic template. This functional form of normalisation has chosen to be largely independent of reddening.
Once generated, a given synthetic template (initially in the rest frame) can be interpolated and shifted to the science velocity frame as follows:
\begin{equation}
f_{\rm temp,~rvs} = f_{\rm t}\big[\lambda \times \big(1-\frac{v_r-v_b}{c}\big)\big]
\end{equation}
where $f_{\rm temp,~rvs}$ is the RV shifted normalised template flux, $f_{\rm temp}$ is the template flux in the rest frame, $v_r$ and $v_b$ the radial and barycentric velocities respectively, and $c$ is the speed of light. $v_b$ is computed using the \texttt{ASTROPY} package \citep{astropy:2018} in \texttt{PYTHON}.
Given a grid of $k$ different synthetic template spectra, the final radial velocity value is found by finding the synthetic template that best minimises:
\begin{equation}
R(v_r) = \displaystyle\sum_{j}^{N}\bigg(\frac{{f_{\rm obs,~j} - f_{\rm temp,~rvs,~k,~j}(v_r)}}{\sigma_{{f_{\rm obs,~j}}}}\bigg)^2 M_j
\end{equation}
where $R$ is the total squared residuals as a function of radial velocity offset, $j$ is the pixel index, $N$ the total number of spectral pixels, $f_{\rm obs,~j}$ is the normalised observed flux at pixel $j$, $\sigma_{{f_{\rm obs,~j}}}$ is the uncertainty on $f_{\rm obs,~j}$, and $M_j$ is a masking term set to either 0 or 1 for each pixel. This step is done twice for each template spectrum, initially masking out only pixels affected by telluric contamination (H$_2$O: 6270-6290$\,$\SI{}{\angstrom}, and O2: 6856-6956$\,$\SI{}{\angstrom}), but then additionally masking out further pixels with high fit residuals. This second mask has the effect of excluding any pixels likely to skew the fit such as science target emission not present in the synthetic template (such as H$\alpha$).
Least squares minimisation was done using the leastsq function from \texttt{PYTHON}'s \texttt{SCIPY} library, implemented in the \texttt{PYTHON} package \texttt{plumage}\footnote{\url{https://github.com/adrains/plumage}}. Statistical uncertainties on this approach are on average 430$\,$m$\,$s$^{-1}$, however per the work of \citet{2018MNRAS.480.5099K} we add this in quadrature with an additional 3$\,$km$\,$s$^{-1}$ uncertainty to account for WiFeS varying on shorter timescales than our hourly arcs can account for, and effects of variable star alignment on the slitlets. Note however that we do not employ corrections based on oxygen B-band absorption, demonstrated by \citealt{2018MNRAS.480.5099K} to improve precision, as such additional precision is unnecessary for this work and is difficult for cooler stars.
Comparison of radial velocities for cool dwarf standard stars (e.g. from \citealp{2015ApJ...804...64M} and \citealp{2012ApJ...748...93R}, observed with the same instrument setup as part of Rains et al. in prep) with the Gaia catalogue \citep{2018A&A...616A...6S} shows an offset of WiFeS values for -1.7~$\mathrm{km\,s^{-1}}$ and a standard deviation of 3.2~$\mathrm{km\,s^{-1}}$ (Figure \ref{fig.rvs}). We suspect that most of the outliers are binary stars. Some of them are confirmed by either visual inspection or significally different radial velocities in case of repeated observations while there is not enough information available to investigate the rest of the interlopers.
\begin{figure}
\includegraphics[width=\linewidth]{rv_wifes_gaia_comparison.pdf}
\caption{A comparison between radial velocities from Gaia and from our pipeline for the WiFeS spectra. Standard stars (blue) have high S/N and small uncertainties. Binary star candidates (stars with repeated observations that show standard deviation of radial velocities greater than $5\,\mathrm{km\,s^{-1}}$ and stars that were classified as binaries by visual inspection) are marked in red.
}
\label{fig.rvs}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth]{rv_echelle.pdf}
\caption{A comparison between radial velocities from Gaia and our Echelle pipeline. Stars with the biggest disagreement with Gaia appear to be binary star candidates (red circles) or active (measured by calcium~II~H\&K emission $\mathrm{\log{R'_{HK}}}$, see Section \ref{sec.calcium}). The match with best-fitting template has been visually inspected for all stars in the sample.
}
\label{fig.rvs_echelle}
\end{figure}
\subsubsection{Echelle}
The same routine was utilized for the Echelle spectra on wavelengths between 5000 and 6500~\AA~ using their own synthetic library described in Sec. \ref{sec.synthetic}. As the correction for the blaze function and flux calibration were not performed in the data reduction step, each order within the relevant wavelength range was continuum normalized with a low order polynomial. Orders were then combined together into one spectrum in the range between 5000 and 6500~\AA. To match the continua of measured and synthetic libraries, fluxed model spectra were cut into wavelengths corresponding to Echelle orders, normalized and stitched back together with the same procedure. Finally, synthetic spectra were scaled to match 90th percentile of Echelle continua.
All spectra were visually inspected for any major reduction issues or other sources of peculiarity. Obvious double-lined binaries were flagged and their radial velocities are not reported in this work. Binary detection is reported in a separate column in Table \ref{tab.results}.
Median internal uncertainty of derived radial velocity is 0.06~$\mathrm{km\,s^{-1}}$, but a combination of the systematic uncertainty and radial velocity jitter characteristic to young stars account for 1.5~$\mathrm{km\,s^{-1}}$.
Most of the stars have radial velocities consistent with Gaia (Figure \ref{fig.rvs_echelle}). Mean absolute deviation for stars with difference less than 10$\mathrm{km\,s^{-1}}$ is 0.6$\mathrm{km\,s^{-1}}$. There are a handful of outliers, and they all have large uncertainties in Gaia values. Some of those appear to be binary stars discovered either by visual inspection or large radial velocity difference in case of the repeated measurements. At the same time, a lost of such stars show high activity level (depicted by a measure of activity in calcium HK lines) that might dominate Gaia's calcium infrared triplet region used to determine radial velocities and cause systematic offsets. All Echelle stars have been visually inspected for possible peculiarity and their match with the best-fitting template.
\subsection{Reddening} \label{sec.reddening}
The M dwarf candidates are too close to be significantly reddened (<200~pc), but on the other hand they could remain embedded in their birth cocoons. At the same time, the sample is contaminated with hotter stars that lie in regions of more heavy extinction within the Galactic plane.
To derive an estimate for the intrinsic colour index (BP-W1)$_0$, temperatures of the best-matching templates were used as an input in the colour-temperature relation derived from the synthetic spectral library. Although Solar values were used to calibrate the zero point, a degree of uncertainties remains (increasing with colour) and the relation is only approximate.
The resulting E(BP-W1) reveals a number of interlopers with temperatures higher than 4500~K. In particular, 156 WiFeS stars have E(BP-W1)$>$1 (20\% of the entire sample).
The estimated reddening E(BP-W1) is presented in Figure \ref{fig.cmd} together with the reddening vector\footnote{Reddening vector is determined for $A_V = 1$ and $R_V = 3.1$ from the \citet{1989ApJ...345..245C} model - \texttt{ccm89} in \url{https://extinction.readthedocs.io/en/latest/index.html}.}. Most interlopers with high reddening are found in the two regions in the Galactic plane with the highest concentration of stars in our sample: the Hyades and the Scorpius-Centaurus OB2 region (Fig. \ref{fig.galaxy}). Further analysis revealed that these stars do not show signs of youth and are likely located behind the local dust clouds associated with star-forming regions.
\begin{figure}
\includegraphics[width=\columnwidth]{gal.pdf}
\caption{The distribution of young candidates in the Galaxy. The majority of stars is found in clumps suggesting that they still reside close to their birth sites. The biggest group is found in the direction of the Scorpius-Centaurus OB2 region ($l>280$~deg). The second clump is likely the Hyades stars ($l\sim180$~deg). Colours show the interstellar reddening E(BP-W1).}
\label{fig.galaxy}
\end{figure}
\section{Youth indicators} \label{sec.youth_indicators}
The following subsections address the characterization of the lithium absorption line and the excess emission in H$\alpha$ and Ca~II~H\&K lines for stars in our sample. A combination of all three values provides a robust indicator of the stellar youth. Algorithms used to measure the strengths of lithium and H$\alpha$ lines in this work are similar for data from both instruments WiFeS and Echelle. Excess emission in calcium is measured differently for Echelle due to low signal in the blue.
All spectra, except the WiFeS calcium region, were locally normalized so that the youth features are surrounded by the continuum at 1 (and pseudo-continuum in M dwarfs). Binaries were not treated separately in this work and we provide youth indicators regardless of stars' multiplicity. All spectra were visually inspected for multiplicity and high rotation rate. We flag such cases in the final table and emphasize that this is qualitative inspection only and it is not complete.
\subsection{Lithium}
The primary and most reliable spectroscopic feature sensitive to the age of the pre-main sequence dwarfs in the temperature range observed in our sample is the lithium 6708~\AA~ line. This absorption line is observed in low-mass pre-main sequence stars before the ignition of lithium in their interiors. Since these stars appear to be fully convective before their onset on the main sequence, the depletion of lithium throughout the entire star occurs almost instantly. Lithium is observed in F, G and early K dwarfs for up to $\sim$100~Myr (mass dependent), but late K and early M-type dwarfs deplete lithium much faster. For further information see \cite{2014prpl.conf..219S} and references therein. Both data and theoretical predictions show that at the age between $\sim$15-40~Myr there is practically no lithium left in these stars \citep{2015A&A...577A..42B, 2019MNRAS.484.4591Z}.
The strength of the lithium absorption lines in this work was characterized with the equivalent widths measured within 6707.8$\pm$1.4~\AA.
Our spectra were pseudo-continuum normalized with a second order polynomial between 6700 and 6711~\AA. The lithium line was excluded from the continuum fit. The equivalent width was defined to be positive for lines in the absorption and was measured from the continuum level of 1.
In contrast to the emission-related features superimposed on the photospheric spectrum, the lithium absorption line shows a certain degree of correlation with the stellar rotation rate, e.g. \citet{2018A&A...613A..63B}. Fast rotators found by visual inspection are flagged in the table with results. While it appears to be fairly insensitive to the chromospheric activity \citep[e.g.][]{2019MNRAS.490L..86Y}
it might in some cases be affected by strong veiling present in the classical T~Tauri stars \citep{1989AJ.....98.1444S}. Veiling is an extra source of continuum that causes absorption lines to appear weaker \citep{1990ApJ...363..654B}. However, measurements of H$\alpha$ emission described below reveal that no classical T~Tauri stars are present in the sample. Figure \ref{fig.youth_indicators} confirms a robust correlation between all three measures of the youth.
The distribution of EW(Li) shows a concentration of stars below 0.05~\AA, though we only consider positive detections in stars with values above this level.
Repeated observations (45 stars) show 0.02~\AA~ of variation between individual measurements of the same object.
\subsection{Calcium~II~H\&K} \label{sec.calcium}
It has long been known that atmospheric features associated with stellar activity in solar-like dwarfs anticorrelate with their age \citep{1972ApJ...171..565S, 1991ApJ...375..722S}. Empirical relations derived from chromospheric activity proxies enable age estimation of stars between $\sim$0.6-4.5~Gyr to a precision of $\sim$0.2~dex \citep{2008ApJ...687.1264M}. However, a combination of saturation \citep{2010ApJ...709..332B} and high variability \citep{1995ApJ...438..269B} of activity in younger stars prevents this technique yielding reliable results before the age of $\sim$200~Myr.
Nevertheless, a detection of a strong excess emission in the calcium~II~H\&K lines (Ca~II~H\&K; 3968.47 and 3933.66~\AA, respectively) -- a proxy for chromospheric activity -- helps to distinguish between active young stars and older stars with significantly lower emission rates.
A commonly used measure of stellar activity in solar-type stars is S-index introduced by \citealp{1978PASP...90..267V} and derived as
\begin{equation}
S = \alpha \frac{N_H + N_K}{N_V + N_R},
\end{equation}
where $N_H$ and $N_K$ are the count rates in a bandpass with a width of 1.09~\AA~ in the center of the Ca~II~H and K line, respectively. To match the definition of the first measurements obtained by a spectrometer at Mount Wilson Observatory \citep{1978ApJ...226..379W} and make the measurements directly comparable, counts are adjusted to the triangular instrumental profile as described in \citealp{1978PASP...90..267V}.
$N_V$ and $N_R$ are the count rates in 20~\AA-wide continuum bands outside the lines, centered at 3901.07~\AA~ and 4001.07~\AA.
Constant $\alpha$ is a calibration factor that accounts for different instrument sensitivity and is derived by a comparison with literature S values. For WiFeS we provide a linear relation that converts measured S value on a scale directly comparable with the literature. For derivation see Appendix \ref{sec.appendix_s_index}.
Since $N_V$+$N_R$ has a color term due to nearby continuum shape varying with temperature, and because $N_H + N_K$ accounts for both chromospheric and photospheric contribution, it is more convenient to use the $\mathrm{R'_{HK}}$ index (first introduced by \citealp{1979ApJS...41...47L}) that represents a ratio between the chromospheric and bolometric flux and enables a direct comparison of activity in different stellar types. Using the conversion factor $C_{cf}$ that describes the colour-dependent relation between the S-index and the total flux emitted in the calcium lines, and $\mathrm{R_{phot}}$ that removes the photospheric contribution from the total flux in calcium, $\mathrm{R'_{HK}}$ is obtained as
\begin{equation}
\mathrm{R'_{HK} = R_{HK} - R_{phot}}
\end{equation}
where $\mathrm{R_{HK}} = 1.887 \times 10^{-4} \times C_{cf} \times S$. The constant in the equation is taken from \citealp{2017A&A...600A..13A}.
\citealp{1982A&A...107...31M} and \citealp{1984A&A...130..353R} provide the calibration of $C_{cf}$ and \citealp{1984ApJ...279..763N} and \citealp{1984ApJ...276..254H} for $\mathrm{R_{HK}}$ for the main sequence stars, but their relations become increasingly uncertain above B-V$>$1.2. \citealp{2017A&A...600A..13A} have recently extended the relation to M6 dwarfs (B-V$\sim$1.9) using HARPS data and calibrated the relation for colours that are more suitable for cool stars:
\begin{align}
\log_{10}{C_{cf}} & = - 0.005c^3 + 0.071c^2 - 0.713c + 0.973 \\
\log_{10}{R_{phot}} & = - 0.003c^3 + 0.069c^2 - 0.717c - 3.498.
\end{align}
where $c=$V-K was determined from a low-order polynomial fit to the relation between synthetic BP-RP and V-K from \citealp{2018MNRAS.479L.102C}.
There are 26 stars in the sample with repeated observations. In general more active stars show higher variability rates. We provide a median value of 1.1$ \times 10^{-5}$ for the $\mathrm{R'_{HK}}$ variability. Stars with low levels of activity have measured $\log{\mathrm{R'_{HK}}}=$ -5 or lower and we consider them inactive.
Activity in the Echelle spectra was evaluated in the same way as WiFeS stars. The calibration of the S-index was done using 19 stars observed with both instruments. For more details on the calibration see Appendix \ref{sec.appendix_s_index}.
The distribution of $\log{\mathrm{R'_{HK}}}$ is known to be bimodal for the main sequence stars in the Solar neighbourhood (e.g. \citealp{2003AJ....126.2048G}). Figure \ref{fig.rhk} shows two peaks, but they are centered at higher levels of activity due to our focus on the pre-main sequence stars. The more active peak is found at $\sim -4$ where $\log{\mathrm{R'_{HK}}}$ saturates for stars with rotation rates less than 10~days \citep{2017A&A...600A..13A}. According to \citealp{2008ApJ...687.1264M}, such high activity levels occur at ages of $\sim$10~Myr. We also plot $\log{\mathrm{R'_{HK}}}$ versus colour (the same figure) to confirm that the colour term is minimized.
There are two sets of lines that cause strong emission in this wavelength range: calcium~II~H\&K lines and Balmer emission lines in the youngest stars. Calcium H line is in some cases strongly blended by the Balmer emission line in the WiFeS spectra but count rate was measured within 1.09~\AA~ (see Fig. \ref{fig.wifes_calcium_balmer}).
\begin{figure*}
\includegraphics[width=\linewidth]{echelle_calcium.pdf}
\caption{Calcium lines in the Echelle spectra. Strong emission lines are detectable despite a low signal-to-noise ratio. There is an indication of a weak Balmer emission line at 3970~\AA. The red line is an average spectrum with a marginally detectable calcium emission while the blue line represents an average very active spectrum. Thick black line is a median inactive spectrum. Spectra in this plot were convolved with a smoothing kernel with the of width 7 for noise reduction purposes.}
\label{fig.echelle_calcium}
\end{figure*}
\begin{figure*}
\includegraphics[width=\linewidth]{wifes_calcium_and_balmer.pdf}
\caption{Calcium lines in the WiFeS spectra. Ca~II~H line appears to be wider than Ca~II~K due to the presence of the Balmer emission line at 3970~\AA. Red spectrum is a median spectrum with $\mathrm{logR'_{HK}}<$-4.9. Very active spectra with $\mathrm{logR'_{HK}}>$-4.4 (green) are young and show Balmer emission.
}
\label{fig.wifes_calcium_balmer}
\end{figure*}
\begin{figure}
\includegraphics[width=\columnwidth]{rhk_distribution.pdf}
\caption{\textit{Upper panel:} The introduction of the $\mathrm{logR'_{HK}}$ index minimises the color term and allows for comparison of activity rates among different spectral types. \textit{Lower panel:} Distribution of $\mathrm{logR'_{HK}}$ index for 680 stars. Nearly all stars with a detectable lithium show very strong calcium emission.
}
\label{fig.rhk}
\end{figure}
\subsection{Balmer series}
While weak and moderate excess emission rates in the H$\alpha$ line (6562.8~\AA) are associated with chromospheric activity \cite[e.g.][]{2004AJ....128..426W, 2008AJ....135..785W}, strong emission in the entire Balmer series, with H$\alpha$ being especially prominent ($>$10~\AA), is typically observed in classical T~Tauri stars that are low-mass objects younger than $\sim$10~Myr \citep{1989ARA&A..27..351B, 1989A&ARv...1..291A, 1998AJ....115..351M, 2006MNRAS.370..580K, 2014prpl.conf..219S}. It is widely accepted that there is a tight correlation between the average chromospheric fluxes emitted by the Ca~II~H\&K and H$\alpha$ lines \cite[e.g.][]{1995A&A...294..165M}. Although \citet{2007A&A...469..309C} report that this relation is more complicated, emission in H$\alpha$ represents a robust indicator of stellar youth.
Characterisation of stellar activity from the H$\alpha$ line is especially convenient in late-type dwarfs that only present a weak photosphere in the blue where Ca~II~H\&K are located.
The equivalent width of H$\alpha$ lines was measured between 6555 and 6567~\AA~ relative to the continuum, e.g. (1~-~flux) in the H$\alpha$ region. Negative values thus indicate absorption while positive values denote emission above the continuum. Interpretation of these results is not straightforward due to a wide range of the H$\alpha$ line profiles being strongly affected not only by the temperature but also the surface gravity. However, most of the stars show strong emission that is in any case an indicator for extreme stellar youth. We make a conservative estimate and only treat spectra with EW(Ha)$>$-0.5~\AA~ as active (see Fig. \ref{fig.wifes_calcium_balmer}).
Repeated observations of 45 stars reveal a typical difference between the maximal and minimal EW(Ha) value of 0.2~\AA. This uncertainty might also include a variability component of stellar activity.
Based on equivalent widths of H$\alpha$, most of the stars with excess emission belong to either weak (EW(Ha)$<$5~\AA) or post-T~Tauri stars. One third of the entire sample shows emission in the entire Balmer series. Column \texttt{Balmer} in Table \ref{tab.results} lists objects with clear Balmer emission that was detected by visual inspection.
\section{Discussion} \label{sec.discussion}
A combination of the three complementary youth features -- excess emission in Ca~II~H\&K and H$\alpha$ associated with magnetic fields active but declining for billions of years, and lithium absorption line present for a few 10~Myr in late K and early M dwarfs -- maximises the estimated age range and the robustness of our young star identification.
\begin{figure}
\includegraphics[width=\columnwidth]{activity_correlation.pdf}
\caption{Youth indicators studied in this work show a high degree of correlation. Chromospheric activity in young stars shows a high level of variability over time, but there appears to be a lower limit for H$\alpha$ emission with respect to the strength of the lithium line. Stars with no $\log{\mathrm{R'_{HK}}}$ available are marked with circles.
}
\label{fig.youth_indicators}
\end{figure}
This work uncovered 549 sources with at least one of the three indicators above the detection limit:
EW(Li)$>$0.1~\AA~ or EW(Ha)$>$-0.5~\AA~ or $\log{\mathrm{R'_{HK}}}>-4.75$. The strategy is thus 70\% successful. In particular, there are 281 stars with all three indicators above the detection limit.
There are 346 stars with a detectable lithium line (44\%), 479 with $\mathrm{EW(H}\alpha)>-0.5$ (60\% of the sample) and 464 objects (60\%) with a detectable calcium emission. Not surprisingly, there are 409 stars that show both calcium and H$\alpha$ youth features, as these two indicators are well correlated due to their common origin in chromospheric activity.
The lithium absorption line undergoes a different mechanism (lithium depletion in the pre-main sequence phase) and is much more short-lived. This causes an overdensity of chromospherically active stars with high H$\alpha$ but no lithium left (Fig. \ref{fig.youth_indicators}).
There are 10 stars in the sample that display lithium absorption but show no chromospheric activity.
The figure also shows that all stars with strong lithium emit excess flux in their chromospheres. This explains the void in the bottom right part of this figure.
Note that a small subset of individual stars only has one or two youth indicators measured due to noise in the respective spectral regions.
All youth indicators, radial velocities and flags denoting Balmer emission, binarity and fast rotation are listed in Table \ref{tab.results}, together with their 2MASS identifiers \citep{2003yCat.2246....0C}.
Even though our selection avoided known young stars, we cross-matched our catalogue with the literature. We found 15 stars in common with the list of association members described by \citet{2018ApJ...856...23G} and 6 from \citet{2018ApJ...860...43G}. We found 9 objects from our list in the work by \citet{2009A&A...508..833D} measuring lithium lines of $\sim$400 objects, and 3 overlapping stars with \citealp{2015MNRAS.448.2737R} who targeted stars from Upper Scorpius that were mostly fainter than our magnitude limit.
In total, 33 unique objects out of 766 from our list (4\%) are known association members or have lithium measured in the literature, and the rest are considered new detections.
The occurrence rate for all youth features is color dependent (Fig. \ref{fig.strategy_success}). Cooler stars in general more likely show signs of youth. Due to their slower evolution they spend more time above the main sequence and display signs of their youth much longer. However, we observe a drop in the occurrence rate of the lithium line in M dwarfs. This is because they deplete lithium the fastest and soon fall below the detection limit.
Lithium isochrones enable age estimation for late K and early M dwarfs younger than 15-40~Myr. We follow \citet{2019MNRAS.484.4591Z} and take indicative non-LTE equivalent widths from \citet{1996A&A...311..961P} for Solar metallicity and $\log{g} = 4.5$. We combine them with the \citet{2015A&A...577A..42B} models of lithium depletion (assuming the initial absolute abundance of 3.26 from \citealp{2009ARA&A..47..481A}) to compute lithium isochrones (Fig. \ref{fig.isochrones}).
Lines indicating abundances in the plot show that EW(Li) in our temperature range practically traces any amount of lithium left in the atmosphere.
There appears to be an overdensity of 278 objects above EW(Li)$>$0.3~\AA~ corresponding to the ages of 15~Myr and younger. Moreover, there are 325 stars lying above the 20~Myr isochrone and the 0.1~\AA~ detection limit.
Figure \ref{fig.youth_indicators} confirms that stars with the strongest lithium have the highest $\log{\mathrm{R'_{HK}}}$ values of -4 which corresponds to $\sim$10~Myr according to the \citealp{2008ApJ...687.1264M} activity-age relation.
These objects likely belong to the Scorpius-Centaurus association -- especially because their $(l,b)$ location overlaps with this region in the sky. However, further kinematic analysis is needed to confirm their membership.
Since our selection encompass all stars above the main sequence, the sample is contaminated with stars with bad astrometric solutions. 45\% of our observed objects have \textit{re-normalised unit weight error} (the \texttt{RUWE} parameter from the Gaia~DR2 tables describing the goodness of fit to the astrometric observations for a single star) greater than 1.4. Gaia~DR2 documentation suggests that such stars either have a companion or their astrometric solution is problematic. There is no detectable lithium left in these stars and they appear to be old in our context with low or zero emission in calcium and H$\alpha$. When stars with \texttt{ruwe}$>$1.4 and high reddening are removed from our catalog, 80\% of stars left show at least one spectroscopic sign of stellar youth. This suggests a high efficiency in selection of young stars from the Gaia catalog based on their overluminosity and a reliable astrometric single star solution.
\begin{figure}
\includegraphics[width=\columnwidth]{strategy_success.pdf}
\caption{Strategy success as a fraction of young stars with detectable spectroscopic features of youth versus their color. Detection rate for calcium and H$\alpha$ increase towards redder stars with different slopes. This might be due to a dependence of EW(H$\alpha$) on the temperature. Lithium absorption line is observed only in the youngest stars. Detection rate drops for early M dwarfs because they deplete lithium the fastest.
The number of all candidates in each colour bin is shown in the plot.
}
\label{fig.strategy_success}
\end{figure}
\begin{figure*}
\includegraphics[width=\linewidth]{lithium_color.pdf}
\caption{Lithium isochrones (blue lines) reveal a number of very young stars in the sample ($<$25~Myr). 349 stars have a detectable lithium with EW(Li)$>$0.1~\AA. Black lines show lithium abundances with their uncertainties (dashed). Lithium strength correlates well with the excess emission in the H$\alpha$ line.
}
\label{fig.isochrones}
\end{figure*}
\section{Conclusions} \label{sec.conclusions}
We selected and observed 766 overluminous late K and early M dwarfs with at least 1~magnitude above the main sequence and with Gaia G magnitude between 12.5 and 14.5. The kinematic cut was wide enough to avoid a bias towards higher-mass stars and include low-mass dwarfs. Observations were carried out over 64 nights with the Echelle and Wide Field Spectrographs at the ANU 2.3m telescope in Siding Spring observatory.
The analysis revealed 544 stars with at least one feature of stellar youth, i.e. the lithium absorption line or excess emission in H$\alpha$ or calcium~H\&K lines. The strength of the lithium absorption line indicates that 349 stars are younger than 25~Myr.
This sample significantly expands the census of nearby young stars and adds 512 new young stars to the list. Only 33 out of 544 objects with at least one youth indicator are listed in external catalogs of young stars. For example, \citealp{2018ApJ...856...23G} characterised known nearby associations and provided a list of 1400 young stars from a wide variety of sources. Our catalog has only 15 stars in common with theirs and thus expands the sample by 35\%. Although a further kinematic analysis is needed to confirm their membership, it is likely that a great fraction of stars from our sample belong to the Scorpius-Centaurus association because they are found in that direction in the sky and all have lithium ages $<$20~Myr. However, we only find 3 stars in common with \citealp{2015MNRAS.448.2737R} who kinematically and photometrically selected and observed mostly fainter stars in Upper-Scorpius.
Strong lithium absorption lines and excess emission in calcium in these objects consistently indicate likely stellar ages of roughly 10~Myr, according to the activity--age relation \citep{2008ApJ...687.1264M} and lithium isochrones (see Fig. \ref{fig.isochrones}). The latter reveal 325 stars with $\mathrm{EW(Li)>0.1}$~\AA~ and above the 20~Myr isochrone.
We report on a high success rate in search for young stars by selecting overluminous objects in the Gaia catalog. After stars with unreliable astrometry ($\mathrm{ruwe}>1.4$ that indicates bad astrometry or multiplicity) and high reddening are removed, the success rate is 80\%.
Radial velocities are determined for spectra from both instruments, with average uncertainties of 3.2~$\mathrm{km\,s^{-1}}$ for WiFeS and 1.5~$\mathrm{km\,s^{-1}}$ for Echelle stars.
This catalog of nearby young stars now has all kinematic measurements available to improve the analysis of young associations and help to find their birthplace. For example, \citealp{2020MNRAS.499.5623Q} have recently shown that stellar associations come from different places in the Galaxy. Follow up work may include e.g. using Chronostar \citep{2019MNRAS.489.3625C} to provide kinematic ages, robust membership estimates and orbital models of young associations to infer the origins of this sample, as well as the extraction and analysis of rotational periods using TESS to obtain ages using gyrochronology where possible.
\section*{Acknowledgements}
We acknowledge the traditional owners of the land on which the telescope stands, the Gamilaraay people, and pay our respects to elders past and present. This work has made use of data from the European Space Agency (ESA) mission Gaia (https://www.cosmos.esa.int/gaia), processed by the Gaia Data Processing and Analysis Consortium (DPAC, https://www.cosmos.esa.int/web/gaia/dpac/consortium). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the Gaia Multilateral Agreement. This publication makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration. M{\v Z} acknowledges funding from the Australian Research Council (grant DP170102233). ADR acknowledges support from the Australian Government Research Training Program, and the Research School of Astronomy \& Astrophysics top up scholarship. This research made use of Astropy, a community-developed core Python package for Astronomy (Astropy Collaboration 2013, 2018).
Parts of this research were supported by the Australian Research Council Centre of Excellence for All Sky Astrophysics in 3 Dimensions (ASTRO 3D), through project number CE170100013.
Parts of this research were conducted by the Australian Research Council Centre of Excellence for Gravitational Wave Discovery (OzGrav), through project number CE170100004. S.-W. Chang acknowledges support from the National Research Foundation of Korea (NRF) grant, No. 2020R1A2C3011091, funded by the Korea government (MSIT).
Software: \texttt{numpy} \citep{harris2020array}, \texttt{scipy} \citep{2020SciPy-NMeth}, \texttt{ipython} \citep{doi:10.1109/MCSE.2007.53}, \texttt{pandas} \citep{mckinney-proc-scipy-2010}, \texttt{matplotlib}
\citep{doi:10.1109/MCSE.2007.55} and \texttt{astropy} \citep{astropy:2018}.
\section*{Data Availability}
This work is based on publicly available databases. Gaia data is available on \href{https://gea.esac.esa.int/archive/}{https://gea.esac.esa.int/archive/} together with the crossmatch with 2MASS and WISE catalogs.
A compilation of known young stars with S-indices from \citealp{2013A&A...551L...8P} is available on \href{http://vizier.u-strasbg.fr/viz-bin/VizieR?-source=J/A+A/551/L8&-to=3}{http://vizier.u-strasbg.fr/viz-bin/VizieR?-source=J/A+A/551/L8\&-to=3}.
All measurements from this work are provided in the appendix with a full table available online.
\bibliographystyle{mnras}
| dcf572c93bb05e6e8686a4d2d160908c8955e924 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Adding nodes and edges through Named-Entity Recognition}
\mysection{Named-Entity Recognition}
\label{sec:entity-extraction}
We enrich our graphs by leveraging Machine Learning (ML) tools for Information Extraction.
{\em Named entities} (NEs) \cite{nadeau2007ner} are words or phrases which, together, designate certain real-world entities. Named entities include common concepts such as people, organizations, and locations.
The {\em Named-Entity Recognition} (NER) task consists of $(i)$~identifying NEs in a natural language text, and ($ii$)~classifying them according to a pre-defined set of NE types.
Let $n_t$ be a text node. We feed $n_t$ as input to a NER module and create, for each entity occurrence $E$ in $n_t$, an \textbf{entity occurrence node} (or entity node, in short) $n_E$; as explained below, we extract \textbf{Person, Organization} and \textbf{Location} entity nodes. Further, we add an edge from $n_t$ to $n_E$ whose label is cl:extract$T$, where $T$ is the type of $E$, and whose confidence is $c$, the {\em confidence of the extraction}. In Figure~\ref{fig:example-graph}, the blue, round-corner rectangles {\sf Centrafrique, Areva, P. Balkany, Levallois-Perret} correspond to the entities recognized from the text document, while the {\sf Marrakech} entity is extracted from the identical-label value node originating from the CSV file.
\vspace{3mm}
\noindent\textbf{Named-Entity Recognition}
We describe here the NER approach we devised for our framework, for English and French.
While we have used Stanford NER~\cite{finkel2005incorporating} in~\cite{Chanial2018}, we have subsequently developed a more performant module based on the Deep Learning Flair NLP framework~\cite{akbik2019flair}.
Flair and similar frameworks rely on {\em embedding} words into vectors in a multi-dimensional space. Traditional word embeddings, e.g., Word2Vec \cite{mikolov2013efficient}, Glove \cite{pennington2014glove} and fastText \cite{bojanowski2017enriching}, are \emph{static}, meaning that a word's representation does not depend on the context where it occurs. New embedding techniques are \emph{dynamic}, in the sense that the word's representation also depends on its context. In particular, the Flair dynamic embeddings~\cite{akbik2018contextual} achieve state-of-the-art NER performance. The latest Flair architecture~\cite{akbik2019flair} facilitates {\em combining} different types of word embeddings, as a better performance might be achieved by combining dynamic with static word embeddings.
For English, we rely on a model\footnote{https://github.com/flairNLP/flair} pre-trained using the English CoNLL-2003\footnote{https://www.clips.uantwerpen.be/conll2003/ner/} news articles dataset.
The model combines Glove embeddings~\cite{pennington2014glove} and so-called {\em forward and backward pooled} Flair embeddings, that evolve across subsequent extractions.
As such a model was missing for French, we trained a Flair one on WikiNER~\cite{nothman2013learning}, a multilingual NER dataset automatically created using the text and structure of Wikipedia.
The dataset contains $132$K sentences, $3.4$M tokens and $216$K named-entities, including $74$K Person, $116$K Location and $25$K Organization entities.
The model uses stacked forward and backward French Flair embeddings with French fastText~\cite{bojanowski2017enriching} embeddings.
\noindent\textbf{Entity node creation.} Similarly to the discussion about value node factorization (Section~\ref{sec:graph-refine-optim}), we have the choice of creating an entity node $n_E$ of type $t$ once per occurrence, or (in hierarchical datasets) {\em per-path}, {\em per-dataset} or {\em per-graph}.
We adopt the per graph method, with the mention that we will create one entity node for each disambiguated entity and one entity node for each non-disambiguated entity.
\subsubsection{Disambiguation quality}
\label{sec:exp-disambiguation}
We now evaluation the quality of the disambiguation module.
As mentioned in Section~\ref{sec:disambig}, our module works for both English and French.
\begin{wrapfigure}{L}{0.55\textwidth}
\begin{tabular}{|l|c|c|c|}\cline{2-4}
\multicolumn{1}{l|}{}
& Precision & Recall & $F1$\\\hline
LOC & 99.00\% & 97.05\% & 98.01\% \\
ORG & 92.38\% & 75.19\% & 82.90\% \\
PER & 75.36\% & 77.94\% & 76.62\% \\\hline
Micro & 90.51\% & 82.94\% & 86.55\% \\\hline
\end{tabular}
\caption{Quality of disambiguation for French.\label{fig:ned}}
\vspace{-4mm}
\end{wrapfigure}
The performance for English has been measured on the CoNLL-YAGO dataset~\cite{hoffart2011robust}, by the developers of Ambiverse. They report a micro-accuracy of $84.61\%$ and a macro-accuracy of $82.67\%$.
To the best of our knowledge, there is no labeled corpus for entity disambiguation in French, thus we evaluate the performance of the module on the FTBNER dataset previously introduced.
FTBNER consists of sentences annotated with named entities. The disambiguation module takes a sentence, the type, and offsets of the entities extracted from it, and returns for each entity either the URI of the matched entity or an empty result if the entity was not found in the KB. In our experiment, $19\%$ of entities have not been disambiguated, more precisely $22\%$ of organizations, $29\%$ of persons, and $2\%$ of locations.
For a fine-grained error analysis, we sampled 150 sentences and we manually verified the disambiguation results (Figure~\ref{fig:ned}). The module performs very well, with excellent results for locations ($F1 = 98.01\%$), followed by good results for organizations ($F1 = 82.90\%$) and for persons ($F1 = 76.62\%$). In addition to these results, we obtain a micro-accuracy of $90.62\%$ and a macro-accuracy of $90.92\%$. The performance is comparable with the one reported by the Ambiverse authors for English. We should note though that the improvement for French might be due to our smaller test set.
\subsubsection{Graph construction}
\label{sec:exp-construction}
We start by studying the impact of \textbf{node factorization} (Section~\ref{sec:graph-refine-optim}) on the number of graph nodes and the graph storage time. For that, we rely on the XML dataset, and {\em disable entity extraction, entity disambiguation, and node matching}. Its (unchanged) number of edges $|E|$ is $1.588.839$.
For each type of loading, we report the number of nodes $|N|$, the time spent storing nodes and edges to disk $T_{DB}$, and the total running time $T$ in Table~\ref{tab:impact}.
\begin{wrapfigure}{L}{0.62\textwidth}
\vspace{-3.5mm}
\scalebox{0.9}{\begin{tabular}{p{3.7cm}rrrrr}
Value node creation policy & $|N|$ & $T_{DB}$ (s) & $T$ (s)\\\hline
Per-instance & $1.588.840$ & $318$ & $319$ \\
Per-path & $1.093.060$ & $271$& $318$ \\
Per-path w/ null code detection & $1.207.951$ & $276$ & $323$ \\
Per-dataset & $1.084.918$ & $265$ & $307$\\
Per-graph & $1.084.918$ & $179$ & $228$ \\
Per-graph w/ null code detection & $1.199.966$ & $179$ & $229$ \\
\end{tabular}}
\vspace{-2.5mm}
\caption{Impact of node factorization.}
\label{tab:impact}
\vspace{-3mm}
\end{wrapfigure}
Moving from per-instance to per-path node creation reduces the number of nodes by a third. However, this introduces some errors, as the dataset features many \textbf{null codes} (Section~\ref{sec:graph-refine-optim}); for instance, with per-instance value creation, there are
$1.113$ nodes labeled {\em ``n\'{e}ant''} (meaning ``non-existent''),
$32.607$ nodes labeled {\em ``Donn\'{e}es non publi\'{e}es''} (unavailable information) etc. Using per-path, the latter are reduced to just $1.154$, which means that in the dataset, {\em ``Donn\'{e}es non publi\'{e}es''} appears $32.607$ times on $1.154$ different paths. However, this factorization, which introduces connections between the XML nodes which are parents of this ``null'' value, may be seen as wrong. When the null codes were input to ConnectionLens, such nodes are no longer unified; the number of nodes increases, and so does the storage time.
In this graph, consisting of one data source, per-dataset and per-graph produce the same number of nodes, overall the smallest; it also increases when null codes are not factorized.
We conclude that {\em per-graph value creation combined with null code detection} is a practical alternative.
Next, we study the impact of \textbf{named entity extraction} (Section~\ref{sec:entity-extraction}) and \textbf{disambiguation} (Section~\ref{sec:disambig}) on the graph construction time.
\begin{wrapfigure}{L}{0.6\textwidth}
\includegraphics[width=0.6\textwidth]{images/updated/constructionbw.png}
\vspace{-10mm}
\caption{Graph construction time (seconds).\label{fig:impact-extract-disambig}}
\vspace{-5mm}
\end{wrapfigure}
For this, we load $100.000$ triples from our Yago subset, with per-graph factorization, natural for the RDF data model where each literal or URI denotes one node in the RDF graph. In Figure~\ref{fig:impact-extract-disambig}, we load the triples using several configurations: without any entity extraction (NONE); with SNER entity extraction, without and then with disambiguation; with FLAIR entity extraction, without and then with disambiguation.
While Flair is slower, its results are qualitatively much better than those obtained with SNER (see Section~\ref{sec:exp-ner} below). To make it faster,
we also implement a \textbf{batch extraction} mechanism whereas $l_B$ labels are input a time to each extraction service, to take advantage of the parallel processing capacity available in current CPUs. In Figure~\ref{fig:impact-extract-disambig}, in the ``FLAIR, BATCH'' column, we used $l_B=128$ which maximized performance in our experiments.
A second optimization leverages the fact that so-called {\em sequence to sequence (seq2seq)} models such as that used in our Flair extractor, when given a batch of inputs, pad the shortest ones to align them to the longest input, and some computation effort is lost on useless padding tokens. Instead, if {\em several batches}, say $n_B$, are received by the extraction service, it can {\em re-group} the inputs so that one call to the seq2seq model is made over inputs of very similar length, thus no effort is wasted. In our experiments, we used $n_B=10$.
Figure~\ref{fig:impact-extract-disambig} shows the time taken by storage, extraction and disambiguation (when applied) and the total graph creation time; note the logarithmic $y$ axis. Storing the graph PostgreSQL dominates the loading time in the absence of extraction, or when using SNER. In contrast, Flair extraction takes more than one order of magnitude longer; batching reduces it by a factor of two. Finally, disambiguation, relying on computationally complex operations, takes longest; it also incurs a modest invocation overhead as it resides on a different server (with the regular server hardware described in Section~\ref{sec:evaluation:setup}), in the same fast network.
\noindent Next, we study \textbf{node matching} (Section~\ref{sec:matching}). For this, we loaded the XML dataset, which comes from individual declaration of interest, filled in by hundreds of users, with numerous spelling variations, small errors and typos. Loaded per-graph mode, with batched Flair extraction, the resulting graph has $1.102.209$ nodes and $1.615.291$ edges.
We test two configurations: comparing {\em all leaf node pairs} to find possible similarity edges, respectively, comparing {\em only entity pairs}.
On the regular server, in both cases, data storage took $3$ minutes, and extraction $13$ minutes. When all comparisons are made, they take $39$ minutes, dominating the total time of $56$ minutes. A total of $28.875$ similar pairs are found to be above our similarity thresholds,
and lead to the same number of cl:sameAs edges stored in the graph, together with the respective similarity values. Only $748$ among these are entity pairs; the others are pairs of similar strings. This confirms the interest of node matching; we hope to reduce its cost further by using techniques such as Locality-Sensitive Hashing (LSH).
\begin{wrapfigure}{L}{0.6\textwidth}
\vspace{-4mm}
\includegraphics[width=0.6\textwidth]{images/updated/scalability.png}
\vspace{-8mm}
\caption{YAGO loading time (minutes) using Flair. \label{fig:yago}}
\vspace{-4mm}
\end{wrapfigure}
To study the \textbf{scalability} of our loading process, we loaded our YAGO subset by slices of 1M triples and measured the running time for these increasing data sizes, using the best extractor (Flair) with the same batch size(s) as above. Figure~\ref{fig:yago} shows the loading time as the data grows, in three different hardware settings: on our regular server, our more powerful GPU server {\em disabling GPU use}, and the same {\em exploiting the GPU}.
Figure~\ref{fig:yago} shows that loading time scales linearly in the data volume on all configurations, and batching helps make the most out of our regular server.
\noindent \textbf{Complete graph} Finally, we loaded all the data sources described in Section~\ref{sec:evaluation:setup} in a single graph, using per-graph node creation, batched Flair extraction, and disambiguation for HTML and XML (not for the RDF Yago subset, whose literals are already associated to URIs).
The graph has $|N|=8.019.651$ nodes (including $677.459$ person entities, $275.316$ location entities, and $61.452$ organization entities), and $|E|=20.642.207$ edges. Many entities occur across data sources, e.g., $330$ person entities occur, each, in at least $10$ sources; the French president E.~Macron occurs in $183$ distinct sources. All these lead to interconnections across sources. On our fastest (GPU) hardware configuration, loading the RDF data took $128$ minutes; the HTML articles another $260$, and the XML document $96$ minutes more. The last two times reflect the relatively high cost of disambiguation (recall Figure~\ref{fig:impact-extract-disambig}). ConnectionLens users can turn it off when it is not needed (e.g., if users feel they know the real-world entities behind entity labels encountered in the graph), or trigger it {\em selectively}, e.g., on organizations but not on people nor locations etc.
\subsubsection{Named-Entity Recognition quality}
\label{sec:exp-ner}
Due to the unavailability of an off-the-shelf, good-quality entity extractor for French text, we decided to train a new model.
To decide the best NLP framework to use, we experimented with the Flair~\cite{akbik2019flair} and SpaCy (\url{https://spacy.io/}) frameworks. Flair allows {\em combining} several embeddings, which can lead to significant quality gains. Following~\cite{akbik2019flair}, after testing different word embedding configurations,
we trained a Flair model using {\em stacked forward and backward French Flair embeddings} with {\em French fastText embeddings} on the WikiNER dataset. We will refer to this model as \textit{Flair-SFTF}.
Below, we describe a \emph{qualitative comparison} of \textit{Flair-SFTF} with the French Flair and SpaCy {\em pre-trained} models.
The French pre-trained Flair model
is trained with the WikiNER dataset, and uses French character embeddings trained on Wikipedia, and French fastText embeddings.
As for SpaCy, two pre-trained models are available for French: a medium (\textit{SpaCy-md}) and a small one (\textit{SpaCy-sm}). They are both trained with the WikiNER dataset and the same parameterization. The difference is that \textit{SpaCy-sm} does not include word vectors, thus, in general, \textit{SpaCy-md} is expected to perform better, since word vectors will most likely impact positively the model performance.
Our evaluation also includes the model previously present in ConnectionLens~\cite{Chanial2018}, trained using \textit{Stanford NER}~\cite{finkel2005incorporating}, with the Quaero Old Press Extended Named Entity corpus~\cite{galibert2012extended}.
We measured the precision, recall, and $F1$-score of each model using the \textit{conlleval} evaluation script, previously used for such tasks\footnote{The script \url{https://www.clips.uantwerpen.be/conll2002/ner/} has been made available in conjunction with the CoNLL (Conference on Natural Language Learning).}. \textit{conlleval} evaluates \emph{exact matches}, i.e., both the text segment of the proposed entity and its type, need to match ``gold standard'' annotation,
to be considered correct. Precision, recall, and $F1$-score (harmonic mean of precision and recall) are computed for each named-entity type.
To get an aggregated, single quality measure, \textit{conlleval} computes the {\em micro-average} precision, recall, and $F1$-score over all recognized entity instances, of all named-entity types.
For evaluation, we used the entire FTBNER dataset~\cite{sagot-etal-2012-annotation}.
We pre-processed it to convert its entities from the seven types they used, to the three we consider, namely, persons, locations and organizations.
After pre-processing, the dataset contains $12$K sentences
and $11$K named-entities ($2$K persons, $3$K locations and $5$K organizations).
\begin{table}[h!]
\begin{tabular}{|p{0.14\columnwidth}|p{0.12\columnwidth}|p{0.14\columnwidth}|p{0.14\columnwidth}|p{0.14\columnwidth}|p{0.14\columnwidth}|}\cline{1-6}
Entities & Flair-SFTF & Flair-pre-trained & SpaCy-md & SpaCy-sm & Stanford NER \\\hline
LOC-P & 59.52\% & 53.26\% & 55.77\% & 54.92\% & 62.17\% \\
LOC-R & 79.36\% & 77.71\% & 78.00\% & 79.41\% & 69.05\% \\
LOC-$F1$ & 68.02\% & 63.20\% & 65.04\% & 64.93\% & 65.43\% \\\hline
ORG-P & 76.56\% & 74.57\% & 72.72\% & 71.92\% & 15.82\% \\
ORG-R & 74.55\% & 75.61\% & 54.85\% & 53.23\% & 5.39\% \\
ORG-$F1$ & 75.54\% & 75.09\% & 62.53\% & 61.18\% & 8.04\% \\\hline
PER-P & 72.29\% & 71.76\% & 53.09\% & 57.32\% & 55.31\% \\
PER-R & 84.94\% & 84.89\% & 74.98\% & 79.19\% & 88.26\% \\
PER-$F1$ & 78.10\% & 77.78\% & 62.16\% & 66.50\% & 68.00\% \\\hline
Micro-P & 69.20\% & 65.55\% & 61.06\% & 61.25\% & 50.12\% \\
Micro-R & 77.94\% & 77.92\% & 65.93\% & 66.32\% & 40.69\% \\
Micro-$F1$ & 73.31\% & 71.20\% & 63.40\% & 63.68\% & 44.91\% \\\hline
\end{tabular}
\caption{Quality of NER from French text.\label{fig:ner-results-table}}
\vspace{-4mm}
\end{table}
\begin{comment}
\begin{table}[h!]
\begin{tabular}{|l|c|c|c|}\cline{2-4}
\multicolumn{1}{l|}{Flair-SFTF}
& Precision & Recall & $F1$ \\\hline
LOC & 59.52\% & 79.36\% & 68.02\% \\
ORG & 76.56\% & 74.55\% & 75.54\% \\
PER & 72.29\% & 84.94\% & 78.10\% \\\hline
Micro & 69.20\% & 77.94\% & \\\hline
\multicolumn{4}{c}{}\\\cline{2-4}
\multicolumn{1}{l|}{Flair-pre-trained}
& Precision & Recall & $F1$ \\\hline
LOC & 53.26\% & 77.71\% & 63.20\% \\
ORG & 74.57\% & 75.61\% & 75.09\% \\
PER & 71.76\% & 84.89\% & 77.78\% \\\hline
Micro & 65.55\% & 77.92\% & 71.20\% \\\hline
\multicolumn{4}{c}{}\\\cline{2-4}
\multicolumn{1}{l|}{SpaCy-md}
& Precision & Recall & $F1$ \\\hline
LOC & 55.77\% & 78.00\% & 65.04\% \\
ORG & 72.72\% & 54.85\% & 62.53\% \\
PER & 53.09\% & 74.98\% & 62.16\% \\\hline
Micro & 61.06\% & 65.93\% & 63.40\% \\\hline
\multicolumn{4}{c}{}\\\cline{2-4}
\multicolumn{1}{l|}{SpaCy-sm}
& Precision & Recall & $F1$ \\\hline
LOC & 54.92\% & 79.41\% & 64.93\% \\
ORG & 71.92\% & 53.23\% & 61.18\% \\
PER & 57.32\% & 79.19\% & 66.50\% \\\hline
Micro & 61.25\% & 66.32\% & 63.68\% \\\hline
\multicolumn{4}{c}{}\\\cline{2-4}
\multicolumn{1}{l|}{Stanford NER}
& Precision & Recall & $F1$ \\\hline
LOC & 62.17\% & 69.05\% & 65.43\% \\
ORG & 15.82\% & 5.39\% & 8.04\% \\
PER & 55.31\% & 88.26\% & 68.00\% \\\hline
Micro & 50.12\% & 40.69\% & 44.91\% \\\hline
\end{tabular}
\caption{Quality of NER from French text.\label{fig:ner-results-table}}
\end{table}
\end{comment}
The evaluation results are shown in Table \ref{fig:ner-results-table}.
All models perform better overall than the \textit{Stanford NER} model previously used in ConnectionLens~\cite{Chanial2018}, which has a micro $F1$-score of about 45\%.
The \textit{SpaCy-sm} model has a slightly better overall performance than \textit{SpaCy-md}, with a small micro $F1$-score difference of $0.28\%$. \textit{SpaCy-md} shows higher $F1$-scores for locations and organizations, but is worse on people, driving down its overall quality.
All Flair models surpass the micro scores of SpaCy models. In particular, for people and organizations, Flair models show more than $11\%$ higher $F1$-scores than SpaCy models. Flair models score better on all named-entity types, except for locations when comparing the SpaCy models, specifically, with the \textit{Flair-pre-trained}.
\textit{Flair-SFTF} has an overall $F1$-score of $73.31\%$ and has better scores than the \textit{Flair-pre-trained} for all metrics and named-entity types, with the exception of the recall of organizations, lower by $1.06\%$.
In conclusion, {\em Flair-SFTF} is the best NER model we evaluated.
\subsection{Construction evaluation}
\label{sec:evaluation:construction}
We present results of our experiments measuring the performance and the quality of the modules involved in graph construction.
We study graph construction performance in
Section~\ref{sec:exp-construction}, the quality of our
information extraction in Section~\ref{sec:exp-ner}, and that of
disambiguation in Section~\ref{sec:exp-disambiguation}.
\input{construction/exp-construction}
\input{construction/exp-ner}
\input{construction/exp-disambiguation}
\subsection{Software, hardware, and datasets}
\label{sec:evaluation:setup}
ConnectionLens is a \textbf{Java application} (44.700 lines) which relies on a relational database to store the constructed graphs and as a back-end used by the query algorithm. It features controllable-size caches to keep in memory as many nodes and edges as possible; this allows adapting to different memory sizes.
It also comprises \textbf{Python} code (6.300 lines) which implements entity extraction (Section~\ref{sec:entity-extraction}) and content extraction from PDF documents to JSON (see~\cite{construction-paper}), tasks for which the most suitable libraries are in Python.
The Flair extractor (Section~\ref{sec:entity-extraction}) and the disambiguator (Section~\ref{sec:disambig}) are Web services which ConnectionLens calls. The former is deployed on the machine where ConnectionLens runs. We deployed the latter on a dedicated Inria server, adapting the original Ambiverse code to our new pipeline introduced in Section~\ref{sec:disambig}; the disambiguator consists of $842$ Java classes.
For our experiments, we used a \textbf{regular server} from 2016, equipped with 2x10-core Intel Xeon E5-2640 (Broadwell) CPUs clocked at 2.40GHz, and 128GB DRAM, which uses PostgreSQL 12.4 to store the graph content in a set of tables.
This is a medium-capacity machine, without special capabilities; recall our requirement \textbf{R3} that our algorithms be feasible on off-the-shelf hardware.
We also used a \textbf{GPU server} from 2020, with a 2x16-core Intel Xeon Gold 5218 (Skylake) CPUs clocked at 2.30GHz, an NVIDIA Tesla V100 GPU and 128GB DRAM.
To show the applicability of our software to standard hardware configurations, we focus on the results that we obtained with our regular server.
However, we also include some results on the more advanced server to show that our platform adapts seamlessly to modern as well as heterogeneous hardware, which includes both CPUs and GPUs.
When needed to separate them, we will refer to each server with its CPU generation name.
\noindent\textbf{Data sources} Most of our evaluation is on {\em real-world} datasets, described below from the smallest to the largest (measuring their size on disk before being input to ConnectionLens).
\noindent\textbf{1.} We crawled the French online newspaper Mediapart and obtained 256 articles for the search keywords ``{\em corona, chloroquine, covid}'' (256 documents), and 886 for ``economie, chomage, crise, budget'' (1142 documents and \textbf{18.4 MB} overall).
\noindent\textbf{2.} An \textbf{XML document}\footnote{https://www.hatvp.fr/livraison/merge/declarations.xml} comprising business interest statements of French public figures, provided by HATVP ({\em Haute Autorit\'{e} pour la Transparence de la Vie Publique}); the file occupies \textbf{35 MB}.
\noindent\textbf{3.} A subset of the \textbf{YAGO 4}~\cite{yago4} RDF knowledge base, comprising entities present in the French Wikipedia and their properties; this takes \textbf{2.49 GB} on disk (17.36 M triples).
For a fine-granularity, controlled study of our query algorithm, we also devised a set of {\em small synthetic graphs}, described
in Section~\ref{sec:evaluation:query}.
\section{Experimental evaluation}
\label{sec:evaluation}
The algorithms described above are implemented in the \textbf{ConnectionLens} prototype, available
\href{https://gitlab.inria.fr/cedar/connectionlens}{online}.
Below, we report the results of an experimental evaluation we carried out to study the performance of its algorithms, as well as quality aspects of the constructed graphs.
Section~\ref{sec:evaluation:setup} describes the software and hardware setup, and our datasets. Section~\ref{sec:evaluation:construction} focuses on graph construction, while Section~\ref{sec:evaluation:query} targets keyword query answering.
\input{eval-setup}
\input{construction/experiments}
\input{query/evaluation}
\section{#1}\vspace{-0.5mm}}
\newcommand\mysubsection[1]{\vspace{-1mm}\subsection{#1}\vspace{-0.5mm}}
\newcommand\va{\vspace{-0mm}}
\newcommand\grow{\textsc{Grow}}
\newcommand\mergecur{\textsc{Merge}}
\newcommand\growac{\textsc{GrowAcross}}
\newcommand{\extVersion}{false}
\newcommand{\printIfExtVersion}[2]
{
\ifthenelse{\equal{\extVersion}{true}}{#1}{}
\ifthenelse{\equal{\extVersion}{false}}{#2}{}
}
\section{Answering keyword queries}
\label{sec:algo}
We now present our approach for computing query answers, on the graph which integrates the heterogeneous datasets.
\subsection{\grow\ and \mergecur}
\label{sec:algorithm}
Our algorithm relies on concepts from prior literature~\cite{dpbf,blinks} while exploring many more trees.
Specifically, it starts from the sets of nodes $N_1,\ldots,N_m$ where the nodes in $N_i$ all match the query keyword $w_i$; each node $n_{i,j}\in N_i$ forms a one-node partial tree.
For instance, in Figure~\ref{fig:example-graph}, one-node trees are built from the nodes with boldface text, labeled ``Africa'', ``Real Estate'' and ``I. Balkany''.
We identify two transformations that can be applied to form increasingly larger trees, working toward query answers:
\noindent\textbf{\grow($t,e$)}, where $t$ is a tree, $e$ is an edge \emph{adjacent to the root} of $t$, and $e$ does not close a loop with a node in $t$, creates a new tree $t'$ having all the edges of $t$ plus $e$; the root of the new tree is the other end of the edge $e$. For instance, starting from the node labeled ``Africa'', a \grow\ can add the edge labeled {\small \texttt{dbo:name}}.
\noindent\textbf{\mergecur($t_1,t_2$)}, where $t_1,t_2$ are trees with the same root, whose other nodes are disjoint, and
matching disjoint sets of keywords, creates a tree $t''$ with the same root and with all edges from $t_1$ and $t_2$.
Intuitively, \grow\ moves away from the keywords, to explore the graph; \mergecur\ fuses two trees into one that matches more keywords than both $t_1$ and $t_2$.
In a {\em single-dataset} context, \grow\ and \mergecur\ have the following properties.
($gm_1$)~\grow\ alone is \textbf{complete} (guaranteed to find all answers) for $k=1,2$ only; for higher $k$, \grow\ and \mergecur\, together are complete.
($gm_2$)~Using \mergecur\ steps helps to find answers faster than using just \grow~\cite{blinks}: partial trees, each starting from a leaf that matches a keyword, are merged into an answer as soon as they have reached the same root.
($gm_3$)~An answer can be found through \textbf{multiple combinations of \grow\ and \mergecur}. For instance, consider a linear graph $n_1\rightarrow n_2 \rightarrow \ldots n_p$ and the two-keyword query $\{a_1, a_p\}$ where $a_i$ matches the label of $n_i$. The answer is obviously the full graph. It can be found: starting from $n_1$ and applying $p-1$ \grow\ steps; starting from $n_p$ and applying $p-1$ \grow\ steps; and in $p-2$ ways of the form \mergecur(\grow(\grow\ldots), \grow(\grow\ldots)), each merging in an intermediary node $n_2,\ldots,n_{p-1}$. These are all the same according to our definition of an answer (Section~\ref{sec:search-problem}), which does not distinguish a root in an answer tree; this follows users' need to know how things are connected, and for which the tree root is irrelevant.
\subsection{Adapting to multi-datasets graphs}
The changes we brought for our harder problem (bidirectional edges and multiple interconnected datasets) are as follows.
\noindent\textbf{1. Bidirectional growth.} We allow \grow\ to traverse an edge both going from the source to the target, and going from the target to the source.
\noindent\textbf{2. Many-dataset answers.} As defined in a single-dataset scenario, \grow\ and \mergecur\ do not allow to connect multiple datasets. To make that possible, we need to enable one, another, or both to also traverse similarity and equivalence edges (shown in solid or dotted red lines in Figure~\ref{fig:example-graph}). We decide to simply extend \grow\ to allow it to traverse not just data edges, but also {\em similarity} edges between nodes of the same or different datasets.
We handle {\em equivalence} edges as follows:
\newcommand{\textsc{Grow2Rep}}{\textsc{Grow2Rep}}
\noindent\textbf{\grow-to-representative}
Let $t$ be a partial tree developed during the search, rooted in a node $n$, such that the representative of $n$ is a node $n_{rep}\neq n$. \textsc{Grow2Rep}\ creates a new tree by adding to $t$ the edge $n\xrightarrow{\equiv}n_{rep}$; this new tree is rooted in $n_{rep}$. If $n$ is part of a group of $p$ equivalent nodes, only {\em one} \textsc{Grow2Rep}\ step is possible from $t$, to the unique representative of $n$; \textsc{Grow2Rep}\ does not apply again on \textsc{Grow2Rep}($t$), because the root of this tree is $n_{rep}$, which is its own representative.
Together, \grow, \textsc{Grow2Rep}\ and \mergecur\ enable finding answers that span multiple data sources, as follows:
\grow\ allows exploring data edges within a dataset, and similarity edges within or across datasets.
\textsc{Grow2Rep}\ goes from a node to its representative when they differ; the representative may be in a different dataset.
\mergecur\ merges trees with a same root: when that root represents $p$ equivalent nodes, this allows connecting partial trees, including \textsc{Grow2Rep}\ results, containing nodes from different datasets. Thus, \mergecur\ can build trees spanning multiple datasets.
One potential performance problem remains. Consider again $p$ equivalent nodes $n_1,\ldots,n_p$; assume without loss of generality that their representative is $n_1$. Assume that during the search, a tree $t_i$ is created rooted in each of these $p$ nodes. \textsc{Grow2Rep}\ applies to all but the first of these trees, creating the trees $t_2', t_3', \ldots, t_p'$, all rooted in $n_1$. Now, \mergecur\ can merge any pair of them, and can then repeatedly apply to merge three, then four such trees etc., as they all have the same root $n_1$. The exponential explosion of \grow\ trees, avoided by introducing \textsc{Grow2Rep}, is still present due to \mergecur.
We solve this problem as follows.
Observe that in an answer, a {\em path of two or more equivalence edges} of the form $n_1\xrightarrow{\equiv}n_2\xrightarrow{\equiv}n_3$ such that {\em a node internal to the path}, e.g. $n_2$, {\em has no other adjacent edge}, even if allowed by our definition, is \emph{redundant}. Intuitively, such a node brings nothing to the answer, since its neighbors, e.g., $n_1$ and $n_3$, could have been connected directly by a single equivalence edge, thanks to the transitivity of equivalence. We call {\em non-redundant} an answer that does not feature any such path, and decide to \textbf{search for non-redundant answers} only.
The following properties hold on non-redundant answers:
\begin{property}
There exists a graph $G$ and a $k$-keyword query $Q$ such that a non-redundant answer contains $k-1$ adjacent equivalence edges
(edges that, together, form a single connected subtree).
\end{property}
\begin{wrapfigure}{L}{0.6\textwidth}
\vspace{-5mm}
\centering
\includegraphics[width=.58\textwidth]{images/rep-trees.png}
\vspace{-2mm}
\caption{Sample graph and answer trees.\label{fig:prop-example}}
\vspace{-5mm}
\end{wrapfigure}
We prove this by exhibiting such an instance. Let $G$ be a graph of $2k$ nodes shown in Figure~\ref{fig:prop-example} (a), such that all the $x_i$ are equivalent,
and consider the $k$-keyword query $Q=\{a_1,\ldots,a_k\}$ (each keyword matches exactly the respective $a_i$ node). An answer needs to traverse all the $k$ edges from $a_i$ to $x_i$, and then connect the nodes $x_i,\ldots,x_k$; we need $k-1$ equivalence edges for this.
Next, we show:
\begin{property}\label{prop:2}
Let $t$ be a non-redundant answer to a query $Q$ of $k$ keywords. A group of adjacent equivalence edges contained in $t$ has at most $k-1$ edges.
\end{property}
We prove this by induction over $k$.
For $k=1$, each answer has $1$ node and $0$ edge (trivial case).
Now, consider this true for $k$ and let us prove it for $k+1$. Assume by contradiction that a non-redundant answer $t_Q$ to a query $Q$ of $k+1$ keywords comprises $k+1$ adjacent equivalence edges. Let $Q'$ be the query having only the first $k$ keywords of $Q$, and $t'$ be a subtree of $t$ that is a non-redundant answer to $Q'$:
\begin{itemize}
\item $t'$ exists, because $t$ connects all $Q$ keywords, thus also the $Q'$ keywords;
\item $t'$ is non-redundant, because all its edges are in the (non-redundant) $t$.
\end{itemize}
By the induction hypothesis, $t'$ has at most $k-1$ adjacent equivalence edges. This means that there are {\em two adjacent equivalent edges} in $t\setminus t'$.
\begin{enumerate}
\item If these edges, together, lead to two distinct leaves of $t$, then $t$ has {\em two} leaves not in $t'$. This is not possible, because by definition of an answer, $t$ has $k+1$ leaves (each matching a keyword) and similarly $t'$ has $k$ leaves.
\item It follows, then, that the two edges lead to a single leaf of $t$, therefore the edges form a redundant path. This contradicts the non-redundancy of $t$, and
concludes our proof.
\end{enumerate}
Property~\ref{prop:2} gives us an important way to control the exponential development of trees due to $p$ equivalent nodes. \grow, \textsc{Grow2Rep}\ and \mergecur, together, can generate trees with up to $k$ (instead of $k-1$) adjacent equivalence edges.
This happens because \textsc{Grow2Rep}\ may ``force'' the search to visit the representative of a set of $k$ equivalent nodes (see Figure~\ref{fig:prop-example}(b), assuming $x_1$ is the representative of all the equivalent $x_i$s, and the query $\{a_2,\ldots,a_k\}$). The resulting answer may be redundant, if the representative has no other adjacent edges in the answer other than equivalence edges. In such cases,
in a \textbf{post-processing step}, we remove from the answer the representative and its equivalence edges, then reconnect the respective equivalent nodes using $k-1$ equivalence edges. This guarantees obtaining a non-redundant tree, such as the one in Figure~\ref{fig:prop-example}(c).
\subsection{The GAM algorithm}
\begin{figure*}[t!]
\fbox{
\begin{minipage}{\textwidth}
Procedure \textbf{process}(tree $t$)
\begin{itemize}[noitemsep,nolistsep]
\item if $t$ is not already in $E$
\item then
\begin{itemize}[noitemsep,nolistsep]
\item add $t$ to $E$
\item if $t$ has matches for all the query keywords
\item then post-process $t$ if needed; output the result as an answer
\item else insert $t$ into $K$
\end{itemize}
\end{itemize}
Algorithm \textbf{GAMSearch}(query $Q=\{w_1,w_2,\ldots, w_k\}$)
\begin{enumerate}[noitemsep,nolistsep]
\item For each $w_i$, $1\leq i \leq k$
\begin{itemize}[noitemsep,nolistsep]
\item For each node $n_i^j$ matching $w_i$, let $t_i^j$ be the 1-node tree consisting of $n_i^j$; process($t_i^j$) \label{item:1node}
\end{itemize}
\item Initial \textsc{merge}$^*$: try to merge every pair of trees from
$E$, and process any resulting answer tree. \item Initialize $U$ (empty so far): \label{item:init-Q}
\begin{enumerate}[noitemsep,nolistsep]
\item Create \grow\ opportunities: Insert into $U$ the pair $(t,e)$, for each $t\in E$ and $e$ a data or similarity edge adjacent
to $t$'s root.
\item Create \textsc{Grow2Rep}\ opportunities: Insert into $U$ the pair $(t, n\rightarrow n_{rep})$ for each $t\in E$ whose root is $n$, such that the representative of $n$ is $n_{rep}\neq n$.
\end{enumerate}
\item While ($U$ is not empty)
\begin{enumerate}[noitemsep]
\item Pop out of $U$ the highest-priority pair $(t,e)$.
\item Apply the corresponding \grow\ or \textsc{Grow2Rep}, resulting in a new tree $t''$; process($t''$).
\item If $t''$ was not already in $E$, agressively \mergecur:
\begin{enumerate}[noitemsep,nolistsep]
\item Let $NT$ be a set of new trees obtained from the \mergecur\ (initially $\emptyset$).
\item Let $\mathbf{p_1}$ be the keyword set of $t''$
\item For each keyword subset $\mathbf{p_2}$ that is a
key within $K$, and such that $\mathbf{p_1}\,\cap\, \mathbf{p_2} =\emptyset$
\begin{enumerate}[noitemsep,nolistsep]
\item For each tree $t^i$ that corresponds to $\mathbf{p_2}$, try to merge $t''$ with $t^i$. Process any possible result; if it is new (not in $E$ previously), add it to $NT$. \label{item:merge-candidates}
\end{enumerate}
\end{enumerate}
\item Re-plenish $U$ (add more entries in it) as in step~\ref{item:init-Q}, based on the trees $\{t''\} \, \cup \, NT$. \label{item:replenish-Q}
\end{enumerate}
\end{enumerate}
\end{minipage}}
\caption{Outline of GAM algorithm\label{fig:algo}}
\end{figure*}
We now have the basic exploration steps we need: \grow, \textsc{Grow2Rep}\ and \mergecur.
In this section, we explain how we use them in our integrated keyword search algorithm.
We decide to apply in sequence: one \grow\ or \textsc{Grow2Rep}\ (see below), leading to a new tree $t$, immediately followed by all the \mergecur\ operations possible on $t$. Thus, we call our algorithm \textbf{Grow and Aggressive Merge} (GAM, in short). We merge aggressively in order to detect as quickly as possible when some of our trees, merged at the root, form an answer.
Given that every node of a currently explored answer tree can be connected with several edges, we need to decide which \grow\ (or \textsc{Grow2Rep}) to apply at a certain point. For that, we use a \textbf{priority queue} $U$ in which we add (tree, edge) entries: for \grow, with the notation above, we add the $(t, e)$ pair, while for \textsc{Grow2Rep}, we add $t$ together with the equivalence edge leading to the representative of $t$'s root. In both cases, when a $(t, e)$ pair is extracted from $U$, we just extend $t$ with the edge $e$ (adjacent to its root), leading to a new tree $t_G$, whose root is the other end of the edge $e$. Then we aggressively merge $t_G$ with all compatible trees explored so far, finally we read from the graph the (data, similarity or equivalence) edges adjacent to $t_G$'s root and add to $U$ more (tree, edge) pairs to be considered further during the search. The algorithm then picks the highest-priority pair in $U$ and reiterates; it stops when $U$ is empty, at a timeout, or when a maximum number of answers are found (whichever comes first).
The last parameter impacting the exploration order is the priority used in $U$: at any point, $U$ gives the highest-priority $(t, e)$ pair, which determines the operations performed next.
\begin{enumerate}
\item Trees matching \emph{many query keywords} are preferable, to go toward complete query answers;
\item At the same number of matched keywords, \emph{smaller trees} are preferable in order not to miss small answers;
\item Finally, among $(t_1,e_1)$, $(t_2,e_2)$ with the same number of nodes and matched keywords, we prefer the pair with the \emph{higher specificity edge}.
\end{enumerate}
\noindent\textbf{Algorithm details} Beyond the priority queue $U$ described above, the algorithm also uses a {\em memory of all the trees explored}, called $E$. It also organizes all the (non-answer) trees into a map $K$ in which they can be accessed by the subset of query keywords that they match. The algorithm is shown in pseudocode in Figure~\ref{fig:algo}, following the notations introduced in the above discussion.
While not shown in Figure~\ref{fig:algo} to avoid clutter, the algorithm {\em only develops minimal trees} (thus, it only finds minimal answers). This is guaranteed:
\begin{itemize}
\item When creating \grow\ and \textsc{Grow2Rep}\ opportunities (steps \ref{item:init-Q} and \ref{item:replenish-Q}): we check not only that the newly added does not close a cycle, but also that the matches present in the new tree satisfy our minimality condition (Section~\ref{sec:search-problem}).
\item Similarly, when selecting potential \mergecur\ candidates (step \ref{item:merge-candidates}).
\end{itemize}
\subsection{Answering keyword queries}
\label{sec:evaluation:query}
This section presents the results that we obtained by using synthetic and real-world datasets.
In each case, we first describe the datasets, and then we present and explain our findings.
We bound the query execution time to 120 seconds, after which the algorithm stops searching for matches.
\subsubsection{Queries on synthetic datasets}
We first study the performance of our algorithm on different types of synthetic datasets.
The first type is the \emph{line graph}, where every node is connected with two others, having one edge for each node, except two nodes which are connected with only one.
We use the line graph to clearly show the performance of \grow~and \mergecur~operations with respect to the graph size.
The second type is the \emph{chain graph}, which is the same as the line, but it has two edges (instead of one) connecting every pair of nodes.
We use the chain graph to show the performance of the algorithm as we double the amount of edges of the line graph and we give more options to \grow~and \mergecur.
The third type is the \emph{star graph}, where we have several line graphs connected through a strongly connected cluster of nodes with a representative.
We use this type to show the performance of \textsc{Grow2Rep}, by placing the query keywords on different line graphs.
The fourth type is a random graph based on the \emph{Barabasi-Albert}
(BA) model~\cite{Barabasi509}, which generates scale-free networks with only a few nodes (hubs) of the graph having much higher degree than the rest.
The graph in this model is created in a two-staged process.
During the first stage, a network of some nodes is created.
Then, during the second stage, new nodes are inserted in the graph and they are connected to nodes created during the first stage.
We set every node created at the second stage to be connected with exactly one node created at the first stage.
In the following, The black line shows the time elapsed until the first answer is found, whereas the grey line shows the overall execution time.
\begin{figure*}[t!]
\centering
\subfloat[Line graph\label{fig:eval/line}]{
\begin{minipage}{0.45\columnwidth}
\includegraphics[width=\columnwidth]{images/updated/line.png}
\end{minipage}
}
\quad
\subfloat[Chain graph\label{fig:eval/chain}] {
\begin{minipage}{0.45\columnwidth}
\includegraphics[width=\columnwidth]{images/updated/chain.png}
\end{minipage}
}
\vspace{-4mm}
\subfloat[Star graph\label{fig:eval/star}] {
\begin{minipage}{0.45\columnwidth}
\includegraphics[width=\columnwidth]{images/updated/star.png}
\end{minipage}
}
\quad
\subfloat[Barabasi-Albert\label{fig:eval/ba}] {
\begin{minipage}{0.45\columnwidth}
\includegraphics[width=\columnwidth]{images/updated/ba.png}
\end{minipage}
}
\vspace{-4mm}
\caption{Query execution time on different synthetic graph types.}
\label{fig:eval/synthetic}
\vspace{-6mm}
\end{figure*}
Figure~\ref{fig:eval/synthetic} includes the results that we obtained by querying the synthetic datasets.
The black line shows the time elapsed until the first answer is found, whereas the grey line shows the overall execution time.
Figure~\ref{fig:eval/line} shows the execution time of our algorithm when executing a query with two keywords on a line graph, as we vary the number of nodes of the graph.
We place the keywords on the two ``ends'' of the graph to show the impact of the distance on the execution time.
The performance of our algorithm is naturally affected by the size of the graph, as it generates $2\cdot N$ answer trees, where $N$ is the number of nodes.
Given that this is a line graph, there is only one answer, which is the whole graph, and, therefore, the time to find the first answer is also the overall execution time.
Figure~\ref{fig:eval/chain} shows the performance of our algorithm on a chain graph.
The execution times for the first answer are almost the same, as the graph size increases slowly.
Instead, the overall execution times increase at a much
higher (exponential) rate; note the logarithmic scale of the $y$ axis.
The reason is that every pair of nodes is connected with two edges, which increases the amount of answers exponentially with the amount of nodes in the graph.
In Figure~\ref{fig:eval/star}, we report the execution time for the star graph.
We place keywords in two different lines connected through the center of the graph, forcing the algorithm to use \textsc{Grow2Rep}, whereas in the previous cases it only had to use \grow~and \mergecur.
The number of branches, depicted on the $x$ axis, corresponds to the number of line graphs connected in the star.
Each line graph has 10 nodes and we place the query keywords at the extremities of two different line graphs.
The number of merges is exponential to the number of branches, that is $\mathcal{O}(2^K)$ where $K$ is the number of branches, since the algorithm will check all possible answers.
This behaviour is clearly shown in both lines of
Figure~\ref{fig:eval/star}, where on the $y$ axis (in logarithmic
scale) we show the times to find the first, and, respectively, all
answers.
Above 12 branches, the timeout of 120 seconds that we have set is hit and, thus, search is terminated, as shown when we search for all answers.
Figure~\ref{fig:eval/ba} depicts the query performance with the Barabasi-Albert model.
We fix the graph size to 2000 nodes and we vary the position of two keywords, by choosing nodes which have a
distance, as given in the $x$ axis; note the logarithmic $y$ axis.
As the graph is randomly generated within the BA model, we note some irregularity in the time to the first solution, which however grows at a moderate pace as the distance between the keyword node grows.
The overall relation between the time to the first solution and the total time confirms that the search space is very large but that most of the exploration
is not needed, since the first solution is found quite fast.
\vspace{-1.5mm}
\subsubsection{Queries on the complete, real-world graph}
\vspace{-1mm}
\begin{table*}[h!]
\begin{adjustbox}{width=\textwidth}
\centering
\begin{tabular}{ |c|r|r|r| }
\hline
Query keyword(s) & Answers & Answer trees & Time to 1st (ms) \\
\hline\hline
{\em a\'eronautique}, {\em Macron} & 2779 & 1152577 & 7225 \\
\hline
Brigitte Macron, Clara Gaymard & 4584 & 545020 & 1412 \\
\hline
{\em Chine}, {\em covid}, {\em France} & 17 & 25974 & 8380 \\
\hline
{\em ch\^omage}, {\em covid} & 108 & 205584 & 4476 \\
\hline
{\em covid}, El Khomri & 16 & 215952 & 52486 \\
\hline
{\em confinement}, Christophe Castaner & 4 & 120367& 3820 \\
\hline
Chine, France, Didier Raoult & 36 & 146261 & 14666 \\
\hline
Ebola, Raoult & 1 & 37751 & 75759 \\
\hline
{\em entreprise}, {\em Raffarin} & 6336 & 1174822 & 6589 \\
\hline
Julien Denormandie, Macron & 464 & 20181 & 2661 \\
\hline
Khalid al-Falih, Kristalina Georgieva & 1 & 1775 & 765 \\
\hline
Kristalina Georgieva, Walter Butler & 1 & 3224 & 353 \\
\hline
Louis Beam, Ku Klux Klan, Trump & 1 & 24172 & 15207 \\
\hline
{\em Macron}, {\em Royal} & 102 & 6413 & 4107 \\
\hline
Marisol Touraine, Jean-Fran\c{c}ois Delfraissy & 3 & 1497 & 1183 \\
\hline
{\em masque}, {\em France} & 35 & 15082 & 7126 \\
\hline
Michael Ryan, Anthony Fauci & 2 & 6653 & 12215 \\
\hline
Pascale Gruny, Jean-Fran\c{c}ois Delfraissy & 2 & 1332 & 91591 \\
\hline
{\em vaccination}, {\em Trump} & 12 & 1188 & 6518 \\
\hline
Yazdan Yazdanpanah, Jean-Fran\c{c}ois Delfraissy & 29 & 5446 & 1687 \\
\hline
\end{tabular}
\end{adjustbox}\vspace{-1.5mm}
\caption{Query results on the complete graph.}
\label{tbl:eval/realworld}
\vspace{-5mm}
\end{table*}
Next, we describe results that we obtained querying a graph obtained by loading all the real-world data sources decribed in Section~\ref{sec:evaluation:setup}.
We report our findings in Table~\ref{tbl:eval/realworld}. The queries feature terms that appear in recent French news; they are related to the economy, the Covid crisis, world events, and/or French politics. In most queries, keywords are exact entity names, with their most common spelling. In other cases, we allowed node labels to {\em approximately} match a keyword (based on PostgreSQL' stemming and string pattern matching); such keywords are shown in italic in the table. Again, we gave a timeout of 2 minutes, and on this large graph, all the GAM searches stopped at a time-out. The table shows that queries return varied number of answers, but many ATs are developed in all cases, and the first is found quite before the timeout.
Finally, an inspection of the results showed that most are obtained from different datasets, confirming the interest of linking datasets in ConnectionLens. These results show the feasibility and interest of GAMSearch on large heterogeneous graphs.
\section{Querying the graph}
\label{sec:pb-statement}
We formalize the keyword search problem over a graph built out of heterogeneous datasets as previously described.
\subsection{Search problem}
\label{sec:search-problem}
We consider a graph $G=(N, E)$ and we denote by $\mathcal{L}$ the set of all the labels of $G$ nodes, plus the empty label $\epsilon$ (see Figure~\ref{fig:example-graph}). Let $W$ be the set of {\em keywords}, obtained by stemming the label set $\mathcal{L}$; a {\em search query} is a set of keywords $Q=\{w_1,...,w_m\}$, where $w_i\in W$.
We define an \textbf{answer tree} (AT, in short) as a set $t$ of $G$ edges which ($i$)~together, form a tree (each node is reachable from any other through exactly one path), ($ii$)~for each $w_i$, contain at least one node whose label matches $w_i$. Here, the edges are \textbf{considered undirected}, that is: $n_1\xrightarrow{a}n_2\xleftarrow{b}n_3\xrightarrow{c}n_4$ is a sample AT, such that for all $w_i \in Q$, there is a node $n_i\in t$ such that $w_i \in \lambda(n_i)$.
We treat the edges of $G$ as undirected when defining the AT in order to allow more query results, on a graph built out of heterogeneous content whose structure is not well-known to users.
Further, we are interested in \textbf{minimal} answer trees, that is:
(i) removing an edge from the tree should make it lack one or more of the query keywords $w_i$;
(ii) if a query keyword $w_i$ matches the label of more than one node in the answer tree, then all these matching nodes must be equivalent.
Condition~(ii) is specific to the graph we consider, built from {\em several data sources connected by equivalence or similarity edges}. In classical graph keyword search problems, each query keyword is matched {\em exactly once} in an answer (otherwise, the tree is considered non-minimal). In contrast, our answer trees \emph{may need to traverse equivalence edges}, and if $w_i$ is matched by one node connected by such an edge, it is also matched by the other. For instance, consider the three-keyword query ``Gyucy Balkany Levallois'' in Figure~\ref{fig:example-graph}: the keyword Balkany is matched by the two nodes labeled ``P. Balkany'' which are part of the answer.
As a counter-example, consider the query ``Balkany Centrafrique'' in Figure~\ref{fig:example-graph}, assuming the keyword Centrafrique is also matched in the label ``Central African Republic''\footnote{This may be the case using a more advanced indexing system that includes some natural language understanding, term dictionaries etc.}.
Consider the tree that connects a ``P. Balkany'' node with ``Centrafrique'', and also traverses the edge between ``Centrafrique'' and ``Central African Republic'': this tree is not minimal, thus it is not an answer. The intuition for rejecting it is that ``Centrafrique'' and ``Central African Republic'' are not necessarily equivalent (we have a similarity, not an equivalence edge), therefore the query keyword ``Centrafrique'' is matched by two potentially different things in this answer, making it hard to interpret.
A direct consequence of minimality is that {\em in an answer, each and every leaf matches a query keyword}.
A graph may hold several minimal answer trees for a given query. We consider available a {\em scoring function} which assigns a higher value to more interesting answer trees (see Section~\ref{sec:score}).
\textbf{Problem Statement.}
Given the graph $G$ built out of the datasets $\mathcal{D}$ and a query $Q$, return the $k$ highest-score minimal answer trees.~\qed
An AT may potentially span over the whole graph, (also) because it can traverse $G$ edges in any direction; this makes the problem challenging.
\subsection{Search space and complexity}
\label{sec:steiner}
\newcommand\pbma{$\diamond$}
\newcommand\pbmb{$\rhd$}
\newcommand\pbmc{$\lhd$}
\newcommand\pbmd{$\circ$}
\newcommand\pbme{$\Box$}
The problem that we study is related to the (Group) Steiner Tree Problem, which we recall below.
Given a graph $G$ with weights (costs) on edges, and a set of $m$ nodes $n_1,\ldots,n_m$, the \emph{Steiner Tree Problem (STP)} \cite{garey2011} consists of finding the smallest-cost tree in $G$ that connects all the nodes together. We could answer our queries by solving one STP problem for each combination of nodes matching the keywords $w_1,\ldots,w_m$. However, there are several obstacles left: (\pbma)~STP is a known NP-hard problem in the size of $G$, denoted $|G|$; (\pbmb)~as we consider that each edge can be taken in the direct or reverse direction, this amounts to ``doubling'' every edge in $G$. Thus, our search space is \textbf{$2^{|G|}$ larger than the one of the STP, or that considered in similar works}, discussed in Section~\ref{sec:related}. This is daunting even for small graphs of a few hundred edges; (\pbmc)~we need the $k$ smallest-cost trees, not just one; (\pbmd)~each keyword may match several nodes, not just one.
The closely related {\em Group STP} (GSTP, in short) \cite{garey2011} is: given $m$ {\em sets of nodes} from $G$, find the minimum-cost subtree connecting one node from each of these sets. GSTP does not raise the problem (\pbmd), but still has all the others.
In conclusion, the complexity of the problem we consider is extremely high. Therefore, solving it fully is unfeasible for large and/or high-connectivity graphs. Instead, our approach is:
($i$)~{\em Attempt to find all answers from the smallest} (fewest edges) {\em to the largest}.
Enumerating small trees first is both a practical decision (we use them to build larger ones) and fits the intuition that we should not miss small answers that a human could have found manually.
However, as we will explain, we still ``opportunistically'' build some trees before exhausting the enumeration of smaller ones, whenever this is likely to lead faster to answers. The strategy for choosing to move towards bigger instead of smaller tress leaves rooms for optimizations on the search order.
($ii$)~{\em Stop at a given time-out or when $m$ answers have been found}, for some $m\geq k$;
($iii$)~{\em Return the $k$ top-scoring answers} found.
\section{Scoring answer trees}
\label{sec:score}
We now discuss how to evaluate the quality of an
answer. Section~\ref{sec:generic-score} introduces the general notion
of score on which we base our approach. Section~\ref{sec:specif}
describes the metric that we attach to edges in order to
instantiate this score, and Section~\ref{sec:concrete-score}
details the actual score function we used.
\subsection{Generic score function}
\label{sec:generic-score}
We have configured our problem setting to allow {\em any scoring function}, which enables the use of different scoring schemes fitting the requirements of different users.
As a consequence, this approach allows us to study the interaction of the scoring function with different properties of the graph.
Given an answer tree $t$ to a query $Q$, we consider a score function consisting of (at least) the following two components.
First, the {\em matching score} $ms(t)$, which reflects the quality of the answer tree, that is, how well its leaves match the query terms.
Second, the {\em connection score} $cs(t)$, which reflects the quality of
the tree connecting the edges.
Any formula can be used here, considering the number of edges, the confidence or any other property attached to edges, or a query-independent property of the nodes, such as their PageRank or betweenness centrality score etc.
The score of $t$ for $Q$, denoted $s(t)$, is computed as a
combination of the two independent components $ms(t)$ and $cs(t)$.
Popular combinations functions (a weighted sum, or product etc.) are monotonous in both components, however, our framework does not require monotonicity. Finally, both $ms(t)$ and $cs(t)$ can be tuned based on a given user's preferences, to personalize the score, or make them evolve in time through user feedback etc.
\mysubsection{Edge specificity}\label{sec:specif}
We now describe a metric on edges, which we used (through the
connection score $cs(t)$) to favor edges that are ``rare'' for both nodes they
connect. This metric was inspired by our experiments with real-world
data sources, and it helped return interesting answer trees in our
experience.
For a given node $n$ and label $l$, let $\edgesin{n}{l}$ be the number of $l$-labeled edges entering $n$, and $\edgesout{n}{l}$ the number of $l$-labeled edges exiting $n$.
\noindent The \textbf{specificity} of an edge $e=n_1\xrightarrow{l}n_2$ is defined as:
\begin{center}
\va\va
$s(e)=2/(\edgesout{n_1}{l} + \edgesin{n_2}{l})$.
\va\va
\end{center}
Specificity is $1.0$ for edges that are ``unique'' for both their source and their target, and decreases when the edge does not ``stand out'' among the edges of these two nodes.
For instance, the city council of Levallois-Perret comprises only one mayor (and one individual cannot be mayor of two cities in France).
Thus, the edge from the city council to P.~Balkany has a specificity of $2/(1.0+1.0)=1.0$. In contrast, there are 54 countries in Africa (we show only two), and each country is in exactly one continent; thus,
the specificity of the {\sf dbo:partOf } edges in the DBPedia fragment, going from the node named Morocco (or the one named Central African Republic) to the node named Africa is $2/(1+54)\simeq .036$.
\begin{comment}
\textbf{Specificity computation.}
When registering the first dataset $D_1$, computing the specificity of its edges is trivial.
However, when registering subsequent datasets $D_2,D_3$ etc., if some
node, say $n_2\in D_2$ is found to be equivalent to a node $n_1\in
D_1$,
{\em all} the $D_1$ edges adjacent to $n_1$ {\em and} the $D_2$ edges adjacent to $n_2$ should be reflected in the specificity of {\em each} of these edges. Thus, in particular, the specificity of $D_1$ edges needs to be {\em recomputed} when a node in a source added after $D_1$ is equivalent to one of its nodes.
A {\em na\"ive approach} would be: when the edges of $D_2$ are traversed (when we add this dataset to the graph), re-traverse the edges of $n_1$ in $D_1$ in order to (re)compute their specificity. However, that would be quite inefficient.
Instead, below, we describe an {\em efficient incremental algorithm} to compute specificity. We introduce two notations.
For any edge $e$, we denote $N^e_{\rightarrow \bullet}$, respectively $N^e_{\circ \rightarrow}$, the two numbers out of which the specificity of $e$ has been {\em most recently} computed\footnote{This can be either during the first specificity computation of $e$, or during a recomputation, as discussed below.}. Specifically, $N^e_{\rightarrow \bullet}$ counts $l$-labeled edges incoming to the target of $e$, while $N^e_{\circ \rightarrow}$ counts $l$-labeled edges outgoing the source of $e$. In Figure~\ref{fig:edge-recomp}, if $e$ is the edge $x \xrightarrow{l}n_1$, then $N^e_{\rightarrow \bullet}=3$ (blue edges) and $N^e_{\circ \rightarrow}=1$, thus $s(e)=2/4=.5$.
\begin{figure}[t!]
\begin{center}
\input{images/edge-specif-tikz}
\caption{Illustration for specificity (re)computation. The specificity of the edge $x \xrightarrow{l}n_1$, $s(e)$ is initially computed out of the blue edges; when $n_2$ joins the equivalence set $es_1$, it is recomputed to also reflect the violet edges. \label{fig:edge-recomp}}
\end{center}
\vspace{-7mm}
\end{figure}
Let $n_1\in D_1$ be a node, $es_1$ be the set of all nodes equivalent to
$n_1$, and $n_2\in D_2$ be a node in a dataset we currently register,
and which has just been found to be equivalent to $n_1$, also.
Further, let $l$ be a label of an edge incoming or outgoing (any) node from $es_1$, and/or $n_2$.
We denote by $\edgesin{es_1}{l}$ the sum $\sum_{n\in es_1}(\edgesin{n}{l})$ and similarly by $\edgesout{es_1}{l}$ the sum $\sum_{n\in es_1}(\edgesout{n}{l})$; they are the numbers of $l$-labeled outgoing (resp., incoming) $l$-labeled edges of any node in $es_1$.
When $n_2$ joins the equivalence set $es_1$ of $n_1$ (see Figure~\ref{fig:edge-recomp}):
\va\va
\begin{enumerate}
\item \label{item:incremental-1}
If $\edgesin{es_1}{l}\neq 0$ and $\edgesin{n_2}{l}\neq 0$, the specificity of every $l$-labeled edge $e$ {\em incoming} either a node in $es_1$ or the node $n_2$ must be recomputed.\\
Let $e$ be such an {\em incoming} edge labeled $l$. When $n_2$ is added to the set $es_1$, the specificity of $e$ becomes
$2/((N^e_{\rightarrow \bullet} + \edgesin{n_2}{l}) + N^e_{\circ \rightarrow})$,
to reflect that $n_2$ brings more incoming $l$-labeled edges. This amounts to $2/(3+2+1)=.33$ in Figure~\ref{fig:edge-recomp}: the violet edges have joined the blue ones. Following this adjustment, the numbers out of which $e$'s specificity has been most recently computed are modified as follows:
$N^e_{\rightarrow \bullet}$ becomes $N^e_{\rightarrow \bullet} + \edgesin{n_2}{l}$, thus $3+2=5$ in Figure~\ref{fig:edge-recomp};
$N^e_{\circ \rightarrow}$ remains unchanged.
\va
\item \label{item:incremental-2}
If $\edgesin{es_1}{l}=0$ and $\edgesin{n_2}{l}\neq 0$, the specificity of every $l$-labeled edge $e$ {\em incoming} $n_2$
does not change when $n_2$ joins the equivalence set $es_1$.
\item \label{item:incremental-3}
If $\edgesin{es_1}{l}\neq 0$ and $\edgesin{n_2}{l}=0$, the newly added node $n_2$ does not change the edges adjacent to the nodes of $es_1$, nor their specificity values.
\printIfExtVersion{
\noindent\textbf{4. }
Similarly to \textbf{1.}
above, if $\edgesout{es_1}{l}\neq 0$ and $\edgesout{n_2}{l}\neq 0$, the specificity of every $l$-labeled edge $e$ {\em outgoing} either a node in $es_1$ or the node $n_2$ is recomputed, and it becomes
$2/(N^e_{\rightarrow \bullet} + (N^e_{\circ \rightarrow}+\edgesout{n_2}{l}))$.
The numbers out of which $e$'s specificity has been most recently computed become: $N^e_{\rightarrow \bullet}$ and $(N^e_{\circ \rightarrow}+\edgesout{n_2}{l})$.
\noindent\textbf{5. }
Finally, if $\edgesout{es_1}{l}=0$ and $\edgesout{n_2}{l}\neq 0$, the reasoning is similar to the one in \textbf{2. }
above, while if $\edgesout{es_1}{l}\neq 0$ and $\edgesout{n_2}{l}=0$, it is similar to \textbf{3.};
in both cases, we consider outgoing (instead of incoming) edges and edge count numbers.
}{
\end{enumerate}
\va
The last two cases, when $\edgesout{es_1}{l}\neq 0$ and $\edgesout{n_2}{l}\neq 0$, respectively, $\edgesout{es_1}{l}=0$ and $\edgesout{n_2}{l}\neq 0$, are handled in a similar manner.
}
The above method only needs, for a given node $n_2$ newly added to the graph, and label $l$, the number of edges adjacent to $n_2$ in its dataset, and the \emph{number} of $l$ edges adjacent to a node equivalent to $n_2$. Unlike the na\"ive specificity computation method, it does not need to actually {\em traverse} these edges previously registered edges, making it more efficient.
Concretely, for each edge $e\in E$, we store three attributes: $N^e_{\rightarrow \bullet}$, $N^e_{\circ \rightarrow}$ and $s$, the last-computed specificity, and we update $N^e_{\rightarrow \bullet}$, $N^e_{\circ \rightarrow}$ as explained above.
\end{comment}
\begin{comment}
\textbf{Where is specificity information stored}
We store a {\it specificity (\underline{representative, edgelabel},
nIn, nOut)} table. As the specificities are computed jointly for all
the edges adjacent to a set of equivalent nodes, this information is
stored per representative of an equivalence class. {\it edgeLabel} is
the label of some edge(s) adjacent to this equivalence class; {\it
nIn} and {\it nOut} are the numbers out of which specificity has
been most recently computed. As an extra optimization, we only store a
tuple in this table if $nIn>1$ or $nOut>1$. Thus, for tree or
tree-like datasets, there are much fewer tuples in the specificity
table than there are in the edges table.
\end{comment}
\subsection{Concrete score function}
\label{sec:concrete-score}
We have implemented the following prototype scoring function in our system.
For an answer $t$ to the query $Q$, we compute the matching score $ms(t)$ as the \emph{average}, over all query keywords $w_i$, of the similarity between the $t$ node matching $w_i$ and the keyword $w_i$ itself; we used the edit distance.
We compute the connection score $cs(t)$ based on edge confidence, on one hand, and edge specificity on the other. We {\em multiply} the confidence values, since we consider that uncertainty (confidence $<1$) multiplies; and we also {\em multiply} the specificities of all edges in $t$, to discourage many low-specificity edges. Specifically, our score is computed as:
\begin{center}
$score(t,Q)=\alpha \cdot ms(t,Q) + \beta \cdot \prod_{e\in E} c(e) + (1 - \alpha - \beta) \cdot \prod_{e\in E} s(e)$
\end{center}
\noindent where $\alpha$, $\beta$ are parameters of the system such that $0\leq \alpha, \beta <1$ and $\alpha +\beta\leq 1$.
\begin{comment}
\subsection{Orthogonality between the score and the algorithm}
Before we describe the search algorithm, we make a few more remarks on the connection between the score function and the search algorithm.
We start by considering the classical Steiner Tree and Group Steiner Tree Problems (Section~\ref{sec:steiner}). These assume that the tree cost is \textbf{monotonous}, that is: for any query $Q$ and all trees $T, \hat{T}$ where $T$ is a subtree of $\hat{T}$ it follows that the cost of $T$ is higher (in our terminology, its score is lower) than the cost of $\hat{T}$. This is naturally satisfied if the cost is the addition of edge weights.
In contrast,
the score, in its general form (Section~\ref{sec:generic-score}), and in particular our concrete one (Section~\ref{sec:concrete-score}), is \textbf{not monotonous}, as illustrated in Figure~\ref{fig:nonmono}, where on each edge, $c$ is the confidence and $s$ is the specificity.
Denoting $T$ the four-edge tree rooted in $n_1$, the connection score $cs(T)=\beta + (1-\alpha-\beta)\cdot (.5)^4$, while $cs(T')=\beta \cdot .5 + (1-\alpha-\beta) (.5)^4 \cdot .25$. If we assume $\alpha=\beta=\frac{1}{3}$,
then $cs(T)=\frac{1}{3}(1+(.5)^4)\simeq .35$ while $cs(T')=\frac{1}{3}\cdot (.5 + (.5)^5)\simeq .17$, which is clearly smaller.
Assuming $T'$ has the same matching score as $T$, the global score of $T'$ is smaller than that of $T$, contradicting the monotonicity assumption.
\begin{figure*}[t!]
\begin{center}
\tikzstyle{node} = [text centered, fill=white
\tikzstyle{arrow} = [thick,->,>=stealth]
\begin{tikzpicture}[node distance=34mm and 34mm
\node (zero) {~};
\node (n1) [node, right of=zero, xshift=-14mm] {$n_1$};
\node (n2) [node, left of=n1, yshift=4mm] {$n_2$};
\node (n3) [node, left of=n1, yshift=-4mm] {$n_3$};
\node (n4) [node, right of=n1, yshift=4mm] {$n_4$};
\node (n5) [node, right of=n1, yshift=-4mm] {$n_5$};
\draw [arrow] (n1) -> (n2) node [midway, fill=white] {$s=.5,c=1$};
\draw [arrow] (n1) -> (n3) node [midway, fill=white] {$s=.5,c=1$};
\draw [arrow] (n1) -> (n4) node [midway, fill=white] {$s=.5,c=1$};
\draw [arrow] (n1) -> (n5) node [midway, fill=white] {$s=.5,c=1$};
\node (n6) [node, right of=n4] {$n_6$};
\node (n7) [node, right of=n5] {$n_7$};
\draw [arrow, dashed] (n4) -> (n6) node [midway, fill=white] {$s=.25,c=.5$};
\draw [arrow, dashed] (n5) -> (n7) node [midway, fill=white] {$s=1,c=1$};
\end{tikzpicture}
\caption{Example (non-monotonicity of the tree score). $T$ is the four-edges tree rooted in $n_1$.}
\label{fig:nonmono}
\end{center}
\end{figure*}
Another property sometimes assumed by score functions is the called \textbf{optimal substructure}, that is: the best solution for a problem of size $p$ is part of the best solution for a problem of size $p+1$ that is an extension of $p$, for some problem size $p$. When this holds, the problem can be efficiently solved in a dynamic programming fashion.
However, STP does not enjoy this property: the smallest-cost tree connecting two nodes $n_1,n_2$ is not necessarily part of the smallest-cost tree that connects $n_1,n_2,n_3$ (and the same holds for GSTP). Some existing algorithms also assume a variant of the optimal substructure property (see Section~\ref{sec:related}). In contrast, our score function (both in its general and its concrete form) does not ensure such favorable properties. This is why the search algorithm we describe next has to find as many answers as possible, as quickly as possible.
\end{comment}
| 8ed94ce204db18e4b18758131538bd6e21d80fbf | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Acknowledgement}
We would like to thank the anonymous reviewers for their helpful comments. The work was supported by NSF DBI-1565137, DGE-1829071, NIH R35-HL135772, NSF III-1705169, NSF CAREER Award 1741634, NSF \#1937599, DARPA HR00112090027, Okawa Foundation Grant, and Amazon Research Award.
{ \small
\section{Related Work}
\subsection{Clinical Temporal Relation Extraction}
\noindent \textbf{Corpora. }
Different from the datasets in the news domain~\cite{pustejovsky2003timebank,AQUAINT}, the corpora in the clinical domain require rich domain knowledge for annotating the temporal relations.
I2b2-2012~\cite{sun2013evaluating} and Clinical TempEval~\cite{bethard-etal-2015-semeval,bethard-etal-2016-semeval,bethard-etal-2017-semeval} are some great efforts of building clinical datasets with extensive annotations including labels of clinical events and temporal relations, the second of which was not tested in our paper due to lack of access to the data.
\noindent \textbf{Models.}
Some early efforts to solve the clinical relation extraction problem leverage conventional machine learning methods~\cite{llorens2010tipsem,sun2013evaluating,xu2013end,tang2013hybrid,lee-etal-2016-uthealth,chikka-2016-cde} such as SVMs, MaxEnt and CRFs, and neural network based methods~\cite{lin2017representations,lin2018self,dligach2017neural,tourille2017neural,lin-etal-2019-bert,guan2020robustly,lin-etal-2020-bert,galvan2020empirical}. They either require expensive feature engineering or fail to consider the dependencies among temporal relations within a document. \cite{leeuwenberg-moens-2017-structured,han-etal-2019-deep,han-etal-2019-joint,ning-etal-2017-structured} formulate the problem as a structured prediction problem to model the dependencies but can not globally predict temporal relations. Instead, our method can infer the temporal relations at document level.
\subsection{Probabilistic Soft Logic}
In recent years, PSL rules have been applied to various machine learning topics such as Fairness~\cite{farnadi2019declarative}, Model Interpretability~\cite{hu-etal-2016-harnessing}, Probabilistic Reasoning~\cite{augustine2019tractable,dellert2020exploring}, Knowledge Graph Construction~\cite{pujara2013knowledge,chen2019embedding} and Sentiment Analysis~\cite{deng-wiebe-2015-joint,gridach2020framework}. We are the first to model the temporal dependencies with PSL.
\section{Preliminaries}\label{sec:psl}
\subsection{Problem Statement}
Document $D$ contains sequences $[s_1,s_2, ..., s_M]$ and named entities $x_i \in \mathcal{E} \bigcup \mathcal{T}, 1 \leq i \leq N$, where $M,N$ are the total number of sequences and entities in $D$. $\mathcal{E}$ and $\mathcal{T}$ represent the set of events and time expressions, respectively. There is a potential temporal relation between any pair of annotated named entities $(x_j, x_k)$, where $1\leq j,k \leq N$. Formally, the task is modeled as a classification problem with a set of temporal relation types $\mathcal{Y}$.
Given a sequence $s_i$ together with two named entities $x_{i,1},x_{i,2}$ included, we predict the temporal relation $y_i\in \mathcal{Y}$ from $x_{i,1}$ to $x_{i,2}$.
In practice, we create a triplet with three pairs of entities to be one training instance $\mathcal{I}$, to enable the PSL rule grounding, as explained in the following section
\subsection{Probabilistic Soft Logic and Temporal Dependencies in Clinical Narratives}\label{sec:psld}
Here, we introduce some concepts and notations for the language PSL and illustrate how PSL is applicable to define templates for temporal dependencies and to help jointly learn a relation classifier.
\begin{mydef}
A \textbf{predicate} $\tilde{p}$ is a relation defined by a unique identifier and an \textbf{atom} $\tilde{l}$ is a predicate combined with a sequence of terms of length equal to the predicate’s argument number. Atoms in PSL take on continuous values in the unit interval $[0, 1]$.
\end{mydef}
\begin{myex}
$\underline{\text{Before}}/2$ indicates a predicate taking two arguments, and the atom $\underline{\text{Before}}(A,B)$ represents whether $A$ happens before $B$.
\end{myex}
\begin{mydef}\label{def:rule}
A \textbf{PSL rule} $\tilde{r}$ is a disjunctive clause of atoms or negative atoms:
\begin{equation}
\eta_r: T_1 \land T_2 \land ... \land T_{m} \rightarrow H_1 \lor H_2 \lor ... \lor H_{n},
\end{equation}
where $T_1,T_2,...,T_m,H_1,H_2,...,H_n$ are atoms or negative atoms.
\end{mydef}
We name $T_1,T_2,...,T_m$ as $r_{body}$ and $H_1,H_2,...,H_n$ as $r_{head}$. $\eta_r\in [0,1]$ is the weight of the rule $r$, denoting the prior confidence of this rule. To the opposite, an unweighted PSL rule is to describe a constraint that is always true. The unweighted logical clauses in Table~\ref{tab:psl} describe the common temporal transitivity and symmetry dependencies we summarize from the clinical narratives.
\begin{table}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|c|c|}
\hline
Abbrev. & PSL rules \\
\hline
\hline
\multicolumn{2}{|l|}{Transitivity Dependencies} \\
\hline
BBB & Before$(A,B)$ $\land$ Before$(B,C)$ $\rightarrow$ Before$(A,C)$ \\
BOB & Before$(A,B)$ $\land$ Overlap$(B,C)$ $\rightarrow$ Before$(A,C)$ \\
OBB & Overlap$(A,B)$ $\land$ Before$(B,C)$ $\rightarrow$ Before$(A,C)$ \\
OOO & Overlap$(A,B)$ $\land$ Overlap$(B,C)$ $\rightarrow$ Overlap$(A,C)$ \\
AAA & After$(A,B)$ $\land$ After$(B,C)$ $\rightarrow$ After$(A,C)$ \\
AOA & After$(A,B)$ $\land$ Overlap$(B,C)$ $\rightarrow$ After$(A,C)$ \\
OAA & Overlap$(A,B)$ $\land$ After$(B,C)$ $\rightarrow$ After$(A,C)$ \\
\hline
\hline
\multicolumn{2}{|l|}{Symmetry Dependencies} \\
\hline
BA & Before$(A,B)$ $\rightarrow$ After$(B,A)$ \\
AB & After$(A,B)$ $\rightarrow$ Before$(B,A)$\\
OO & Overlap$(A,B)$ $\rightarrow$ Overlap$(B,A)$ \\
\hline
\end{tabular}}
\caption{Temporal transitivity and symmetry PSL rules $\mathcal{R}$. $A,B,C$ are three terms representing either events or time expressions.}
\label{tab:psl}
\end{table}
\begin{mydef}
The \textbf{ground atom} $l$ and \textbf{ground rule} $r$ are particular variable instantiation of some atom $\tilde{l}$ and rule $\tilde{r}$, respectively.
\end{mydef}
\begin{myex}\label{ex:rule}
That \underline{Overlap} (\texttt{e}, \texttt{f}) $\land$ \underline{Overlap} (\texttt{f}, \texttt{g}) $\rightarrow$ \underline{Overlap} (\texttt{e}, \texttt{g}) from Figure~\ref{fig:case_report.pdf} is a ground rule composed of three ground atoms, denoted as $l_1,l_2$, and $l_3$, respectively. It is grounded from the OOO rule, as shown in Table~\ref{tab:psl}.
\end{myex}
\begin{mydef}
The interpretation $I(l)$ denotes the soft truth value of an atom $l$.
\end{mydef}
\begin{mydef}
Łukasiewicz t-norm~\cite{klir1995fuzzy} is used to define the basic
logical operations in PSL, including logical conjunction ($\land$), disjunction ($\lor$), and negation ($\neg$):
\begin{align}
& I(l_1 \land l_2) = \max\{I(l_1) + I(l_2) - 1, 0 \}\label{eq:luk} \\
& I(l_1 \lor l_2) = \min\{I(l_1) + I(l_2), 1 \} \\
& I(\neg l_1) = 1 - I(l_1)
\end{align}
\end{mydef}
The PSL rule in Definition~\ref{def:rule} can also be represented as:
\begin{align*}
I(r_{body}\rightarrow r_{head}) = I(\neg r_{body} \lor r_{head}),
\end{align*}
so we can induce the distance to satisfaction for rule $r$.
\begin{mydef}
The \textbf{distance to satisfaction} $d_r(I)$ of rule $r$ under an interpretation $I$ is defined as:
\begin{align}\label{eq:dis}
d_r(I) = \max\{0, I(r_{body})-I(r_{head})\}
\end{align}
\end{mydef}
PSL program determines a rule $r$ as satisfied when the truth value of $I(r_{head})-I(r_{body})\geq 0$.
\begin{myex}
Given that $I(l_1) = 0.7, I(l_2) = 0.8, $ and $I(l_3) = 0.3$, we can compute the distance according to Equation~\eqref{eq:luk}-\eqref{eq:dis}:
\begin{align*}
& d_r = \max\{0, I(l_1 \land l_2) - I(l_3)\} \\
& = \max\{0, 0.7 + 0.8 -1 - I(l_3)\} \\
& = \max\{0, 0.5 -0.3\} \\
& = 0.2
\end{align*}
\end{myex}
This equation indicates that the ground rule in Example~\ref{ex:rule} is completely satisfied when $I(l_3)$ is above $0.5$. Otherwise, a penalty factor will be raised ($0.2$ in this case). When $I(l_3)$ is under $0.5$, the smaller $I(l_3)$ is, the larger penalty we have.
In short, we compute the distance to satisfaction for each ground rule as a loss regularization term to jointly learn a relation classification model.
We finally use the smallest one as the penalty because we only need one of the rules to be satisfied.
\input{figures/framework}
\section{Conclusion}
In this paper, we propose \texttt{CTRL-PG}\xspace that leverages the PSL rules to model the temporal dependencies as a regularization term to jointly learn a relation classification model. Extensive experiments show the efficacy of the PSL regularization and global temporal inference with time graphs.
\section{Introduction}
Clinical case reports (CCRs) are written descriptions of the unique aspects of a particular clinical case~\citep{Caban-Martinez2012, caufield2018reference}.
They are intended to serve as educational aids to science and medicine, as they play an essential role in sharing clinical experiences about atypical disease phenotypes and new therapies~\citep{caufield2018reference}.
There is a perennial need to automatically and precisely curate the clinical case reports into structured knowledge, i.e. extract important clinical named entities and relationships from the narratives~\cite{aronson2010overview,savova2010mayo,soysal2018clamp,caufield2019comprehensive,alfattni2020extraction}.
This would greatly enable both doctors and patients to retrieve related case reports for reference and provide a certain degree of technical support for resolving public health crises like the recent COVID-19 pandemic.
Clinical reports describe chronicle events, elucidating a chain of clinical observations and reasoning~\cite{sun2013evaluating,chen2016orderrex}.
Extracting temporal relations between clinical events is essential for the case report retrieval over the patient chronologies. Besides, medical question answering systems require the precise ordering of clinical events in a time series within each document.
In this paper, we tackle the temporal relation extraction problem in clinical case reports.
\input{figures/case-report}
Figure~\ref{fig:case_report.pdf} illustrates a paragraph from a typical CCR document with three common types of temporal relations, ``Before'', ``After'', and ``Overlap''. \textit{Glucocortocoids} was described as the medicine history of this patient, which happened before \textit{confirmed with COVID-19} and \textit{positive of antibody}.
An ``Overlap'' temporal relation exists between \textit{nasal congestion} and \textit{a mild cough}.
We consider the aforementioned clinical concepts as events, while regarding \textit{a day later} as a time expression. A temporal relation may exist between event and event ($\texttt{E-E}$), event and time expression ($\texttt{E-T}$) or time expression and time expression ($\texttt{T-T}$).
There is a consensus within the clinical community regarding the difficulty of temporal information extraction, due to the high demand for domain knowledge and high complexity of clinical language representations
~\cite{galvan2018investigating}.
\citet{meng-rumshisky-2018-context,lee-etal-2016-uthealth} apply machine learning models with lexical, syntactic features, or pre-trained word representations to tackle the problem but neglect the strong dependencies between narrative containment and temporal order, thus predicting inconsistent output labels and garbled time-lines~\cite{leeuwenberg-moens-2017-structured}.
The dependency is the key enabler of classifying the temporal relations.
For instance in Figure~\ref{fig:case_report.pdf}, given that \texttt{b} happened before \texttt{d}, \texttt{e} happened after \texttt{d} and \texttt{e} happened simultaneously with \texttt{f} , we can infer according to the temporal transitivity rule that \texttt{b} was before \texttt{f}.
Some recent studies~\cite{leeuwenberg-moens-2017-structured,ning-etal-2017-structured,han2020knowledge}
convert the task to a structured prediction problem
and solve it with Maximum a posteriori Inference.
Integer Linear Programming (ILP) with hard constraints is deployed for optimization, which however needs an off-the-shelf solver to tackle the NP-hard optimization problem and can only approximate the optimum via relaxation. Besides, globally inferring the relations at the document level would also be intractable for them due to the high complexity and low scalability~\cite{bach2017hinge}.
Recently, some researchers~\cite{deng-wiebe-2015-joint,chen2019embedding,hu-etal-2016-harnessing} have explored
Probabilistic Soft Logic (PSL)~\cite{bach2017hinge} to tackle the structured prediction problem.
Inspired by them, we propose to leverage the PSL rules to model
relation extraction more flexibly and efficiently.
In specific, we summarize common transitivity and symmetry patterns of temporal relations as PSL rules and penalize the training instances that violate any of those rules. Different from ILP solutions, no off-the-shelf solver is required and the algorithm conducts the training process with linear time complexity. Besides, logical propositions in PSL can be interpreted not just as $true$ or $false$, but as continuously valued in the $[0, 1]$ interval. We also propose a simple but effective time-anchored global temporal inference algorithm to classify the relations at the document level. With such a mechanism, we can easily verify some relations, such as the relation between \texttt{b} and \texttt{f}, with long-term dependencies which are intractable with existing approaches.
As a summary, our main contributions are list as follows:
\begin{itemize}
\item To the best of our knowledge, this is the first work to formulate the probabilistic soft logic rules of temporal dependencies as a regularization term to jointly learn a relation classification model,
\item We show the efficacy of globally inferring the temporal relations with the time graphs,
\item We release the codes\footnote{\url{https://github.com/yuyanislearning/CTRL-PG}.} to facilitate further developments by the research community.
\end{itemize}
Next, we give the problem definition and explain how we leverage PSL rules to model the temporal dependencies. We then describe the overall architecture of our clinical temporal relation extraction model, \texttt{CTRL-PG}\xspace and show the extensive experimental results in the following sections.
\section{Clinical Temporal Relation Extraction}\label{sec:model}
Figure~\ref{fig:framework} shows the overall framework of the proposed \texttt{CTRL-PG}\xspace model. The framework consists of three components, (i) a temporal relation classifier composed of a deep language encoder and a Feed-Forward Network (FFN), (ii) a Cross-Entropy loss function with PSL regularization, and (iii) a time-anchored global temporal inference module. We will introduce the details of the three modules in the following subsections.
\subsection{Temporal Relation Classifier}
The context is essential for capturing the syntactic and semantic features of each word in a sequence. Hence, we propose to apply the contextualized language model, BERT~\cite{devlin2018bert}, to derive the sentence representation $v_i$ of $d_s$-dimension to encode the input sequence $s_i$ including two marked named entities $x_{i,1}, x_{i,2}$ from the instance $\mathcal{I}$, where $i\in\{1,2,3\}$. We group three sequences together to facilitate the computation of regularization term introduced in the next subsection.
By feeding the sentence embedding $v_i$ to a layer of FFN, we can predict the relation type $\hat{y_i}$ with the softmax function:
\begin{align}
& \hat{y}_i = \argmax_{y \in \mathcal{Y}} \; \mathbb{P} (y|s_i)\\
& \mathbb{P}(y|s_i) = \text{softmax}( W_f \cdot v_i + b_f),
\end{align}
where $W_f$ and $b_f$ are the weights and bias in the FFN layer.
To learn the relation classification model, we first compute a loss with the Cross-Entropy objective for each instance $\mathcal{I}$:
\begin{align}\label{eq:ce}
& \mathcal{L}_{ce} = -\sum_{i\in\{1,2,3\}}\sum_{y\in\mathcal{Y}} y\log\mathbb{P}(y|s_i)
\end{align}
\subsection{Learning with Probabilistic Soft Logic Regularization}\label{sec:pslr}
We also aim to minimize the distance to rule satisfaction for each instance. We compute the distance with function $\mathcal{F}(\cdot,\cdot)$, as described in Algorithm~\ref{algo:PSL},
by finding the minimum of all possible PSL rule grounding results, i.e., when one PSL rule is satisfied, $\mathcal{F(\cdot,\cdot)}$ should return $0$.
In specific, we first ground the three relation predictions $\hat{y}_{i}$ with potential PSL rules.
We then incorporate Equation~\eqref{eq:luk}-\eqref{eq:dis} for distance computation. The prediction probabilities are regarded as the interpretation of the ground atoms $l_i$.
If none of the rules can be grounded, the distance will be set as 0.
\begin{algorithm}[t]
\SetAlgoLined
\textbf{Input:} PSL Rules $\mathcal{R}$, Prediction $\hat{y}_i$, and Probability $\mathbb{P}(y|s_i)$, $i=\{1,2,3\}$\;
\textbf{Output:} Distance $d_r$\;
Set $d_r=1$; $d_t=0$; IsGround $=false$\;
\For{each $l_1\land l_2 \to l_3 \in \mathcal{R}$}{
\uIf{$\hat{y}_{1}$ matches $l_1$ and $\hat{y}_{2}$ matches $l_2$}{
Determine $\bar{y}_{3}$ with $l_3$\;
$d_t \leftarrow \max\{\mathbb{P}(y=\hat{y}_{1}|s_1) + \mathbb{P}(y=\hat{y}_{2}|s_2) - 1, 0\}$\;
$d_t \leftarrow \max\{d_t - \mathbb{P}(y=\bar{y}_{3}|s_3) , 0\}$\;
$d_r \leftarrow \min\{d_r, d_t\}$\;
IsGround $\leftarrow true$\;
}
}
\uIf{IsGround $==false$}{$d_r \leftarrow 0$\;}
\caption{Function $\mathcal{F}$ for PSL Rule Grounding and Distance Calculation.}
\label{algo:PSL}
\end{algorithm}
Then, we formulate the distance to satisfaction as a regularization term to penalize the predictions that violate any PSL rule:
\begin{align}\label{eq:psl}
& \mathcal{L}_{psl} = \mathcal{F}(\mathcal{R}; \{(\mathbb{P}(y|s_i),\hat{y}_i)\}), i=\{1,2,3\}
\end{align}
and finalize the loss function by summing up \eqref{eq:ce} and \eqref{eq:psl}:
\begin{align}\label{eq:loss}
& \mathcal{L} = \mathcal{L}_{ce} + \lambda\cdot \mathcal{L}_{psl},
\end{align}
where $\lambda$ is a hyperparameter as the weight for PSL regularization term.
We apply gradient descent to minimize the loss function~\eqref{eq:loss} and to update the parameters of our model.
\subsection{Global Temporal Inference}
In the inference stage, we leverage the Timegraph algorithm~\cite{miller1990time} to resolve the conflicts in the temporal relation predictions $\bm{\hat{y}}$.
Timegraph is a widely used algorithm of time complexity $\mathcal{O}(v+e)$ for deriving the temporal relation for any two nodes in a connected graph, where $v$ and $e$ denote the numbers of nodes and edges. Nodes and edges represent the named entities and temporal relations, respectively.
Our goal is to construct a conflict-free time graph $\mathcal{G}$ for each document $D$ through a greedy Check-And-Add process, described as $4$ steps in Algorithm~\ref{algo:gri}.
Intuitively, we want to rely on some trustworthy edges to resolve the conflicts in the time graph with the transitivity and symmetry dependencies listed in Table~\ref{tab:psl}. As illustrated in Figure~\ref{fig:framework}, the probabilities of predictions \underline{Overlap}(e1, e2) and \underline{Overlap}(e2, e3) are $0.7$ and $0.8$, which are higher than that of \underline{Overlap}(e1, e3). When we trust the first two predictions, the third prediction could be neglected considering the relation between e1 and e3 can already be inferred with the transitivity dependency. In this way, the predicting mistakes with low confidence scores can be ruled out, leading to better model performance in the closure evaluation.
We believe that the relations between time expressions are the easiest ones to predict. For example, the ground atom \underline{Before} (\textit{06-15-91}, \textit{July 1st 1991}) is obviously $true$. Therefore, we try to build up a base time graph on top of the relations of type \texttt{T-T}. Next, we rank the rest of the predictions according to their probabilities in decreasing order and then check whether each of the predictions is inconsistent with the current time graph iteratively. The relation will be dropped if it raises a conflict, otherwise added to the graph as a new edge.
\begin{algorithm}[t]
\SetAlgoLined
Step 1: Predict temporal relations $P_1$ on pairs of the time expressions \texttt{T-T}\;
Step 2: Construct a time graph $\mathcal{G}$ with $P_1$\;
Step 3: Rank all other predictions $P_2$ on the relations of type \texttt{E-E} and \texttt{E-T} according to the predicting probabilities in decreasing order, naming $P_2^{ranked}$\;
Step 4: \\
\For{each $p$ in $P_2^{ranked}$}{
Apply Timegraph algorithm to check the conflict between $p$ and $\mathcal{G}$\;
\uIf{there exists a conflict}{Drop $p$\;}
\Else{Add the edge $p$ to $\mathcal{G}$\;}
}
\caption{Check-And-Add Process for Constructing a Conflict-free Time Graph $\mathcal{G}$}
\label{algo:gri}
\end{algorithm}
\section{Experiments}\label{sec:exp}
In this section, we develop experiments on two benchmark datasets to prove the effectiveness of both PSL regularization and global temporal inference. We also discuss the limitation and perform error analyses for \texttt{CTRL-PG}\xspace.
\begin{table}[ht!]
\centering
\resizebox{.95\linewidth}{!}{
\begin{tabular}{|c|l|r|r|r|}
\hline
\multicolumn{2}{|c|}{Dataset} & Train & Dev & Test \\
\hline
\hline
\multirow{2}{*}{I2B2-2012} & \# doc & 181 & 9 & 120 \\
\cline{2-5}
& \# relation & 29,736 & 1,165 & 24,971 \\
\hline
\multirow{2}{*}{TB-Dense} & \# doc & 22 & 5 & 9 \\
\cline{2-5}
& \# relation & 4,032 & 629 & 1,427 \\
\hline
\end{tabular}}
\caption{Dataset Statistics.}
\label{tab:data}
\end{table}
\subsection{Datasets}
Experiments are conducted on I2B2-2012 and TB-Dense datasets and an overview of the data statistics is shown in Table~\ref{tab:data}. The datasets have diverse annotation densities and instance numbers.
\noindent \textbf{I2B2-2012.} The I2B2-2012 challenge corpus~\cite{sun2013evaluating} consists of 310 discharge summaries. Two categories of temporal relations, \texttt{E-T} and \texttt{E-E}, were annotated in each document.
Three temporal relations\footnote{\citet{sun2013evaluating} merged 7 original temporal relations to 3 to increase Inter-annotator agreement.}, \underline{Before}, \underline{After}, and \underline{Overlap}, were used. I2B2-2012 has a relatively low annotation density\footnote{Annotation density denotes the percentage of annotated pairs of event/time expressions.}, which is $0.21$.
\noindent \textbf{TB-Dense.} To prove that our PSL regularization is a generic algorithm and can be easily adapted to other domains, we also test it on the TB-dense~\cite{cassidy-etal-2014-annotation} dataset, which is based on TimeBank News Corpus~\cite{pustejovsky2003timebank}. Annotators were required to label all pairs of events/times in a given window to address the sparse annotation issue in the original data. Thus the annotation density is 1. This dataset has six relation types, \underline{Simultaneous}, \underline{Before}, \underline{After}, \underline{Includes}, \underline{Is\_Include}, and \underline{Vague}.
\subsection{Baseline Models}
We employ different baseline models for the two datasets to compare our method with the SOTA models in both clinical and news domains.
\\
\noindent \textbf{I2B2-2012} (1) Feature-engineering based statistic models from I2B2-2012 challenge, \texttt{MaxEnt-SVM}~\cite{xu2013end} incorporating Maximum Entropy with Support Vector Machine (SVM), \texttt{CRF-SVM}~\cite{tang2013hybrid} using Conditional Random Fields and SVM, \texttt{RULE-SVM}~\cite{nikfarjam2013towards} relying on rule-based algorithms; (2) Neural network based model, \texttt{RNN-ATT}~\cite{liu-etal-2019-attention}, which applies Recurrent Neural Network plus attention mechanism; (3) Structured Prediction method, \texttt{SP-ILP}~\cite{han-etal-2019-deep,leeuwenberg-moens-2017-structured} leveraging the ILP optimization; (4) Basic version of our model, \texttt{CTRL}, which only fine-tunes a BERT-BASE~\cite{devlin2018bert} language model with one layer of FFN, similar to the implementations in \citet{lin-etal-2019-bert,guan2020robustly}.
\\
\noindent \textbf{TB-Dense.} (1) \texttt{CAEVO}~\cite{chambers2014dense} with a cascade of rule-based classifiers; (2) \texttt{LSTM-DP}~\cite{cheng-miyao-2017-classifying} using LSTM-based network and cross-sentence dependency paths; (3) \texttt{GCL}~\cite{meng-rumshisky-2018-context} incorporating LSTM-based network with discourse-level contexts; (4) \texttt{SP-ILP} and \texttt{CTRL}, same as the baselines for I2B2-2012. Note that the results of \texttt{CAEVO}, \texttt{LSTM-DP}, \texttt{GCL}, and \texttt{SP-ILP} are collected from \citet{han-etal-2019-deep}.
\subsection{Evaluation Metrics}
To be consistent with previous work for a fair comparison, we adopt two different evaluation metrics. For TB-Dense dataset, we compute the Precision, Recall, and Micro-average F1 scores. Following \cite{han-etal-2019-joint,meng-rumshisky-2018-context}, we only predict the \texttt{E-E} relations and exclude all other relations from evaluation. Note that Micro-averaging in a multi-class setting will lead to the same value for Precision, Recall, and F1.
For I2B2-2012, we leverage the TempEval evaluation metrics used by the official challenge~\cite{sun2013evaluating}, which also calculates the Precision, Recall, and Micro-average F1 scores.
This evaluation metrics differ from the standard F1 used for TB-Dense in a way that it computes the Precision by verifying each prediction in the closure of the ground truths and computes the Recall by verifying each ground truth in the closure of the predictions. We explore all types of temporal relations in I2B2-2012 dataset.
\subsection{Implementation Details}
In the framework of \texttt{CTRL-PG}\xspace, any contextualized word embedding method, such as BERT~\cite{devlin2018bert}, ELMo~\cite{peters2018deep}, and RoBERTa~\cite{liu2019roberta}, can be utilized. We choose BERT~\cite{devlin2018bert} to derive contextualized sentence embeddings without loss of generality. BERT adds a special token \texttt{[CLS]} at the beginning of each tokenized sequence and learns an embedding vector for it. We follow the experimental settings in \cite{devlin2018bert} to use $12$ Transformer layers and attention heads and set the embedding size $d_s$ as 768.
The \texttt{CTRL-PG}\xspace is implemented in PyTorch and we use the fused Adam optimizer~\cite{kingma2014adam} to optimize the parameters. We follow the experimental settings in \cite{devlin2018bert} to set the dropout rate, and batch size as $10^{-1}$ and 8. We perform grid search for the initial learning rate from a range of $\{1\times 10^{-5},2\times 10^{-5},4\times 10^{-5},8\times 10^{-5}\}$ and finally select $2\times 10^{-5}$ for both datasets.
We train 10 epochs for each experiment on two datasets, which can all be completed within 2 hours on single DGX1 Nvidia GPU.
\input{figures/lambda}
To tune the hyperparameters, we search the PSL regularization term $\lambda$ from $\{0.1, 0.5 , 1, 2, 5, 10\}$ as shown in Figure~\ref{fig:hyper}. For I2B2-2012 and TB-Dense datasets, we set $\lambda$ as $5$ and $0.5$, respectively. The hyperparameters are selected by observing the best F1 performance on the validation set. More implementation details can be found in the Appendix.
\begin{table}[h]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|c|R{1.5cm}|R{1.5cm}|R{1.5cm}|}
\hline
Model & P & R & F1 \\
\hline
\hline
\texttt{RULE-SVM} &71.09&58.39&64.12 \\
\texttt{MaxEnt-SVM} &74.99&64.31&69.24 \\
\texttt{CRF-SVM} &72.27&66.81&69.43\\
\hline
\hline
\texttt{RNN-ATT} &71.96&69.15&70.53\\
\texttt{SP-ILP} &78.15&\textbf{78.29}&78.22\\
\texttt{CTRL} &84.88&73.28&78.65 \\
\hline
\hline
\texttt{CTRL-PG}\xspace &\textbf{86.80}&74.53&\textbf{80.20} \\
\hline
\end{tabular}}
\caption{Performance of temporal relation extraction on I2B2-2012 datasets. All improvements of \texttt{CTRL-PG}\xspace over baseline methods are statistically significant at a 99\% confidence level in paired \textit{t}-tests. Results show that \texttt{CTRL-PG}\xspace outperforms all the baselines. }
\label{tab:result1}
\end{table}
\begin{table}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|l|R{1.4cm}|R{1.4cm}|R{1.4cm}|R{1cm}|}
\hline
Feature & P & R & F1 & Lift\\
\hline
\hline
Best & 86.80 & 74.53 & 80.20 & -\\
\hline
w/o PSL & 85.78 & 73.31 & 79.06 & \textbf{1.44}\% \\
\hline
w/o GTI & 85.08 & 73.31 & 78.76 & \textbf{1.83}\% \\
\hline
\end{tabular}}
\caption{Ablation study on I2B2-2012 dataset. GTI denotes the global temporal inference. Results show significant performance lifts from both PSL and GTI modules.}
\label{tab:abl}
\end{table}
\begin{table}[h]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|l|R{1.3cm}|R{1.3cm}|R{1.3cm}|R{1cm}|}
\hline
Strategy & P & R & F1 & Lift\\
\hline
\hline
Random & 85.08 & 73.93 & 79.21 & -\\
\hline
Confidence & 86.07 & 73.76 & 79.44 & \textbf{0.29\%}\\
\hline
Confidence + & \multirow{2}{*}{86.80} & \multirow{2}{*}{74.53} & \multirow{2}{*}{80.20} & \multirow{2}{*}{\textbf{1.25}\%}\\
Time Anchor & & & &\\
\hline
\end{tabular}}
\caption{Comparison of different ranking methods applied in the global inference on I2B2-2012 dataset.}
\label{tab:gti}
\end{table}
\begin{table*}[h]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|c|c|c|c|c|}
\hline
\multirow{7}{*}{1} & \multirow{2}{*}{Text} & \multicolumn{3}{l|}{{\color{red} Her acute bradycardic event} was felt likely secondary to {\color{blue}her new beta blocker} in conjunction with} \\
& & \multicolumn{3}{l|}{a vagal response . It was determined to stop {\color{green}the beta blocker} , and atropine was placed at the bedside .}\\
\cline{2-5}
& (e1, e2) & ({\color{red}Her...event}, {\color{blue}her...blocker}) & ({\color{blue}her...blocker}, {\color{green}the beta blocker}) & ({\color{red}Her...event}, {\color{green}the beta blocker})\\
\cline{2-5}
& True Label & After & Overlap & After\\
\cline{2-5}
& \texttt{CRTL} & After & Overlap & Overlap \\
\cline{2-5}
& \texttt{CTRL-PG}\xspace & After & Overlap & After \\
\cline{2-5}
&Rule & \multicolumn{3}{c|}{After$(A,B)$ $\land$ Overlap$(B,C)$ $\rightarrow$ After$(A,C)$}\\
\hline
\hline
\multirow{7}{*}{2} & \multirow{2}{*}{Text} & \multicolumn{3}{l|}{The patient was given an aspirin and Plavix and in addition {\color{red}started} on {\color{blue}a beta Elmore} , {\color{green}Maxine ACE}} \\
& & \multicolumn{3}{l|}{{\color{green}inhibitor} , and these were titrated up as her blood pressure tolerated .}\\
\cline{2-5}
& (e1, e2) & ({\color{red}started}, {\color{blue}a beta Elmore}) & ({\color{blue}a beta Elmore}, {\color{green}Maxine ACE}) & ({\color{red}started}, {\color{green}Maxine ACE})\\
\cline{2-5}
& True Label & Before & Overlap & Before\\
\cline{2-5}
& \texttt{CRTL} & After & Overlap & Before \\
\cline{2-5}
& \texttt{CTRL-PG}\xspace & After & Overlap & After \\
\cline{2-5}
&Rule & \multicolumn{3}{c|}{After$(A,B)$ $\land$ Overlap$(B,C)$ $\rightarrow$ After$(A,C)$}\\
\hline
\hline
\multirow{6}{*}{3} & Text & \multicolumn{3}{l|}{She has had attacks treated with {\color{red}antibiotics} in the past notably in {\color{blue}12/96} and {\color{green}08/97} .} \\%Lipase was within normal limits.\\
\cline{2-5}
& (e1, e2) & ({\color{red}antibiotics}, {\color{blue}12/96}) & ({\color{blue}12/96}, {\color{green}08/97}) & ({\color{red}antibiotics}, {\color{green}08/97})\\
\cline{2-5}
& True Label & Overlap & Before & Overlap\\
\cline{2-5}
& \texttt{CRTL} & Overlap & Before & Overlap \\
\cline{2-5}
& \texttt{CTRL-PG}\xspace & Overlap & Before & Before \\
\cline{2-5}
&Rule & \multicolumn{3}{c|}{Overlap$(A,B)$ $\land$ Before$(B,C)$ $\rightarrow$ Before$(A,C)$}\\
\hline
\end{tabular}}
\caption{Case study and error analysis of the model predictions on I2B2-2012 Dataset.}
\label{tab:error}
\end{table*}
\begin{table}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{|c|r|r|r|r|r|r|}
\hline
\multirow{2}{*}{} & \multicolumn{3}{c|}{\texttt{SP-ILP}} & \multicolumn{3}{c|}{\texttt{CTRL-PG}\xspace} \\
\cline{2-7}
& P & R & F1 & P & R & F1 \\
\hline
\hline
Before &71.1&58.9&64.4&52.6&74.8&61.7 \\
After &75.0&55.6&63.5&69.0&72.5&70.7 \\
Includes &24.6&4.2&6.9&60.9&29.8&40.0 \\
Is\_Include &57.9&5.7&10.2&34.7&27.7&30.8 \\
Simultaneous &-&-&-&-&-&-\\
Vague &58.3&81.2&67.8&72.8&64.8&68.6 \\
\hline
Micro-average &\multicolumn{3}{r|}{63.2}&\multicolumn{3}{r|}{\textbf{65.2}} \\
\hline
\hline
\texttt{CAEVO} &\multicolumn{6}{r|}{49.4} \\
\texttt{LSTM-DP} &\multicolumn{6}{r|}{52.9} \\
\texttt{GCL} &\multicolumn{6}{r|}{57.0} \\
\texttt{CTRL} &\multicolumn{6}{r|}{63.6} \\
\hline
\end{tabular}}
\caption{Performance of temporal relation extraction on TB-dense datasets. All improvements of \texttt{CTRL-PG}\xspace over baseline methods are statistically significant at a 99\% confidence level in paired \textit{t}-tests. We also compare the breakdown performance for each relation class between \texttt{CTRL-PG}\xspace and \texttt{SP-ILP}.}
\label{tab:result2}
\end{table}
\subsection{Experimental Results}
Table~\ref{tab:result1} and Table~\ref{tab:result2} contains our main results. As we observe, our \texttt{CTRL-PG}\xspace enhanced by PSL regularization and global inference achieve the best relation extraction performances per F1 score. Compared with the baseline models, the F1 score improvements are 2.0\% and 2.5\% on I2B2-2012 and TB-Dense data respectively, which are all statistically significant.
\\
\noindent \textbf{I2B2-2012.}
As shown in Table~\ref{tab:result1}, our model \texttt{CTRL-PG}\xspace outperforms the best baseline method \texttt{CTRL} by 2\% and outperforms the structured prediction method \texttt{SP-ILP} by 2.5\% per F1 score. \texttt{SP-ILP} gets the highest Recall score, but sacrifice the predicting precision instead. We also observe that by simply fine-tuning the BERT to generate the sentence embeddings and then feeding them into one layer of FFN for classification, \texttt{CTRL} can achieve an impressive F1 score of 78.65\%. This proves the advantage of contextualized embeddings over static embeddings used by other baseline models. Besides, \texttt{CTRL-PG}\xspace outperforms the feature-based systems, \texttt{CRF-SVM} and \texttt{MaxEnt-SVM}, by over 10\% per F1 score.
We develop an ablation study to test different features, as shown in Table~\ref{tab:abl}. We see that PSL regularization and global temporal inference modules lift the performance by 1.44\% and 1.83\% separately. Both Precision and Recall performances are improved. We can clearly conclude that learning the relations with the proposed algorithms improves our model significantly (also at a 99\% level in paired \textit{t}-tests).
We also show the comparisons among different ranking strategies for the global inference module in Table~\ref{tab:gti}. Random denotes that we randomly add a new prediction to the time graph and resolve the conflict. Confidence denotes we rank the predictions per the prediction probabilities and then add them to the graph in decreasing order. Time Anchor represents that we first construct the time graph based on the predictions for temporal relations of type \texttt{T-T}. In the results, we see a 0.29\% improvement per F1 score when switching from the Random to the Confidence strategy. After adding the Time Anchor method, we observe a 1.25\% performance lift, compared to Random strategy.
This proves the effectiveness of the time-anchored global temporal inference module.
\\
\noindent \textbf{TB-Dense.}
We show the experimental results on TB-Dense dataset in Table~\ref{tab:result2}. Our model outperforms the best baseline model \texttt{CTRL} by 2.5\% and outperforms the structured prediction method \texttt{SP-ILP} by 3.2\% per Micro-average F1 score. We observe that in the performance breakdown for each relation class, \texttt{CTRL-PG}\xspace obtains similar scores on \underline{Before}, \underline{After}, and \underline{Vague} as \texttt{SP-ILP} and gets much better performances on \underline{Is\_Include} and \underline{Includes}. These two types only occupy 5.7\% and 4.5\% of all the instances. \texttt{CTRL-PG}\xspace and \texttt{SP-ILP} both fail to label any instance as \underline{Simultaneous} because of its even fewer instances (1.5\%) for training.
Besides, we observe \texttt{CTRL-PG}\xspace achieves higher Recall values in all the categories of temporal relations, which prove that incorporating the dependency rules into model training can dramatically lift the coverage of predictions.
\subsection{Case Study and Error Analysis.}
Table~\ref{tab:error} shows the results of a case study with the outputs of \texttt{CTRL} and \texttt{CTRL-PG}\xspace. In the first case, the temporal relation between \textit{Her acute bradycardic event} and \textit{the beta blocker} is hard to predict due to the noise brought by the long context. \texttt{CTRL} predicts it as \underline{Overlap}, while \texttt{CTRL-PG}\xspace corrects it to \underline{After} according to the potential PSL rule that can be matched with the first two correct predictions. In some cases, however, \texttt{CTRL-PG}\xspace will make new mistakes. For example in case 2, if our model initially predicts the relation between \textit{started} and \textit{a beta Elmore} wrong, a potential PSL rule sometimes will lead to an extra mistake when predicting the relation between \textit{started} and \textit{Maxine ACE}. In the case 3, \textit{antibiotics} treated the \textit{attacks} twice in both $12/96$ and $08/97$, where the PSL rule is no longer valid since the \textit{antibiotics} in fact denote two occurrences of this event. In such special cases with invalid rules, \texttt{CTRL-PG}\xspace may make a mistake.
\section{Appendices}
\label{sec:appendix}
\subsection{Dataset}
The I2B2-2012 dataset\footnote{\url{https://www.i2b2.org/NLP/DataSets/}} is officially split into training and test sets, containing 190 and 120 documents, separately. We randomly sampled 5\% of training data as a validation set.
For
TB-Dense\footnote{\url{https://www.usna.edu/Users/cs/nchamber/caevo/}}, the training/validation/test sets are given. The statistics are shown in Table~\ref{tab:data} and we plot the detailed relation type distributions of two datasets in Figure~\ref{fig:type}. We observe that
TB-Dense is a relatively unbalanced dataset, where \underline{Vague} dominates the dataset. We compute the density by (\# existing relations)/(\# possible pairs of entities).
Following \cite{han-etal-2019-joint,meng-rumshisky-2018-context}, we only predict the \texttt{E-E} relations and exclude all other relations from evaluation for TB-Dense dataset evaluation. Therefore, we cannot apply the global inference module to this dataset.
\begin{figure}[ht]
\centering
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=\linewidth]{images/i2b2data.pdf}
\caption{I2B2-2012}
\label{fig:i2b2data}
\end{subfigure}
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=\linewidth]{images/tbddata.pdf}
\caption{TB-Dense
}
\label{fig:tbddata}
\end{subfigure}
\caption{Relation Type Distribution of I2B2-2012 and TB-Dense}
\label{fig:type}
\end{figure}
\subsection{Data Preprocessing}
To facilitate the PSL rule grounding and distance calculation,
we arrange the training data as a collection of instances, each of which contains three pairs of relations.
We first traverse the training dataset to match the PSL rules and make sure that each ground rule is treated as one instance to further calculate the PSL loss term. Besides, relations that are not involved in any rules are packed together and we do not have to compute a PSL loss for them.
We also augment the training data by flipping every pair according to the symmetry rules, i.e. if \underline{Before}($A,B$) is $true$, \underline{After}($B,A$) should also be $true$. We incorporate the new relations into the training dataset to alleviate the unbalanced data issue.
\subsection{Model Training Details}
We leverage the pretrained BERT-BASE model~\cite{devlin2018bert} to generate the sentence embeddings, which contains 110M parameters to fine-tune.
In the experiments, we save the checkpoint with the highest validation performance for final testing. In Table~\ref{tab:dev}, we list the best validation performance of different datasets and the corresponding test performance that we also reported in Table~\ref{tab:result1} and Table~\ref{tab:result2} for completeness.
Data and codes are attached.
\begin{table}[h]
\centering
\resizebox{.95\linewidth}{!}{
\begin{tabular}{|c|l|r|r|r|}
\hline
\multicolumn{2}{|c|}{Dataset} & Precision & Recall & F1 \\
\hline
\hline
\multirow{2}{*}{I2B2-2012} & Test & 86.80 & 74.53 & 80.20 \\
\cline{2-5}
& Val & 86.03 & 86.03 & 86.03 \\
\hline
\multirow{2}{*}{TB-Dense} & Test & 65.20 & 65.20 & 65.20 \\
\cline{2-5}
& Val & 58.80 & 58.80 & 58.80 \\
\hline
\end{tabular}}
\caption{Corresponding validation performance for each reported test result of \texttt{CTRL-PG}\xspace in Table~\ref{tab:result1} and Table~\ref{tab:result2}.}
\label{tab:dev}
\end{table}
| 7148c014b499fef7880c49a1ec037f4699a011bd | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\subsection{Astrophysical background}
Nearly all astrophysical objects, ranging from planets,
accretion disks, stars and galaxies to the intergalactic
medium, host magnetic fields coherent over the largest
scales of the system. At the same time, the matter in
which the magnetic fields originate, is in a vigorously
turbulent state, and often the driving scale of the
turbulence is at small or intermediate scales with
respect to the system scale
\citep[see, e.g.,][]{rincon2019}.
Hence, a theoretical
explanation of how these objects can sustain large-scale
magnetic fields, driven by small-scale turbulence, is required.
One such theoretical framework is the theory
of $\alpha \Omega$ dynamos, where helical turbulence
together with large-scale non-uniformities in the
rotation profile
excite magnetic field at the largest
scales \citep[as originally proposed by][]{Parker55a}.
Here, kinetic helicity is thought to arise from stratification and rotation.
Large-scale dynamos (hereafter LSD) do not necessarily
need rotational non-uniformities, but can also work
solely based on helical turbulence, then denoted as
$\alpha^2$ dynamos
\citep[see, .e.g., ][]{KR80}.
As the $\alpha$ effect is such a
fundamental building block of LSDs, studying it in
isolation has been a persistent task.
Many details and questions, however, still remain open, especially at vigorously turbulent regimes, which to reach numerical models still struggle.
Another dynamo instability, namely the fluctuation dynamo (or small-scale dynamo, hereafter SSD) is excited in
astrophysical flows for magnetic Reynolds
numbers ($\mathrm{Re}_\mathrm{M}$) exceeding a threshold value, which is thought
to happen in most astrophysical settings
\citep[as originally proposed by][]{Kazantsev1968}.
SSD generates
random magnetic fields primarily below the scales of the
forcing, and their growth rate is high, providing a
plausible explanation for magnetic fields in galaxy
clusters, or fluctuating fields seen on
the solar surface.
The latter, however, remains under
some debate, as SSD in a low magnetic Prandtl number ($\mathrm{Pr}_\mathrm{M}$)
environment like the Sun is notoriously hard to excite in numerical
experiments.
In nearly all astrophysical objects, LSD and SSD instabilities may co-exist. Their interactions,
however, are poorly understood, mainly because it is very
challenging numerically to include them both in one
and the same model, and only quite recently, such
modelling efforts have become feasible. Also, whenever
turbulent enough regimes can be reached, it becomes
very difficult to disentangle the two dynamos, as
also LSD produces fluctuating magnetic fields by
turbulent tangling of the large-scale field.
One of the earliest theoretical scenarios was
catastrophic quenching of the LSD by the growing
magnetic fluctuations, resulting in the suppression
of the $\alpha$ effect proportional to ${\mathrm{Re}_\mathrm{M}}^{-1}$,
meaning in practise that no $\alpha$-effect related LSD could be excited in astrophysical objects \citep{CV91}. This is
now understood to be a special case, detrimentally constrained by magnetic helicity
conservation, e.g.\ due to closed boundaries, such that
helicity fluxes cannot occur \citep{Bran2005review}. How these fluxes,
which alleviate catastrophic quenching, occur in
cosmic objects, however, is not known in detail.
It has also been
proposed that SSD can help LSD in shear dynamos
in the absence of the $\alpha$ effect, through
the so-called
magnetic shear-current effect
\citep[e.g.][]{SB15a}, but its potential
still remains debated
\citep[e.g.][]{SMHD}.
\cite{Hotta2016} claimed that, in simulations of turbulent magnetoconvection, the SSD would first suppress LSD at intermediate $\mathrm{Re}_\mathrm{M}$, but would let it recover at higher $\mathrm{Re}_\mathrm{M}$. This result was based on measuring the strength of the large--scale field at a few $\mathrm{Re}_\mathrm{M}$ values, but the diffusion scheme was changed in between the different runs, due to which a straightforward interpretation is difficult.
Many numerical studies have been undertaken concentrating on LSDs by helical forcing or SSDs by non-helical forcing,
in the former case also including
unintentionally or intentionally both dynamos.
\citet{Brandenburg2001} demonstrated how helically driven turbulence could give rise to large-scale structures based on non-local interactions at the forcing scale. Their
simulations employed isotropically
driven forced turbulence with various
resolutions and Reynolds numbers, to systematically
demonstrate the inverse cascade of MHD-turbulence.
In some of these runs both LSD and SSD were present, but even
though the evolution of mean and fluctuating fields was
monitored separately, no attempt to study the LSD-SSD
interactions was made.
In the galactic context,
\citet{Gent2013a} studied supernova-driven flows, where the
magnetic Reynolds number again permitted both dynamos
together. They made an attempt to separate the growth
rates of SSD and LSD by using a Gaussian smoothing
procedure, but their setup was too complex to derive
any reliable information on the two dynamo processes.
Moreover, their viscosity scheme allowed for
a spatially varying
$\mathrm{Pr}_\mathrm{M}$, hence the excitation conditions
for the different dynamos were more favourable in regions with hot
gas, in which most likely, most of the SSD action occurred.
Both dynamos together have also been seen in turbulent
convection simulations \cite{KKB08}, and there different
growth rates were detected for mean and fluctuating
magnetic fields, following a $\mathrm{Re}_\mathrm{M}^{1/2}$ scaling.
\cite{Brandenburg2018} compare the results from
several studies of helically and non-helically forced
turbulence models, and show that the
growth rate of the SSD is following the same
scaling. Moreover, in helical system possessing both dynamos,
the same growth rate is observed. Their
data, however, is sparse, hence
the $\mathrm{Re}_\mathrm{M}$-dependence not very certain.
There is a rich literature on SSDs, studied in
isolation in non-helical setups. Most of it concentrates
either on high $\mathrm{Pr}_\mathrm{M}$, relevant for
ISM and intergalactic medium, or low $\mathrm{Pr}_\mathrm{M}$, important for accretion disks and stellar
convection zones. Both regimes are numerically extremely
challenging, as the magnetic diffusivity has to be
set to values much lower/higher than viscosity, and
resolving such systems numerically is difficult. Here
we avoid these complications by concentrating
on the regime $\mathrm{Pr}_\mathrm{M}=1$ and make an
effort to analyze the interaction of
LDS and SSD.
The objective of this study is twofold. First, we aim at
replicating the work of \citet{Brandenburg2001},
and then at exploring a
wider range of resolution and $\mathrm{Re}_\mathrm{M}$. Second, we provide the first
physical application of the multi-GPU magnetohydrodynamics code \textit{Astaroth} \citep{Astaroth2017, vaisala2017thesis, Pekkila2019} which features novel methods for efficiently calculating high-order finite-difference derivatives, based on large stencils.
\subsection{Emergence of GPU computation} \label{sec:acbg}
In the last ten years, the emergence of graphics processing units (GPUs) has
enabled several times higher throughput in data-parallel tasks, compared
with central processing units (CPUs) traditionally used in high-performance
computing\footnote{
A Tesla V100-SXM2-32GB GPU provides an arithmetic performance of $7.83$ TFLOPS
(floating-point operations per second) and $863$ GiB/s off-chip memory
bandwidth~\citep{volta-whitepaper, Jia2018}, whereas an Intel Xeon Gold 6230
CPU has the theoretical peak performance of $1.25$ TFLOPS and $131$ GiB/s
bandwidth~\citep{intel-whitepaper}. Therefore a GPU could theoretically
provide roughly $6\times$ improved throughput in data-parallel tasks.
}. GPUs excel in tasks, where the same operation can be
executed on a very high number of data elements in parallel. In contrast to
CPUs, GPUs have been designed to maximize the throughput of memory systems
with the cost of higher memory access latency~\citep{hennessy_computer}.
Therefore, ensuring there is a sufficient amount of parallel work to hide
latencies is critical for obtaining high performance.
GPUs provide an attractive platform for stencil codes, where each grid point in
the problem domain can be updated in parallel. Stencil codes are commonly used
in, for example, finite-difference fluid simulations~\citep{Brandenburg2001} and image processing~\citep{mullapudi_polymage, kelley_halide}. However, following the multi-core revolution and the introduction of highly-parallel accelerators to general-purpose computing, converting existing codes to use all of the capabilities of the hardware has been a significant challenge~\citep{Asanovic2009}. Writing efficient programs for these architectures often requires deep knowledge in their hardware and execution models.
GPUs are programmed using the stream programming model, where the programmer defines a stream of instructions to be executed in parallel on a multitude of stream processors. Each
individual stream
can access data from different memory locations or follow different execution paths.
A notable complication in GPU programming is finding efficient caching techniques to reduce pressure onto off-chip memory. As GPUs are capable of high arithmetic throughput, it is paramount to ensure that the stream processors do not become starved of data. This is especially an issue in high-order finite-difference codes, where the ratio of arithmetic operations to bytes transferred is generally low. Because the optimal caching technique depends on several factors, such as the problem size and stencil shape, the optimal implementation for one workload does usually not carry to another. This presents a major obstacle in studying physical phenomena, or developing new mathematical models by GPUs, as a significant amount of time must be spent on writing, debugging and optimizing the code.
There have been several proposals to make GPU programming more convenient, e.g. high-level language extensions, such as \textit{OpenACC}~\citep{openacc}. However, high-level programming models are argued to lack the expressiveness to translate more complex tasks, especially those that require advanced caching techniques, into efficient code~\citep{sujeeth_delite, edwards2014}. More specialized approaches have also been suggested. Frameworks focusing on solving PDEs in structured grids include \textit{SBLOCK}~\citep{brandvik_sblock}, \textit{Fargo3D}~\citep{benitez_fargo3d} and \textit{Cactus}~\citep{goodale_cactus}. Alternative approaches focused on achieving near hand-tuned performance and performance-portability have been demonstrated by \textit{Lift}~\citep{hagedorn_high} and \textit{Delite}~\citep{sujeeth_delite}. \textit{Lift} translates high-level algorithmic primitives into lower-level code based on rewrite rules, while \textit{Delite} provides an intermediate language, which can be used as a basis for building domain-specific languages (DSLs).
Yet another approach is to provide a compiler for generating efficient code
from sources written in a DSL. For example, \textit{Polymage}~\citep{mullapudi_polymage} and \textit{Halide}~\citep{kelley_halide} provide a DSL and a compiler for generating two-dimensional image processing pipelines. In contrast to the above approaches, \textit{Astaroth} has been tailored for high-order stencil computations with special consideration of caching coupled fields
commonly found in multiphysics simulations,
such as velocity coupled with magnetic field in MHD induction equation (See Equation \ref{eq:magnetic}).
By caching the results of intermediate stencil operations, traffic to main memory is reduced significantly when these intermediate results are used to update multiple fields.
In previous work, we have presented the \textit{Astaroth} library~\citep{Pekkila2019}.
In the case of this study, we have extended it to work on multiple GPUs and utilized it to simulate resistive MHD turbulence and the emergence of dynamos. Inter-GPU communication is carried out using peer-to-peer memory copy functions provided by the CUDA API. However, as these functions do not support inter-node transfers, our implementation is limited to computations within a single node. An implementation for multiple nodes is the subject of ongoing work.
\section{Core methods}\label{sec:astaroth}
In this study, we use the model of magnetohydrodynamics to examine the growth of small- and large-scale dynamos (SSD and LSD respectively).
Hence, the systems examined are essentially non-linear.
To excite a dynamo, turbulence is a highly suitable (but not necessary) ingredient.
Some properties of dynamo-capable systems can be described qualitatively:
In a small-scale dynamo, the magnetic field grows through cascading turbulence and the resulting entanglements, at scales smaller than the turbulence driving scale. In such a situation merely
sufficiently high magnetic Reynolds numbers are required \citep{Haugen2004}. In a large-scale dynamo, turbulent helical flows result in growth of magnetic field at scales larger than the flow scales, in the extreme at the largest scales possible \citep{Brandenburg2001}.
To understand LSD action, often the perspective of mean-field theory is taken which is discussed more closely in Section \ref{sec:soca}.
Because the dynamo processes are non-linear, exploring them requires direct numerical simulations (DNS).
They can mimic laboratory experiments by studying how the systems react to changing parameters.
For such DNS, codes are required which can support two central features. The first is resistive MHD, as a dynamo is commonly supposed to be impossible under ideal-MHD conditions
because magnetic reconnections are thought to be a necessary part of the self-amplification of the magnetic field \citep{rincon2019}.
The second important feature is a high-order numerical PDE solver to effectively minimize uncontrolled numerical diffusion and to resolve the fine structure of turbulence with high accuracy \citep{axelnum}.
There are many openly accessible codes which meet these requirements, such as the \textit{Pencil Code}~\citep{Pencil2020}. However, the Pencil Code works presently only with traditional CPU-parallelism via MPI communication.
GPU acceleration can significantly reduce the computational costs, and here we demonstrate the use of the GPU code \textit{Astaroth} with similar properties as the \textit{Pencil Code}. \textit{Astaroth} is a software library developed for accelerating stencil computations especially in high-order accurate simulations. In such tasks, \textit{Astaroth} has been shown to to provide higher throughput and energy-efficiency than CPU-based solvers~\citep{Pekkila2019}. The \textit{Astaroth} library consists of an application-programming interface (API), a domain-specific language (DSL), an optimizing compiler that performs source-to-source translation from DSL sources to CUDA kernels, and a toolbox for carrying out common tasks, such as executing reductions on GPUs.
However, this study uses \textit{Astaroth} in a more specific way.
The \textit{Astaroth} API is surrounded by supplementary code, like tools for purposes of input and output, testing, interfacing, data analysis and data processing.
These additions allow \textit{Astaroth} to be used as a self-sufficient MHD code. As a self-sufficient MHD code, \textit{Astaroth} has following properties, in addition to the general features provided by the library:
\begin{itemize}
\item Physics (with DSL)
\begin{itemize}
\item Continuity equation.
\item Momentum equations, with full description of viscosity.
\item Resistive induction equation.
\item Energy equation in terms of entropy and ideal gas equation of state. (Not used in this study.)
\item Isotropic random forcing.
\end{itemize}
\item Numerical methods
\begin{itemize}
\item 6th-order finite difference scheme for calculating derivatives.
\item 3rd-order 2N-Runge-Kutta time integration.
\end{itemize}
\item Auxiliary tools
\begin{itemize}
\item Simulation suite for running DNS, which handles the tasks required on the CPU host side.
\item Autotest suite to check coherence between GPU and CPU operations.
\item Limited live rendering features for testing and demonstration purposes.
\item Python toolbox for data post-processing, visualization and analysis.
\end{itemize}
\end{itemize}
What is novel about \textit{Astaroth}, is the versatile GPU implementation. Because \textit{Astaroth} can be directed using the Domain Specific Language (DSL), it is flexible with adding new physics operations additionally to the ones that exist without making demanding case-by-case implementations on the level of CUDA. In the following Section \ref{sec:multigpu}, we will discuss GPU implementation aspect of this work before returning to the physical problem (Sections \ref{sec:physics} and \ref{sec:results}).
\textit{Astaroth} outputs time series and binary datacubes, which were reduced,
Fourier transformed, imaged and fitted with \textit{Python} using tools from
the \textit{SciPy}, \textit{NumPy} and \textit{Matplotlib} packages
\citep{2020SciPy-NMeth, harris2020array,Hunter:2007}.
Our analysis resulted in several values from performed post-processing,
and for this, the \textit{Pandas} tool \citep{reback2020pandas,mckinney-proc-scipy-2010} proved to be useful. \textit{Pandas} made it possible to
store results from several post-processing routines into an extended table. This table could then be accessed to organize and cross-reference our results with relative ease.
\textit{Paraview} was utilized for 3D visualization \citep{paraview_ref}.
\section{Multi-GPU implementation}\label{sec:multigpu}
Next, we describe our approach of distributing the workload to multiple GPUs on a single computational node.
\subsection{Terminology}
\label{sec:terminology}
The finite-difference method belongs to the class of stencil schemes, where
data values assigned to points in a structured grid are updated by sampling the
neighborhood of each grid point according to a specific
pattern, called a {\it stencil}, see Figure~\ref{fig:stencils}. The radius of a symmetric stencil is denoted as $r$, which is the Chebyshev distance in grid indices
from the center point to the furthest points of the stencil. We use the term $k$th-order stencil for a stencil used to calculate derivatives with $k$th-order accurate finite differences.
The number of grid points in the computational domain is denoted by the triple $\mathbf{N} = (N_x, N_y, N_z)$. In this work, we split the grid along the $z$-axis for $p$ devices
such that the size of the computational domain, local to each device is $\mathbf{n} = (n_x, n_y, n_z) = (N_x, N_y, N_z / p)$. We use the term \textit{device} to refer to a GPU controlled by a \textit{host} CPU.
Some stencil points, required for updating grid points near its boundaries, fall outside the local computational domain. The entirety of those points is called the halo. The total size of the grid, including the halo, is therefore $\mathbf{m} = (n_x + 2r, n_y + 2r, n_z + 2r)$. For $p >= 2$, some of the halo points map to the computational domain of a neighboring device. The data values at these points must be communicated between the contributing devices after each update step. The structure of the grid is visualized in Figure~\ref{fig:halo}.
To make a distinction between the halo and the area which is subject to boundary conditions (BCs), we
call the latter the ghost zone. The ghost zone exists for all
devices assigned to the boundaries of the global computational domain. Our discussion here focuses
specifically on communicating non-ghost-zone halos between neighboring devices,
which we refer to as halo exchange. Halo exchange can be easily adapted to
support periodic BCs by wrapping ghost zones around the global computational domain instead of excluding them from communication. Other BCs
may require an additional communication step to update the ghost zone. These
are left out of scope.
\begin{figure}
\centering
\begin{tikzpicture}[scale=0.5]
\foreach \i in {-3,...,3}
{
\draw[pattern = north west lines, shift = {(\i, 0)}] (0,0) -- (1,0)--(1,1)--(0,1)--cycle;
\draw[pattern = north west lines, shift = {(0, \i)}] (0,0) -- (1,0)--(1,1)--(0,1)--cycle;
\draw[pattern = north west lines, shift = {(\i, \i)}] (0,0) -- (1,0)--(1,1)--(0,1)--cycle;
\draw[pattern = north west lines, shift = {(\i, -\i)}] (0,0) -- (1,0)--(1,1)--(0,1)--cycle;
}
\draw[fill = white, shift = {(0, 0)}] (0,0) -- (1,0)--(1,1)--(0,1)--cycle;
\draw[<->, fill = white, shift = {(-4, 1)}] (0, 0) -- (0, 3) node[midway, left] {$r$};
\end{tikzpicture}
\caption{
Two-dimensional cut of a sixth-order stencil used for computing first- and second-order derivatives in this work. Here $r = 3$.
}
\label{fig:stencils}
\end{figure}
\begin{figure}
\centering
\begin{tikzpicture}[scale=0.35]
\foreach \y in {-9,...,8}
{
\foreach \x in {-9,...,8}
{
\draw[pattern = north west lines, shift = {(\x,\y)}] (0,0) -- (1,0)--(1,1)--(0,1)--cycle;
}
}
\foreach \y in {-6,...,5}
{
\foreach \x in {-6,...,5}
{
\draw[fill=white, shift = {(\x,\y)}] (0,0) -- (1,0)--(1,1)--(0,1)--cycle;
}
}
\foreach \y in {-3,...,2}
{
\foreach \x in {-9,...,8}
{
\draw[pattern = north east lines, opacity = 0.5, shift = {(\x,\y)}] (0,0) -- (1,0)--(1,1)--(0,1)--cycle;
}
}
\draw[ultra thick] (-9, 0) -- (9, 0);
\draw[pattern = north east lines](-9, 12) rectangle ++(1,1) node[align = right, anchor = west, midway] {\ Exchanged halo};
\draw[pattern = north west lines](-9, 11) rectangle ++(1,1) node[align = right, anchor = west, midway] {\ Ghost zone};
\draw[] (-9, 10) rectangle ++(1,1) node[align = right, anchor = west, midway] {\ Computational domain};
\draw[<->, fill = white, shift = {(-10, 6)}] (0, 0) -- (0, 3) node[midway, left] {$r$};
\draw[<->, fill = white, shift = {(10, 0)}] (0, 0) -- (0, 6) node[midway, right] {$n_z$};
\draw[<->, fill = white, shift = {(11.5, -6)}] (0, 0) -- (0, 12) node[midway, right] {$N_z$};
\draw[<->, fill = white, shift = {(-10, -9)}] (0, 0) -- +(0, 12) node[midway, left] {$m_z$};
\end{tikzpicture}
\caption{Visualization of the structured grid used in this work. Each cell represents a grid point.
The thick line represents the separation of the computational domain to two devices.}
\label{fig:halo}
\end{figure}
\subsection{Astaroth domain-specific language}
\label{sec:DSL}
The \textit{Astaroth} domain-specific language (DSL) is a stream programming language designed to facilitate the writing of stencil kernels for GPUs. It provides an abstraction level similar to graphics shading languages, such as GLSL~\citep{glsl}. The syntax is an extended subset of C-like languages, providing basic datatypes, operators and tools for control flow, and extending the syntax by adding stream
programming constructs and, for example, function type qualifiers for specifying reusable data. Precision of real numbers is not specified by the DSL, but instead passed as a compilation parameter when building the \textit{Astaroth} executable.
In previous work, we made three assumptions when designing our DSL. Firstly, we assumed that computations are carried out on a structured grid. Secondly, we assumed that each grid point is updated using the same memory access pattern. Finally, we assumed that the result of intermediate operations can be cached and used multiple times when updating a grid point.
This assumption is the most significant one in terms of performance, as it avoids the
repetition of expensive intermediate operations which involve reading from slow off-chip memory.
In our case, for example, the current density can be reused to update velocity and entropy.
\textit{Astaroth} provides a source-to-source compiler for generating efficient CUDA kernels from functions written in the DSL. In addition to optimizations applied during code generation, an automatic optimization is performed at runtime to find the most efficient problem decomposition for the given problem size and hardware.
For this work, we have written the integration kernel used for the MHD simulations solely with the DSL. For further discussion on its syntax and implementation details, we refer the reader to~\cite{Pekkila2019}.
\subsection{Domain decomposition}
In this work, we decompose the computational domain along a single axis before distributing
the subdomains to multiple GPUs. The major benefit of this approach is, that it is simple to
implement, while providing sufficiently efficient scaling within a single node. The main
drawback is, that one-dimensional decomposition is not suitable for large-scale applications,
due to the fact that the size of the
exchanged
halo
decreases at a
much slower rate as a function of the number of devices compared with multi-dimensional
decomposition schemes. The benefits of the latter for high-order stencil codes will be discussed in more
detail in upcoming work.
\subsection{Functions and data dependencies}
In order to hide communication latency, it is critical to carry out computations in parallel with communication. For this, the computational domain must be divided into inner and outer subdomains. The inner computational domain consists of the grid points which can be updated without sampling points in the halo, hence the update can be carried out in parallel with halo exchange. The outer computational domain is formed by the remaining points which can only be updated if the data in the halos is up-to-date. The CUDA API provides concurrency primitives, called streams, which can be used to achieve parallel execution of asynchronous kernels and memory transfers.
We use two buffers for storing the state of the system in order to avoid data races. During integration, we read the data from an input buffer and store the result in a separate output buffer. The buffers are swapped after each substep.
A single simulation step
comprises the execution of the following functions. Here we use the term \textit{local} to refer to computations or memory operations which do not require
halo exchange.
The term \textit{global} is used for operations depending on non-local data.
\begin{itemize}
\item \textbf{Local boundary transfer}.
Update the portion of the halo, which
depends exclusively on
data resident in the local memory system
according to the BCs.
\item \textbf{Local update}. Advance the state of the points in the inner computational domain in time.
\item \textbf{Halo exchange}. Exchange a portion of the halos between neighboring devices.
\item \textbf{Global update}.
Advance the state of the points in the outer computational domain in time.
\item \textbf{Buffer swap}. Swap the input and output buffers in preparation for the next substep.
\item \textbf{Barrier synchronization}. Synchronize the execution state of all devices.
\end{itemize}
The dependencies between these functions and their execution order are visualized in Figure~\ref{fig:concurrency}. Performance enhancements provided by this implementation are listed in Appendix~\ref{sec:performance}.
\tikzstyle{process} = [rectangle, minimum width = 3cm, minimum height = 1cm, align = center, draw]
\tikzstyle{arrow} = [draw, -latex']
\begin{figure}
\centering
\begin{tikzpicture}[node distance = 1.5cm]
\node (local-bound) [process] {Local\\boundary transfer};
\node (local-upd) [process, below of = local-bound, xshift = -2cm] {Local update};
\node (global-bound) [process, below of = local-bound, xshift = 2cm] {Halo exchange};
\node (global-upd) [process, below of = global-bound] {Global update};
\node (swap-buffers) [process, below of = global-upd, xshift = -2cm] {Buffer swap};
\node (synchronization) [process, below of = swap-buffers, xshift = 0cm] {Barrier\\ synchronization};
\draw [<-, thick] (local-bound) -- ++(0, 1cm);
\draw [->, thick] (local-bound) -| (local-upd);
\draw [->, thick] (local-bound) -| (global-bound);
\draw [->, thick] (global-bound) -- (global-upd);
\draw [<-, thick] (swap-buffers) -| (global-upd);
\draw [<-, thick] (swap-buffers) -| (local-upd);
\draw [->, thick] (swap-buffers) -- ++(0, -1cm);
\draw [->, thick] (synchronization) -- ++(0, -1cm);
\end{tikzpicture}
\caption{
Flowchart of the functions executed during a single integration
substep. Dependencies are indicated with arrows.}
\label{fig:concurrency}
\end{figure}
\iffalse
\begin{figure}
\begin{algorithmic}[1]
\Procedure{IntegrateStepP2P}{node, dt}
\For {$\text{isubstep} = 0 \to 3$}
\State {Perform local boundary transfer}
\State {Synchronize all devices on the node}
\\
\State {Compute local update on stream $0$}
\State {Perform halo exchange on stream $1$}
\State {Synchronize stream $1$ on node}
\\
\State {Compute global update on stream $1$}
\State {Swap buffers}
\EndFor
\EndProcedure
\end{algorithmic}
\caption{Single-node multi-GPU implementation using peer-to-peer communication.}
\label{fig:pseudocode}
\end{figure}
\fi
\section{Magnetohydrodynamical model}\label{sec:physics}
We used the continuity, momentum and induction equations of isothermal resistive MHD, corresponding with \citet{Brandenburg2001}:
\begin{equation}\label{eq:continuity}
\frac{D \ln\rho}{D t} = - \nabla \cdot \mathbf{u}
\end{equation}
\begin{equation}\label{eq:navierstokes}
\begin{split}
\frac{D \mathbf{u}}{D t} = & -c_s^2 \nabla \ln\rho + \frac{\mathbf{j} \times \mathbf{B}}{\rho} \\
& +\nu \bigg[ \nabla^2\mathbf{u}
+ \frac{1}{3}\nabla(\nabla \cdot \mathbf{u})
+ 2 \mathbf{S} \cdot \nabla \ln \rho \bigg] + \mathbf{f}
\end{split}
\end{equation}
\begin{equation}\label{eq:magnetic}
\frac{\partial \mathbf{A}}{\partial t} = \mathbf{u} \times \mathbf{B} + \eta \nabla^2 \mathbf{A},
\end{equation}
where, $\rho$ is density, $\mathbf{u}$ is velocity,
$\mathbf{A}$ is the magnetic vector potential with
$\mathbf{B} = \nabla\times\mathbf{A}$ being the magnetic field, $\mathbf{j} = \nabla\times\mathbf{B}/\mu_0$ is the current density, $\mathbf{S}$ is the traceless rate-of-strain tensor and $\mathbf{f}$ is an external forcing. Of constants, $c_s$ is the isothermal speed of sound, $\mu_0$ the magnetic vacuum permeability, $\nu$ the kinematic viscosity and $\eta$ the ohmic diffusivity.
Note the use of the {\it diffusive gauge} in \eq{eq:magnetic} for enhanced numerical stability.
In the continuity \eq{eq:continuity} we have used the high-order upwinding method of \citet{Dobler2006} to enhance numerical stability.
We included a forcing function similar to the one in Pencil Code \citep{Pencil2020}
to generate turbulence. It can be described as,
\begin{equation}\label{eq:forcing}
\mathbf{f}(\mathbf{x},t) = \mathrm{Re}\bigg\{ N \mathbf{f}_{\mathbf{k}(t)} \exp{\big[i \mathbf{k}(t)\cdot \mathbf{x} + i \phi(t)}\big] \bigg\}.
\end{equation}
Here $\textbf{k}(t) = (k_x, k_y, k_z)$ is a wave vector that changes randomly in each time step, $\textbf{x}$ is a position on the grid and $\phi(t)$ is a random phase in range $[-\pi, \pi]$. The normalization factor is set as
\begin{equation}\label{eq:fnorm}
N = f_0 c_s \sqrt{\frac{k c_s}{\delta t}},
\end{equation}
where $k = |\mathbf{k}|$ and $f_0$ is a scaling factor. For each given time step we randomly generate vectors where $4.5\le|k|\le 5.5$, such that $k_x$, $k_y$ and $ k_z$ are integers.
We determine eigenfunctions of the curl operator as
\begin{equation}\label{eq:feigen}
\mathbf{f}_\mathbf{k} = \frac{\mathbf{k} \times (\mathbf{k} \times \mathbf{\hat{e}}) - i \sigma |\mathbf{k}|(\mathbf{k} \times \mathbf{\hat{e}})}{\sqrt{1 + \sigma^2} \mathbf{k}^2 \sqrt{1 - (\mathbf{k}\cdot\mathbf{\hat{e}})^2/\mathbf{k}^2}},
\end{equation}
where $\mathbf{\hat{e}}$ is a
random unit vector perpendicular to $\mathbf{k}$.
The forcing function is almost identical to the one presented in \citet{Brandenburg2001}. However, the factor $\sigma \in [0, 1]$ is included to control the degree of helicity, so that with $\sigma = 1$ we get $\sqrt{2}$ instead of $2$ of \citet{Brandenburg2001} in the denominator of \eq{eq:feigen}. Therefore, to match the normalizations, we have set $f_0 = 0.08$ instead of $f_0 = 0.1$ in our models.
For the numerical domain we have adopted the size $L_{x,y,z} = 2\pi$ so that the smallest wave number in the domain is $k_1 = 1$, hence the unit length of $\mathbf{x}$ was set to unity.
We set the unit of velocity $\mathbf{u}$ to be $c_s = 1$ ,
and the unit of density $\rho$ to $\rho(\mathbf{x},0)=\rho_0 = 1$, where
$\rho_0$ is the uniform initial density. For the magnetic field $\mathbf{B} $
we choose a unit system in which $\mu_0=1$ and hence set its unit as
$\sqrt{\mu_0\rho_0} c_s = 1$. Therefore the units are equivalent to the ones in
\citet{Brandenburg2001}.
For describing the results, the nondimensional kinetic and magnetic Reynolds numbers,
\begin{equation}\label{eq:reynolds}
\mathrm{Re} = \frac{u_\mathrm{rms}}{\nu k_\mathrm{f}}
\quad \mathrm{and} \quad
\mathrm{Re}_\mathrm{M} = \frac{u_\mathrm{rms}}{\eta k_\mathrm{f}}
\end{equation}
respectively, are useful.
Here $k_\mathrm{f}$ is the average wave number of the forcing function and $u_\mathrm{rms}$ is the root mean square of the velocity. For the Reynolds numbers presented, we have used measures of $u_\mathrm{rms}$ at the growth stage of dynamo, hereafter $u_\mathrm{rms,0}$. In most of our simulations, $k_\mathrm{f} = 5$ while a few have $\kf=15$, and we keep the magnetic Prandtl number $\mathrm{Pr}_\mathrm{M}=\nu/\eta$ at unity.
Tables \ref{tab:param_LSD} and \ref{tab:param_SSD} list our helical and non-helical turbulence simulation setups. We run our forcing function with both full helicity ($\sigma = 1$) and without helicity ($\sigma = 0$). This allows us to compare effects which are either due to LSD or SSD. We started our simulations with uniform density, zero velocity and a weak Gaussian random magnetic field, which was $\delta$-correlated in space, as a seed field. We also run our models with multiple resolutions to monitor convergence.
Apart from the forcing helicity, the other physical parameters varied were $\nu=\eta$. At the high end, they correspond to the ones featured in \citet{Brandenburg2001}. At the low end however, the limit for $\nu$ and $\eta$ and therefore $\mathrm{Re}_\mathrm{M}$ is set by the maximum available resolution, limited by the available total GPU memory in a computational node. A single node with 4 Tesla P100 devices was able to support $512^3$ grid resolution at maximum. A significant part of the analysis was dependent on global averages calculated during runtime, and the reduction method required a computational domain resolution with $2^n$ grid points.
The version of \textit{Astaroth} used for this work did not yet provide support for multiple nodes, which would have overcome this issue.
The runs with insufficient resolution tended to crash very early. However, those that kept stable after beginning would keep stable until the end. We avoided using $512^3$ resolution unless necessary to save hard drive space in the computing cluster and for avoiding unnecessary post-processing time. In addition we set $t_\mathrm{max} = 2000$ instead of $4000$ for $512^3$, which might affect our estimates related to the saturation stage. In addition, we run a smaller set of runs to $t_\mathrm{max} = 600$ and a short snapshot interval to perform some more advanced analysis at the exponential growth stage.
During the growth stage, we fit an exponential function to estimate the growth rate $\lambda$ or $\lambda_k$ of our models
\begin{equation}
B_\mathrm{rms} \propto \exp{(\lambda t)}
\quad \mathrm{or} \quad
E_{B, k} \propto \exp{(2\lambda_k t)},
\end{equation}
depending on whether we estimate the growth rate of a global average (like the rms value) or of a spectral channel
of $\mathbf{B}$, $E_{B, k}= B_k^2/2\mu_0$, where $k$ is the wavenumber of the channel. The factor of 2 is required for both $\lambda$ or $\lambda_k$ to agree.
Error estimates are based on Equation (10) of \citet{Morr2014}.
For $\lambda$ we set the time ranges for the fitting by hand, whereas for $\lambda_k$ we find them by fitting in multiple ranges and picking the range with smallest error.
Effectively the fitting errors are negligible, but $\lambda_k$ can have some uncertainties due to automatically picked time ranges.
In Section \ref{sec:soca} we show results for $\lambda$ based on estimated mean-field turbulent transfer coefficients.
\begin{deluxetable}{rccccc}
\tablecaption{Simulation setups and their properties, with
helical forcing ($\sigma = 1$) \label{tab:param_LSD}}
\tablewidth{0pt}
\tablehead{
$N$ & $\eta$ & $\lambda$ & $\mathrm{Re}_\mathrm{M}$ & $\alpha$ & $\eta_\mathrm{t}$
}
\startdata
64 & 5.00e-03 & 1.73e-02 & 3.60e+00 & 2.29e-02 & 4.68e-03 \\
128 & 5.00e-03 & 1.85e-02 & 3.61e+00 & 2.28e-02 & 4.66e-03 \\
256 & 5.00e-03 & 1.79e-02 & 3.60e+00 & 2.27e-02 & 4.65e-03 \\
64 & 3.00e-03 & 2.82e-02 & 7.74e+00 & 2.57e-02 & 5.24e-03 \\
128 & 3.00e-03 & 2.90e-02 & 7.70e+00 & 2.59e-02 & 5.29e-03 \\
256 & 3.00e-03 & 2.85e-02 & 7.69e+00 & 2.58e-02 & 5.26e-03 \\
64 & 2.00e-03 & 3.11e-02 & 1.37e+01 & 2.83e-02 & 5.73e-03 \\
128 & 2.00e-03 & 3.08e-02 & 1.37e+01 & 2.84e-02 & 5.74e-03 \\
256 & 2.00e-03 & 3.12e-02 & 1.36e+01 & 2.84e-02 & 5.75e-03 \\
64 & 1.50e-03 & 2.78e-02 & 1.93e+01 & 3.00e-02 & 6.07e-03 \\
128 & 1.50e-03 & 2.90e-02 & 1.94e+01 & 3.01e-02 & 6.48e-03 \\
256 & 1.50e-03 & 2.90e-02 & 1.95e+01 & 3.05e-02 & 6.11e-03 \\
64 & 1.00e-03 & 2.96e-02 & 3.14e+01 & 3.21e-02 & 6.50e-03 \\
128 & 1.00e-03 & 2.95e-02 & 3.16e+01 & 3.31e-02 & 6.57e-03 \\
256 & 1.00e-03 & 3.09e-02 & 3.10e+01 & 3.31e-02 & 6.58e-03 \\
128 & 7.50e-04 & 3.04e-02 & 4.21e+01 & 3.46e-02 & 6.85e-03 \\
256 & 7.50e-04 & 3.20e-02 & 4.23e+01 & 3.47e-02 & 6.84e-03 \\
128 & 5.00e-04 & 3.17e-02 & 6.56e+01 & 3.63e-02 & 7.19e-03 \\
256 & 5.00e-04 & 3.24e-02 & 6.56e+01 & 3.67e-02 & 7.27e-03 \\
256 & 4.00e-04 & 3.46e-02 & 8.15e+01 & 3.76e-02 & 7.43e-03 \\
256 & 2.50e-04 & 3.54e-02 & 1.31e+02 & 3.80e-02 & 7.60e-03 \\
512 & 2.50e-04 & 3.55e-02 & 1.30e+02 & 3.67e-02 & 7.51e-03 \\
512 & 2.00e-04 & 3.76e-02 & 1.62e+02 & 3.70e-02 & 7.55e-03 \\
512 & 1.50e-04 & 4.09e-02 & 2.20e+02 & 3.69e-02 & 7.63e-03 \\
512 & 1.25e-04 & 4.40e-02 & 2.66e+02 & 3.62e-02 & 7.53e-03 \\
\enddata
\end{deluxetable}
\begin{deluxetable}{rccc}
\tablecaption{Simulation setups and their properties, with non-helical forcing ($\sigma = 0$). \label{tab:param_SSD}}
\tablewidth{0pt}
\tablehead{
$N$ & $\eta$ & $\lambda$ & $\mathrm{Re}_\mathrm{M}$
}
\startdata
64 & 5.00e-03 & -1.35e-02 & 3.53e+00 \\
128 & 5.00e-03 & -1.37e-02 & 3.52e+00 \\
256 & 5.00e-03 & -1.43e-02 & 3.52e+00 \\
64 & 3.00e-03 & -1.65e-02 & 7.16e+00 \\
128 & 3.00e-03 & -1.56e-02 & 7.14e+00 \\
256 & 3.00e-03 & -1.62e-02 & 7.16e+00 \\
64 & 2.00e-03 & -1.22e-02 & 1.21e+01 \\
128 & 2.00e-03 & -1.15e-02 & 1.21e+01 \\
256 & 2.00e-03 & -1.09e-02 & 1.21e+01 \\
64 & 1.50e-03 & -6.70e-03 & 1.73e+01 \\
128 & 1.50e-03 & -7.33e-03 & 1.73e+01 \\
256 & 1.50e-03 & -5.64e-03 & 1.73e+01 \\
64 & 1.00e-03 & 1.60e-03 & 2.80e+01 \\
128 & 1.00e-03 & 1.64e-03 & 2.79e+01 \\
256 & 1.00e-03 & 2.04e-03 & 2.79e+01 \\
128 & 7.50e-04 & 7.38e-03 & 3.87e+01 \\
256 & 7.50e-04 & 7.50e-03 & 3.86e+01 \\
128 & 5.00e-04 & 1.51e-02 & 6.05e+01 \\
256 & 5.00e-04 & 1.46e-02 & 6.00e+01 \\
128 & 4.00e-04 & 1.74e-02 & 7.65e+01 \\
256 & 4.00e-04 & 1.96e-02 & 7.62e+01 \\
256 & 2.50e-04 & 2.71e-02 & 1.24e+02 \\
512 & 2.50e-04 & 2.87e-02 & 1.25e+02 \\
256 & 2.00e-04 & 3.14e-02 & 1.56e+02 \\
512 & 2.00e-04 & 3.27e-02 & 1.57e+02 \\
512 & 1.50e-04 & 3.92e-02 & 2.10e+02 \\
512 & 1.25e-04 & 4.30e-02 & 2.53e+02 \\
\enddata
\end{deluxetable}
\begin{figure*}[htb!]
\plotone{relhel1_eta1em30040.png}
\plotone{relhel0_eta1em30040.png}
\caption{Snapshots of dynamo fields in helically (top) and non-helically
(bottom) driven system with $\eta = \nu = 1\dee{-3}$ and $N=256$. Colours
represent $B_z$, normalized with the equipartition magnetic field
$B_\mathrm{eq} = \sqrt{\mu_0\rho_0} u_\mathrm{rms,0}$. (Animated figures
display growth of the magnetic field from the initial seed field.
During the early phases of evolution, magnetic fluctuations grow quickly. In
the helical case (top) dominant large-scale structure with $k/k_1=1$ grows more
slowly, but eventually dominates, whereas non-helical case (bottom)
fluctuations grow without large-scale structure.) \ \label{fig:3deta1em3}}
\end{figure*}
\begin{figure*}[htb!]
\plotone{relhel1_eta25em40040.png}
\plotone{relhel0_eta25em40040.png}
\caption{As Figure \ref{fig:3deta1em3}, but with $\eta = \nu = 2.5\dee{-4}$. (Animations behave as in Figure \ref{fig:3deta1em3} but turbulent
fluctuations cascade into smaller scales.) \ \label{fig:3deta25em4}}
\end{figure*}
\section{Results} \label{sec:results}
The time development of the runs with helical forcing (see Table \ref{tab:param_LSD}; Figures \ref{fig:3deta1em3}, \ref{fig:3deta25em4} and \ref{fig:timeseries}) exhibits three stages: initial decay, exponential growth, and possible slow growth leading to saturation.
The initial transient growth of the velocity field, and the contemporary decay of the initial magnetic field, are short for all runs. It is followed by saturation of the rms velocity and exponential growth of the magnetic field. The growing magnetic field starts eventually to quench the velocity, when their energy densities become comparable. The rms velocities then saturate at lower levels, the quenching being the strongest and taking place most slowly the lower $\mathrm{Re}_\mathrm{M}$ is.
The rms velocities used for calculated $\mathrm{Re}_\mathrm{M}$ have been measured from the saturated values before quenching starts. Magnetic field values presented in the figures have been normalized with the equipartition magnetic field $B_\mathrm{eq} = \sqrt{\mu_0\rho_0} u_\mathrm{rms,0}$.
For helical forcing, the dynamo will always exhibit a large-scale ($k=1$) magnetic field.
If there is no simultaneous SSD operating, this field is
well visible during exponential growth and saturation, otherwise it is
fully emerging only during the saturated stage, with weak signatures during exponential growth.
Typical field geometries are shown in Figures \ref{fig:3deta1em3} and \ref{fig:3deta25em4} for low and high $\mathrm{Re}_\mathrm{M}$, respectively. The presence of LSD
is
expected because in the helically turbulent regime,
the critical dynamo number is close to unity. The emerging large-scale field is of the form of a Beltrami field,
\begin{equation}
\mathbf{B}\big(\mathbf{x}) = (B_x \sin (k z + \phi), B_y \cos( k z + \phi), 0\big),
\end{equation}
for alignment along $z$, analogously for $x$ and $y$;
$\phi$ is an arbitrary phase.
Alignment and phase are unpredictable as due to the non--linear nature of the MHD system, even tiny differences in initial conditions or round--off errors may lead to different
orientation of the Beltrami field being realized in the simulation.
\begin{figure*}[ht!]
\plotone{helical_paperubtotrms.pdf}
\caption{Behaviour of different helical setups as a function of time. $u_\mathrm{rms,0}$ is the time average of $u_\mathrm{rms}$ during the expotential growth stage.
Thick dotted lines within the insets display the exponential fits.
\label{fig:timeseries}}
\end{figure*}
\begin{figure*}[ht!]
\plotone{nohel_paperubtotrms.pdf}
\caption{As Figure \ref{fig:timeseries}, but for non-helical forcing. \ \label{fig:timeseries_nohel}}
\end{figure*}
For non-helical forcing, magnetic field growth is not seen for all of our $\mathrm{Re}_\mathrm{M}$ values, but only above a critical value $\mathrm{Re}_\mathrm{M, crit} \sim 25$.
Below it, the magnetic field decays exponentially. In the case of non-helical turbulence, the
overall structure and geometry of the magnetic field at the saturation stage retains similar form to the growth stage,
but the magnetic field strength no longer increases (such as Figure \ref{fig:3deta25em4}, bottom).
After the exponential growth, if an LSD is present, there can be still gradual growth of the magnetic field until full saturation is reached.
At low $\mathrm{Re}_\mathrm{M}$ where our diffusivity parameters are within the same range as theirs, our results generally agree qualitatively with \citet{Brandenburg2001} with the emergence of $k = 1$ Beltrami field,
however we do get generally weaker growth rates than they do for unidentifiable reason, with theirs being $\sim 2-3 \times$ larger with the points having comparable magnetic diffusivity \citep[][Run 2 and Run 3]{Brandenburg2001}.
-- and at higher $\mathrm{Re}_\mathrm{M}$ similar principles apply, with the large-scale magnetic field forming, despite increased randomness at smaller scales.
However, despite the chaotic nature of the system, the resolution does not appear to make a significant difference. As long as $\mathrm{Re}_\mathrm{M}$ is not too high for the given resolution, practically identical results are produced.
Therefore, for the figures we have chosen representative samples from the highest resolution runs. Agreement across resolutions also indicates that any effects caused by the numerical grid are not significant.
Growth and self-organization of the magnetic field can be seen at different resolutions in the animated Figures \ref{fig:3deta1em3} and \ref{fig:3deta25em4}. With increasing $\mathrm{Re}_\mathrm{M}$ the small-scale substructures accompanying the coherent mean-field tend to get finer. As we further discuss in Section \ref{sec:spectra}, this could be an indication that an SSD is acting in parallel with the LSD.
\subsection{Growth rates} \label{sec:growth}
Figures \ref{fig:timeseries} and \ref{fig:timeseries_nohel} show the early exponential growth stage
as well as saturation, cf.
\citet{Brandenburg2001} (their Fig. 1) for helical turbulence and \citet{Haugen2004} (their Fig. 6) for non-helical turbulence.
Our estimated growth rates are shown in Figures \ref{fig:growthrate} and \ref{fig:growthrate_nohel}.
Those of the helically forced simulations with LSD are positive in all cases. Mildly higher values appear for $\mathrm{Re}_\mathrm{M}=7.7,13.7$,
but otherwise the normalized growth rate curve appears flat until $\mathrm{Re}_\mathrm{M} \approx 100$.
For the two highest values of $\mathrm{Re}_\mathrm{M}$, however, the growth rates become similar to those of the non-helical cases. This is indicative of simultaneous SSD action.
Our results
agree reasonably well with \citet[Figure 3]{Brandenburg2009}, where they display growth rates at various $\mathrm{Re}_\mathrm{M}$ for helical forcing, combined with data from \citet{Haugen2004} for the non-helical forcing. Their helical growth rate curve is also flat at low $\mathrm{Re}_\mathrm{M}\le 70$, but approaches the non-helical growth rates at $\mathrm{Re}_\mathrm{M}=670$. The helical and non-helical growth rates shown in \citet{Brandenburg2009} align with comparable numerical range to ours.
However, the comparison can be problematic for two reasons. First, their results are more limited with the respect to the number of data points at with 3 points for helical and 4 for non-helical turbulence.
Therefore they did not truly resolve the shape of the curve.
Second, $\mathrm{Re}_\mathrm{M}$ is varied but $\mathrm{Re}$ is kept the same leading to variable $\mathrm{Pr}_\mathrm{M}$, which means that our setups are not completely equal type.
\citet{Brandenburg2009} suspected that there is a point where a system would switch from exhibiting merely a LSD to a dynamo combined of LSD and SSD. Our results clearly support this interpretation
as can been seen when comparing Figures \ref{fig:growthrate} and \ref{fig:growthrate_nohel} (see also the power spectra based growth rates in Figure \ref{fig:growthrate_channel} and magnetic field distributions in Figure \ref{fig:BPDF}). It should be noted that if the basic phenomenon observed by \citet{Brandenburg2009} is the same as we observe, this could imply that it is more dependent on $\mathrm{Re}_\mathrm{M}$ than $\mathrm{Re}$. However, our results cannot substantiate this claim, as we only examine $\mathrm{Pr}_\mathrm{M} = 1 $ regime.
\begin{figure}[ht!]
\plotone{lsd_growth.pdf}
\caption{Growth rate as a function of magnetic Reynolds number for helical forcing. \label{fig:growthrate}}
\end{figure}
\begin{figure}[ht!]
\plotone{ssd_growth.pdf}
\caption{Growth rate as a function of magnetic Reynolds number for non-helical forcing. \label{fig:growthrate_nohel} }
\end{figure}
There are two analytical predictions for the SSD growth rate: the more common $\sqrt{\mathrm{Re}_\mathrm{M}}$ scaling \citep[see e.g.][]{Haugen2004} in contrast to the logarithmic scaling $\ln (\mathrm{Re}_\mathrm{M}/\mathrm{Re}_\mathrm{M, crit})$, where $\mathrm{Re}_\mathrm{M, crit}$ is the critical Reynolds number for the SSD \citep{Kleeorin2012}, and validity is restricted to low magnetic Prandtl numbers and $\mathrm{Re}_\mathrm{M} \approx \mathrm{Re}_\mathrm{M, crit}$.
In Fig. \ref{fig:growthrate_nohel}, the SSD growth rates appear to be highly consistent with the logarithmic scaling, except at the lowest $\mathrm{Re}_\mathrm{M}$.
In contrast, the $\sqrt{\mathrm{Re}_\mathrm{M}}$ scaling does not really apply, apart from high $\mathrm{Re}_\mathrm{M}$.
From Figure \ref{fig:growthrate_nohel}, $\mathrm{Re}_\mathrm{M, crit}\sim 25$ has been estimated, while \citet[Fig. 1]{Haugen2004} provide $\sim 35$.
However, given the uncertainty due to their low number of $\mathrm{Re}_\mathrm{M}$ data points,
the estimates might not be significantly different. Figure 2 of \citet{Iskakov2007} shows that for
incompressible turbulence with $\mathrm{Pr}_\mathrm{M}=1$,
$\mathrm{Re}_\mathrm{M, crit} \approx 60$ based on $k_1 = 2\pi$,
or $\approx 42$ if their $\mathrm{Re}_\mathrm{M}$ is scaled with $\kf/k_1 = \sqrt{2}k_1/k_1 = \sqrt{2}$
instead, as noted by \citet{Brandenburg2018}.
\subsection{Saturation}
\label{sec:saturation}
Our main focus is on examining the
kinematic growth stage of the SSD, with (helical forcing) and without (non-helical forcing) a co-existing LSD.
Some conclusions about the saturated stage can also be drawn, but
for its complete study, many of the helical runs would need to be continued
longer, as the saturation of the LSD is known to occur on a resistive time scale only \citep[e.g.][]{Brandenburg2001}.
Unfortunately, for the highest
$\mathrm{Re}_\mathrm{M}$, our datasets are not long enough to determine
their saturation field strength.
To compensate this, we performed a prolonged run with $\nu = \eta = 1.5\dee{-4}$, $\mathrm{Re}_\mathrm{M} = 210$,
extending to one diffusion time based on the forcing scale. This was the highest $\mathrm{Re}_\mathrm{M}$ allowing numerical stability towards saturation. However, otherwise we have to restrict our analysis on low and intermediate $\mathrm{Re}_\mathrm{M}$.
The time development of the large-scale fields
can be fitted well with a function
$B_\mathrm{sat}\tanh{(t/d_0+d_1)}$, where $d_0$ and $d_1$ are
fitting parameters,
and we use it to
determine the saturation magnetic field $B_\mathrm{sat}$.
Restricting to the helical runs with intermediate $\mathrm{Re}_\mathrm{M}$,
we find that $B_\mathrm{sat}/B_\mathrm{eq}$
increases roughly
logarithmically as a function of $\mathrm{Re}_\mathrm{M}$ (See Figure \ref{fig:saturation}).
The saturation values for the non-helical cases are easier to
determine as they saturate quickly after their exponential growth, and compared to the helical cases, the field strengths are roughly by a factor of six smaller.
Their $B_\mathrm{sat}/B_\mathrm{eq}$ grows also in with $\mathrm{Re}_\mathrm{M}$.
For helical forcing, in both the cases with only an LSD ($\mathrm{Re}_\mathrm{M}<40$) and a combined LSD-SSD, $B_\mathrm{sat}/B_\mathrm{eq}$ as a function of $\mathrm{Re}_\mathrm{M}$ obeys a logarithmic law as shown in Figure \ref{fig:saturation}, albeit with different slopes. For combined LSD-SSD, the slope is roughly the same as in
the non-helical cases. Therefore we hypothezise that the emergence of the SSD is to the disadvantage of the LSD, most likely due to a reduction of $\alpha$, and thus prevents its saturation strength from further growing with $\mathrm{Re}_\mathrm{M}$.
Consequently, the observed growth of $B_\mathrm{sat}/B_\mathrm{eq}$ with $\mathrm{Re}_\mathrm{M}$ would be exclusively due to the SSD.
\begin{figure}[htb!]
\plotone{b_saturation.pdf}
\caption{Estimated saturation magnetic field $B_\mathrm{sat}$ normalized to
$B_\mathrm{eq}$. The solid lines represent logarithmic fits. We have included
those datasets with $256^3$ and $512^3$ resolution from which a valid estimate
could be obtained. \label{fig:saturation} }
\end{figure}
\subsection{Power spectra}\label{sec:spectra}
One substantial difference between SSD and LSD consists in the scale distribution of the magnetic energy during growth and saturation. To investigate it, we have calculated magnetic power spectra $E_B(k)$ (with normalization $\int E_B(k) dk = \int \mathbf{B}^2 dV/2\\mu_0$
for individual simulation snapshots.
\begin{figure}[htb!]
\plotone{Btot_powerspectra_sigma1_512_eta125e-04.pdf}
\plotone{Btot_powerspectra_sigma0_512_eta125e-04.pdf}
\caption{Time-dependent powerspectra of magnetic energy, $E_B(k)$, for $\eta = 1.25\dee{-4}$ with resolution $512^3$ for helical (top) and non-helical (bottom) forcing
and $0\le t \le600$.
Both show Kazantsev scaling $\propto k^{3/2}$ during growth (red dashed). Time difference between spectra $\Delta t=30$.
For $t=0$, the spectrum reflects the initial random (Gaussian) field, hence $\propto k^2$.
The dashed curves correspond to $t=0$.
\label{fig:pspec_a}
}
\end{figure}
\begin{figure}[thb!]
\plotone{Btot_powerspectra_sigma1_256_eta200e-03.pdf}
\caption{Time-dependent powerspectra of magnetic energy, $E_B(k)$, for $\eta = 2\dee{-3}$ with resolution $256^3$ for helical forcing for $0\le t \le600$. Time difference between spectra $\Delta t=30$.
\label{fig:pspec_flat}
}
\end{figure}
Figures \ref{fig:pspec_a} and \ref{fig:pspec_flat} show time-dependent spectra $E_B(k;t)$ for the highest studied $\mathrm{Re}_\mathrm{M}$ and a low one, respectively.
In the case of helical forcing (LSD) with high $\mathrm{Re}_\mathrm{M}>\mathrm{Re}_\mathrm{M, crit}$ (Figure \ref{fig:pspec_a}, top), we witness a persistent peak at the forcing wavenumber $\kf$ and at late times gradually growing power for $k\gtrsim 1$, i.e. in the large-scale field. At late times, there is a typical forward energy cascade towards small scales $k>\kf$,.
For SSD, (Figure \ref{fig:pspec_a}, bottom), large scales follow persistently the Kazantsev scaling $\propto k^{3/2}$ \citep{Kazantsev1968}, while the spectrum peaks above $\kf$ at $k = 9$; beyond that a similar forward cascade exists as in LSD.
Qualitatively, this picture is the same for all $\mathrm{Re}_\mathrm{M}>\mathrm{Re}_\mathrm{M, crit}$.
For high $\mathrm{Re}_\mathrm{M}$, as shown in Figure \ref{fig:pspec_a}, even with helical forcing we can clearly see SSD-type (Kazantsev) spectra during the growth stage, which signifies the presence of SSD during the growth. The SSD-type spectra can also appear during the growth of medium $\mathrm{Re}_\mathrm{M}$ cases, which are not too diffusive.
However, at low $\mathrm{Re}_\mathrm{M}$, SSD and LSD spectral shapes are clearly different during this stage:
In helically driven systems, the spectrum is essentially flat at large scales until the emergence of the $k=1$ mode, see Figure \ref{fig:pspec_flat}.
This resembles the spectral growth of \citet{Brandenburg2001} with a flattened curve during the growth stage and the eventual emergence of $k=1$ mode (Their Figures 2 ans 3 respectively).
\begin{figure}[h!]
\plotone{spect_sigma1growth.pdf}
\plotone{spect_sigma0growth.pdf}
\plotone{spect_growth.pdf}
\caption{Growth rates $\lambda_k$ of the power spectra channels $k=1$ and $k=9$ for helical (top) and non-helical (middle) forcing. Bottom: combination of top and middle panel. \label{fig:growthrate_channel} }
\end{figure}
As helical and non-helical growth rates converge at high $\mathrm{Re}_\mathrm{M}$ (see Sec. \ref{sec:growth} and Figures \ref{fig:growthrate} and \ref{fig:growthrate_nohel}),
we may explain this behavior by the assumption that in this range the SSD growth rate is higher than the LSD one, thus the former is dominating the latter during growth. But if SSD saturates earlier than LSD and also at lower magnitude, the spectrum has to undergo the observed change in its shape with finally dominating large scales.
For medium $\mathrm{Re}_\mathrm{M}$ the situation would be less clear.
The Kazantsev scaling is kept during most of the growth stage,
but the growth of the magnetic energy is more rapid than for the respective non-helical runs.
Therefore helical turbulence can enhance the accumulation of magnetic energy without affecting the spectral shape during the initial growth.
In an attempt to separate SSD and LSD behaviour, we have estimated the growth
rates as functions of wavenumber. The resulting $\lambda_k$, displayed in
Figure \ref{fig:growthrate_channel} for $k=1$ and $k=9$, are
in general proportional to those estimated from $B_{\mathrm{rms}}$.
Again, the growth rates of SSD and LSD merge at high $\mathrm{Re}_\mathrm{M}$ and the SSD ones follow the logarithmic scaling.
For both helical and non-helical forcing, both of the referred scales ($k=1$ and $k=9$)
grow at similar rates,
with a mild tendency of
$\lambda_1$ being marginally higher than $\lambda_9$. In addition, the helical runs seem to show higher growth rates at high $\mathrm{Re}_\mathrm{M}$ than the nonhelical runs.
Yet, because of fitting uncertainties, all this should be taken with caution.
Figure \ref{fig:psts_a} displays the corresponding time development of the spectral channels $k=1,5,9$ for the high $\mathrm{Re}_\mathrm{M}$ helical and non-helical runs featured in Fig. \ref{fig:pspec_a}. In both cases, the exponential growth occurs at similar rates for all three channels
with the $k=5,9$ ones dominating. The only difference is that when LSD is present, the $k=1$ channel continues to grow past the exponential growth stage,
while the other channels are almost saturated, to become finally dominating.
At low $\mathrm{Re}_\mathrm{M}$, with LSD alone (Figure \ref{fig:psts_flat}), all three channels have approximately equal magnitude with minor deviations in the beginning. After the exponential growth phase the spectral channels differ with $k=1$ becoming the strongest,
as it happens in the spectra of Figure \ref{fig:pspec_flat}.
\begin{figure}[tb!]
\plotone{Btot_pstimeline_sigma1_512_eta125e-04.pdf}
\plotone{Btot_pstimeline_sigma0_512_eta125e-04.pdf}
\caption{Time development of selected spectral channels for $\eta = 1.25\dee{-4}$ with resolution $512^3$ for helical (top) and non-helical (bottom) forcing. \label{fig:psts_a}
}
\end{figure}
\begin{figure}[htb!]
\plotone{Btot_pstimeline_sigma1_256_eta200e-03.pdf}
\caption{Time development of selected spectral channels for $\eta = 2\dee{-3}$ with resolution $256^3$ for helical forcing. \label{fig:psts_flat}
}
\end{figure}
\begin{figure}[htb!]
\plotone{k15Btot_powerspectra_sigma1_256_eta250e-04.pdf}
\caption{Time-dependent power spectra for $\eta = 2.5\dee{-4}$ with resolution $256^3$ for helical forcing with $\kf = 15$. Time difference between spectra $\Delta t=100$. \label{fig:pspec_k15}
}
\end{figure}
For the sake of testing and comparison, we changed the forcing scale to $\kf= 15$ to separate it more safely from the largest scale of the emergent mean field. During growth we see a more gradual buildup of the
inverse energy cascade towards large scales, see Figure \ref{fig:pspec_k15}:
the energy peak of the growing large-scale magnetic field moves gradually towards larger scales until reaching its largest values at $k = 1$. This is similar to the same phenomena visible in Figure 7 of \citet{Brandenburg2012review}.
The time development of the spectral channels $k=1,5,9$ is shown in Figure \ref{fig:psts_k15}, indicating a clearly different growth rate of the $k = 1$ channel.
The $k = 1$ channel curve consists of two different exponentials at concurrent stages. However, the data is too sparse in that range to produce meaningful fit estimates.
There is even decay visible
in the $k=5$ channel which is explainable with the magnetic energy inverse-cascading onto larger scale over time as visible in Figure \ref{fig:pspec_k15}.
\begin{figure}[htb!]
\plotone{k15Btot_pstimeline_sigma1_256_eta250e-04.pdf}
\caption{As Figure \ref{fig:pspec_k15}, but time evolution of the spectral channels $k=1,5,9$. \label{fig:psts_k15}
}
\end{figure}
\subsection{Mean-field analysis}\label{sec:soca}
For LSD, mean-field theory provides some testable predictions: First, an estimate of the growth rate based on turbulent transport coefficients. Second, quenching of the $\alpha$-effect as a function of $\mathrm{Re}_\mathrm{M}$.
Unfortunately, at this moment \textit{Astaroth} is not capable of handling the test-field method \citep{Schrinetal07} or other method
for measuring the turbulent transport coefficents
at runtime.
Therefore, we calculated estimates for the coefficients of $\alpha$-effect and turbulent diffusion, $\eta_\mathrm{t}$, using results from the second-order correlation approximation (SOCA) and other closures
which have proven to be surprisingly useful \citep{Sur2008, Vaisala2014}.
According to the mean-field approximation, the growth rate of the mean field $\overline{\mathbf{B}}$ for isotropic stationary turbulence, hence constant $\alpha$ and $\eta_\mathrm{t}$, is
\begin{equation}\label{eq:MFgrowth}
\lambda = |\alpha| k - (\eta_\mathrm{t} + \eta) k^2
\end{equation}
where $k$ is the wavenumber of the mean field and $\alpha = \alpha_K + \alpha_M$. In the limit of ideal MHD, the constituents of $\alpha$ are related to kinetic and current helicity, respectively, by
\begin{equation}\label{eq:alphaFOSA}
\alpha_K = -\frac{1}{3} \tau \left\langle \boldsymbol{\omega} \cdot \mathbf{u} \right\rangle \quad \mathrm{and} \quad
\alpha_M = \frac{1}{3} \tau \langle \mathbf{j}'\cdot \mathbf{b}' \rangle /\mu_0 \rho_0,
\end{equation}
with correlation time $\tau$, and vorticity $\boldsymbol\omega=\nabla\times \mathbf{u}$ and primes indicating the fluctuating parts.
Note two possible interpretations for $\alpha_M$:
First, it reflects the contribution of a magnetic background turbulence, like that provided by an SSD, to $\alpha$. Here, $\alpha_M$ can be obtained already by SOCA \citep{RaeRhei07}. Second, it can be interpreted as reflecting the quenching of $\alpha$ by $\overline{\mathbf{B}}$ such that with its magnitude growing, $\alpha_M$ also grows, but opposite in sign to $\alpha_K$, resulting in a reduced total $\alpha$. This can be obtained via closure approaches like the $\tau$
or eddy-damped quasi-normal Markovian approximations \citep{Pouquet1976}.
Turbulent diffusivity in incompressible flows is estimated as
\begin{equation}\label{eq:etaFOSA}
\eta_\mathrm{t} = \frac{1}{3} \tau \langle {\mathbf{u}}^2 \rangle.
\end{equation}
If the Strouhal number $u_{\mathrm rms}\tau/\ell$, $\ell$ a characteristic scale of the flow, is assumed to be unity, we can estimate $\tau = 1/\kf u_\mathrm{rms}$. To obtain the fluctuating fields $\mathbf{b}'$, $\mathbf{j}'$ we have removed the large scale field via filtering out
the contributions of the $k = 1$ mode from the magnetic field snapshots.
\begin{figure}[th!]
\plotone{soca_growth.pdf}
\caption{Growth rates \eqref{eq:MFgrowth} based on closure estimates of $\alpha$ and $\eta_\mathrm{t}$ at the exponential growth stage, compared to directly estimated ones. \label{fig:socagrowth} }
\end{figure}
We estimated the growth rates by first calculating $\alpha$ and $\eta_\mathrm{t}$ for individual snapshots using data from runs with high snapshot frequency during exponential growth. Then we used Equation (\ref{eq:MFgrowth}) to get $\lambda$ for an individual snapshot, and subsequently time-averaged over the growth phase.
We find results which are at least approximately aligned with the directly measured values
with better agreement at low $\mathrm{Re}_\mathrm{M}$,
see Fig. \ref{fig:socagrowth}. Towards high $\mathrm{Re}_\mathrm{M}$, the results begin to diverge, most strongly at the highest $\mathrm{Re}_\mathrm{M}$.
This appears to indicate that the closure estimates can work surprisingly well, but do not catch all of the details.
But how does $\alpha$ behave at the saturation stage? \citet{Brandenburg2008} measured $\alpha$ and $\eta_\mathrm{t}$ quenching, with both decreasing as functions of $\mathrm{Re}_\mathrm{M}$.
\begin{figure}[ht!]
\plotone{alpha_re.pdf}
\caption{$\alpha_{\rm K}$ and $\alpha_{\rm M}$ at the saturated stage as functions of $\mathrm{Re}_\mathrm{M}$, normalized by $\alpha_0 = u_\mathrm{rms}/3$.
Bars represent the fluctuation level of $\alpha$ derived from their standard deviations.
The jump at $\mathrm{Re}_\mathrm{M} \sim 20$ is due to an unclear disagreement between resolutions $128^3$, $64^3$ and $256^3$, that at $\mathrm{Re}_\mathrm{M} \sim 100$ due to different saturation stage lengths in the runs with resolutions $256^3$ and $512^3$ .
\label{fig:alphaquench} }
\end{figure}
For comparison,
we calculated $\alpha_{\rm K,M}$ from the saturation stages of our runs and discovered similar results, see Figure \ref{fig:alphaquench}. The normalized $\alpha_K$ approaches a constant whereas $\alpha_M$ keeps getting stronger with $\mathrm{Re}_\mathrm{M}$.
Our results are close to \citet{Brandenburg2008} (their Fig. 3) within our range of $\mathrm{Re}_\mathrm{M}$, including substantially more points though. The total $\alpha$ decreases with growing $\mathrm{Re}_\mathrm{M}$ reaching eventually a tentative plateau.
\subsection{Magnitude distribution of the magnetic field}
\begin{figure*}[ht!]
\plotone{BPDF_collection.pdf}
\caption{Probability density functions of $B_{x,y,z}$ for non-helical forcing (first column) and for helical forcing with low, intermediate and high $\mathrm{Re}_\mathrm{M}$ (second to fourth column), as a function of time (rows). \label{fig:BPDF}
$K(B_x)$, $K(B_y)$ and $K(B_z)$ denote the respective kurtoses using Fisher’s definition.}
\end{figure*}
Next we investigate how the probability density functions (PDFs) of the
magnetic field differ in the regimes where SSD or LSD alone and SSD and LSD
together are acting. While spectra gave us information about the distribution
of the magnetic field across different scales, PDFs can further reveal details
of the structure of the field in different spatial directions. Additionally, we
calculated kurtoses of these distributions using the
Fisher’s definition, for which the kurtosis of a normal distribution is 0.
Figure \ref{fig:BPDF} shows the temporal evolution of the PDFs of all three components of the dynamo-generated $\mathbf{B}$ for non-helical forcing and helical forcing with three different $\mathrm{Re}_\mathrm{M}$.
Two basic types of PDFs are obtained:
A SSD produces a symmetric exponential distribution, which expands
over time, but keeps otherwise the same shape
and eventually stops changing at the saturation stage, with its kurtosis ranging from $\sim 4$ to $\sim 6$.
An LSD produces a more Gaussian--shaped distribution.
As a pure Beltrami field has a PDF of top-hat shape,
this can result in a PDF of the total field with a both widened and flattened peak if $\overline{\mathbf{B}}$ is strong enough.
In Figure \ref{fig:BPDF}, such an indication of a top-hat profile is best
visible in the PDFs of $B_y$ at saturation for helically forced cases with
$\eta = 0.002$ and $0.0005$, with their kurtoses approaching 0, whether or not
the top is flattened. This is not surprising because the top hat effect happens
close to the peak and the PDF has still significant tails.
The widened Gaussian PDF profile of LSD appears most pronouncedly at (or just before) the saturation stage, whereas during the exponential growth stage, the LSD cases show at high and intermediate $\mathrm{Re}_\mathrm{M}$ rather an SSD type of PDF, including similar values of the kurtosis.
Even after initial growth, there is a tendency of the LSD-type PDFs to develop a sharp tip, reminiscent of the SSD type, because the emergence of fully saturated large scale Beltrami field takes time.
These observations support the finding of Section \ref{sec:spectra}, that at high $\mathrm{Re}_\mathrm{M}$ exponential growth is seemingly dominated by the SSD, present simultaneously with the LSD.
However, even with high $\mathrm{Re}_\mathrm{M}$, where small-scale fluctuations are strong, the LSD will turn the field profile into a Gaussian type over time. In Figure \ref{fig:BPDF}, the highest $\mathrm{Re}_\mathrm{M}$ simulation is not depicted at its most saturated state, but the $k=1$ mode will keep growing mere over time and its feature will soften.
\section{Discussion}\label{sec:discussion}
We examined emergence and growth of both large and small scale dynamos and found that with helical forcing, approaching high $\mathrm{Re}_\mathrm{M}$, both SSD and LSD become clearly simultaneous phenomena. This is visible both from the time evolution of the powerspectra and from the probability density functions of $\mathbf{B}$. For high $\mathrm{Re}_\mathrm{M}$, the growth rates of helical and non-helical cases converge, indicating the dominance of SSD in the helical ones.
\begin{figure}[ht!]
\plotone{LSDvsSSD.pdf}
\caption{Schematic depiction of our hypothesis for the change of growth rate as a function of $\mathrm{Re}_\mathrm{M}$ in helically forced simulations, cf. Figure \ref{fig:growthrate}. \label{fig:LSDvsSSD} }
\end{figure}
Under helically driven turbulence, an exact delimitation between LSD and SSD is difficult. At medium $\mathrm{Re}_\mathrm{M} > \mathrm{Re}_\mathrm{M, crit}$, there is a tentative indication of an SSD in the powerspectra during the dynamo growth. While the growth rate is dominated by the LSD, an SSD can already be operating, so that SSD and LSD are coupled at this stage.
Based on the dependency of the growth rates of the helical simulations on $\mathrm{Re}_\mathrm{M}$, we attempt a hypothetical explanation of the interaction of the two dynamos, see Figure \ref{fig:LSDvsSSD}: At low $\mathrm{Re}_\mathrm{M}$, the system is diffusive enough to prevent any presence of SSD, and the -- pure LSD -- growth rate increases as a function of $\mathrm{Re}_\mathrm{M}$.
In the mid range of $\mathrm{Re}_\mathrm{M}\gtrsim 10$, the growth rate decreases as a function of $\mathrm{Re}_\mathrm{M}$. We suggest that now SSD becomes effective and starts to inhibit the growth of the LSD by reducing $\alpha$ via, in turn, increasing $\alpha_M$. However, to prove this we had to show that the $\mathrm{Re}_\mathrm{M, crit}$ of helical turbulence is smaller than that of non-helical turbulence and that $\alpha$ is really reduced. While possible, recognizing both of these affects conclusively would require more thorough mapping of $\alpha$ than what we have available.
At high $\mathrm{Re}_\mathrm{M}$, SSD dominates the exponential growth, which is subtantiated by our results.
We can find a point of comparison in \citet{Hotta2016}, who have
reported on the effect of SSD on LSD in the context of
turbulent convection
in stars with a solar--like convective envelope.
They state that SSD shows both inhibiting and enhancing effects to the emerging magnetic field
depending on $\mathrm{Re}_\mathrm{M}$. In their low $\mathrm{Re}_\mathrm{M}$ case, a large-scale magnetic field emerges, but in their medium $\mathrm{Re}_\mathrm{M}$ case, the large-scale magnetic field is suppressed,
while emerging again in their high $\mathrm{Re}_\mathrm{M}$ case.
\citet{Hotta2016} explain their medium $\mathrm{Re}_\mathrm{M}$ case as the suppression of LSD by SSD, whereas in their high $\mathrm{Re}_\mathrm{M}$ case the SSD would enhance the magnetic energy generation by LSD.
Our hypothesis, illustrated by Figure \ref{fig:LSDvsSSD} would be congruent with their scenario. However, caution should be shown because our methods are not completely equivalent to theirs.
We have focused on exponential growth, and we cannot compare the saturated stages with equal detail.
Our limited saturation data at low, medium and high $\mathrm{Re}_\mathrm{M}$ shows that the
saturated magnetic field increases towards high $\mathrm{Re}_\mathrm{M}$, with is a tentative
indication that SSD would exclusively contribute to the the increase of
saturated field strength, but with SSD being a disadvantage of the LSD itself.
Another caveat with respect to the \citet{Hotta2016} results is that their
model diffusivities are implicit and their diffusion schemes changes between
different simulation while ours are explicit and have an uniform scheme, which
makes a direct comparison difficult.
When it comes to the pure SSD, our results show that the $\mathrm{Re}_\mathrm{M}$ dependence of the growth rate matches the logarithmic prediction of \citet{Kleeorin2012} very well. In contrast, the $\propto \sqrt{\mathrm{Re}_\mathrm{M}}$ relation \citep{Haugen2004} does not seem functional at low $\mathrm{Re}_\mathrm{M}$. The logarithmic scaling appears to be empirically valid, although it has been derived for low magnetic $\mathrm{Pr}_\mathrm{M}$, whereas in our simulations $\mathrm{Pr}_\mathrm{M}$ was unity.
We also estimated the turbulent transport coefficients $\alpha$ and $\eta_\mathrm{t}$ based on closure approaches. We found an $\alpha$ quenching behaviour comparable to \citet{Brandenburg2008} and calculated the growth rates based on $\alpha$ and $\eta_\mathrm{t}$.
They appear to be in a
similar approximate range, but there is a number of differences when compared to the direct measurements.
The difference between closure and direct estimates are a possible result of the fact that the former are very rough. To improve and check the goodness of the estimate, a more refined method such as the test-field method \citep{Schrinetal07} would be required, which is not currently supported by \textit{Astaroth}. In addition, in further studies the number of points on the $\mathrm{Re}_\mathrm{M}$ axis should be increased.
As we are also used this study as a way for exploring the scientific potential of the \textit{Astaroth} API, some remarks should be made on the practical aspects of computation. The work presented here has benefited significantly from the performance enhancement provided by \textit{Astaroth}. It was feasible to perform the dynamo simulations and related tests with two computing nodes, with four Tesla P100 devices per node. This made our simulation very affordable within the limits of the ASIAA high-performance computing cluster. More discussion of GPU performance can be found in the Appendix \ref{sec:performance}.
Additionally, it should be noted that an efficient GPU code can produce substantial amounts of data.
Therefore, benefits of the performance will come in contact with the limitation of the data processing tools, that might not be as efficient and/or optimized as the GPU code.
Astrophysicists rely on data analysis libraries, and we should note that for the maximal benefit of GPU performance, also connected data processing tools should be improved in efficiency.
For future enhancements to this work, there are several possibilities. First, more points in the $\mathrm{Re}_\mathrm{M}$-space could be covered. To better understand the scenario illustrated by our hypothesis depicted in Figure \ref{fig:LSDvsSSD}.
As the computation is efficient, this approach is basically limited by the available data storage. Second, a test-field method could be implemented to estimate the turbulent transfer coefficients $\alpha$ and $\eta_\mathrm{t}$ in a more precise manner.
The third possibility is to increase resolution and therefore $\mathrm{Re}_\mathrm{M}$ with the multi-node MPI implementation of \textit{Astaroth} becoming available.
\section{Conclusions}\label{sec:conclusion}
In this paper, we extended the \textit{Astaroth} library to work on multiple GPUs and applied it to study the turbulent dynamo problem.
Our implementation scaled from one to four GPUs with at least $90\%$ efficiency and exhibited a speedup of $35$ in single-node performance on four V100-SXM2-32GB GPUs compared with \textit{Pencil Code} runs on two 20-core Intel Xeon Gold 6230 Cascade Lake CPUs. Because of the limitations of CUDA peer-to-peer memory transfers, our implementation was restricted to a single node. Our results demonstrate that one-dimensional decomposition is sufficient to hide communication latencies within a node when carrying out computation and communication in parallel on current hardware. However, we expect that the use of MPI and multidimensional decomposition schemes are required for witnessing further scaling.
We simulated helical and non-helical MHD turbulence with homogeneous random forcing, and by modifying diffusivity and viscosity
within the resolution limits
to investigate the dependence of the dynamo growth
on the magnetic Reynolds number, while keeping the magnetic Prandtl number fixed to unity.
We were able to extend the $\mathrm{Re}_\mathrm{M}$ range to
somewhat larger values than in some of the older studies, but most importantly, produce a large set of simulations to determine the dependence more accurately than before.
We estimated growth rates from the simulations and saw that with helical turbulence an LSD would
grow
at any $\mathrm{Re}_\mathrm{M}\gtrsim 1$,
while SSD would appear only beyond a critical value $\mathrm{Re}_\mathrm{M} \sim 25$. SSD growth rates
followed a clear logarithmic $\mathrm{Re}_\mathrm{M}$ dependence.
Earlier studies have either not been able to determine a clear dependency due to the small amount of data points, or reported consistency with a $\sqrt{\mathrm{Re}_\mathrm{M}}$ dependency.
In helical simulations with $\mathrm{Re}_\mathrm{M} >$ 25, both dynamo instabilities are evidently present simultaneously.
To further inspect and separate the signatures of the SSD and LSD,
we determined magnetic powerspectra.
LSD spectra displayed growth at the largest scales while SSD ones showed a Kazantsev profile
at low to intermediate wavenumbers, and peaked at scales smaller than the forcing scale.
Spectra in the regime where both dynamo instabilities act together retains characteristic of the LSD at low wavenumbers, and those of the SSD at high wavenumbers,
although there is always a peak at the forcing scale.
In such circumstances, the powerspectra display SSD features during the exponential growth stage.
We computed probability density functions of the magnetic field, which showed
exponential shapes in the case of SSD, and
a Gaussian distribution deformed by a top-hat profile from the mean (Beltrami) field for LSD.
They also showed evidence for the coexistence of SSD
with LSD towards high $\mathrm{Re}_\mathrm{M}$.
To explain the behaviour of the growth rate in helically driven simulations we presented a hypothesis that the growth of LSD is inhibited by a budding SSD around medium $\mathrm{Re}_\mathrm{M}$.
We analyzed LSD
using closure estimates of turbulent transport coefficients. These estimated
growth rates agreed at low and medium $\mathrm{Re}_\mathrm{M}$ with the direct measurements,
but diverged
at high $\mathrm{Re}_\mathrm{M}$. The closure--estimated $\alpha$ displayed
clear signs of quenching at high $\mathrm{Re}_\mathrm{M}$.
\acknowledgments
\textit{Astaroth} is open source and available under GPL 3 license at \url{https://bitbucket.org/jpekkila/astaroth/}.
This work utilized tools developed by the CHARMS group and high-performance computing resources and cluster in ASIAA. This research has made use of {SAO/NASA} Astrophysics Data System. Additional compute resources for this work were provided by CSC -- IT Center for Science.
Authors thank Dr. Chun-Fan Liu
for an useful insight.
J.P., M.J.K., and M.R.\ acknowledge the support of the Academy of Finland
ReSoLVE Centre of Excellence (grant number 307411).
This project has received funding from the European Research Council (ERC)
under the European Union's Horizon 2020 research and innovation
programme (Project UniSDyn, grant agreement n:o 818665).
M.V., H.S., and R.K acknowledge funding support for Theory within ASIAA from Academia Sinica.
H.S. acknowledges grant support from Ministry of Science and Technology (MoST)
in Taiwan through 105-2119-M-001-044-MY3, and 108-2112-M-001-009-.
| 0f0ee13fe5a93b65e08f545d2afd2f9af672e309 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{abstract}
This paper argues that data of strategic individuals with heterogeneous privacy valuations in a distributed online social network (e.g., Facebook) will be under-priced, if traded in a monopoly buyer setting, and will lead to diminishing utilitarian welfare. This result, for a \textit{certain family} of online community data trading problems, is in stark contrast to a popular information economics intuition that increased amounts of end-user data signals in a data market improves its efficiency. Our proposed theory paves the way for a future (counter-intuitive) analysis of data trading oligopoly markets for online social networks (OSNs).
\section*{keywords}
distributed community, monopoly, social welfare
\vspace{-3 mm}
\section{Introduction}
Data of billions of online individuals are currently gathered, processed, and analyzed for personalized advertising or other online service\footnote{Facebook itself has approximately 2.5 billion monthly active individual users.}. This trend is on the high rise with a perennial increase in online apps, IoT technologies, and advanced AI/ML methodologies. It is a common and age-old notion in economics (see \cite{posner1,posner2,stigler,Laudon:1996:MP:234215.234476,acquisti2016economics,odlyzko,samuelson2000privacy,schwartz2003property,posner2018radical}) that the benefits and use of sharing individual information with the demand side of an information market is beneficial to targeted customization, demand side profit, and the growth of data-`hungry' AI/ML controlled businesses. It has also been argued by economists \cite{varian2009economic,farboodi2019big} that because of the above-mentioned benefits that individual data brings to a market setting, a competitive market mechanism might generate too little data sharing from the supply side.
In this letter, we rigorously argue, through a counter-example of a simple application type, that the popularly known economic intuition \emph{does not hold} in general, atleast for a monopoly information market setting. More specifically, we show that for some community settings (e.g., Facebook) trading end-user data/information signals in a monopoly market leads to diminishing economic utilitarian social welfare. \emph{The intuition behind this result primarily lies behind the negative externalities created via trading statistically correlated end-user signals when these heterogeneous users have varying privacy valuations of their data signals.} The result is contrary to recent results that intuit/prove that privacy can be detrimental to information market efficiency \cite{acquisti2016economics,posner2018radical,pal2019privacy, calo2015privacy,pal2020preference, acemoglu2019too, laoutaris2019online} if \emph{ideally, one's value of privacy is not high, or one's data is mildly correlated with others.} Specifically, we prove that information markets will be inefficient in \emph{non-ideal} community settings - formally hinted earlier in \cite{pal2020preference}.
Our analysis is complete for a monopoly structure, with a major takeaway being that in practical social community settings, sub-population privacy will jeopardized at a monopoly data trading market equilibrium. To the best of our knowledge, we are the first to mathematically dispel traditional information economics intuition, albeit for social network settings only. Moreover, it paves the way for a future (counter-intuitive) analysis of general community data trading oligopoly markets.
The rest of the paper is organized as follows. In Section II, we provide an intuition, via an example, towards proving our claim. We then follow this up in Section III with the description of a formal monopoly market model. In Section IV, we analyze this market model and formally prove our claim. We provide illustrative examples of our theory in Section V. We conclude the letter in Section VI.
\section{Intuition}
We provide an example-driven intuition that leads us to formally investigate the validity of the hypothesis that OSN user information promotes efficient data trading markets.
We focus on the widely popular \emph{Cambridge Analytica} scandal. The company acquired private information of millions of individuals from data shared by 270,000 Facebook users who voluntarily downloaded an app for mapping personality traits, called \emph{This is your digital life}. The app accessed users' news feed, timeline, posts, and messages, and revealed information about other Facebook users. The company was finally able to infer valuable information about more than 50 million Facebook users, which it deployed for designing personalized political messages and advertising in the Brexit referendum and the 2016 US presidential election. This scandal highlighted two important facets: (i) private information (e.g., behavior, habits, preferences) of users part of an online social community such as Facebook are correlated and results in knowing such information about other users\footnote{Habits and preferences of a highly educated gay from a particular locality is informative about others with the same profile and residing in the same area.} whose data is not leaked, and (ii) once it is openly publicized that valuable user information has been breached to satisfy external objectives, users are often miffed resulting in a huge social uproar as did happen in the case of the Cambridge Analytica scandal. These observations motivated us to develop a skepticism regarding the popular economic notion that more data implies increased information market efficiency. It could also be that trading in return for incentives for such community settings might not go down well\footnote{This is a high chance in scenarios of social uproar post publicly known data breaches, if not in cases where data breaches go un-noticed.} with the user privacy preferences, consequently hampering societal welfare (see Section III for a definition).
To state our intuition in a relatively more formal manner (courtesy of \cite{acemoglu2019too}), consider a community platform with two users, $i$ = 1, 2. Each user owns its own personal data, which we represent with a random variable $X_{i}$ (from the viewpoint of the platform). The relevant data of the two users are related, which we capture by assuming that their random variables are jointly normally distributed with mean zero and correlation coefficient $\rho$. The community platform can acquire or buy the data of a user in order to better estimate her preferences or actions. Its objective is to minimize the mean square error of its estimates of user types, or maximize the amount of leaked information about them.
Suppose that the valuation (in monetary terms) of the platform for the users’ leaked information is one, while the value that the first user attaches to her privacy, again in terms of leaked information about her, is $\frac{1}{2}$ and for the second user it is $v > 0$. We also assume that the platform makes take-it-or-leave-it offers to the users to purchase their data. In the absence of any restrictions on data markets or transaction costs, the first user will always sell her data (because her valuation of privacy, $\frac{1}{2}$, is less than the value of information to the platform, 1). But given the correlation between the types of the two users, this implies that the platform will already have a fairly good estimate of the second user’s information. Suppose, for illustration, that $\rho \simeq 1$. In this case, the platform will know almost everything relevant about user 2 from user 1’s data, and this undermines the willingness of user 2 to protect her data.
In fact, since user 1 is revealing almost everything about her, she would be willing to sell her own data for a very low price (approximately 0 given $\rho \simeq 1$). But once the second user is selling her data, this also reveals the first user’s data, so the first user can only charge a very low price for her data.
Therefore in this simple example, the community platform will be able to acquire both users’ data at approximately zero price. Critically, however, this price does not reflect the users’ valuation of privacy. When $v \le 1$, the equilibrium is efficient because data are socially beneficial in this case (even if data externalities change the distribution of economic surplus between the platform and users). However, it can be arbitrarily inefficient when $v$ is sufficiently high. This is because the first user, by selling her data, is creating a negative externality on the second user.
\section{System Model}
A simple example, such as the one aforementioned, clearly provides an intuition regarding the inefficiency of information trading in community settings with heterogeneous privacy valuations. In this section. en route to generalizing the validity (or invalidity) of our intuition, we propose a monopoly information trading market model (reproduced from \cite{acemoglu2019too})\footnote{We use the same notation for consistency purposes.} consisting of $n$ platform users and a profit-maximizing community platform (e.g., Facebook).
We consider \(n\) community users represented by the set \(\mathcal{V}=\{1, \ldots, n\} .\) Each user \(i \in \mathcal{V}\) has a type denoted by
\(x_{i}\) which is a realization of a random variable \(X_{i} .\) We assume that the vector of random variables
\(\mathbf{X}=\left(X_{1}, \ldots, X_{n}\right)\) has a joint normal distribution \(\mathcal{N}(0, \Sigma),\)
where $\Sigma \in \mathbf{R}^{n \times n}$ is the publicly known covariance
matrix of \(\mathbf{X} .\) Let \(\Sigma_{i j}\) designate the \((i, j)\) -th entry of \(\Sigma\) and \(\Sigma_{i i}=\sigma_{i}^{2}>0\) denote the variance of
individual \(i^{\prime}\)s type. Each user has some personal data, \(S_{i}\), which is informative about its type, i.e., the `DNA' that drives the user's tastes (for example, based
on her past behavior, preferences, or contacts). We suppose that \(S_{i}=X_{i}+Z_{i}\) where \(Z_{i}\) is an
independent random variable with standard normal distribution\footnote{This has taken various forms in the information privacy literature \cite{sarwate2013signal}.}, i.e., \(Z_{i} \sim \mathcal{N}(0,1)\). For any user joining the community platform, the platform can derive additional revenue (e.g., due to benefits of targeted advertising) if it can predict the user's type. We simply assume that the community platform’s revenue from each user is a deceasing function of the mean square error of its forecast of the user’s type, minus what the platform pays to users to acquire their information. More specifically, the objective of the platform is to minimize
\begin{equation}
\sum_{i \in \mathcal{V}}
\left(\mathbb{E}
\left[\left(\hat{x}_{i}(\mathbf{S})-X_{i}\right)^{2}\right]-\sigma_{i}^{2}+p_{i}\right)
\end{equation}
where \(\mathbf{S}\) is the vector of data the platform acquires, \(\hat{x}_{i}(\mathbf{S})\) is the platform's estimate of the user's type given this information, \(-\sigma_{i}^{2}\) is included as a convenient normalization, and \(p_{i}\) denotes payments (be it explicit or implicit) to user \(i\) from the platform for their data (we ignore for simplicity any other transaction costs incurred by the platform).
Users value their privacy, which we also model in a reduced-form manner (reflecting both pecuniary and non-pecuniary\footnote{As example, the fact that a user may receive a greater consumer surplus when the platform knows less about her or she may have a genuine demand for keeping her preferences, behavior, and information private. There may also be political and social reasons for privacy, for example, for concealing dissident activities or behaviors disapproved by some groups.} motives) as a function of the same mean square error.
We assume, specifically, that user \(i^{\prime}\) s value of privacy is \(v_{i} \geq 0,\) and her payoff is $
v_{i}\left(
\mathbb{E}
\left[\left(\hat{x}_{i}(
\mathbf{S}
)-X_{i}\right)^{2}\right]-\sigma_{i}^{2}\right)+p_{i}
$
This expression and its comparison with objective (1) clarifies that the platform and users have potentially-
opposing preferences over information about user type. We have again subtracted \(\sigma_{i}^{2}\) as a normalization, which ensures that if the platform acquires no additional information about the user and makes no payment to it, the payoff is zero.
Critically, users with \(v_{i}<1\) value their privacy less than the valuation that the platform attaches to information about them, and thus reducing the mean square error of the estimates of their types is socially beneficial. In contrast, users with \(v_{i}>1\) value their privacy more, and reducing their mean square error is socially costly. In settings without data externalities (where data about one user have no relevance to the information about other users - an example being collection agencies not gathering addresses locations), the first group of
users should allow the platform to acquire (buy) their data, while the second group should not. A simple market mechanism based on prices for data can implement this efficient outcome, in accordance to the traditional economic notion that more information implies better market efficiency. However, the situation could be very different in the presence of data externalities (e.g., online community settings such as Facebook).
A key notion for our analysis is breached information, which captures the reduction in the mean square error of the platform’s estimate of the type of a user. When the platform has no information about user $i$, its estimate satisfies $
\mathbb{E}\left[\left(\hat{x}_{i}-X_{i}\right)^{2}\right]=\sigma_{i}^{2}$. As the platform receives data from this and other
users, its estimate improves and the mean square error declines. The notion of breached information captures this reduction in mean square error (MSE). Specifically, let \(a_{i} \in\{0,1\}\) denote the data sharing action of user \(i \in \mathcal{V}\) with \(a_{i}=1\) corresponding to sharing. Denote the the profile of sharing decisions by \(\mathbf{a}=\left(a_{1}, \ldots, a_{n}\right)\) and the decisions of agents other than \(i\) by \(\mathbf{a}_{-i} .\) We also use the notation \(\mathbf{S}_{\mathbf{a}}\) to denote the data of all individuals for whom \(a_{j}=1,\) i.e., \(\mathbf{S}_{\mathbf{a}}=\left(S_{j}: j \in \mathcal{V}\, \mathrm{s.t.}\, a_{j}=1\right) .\) Given a profile of actions \textbf{a}, the breached information of (or about) user \(i \in \mathcal{V}\) is the reduction in the MSE of the best estimator of the type
of user \(i\) :
$\mathcal{I}_{i}(\mathbf{a})=\sigma_{i}^{2}-\min _{\hat{x}_{i}(\cdot)}
\mathbb{E}
\left[\left(X_{i}-\hat{x}_{i}\left(\mathbf{S}_{\mathbf{a}}\right)\right)^{2}\right].$
Notably, because of data externalities, breached information about user $i$ depends not just on her decisions but also on the sharing actions taken by all users. With this notion at hand, we can write the payoff of user $i$ given the price vector $p = (p_1, . . . , p_n)$
as
{\setlength\abovedisplayskip{0pt}
\setlength\belowdisplayskip{2pt}$$
u_{i}\left(a_{i}, \mathbf{a}_{-i}, \mathbf{p}\right)=\left\{\begin{array}{ll}p_{i}-v_{i} \mathcal{I}_{i}\left(a_{i}=1, \mathbf{a}_{-i}\right), & a_{i}=1 \\ -v_{i} \mathcal{I}_{i}\left(a_{i}=0, \mathbf{a}_{-i}\right), & a_{i}=0\end{array}\right.
$$}
where recall that $v_{i} \ge 0$ is user’s value of privacy. We also express the monopoly platform’s payoff more compactly as
{\setlength\abovedisplayskip{1 pt}
\setlength\belowdisplayskip{1 pt}\begin{equation}
U(\mathbf{a}, \mathbf{p})=\sum_{i \in \mathcal{V}} \mathcal{I}_{i}(\mathbf{a})-\sum_{i \in \mathcal{V}: a_{i}=1} p_{i}
\end{equation}
}
An action profile $a = (a_1,...,a_n)$ of the strategic users and a price vector $p = (p_1,...,p_n)$ for the users constitute a pure strategy equilibrium of the user-platform game if both users and the community platform maximize their payoffs given other players’ strategies. More formally, we define an equilibrium of this game as a \emph{Stackelberg equilibrium} in which the monopoly platform chooses the price vector recognizing the user equilibrium that will result following this choice.
\begin{definition}
\emph{Given the price vector \(\mathbf{p}=\left(p_{1}, \ldots, p_{n}\right),\) an action profile \(\mathbf{a}=\left(a_{1}, \ldots, a_{n}\right)\) is user equilibrium if for all \(i \in \mathcal{V}\),
{\setlength\abovedisplayskip{-2pt}
\setlength\belowdisplayskip{2pt}$$
a_{i} \in \operatorname{argmax}_{a \in\{0,1\}} u_{i}\left(a_{i}=a, \mathbf{a}_{-i}, \mathbf{p}\right).
$$}}
\end{definition}
We denote the set of user equilibria at a given price vector \(\mathbf{p}\) by \(\mathcal{A}(\mathbf{p})\). A pair \(\left(\mathbf{p}^{\mathrm{E}}, \mathbf{a}^{\mathrm{E}}\right)\) of price and
action vectors is a pure strategy Stackelberg equilibrium if \(\mathbf{a}^{\mathrm{E}} \in \mathcal{A}\left(\mathbf{p}^{\mathrm{E}}\right)\) and there is no profitable
deviation for the platform, i.e.,
{\setlength\abovedisplayskip{1pt}
\setlength\belowdisplayskip{0pt}$$
U\left(\mathbf{a}^{\mathrm{E}}, \mathbf{p}^{\mathrm{E}}\right) \geq U(\mathbf{a}, \mathbf{p}), \; {\rm for \; all \;} \mathbf{p} {\rm \;and \; for \; all \; } \mathbf{a} \in \mathcal{A}(\mathbf{p})
$$}
In what follows, we refer to a pure strategy Stackelberg equilibrium simply as an equilibrium.
We now characterize two important properties of the breached information function \(\mathcal{I}_{i}\) :
\(\{0,1\}^{n} \rightarrow \mathbf{R}\). \\
1. \emph{Monotonicity:} for two action profiles $\textbf{a}$ and $\textbf{a}'$ with $\textbf{a}$ \(\geq\) $\textbf{a}'$
{\setlength\abovedisplayskip{2pt}
\setlength\belowdisplayskip{2pt}$$
\mathcal{I}_{i}(\mathbf{a}) \geq \mathcal{I}_{i}\left(\mathbf{a}^{\prime}\right), \quad \forall i \in\{1, \ldots, n\}
$$}
2. \emph{Submodularity:} for two profiles $\textbf{a}$ and $\textbf{a}'$ with $\mathbf{a'}_{-i} \ge \mathbf{a}_{-i}$,
{\setlength\abovedisplayskip{2pt}
\setlength\belowdisplayskip{2pt}{\small
$$
\mathcal{I}_{i}\left(a_{i}=1, \mathbf{a}_{-i}\right)-\mathcal{I}_{i}\left(a_{i}=0, \mathbf{a}_{-i}\right) \geq \mathcal{I}_{i}\left(a_{i}=1, \mathbf{a}_{-i}^{\prime}\right)-\mathcal{I}_{i}\left(a_{i}=0, \mathbf{a}_{-i}^{\prime}\right)
$$
}}
The monotonicity property states that as the set of community users who share their information expands, the breached information about each user (weakly) increases. This is an intuitive consequence of the fact that more information always facilitates the estimation problem of the platform and reduces the mean square error of its estimates. More important for the rest of our analysis is the submodularity property, which implies that the marginal increase in the breached information from individual $i$’s sharing decision is decreasing in the information shared by others. This too is intuitive and follows from the fact that when others’ actions reveal more information, there is less to be revealed by the sharing decision of any given individual. Thus, from the celebrated result due to Topkis \cite{topkis1978minimizing}, for any $\mathbf{p}$, the set \(\mathcal{A}(\mathbf{p})\) is a complete lattice, and thus has a least and a greatest element. This implies that the set of user equilibria is always non-empty.
\section{Monopoly Market Analysis}
Enroute to analyzing the market welfare generated via the aforementioned game setting, we first define the benchmark \emph{first best} welfare outcome as the data sharing decisions that maximize utilitarian social welfare or social surplus given by the sum of the payoffs of the platform and users. Social surplus (SoS) from an action profile \textbf{a} is
\[
SoS(\textbf{a})=U(\mathbf{a}, \mathbf{p})+\sum_{i \in \mathcal{V}} u_{i}(\mathbf{a}, \mathbf{p})=\sum_{i \in \mathcal{V}}\left(1-v_{i}\right) \mathcal{I}_{i}(\mathbf{a})\]
Note that prices do not appear in this expression because they are transfers from the community platform to users. The first-best action profile, $\textbf{a}^{W}$, maximizes this expression. The following theorem (built on \cite{acemoglu2019too}) characterizes the first-best action profile.
\begin{theorem}\label{Proposition-one} (due to \cite{acemoglu2019too})
\emph{The first best involves \(a_{i}^{\mathrm{W}}=1\) if
{\setlength\abovedisplayskip{1 pt}
\setlength\belowdisplayskip{1 pt}
\begin{equation}
\sum_{j \in \mathcal{V}}\left(1-v_{j}\right) \frac{\left({\rm Cov}\left(X_{i}, X_{j} | a_{i}=0, \mathbf{a}_{-i}^{\mathrm{W}}\right)\right)^{2}}{1+\sigma_{j}^{2}-\mathcal{I}_{j}\left(a_{i}=0, \mathbf{a}_{-i}^{\mathrm{W}}\right)} \geq 0
\end{equation}
}
and \(a_{i}^{\mathrm{W}}=0\) if (3) is negative.}
\end{theorem}
\textbf{Implication} - To understand this result, consider first the case in which there are no data externalities so that the covariance terms in (3) are zero, except
\({\rm Cov}\left(X_{i}, X_{i} | a_{i}=0, \mathbf{a}_{-i}^{\mathrm{W}}\right)=\sigma_{i}^{2}\), so that the left-hand side is simply \(\sigma_{i}^{4} /\left(1+\sigma_{i}^{2}\right)\) times $1- v_i$. This yields $a^W_i = 1$ if $v_i \le 1$ (thus a no externality setting becomes mathematically equivalent to case when all users do not value their privacy enough) . The situation is different in the presence of data externalities, because now the covariance terms are non-zero. In this case, an individual should optimally share her data only if it does not reveal too much about users with $v_j > 1$. Note here that the covariance matrix can be robustly estimated from publicly observed $S_{i}$ values as dependencies are usually preserved in the addition of noise within a threshold.
In this section, we adopt the more realistic assumption that, to start with, the monopoly platform does not know the exact valuations of users in a community (in contrast to assumptions made in existing works such as \cite{wang2016value}) that are only private to them, but are informed that the valuations $v_{i}$ come from the cumulative distribution function $F_{i}$ and density function $f_{i}$ (with upper support denoted by $v_{\max}$). We allow the platform to design a pricing mechanism that elicits the true privacy valuations from the users \emph{(as Step 1 of the economy)}, somewhat similar to the seminal Vickrey-Clarkes-Groves (VCG) mechanism with a minor variation. More specifically, for any user $i \in \mathcal{V}$ the price offered to user $i$ \emph{(as Step 2 of the economy)} is equal to the surplus of all other users on the platform when user $i$ is present minus by the surplus when user $i$ is absent. We consequently have the following result.
\begin{theorem}\label{eight} (due to \cite{acemoglu2019too})
\emph{Let $\mathbf{v}$ be the reported vector of values of privacy. Then the non-negative pricing scheme}
\emph{
\begin{equation*}
p_{i}(\mathbf{v})=\left(\mathcal{I}_{i}(\mathbf{a}(\mathbf{v}))+\sum_{j \neq i}\left(1-v_{j}\right) \mathcal{I}_{j}(\mathbf{a}(\mathbf{v}))\right)
-\min _{\mathbf{a} \in\{0,1\}^{n}}\left(\mathcal{I}_{i}(\mathbf{a})+\sum_{j \neq i}\left(1-v_{j}\right) \mathcal{I}_{j}(\mathbf{a})\right)
\end{equation*}}
\emph{where $\mathbf{a}(\mathbf{v})=\arg \max _{\mathbf{a} \in\{0,1\}^{n}} \sum_{i \in \mathcal{V}}\left(1-v_{i}\right) \mathcal{I}_{i}(\mathbf{a})$ incentivizes users to report their value of privacy truthfully.}
\end{theorem}
\begin{definition}
\emph{An equilibrium is a pair \(\left(\mathbf{a}^{\mathrm{E}}, \mathbf{p}^{\mathrm{E}}\right)\) of functions of the reported valuations \(\mathbf{v}=\)
\(\left(v_{1}, \ldots, v_{n}\right)\) such that each user reports its true value and the expected payoff of the platform
is maximized. That is},
{{
\begin{equation*}
\begin{split}&\left(\mathbf{a}^{\mathrm{E}}, \mathbf{p}^{\mathrm{E}}\right)= \max _{\mathbf{a}:
\mathbb{R}^{n} \rightarrow\{0,1\}^{n}, \mathbf{p}: \mathbb{R}^{n}
\rightarrow \mathbb{R}^{n}} \mathbb{E}_{\mathbf{v}}\left[\sum_{i=1}^{n} \mathcal{I}_{i}(\mathbf{a}(\mathbf{v}))-\sum_{i: a_{i}(\mathbf{v})=1} p_{i}(\mathbf{v})\right] \\ & p_{i}(\mathbf{v})-v_{i} \mathcal{I}_{i}(\mathbf{a}(\mathbf{v})) \geq p_{i}\left(\mathbf{v}_{-i}, v_{i}^{\prime}\right)-v_{i} \mathcal{I}_{i}\left(\mathbf{a}\left(\mathbf{v}_{-i}, v_{i}^{\prime}\right)\right), \forall v_{i}^{\prime}, \mathbf{v}: i \in \mathcal{V} \end{split}
\end{equation*}}}
\end{definition}
We now have the following theorem characterizing the equilibrium of the monopoly market setting.
\begin{theorem}\label{eleven} (due to \cite{acemoglu2019too})
\emph{For any reported vector of values v, the market equilibrium is given by}
\emph{
\begin{equation*}
\mathbf{a}^{\mathrm{E}}(\mathbf{v})=\operatorname{argmax}_{\mathbf{a} \in\{0,1\}^{n}} \sum_{i=1}^{n}\left(1-\Phi_{i}\left(v_{i}\right)\right) \mathcal{I}_{i}(\mathbf{a})
+\Phi_{i}\left(v_{i}\right) \mathcal{I}_{i}\left(\mathbf{a}_{-i}, a_{i}=0\right)
\end{equation*}
}
\emph{and}
{
\begin{equation*}
p_{i}^{\mathrm{E}}\left(v_{i}\right)= \int_{v}^{v_{\max }}
\left(\mathcal{I}_{i}\left(\mathbf{a}^{\mathrm{E}}\left(x,
\mathbf{v}_{-i})\right)-\mathcal{I}_{i}\left(\mathbf{a}_{-i}^{\mathrm{E}}\left(x,
\mathbf{v}_{-i}
\right), a_{i}=0\right)\right)\right) d x
+v_{i}\left(\mathcal{I}_{i}\left(\mathbf{a}^{\mathrm{E}}\left(v_{i}, \mathbf{v}_{-i}\right)\right)-\mathcal{I}_{i}\left(\mathbf{a}_{-i}^{\mathrm{E}}\left(v_{i}, \mathbf{v}_{-i}\right), a_{i}=0\right)\right)
\end{equation*}
}
\emph{Moreover, all users report truthfully and thus the expected payoff of the platform is}
{$$
\mathbb{E}_{\mathbf{v}}\left[\max _{\mathbf{a} \in\{0,1\}^{n}} \sum_{i=1}^{n}\left(1-\Phi_{i}\left(v_{i}\right)\right) \mathcal{I}_{i}(\mathbf{a})+\Phi_{i}\left(v_{i}\right) \mathcal{I}_{i}\left(\mathbf{a}_{-i}, a_{i}=0\right)\right]
$$}
\emph{where, $\Phi_{i}(v) = v + \frac{F_{i}(v)}{f_{i}(v)}$ is a non-decreasing function representing the additional rent that a user will capture in incentive compatible mechanisms.}
\end{theorem}
\textbf{Implication} - The theorem guarantees the existence of a unique monopoly market equilibrium where community platform users elicit their true valuations. A sufficient condition for $\Phi_{i}(v)$ to be non-decreasing is for $\frac{f_{i}(v)}{F_{i}(v)}$ to be non-increasing. This requirement is satisfied for a variety of distributions such as uniform and exponential \cite{burkschat2014reversed}.
We now investigate whether the reachable market equilibrium is efficient. We have the following result, via \cite{acemoglu2019too}, in this regard.
\begin{theorem}\label{twelve} (due to \cite{acemoglu2019too})
\emph{1. Suppose high-value users are uncorrelated with all other users and $\mathcal{V}^{(l)}=\mathcal{V}_{\Phi}^{(l)}$, where $\mathcal{V}_{\Phi}^{(l)}=\left\{i \in \mathcal{V}: \Phi_{i}\left(v_{i}\right) \leq 1\right\}$ denotes the set of users with $\Phi_{i}(v_{i}) \le 1$. Then
the market equilibrium is \textbf{efficient}.\\
2. Suppose some high-value users (those in $\mathcal{V}^{(h)}$) are correlated with users in $v_{\phi}^{(l)}$. Then there exists $\overline{\mathfrak{v}} \in \mathbb{R}^{| V(n) |}$ such that for $\mathbf{v}^{(h)} \geq \overline{\mathbf{v}}$ the market equilibrium is \textbf{inefficient}.\\
3. Suppose every high-value user is uncorrelated with all users in $\mathcal{V}_{\phi}^{(l)},$ but users in a nonempty subset
$\hat{\mathcal{V}}^{(l)}$ of $\mathcal{V}^{(l)} \backslash \mathcal{V}_{\phi}^{(l)}$ are correlated with at least one high-value user. Then there exist $\overline{\mathbf{v}}$ and $\tilde{v}$ such that if $\mathbf{v}^{(h)} \geq \overline{\mathbf{v}}$ and $v_{i}<\tilde{v}$ for some $i \in \hat{\mathcal{V}}^{(l)}$, the market equilibrium is \textbf{inefficient}.\\
4. Suppose every high-value user is uncorrelated with all low-value users and at least one high-value user is correlated with another high-value user. Let $\tilde{\mathcal{V}}^{(h)} \subseteq \mathcal{V}^{(h)}$ be the subset of high-value users correlated with at least one other high-value user. Then for each $i \in \tilde{\mathcal{V}}^{(h)}$ there exists $\bar{v}_{i}>0$ such that if for any $i \in \tilde{\mathcal{V}}^{(h)} v_{i}<\bar{v}_{i},$ the market equilibrium is \textbf{inefficient}.\\
5.The social surplus at market equilibrium, $\mathbf{a}^{\mathrm{E}}$, for any \textbf{v} (either known truthfully or otherwise)
\[SoS\left(\mathbf{a}^{\mathrm{E}}\right) \leq \sum_{i: v_{i} \in \mathcal{V}^{(l)}}\left(1-v_{i}\right) \mathcal{I}_{i}(\mathcal{V})-\sum_{i: v_{i} \in \mathcal{V}^{(h)}}\left(v_{i}-1\right) \mathcal{I}_{i}\left(\mathcal{V}_{\phi}^{(l)}\right)\]}
\end{theorem}
\textbf{Implication} - The theorem provides the conditions under which market equilibrium in a monopoly community information trading setting is utilitarian welfare (in)efficient. Note that the market efficiency results in the theorem are conservative in the sense we assume user privacy valuations are unknown in the worst case, and the platform does its best to elicit true valuation responses. Inefficiency in this setting would imply inefficiency in the case where community user valuations are untruthful (see point \#5 in the theorem). According to the points in the theorem, it is evident that the information trading market is efficient is when high-value users (those with both $v_{i}, \Phi_{i}(v_{i}) \ge 1$ are uncorrelated with all other users - something practically rare to achieve, and low-value users have $\Phi_{i}(v_{i})$ values less than one (note here that low-value users always have $v_{i} <1$ but could have $\Phi_{i}(v_{i})$ values greater than 1). Not satisfying either will lead to incentive compatibility conditions preventing efficient allocation. In all other cases, the information trading market is inefficient (SoS at equilibrium is not optimal), and the extent of inefficiency depends on whether high-value users are correlated with low-value users with $\Phi_{i}(v_{i})$ values greater or less than one.
\section{Examples}
In this section, we provide numerical examples (as in \cite{acemoglu2019too}) to lucidly illustrate (a) the existence of a data trading market equilibrium, and (b) the social surplus (SoS) zone at market equilibrium.
\noindent\textbf{Example 1. \label{example1}
Suppose there are two users 1 and 2 with covariance matrix
$\Sigma=\left(\begin{array}{lr}1 & \rho \\ \rho & 1\end{array}\right)$
and \(v_{1}=v_{2}=v .\)
When \(p_{1}, p_{2} \in\left[\frac{\left(2-\rho^{2}\right)^{2}}{2\left(4-\rho^{2}\right)}, \frac{1}{2}\right],\) both action profiles \(a_{1}=a_{2}=0\) and \(a_{1}=a_{2}=1\) are user
equilibria. This is a consequence of the submodularity of the leaked information functio
: when user 1 shares her data, she is also revealing a lot about user 2 , and making it less costly
for her to share her data. Conversely, when user 1 does not share, this encourages user 2 not to
share. Despite this multiplicity of user equilibria, there exists a unique (Stackelberg) equilibrium
for this game given by \(a_{1}^{\mathrm{E}}=a_{2}^{\mathrm{E}}=1\) and \(p_{1}^{\mathrm{E}}=p_{2}^{\mathrm{E}}=\frac{\left(2-\rho^{2}\right)^{2}}{2\left(4-\rho^{2}\right)} .\) This uniqueness follows because the platform can choose the price vector to encourage both users to share. \emph{The next example suggests that though there may be multiple equilibria in the Stackelberg game, all of them yield the same payoff for the community platform.}
\noindent\textbf{Example 2. \label{example2}
Suppose there are three users with the same value of privacy and variance: \(v_i = 1.18\) and \(\sigma_{i}^{2}=1\) for \(i=1,2,3 .\) We let all off-diagonal entries of \(\Sigma\) to be \(0.3 .\) Any action profile where two out of three users share their information is an equilibrium, and thus there are three distinct equilibria. But it is straightforward to verify that they all yield the same payoff to the platform.
The following example illustrates the social welfare zone at market equilibrium with variations in the correlation coefficient $\rho$ and privacy valuation for high-value community users.
\noindent\textbf{Example 3. \label{example4}
We consider a setting with two communities, each of size 10. Suppose that all users in
community 1 are low-value and have a value of privacy equal to 0.9, while all users in community
2 are high-value (with \(v_{h}>1\) ). We also take the variances of all user data to be 1, the correlation
between any two users who belong to the same community to be 1/20, and the correlation be-
tween any two users who belong to different communities to be \(\rho\). depicts equilibrium
surplus as a function of \(v_{h}\) and \(\rho\).
\noindent \emph{Two points are worth noting from this example. First, relatively small values of the correlation coefficient \(\rho\) are sufficient for social surplus to be negative. Second, when \(v_{h}\) is very close to 1, the social surplus is always positive because the negative surplus from high-value users is compensated by the social benefits their data sharing creates for low-value users.}
\section{Conclusion}
In this paper, we mathematically argued, using recent developments in \cite{acemoglu2019too}, that social community data trading is not economically welfare efficient in a monopoly market setting, going against the popular economic philosophy/intuition that increased amounts of end-user data signals in a market improves utilitarian social welfare. The primary reason behind our result is the significant negative externality (via user signal correlations) generated by privacy breaches in the information market that cannot be cancelled out via market equilibrium prices handed over to the users for their information.
\section{Acknowledgement}
This work NSF-supported under grants CNS-1616575, CNS-1939006, CNS-2012001, and ARO W911NF1810208.
\bibliographystyle{unsrt}
| 162811ae6cbec8e46df3754fdda6e60c1c2a56e6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Proof of Theorem \ref{thm:master_W2}}\label{sec proof thm 1}
\noindent\textbf{Notation.}~For the reader's convenience, we recall some necessary notation from Section \ref{sec main}. We say that a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$, denoted $f\in\rm{PL}(k)$, if there is a constant $L>0$ such that for all $\vct{x},\vct{y}\in\mathbb{R}^m$,
$
|f(\vct{x}) - f(\vct{y})|\leq L(1+\tn{\vct{x}}^{k-1}+\tn{\vct{y}}^{k-1})\|\vct{x}-\vct{y}\|_2
$
(See also Section \ref{SM useful fact}).
We say that a sequence of probability distributions $\nu_p$ on $\mathbb{R}^m$ {converges in $W_k$} to
$\nu$, written $\nu_p\stackrel{W_k}{\Longrightarrow} \nu$, if $W_k(\nu_p,\nu) \rightarrow 0$ as $p \rightarrow \infty$. An equivalent definition is that, for any $f\in\rm{PL}(k)$, $\lim_{p\rightarrow}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$ (e.g., \cite{montanari2017estimation}).
For a sequence of random variables $\mathcal{X}_{p}$ that converge in probability to some constant $c$ in the limit of Assumption \ref{ass:linear} below, we write $\mathcal{X}_{p}\stackrel{{P}}{\longrightarrow} c$. For a sequence of event $\mathcal{E}_p$ for which $\lim_{p\rightarrow}\mathbb{P}(\mathcal{E}_p) = 1$, we say that $\mathcal{E}_p$ occurs \emph{with probability approaching 1}. For this, we will often use the shorthand ``wpa. 1".
\vspace{5pt}
Let ${\mtx{X}}\in\mathbb{R}^{n\times p}$ have zero-mean and normally distributed rows with a diagonal covariance matrix ${\boldsymbol{{\Sigma}}}=\operatorname{\mathbb{E}}[\vct{x}\x^T]$. Given a ground-truth vector $\betab^\star$ and labels $\vct{y}={\mtx{X}}\betab^\star+\sigma {\vct{z}},~{\vct{z}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, we consider the least-squares problem subject to the minimum Euclidian norm constraint (as $\kappa=p/n>1$) given by
\begin{align}\label{eq:PO_beta}
\min_{{\boldsymbol{\beta}}}\frac{1}{2}\tn{{\boldsymbol{\beta}}}^2\quad\text{subject to}\quad \vct{y}={\mtx{X}}{\boldsymbol{\beta}}.
\end{align}
It is more convenient to work with the following change of variable: $\vct{w}:=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\beta}}-\betab^\star)$. With this, the optimization problem in \eqref{eq:min_norm} can be rewritten as
\begin{align}\label{eq:PO}
\Phi({\mtx{X}})=\min_{\vct{w}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
where we write ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$. First, using standard arguments, we show that the solution of \eqref{eq:PO} is bounded. Hence, we can constraint the optimization in a sufficiently large compact set without loss of generality.
\begin{lemma}[Boundedness of solution]\label{lem:bd_PO}
Let $\hat\vct{w}_n:=\hat\vct{w}_n({\mtx{X}},{\vct{z}})$ be the minimizer in \eqref{eq:PO}. Then, with probability approaching 1, it holds that $\hat\vct{w}_n\in\mathcal{B}$, where
$$\mathcal{B}:=\left\{\vct{w}\,|\,\|\vct{w}\|_2\leq B_{+} \right\},\qquad B_+:=4\sqrt{\Sigma_{\max}}\frac{2(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + \sigma}{\sqrt{\Sigma_{\min}}(\sqrt{\kappa}-1)} + \sqrt{\Sigma_{\max}}\sqrt{\operatorname{\mathbb{E}}\left[B^2\right]}.
$$
\end{lemma}
\begin{proof}
First, we show that the min-norm solution $\hat\boldsymbol{\beta}={\mtx{X}}^T({\mtx{X}}\X^T)^{-1}\vct{y}$ of \eqref{eq:PO_beta} is bounded. Here, we used the fact that $\kappa>1$, thus ${\mtx{X}}\X^T$ is invertible wpa 1. We
have,
\begin{align}
\tn{\hat\boldsymbol{\beta}_n}^2 = \vct{y}^T({\mtx{X}}\X^T)^{-1}\vct{y} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{X}}\X^T)} = \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\boldsymbol{\Sigma}{\mtx{\bar{X}}}^T)} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\Xb^T)\,\Sigma_{\min}} = \frac{\tn{\vct{y}}^2}{\sigma_{\min}^2({\mtx{\bar{X}}})\,\Sigma_{\min}}. \label{eq:Ubb}
\end{align}
But, wpa 1,
$
\sigma_{\min}({\mtx{\bar{X}}})/\sqrt{n} \geq \frac{1}{2}\left(\sqrt{\kappa}-1\right).
$
Furthermore,
$
\|\vct{y}\|_2 \leq \|{\mtx{\bar{X}}}\boldsymbol{\Sigma}^{1/2}\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2 \leq \sigma_{\max}({\mtx{\bar{X}}})\sqrt{\Sigma_{\max}} \|\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2.
$
Hence, wpa 1,
$$
\|\vct{y}\|_2/\sqrt{n} \leq 4(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma,
$$
where we used the facts that wpa 1: $\|z\|_2/\sqrt{n}\stackrel{{P}}{\longrightarrow} 1$, $\sigma_{\max}({\mtx{\bar{X}}})<2(\sqrt{\kappa}+1)$ and by Assumption \ref{ass:mu}:
$$
\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^{p}(\sqrt{p}\betab^\star_i)^2 \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[B^2\right].
$$
Put together in \eqref{eq:Ubb}, shows that
\begin{align}
\tn{\hat\boldsymbol{\beta}_n} < \frac{4(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma}{\sqrt{\Sigma_{\min}}(\sqrt{\kappa}-1)/2} =: \tilde{B}_+.\label{eq:bd_beta}
\end{align}
Recalling that $\hat\vct{w}_n= \sqrt{\boldsymbol{\Sigma}}\boldsymbol{\beta} - \sqrt{\boldsymbol{\Sigma}}\betab^\star$, we conclude, as desired, that wpa 1,
$
\tn{\hat\vct{w}_n} \leq \sqrt{\Sigma_{\max}}\tilde{B}_+ + \sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}[B^2]} =: B_+.
$
\end{proof}
Lemma \ref{lem:bd_PO} implies that nothing changes in \eqref{eq:PO} if we further constrain $\vct{w}\in\mathcal{B}$ in \eqref{eq:PO}. Henceforth, with some abuse of notation, we let
\begin{align}\label{eq:PO_bd}
\Phi({\mtx{X}}):=\min_{\vct{w}\in\mathcal{B}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
Next, in order to analyze the primary optimization (PO) problem in \eqref{eq:PO_bd} in apply the CGMT \cite{thrampoulidis2015lasso}. Specifically, we use the constrained formulation of the CGMT, Theorem \ref{thm closed}. Specifically, the auxiliary problem (AO) corresponding to \eqref{eq:PO_bd} takes the following form with ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$
\begin{align}
\phi({\vct{g}},\vct{h}) = \min_{\vct{w}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_con}
\end{align
We will prove the following result about the AO problem.
\begin{lemma}[Properties of the AO -- Overparameterized regime]\label{lem:AO}
Let $\phi_n=\phi({\vct{g}},\vct{h})$ be the optimal cost of the minimization in \eqref{eq:AO_con}. Define $\bar\phi$ as the optimal cost of the following deterministic min-max problem
\begin{align}\label{eq:AO_det}
\bar\phi:=\max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau):=\frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{N^2}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).
\end{align}
The following statements are true.
\noindent{(i).}~The AO minimization in \eqref{eq:AO_con} is $\frac{1}{\Sigma_{\max}}$-strongly convex and has a unique minimizer $\vct{w}_n:=\vct{w}_n({\vct{g}},\vct{h})$.
\noindent{(ii).}~In the limit of $n,p\rightarrow\infty, p/n=\kappa$, it holds that $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$, i.e., for any $\varepsilon>0$:
$$
\lim_{n\rightarrow\infty}\P\left(|\phi({\vct{g}},\vct{h})-\bar\phi|>\varepsilon\right) = 0.
$$
\noindent{(iii).} The max-min optimization in \eqref{eq:AO_det} has a unique saddle point $(u_*,\tau*)$ satisfying the following:
$$
u_*/\tau_* = \xi\quad\text{and}\quad\tau_* = \gamma,
$$
where $\xi, \gamma>0$ are defined in Definition \ref{def:Xi}.
\noindent{(iv).}~Let $f:\mathbb{R}\rightarrow\mathbb{R}$ be a $\rm{PL}({k})$ function. Let $\boldsymbol{\beta}_n=\boldsymbol{\Sigma}^{-1/2}\vct{w}_n + \betab^\star$. Then,
$$
\frac{1}{p}\sum_{i=1}^{p}f\left(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(B,\Lambda,H)\sim\mu\otimes \mathcal{N}(0,1)}\left[f\left(X_{\kappa,\sigma^2}(B,\Lambda,H),B,\Lambda\right) \right].
$$
\end{lemma}
We prove Lemma \ref{lem:AO} in Section \ref{sec:proofAO}. Here, we show how this leads to the proof of Theorem \ref{thm:master_W2} when combined with the CGMT.
\noindent\textbf{Finishing the proof of Theorem \ref{thm:master_W2}:} For convenience, define
$$F_n(\hat\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}\hat\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Fix any $\varepsilon>0$. It suffices to prove that the solution $\hat\boldsymbol{\beta}_n$ of the PO in \eqref{eq:PO_beta} satisfies $\hat\boldsymbol{\beta}_n\not\in\mathcal{S}$ wpa. 1, where
\begin{align}\label{eq:S_set}
\mathcal{S} = \mathcal{S}(\betab^\star,\boldsymbol{\Sigma}) =\{\boldsymbol{\beta}{~\big |~} |F_n(\hat\boldsymbol{\beta}_n,\betab^\star,\boldsymbol{\Sigma})-\alpha_*|\geq 2\varepsilon\}.
\end{align}
In particular, define the ``perturbed" PO and AO problems (compare to \eqref{eq:PO} and \eqref{eq:AO_con}) as:
\begin{align}\label{eq:PO_S}
\Phi_S({\mtx{X}})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
and
\begin{align}
\phi_S({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_S}
\end{align
where recall that ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$, ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$ and we have used the change of variables $\vct{w}:=\sqrt{\boldsymbol{\Sigma}}(\boldsymbol{\beta}-\betab^\star)$ for convenience.
Using \cite[Theorem 6.1(iii)]{thrampoulidis2018precise} it suffices to find costants $\bar\phi, \bar\phi_S$ and $\eta>0$ such that the following hold:
\begin{enumerate}
\item $\bar\phi_S \geq \bar\phi + 3\eta$,
\item $\phi({\vct{g}},\vct{h}) \leq \bar\phi + \eta$, with probability approaching 1,
\item $\phi_S({\vct{g}},\vct{h}) \geq \bar\phi_S - \eta$, with probability approaching 1.
\end{enumerate}
In what follows, we explicitly find $\bar\phi, \bar\phi_S,\eta$ such that the three conditions above hold.
\vspace{5pt}
\noindent\underline{Satisfying Condition 2}: Recall the deterministic min-max optimization in \eqref{eq:AO_det}. Choose $\bar\phi=\mathcal{D}(u_*,\tau_*)$ be the optimal cost of this optimization. From Lemma \ref{lem:AO}(ii), $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$. Thus, for any $\eta>0$, with probability approaching 1:
\begin{align}\label{eq:phi_lim}
\bar\phi + \eta \geq \phi({\vct{g}},\vct{h}) \geq \bar\phi - \eta.
\end{align}
Clearly then, Condition 2 above holds for any $\eta>0$.
\vspace{5pt}
\noindent\underline{Satisfying Condition 3}: Next, we will show that the third condition holds for appropriate $\bar\phi$.
Let $\vct{w}_n=\vct{w}_n({\vct{g}},\vct{h})$ be the unique minimizer of \eqref{eq:AO_con} as per Lemma \ref{lem:AO}(i), i.e., $\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}_n+{\boldsymbol{\beta}}^\star}^2 = \phi({\vct{g}},\vct{h})$. Again from Lemma \ref{lem:AO}, the minimization in \eqref{eq:AO_con} is $1/\Sigma_{\max}$-strongly convex in $\vct{w}$. Here, $\Sigma_{\max}$ is the upper bound on the eigenvalues of $\boldsymbol{\Sigma}$ as per Assumption \ref{ass:inv}. Thus, for any $\tilde\varepsilon>0$ and any feasible $\vct{w}$ the following holds (deterministically):
\begin{align}\label{eq:sc}
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+{\boldsymbol{\beta}}^\star}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2C},~\text{provided that}~ \|\vct{w}-\vct{w}_n({\vct{g}},\vct{h})\|_2 \geq \tilde\epsilon.
\end{align}
Now, we argue that $\boldsymbol{\beta}\in\mathcal{S}$ implies that $\|\vct{w}-\vct{w}_n({\vct{g}},\vct{h})\|_2\geq \tilde\varepsilon$ wpa 1, for
appropriate value of $\tilde\varepsilon$ and $\vct{w}=\sqrt{\boldsymbol{\Sigma}}(\boldsymbol{\beta}-\betab^\star)$. Consider any $\boldsymbol{\beta}\in\mathcal{S}$. First, by definition in \eqref{eq:S_set},
$$
|F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \geq 2\varepsilon.
$$
Second, by Lemma \ref{lem:AO}(iv), with probability approaching 1,
$$
|F(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \leq \epsilon.
$$
Third, we will show that wpa 1, there exists universal constant $C>0$ such that
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| \leq C {\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}\label{eq:dev2show}.
\end{align}
Before proving \eqref{eq:dev2show}, let us argue how combining the above three displays shows the desired. Indeed, wpa. 1,
\begin{align*}
2\varepsilon &\leq |F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \leq |F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| + |F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \\
&\leq \epsilon + C \,\|\boldsymbol{\beta}-\boldsymbol{\beta}_n\|_2. \\
&\qquad\Longrightarrow \|\boldsymbol{\beta}-\boldsymbol{\beta}_n\|_2 \geq {\varepsilon}/{C}=:\hat\varepsilon\\
&\qquad\Longrightarrow \|\vct{w}-\vct{w}_n\|_2 \geq \hat\varepsilon\sqrt{\Sigma_{\min}}=:\tilde\varepsilon.
\end{align*}
In the last line above, we recalled that $\boldsymbol{\beta}=\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star$ and $\boldsymbol{\Sigma}_{i,i}\geq\Sigma_{\min},~i\in[p]$ by Assumption \ref{ass:inv}.
From this and \eqref{eq:sc}, we find that wpa 1,
$
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\boldsymbol{\beta}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2C},~\text{for all}~ \vct{w}\in\mathcal{S}.
$
Thus,
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}.\notag
\end{align}
When combined with \eqref{eq:phi_lim}, this shows that
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta.
\end{align}
Thus, choosing $\bar\phi_S = \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}$ proves the Condition 3 above.
To complete the proof, let us now show \eqref{eq:dev2show}. Henceforth, $C$ is used to denote a universal constant whose value can change from line to line.
To simplify notation, let $\vct{a}_i=(\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}),\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\boldsymbol{\beta}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$, for $i\in[p]$. Then, using $f\in\rm{PL}(k)$,
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| &\leq \frac{L}{p}\sum_{i=1}^p (1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\})\|\vct{a}_i-\vct{b}_i\|_2\notag \\
&= \frac{L}{p}\sum_{i=1}^p \left(1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\}\right)\cdot|\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}) - \sqrt{p}\boldsymbol{\beta}_i|\notag\\
&\leq L\left(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} \right)^{1/2}{\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}\label{eq:LipC}\,.
\end{align}
The last inequality above follows by applying Cauchy-Schwartz.
To reach \eqref{eq:dev2show} from the above, it suffices to show that
\begin{align}
\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} \leq C,\label{eq:leqC}
\end{align}
for some universal constant $C$ (that may depend on $k$, but not on $n,p$).
To prove \eqref{eq:leqC}, note that
using $a^{\ell_1}b^{\ell_2}c^{\ell_3}\leq a^{\ell}+b^{\ell}+c^\ell,~\ell=\ell_1+\ell_2+\ell_3$, for some constant $C=C(k)>0$ it holds:
\begin{align}
\frac{1}{p}\sum_{i=1}^p \|\vct{b}_i\|_2^{2k-2} {\leq} C\left(\frac{1}{p}\sum_{i=1}^p |\sqrt{p}\boldsymbol{\beta}_{i}|^{2k-2} + \frac{1}{p}\sum_{i=1}^p |\sqrt{p}\betab^\star_{i}|^{2k-2} + \frac{1}{p}\sum_{i=1}^p |\boldsymbol{\Sigma}_{i,i}|^{2k-2} \right)\label{eq:bdmom}\,.
\end{align}
\cts{The terms $\sum_{i=1}^p |\sqrt{p}\betab^\star_{i}|^{2k-2} $ and $ \frac{1}{p}\sum_{i=1}^p |\boldsymbol{\Sigma}_{i,i}|^{2k-2}$ are bounded by assumption}\ct{Need to add assumption that $ \frac{1}{p}\sum_{i=1}^p |\boldsymbol{\Sigma}_{i,i}|^{2k-2}<\infty$ and $\frac{1}{p}\sum_{i=1}^p (\sqrt{p}|\betab^\star_{i}|^{2k-2} $. OK by just assuming that $B$ is Definition \ref{def:Xi} is bounded.}. {\color{red}Furthermore we need to argue that,
\begin{align}
\frac{1}{p}\sum_{i=1}^p |\sqrt{p}\boldsymbol{\beta}_{i}|^{2k-2} <\infty
\end{align}
}
These together with \eqref{eq:bdmom} show that $\frac{1}{p}\sum_{i=1}^p \|\vct{b}_i\|_2^{2k-2}<C$ wpa 1. Similarly, we can argue for $\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2}$, completing the proof of \eqref{eq:leqC}.
\vspace{5pt}
\noindent\underline{Satisfying Condition 1:} To prove Condition 1, we simply pick $\eta$ to satisfy the following
\begin{align}
\bar\phi_S > \bar\phi + 3 \eta ~\Leftarrow~ \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta \geq \bar\phi + 3 \eta ~\Leftarrow~ \eta \leq \frac{\tilde\epsilon^2}{8\Sigma_{\max}}.
\end{align}
This completes the proof of Theorem \ref{thm:master_W2}.
\subsection{Proof of Lemma \ref{lem:AO}}\label{sec:proofAO}
~~~~
~~~~
\vspace{3pt}
\noindent\underline{Proof of (i):}~Strong convexity of the objective function in \eqref{eq:PO} is easily verified by the second derivative test. Note here that we use Assumption \ref{ass:inv} that $\bSi_{i,i}\leq\Sigma_{\max},~i\in[p].$ Uniqueness of the solution follows directly from strong convexity. \ct{Strictly speaking we might need to also argue existence, i.e., feasibility of the AO. An indirect way is to show feasibility using the CGMT, but it seems unnecessarily complicated?}
\vspace{5pt}
\noindent\underline{Proof of (ii):}~Using Lagrangian formulation, the solution $\vct{w}_n$ to \eqref{eq:AO_con} is the same as the solution to the following:
\begin{align}
\left(\vct{w}_n,u_n\right) :=\arg\min_{\vct{w}\in\mathcal{B}}\max_{u\geq 0} ~\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + u \left( \sqrt{\tn{\vct{w}}^2+\sigma^2} \tn{\bar{\g}} - \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} + \frac{\sigma h}{\sqrt{n}} \right)\label{eq:AO_2}
\end{align}
where we have: (i) set $\bar{\g} := {\vct{g}}/\sqrt{n}$ and $\bar{\h}:= \vct{h}/\sqrt{p}$; (ii) recalled that $p/n=\kappa$; and, (iii) used $\left(\vct{w}_n,u_n\right)$ to denote the optimal solutions in \eqref{eq:AO_2}. The subscript $n$ emphasizes the dependence of $\left(\vct{w}_n,u_n\right)$ on the problem dimensions. Also note that (even though not explicit in the notation) $\left(\vct{w}_n,u_n\right)$ are random variables depending on the realizations of $\bar{\g},\bar{\h}$ and $h$.
Notice that the objective function above is convex in $\vct{w}$ and linear (thus, concave) in $u$. Thus, strong duality holds and we can flip the order of min-max. Moreover, in order to make the objective easy to optimize with respect to $\vct{w}$, we use the following variational expression for the square-root term $\sqrt{\tn{\vct{w}}^2+\sigma^2}$:
$$
\tn{\bar{\g}}\sqrt{\tn{\vct{w}}^2+\sigma^2} = \tn{\bar{\g}}\cdot\min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau}{2} + \frac{\tn{\vct{w}}^2+\sigma^2}{2\tau} \right\} = \min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau\tn{\bar{\g}}^2}{2} + \frac{\sigma^2}{2\tau} + \frac{\tn{\vct{w}}^2}{2\tau} \right\},
$$
where $B$ is defined in Lemma \ref{lem:bd_PO}. For convenience define the constraint set for the variable $\tau$ as $\mathcal{T}':=[\sigma,\sqrt{\sigma^2+B_+^2}]$. For reasons to be made clear later in the proof (see proof of statement (iii)), we consider the (possibly larger) set:
\[
\mathcal{T}:=[\sigma,\max\{\sqrt{\sigma^2+B_+^2},2\tau_*\}]\,
\]
where $\tau_*$ is as in the statement of the lemma.
The above lead to the following equivalent formulation of \eqref{eq:AO_2}:
\begin{align}
\left(\vct{w}_n,u_n,\tau_n\right) = \max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} + \min_{\vct{w}} \left\{ \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + \frac{u}{2\tau}\tn{\vct{w}}^2 - u \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} \right\} .\label{eq:AO_3}
\end{align}
\ct{To be fully rigorous, need to show here that the unconstrained min over $\vct{w}$ is the same as the constrained $\vct{w}\in\mathcal{B}$.}
The minimization over $\vct{w}$ is easy as it involves a strongly convex quadratic function. The optimal $\vct{w}':=\vct{w}'(\tau,u)$ (for fixed $(\tau,u)$) is given by
\begin{align}\label{eq:w'}
\vct{w}':=\vct{w}'(\tau,u) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right),
\end{align}
and \eqref{eq:AO_3} simplifies to
\begin{align}
\left(u_n,\tau_n\right)=\max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} - \frac{1}{2} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)\,=:\mathcal{R}(u,\tau) .\label{eq:AO_4}
\end{align}
It can be checked by direct differentiation and the second-derivative test that the objective function in \eqref{eq:AO_4} is strictly convex in $\tau$ and strictly concave in $u$ in the domain $\{(u,\tau)\in\mathbb{R}_+\times\mathbb{R}_+\}$. Thus, the saddle point $(u_n,\tau_n)$ is unique. Specifically, this implies that the optimal $\vct{w}_n$ in \eqref{eq:AO_3} is given by (cf. \eqref{eq:w'})
\begin{align}\label{eq:w_n}
\vct{w}_n=\vct{w}'(\tau_n,u_n) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).
\end{align}
In what follows, we characterize the high-dimensional limit of the optimal pair $(u_n,\tau_n)$ in the limit $n,p\rightarrow\infty,~p/n\rightarrow\kappa$.
We start by analyzing the (point-wise) convergence of $\mathcal{R}(u,\tau)$.
For the first three summands in \eqref{eq:AO_4}, we easily find that
$$
\left\{\frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}}\right\}~ \stackrel{{P}}{\longrightarrow}~\left\{
\frac{u\tau}{2} + \frac{u\sigma^2}{2\tau} \right\}.
$$
Next, we study the fourth summand. First, note that
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}(u\sqrt{\kappa}\bar{\h}) &= u^2\kappa\,\frac{1}{p}\vct{h}^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\vct{h} \notag\\
&= u^2\kappa\,\frac{1}{p}\sum_{i=1}^{p}\frac{\vct{h}_i^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}} \notag\\
&\stackrel{{P}}{\longrightarrow} u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right].
\end{align}
In the last line, $\Lambda$ is a random variable as in Definition \ref{def:Xi}. \cts{Also, we used Assumption \ref{ass:mu} together with the facts that $\vct{h}$ is independent of $\boldsymbol{\Sigma}$ and that the function $(x_1,x_2)\mapsto f(x_1,x_2)=x_1^2(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(2)$ assuming $x_2$ is bounded.} \ct{Question: Is it immediate that the empirical distribution of $(\beta,\boldsymbol{\Sigma},\vct{h})$ converges in $W_k$ to $\mu\otimes\mathcal{N}(0,1)$ given that $(\beta,\boldsymbol{\Sigma})$ converges to $\mu$ and $\vct{h}$ is independent???}
Second, we find that
\begin{align}
(\betab^\star)^T\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star &= \frac{1}{p}(\sqrt{p}\betab^\star)^T\left({\mtx{I}}+\frac{u}{\tau}\boldsymbol{\Sigma}\right)^{-1}(\sqrt{p}\betab^\star) \notag\\
&= \frac{1}{p}\sum_{i=1}^{p}\frac{\left(\sqrt{p}\betab^\star_i/\sqrt{\bSi_{i,i}}\right)^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}}\notag\\
&\stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right].
\end{align}
Here, $\Lambda,B$ are random variables as in Definition \ref{def:Xi} and \cts{we also used Assumption \ref{ass:mu} together with the fact that the function $(x_1,x_2)\mapsto f(x_1,x_2)=x_1^2x_2^{-1}(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(2)$ assuming $x_2$ is bounded.}
Third, \cts{by independence of $(\betab^\star, \boldsymbol{\Sigma})$ from $\vct{h}$}
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star = u\sqrt{\kappa} \cdot \frac{1}{p}\sum_{i=1}
^p \frac{\vct{h}_i\boldsymbol{\Sigma}_{i,i}^{-1/2}(\sqrt{p}\betab^\star_i)}{\boldsymbol{\Sigma}_{i,i}^{-1}+\frac{u}{\tau}{\mtx{I}}} ~\stackrel{{P}}{\longrightarrow}~ 0.
\end{align}
Putting these together, the objective $\mathcal{R}(u,\tau)$ in \eqref{eq:AO_4} converges point-wise in $u,\tau$ to
\begin{align}
\mathcal{R}(u,\tau)\stackrel{{P}}{\longrightarrow}\mathcal{D}(u,\tau) := \frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).\label{eq:conv_pt}
\end{align}
Note that $\mathcal{R}(u,\tau)$ (and thus, $\mathcal{D}(u,\tau)$) is convex in $\tau$ and concave in $u$. Thus, the convergence in \eqref{eq:conv_pt} is in fact uniform (e.g., \cite{AG1982}) and we can conclude that
\begin{align}\label{eq:Dc0}
\phi({\vct{g}},\vct{h}) \stackrel{{P}}{\longrightarrow} \max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
and
\begin{align}\label{eq:Dc}
(u_n,\tau_n) \stackrel{{P}}{\longrightarrow} (u_*,\tau_*):=\arg\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
In the proof of statement (iii) below, we show that the saddle point of \eqref{eq:Dc0} is $(u_*,\tau_*)$. In particular, $\tau_*$ is strictly in the interior of $\mathcal{T}$, which combined with convexity implies that
$$
\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau) = \max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau) =: \bar\phi.
$$
This, together with the first display above proves the second statement of the lemma.
\vspace{5pt}
\noindent\underline{Proof of (iii):}
Next, we compute the saddle point $(u_*,\tau_*)$ by studying the first-order optimality conditions of the srtictly concave-convex $\mathcal{D}(u,\tau)$. Specifically, we consider unconstrained minimization over $\tau$ and we will show that the minimum is achieved in the strict interior of $\mathcal{T}$. Direct differentiation gives
\begin{subequations}
\begin{align}
{\tau} + \frac{\sigma^2}{\tau} - 2u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] + \frac{u^2}{\tau}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] + \frac{1}{\tau}\operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0, \label{eq:fo1}\\
{u} - \frac{u\sigma^2}{\tau^2} - \frac{u^3}{\tau^2}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} + \frac{u}{\tau}\right)^2}\right] - \frac{u}{\tau^2} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0,\label{eq:fo2}
\end{align}
\end{subequations}
Multiplying \eqref{eq:fo2} with $\frac{\tau}{u}$ and adding to \eqref{eq:fo1} results in the following equation
\begin{align}
\tau = u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] ~\Leftrightarrow ~
\operatorname{\mathbb{E}}\left[ \frac{1}{(\frac{u}{\tau}\Lambda)^{-1}+1} \right] = \frac{1}{\kappa} \label{eq:tauu}\,.
\end{align}
Thus, we have found that the ratio $\frac{u_*}{\tau_*}$ is the unique solution to the equation in \eqref{eq:tauu}. Note that this coincides with the Equation \eqref{eq:ksi} that defines the parameter $\xi$ in Definition \ref{def:Xi}. The fact that \eqref{eq:tauu} has a unique solution for all $\kappa>1$ can be easily seen as $F(x)=\operatorname{\mathbb{E}}\left[ \frac{1}{(x\Lambda)^{-1}+1} \right], x\in\mathbb{R}_+$ has range $(0,1)$ and is strictly increasing (by differentiation).
Thus, we call $\xi=\frac{u_*}{\tau_*}$. Moreover, multiplying \label{eq:fo2} with $u$ leads to the following equation for $\tau_*$:
\begin{align}
u_*^2 = \sigma^2\xi^2 + u_*^2 \xi^2 \kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right] + \xi^2 \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right] ~\Rightarrow~ \tau_*^2 = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\xi^2\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right]} = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left((\xi\Lambda)^{-1} +1\right)^2}\right]}.
\end{align}
{Again, note that this coincides with Equation \eqref{eq:gamma} that determines the parameter $\gamma$ in Definition \ref{def:Xi}, i.e., $\tau_*^2 = \gamma.$ }
\vspace{3pt}\noindent\underline{Proof of (iv):}
For convenience, define
$$F_n(\hat\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}\hat\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Recall from \eqref{eq:w_n} the explicit expression for $\vct{w}_n$, repeated here for convenience.
\begin{align}
\vct{w}_n = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).\notag
\end{align}
Also, recall that $\boldsymbol{\beta}_n = \boldsymbol{\Sigma}^{-1/2}\vct{w}_n+\betab^\star$. Thus, (and using the fact that $\bar{\h}$ is distributed as $-\bar{\h}$),
\begin{align}
\boldsymbol{\beta}_n &=\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}u_n\sqrt{\kappa}\bar{\h} + \left({\mtx{I}}-\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\right)\betab^\star\notag\\
\Longrightarrow\boldsymbol{\beta}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\bar{\h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i.
\end{align}
For $i\in[p]$, define
\begin{align}
\vct{v}_{n,i} = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\bar{\h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i
\end{align}
In the above, for convenience, we have denoted $\xi_n:=u_n/\tau_n$ and recall that $\xi_*:=u_*/\tau_*$.
The proof proceeds in two steps. In the first step, we use the fact that $\xi_n\stackrel{{P}}{\longrightarrow}\xi_*$ and $u_n\stackrel{{P}}{\longrightarrow} u_\star$ (see \eqref{eq:Dc}) to prove that for any $\varepsilon\in(0,\xi_*/2)$, there exists absolute constant $C>0$ such that wpa 1:
\begin{align}\label{eq:ivstep1}
|F_n(\sqrt{p}\boldsymbol{\beta}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.
\end{align}
In the second step, we use Lipschitzness of $f$ and Assumption \ref{ass:mu} to prove that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:ivstep2}
\end{align}
The desired follows by combining \eqref{eq:ivstep1} and \eqref{eq:ivstep2}. Thus, in what follows, we prove \eqref{eq:ivstep1} and \eqref{eq:ivstep2}.
\vspace{2pt}
\noindent\emph{Proof of \eqref{eq:ivstep1}.}~~Fix some $\varepsilon\in(0,\xi_*/2)$. From \eqref{eq:Dc}, we know that w.p.a. 1 $|\xi_n-\xi_*|\leq \varepsilon$ and $|u_n-u_*|\leq \varepsilon$. Thus, $\vct{w}_n$ is close to $\vct{v}_n$. Specifically, in this event, for every $i\in[p]$, it holds that:
\begin{align}
\notag|\boldsymbol{\beta}_{n,i} - \vct{v}_{n,i}| &\leq {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{\tau_n}{1+(\xi_n\bSi_{i,i})^{-1}}-\frac{\tau_*}{1+(\xi_*\bSi_{i,i})^{-1}}\right| \\
\notag&= {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{u_n}{\xi_n+\bSi_{i,i}^{-1}}-\frac{u_*}{\xi_*+\bSi_{i,i}^{-1}}\right| \\
\notag& \leq
{|\betab^\star_i|}\frac{|\bSi_{i,i}||\xi_n-\xi_*|}{|1+\xi_n\bSi_{i,i}||1+\xi_*\bSi_{i,i}|} + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}} \frac{u_*|\xi_n-\xi_*|}{(\xi_n+\bSi_{i,i}^{-1})(\xi_*+\bSi_{i,i}^{-1})} +
\sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}}\frac{|u_n-u_*|}{\xi+\bSi_{i,i}^{-1}}
\\
\notag& \leq
{|\betab^\star_i|}{\Sigma_{\max}\varepsilon} + \sqrt{\kappa}{|\bar{\h}_i|} u_*\Sigma_{\max}^{3/2} \varepsilon+
\sqrt{\kappa}|\bar{\h}_i|\Sigma_{\max}^{1/2}\varepsilon
\\
&\leq \varepsilon \cdot \max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\, \left(|\bar{\h}_i| + |\betab^\star_i|\right).\label{eq:betav}
\end{align}\som{$\varepsilon$ missing}
In the second line above, we recalled that $u_n=\tau_n\xi_n$ and $u_*=\tau_*\xi_*$. In the third line, we used the triangle inequality. In the fourth line, we used that $\xi_*>0$, $0<\bSi_{i,i}\leq\Sigma_{\max}$ and $\xi_n\geq \xi_*-\varepsilon \geq \xi_*/2 >0$.
Now, we will use this and Lipschitzness of $f$ to argue that there exists absolute constant $C>0$ such that wpa 1:
\begin{align}
|F_n(\sqrt{p}\boldsymbol{\beta}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.\notag
\end{align}
Denote, $\vct{a}_i=(\sqrt{p}\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\vct{v}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$. Following the exact same argument as in \eqref{eq:LipC}, we have
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| &\leq \frac{L}{p}\sum_{i=1}^p (1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\})\|\vct{a}_i-\vct{b}_i\|_2\notag \\
&\leq \frac{L}{p}\sum_{i=1}^p \left(1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\}\right)\cdot|\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}) - \sqrt{p}\vct{v}_{n,i}|\notag\\
&\leq L\left(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} \right)^{1/2}{\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \vct{v}_n\|_2}\notag.
\end{align}
From this and \eqref{eq:betav}, we find
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| &\leq
L(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} )^{1/2}
\\
&\qquad\qquad\varepsilon\cdot\sqrt{2}\max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\sqrt{\|\betab^\star\|_2^2 + \|\bar{\h}\|_2^2}.\label{eq:epsS}
\end{align}
As in \eqref{eq:bdmom}, it can be shown that wpa 1: $\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2}<\infty$ and $\frac{1}{p}\sum_{i=1}^p \|\vct{b}_i\|_2^{2k-2}<\infty$, as $p\rightarrow\infty$. Similarly, $\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^2<\infty$, as $p\rightarrow \infty$ {\color{red} by assumption on second moment convergence of $\sqrt{p}\betab^\star$.} Finally, $\|\bar{\h}\|_2^2\leq 2$, wpa 1 as $p\rightarrow\infty$. Therefore, from \eqref{eq:epsS}, wpa 1, there exists constant $C>0$ such that
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \leq C \cdot\varepsilon,
\end{align}
as desired.
\vspace{2pt}
\noindent\emph{Proof of \eqref{eq:ivstep2}.}~~Next, we will use Assumption \ref{ass:mu} to show that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:AO_conv}
\end{align}
Notice that $\vct{v}_n$ is a function of $\betab^\star,\boldsymbol{\Sigma},\bar{\h}$. Concretely, define $g:\mathbb{R}^3\rightarrow\mathbb{R}$, such that
$$
g(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1,
$$
and notice that
$$
\sqrt{p}\vct{v}_{n,i} = g\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right) = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\vct{h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\sqrt{p}\betab^\star_i.
$$
Thus,
$$
F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma}) = \frac{1}{p}\sum_{i=1}^p f\left(g\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right),\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) =: \frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right),
$$
where we have defined $h:\mathbb{R}^3\rightarrow\mathbb{R}$:
\begin{align}
h(x_1,x_2,x_3) := f\left(g(x_2,x_3,x_1),x_2,x_3\right).\label{h func}
\end{align}
It will suffice to prove that $h\in\rm{PL}(k)$. Indeed, if that were the case, then Assumption \ref{ass:mu} would give
\begin{align}
\frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[h(H,B,\Lambda)\right] &= \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[f\left(g(B,\Lambda,H),B,\Lambda\right))\right]\\
& = \operatorname{\mathbb{E}}[f\left(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda\right)] = \alpha_*,
\end{align}
where the penultimate equality follows by recognizing that (cf. Eqn. \eqref{eq:X})
$$
g(B,\Lambda,H) = (1-(1+\xi_*\Lambda)^{-1})B + \sqrt{\kappa}\frac{\tau_*\Lambda^{-1/2}}{1+(\xi_*\Lambda)^{-1}}H = X_{\kappa,\sigma^2}(\Lambda,B,H).
$$
In the remaining of the proof, \cts{we show that $h\in\rm{PL}(k)$} which is formalized below.
\begin{lemma} The function $h$ in \eqref{h func} is $\rm{PL}(k)$ as long as $f$ is $\rm{PL}(k)$.
\end{lemma}
\begin{proof} Since $f$ is $\rm{PL}(k)$, for some $L>0$, \eqref{PL func} holds. Fix $\vct{a}=[\vct{x},\vct{y},{\vct{z}}]\in\mathbb{R}^{3p},\vct{a}'=[\vct{x}',\vct{y}',{\vct{z}}']\in\mathbb{R}^{3p}$. Let $\vct{b}=[{\vct{g}},\vct{y},{\vct{z}}]$ where ${\vct{g}}=g(\vct{y},{\vct{z}},\vct{x})$. We have that
\begin{align}
|h(\vct{a})-h(\vct{a}')|&\leq |f(\vct{b})-f(\vct{b}')|\\
&\leq L\left(1+\|\vct{b}\|_2^{k-1}+\|\vct{b}'\|_2^{k-1}\right)\|\vct{b}-\vct{b}'\|_2\\
&\lesssim L\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+\|{\vct{g}}\|_2^{k-1}+\|{\vct{g}}'\|_2^{k-1}\right)(\|\vct{a}-\vct{a}'\|_2+\|{\vct{g}}-{\vct{g}}'\|_2).
\end{align}
Next, we need to bound the ${\vct{g}}$ term in terms of $\vct{a}$. This is accomplished as follows
\begin{align}
\|{\vct{g}}\|_2^{k-1}&\lesssim \frac{1}{p}\sum_{i=1}^p|g_i|^{k-1}\\
&\lesssim \frac{1}{p}\sum_{i=1}^p\left|\frac{y_i^{-1/2}}{1+(\xi y_i)^{-1}}\sqrt{\kappa}\tau_* z_i + (1-(1+\xi_*y_i)^{-1})x_i\right|\\
&\lesssim \frac{1}{p}\sum_{i=1}^p(|z_i|+|x_i|)^{k-1}\lesssim \|\vct{x}\|_2^{k-1}+\|{\vct{z}}\|_2^{k-1}\\
&\lesssim \|\vct{a}\|^{k-1}.
\end{align}
Secondly and similarly, we have the following perturbation bound on the ${\vct{g}}-{\vct{g}}'$. \so{Suppose the triples $(x_i,y_i,z_i)$ takes values in a fixed bounded compact set $\mathcal{M}$. We will prove the following sequence of inequalities}
\begin{align}
\tn{{\vct{g}}-{\vct{g}}'}^2&\leq \sum_{i=1}^p|g(x_i,y_i,z_i)-g(x'_i,y'_i,z'_i)|^2\\
&\leq C^2_{\mathcal{M}} \sum_{i=1}^p(|x_i-x'_i|^2+|y_i-y'_i|^2+|z_i-z'_i|^2)\label{sec line eq}\\
&\leq C^2_{\mathcal{M}} (\tn{\vct{x}-\vct{x}'}^2+\tn{\vct{y}-\vct{y}'}^2+\tn{{\vct{z}}-{\vct{z}}'}^2)\\
&\lesssim \tn{\vct{a}-\vct{a}'}^2.
\end{align}
In what follows, we prove the second line \eqref{sec line eq} i.e.~the fact that for any triples $a=(x,y,z),a'=(x',y',z')$ (with $a,a'\in\mathbb{R}^3$), we have that
\[
|g(a)-g(a')|\leq C_{\mathcal{M}}\tn{a-a'}.
\]
Set $C_\nabla=\sup_{a\in \mathcal{M}}\tn{\nabla g(a)}$. By definition of gradient, the inequality above holds with $C_\mathcal{M}=C_\nabla$. Thus, all that remains is proving that $C_\nabla$ is upper bounded by a constant. \so{However, this automatically holds because from the definition of $g(\dots)$ function, it is clear that $C_\nabla$ is defined everywhere and continuous thus it has a finite maximum over a compact set.}
\end{proof}
\section{Facts about Lipschitz functions}\label{SM useful fact}
For $k\geq 1$ we say a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$ and denote it by $f\in \rm{PL}(k)$ if there exists a cosntant $L>0$ such that, for all $\vct{x},\vct{y}\in\mathbb{R}^m$:
\begin{align}
|f(\vct{x})-f(\vct{y})|\leq L\left(1+\|\vct{x}\|_2^{k-1}+\|\vct{y}\|_2^{k-1}\right)\|\vct{x}-\vct{y}\|_2.\label{PL func}
\end{align}
In particular, when $f\in\rm{PL}(k)$, the following properties hold:
\begin{enumerate}
\item There exists a constant $L'$ such that for all $\vct{x}\in\mathbb{R}^n$: $|f(\vct{x})|\leq L'(1+\|\vct{x}\|_2^k).$
\item $f$ is locally Lipschitz, that is for any $M>0$, there exists a constant $L_{M,m}<\infty$ such that for all $x,y\in[-M,M]^m$,
$
|f(\vct{x})-f(\vct{y})| \leq L_{M,m}\|\vct{x}-\vct{y}\|_2.
$
Further, $L_{M,m}\leq c(1+(M\sqrt{m})^{k-1})$ for some costant $c$.
\end{enumerate}
\begin{lemma}
Let $g:\mathbb{R}\rightarrow\mathbb{R}$ be a bounded, Lipschitz function. Consider the function $f:\mathbb{R}^3\rightarrow\mathbb{R}$ defined as follows:
$$
f(\vct{x}) = x_1(x_2-g(x_3))^2.
$$
Then, $f\in\rm{PL}(3).$
\end{lemma}
\begin{proof}
Let $h:\mathbb{R}^2\rightarrow\mathbb{R}$ defined as $h({\vct{u}})=({\vct{u}}_1-g({\vct{u}}_2))^2$. The function $({\vct{u}}_1,{\vct{u}}_2)\mapsto{\vct{u}}_1-g({\vct{u}}_2)$ is clearly Lipschitz. Thus, $h\in\rm{PL}(2)$, i.e.,
\begin{align}
|h({\vct{u}})-h(\vct{v})| \leq C(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2\quad\text{and}\quad |h(\vct{v})|\leq C'(1+\|\vct{v}\|_2^2).\label{eq:h_pl}
\end{align}
Thus, letting $\vct{x}=(x_1,{\vct{u}})\in\mathbb{R}^3$ and $\vct{y}=(y_1,\vct{v})\in\mathbb{R}^3$, we have that
\begin{align}
|f(\vct{x})-f(\vct{y})| &= |x_1h({\vct{u}}) - y_1h(\vct{v})| \leq |x_1||h({\vct{u}})-h(\vct{v})| + |h(\vct{v})| |x_1-y_1|\notag\\
&\leq C|x_1|(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(|x_1|^2+(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|\vct{x}-\vct{y}\|_2.
\end{align}
In the second line, we used \eqref{eq:h_pl}. In the third line, we used $2xy\leq x^2+y^2$. In the fourth line, we used Cauchy-Schwarz inequality. $C,C'>0$ are absolute constants that may change from line to line.
This completes the proof of the lemma.
\end{proof}
\begin{lemma}
Let functions $f,g:\mathbb{R}^3\rightarrow\mathbb{R}$ such that $f\in\rm{PL}(3)$ and
$$
g(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi_* x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1.
$$
Here, $\xi_*,\tau_*,\kappa$ are positive constants. Also, $x_2$ is bounded, say $x_2\in[m,M]\subset(0,\infty)$.
Further define
\begin{align}
h(x_1,x_2,x_3) := f\left(g(x_2,x_3,x_1),x_2,x_3\right).
\end{align}
Then, it holds that $h\in\rm{PL}(k+?)$.
\end{lemma}
\section{Proofs for overparameterized least-squares}\label{sec proof thm 1}
In this section, we assume the linear Gaussian problem (LGP) of Definition \ref{def LGP}, the overparameterized regime $k=p>n$ and the min-norm model $\hat\boldsymbol{\beta}$ of \eqref{eq:min_norm}. We prove Theorem \ref{thm:master_W2} that derives the asymptotic DC of $\hat\boldsymbol{\beta}$ and we show how this leads to sharp formulae for the risk of the Magnitude- and Hessian-pruned models.
\subsection{Notation and Assumptions}\label{sec:ass_app}
For the reader's convenience, we recall some necessary notation and assumptions from Section \ref{sec main}. We say that a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$, denoted $f\in\rm{PL}(k)$, if there is a constant $L>0$ such that for all $\vct{x},\vct{y}\in\mathbb{R}^m$,
$
|f(\vct{x}) - f(\vct{y})|\leq L(1+\tn{\vct{x}}^{k-1}+\tn{\vct{y}}^{k-1})\|\vct{x}-\vct{y}\|_2
$
(See also Section \ref{SM useful fact}).
We say that a sequence of probability distributions $\nu_p$ on $\mathbb{R}^m$ {converges in $W_k$} to
$\nu$, written $\nu_p\stackrel{W_k}{\Longrightarrow} \nu$, if $W_k(\nu_p,\nu) \rightarrow 0$ as $p \rightarrow \infty$. An equivalent definition is that, for any $f\in\rm{PL}(k)$, $\lim_{p\rightarrow\infty}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$ (e.g., \cite{montanari2017estimation}).
Finally, recall that a sequence of probability distributions $\nu_n$ on $\mathbb{R}^m$ \emph{converges weakly} to $\nu$, if for any bounded Lipschitz function $f$: $\lim_{p\rightarrow\infty}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$.
Throughout, we use $C,C',c,c'$ to denote absolute constants (not depending on $n,p$) whose value might change from line to line.
We focus on a double asymptotic regime where:
$$n,p,s\rightarrow\infty \text{ at fixed overparameterization ratio } \kappa:=p/n>1 \text{ and sparsity level } \alpha:=s/p\in(0,1).$$
For a sequence of random variables $\mathcal{X}_{p}$ that converge in probability to some constant $c$ in the limit of Assumption \ref{ass:linear} below, we write $\mathcal{X}_{p}\stackrel{{P}}{\longrightarrow} c$. For a sequence of event $\mathcal{E}_p$ for which $\lim_{p\rightarrow}\mathbb{P}(\mathcal{E}_p) = 1$, we say that $\mathcal{E}_p$ occurs \emph{with probability approaching 1}. For this, we will often use the shorthand ``wpa 1".
\vspace{5pt}
Next, we recall the set of assumption under which our analysis applies:
\vspace{3pt}
\textbf{(A1)[Diagonal covariance].}~~The covariance matrix $\boldsymbol{\Sigma}$ is diagonal.
\vspace{3pt}
\textbf{(A2)[Boundedness and empirical distribution].}~~There exist constants $\Sigma_{\min},\Sigma_{\max}\in(0,\infty)$ such that:
$
\Sigma_{\min}\leq{\boldsymbol{{\Sigma}}}_{i,i}\leq \Sigma_{\max}.
$
for all $i\in[p].$
\cts{Furthermore, the joint empirical distribution of $\{(\bSi_{i,i},\sqrt{p}\betab^\star_i)\}_{i\in[p]}$ converges weakly to a probability distribution $\mu$ on $\mathbb{R}_{>0}\times\mathbb{R}$ with bounded $4$th moment
and assume that as $p\rightarrow\infty$, the $4$th moment of the empirical distribution converges to the $4$th moment of $\mu$, i.e.,
$$
\frac{1}{p}\sum_{i=1}^p\left(\sqrt{(\sqrt{p}\betab^\star_i)^{2}+\bSi_{i,i}^2}\right)^4 \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(\Lambda,B)\sim\mu}\left[\left(\sqrt{\Lambda^2+B^2}\right)^4\right]<\infty.
$$
}
\noindent We remark that Assumption (A2) above implies (see \cite[Lem.~4]{bayati2011dynamics} and \cite[Lem.~A3]{javanmard2013state}) that for any pseudo-Lipschitz function $\psi:\mathbb{R}^2\rightarrow\mathbb{R}$ of order $4$, i.e., $\psi\in\rm{PL}(4)$:
$$
\frac{1}{p}\sum_{i=1}^p\psi(\bSi_{i,i},\sqrt{p}\betab^\star_i) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(\Lambda,B)\sim\mu}\left[ \psi(\Lambda,B)\right].
$$
\subsection{Asymptotic distribution and risk characterizations}
In this section, we prove the following main result, which is a special case of Theorem \ref{thm:master_W2}, tailored to our specific need in this paper, that is, characterizing the risk of pruned solutions.
Let ${\boldsymbol{\hat{\beta}}}^P={\cal{P}}({\boldsymbol{\hat{\beta}}})$ be a pruned version of the min-norm solution ${\boldsymbol{\hat{\beta}}}$. Recall from Section \ref{sec:risk}, that the first crucial step in characterizing the risk ${\cal{L}}({\boldsymbol{\hat{\beta}}}^P)$ is studying the risk of a threshold-based pruned vector ${\cal{L}}({\boldsymbol{\hat{\beta}}}^\mathcal{T}_t)$. Theorem \ref{thm:app_W2} below shows how this is possible.
To keep things slightly more general, consider ${\boldsymbol{\hat{\beta}}}^{g}$ defined such that $\sqrt{p}{\boldsymbol{\hat{\beta}}}^{g}=g(\sqrt{p}{\boldsymbol{\hat{\beta}}})$, where $g$ is a Lipschitz function acting entry-wise on ${\boldsymbol{\hat{\beta}}}$ (for example, $g$ can be the \fx{(arbitrarily close Lipschitz approximation of the)} thresholding operator $\mathcal{T}_t$ of Section \ref{sec:risk}). Then, the risk of ${\boldsymbol{\hat{\beta}}}^g$ can be written as
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^g) &= \operatorname{\mathbb{E}}_{\mathcal{D}}[(\vct{x}^T({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^g) + \sigma z)^2] = \sigma^2 + ({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^g)^T\boldsymbol{\Sigma}({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^g) \notag \\
&= \sigma^2 + \frac{1}{p}\sum_{i=1}^p\boldsymbol{\Sigma}_{i,i}\big(\sqrt{p}\betab^\star_i-g(\sqrt{p}\hat\boldsymbol{\beta}_i)\big)^2\notag
\\
&=: \sigma^2 + \frac{1}{p}\sum_{i=1}^p f\big(\sqrt{p}{\boldsymbol{\hat{\beta}}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}\big),\label{eq:risk_app_f}
\end{align}
where in the last line, we defined function $f:\mathbb{R}^3\rightarrow\mathbb{R}$ as follows:
\begin{align}\label{eq:fdef}
f_{\cal{L}}(x,y,z) := \fx{z(y-g(x))}^2.
\end{align}
The following theorem establishes the asymptotic limit of \eqref{eq:risk_app_f}. For the reader's convenience, we repeat the notation introduced in Definition \ref{def:Xi}.
Let random variables $(\Lambda,B)\sim \mu$ (where $\mu$ is defined in Assumption (A2)) and fix $\kappa>1$. Define parameter $\xi$ as the unique positive solution to the following equation
$$
\operatorname{\mathbb{E}}_{\mu}\Big[ \big({1+(\xi\cdot\Lambda)^{-1}}\big)^{-1} \Big] = {\kappa^{-1}}\,.
$$
Further define the positive parameter $\gamma$ as follows:
$$
\hspace{-0.1in}\gamma :=
\Big({\sigma^2 + \operatorname{\mathbb{E}}_{\mu}\Big[\frac{B^2\Lambda}{(1+\xi\Lambda)^2}\Big]}\Big)\Big/\Big({1-\kappa\operatorname{\mathbb{E}}_{\mu}\Big[\frac{1}{\left(1+(\xi\Lambda)^{-1}\right)^2}\Big]}\Big).
$$
With these and $H\sim\mathcal{N}(0,1)$, define the random variable
$$
X_{\kappa,\sigma^2}:=X_{\kappa,\sigma^2}(\Lambda,B,H) := \Big(1-\frac{1}{1+ \xi\Lambda}\Big) B + \sqrt{\kappa}\frac{\sqrt{\gamma}\,\Lambda^{-1/2}}{1+(\xi\Lambda)^{-1}} H,
$$
and let $\Pi_{\kappa,\sigma^2}$ be its distribution.
\cts{
\begin{theorem}[Asymptotic DC -- Overparameterized LGP]\label{thm:app_W2}
Fix $\kappa>1$. Let Assumptions (A1) and (A2) in Section \ref{sec:ass_app} hold for some $k\geq 3$. Consider $\hat{{\boldsymbol{\beta}}}$ as in \eqref{eq:min_norm} and $\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p}\sum_{i=1}^{p}\delta_{\sqrt{p}\hat{{\boldsymbol{\beta}}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}}$, the joint empirical distribution of $(\sqrt{p}\hat\boldsymbol{\beta},\sqrt{p}\betab^\star,\boldsymbol{\Sigma})$. Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be any of the following two: (a) $f\in\rm{PL}(2)$, or, (b) $f=f_{\cal{L}}$ defined in \eqref{eq:fdef} \fx{where $g$ is a Lipschitz function}. In either case, it holds that
\begin{align}\label{eq:thm_app_f}
\frac{1}{p} \sum_{i=1}^{p} f(\sqrt{p}\hat\boldsymbol{\beta}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)}\left[f(X_{\kappa,\sigma^2},B,\Lambda) \right].
\end{align}
\end{theorem}
}
Before we prove the theorem, let us show how it immediately leads to a sharp prediction of the risk behavior. Indeed, a direct application of \eqref{eq:thm_app_f} for $f=f_{\cal{L}}$ to \eqref{eq:risk_app_f} shows that
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^g)\stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}_{(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)}\left[f_{\cal{L}}(X_{\kappa,\sigma^2},B,\Lambda) \right] = \sigma^2 + \operatorname{\mathbb{E}}_{(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)}\left[\Sigma\left(B-g(X_{\kappa,\sigma^2})\right)^2 \right].\label{eq:risk_app_f2}
\end{align}
We further remark on the following two consequences of Theorem \ref{thm:app_W2}.
First, since \eqref{eq:thm_app_f} holds for any $\rm{PL}(2)$ function, we have equivalently that $\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma})$ converges in Wasserstein-2 distance to $\Pi_{\kappa,\sigma^2}\otimes\mu$, where recall that $\Pi_{\kappa,\sigma^2}$ is the distribution of the random variable $X_{\kappa,\sigma^2}$.
Second, the theorem implies that the empirical distribution of $\sqrt{p}\boldsymbol{\beta}_n$ converges weakly to $\Pi_{\kappa,\sigma^2}$. To see this, apply \eqref{eq:thm_app_f} for the $\rm{PL}(2)$ function $f(x,y,z) = \psi(x)$ where $\psi:\mathbb{R}\rightarrow\mathbb{R}$ is a bounded Lipschitz test function.
\subsection{Proof of Theorem \ref{thm:app_W2}}
\vspace{5pt}
Let ${\mtx{X}}\in\mathbb{R}^{n\times p}$ have zero-mean and normally distributed rows with a diagonal covariance matrix ${\boldsymbol{{\Sigma}}}=\operatorname{\mathbb{E}}[\vct{x}\x^T]$. Given a ground-truth vector $\betab^\star$ and labels $\vct{y}={\mtx{X}}\betab^\star+\sigma {\vct{z}},~{\vct{z}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, we consider the least-squares problem subject to the minimum Euclidian norm constraint (as $\kappa=p/n>1$) given by
\begin{align}\label{eq:PO_beta}
\min_{{\boldsymbol{\beta}}}\frac{1}{2}\tn{{\boldsymbol{\beta}}}^2\quad\text{subject to}\quad \vct{y}={\mtx{X}}{\boldsymbol{\beta}}.
\end{align}
It is more convenient to work with the following change of variable:
\begin{align}\label{eq:w}
\vct{w}:=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\beta}}-\betab^\star).
\end{align}
With this, the optimization problem in \eqref{eq:min_norm} can be rewritten as
\begin{align}\label{eq:PO}
\Phi({\mtx{X}})=\min_{\vct{w}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
where we write ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$. First, using standard arguments, we show that the solution of \eqref{eq:PO} is bounded. Hence, we can constraint the optimization in a sufficiently large compact set without loss of generality.
\begin{lemma}[Boundedness of the solution]\label{lem:bd_PO}
Let $\hat\vct{w}_n:=\hat\vct{w}_n({\mtx{X}},{\vct{z}})$ be the minimizer in \eqref{eq:PO}. Then, with probability approaching 1, it holds that $\hat\vct{w}_n\in\mathcal{B}$, where
$$\mathcal{B}:=\left\{\vct{w}\,|\,\|\vct{w}\|_2\leq B_{+} \right\},\qquad B_+:=5\sqrt{\frac{\Sigma_{\max}}{\Sigma_{\min}}}\frac{\sqrt{\kappa}+1}{\sqrt{\kappa}-1}(\sqrt{\Sigma_{\max}\operatorname{\mathbb{E}}\left[B^2\right]} + \sigma).
$$
\end{lemma}
\begin{proof}
First, we show that the min-norm solution $\hat\boldsymbol{\beta}={\mtx{X}}^T({\mtx{X}}\X^T)^{-1}\vct{y}$ of \eqref{eq:PO_beta} is bounded. Note that $\kappa>1$, thus ${\mtx{X}}\X^T$ is invertible wpa 1. We
have,
\begin{align}
\tn{\hat\boldsymbol{\beta}_n}^2 = \vct{y}^T({\mtx{X}}\X^T)^{-1}\vct{y} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{X}}\X^T)} = \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\boldsymbol{\Sigma}{\mtx{\bar{X}}}^T)} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\Xb^T)\,\Sigma_{\min}} = \frac{\tn{\vct{y}}^2}{\sigma_{\min}^2({\mtx{\bar{X}}})\,\Sigma_{\min}}. \label{eq:Ubb}
\end{align}
But, wpa 1,
$
\sigma_{\min}({\mtx{\bar{X}}})/\sqrt{n} \geq \frac{1}{2}\left(\sqrt{\kappa}-1\right).
$
Furthermore,
$
\|\vct{y}\|_2 \leq \|{\mtx{\bar{X}}}\boldsymbol{\Sigma}^{1/2}\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2 \leq \sigma_{\max}({\mtx{\bar{X}}})\sqrt{\Sigma_{\max}} \|\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2.
$
Hence, wpa 1,
$$
\|\vct{y}\|_2/\sqrt{n} \leq 2(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma,
$$
where we used the facts that wpa 1: $\|z\|_2/\sqrt{n}\stackrel{{P}}{\longrightarrow} 1$, $\sigma_{\max}({\mtx{\bar{X}}})<\sqrt{2n}(\sqrt{\kappa}+1)$ and \cts{by Assumption (A2)}:
$$
\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^{p}(\sqrt{p}\betab^\star_i)^2 \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[B^2\right].
$$
Put together in \eqref{eq:Ubb}, shows that
\begin{align}
\tn{\hat\boldsymbol{\beta}_n} < \frac{2(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma}{\sqrt{\Sigma_{\min}}(\sqrt{\kappa}-1)/2} =: \tilde{B}_+.\label{eq:bd_beta}
\end{align}
Recalling that $\hat\vct{w}_n= \sqrt{\boldsymbol{\Sigma}}\boldsymbol{\beta} - \sqrt{\boldsymbol{\Sigma}}\betab^\star$, we conclude, as desired, that wpa 1,
$
\tn{\hat\vct{w}_n} \leq \sqrt{\Sigma_{\max}}\tilde{B}_+ + \sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}[B^2]} \leq B_+.
$
\end{proof}
Lemma \ref{lem:bd_PO} implies that nothing changes in \eqref{eq:PO} if we further constrain $\vct{w}\in\mathcal{B}$ in \eqref{eq:PO}. Henceforth, with some abuse of notation, we let
\begin{align}\label{eq:PO_bd}
\Phi({\mtx{X}}):=\min_{\vct{w}\in\mathcal{B}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
Next, in order to analyze the primary optimization (PO) problem in \eqref{eq:PO_bd} in apply the CGMT \cite{thrampoulidis2015lasso}. Specifically, we use the constrained formulation of the CGMT given by Theorem \ref{thm closed}. Specifically, the auxiliary problem (AO) corresponding to \eqref{eq:PO_bd} takes the following form with ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$
\begin{align}
\phi({\vct{g}},\vct{h}) = \min_{\vct{w}\fx{\in\mathcal{B}}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_con}
\end{align
We will prove the following techincal result about the AO problem.
\somm{Re-specify the assumptions of this lemma!}
\begin{lemma}[Properties of the AO -- Overparameterized regime]\label{lem:AO}
Let $\phi_n=\phi({\vct{g}},\vct{h})$ be the optimal cost of the minimization in \eqref{eq:AO_con}. Define $\bar\phi$ as the optimal cost of the following deterministic min-max problem
\begin{align}\label{eq:AO_det}
\bar\phi:=\max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau):=\frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{N^2}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).
\end{align}
The following statements are true.
\noindent{(i).}~The AO minimization in \eqref{eq:AO_con} is $\frac{1}{\Sigma_{\max}}$-strongly convex and has a unique minimizer $\vct{w}_n:=\vct{w}_n({\vct{g}},\vct{h})$.
\noindent{(ii).}~In the limit of $n,p\rightarrow\infty, p/n=\kappa$, it holds that $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$, i.e., for any $\varepsilon>0$:
$$
\lim_{n\rightarrow\infty}\P\left(|\phi({\vct{g}},\vct{h})-\bar\phi|>\varepsilon\right) = 0.
$$
\noindent{(iii).} The max-min optimization in \eqref{eq:AO_det} has a unique saddle point $(u_*,\tau*)$ satisfying the following:
$$
u_*/\tau_* = \xi\quad\text{and}\quad\tau_* = \gamma,
$$
where $\xi, \gamma$ are defined in Definition \ref{def:Xi}.
\noindent{(iv).}~Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be a $\rm{PL}({k})$ function. Let $\boldsymbol{\beta}_n=\boldsymbol{\Sigma}^{-1/2}\vct{w}_n + \betab^\star$. Then,
$$
\frac{1}{p}\sum_{i=1}^{p}f\left(\sqrt{p}\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\sqrt{p}\betab^\star,\boldsymbol{\Sigma}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(B,\Lambda,H)\sim\mu\otimes \mathcal{N}(0,1)}\left[f\left(X_{\kappa,\sigma^2}(B,\Lambda,H),B,\Lambda\right) \right].
$$
\noindent{(v).}~\cts{The empirical distribution of $\boldsymbol{\beta}_n$ converges weakly to the measure of $X_{\kappa,\sigma^2}$, and also, for large enough absolute constant $C>0$:
\begin{align}\label{eq:k_AO}
\frac{1}{p}\sum_{i=1}^{p}|\sqrt{p}\boldsymbol{\beta}_n({\vct{g}},\vct{h})|^{2} < C.
\end{align}
}
\end{lemma}
We prove Lemma \ref{lem:AO} in Section \ref{sec:proofAO}. Here, we show how this leads to the proof of Theorem \ref{thm:master_W2} when combined with the CGMT.
\cts{Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be a $\rm{PL}(2)$ function, or, the function $f_{\cal{L}}$ defined in \eqref{eq:fdef}.} For convenience,
define
$$F_n(\hat\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}\hat\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Fix any $\varepsilon>0$ and define the set
\begin{align}\label{eq:S_set}
\mathcal{S} = \mathcal{S}(\betab^\star,\boldsymbol{\Sigma}) =\{\boldsymbol{\beta}{~\big |~} |F_n(\hat\boldsymbol{\beta}_n,\betab^\star,\boldsymbol{\Sigma})-\alpha_*|\geq 2\varepsilon\}.
\end{align}
It suffices to prove that the solution $\hat\boldsymbol{\beta}_n$ of the PO in \eqref{eq:PO_beta} satisfies $\hat\boldsymbol{\beta}_n\not\in\mathcal{S}$ wpa 1. To see that this is sufficient, note the following. On the one, setting $f=f_{\cal{L}}$, directly proves \eqref{eq:thm_app_f}. On the other hand, recall that $W_2$-convergence is equivalent to convergence of any $\rm{PL}(2)$ test function $f$.
To prove the desired, we need to consider the ``perturbed" PO and AO problems (compare to \eqref{eq:PO} and \eqref{eq:AO_con}) as:
\begin{align}\label{eq:PO_S}
\Phi_S({\mtx{X}})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
and
\begin{align}
\phi_S({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_S}
\end{align
Recall here, that ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$, ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$ and we have used the change of variables $\vct{w}:=\sqrt{\boldsymbol{\Sigma}}(\boldsymbol{\beta}-\betab^\star)$ for convenience.
Using \cite[Theorem 6.1(iii)]{thrampoulidis2018precise} it suffices to find costants $\bar\phi, \bar\phi_S$ and $\eta>0$ such that the following three conditions hold:
\begin{enumerate}
\item $\bar\phi_S \geq \bar\phi + 3\eta$,
\item $\phi({\vct{g}},\vct{h}) \leq \bar\phi + \eta$, with probability approaching 1,
\item $\phi_S({\vct{g}},\vct{h}) \geq \bar\phi_S - \eta$, with probability approaching 1.
\end{enumerate}
In what follows, we explicitly find $\bar\phi, \bar\phi_S,\eta$ such that the three conditions above hold.
\vspace{5pt}
\noindent\underline{Satisfying Condition 2}: Recall the deterministic min-max optimization in \eqref{eq:AO_det}. Choose $\bar\phi=\mathcal{D}(u_*,\tau_*)$ be the optimal cost of this optimization. From Lemma \ref{lem:AO}(ii), $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$. Thus, for any $\eta>0$, with probability approaching 1:
\begin{align}\label{eq:phi_lim}
\bar\phi + \eta \geq \phi({\vct{g}},\vct{h}) \geq \bar\phi - \eta.
\end{align}
Clearly then, Condition 2 above holds for any $\eta>0$.
\vspace{5pt}
\noindent\underline{Satisfying Condition 3}: Next, we will show that the third condition holds for appropriate $\bar\phi$.
Let $\vct{w}_n=\vct{w}_n({\vct{g}},\vct{h})$ be the unique minimizer of \eqref{eq:AO_con} as per Lemma \ref{lem:AO}(i), i.e., $\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}_n+\boldsymbol{\beta}}^2 = \phi({\vct{g}},\vct{h})$. Again from Lemma \ref{lem:AO}, the minimization in \eqref{eq:AO_con} is $1/\Sigma_{\max}$-strongly convex in $\vct{w}$. Here, $\Sigma_{\max}$ is the upper bound on the eigenvalues of $\boldsymbol{\Sigma}$ as per Assumption (A2). Thus, for any $\tilde\varepsilon>0$ and any feasible $\vct{w}$ the following holds (deterministically):
\begin{align}\label{eq:sc}
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\boldsymbol{\beta}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2C},~\text{provided that}~ \|\vct{w}-\vct{w}_n({\vct{g}},\vct{h})\|_2 \geq \tilde\epsilon.
\end{align}
Now, we argue that
\begin{align}\label{eq:dev_arg}
\boldsymbol{\beta}\in\mathcal{S} \text{ implies that } \|\vct{w}-\vct{w}_n({\vct{g}},\vct{h})\|_2\geq \tilde\varepsilon \text{ wpa 1, }
\end{align}
for
appropriate value of $\tilde\varepsilon$ and $\vct{w}=\sqrt{\boldsymbol{\Sigma}}(\boldsymbol{\beta}-\betab^\star)$.
Consider any $\boldsymbol{\beta}\in\mathcal{S}$.
\noindent First, by definition in \eqref{eq:S_set},
$$
|F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \geq 2\varepsilon.
$$
Second, by Lemma \ref{lem:AO}(iv), with probability approaching 1,
$$
|F(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \leq \epsilon.
$$
Third, we will show that wpa 1, there exists universal constant $C>0$ such that
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| \leq C {\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}\label{eq:dev2show}.
\end{align}
Before proving \eqref{eq:dev2show}, let us argue how combining the above three displays shows the desired. Indeed, in that case, wpa 1,
\begin{align*}
2\varepsilon &\leq |F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \leq |F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| + |F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \\
&\leq \epsilon + C \,\|\boldsymbol{\beta}-\boldsymbol{\beta}_n\|_2. \\
&\qquad\Longrightarrow \|\boldsymbol{\beta}-\boldsymbol{\beta}_n\|_2 \geq {\varepsilon}/{C}=:\hat\varepsilon\\
&\qquad\Longrightarrow \|\vct{w}-\vct{w}_n\|_2 \geq \hat\varepsilon\sqrt{\Sigma_{\min}}=:\tilde\varepsilon.
\end{align*}
In the last line above, we recalled that $\boldsymbol{\beta}=\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star$ and $\boldsymbol{\Sigma}_{i,i}\geq\Sigma_{\min},~i\in[p]$ by Assumption (A2). This proves \eqref{eq:dev_arg}.
Next, combining \eqref{eq:dev_arg} and \eqref{eq:sc}, we find that wpa 1,
$
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\boldsymbol{\beta}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2C},~\text{for all}~ \vct{w}\in\mathcal{S}.
$
Thus,
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}.\notag
\end{align}
When combined with \eqref{eq:phi_lim}, this shows that
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta.
\end{align}
Thus, choosing $\bar\phi_S = \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}$ proves the Condition 3 above.
\vspace{3pt}
To complete the proof, let us now show \eqref{eq:dev2show}. Henceforth, $C$ is used to denote a universal constant whose value can change from line to line.
First, consider the case $f=f_{\cal{L}}$, i.e., $f(x,y,z) = y(x-g(z))^2$. We have the following chain of inequalities:
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| &=
\frac{1}{p}\sum_{i=1}^p|\bSi_{i,i}|\left| (\sqrt{p}\betab^\star_i-g(\sqrt{p}\boldsymbol{\beta}_{n,i}))^2-(\sqrt{p}\betab^\star_i-g(\sqrt{p}\boldsymbol{\beta}_{i}))^2 \right|\notag\\
&\leq
\Sigma_{\max} \frac{1}{p}\sum_{i=1}^p \left| (\sqrt{p}\betab^\star_i-g(\sqrt{p}\boldsymbol{\beta}_{n,i}))^2-(\sqrt{p}\betab^\star_i-g(\sqrt{p}\boldsymbol{\beta}_{i}))^2 \right|\notag\\
&\leq
\Sigma_{\max} L \frac{1}{p}\sum_{i=1}^p (1+ \|\sqrt{p}[\betab^\star_i,\boldsymbol{\beta}_{n,i}]\|_2 + \|\sqrt{p}[\betab^\star_i,\boldsymbol{\beta}_{i}]\|_2) \sqrt{p}|\boldsymbol{\beta}_{n,i}-\boldsymbol{\beta}_{i}|\notag\\
&\leq
\Sigma_{\max} L \Big(1+ \frac{1}{\sqrt{p}}\big(\sum_{i=1}^{p}\|\sqrt{p}[\betab^\star_i,\boldsymbol{\beta}_{n,i}]\|_2^2\big)^{1/2} + \frac{1}{\sqrt{p}}\big(\sum_{i=1}^p\|\sqrt{p}[\betab^\star_i,\boldsymbol{\beta}_{i}]\|_2^2\big)^{1/2}\Big) \|\boldsymbol{\beta}_{n}-\boldsymbol{\beta}\|_2\notag\\
&\leq
C \left(1+ \max\{\|\betab^\star\|_2^2,\|\boldsymbol{\beta}_{n}\|_2^2,\|\boldsymbol{\beta}\|_2^2\}^{1/2} \right) \|\boldsymbol{\beta}_{n}-\boldsymbol{\beta}\|_2.\label{eq:fcase1}
\end{align}
In the second line above, we used boundedness of $\bSi_{i,i}$ as per Assumption (A2). In the third line, we used the fact that the function $\psi(a,b) = (a-g(b))^2$ is $\rm{PL}(2)$. The fourth line follows by Cauchy-Schwartz inequality. Finally, in the last line, we used the elementary fact that $a+b+c\leq 3\max\{a,b,c\}$ for $a=2\sum_{i=1}^p(\betab^\star_i)^2$ and $b=\sum_{i=1}^p\boldsymbol{\beta}_{n,i}^2$ and $c=\sum_{i=1}^p\boldsymbol{\beta}_{i}^2$.
Second, consider the case $f\in\rm{PL}(2)$. Let $\vct{a}_i=(\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}),\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\boldsymbol{\beta}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$, for $i\in[p]$. A similar chain of inequality holds:
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| &\leq \frac{L}{p}\sum_{i=1}^p (1+\max\{\|\vct{a}_i\|_2,\|\vct{b}_i\|_2\})\|\vct{a}_i-\vct{b}_i\|_2\notag \\
&= \frac{L}{p}\sum_{i=1}^p \left(1+\max\{\|\vct{a}_i\|_2,\|\vct{b}_i\|_2\}\right)\cdot|\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}) - \sqrt{p}\boldsymbol{\beta}_i|\notag\\
&\leq L\left(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2}\} \right)^{1/2}{\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}\notag\\
&\leq C\left(1+\max\{\|\boldsymbol{\beta}_{n}\|_2^2,\|\betab^\star\|_2^2,\|\boldsymbol{\beta}\|_2^2,\frac{1}{p}\sum_{i=1}^p\bSi_{i,i}^2\} \right)^{1/2}{\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}\notag\\
&\leq C\Sigma_{\max}^2\left(1+\max\{\|\boldsymbol{\beta}_{n}\|_2^2,\|\betab^\star\|_2^2,\|\boldsymbol{\beta}\|_2^2\} \right)^{1/2}{\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}.\label{eq:fcase2}
\end{align}
The first inequality uses the fact that $f\in\rm{PL}(2)$. The second inequality in the third line follows by Cauchy-Schwartz. The last inequality used Assumption (A2) on boundedness of $\bSi_{i,i}$.
It follows from either \eqref{eq:fcase1} or \eqref{eq:fcase2} that in order to prove \eqref{eq:dev2show}, we need to show boundedness of the following terms: $\|\boldsymbol{\beta}_n\|_2$, $\|\betab^\star\|_2$ and $\|\boldsymbol{\beta}\|_2$. By feasibility of $\boldsymbol{\beta}_n$ and $\boldsymbol{\beta}$, we know that $\boldsymbol{\beta}_n,\boldsymbol{\beta}\in\mathcal{B}$. Thus, the desired $\|\boldsymbol{\beta}\|_2<\infty$ and $\|\boldsymbol{\beta}_n\|_2<\infty$ follow directly by Lemma \ref{lem:bd_PO}. (Alternatively, for $\boldsymbol{\beta}_n$ we conclude the desired by directly applying Lemma \ref{lem:AO}(v)). Finally, to prove $\|\betab^\star\|_2<\infty$, note that
$$
\|\betab^\star\|_2^2 =\frac{1}{p}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^2,
$$
\cts{which is bounded wpa 1 by Assumption (A2), which implies bounded second moments of $\sqrt{p}\betab^\star$. } This completes the proof of \eqref{eq:dev2show}, as desired.
\vspace{5pt}
\noindent\underline{Satisfying Condition 1:} To prove Condition 1, we simply pick $\eta$ to satisfy the following
\begin{align}
\bar\phi_S > \bar\phi + 3 \eta ~\Leftarrow~ \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta \geq \bar\phi + 3 \eta ~\Leftarrow~ \eta \leq \frac{\tilde\epsilon^2}{8\Sigma_{\max}}.\notag
\end{align}
This completes the proof of Theorem \ref{thm:master_W2}.
\subsection{Proof of Lemma \ref{lem:AO}}\label{sec:proofAO}
~~~~
~~~~
\vspace{3pt}
\subsubsection{Proof of (i).} Strong convexity of the objective function in \eqref{eq:PO} is easily verified by the second derivative test. Note here that we use Assumption (A2) that $\bSi_{i,i}\leq\Sigma_{\max},~i\in[p].$ Uniqueness of the solution follows directly from strong convexity. \ct{Strictly speaking we might need to also argue existence, i.e., feasibility of the AO. An indirect way is to show feasibility using the CGMT, but it seems unnecessarily complicated?}
\vspace{5pt}
\subsubsection{Proof of (ii).}Using Lagrangian formulation, the solution $\vct{w}_n$ to \eqref{eq:AO_con} is the same as the solution to the following:
\begin{align}
\left(\vct{w}_n,u_n\right) :=\arg\min_{\vct{w}\in\mathcal{B}}\max_{u\geq 0} ~\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + u \left( \sqrt{\tn{\vct{w}}^2+\sigma^2} \tn{\bar{\g}} - \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} + \frac{\sigma h}{\sqrt{n}} \right)\label{eq:AO_2}
\end{align}
where we have: (i) set $\bar{\g} := {\vct{g}}/\sqrt{n}$ and $\bar{\h}:= \vct{h}/\sqrt{p}$; (ii) recalled that $p/n=\kappa$; and, (iii) used $\left(\vct{w}_n,u_n\right)$ to denote the optimal solutions in \eqref{eq:AO_2}. The subscript $n$ emphasizes the dependence of $\left(\vct{w}_n,u_n\right)$ on the problem dimensions. Also note that (even though not explicit in the notation) $\left(\vct{w}_n,u_n\right)$ are random variables depending on the realizations of $\bar{\g},\bar{\h}$ and $h$.
Notice that the objective function above is convex in $\vct{w}$ and linear (thus, concave) in $u$. Thus, strong duality holds and we can flip the order of min-max. Moreover, in order to make the objective easy to optimize with respect to $\vct{w}$, we use the following variational expression for the square-root term $\sqrt{\tn{\vct{w}}^2+\sigma^2}$:
$$
\tn{\bar{\g}}\sqrt{\tn{\vct{w}}^2+\sigma^2} = \tn{\bar{\g}}\cdot\min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau}{2} + \frac{\tn{\vct{w}}^2+\sigma^2}{2\tau} \right\} = \min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau\tn{\bar{\g}}^2}{2} + \frac{\sigma^2}{2\tau} + \frac{\tn{\vct{w}}^2}{2\tau} \right\},
$$
where $B$ is defined in Lemma \ref{lem:bd_PO}. For convenience define the constraint set for the variable $\tau$ as $\mathcal{T}':=[\sigma,\sqrt{\sigma^2+B_+^2}]$. For reasons to be made clear later in the proof (see proof of statement (iii)), we consider the (possibly larger) set:
\[
\mathcal{T}:=[\sigma,\max\{\sqrt{\sigma^2+B_+^2},2\tau_*\}]\,
\]
where $\tau_*$ is as in the statement of the lemma.
The above lead to the following equivalent formulation of \eqref{eq:AO_2}:
\begin{align}
\left(\vct{w}_n,u_n,\tau_n\right) = \max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} + \min_{\vct{w}\in\mathcal{B}} \left\{ \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + \frac{u}{2\tau}\tn{\vct{w}}^2 - u \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} \right\} .\label{eq:AO_3}
\end{align}
The minimization over $\vct{w}$ is easy as it involves a strongly convex quadratic function. First, note that unconstrained optimal $\vct{w}':=\vct{w}'(\tau,u)$ (for fixed $(\tau,u)$) is given by
\begin{align}\label{eq:w'}
\vct{w}':=\vct{w}'(\tau,u) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right),
\end{align}
and \eqref{eq:AO_3} simplifies to
\begin{align}
\left(u_n,\tau_n\right)=\max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} - \frac{1}{2} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)\,=:\mathcal{R}(u,\tau) .\label{eq:AO_4}
\end{align}
It can be checked by direct differentiation and the second-derivative test that the objective function in \eqref{eq:AO_4} is strictly convex in $\tau$ and strictly concave in $u$ in the domain $\{(u,\tau)\in\mathbb{R}_+\times\mathbb{R}_+\}$. Thus, the saddle point $(u_n,\tau_n)$ is unique. Specifically, this implies that the optimal $\vct{w}_n$ in \eqref{eq:AO_3} is given by (cf. \eqref{eq:w'})
\begin{align}\label{eq:w_n}
\vct{w}_n=\vct{w}'(\tau_n,u_n) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).
\end{align}
In Lemma \ref{lem:AO}(v) we will prove that wpa 1, in the limit of $p\rightarrow\infty$, $\|\vct{w}_n\|_2\leq C$ for sufficiently large absolute constant $C>0$. Thus, by choosing the upper bound in the definition of $\mathcal{B}$ in Lemma \ref{lem:bd_PO} strictly larger than C, guarantees that the unconstrained $\vct{w}_n$ in \eqref{eq:w_n} is feasible in \eqref{eq:AO_3}.
In what follows, we characterize the high-dimensional limit of the optimal pair $(u_n,\tau_n)$ in the limit $n,p\rightarrow\infty,~p/n\rightarrow\kappa$.
We start by analyzing the (point-wise) convergence of $\mathcal{R}(u,\tau)$.
For the first three summands in \eqref{eq:AO_4}, we easily find that
$$
\left\{\frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}}\right\}~ \stackrel{{P}}{\longrightarrow}~\left\{
\frac{u\tau}{2} + \frac{u\sigma^2}{2\tau} \right\}.
$$
Next, we study the fourth summand. First, note that
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}(u\sqrt{\kappa}\bar{\h}) &= u^2\kappa\,\frac{1}{p}\vct{h}^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\vct{h} \notag\\
&= u^2\kappa\,\frac{1}{p}\sum_{i=1}^{p}\frac{\vct{h}_i^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}} \notag\\
&\stackrel{{P}}{\longrightarrow} u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right].
\end{align}
In the last line, $\Lambda$ is a random variable as in Definition \ref{def:Xi}. \cts{Also, we used Assumption (A2) together with the facts that $\vct{h}$ is independent of $\boldsymbol{\Sigma}$ and that the function $(x_1,x_2)\mapsto x_1^2(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(2)$ assuming $x_2$ is bounded.} \ct{Question: Is it immediate that the empirical distribution of $(\beta,\boldsymbol{\Sigma},\vct{h})$ converges in $W_k$ to $\mu\otimes\mathcal{N}(0,1)$ given that $(\beta,\boldsymbol{\Sigma})$ converges to $\mu$ and $\vct{h}$ is independent???}
Second, we find that
\begin{align}
(\betab^\star)^T\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star &= \frac{1}{p}(\sqrt{p}\betab^\star)^T\left({\mtx{I}}+\frac{u}{\tau}\boldsymbol{\Sigma}\right)^{-1}(\sqrt{p}\betab^\star) \notag\\
&= \frac{1}{p}\sum_{i=1}^{p}\frac{\left(\sqrt{p}\betab^\star_i/\sqrt{\bSi_{i,i}}\right)^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}}\notag\\
&\stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right].
\end{align}
Here, $\Lambda,B$ are random variables as in Definition \ref{def:Xi} and \cts{we also used Assumption (A2) together with the fact that the function $(x_1,x_2)\mapsto x_1^2x_2^{-1}(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(2)$ assuming $x_2$ is bounded.}
Third, \cts{by independence of $(\betab^\star, \boldsymbol{\Sigma})$ from $\vct{h}$}
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star = u\sqrt{\kappa} \cdot \frac{1}{p}\sum_{i=1}
^p \frac{\vct{h}_i\boldsymbol{\Sigma}_{i,i}^{-1/2}(\sqrt{p}\betab^\star_i)}{\boldsymbol{\Sigma}_{i,i}^{-1}+\frac{u}{\tau}{\mtx{I}}} ~\stackrel{{P}}{\longrightarrow}~ 0.
\end{align}
Putting these together, the objective $\mathcal{R}(u,\tau)$ in \eqref{eq:AO_4} converges point-wise in $u,\tau$ to
\begin{align}
\mathcal{R}(u,\tau)\stackrel{{P}}{\longrightarrow}\mathcal{D}(u,\tau) := \frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).\label{eq:conv_pt}
\end{align}
Note that $\mathcal{R}(u,\tau)$ (and thus, $\mathcal{D}(u,\tau)$) is convex in $\tau$ and concave in $u$. Thus, the convergence in \eqref{eq:conv_pt} is in fact uniform (e.g., \cite{AG1982}) and we can conclude that
\begin{align}\label{eq:Dc0}
\phi({\vct{g}},\vct{h}) \stackrel{{P}}{\longrightarrow} \max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
and
\begin{align}\label{eq:Dc}
(u_n,\tau_n) \stackrel{{P}}{\longrightarrow} (u_*,\tau_*):=\arg\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
In the proof of statement (iii) below, we show that the saddle point of \eqref{eq:Dc0} is $(u_*,\tau_*)$. In particular, $\tau_*$ is strictly in the interior of $\mathcal{T}$, which combined with convexity implies that
$$
\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau) = \max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau) =: \bar\phi.
$$
This, together with the first display above proves the second statement of the lemma.
\vspace{5pt}
\subsubsection{Proof of (iii).} Next, we compute the saddle point $(u_*,\tau_*)$ by studying the first-order optimality conditions of the srtictly concave-convex $\mathcal{D}(u,\tau)$. Specifically, we consider unconstrained minimization over $\tau$ and we will show that the minimum is achieved in the strict interior of $\mathcal{T}$. Direct differentiation gives
\begin{subequations}
\begin{align}
{\tau} + \frac{\sigma^2}{\tau} - 2u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] + \frac{u^2}{\tau}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] + \frac{1}{\tau}\operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0, \label{eq:fo1}\\
{u} - \frac{u\sigma^2}{\tau^2} - \frac{u^3}{\tau^2}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} + \frac{u}{\tau}\right)^2}\right] - \frac{u}{\tau^2} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0,\label{eq:fo2}
\end{align}
\end{subequations}
Multiplying \eqref{eq:fo2} with $\frac{\tau}{u}$ and adding to \eqref{eq:fo1} results in the following equation
\begin{align}
\tau = u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] ~\Leftrightarrow ~
\operatorname{\mathbb{E}}\left[ \frac{1}{(\frac{u}{\tau}\Lambda)^{-1}+1} \right] = \frac{1}{\kappa} \label{eq:tauu}\,.
\end{align}
Thus, we have found that the ratio $\frac{u_*}{\tau_*}$ is the unique solution to the equation in \eqref{eq:tauu}. Note that this coincides with the Equation \eqref{eq:ksi} that defines the parameter $\xi$ in Definition \ref{def:Xi}. The fact that \eqref{eq:tauu} has a unique solution for all $\kappa>1$ can be easily seen as $F(x)=\operatorname{\mathbb{E}}\left[ \frac{1}{(x\Lambda)^{-1}+1} \right], x\in\mathbb{R}_+$ has range $(0,1)$ and is strictly increasing (by differentiation).
Thus, we call $\xi=\frac{u_*}{\tau_*}$. Moreover, multiplying \label{eq:fo2} with $u$ leads to the following equation for $\tau_*$:
\begin{align}
u_*^2 = \sigma^2\xi^2 + u_*^2 \xi^2 \kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right] + \xi^2 \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right] ~\Rightarrow~ \tau_*^2 = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\xi^2\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right]} = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left((\xi\Lambda)^{-1} +1\right)^2}\right]}.
\end{align}
{Again, note that this coincides with Equation \eqref{eq:gamma} that determines the parameter $\gamma$ in Definition \ref{def:Xi}, i.e., $\tau_*^2 = \gamma.$ }
\vspace{5pt}
\subsubsection{Proof of (iv).} For convenience, define
$$F_n(\hat\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}\hat\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Recall from \eqref{eq:w_n} the explicit expression for $\vct{w}_n$, repeated here for convenience.
\begin{align}
\vct{w}_n = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).\notag
\end{align}
Also, recall that $\boldsymbol{\beta}_n = \boldsymbol{\Sigma}^{-1/2}\vct{w}_n+\betab^\star$. Thus, (and using the fact that $\bar{\h}$ is distributed as $-\bar{\h}$),
\begin{align}
\boldsymbol{\beta}_n &=\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}u_n\sqrt{\kappa}\bar{\h} + \left({\mtx{I}}-\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\right)\betab^\star\notag\\
\Longrightarrow\boldsymbol{\beta}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\bar{\h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i.\label{eq:betan}
\end{align}
For $i\in[p]$, define
\begin{align}
\vct{v}_{n,i} = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\bar{\h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i
\end{align}
In the above, for convenience, we have denoted $\xi_n:=u_n/\tau_n$ and recall that $\xi_*:=u_*/\tau_*$.
The proof proceeds in two steps. In the first step, we use the fact that $\xi_n\stackrel{{P}}{\longrightarrow}\xi_*$ and $u_n\stackrel{{P}}{\longrightarrow} u_\star$ (see \eqref{eq:Dc}) to prove that for any $\varepsilon\in(0,\xi_*/2)$, there exists absolute constant $C>0$ such that wpa 1:
\begin{align}\label{eq:ivstep1}
|F_n(\sqrt{p}\boldsymbol{\beta}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.
\end{align}
In the second step, we use Lipschitzness of $f$ and Assumption \ref{ass:mu} to prove that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:ivstep2}
\end{align}
The desired follows by combining \eqref{eq:ivstep1} and \eqref{eq:ivstep2}. Thus, in what follows, we prove \eqref{eq:ivstep1} and \eqref{eq:ivstep2}.
\vspace{2pt}
\noindent\underline{Proof of \eqref{eq:ivstep1}.}~~Fix some $\varepsilon\in(0,\xi_*/2)$. From \eqref{eq:Dc}, we know that w.p.a. 1 $|\xi_n-\xi_*|\leq \varepsilon$ and $|u_n-u_*|\leq \varepsilon$. Thus, $\vct{w}_n$ is close to $\vct{v}_n$. Specifically, in this event, for every $i\in[p]$, it holds that:
\begin{align}
\notag|\boldsymbol{\beta}_{n,i} - \vct{v}_{n,i}| &\leq {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{\tau_n}{1+(\xi_n\bSi_{i,i})^{-1}}-\frac{\tau_*}{1+(\xi_*\bSi_{i,i})^{-1}}\right| \\
\notag&= {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{u_n}{\xi_n+\bSi_{i,i}^{-1}}-\frac{u_*}{\xi_*+\bSi_{i,i}^{-1}}\right| \\
\notag& \leq
{|\betab^\star_i|}\frac{|\bSi_{i,i}||\xi_n-\xi_*|}{|1+\xi_n\bSi_{i,i}||1+\xi_*\bSi_{i,i}|} + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}} \frac{u_*|\xi_n-\xi_*|}{(\xi_n+\bSi_{i,i}^{-1})(\xi_*+\bSi_{i,i}^{-1})} +
\sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}}\frac{|u_n-u_*|}{\xi+\bSi_{i,i}^{-1}}
\\
\notag& \leq
{|\betab^\star_i|}{\Sigma_{\max}\varepsilon} + \sqrt{\kappa}{|\bar{\h}_i|} u_*\Sigma_{\max}^{3/2} \varepsilon+
\sqrt{\kappa}|\bar{\h}_i|\Sigma_{\max}^{1/2}\varepsilon
\\
&\leq \varepsilon \cdot \max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\, \left(|\bar{\h}_i| + |\betab^\star_i|\right).\label{eq:betav}
\end{align
In the second line above, we recalled that $u_n=\tau_n\xi_n$ and $u_*=\tau_*\xi_*$. In the third line, we used the triangle inequality. In the fourth line, we used that $\xi_*>0$, $0<\bSi_{i,i}\leq\Sigma_{\max}$ and $\xi_n\geq \xi_*-\varepsilon \geq \xi_*/2 >0$.
Now, we will use this and Lipschitzness of $f$ to argue that there exists absolute constant $C>0$ such that wpa 1:
\begin{align}
|F_n(\sqrt{p}\boldsymbol{\beta}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.\notag
\end{align}
Denote, $\vct{a}_i=(\sqrt{p}\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\vct{v}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$. Following the exact same argument as in \eqref{eq:dev2show} (just substitute $\boldsymbol{\beta}\leftrightarrow\vct{v}_n$ in the derivation), we have that for some absolute constant $C>0$ wpa 1:
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| &\leq C \|\boldsymbol{\beta}_n-\vct{v}_n\|_2.
\end{align}
From this and \eqref{eq:betav}, we find that
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})|
&\leq
C\varepsilon\max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2} \right\} \left(\sum_{i=1}^p \left(|\bar{\h}_i|+|\betab^\star_i|\right)^2\right)^{1/2}\notag
\\
&\leq
C\varepsilon\sqrt{2}\max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\sqrt{\|\betab^\star\|_2^2 + \|\bar{\h}\|_2^2}.\label{eq:epsS}
\end{align}
But, \cts{recall that $\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^2<\infty$, as $p\rightarrow \infty$ by Assumption (A2).}
Also, since $\bar{\h}_i\sim\mathcal{N}(0,1/p)$, it holds that $\|\bar{\h}\|_2^2\leq 2$, wpa 1 as $p\rightarrow\infty$. Therefore, from \eqref{eq:epsS}, wpa 1, there exists constant $C>0$ such that
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \leq C \cdot\varepsilon,
\end{align}
as desired.
\vspace{2pt}
\noindent\underline{Proof of \eqref{eq:ivstep2}.}~~Next, we will use (A2) to show that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:AO_conv}
\end{align}
Notice that $\vct{v}_n$ is a function of $\betab^\star,\boldsymbol{\Sigma},\bar{\h}$. Concretely, define ${\widetilde{g}}:\mathbb{R}^3\rightarrow\mathbb{R}$, such that
$$
\widetilde{g}(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi_* x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1,
$$
and notice that
$$
\sqrt{p}\vct{v}_{n,i} = \widetilde{g}\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right) = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\vct{h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\sqrt{p}\betab^\star_i.
$$
Thus,
$$
F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma}) = \frac{1}{p}\sum_{i=1}^p f\left(g\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right),\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) =: \frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right),
$$
where we have defined $h:\mathbb{R}^3\rightarrow\mathbb{R}$:
\begin{align}
h(x_1,x_2,x_3) := f\left(\widetilde{g}(x_2,x_3,x_1),x_2,x_3\right).\label{h func}
\end{align}
\cts{We will prove that $h\in\rm{PL}(4)$. Indeed, if that were the case, then Assumption (A2) gives}
\begin{align}
\frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[h(H,B,\Lambda)\right] &= \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[f\left(\widetilde{g}(B,\Lambda,H),B,\Lambda\right))\right]\\
& = \operatorname{\mathbb{E}}[f\left(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda\right)] = \alpha_*,
\end{align}
where the penultimate equality follows by recognizing that (cf. Eqn. \eqref{eq:X})
$$
\widetilde{g}(B,\Lambda,H) = (1-(1+\xi_*\Lambda)^{-1})B + \sqrt{\kappa}\frac{\tau_*\Lambda^{-1/2}}{1+(\xi_*\Lambda)^{-1}}H = X_{\kappa,\sigma^2}(\Lambda,B,H).
$$
It remains to show that $h\in\rm{PL}(4)$. Lemma \ref{lem:hPL} in Section \ref{SM useful fact} shows that if $f\in\rm{PL}(k)$, then $h\in\rm{PL}(k+1)$ for all integers $k\geq 2$. First, consider the case $f\in\rm{PL}(2)$. Then, $h\in\rm{PL}(3)$; thus, also $h\in\rm{PL}(4)$. Second, consider the case $f=f_{\cal{L}}$. Then, we prove in Lemma \ref{lem:fL} that $f_{\cal{L}}\in\rm{PL}(3)$. Thus, the desired holds in this case too.
\vspace{5pt}
\subsubsection{Proof of (v):} Let $\psi:\mathbb{R}\rightarrow\mathbb{R}$ be any bounded Lipschitz function. The function $f(a,b,c) = \psi(a)$ is trivially $\rm{PL}$ of order $2$. Thus, by directly applying statement (iv) of the lemma, we find that
$$
\frac{1}{p}\sum_{i=1}^p{\psi(\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}))} \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\psi(X_{\kappa,\sigma^2})\right].
$$
Since this holds for any bounded Lipschitz function, we have shown that the empirical convergence of $\boldsymbol{\beta}_n$ converges weakly to the distribution of $X_{\kappa,\sigma^2}$. It remains to prove boundedness of the $2$nd moment as advertised in \eqref{eq:k_AO}. Recall from
\eqref{eq:betan} that
\begin{align}
\sqrt{p}\boldsymbol{\beta}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\vct{h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)(\sqrt{p}\betab^\star_i).\notag
\end{align}
Using this, boundedness of $\boldsymbol{\Sigma}_{i,i}$ from Assumption (A2), and the fact that $\tau_n\stackrel{{P}}{\longrightarrow}\tau_\star, \xi_n\stackrel{{P}}{\longrightarrow}\xi_\star$, there exists constant $C=C(\Sigma_{\max},\Sigma_{\min},k,\tau_\star,\xi_\star)$ such that wpa 1,
\begin{align}
\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\boldsymbol{\beta}_{n,i}|^{2} \leq C\left(\frac{1}{p}\sum_{i=1}^p|\vct{h}_{i}|^{2}+\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\betab^\star_{i}|^{2}\right).\notag
\end{align}
But the two summands in the expression above are finite in the limit of $p\rightarrow\infty$. Specifically, (i) from Assumption (A2), $\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\betab^\star_{i}|^{2}\stackrel{{P}}{\longrightarrow}\operatorname{\mathbb{E}}[B^{2}]<\infty$; (ii) $\frac{1}{p}\sum_{i=1}^p|\vct{h}_{i}|^{2}\stackrel{{P}}{\longrightarrow}\operatorname{\mathbb{E}}[H^{2}]=1$, using the facts that $\vct{h}_i\stackrel{iid}{\sim}\mathcal{N}(0,1)$ and $H\sim\mathcal{N}(0,1)$. This proves \eqref{eq:k_AO}, as desired.
\input{risks_proofs}
\section{Useful results about pseudo-Lipschitz functions}\label{SM useful fact}
For $k\geq 1$ we say a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$ and denote it by $f\in \rm{PL}(k)$ if there exists a cosntant $L>0$ such that, for all $\vct{x},\vct{y}\in\mathbb{R}^m$:
\begin{align}
|f(\vct{x})-f(\vct{y})|\leq L\left(1+\|\vct{x}\|_2^{k-1}+\|\vct{y}\|_2^{k-1}\right)\|\vct{x}-\vct{y}\|_2.\label{PL func}
\end{align}
In particular, when $f\in\rm{PL}(k)$, the following properties hold:
\begin{enumerate}
\item There exists a constant $L'$ such that for all $\vct{x}\in\mathbb{R}^n$: $|f(\vct{x})|\leq L'(1+\|\vct{x}\|_2^k).$
\item $f$ is locally Lipschitz, that is for any $M>0$, there exists a constant $L_{M,m}<\infty$ such that for all $x,y\in[-M,M]^m$,
$
|f(\vct{x})-f(\vct{y})| \leq L_{M,m}\|\vct{x}-\vct{y}\|_2.
$
Further, $L_{M,m}\leq c(1+(M\sqrt{m})^{k-1})$ for some costant $c$.
\end{enumerate}
Using the above properties, we prove the following two technical lemmas used in the proof of Theorem \ref{thm:app_W2}
\begin{lemma}\label{lem:fL}
Let $g:\mathbb{R}\rightarrow\mathbb{R}$ be a Lipschitz function. Consider the function $f:\mathbb{R}^3\rightarrow\mathbb{R}$ defined as follows:
$$
f(\vct{x}) = x_1(x_2-g(x_3))^2.
$$
Then, $f\in\rm{PL}(3).$
\end{lemma}
\begin{proof}
Let $h:\mathbb{R}^2\rightarrow\mathbb{R}$ defined as $h({\vct{u}})=({\vct{u}}_1-g({\vct{u}}_2))^2$. The function $({\vct{u}}_1,{\vct{u}}_2)\mapsto{\vct{u}}_1-g({\vct{u}}_2)$ is clearly Lipschitz. Thus, $h\in\rm{PL}(2)$, i.e.,
\begin{align}
|h({\vct{u}})-h(\vct{v})| \leq C(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2\quad\text{and}\quad |h(\vct{v})|\leq C'(1+\|\vct{v}\|_2^2).\label{eq:h_pl}
\end{align}
Therefore, letting $\vct{x}=(x_1,{\vct{u}})\in\mathbb{R}^3$ and $\vct{y}=(y_1,\vct{v})\in\mathbb{R}^3$, we have that
\begin{align}
|f(\vct{x})-f(\vct{y})| &= |x_1h({\vct{u}}) - y_1h(\vct{v})| \leq |x_1||h({\vct{u}})-h(\vct{v})| + |h(\vct{v})| |x_1-y_1|\notag\\
&\leq C|x_1|(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(|x_1|^2+(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|\vct{x}-\vct{y}\|_2.
\end{align}
In the second line, we used \eqref{eq:h_pl}. In the third line, we used $2xy\leq x^2+y^2$. In the fourth line, we used Cauchy-Schwarz inequality. $C,C'>0$ are absolute constants that may change from line to line.
This completes the proof of the lemma.
\end{proof}
\begin{lemma}\label{lem:hPL}
Let functions $f,g:\mathbb{R}^3\rightarrow\mathbb{R}$ such that $f\in\rm{PL}(k)$ and
$$
g(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi_* x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1.
$$
Here, $\xi_*,\tau_*,\kappa$ are positive constants.
Further define
\begin{align}
h(x,y,z) := f\left(g(y,z,x),y,z\right),
\end{align}
and assume that $y$ take values on a fixed bounded compact set $\mathcal{M}$.
Then, it holds that $h\in\rm{PL}(k+1)$.
\end{lemma}
\begin{proof}
Since $f$ is $\rm{PL}(k)$, for some $L>0$, \eqref{PL func} holds. Fix $x,x'\in\mathbb{R}$, $\vct{a}=[y,z]\in\mathbb{R}^{2}$ and $\vct{a}'=[y',z']\in\mathbb{R}^{2}$. Let $\vct{b}=[{\vct{g}},\vct{a}]=[{\vct{g}},y,z]\in\mathbb{R}^3$ where ${\vct{g}}=g(y,z,x)\in\mathbb{R}$ and define accordingly $\vct{b}'$ and ${\vct{g}}'$. We have that
\begin{align}
|h([x,\vct{a}])-h([x,\vct{a}'])|&= |f(\vct{b})-f(\vct{b}')|\notag\\
&\leq L\left(1+\|\vct{b}\|_2^{k-1}+\|\vct{b}'\|_2^{k-1}\right)\|\vct{b}-\vct{b}'\|_2\notag\\
&\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|{\vct{g}}|^{k-1}+|{\vct{g}}'|^{k-1}\right)(\|\vct{a}-\vct{a}'\|_2+|{\vct{g}}-{\vct{g}}'|),\label{eq:hlip}
\end{align}
for some constant $C>0$. In the last inequality we have repeatedly used the inequality
$
\left(\sum_{i=1}^m \|\vct{v}_i\|_2^2\right)^{\frac{d}{2}} \leq C(m)\cdot\sum_{i=1}^m\|\vct{v}_i\|_2^{d}.
$
Next, we need to bound the ${\vct{g}}$ term in terms of $\vct{a}$. This is accomplished as follows
\begin{align}
|{\vct{g}}|^{k-1} &= \left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x + (1-(1+\xi_*y)^{-1})z\right|^{k-1}\notag\\
&\leq C (|z|+|x|)^{k-1} \notag\\
&\leq C \left(|x|^{k-1}+|z|^{k-1}\right) \leq C (|x|^{k-1} + \|\vct{a}\|_2^{k-1}).\label{eq:gb}
\end{align}
Here, the value of the constant $C>0$ may change from line to line.
Secondly and similarly, we have the following perturbation bound on the ${\vct{g}}-{\vct{g}}'$. Recall that $(x,y)$ and $(x',y')$ are bounded. We will prove the following sequence of inequalities
\begin{align}
|{{\vct{g}}-{\vct{g}}'}|&= |g(y,z,x)-g(y',z',x')| \notag\\
&\leq \left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x - \frac{(y')^{-1/2}}{1+(\xi y')^{-1}}\sqrt{\kappa}\tau_* x'\right|
+ \left|(1-(1+\xi_*y)^{-1})z-(1-(1+\xi_*y')^{-1})z'\right|\notag\\
&\leq
\left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x - \frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x'\right| +\left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x'- \frac{(y')^{-1/2}}{1+(\xi y')^{-1}}\sqrt{\kappa}\tau_* x'\right|\notag\\
&\qquad+ \left|(1-(1+\xi_*y)^{-1})z-(1-(1+\xi_*y)^{-1})z'\right| + \left|(1-(1+\xi_*y)^{-1})z'-(1-(1+\xi_*y')^{-1})z'\right| \notag\\
&\leq C_1|x-x'| + C_2|x'||y-y'| + C_3|z-z'| + C_4|z'||y-y'|\notag\\
&\leq C(1+|x'| + |z'|)(|x-x'| + |z-z'| + |y-y'|)\\
&\leq C\sqrt{3}(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2
.\label{eq:gd}
\end{align}
In the fourth inequality above, we used the fact the assumption that $|y|$ is bounded. In the last line, we used Cauchy-Scwhartz.
Substituting \eqref{eq:gb} and \eqref{eq:gd} in \eqref{eq:hlip} gives:
\begin{align}
|h(x,y,z) - h(x',y',z')| &\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|x|^{k-1}+|x'|^{k-1}\right)(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2\notag \\
&\leq C\left(1+\|[\vct{a},x]\|_2^{k}+\|[\vct{a}',x']\|_2^{k}\right)\|[\vct{a},x]-[\vct{a}',x']\|_2.
\end{align}
Thus, $h\in\rm{PL}(k+1)$, as desired.
\end{proof}
\section{Proofs for overparameterized least-squares}\label{sec proof thm 1}
In this section, we assume the linear Gaussian problem (LGP) of Definition \ref{def LGP}, the overparameterized regime $k=p>n$ and the min-norm model ${\boldsymbol{\hat{\beta}}}$ of \eqref{eq:min_norm}. We prove Theorem \ref{thm:master_W2} that derives the asymptotic DC of ${\boldsymbol{\hat{\beta}}}$ and we show how this leads to sharp formulae for the risk of the Magnitude- and Hessian-pruned models.
\subsection{Notation and Assumptions}\label{sec:ass_app}
For the reader's convenience, we recall some necessary notation and assumptions from Section \ref{sec main}. We say that a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$, denoted $f\in\rm{PL}(k)$, if there is a constant $L>0$ such that for all $\vct{x},\vct{y}\in\mathbb{R}^m$,
$
|f(\vct{x}) - f(\vct{y})|\leq L(1+\tn{\vct{x}}^{k-1}+\tn{\vct{y}}^{k-1})\|\vct{x}-\vct{y}\|_2
$
(See also Section \ref{SM useful fact}).
We say that a sequence of probability distributions $\nu_p$ on $\mathbb{R}^m$ {converges in $W_k$} to
$\nu$, written $\nu_p\stackrel{W_k}{\Longrightarrow} \nu$, if $W_k(\nu_p,\nu) \rightarrow 0$ as $p \rightarrow \infty$. An equivalent definition is that, for any $f\in\rm{PL}(k)$, $\lim_{p\rightarrow\infty}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$ (e.g., \cite{montanari2017estimation}).
Finally, recall that a sequence of probability distributions $\nu_n$ on $\mathbb{R}^m$ \emph{converges weakly} to $\nu$, if for any bounded Lipschitz function $f$: $\lim_{p\rightarrow\infty}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$.
Throughout, we use $C,C',c,c'$ to denote absolute constants (not depending on $n,p$) whose value might change from line to line.
We focus on a double asymptotic regime where:
$$n,p,s\rightarrow\infty \text{ at fixed overparameterization ratio } \kappa:=p/n>1 \text{ and sparsity level } \alpha:=s/p\in(0,1).$$
For a sequence of random variables $\mathcal{X}_{p}$ that converge in probability to some constant $c$ in the limit of Assumption \ref{ass:linear} below, we write $\mathcal{X}_{p}\stackrel{{P}}{\longrightarrow} c$. For a sequence of event $\mathcal{E}_p$ for which $\lim_{p\rightarrow}\mathbb{P}(\mathcal{E}_p) = 1$, we say that $\mathcal{E}_p$ occurs \emph{with probability approaching 1}. For this, we will often use the shorthand ``wpa 1".
\vspace{5pt}
Next, we recall the set of assumption under which our analysis applies:
\asstwo*
\assthree*
\noindent We remark that Assumption \ref{ass:mu} above implies (see \cite[Lem.~4]{bayati2011dynamics} and \cite[Lem.~A3]{javanmard2013state}) that for any pseudo-Lipschitz function $\psi:\mathbb{R}^2\rightarrow\mathbb{R}$ of order $4$, i.e., $\psi\in\rm{PL}(4)$:
$$
\frac{1}{p}\sum_{i=1}^p\psi(\bSi_{i,i},\sqrt{p}\betab^\star_i) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(\Lambda,B)\sim\mu}\left[ \psi(\Lambda,B)\right].
$$
\subsection{Asymptotic distribution and risk characterizations}
\fx{In this section, we prove our main result Theorem \ref{thm:master_W2}. Recall that ${\boldsymbol{\hat{\beta}}}$ is the min-norm solution. Since the distribution of ${\boldsymbol{\hat{\beta}}}$ depends on the problem dimensions (as it is a function of ${\mtx{X}},\vct{y}$), when necessary, we will use ${\boldsymbol{\hat{\beta}}}_n$ notation to make its dimension dependence explicit.} Let ${\boldsymbol{\hat{\beta}}}^P={\cal{P}}({\boldsymbol{\hat{\beta}}})$ be a pruned version of the min-norm solution ${\boldsymbol{\hat{\beta}}}$. Recall from Section \ref{sec:risk}, that the first crucial step in characterizing the risk ${\cal{L}}({\boldsymbol{\hat{\beta}}}^P)$ is studying the risk of a threshold-based pruned vector ${\cal{L}}({\boldsymbol{\hat{\beta}}}^\mathcal{T}_t)$. Theorem \ref{thm:master_W2} below shows how this is possible.
To keep things slightly more general, consider ${\boldsymbol{\hat{\beta}}}^{g}$ defined such that $\sqrt{p}{\boldsymbol{\hat{\beta}}}^{g}=g(\sqrt{p}{\boldsymbol{\hat{\beta}}})$, where $g$ is a Lipschitz function acting entry-wise on ${\boldsymbol{\hat{\beta}}}$ (for example, $g$ can be the \fx{(arbitrarily close Lipschitz approximation of the)} thresholding operator $\mathcal{T}_t$ of Section \ref{sec:risk}). Then, the risk of ${\boldsymbol{\hat{\beta}}}^g$ can be written as
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^g) &= \operatorname{\mathbb{E}}_{\mathcal{D}}[(\vct{x}^T({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^g) + \sigma z)^2] = \sigma^2 + ({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^g)^T\boldsymbol{\Sigma}({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^g) \notag \\
&= \sigma^2 + \frac{1}{p}\sum_{i=1}^p\boldsymbol{\Sigma}_{i,i}\big(\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\hat{\beta}}}_i)\big)^2\notag
\\
&=: \sigma^2 + \frac{1}{p}\sum_{i=1}^p f\big(\sqrt{p}{\boldsymbol{\hat{\beta}}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}\big),\label{eq:risk_app_f}
\end{align}
where in the last line, we defined function $f:\mathbb{R}^3\rightarrow\mathbb{R}$ as follows:
\begin{align}\label{eq:fdef}
f_{\cal{L}}(x,y,z) := \fx{z(y-g(x))}^2.
\end{align}
The following theorem establishes the asymptotic limit of \eqref{eq:risk_app_f}. For the reader's convenience, we repeat the notation introduced in Definition \ref{def:Xi}.
Let random variables $(\Lambda,B)\sim \mu$ (where $\mu$ is defined in Assumption \ref{ass:mu}) and fix $\kappa>1$. Define parameter $\xi$ as the unique positive solution to the following equation
$$
\operatorname{\mathbb{E}}_{\mu}\Big[ \big({1+(\xi\cdot\Lambda)^{-1}}\big)^{-1} \Big] = {\kappa^{-1}}\,.
$$
Further define the positive parameter $\gamma$ as follows:
$$
\hspace{-0.1in}\gamma :=
\Big({\sigma^2 + \operatorname{\mathbb{E}}_{\mu}\Big[\frac{B^2\Lambda}{(1+\xi\Lambda)^2}\Big]}\Big)\Big/\Big({1-\kappa\operatorname{\mathbb{E}}_{\mu}\Big[\frac{1}{\left(1+(\xi\Lambda)^{-1}\right)^2}\Big]}\Big).
$$
With these and $H\sim\mathcal{N}(0,1)$, define the random variable
$$
X_{\kappa,\sigma^2}:=X_{\kappa,\sigma^2}(\Lambda,B,H) := \Big(1-\frac{1}{1+ \xi\Lambda}\Big) B + \sqrt{\kappa}\frac{\sqrt{\gamma}\,\Lambda^{-1/2}}{1+(\xi\Lambda)^{-1}} H,
$$
and let $\Pi_{\kappa,\sigma^2}$ be its distribution.
\mainthm*
Before we prove the theorem, let us show how it immediately leads to a sharp prediction of the risk behavior. Indeed, a direct application of \eqref{eq:thm} for $f=f_{\cal{L}}$ to \eqref{eq:risk_app_f} shows that
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^g)\stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}_{(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)}\left[f_{\cal{L}}(X_{\kappa,\sigma^2},B,\Lambda) \right] = \sigma^2 + \operatorname{\mathbb{E}}_{(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)}\left[\Sigma\left(B-g(X_{\kappa,\sigma^2})\right)^2 \right].\label{eq:risk_app_f2}
\end{align}
We further remark on the following two consequences of Theorem \ref{thm:master_W2}.
First, since \eqref{eq:thm} holds for any $\rm{PL}(2)$ function, we have equivalently that $\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma})$ converges in Wasserstein-2 distance to $\Pi_{\kappa,\sigma^2}\otimes\mu$, where recall that $\Pi_{\kappa,\sigma^2}$ is the distribution of the random variable $X_{\kappa,\sigma^2}$.
Second, the theorem implies that the empirical distribution of $\sqrt{p}{\boldsymbol{\hat{\beta}}}_n$ converges weakly to $\Pi_{\kappa,\sigma^2}$. To see this, apply \eqref{eq:thm} for the $\rm{PL}(2)$ function $f(x,y,z) = \psi(x)$ where $\psi:\mathbb{R}\rightarrow\mathbb{R}$ is a bounded Lipschitz test function.
\subsection{Proof of Theorem \ref{thm:master_W2}}
\vspace{5pt}
Let ${\mtx{X}}\in\mathbb{R}^{n\times p}$ have zero-mean and normally distributed rows with a diagonal covariance matrix ${\boldsymbol{{\Sigma}}}=\operatorname{\mathbb{E}}[\vct{x}\x^T]$. Given a ground-truth vector $\betab^\star$ and labels $\vct{y}={\mtx{X}}\betab^\star+\sigma {\vct{z}},~{\vct{z}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, we consider the least-squares problem subject to the minimum Euclidian norm constraint (as $\kappa=p/n>1$) given by
\begin{align}\label{eq:PO_beta}
\min_{{\boldsymbol{\beta}}}\frac{1}{2}\tn{{\boldsymbol{\beta}}}^2\quad\text{subject to}\quad \vct{y}={\mtx{X}}{\boldsymbol{\beta}}.
\end{align}
It is more convenient to work with the following change of variable:
\begin{align}\label{eq:w}
\vct{w}:=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\beta}}-\betab^\star).
\end{align}
With this, the optimization problem in \eqref{eq:min_norm} can be rewritten as
\begin{align}\label{eq:PO}
\Phi({\mtx{X}})=\min_{\vct{w}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
where we write ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$. First, using standard arguments, we show that the solution of \eqref{eq:PO} is bounded. Hence, we can constraint the optimization in a sufficiently large compact set without loss of generality.
\begin{lemma}[Boundedness of the solution]\label{lem:bd_PO}
Let $\widehat{\vct{w}}_n:=\widehat{\vct{w}}_n({\mtx{X}},{\vct{z}})$ be the minimizer in \eqref{eq:PO}. Then, with probability approaching 1, it holds that $\widehat{\vct{w}}_n\in\mathcal{B}$, where
$$\mathcal{B}:=\left\{\vct{w}\,|\,\|\vct{w}\|_2\leq B_{+} \right\},\qquad B_+:=5\sqrt{\frac{\Sigma_{\max}}{\Sigma_{\min}}}\frac{\sqrt{\kappa}+1}{\sqrt{\kappa}-1}(\sqrt{\Sigma_{\max}\operatorname{\mathbb{E}}\left[B^2\right]} + \sigma).
$$
\end{lemma}
\begin{proof}
First, we show that the min-norm solution ${\boldsymbol{\hat{\beta}}}={\mtx{X}}^T({\mtx{X}}\X^T)^{-1}\vct{y}$ of \eqref{eq:PO_beta} is bounded. Note that $\kappa>1$, thus ${\mtx{X}}\X^T$ is invertible wpa 1. We
have,
\begin{align}
\tn{{\boldsymbol{\hat{\beta}}}_n}^2 = \vct{y}^T({\mtx{X}}\X^T)^{-1}\vct{y} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{X}}\X^T)} = \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\boldsymbol{\Sigma}{\mtx{\bar{X}}}^T)} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\Xb^T)\,\Sigma_{\min}} = \frac{\tn{\vct{y}}^2}{\sigma_{\min}^2({\mtx{\bar{X}}})\,\Sigma_{\min}}. \label{eq:Ubb}
\end{align}
But, wpa 1,
$
\sigma_{\min}({\mtx{\bar{X}}})/\sqrt{n} \geq \frac{1}{2}\left(\sqrt{\kappa}-1\right).
$
Furthermore,
$
\|\vct{y}\|_2 \leq \|{\mtx{\bar{X}}}\boldsymbol{\Sigma}^{1/2}\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2 \leq \sigma_{\max}({\mtx{\bar{X}}})\sqrt{\Sigma_{\max}} \|\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2.
$
Hence, wpa 1,
$$
\|\vct{y}\|_2/\sqrt{n} \leq 2(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma,
$$
where we used the facts that wpa 1: $\|z\|_2/\sqrt{n}\stackrel{{P}}{\longrightarrow} 1$, $\sigma_{\max}({\mtx{\bar{X}}})<\sqrt{2n}(\sqrt{\kappa}+1)$ and \cts{by Assumption \ref{ass:mu}}:
$$
\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^{p}(\sqrt{p}\betab^\star_i)^2 \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[B^2\right].
$$
Put together in \eqref{eq:Ubb}, shows that
\begin{align}
\tn{{\boldsymbol{\hat{\beta}}}_n} < \frac{2(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma}{\sqrt{\Sigma_{\min}}(\sqrt{\kappa}-1)/2} =: \tilde{B}_+.\label{eq:bd_beta}
\end{align}
Recalling that $\widehat{\vct{w}}_n= \sqrt{\boldsymbol{\Sigma}}{\boldsymbol{\hat{\beta}}}_n - \sqrt{\boldsymbol{\Sigma}}\betab^\star$, we conclude, as desired, that wpa 1,
$
\tn{\widehat{\vct{w}}_n} \leq \sqrt{\Sigma_{\max}}\tilde{B}_+ + \sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}[B^2]} \leq B_+.
$
\end{proof}
Lemma \ref{lem:bd_PO} implies that nothing changes in \eqref{eq:PO} if we further constrain $\vct{w}\in\mathcal{B}$ in \eqref{eq:PO}. Henceforth, with some abuse of notation, we let
\begin{align}\label{eq:PO_bd}
\Phi({\mtx{X}}):=\min_{\vct{w}\in\mathcal{B}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
Next, in order to analyze the primary optimization (PO) problem in \eqref{eq:PO_bd} in apply the CGMT \cite{thrampoulidis2015lasso}. Specifically, we use the constrained formulation of the CGMT given by Theorem \ref{thm closed}. Specifically, the auxiliary problem (AO) corresponding to \eqref{eq:PO_bd} takes the following form with ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$
\begin{align}
\phi({\vct{g}},\vct{h}) = \min_{\vct{w}\fx{\in\mathcal{B}}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_con}
\end{align
We will prove the following techincal result about the AO problem.
\somm{Re-specify the assumptions of this lemma!}
\begin{lemma}[Properties of the AO -- Overparameterized regime]\label{lem:AO}
Let $\phi_n=\phi({\vct{g}},\vct{h})$ be the optimal cost of the minimization in \eqref{eq:AO_con}. Define $\bar\phi$ as the optimal cost of the following deterministic min-max problem
\begin{align}\label{eq:AO_det}
\bar\phi:=\max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau):=\frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \fx{\operatorname{\mathbb{E}}\left[\frac{B^2}{1+\frac{u}{\tau}\Lambda}\right]} \right).
\end{align}
The following statements are true.
\noindent{(i).}~The AO minimization in \eqref{eq:AO_con} is $\frac{1}{\Sigma_{\max}}$-strongly convex and has a unique minimizer $\widehat{\vct{w}}_n:=\widehat{\vct{w}}_n({\vct{g}},\vct{h})$.
\noindent{(ii).}~In the limit of $n,p\rightarrow\infty, p/n=\kappa$, it holds that $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$, i.e., for any $\varepsilon>0$:
$$
\lim_{n\rightarrow\infty}\P\left(|\phi({\vct{g}},\vct{h})-\bar\phi|>\varepsilon\right) = 0.
$$
\noindent{(iii).} The max-min optimization in \eqref{eq:AO_det} has a unique saddle point $(u_*,\tau*)$ satisfying the following:
$$
u_*/\tau_* = \xi\quad\text{and}\quad\tau_* = \gamma,
$$
where $\xi, \gamma$ are defined in Definition \ref{def:Xi}.
\noindent{(iv).}~Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be a $\rm{PL}({k})$ function. Let ${\boldsymbol{\hat{\beta}}}_n=\boldsymbol{\Sigma}^{-1/2}\widehat{\vct{w}}_n + \betab^\star$. Then,
$$
\frac{1}{p}\sum_{i=1}^{p}f\left(\sqrt{p}{\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\sqrt{p}\betab^\star,\boldsymbol{\Sigma}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(B,\Lambda,H)\sim\mu\otimes \mathcal{N}(0,1)}\left[f\left(X_{\kappa,\sigma^2}(B,\Lambda,H),B,\Lambda\right) \right].
$$
\noindent{(v).}~\cts{The empirical distribution of ${\boldsymbol{\hat{\beta}}}_n$ converges weakly to the measure of $X_{\kappa,\sigma^2}$, and also, for large enough absolute constant $C>0$:
\begin{align}\label{eq:k_AO}
\sum_{i=1}^{p}{\boldsymbol{\hat{\beta}}}^2_n({\vct{g}},\vct{h}) < C.
\end{align}
}
\end{lemma}
We prove Lemma \ref{lem:AO} in Section \ref{sec:proofAO}. Here, we show how this leads to the proof of Theorem \ref{thm:master_W2} when combined with the CGMT framework \cite{thrampoulidis2015lasso}.
\cts{Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be a $\rm{PL}(2)$ function, or, the function $f_{\cal{L}}$ defined in \eqref{eq:fdef}.} For convenience,
define
$$F_n({\boldsymbol{\hat{\beta}}},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Fix any $\varepsilon>0$ and define the set
\begin{align}\label{eq:S_set}
\mathcal{S} = \mathcal{S}(\betab^\star,\boldsymbol{\Sigma}) =\{{\boldsymbol{\beta}}{~\big |~} |F_n({\boldsymbol{\hat{\beta}}}_n,\betab^\star,\boldsymbol{\Sigma})-\alpha_*|\geq 2\varepsilon\}.
\end{align}
It suffices to prove that the solution ${\boldsymbol{\hat{\beta}}}_n$ of the PO in \eqref{eq:PO_beta} satisfies ${\boldsymbol{\hat{\beta}}}_n\not\in\mathcal{S}$ wpa 1. To see that this is sufficient, note the following. On the one hand, setting $f=f_{\cal{L}}$, directly proves \eqref{eq:thm}. On the other hand, recall that $W_2$-convergence is equivalent to convergence of any $\rm{PL}(2)$ test function $f$.
To prove the desired, we need to consider the ``perturbed" PO and AO problems (compare to \eqref{eq:PO} and \eqref{eq:AO_con}) as:
\begin{align}\label{eq:PO_S}
\Phi_S({\mtx{X}})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
and
\begin{align}
\phi_S({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_S}
\end{align
Recall here, that ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$, ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$ and we have used the change of variables $\vct{w}:=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\beta}}-\betab^\star)$ for convenience.
Using \cite[Theorem 6.1(iii)]{thrampoulidis2018precise} it suffices to find costants $\bar\phi, \bar\phi_S$ and $\eta>0$ such that the following three conditions hold:
\begin{enumerate}
\item $\bar\phi_S \geq \bar\phi + 3\eta$,
\item $\phi({\vct{g}},\vct{h}) \leq \bar\phi + \eta$, with probability approaching 1,
\item $\phi_S({\vct{g}},\vct{h}) \geq \bar\phi_S - \eta$, with probability approaching 1.
\end{enumerate}
In what follows, we explicitly find $\bar\phi, \bar\phi_S,\eta$ such that the three conditions above hold.
\vspace{5pt}
\noindent\underline{Satisfying Condition 2}: Recall the deterministic min-max optimization in \eqref{eq:AO_det}. Choose $\bar\phi=\mathcal{D}(u_*,\tau_*)$ be the optimal cost of this optimization. From Lemma \ref{lem:AO}(ii), $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$. Thus, for any $\eta>0$, with probability approaching 1:
\begin{align}\label{eq:phi_lim}
\bar\phi + \eta \geq \phi({\vct{g}},\vct{h}) \geq \bar\phi - \eta.
\end{align}
Clearly then, Condition 2 above holds for any $\eta>0$.
\vspace{5pt}
\noindent\underline{Satisfying Condition 3}: Next, we will show that the third condition holds for appropriate $\bar\phi$.
Let $\widehat{\vct{w}}_n=\widehat{\vct{w}}_n({\vct{g}},\vct{h})$ be the unique minimizer of \eqref{eq:AO_con} as per Lemma \ref{lem:AO}(i), i.e., $\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\widehat{\vct{w}}_n+{\boldsymbol{\beta}}}^2 = \phi({\vct{g}},\vct{h})$. Again from Lemma \ref{lem:AO}, the minimization in \eqref{eq:AO_con} is $1/\Sigma_{\max}$-strongly convex in $\vct{w}$. Here, $\Sigma_{\max}$ is the upper bound on the eigenvalues of $\boldsymbol{\Sigma}$ as per Assumption \ref{ass:mu}. Thus, for any $\tilde\varepsilon>0$ and any feasible $\vct{w}$ the following holds (deterministically):
\begin{align}\label{eq:sc}
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+{\boldsymbol{\beta}}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2\fx{\Sigma_{\max}}},~\text{provided that}~ \|\vct{w}-\widehat{\vct{w}}_n({\vct{g}},\vct{h})\|_2 \geq \tilde\epsilon.
\end{align}
Now, we argue that \fx{wpa 1,}
\begin{align}\label{eq:dev_arg}
\fx{\text{for all}~{\boldsymbol{\beta}}\in \mathcal{S}~\text{the corresponding}~\vct{w}~\text{obeys}}~\|\vct{w}-\widehat{\vct{w}}_n({\vct{g}},\vct{h})\|_2\geq \tilde\varepsilon,
\end{align}
for an appropriate value of a constant $\tilde\varepsilon>0$ and $\vct{w}=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\beta}}-\betab^\star)$. Consider any ${\boldsymbol{\beta}}\in\mathcal{S}$.
\noindent First, by definition in \eqref{eq:S_set},
$$
|F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \geq 2\varepsilon.
$$
Second, by Lemma \ref{lem:AO}(iv), with probability approaching 1,
$$
|F({\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \leq \epsilon.
$$
Third, we will show that wpa 1, there exists universal constant $C>0$ such that
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})| \leq C {\|{\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}) - {\boldsymbol{\beta}}\|_2}\label{eq:dev2show}.
\end{align}
Before proving \eqref{eq:dev2show}, let us argue how combining the above three displays shows the desired. Indeed, in that case, wpa 1,
\begin{align*}
2\varepsilon &\leq |F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \leq |F_n({\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})| + |F_n({\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \\
&\leq \epsilon + C \,\|{\boldsymbol{\beta}}-{\boldsymbol{\hat{\beta}}}_n\|_2. \\
&\qquad\Longrightarrow \|{\boldsymbol{\beta}}-{\boldsymbol{\hat{\beta}}}_n\|_2 \geq {\varepsilon}/{C}=:\hat\varepsilon\\
&\qquad\Longrightarrow \|\vct{w}-\widehat{\vct{w}}_n\|_2 \geq \hat\varepsilon\sqrt{\Sigma_{\min}}=:\tilde\varepsilon.
\end{align*}
In the last line above, we recalled that ${\boldsymbol{\beta}}=\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star$ and $\boldsymbol{\Sigma}_{i,i}\geq\Sigma_{\min},~i\in[p]$ by Assumption \ref{ass:mu}. This proves \eqref{eq:dev_arg}.
Next, combining \eqref{eq:dev_arg} and \eqref{eq:sc}, we find that wpa 1,
$
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+{\boldsymbol{\beta}}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2C},~\text{for all}~ \vct{w}\in\mathcal{S}.
$
Thus,
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}.\notag
\end{align}
When combined with \eqref{eq:phi_lim}, this shows that
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta.
\end{align}
Thus, choosing $\bar\phi_S = \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}$ proves the Condition 3 above.
\vspace{3pt}
\noindent\fx{\textbf{Perturbation analysis via Pseudo-Lipschitzness:}} To complete the proof, let us now show \eqref{eq:dev2show}. Henceforth, $C$ is used to denote a universal constant whose value can change from line to line.
First, consider the case $f=f_{\cal{L}}$, i.e., $f(x,y,z) = y(x-g(z))^2$. We have the following chain of inequalities:
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})| &=
\frac{1}{p}\sum_{i=1}^p|\bSi_{i,i}|\left| (\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i}))^2-(\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\beta}}_{i}))^2 \right|\notag\\
&\leq
\Sigma_{\max} \frac{1}{p}\sum_{i=1}^p \left| (\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i}))^2-(\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\beta}}_{i}))^2 \right|\notag\\
&\leq
\fx{C} \Sigma_{\max} \frac{1}{p}\sum_{i=1}^p (1+ \|\sqrt{p}[\betab^\star_i,{\boldsymbol{\hat{\beta}}}_{n,i}]\|_2 + \|\sqrt{p}[\betab^\star_i,{\boldsymbol{\beta}}_{i}]\|_2) \sqrt{p}|{\boldsymbol{\hat{\beta}}}_{n,i}-{\boldsymbol{\beta}}_{i}|\notag\\
&\leq
\fx{C} \Sigma_{\max} \Big(1+ \frac{1}{\sqrt{p}}\big(\sum_{i=1}^{p}\|\sqrt{p}[\betab^\star_i,{\boldsymbol{\hat{\beta}}}_{n,i}]\|_2^2\big)^{1/2} + \frac{1}{\sqrt{p}}\big(\sum_{i=1}^p\|\sqrt{p}[\betab^\star_i,{\boldsymbol{\beta}}_{i}]\|_2^2\big)^{1/2}\Big) \|{\boldsymbol{\hat{\beta}}}_n-{\boldsymbol{\beta}}\|_2\notag\\
&\leq
C \left(1+ \max\{\|\betab^\star\|_2^2,\|{\boldsymbol{\hat{\beta}}}_n\|_2^2,\|{\boldsymbol{\beta}}\|_2^2\}^{1/2} \right) \|{\boldsymbol{\hat{\beta}}}_n-{\boldsymbol{\beta}}\|_2.\label{eq:fcase1}
\end{align}
In the second line above, we used boundedness of $\bSi_{i,i}$ as per Assumption \ref{ass:mu}. \somm{Consider explaining the PL(2) sentence.}\fx{In the third line, we used the fact that the function $\psi(a,b) = (a-g(b))^2$ is $\rm{PL}(2)$.} The fourth line follows by Cauchy-Schwartz inequality. Finally, in the last line, we used the elementary fact that $a+b+c\leq 3\max\{a,b,c\}$ for $a=2\sum_{i=1}^p(\betab^\star_i)^2$ and $b=\sum_{i=1}^p{\boldsymbol{\hat{\beta}}}_{n,i}^2$ and $c=\sum_{i=1}^p{\boldsymbol{\beta}}_{i}^2$.
Second, consider the case $f\in\rm{PL}(2)$. Let $\vct{a}_i=(\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i}({\vct{g}},\vct{h}),\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}{\boldsymbol{\beta}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$, for $i\in[p]$. A similar chain of inequality holds:
\somm{The solution for PL(k) is showing that around $\hat{{\boldsymbol{\beta}}}$, there is an arbitarily close $\hat{{\boldsymbol{\beta}}}_{bdd}$ with bounded entries.}
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})| &\leq \frac{\fx{C}}{p}\sum_{i=1}^p (1+\max\{\|\vct{a}_i\|_2,\|\vct{b}_i\|_2\})\|\vct{a}_i-\vct{b}_i\|_2\notag \\
&= \frac{\fx{C}}{p}\sum_{i=1}^p \left(1+\max\{\|\vct{a}_i\|_2,\|\vct{b}_i\|_2\}\right)\cdot|\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i}({\vct{g}},\vct{h}) - \sqrt{p}{\boldsymbol{\beta}}_i|\notag\\
&\leq \fx{C}\left(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2}\} \right)^{1/2}{\|{\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}) - {\boldsymbol{\beta}}\|_2}\notag\\
&\leq C\left(1+\max\{\|{\boldsymbol{\hat{\beta}}}_n\|_2^2,\|\betab^\star\|_2^2,\|{\boldsymbol{\beta}}\|_2^2,\frac{1}{p}\sum_{i=1}^p\bSi_{i,i}^2\} \right)^{1/2}{\|{\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}) - {\boldsymbol{\beta}}\|_2}\notag\\
&\leq C\Sigma_{\max}^2\left(1+\max\{\|{\boldsymbol{\hat{\beta}}}_n\|_2^2,\|\betab^\star\|_2^2,\|{\boldsymbol{\beta}}\|_2^2\} \right)^{1/2}{\|{\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}) - {\boldsymbol{\beta}}\|_2}.\label{eq:fcase2}
\end{align}
The first inequality uses the fact that $f\in\rm{PL}(2)$. The second inequality in the third line follows by Cauchy-Schwartz. The last inequality used Assumption \ref{ass:mu} on boundedness of $\bSi_{i,i}$.
It follows from either \eqref{eq:fcase1} or \eqref{eq:fcase2} that in order to prove \eqref{eq:dev2show}, we need to show boundedness of the following terms: $\|{\boldsymbol{\hat{\beta}}}_n\|_2$, $\|\betab^\star\|_2$ and $\|{\boldsymbol{\beta}}\|_2$. By feasibility of ${\boldsymbol{\hat{\beta}}}_n$ and ${\boldsymbol{\beta}}$, we know that ${\boldsymbol{\hat{\beta}}}_n,{\boldsymbol{\beta}}\in\mathcal{B}$. Thus, the desired $\|{\boldsymbol{\beta}}\|_2<\infty$ and $\|{\boldsymbol{\hat{\beta}}}_n\|_2<\infty$ follow directly by Lemma \ref{lem:bd_PO}. (Alternatively, for ${\boldsymbol{\hat{\beta}}}_n$ we conclude the desired by directly applying Lemma \ref{lem:AO}(v)). Finally, to prove $\|\betab^\star\|_2<\infty$, note that
$$
\|\betab^\star\|_2^2 =\frac{1}{p}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^2,
$$
\cts{which is bounded wpa 1 by Assumption \ref{ass:mu}, which implies bounded second moments of $\sqrt{p}\betab^\star$. } This completes the proof of \eqref{eq:dev2show}, as desired.
\vspace{5pt}
\noindent\underline{Satisfying Condition 1:} To prove Condition 1, we simply pick $\eta$ to satisfy the following
\begin{align}
\bar\phi_S > \bar\phi + 3 \eta ~\Leftarrow~ \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta \geq \bar\phi + 3 \eta ~\Leftarrow~ \eta \leq \frac{\tilde\epsilon^2}{8\Sigma_{\max}}.\notag
\end{align}
This completes the proof of Theorem \ref{thm:master_W2}.
\subsection{Proof of Lemma \ref{lem:AO}}\label{sec:proofAO}
~~~~
~~~~
\vspace{3pt}
\subsubsection{Proof of (i).} Strong convexity of the objective function in \eqref{eq:PO} is easily verified by the second derivative test. Note here that we use Assumption \ref{ass:mu} that $\bSi_{i,i}\leq\Sigma_{\max},~i\in[p].$ Uniqueness of the solution follows directly from strong convexity. \ct{Strictly speaking we might need to also argue existence, i.e., feasibility of the AO. An indirect way is to show feasibility using the CGMT, but it seems unnecessarily complicated?}
\vspace{5pt}
\subsubsection{Proof of (ii).}Using Lagrangian formulation, the solution $\widehat{\vct{w}}_n$ to \eqref{eq:AO_con} is the same as the solution to the following:
\begin{align}
\left(\widehat{\vct{w}}_n,u_n\right) :=\arg\min_{\vct{w}\in\mathcal{B}}\max_{u\geq 0} ~\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + u \left( \sqrt{\tn{\vct{w}}^2+\sigma^2} \tn{\bar{\g}} - \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} + \frac{\sigma h}{\sqrt{n}} \right)\label{eq:AO_2}
\end{align}
where we have: (i) set $\bar{\g} := {\vct{g}}/\sqrt{n}$ and $\bar{\h}:= \vct{h}/\sqrt{p}$; (ii) recalled that $p/n=\kappa$; and, (iii) used $\left(\widehat{\vct{w}}_n,u_n\right)$ to denote the optimal solutions in \eqref{eq:AO_2}. The subscript $n$ emphasizes the dependence of $\left(\widehat{\vct{w}}_n,u_n\right)$ on the problem dimensions. Also note that (even though not explicit in the notation) $\left(\widehat{\vct{w}}_n,u_n\right)$ are random variables depending on the realizations of $\bar{\g},\bar{\h}$ and $h$.
Notice that the objective function above is convex in $\vct{w}$ and linear (thus, concave) in $u$. Thus, strong duality holds and we can flip the order of min-max. Moreover, in order to make the objective easy to optimize with respect to $\vct{w}$, we use the following variational expression for the square-root term $\sqrt{\tn{\vct{w}}^2+\sigma^2}$:
$$
\tn{\bar{\g}}\sqrt{\tn{\vct{w}}^2+\sigma^2} = \tn{\bar{\g}}\cdot\min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau}{2} + \frac{\tn{\vct{w}}^2+\sigma^2}{2\tau} \right\} = \min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau\tn{\bar{\g}}^2}{2} + \frac{\sigma^2}{2\tau} + \frac{\tn{\vct{w}}^2}{2\tau} \right\},
$$
where $B_+$ is defined in Lemma \ref{lem:bd_PO}. For convenience define the constraint set for the variable $\tau$ as $\mathcal{T}':=[\sigma,\sqrt{\sigma^2+B_+^2}]$. For reasons to be made clear later in the proof (see proof of statement (iii)), we consider the (possibly larger) set:
\[
\mathcal{T}:=[\sigma,\max\{\sqrt{\sigma^2+B_+^2},2\tau_*\}]\,
\]
where $\tau_*$ is as in the statement of the lemma.
The above lead to the following equivalent formulation of \eqref{eq:AO_2}:
\begin{align}
\left(\widehat{\vct{w}}_n,u_n,\tau_n\right) = \max_{u\geq 0}\min_{\vct{w}\in\mathcal{B},\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} + \min_{\vct{w}\in\mathcal{B}} \left\{ \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + \frac{u}{2\tau}\tn{\vct{w}}^2 - u \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} \right\} .\label{eq:AO_3}
\end{align}
The minimization over $\vct{w}$ is easy as it involves a strongly convex quadratic function. First, note that unconstrained optimal $\vct{w}':=\vct{w}'(\tau,u)$ (for fixed $(\tau,u)$) is given by
\begin{align}\label{eq:w'}
\vct{w}':=\vct{w}'(\tau,u) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right),
\end{align}
and \eqref{eq:AO_3} simplifies to
\begin{align}
\left(u_n,\tau_n\right)=\max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} - \frac{1}{2} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)\,=:\mathcal{R}(u,\tau) .\label{eq:AO_4}
\end{align}
\somm{Does $u=0$ scenario create a problem in saddle point uniqueness?}
It can be checked by direct differentiation and the second-derivative test that the objective function in \eqref{eq:AO_4} is strictly convex in $\tau$ and strictly concave in $u$ over the domain $\{(u,\tau)\in\mathbb{R}_+\times\mathbb{R}_+\}$ \footnote{\fx{To analyze the matrix-vector product term in \eqref{eq:AO_4} for $(\tau,u)$ one can use the fact that ${\boldsymbol{{\Sigma}}}$ is diagonal. This way, as a function of $u$ and $\tau$ the analysis reduces to the properties of relatively simple functions. For instance, for $\tau$, this function is in the form $f(\tau)=-(a+b/\tau)^{-1}$ for $a,b>0$ which is strictly convex.}}. Thus, the saddle point $(u_n,\tau_n)$ is unique. Specifically, this implies that the optimal $\widehat{\vct{w}}_n$ in \eqref{eq:AO_3} is given by (cf. \eqref{eq:w'})
\begin{align}\label{eq:w_n}
\widehat{\vct{w}}_n=\vct{w}'(\tau_n,u_n) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).
\end{align}
In Lemma \ref{lem:AO}(v) we will prove that wpa 1, in the limit of $p\rightarrow\infty$, $\|\widehat{\vct{w}}_n\|_2\leq C$ for sufficiently large absolute constant $C>0$. Thus, by choosing the upper bound in the definition of $\mathcal{B}$ in Lemma \ref{lem:bd_PO} strictly larger than C, guarantees that the unconstrained $\widehat{\vct{w}}_n$ in \eqref{eq:w_n} is feasible in \eqref{eq:AO_3}.
\noindent\fx{\textbf{Asymptotic limit of the key quantities $\tau_n,u_n$:}} In what follows, we characterize the high-dimensional limit of the optimal pair $(u_n,\tau_n)$ in the limit $n,p\rightarrow\infty,~p/n\rightarrow\kappa$.
We start by analyzing the (point-wise) convergence of $\mathcal{R}(u,\tau)$.
For the first three summands in \eqref{eq:AO_4}, we easily find that
$$
\left\{\frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}}\right\}~ \stackrel{{P}}{\longrightarrow}~\left\{
\frac{u\tau}{2} + \frac{u\sigma^2}{2\tau} \right\}.
$$
Next, we study the fourth summand. First, note that
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}(u\sqrt{\kappa}\bar{\h}) &= u^2\kappa\,\frac{1}{p}\vct{h}^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\vct{h} \notag\\
&= u^2\kappa\,\frac{1}{p}\sum_{i=1}^{p}\frac{\vct{h}_i^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}} \notag\\
&\stackrel{{P}}{\longrightarrow} u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right].
\end{align}
\somm{Explain/justify PL(2)'ness of these functions. Clearly explain how A2 is used\dots}
In the last line, $\Lambda$ is a random variable as in Definition \ref{def:Xi}. \fx{Also, we used Assumption \ref{ass:mu} together with the facts that $\vct{h}$ is independent of $\boldsymbol{\Sigma}$ and that the function $(x_1,x_2)\mapsto x_1^2(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(3)$ assuming $x_2$ is bounded (see Lemma \ref{lem:PLbdd} for proof).}
\ct{Question: Is it immediate that the empirical distribution of $(\beta,\boldsymbol{\Sigma},\vct{h})$ converges in $W_k$ to $\mu\otimes\mathcal{N}(0,1)$ given that $(\beta,\boldsymbol{\Sigma})$ converges to $\mu$ and $\vct{h}$ is independent???}
Second, we find that
\begin{align}
(\betab^\star)^T\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star &= \frac{1}{p}(\sqrt{p}\betab^\star)^T\left({\mtx{I}}+\frac{u}{\tau}\boldsymbol{\Sigma}\right)^{-1}(\sqrt{p}\betab^\star) \notag\\
&= \frac{1}{p}\sum_{i=1}^{p}\frac{\left(\sqrt{p}\betab^\star_i/\sqrt{\bSi_{i,i}}\right)^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}}\notag\\
&\stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right].
\end{align}
Here, $\Lambda,B$ are random variables as in Definition \ref{def:Xi} and \fx{we also used Assumption \ref{ass:mu} together with the fact that the function $(x_1,x_2)\mapsto x_1^2x_2^{-1}(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(2)$ assuming $x_2$ is bounded (see Lemma \ref{lem:PLbdd} for proof).}
Third, \cts{by independence of $(\betab^\star, \boldsymbol{\Sigma})$ from $\vct{h}$}
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star = u\sqrt{\kappa} \cdot \frac{1}{p}\sum_{i=1}
^p \frac{\vct{h}_i\boldsymbol{\Sigma}_{i,i}^{-1/2}(\sqrt{p}\betab^\star_i)}{\boldsymbol{\Sigma}_{i,i}^{-1}+\frac{u}{\tau}{\mtx{I}}} ~\stackrel{{P}}{\longrightarrow}~ 0.
\end{align}
Putting these together, the objective $\mathcal{R}(u,\tau)$ in \eqref{eq:AO_4} converges point-wise in $u,\tau$ to
\begin{align}
\mathcal{R}(u,\tau)\stackrel{{P}}{\longrightarrow}\mathcal{D}(u,\tau) := \frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).\label{eq:conv_pt}
\end{align}
Note that $\mathcal{R}(u,\tau)$ (and thus, $\mathcal{D}(u,\tau)$) is convex in $\tau$ and concave in $u$. Thus, the convergence in \eqref{eq:conv_pt} is in fact uniform (e.g., \cite{AG1982}) and we can conclude that
\begin{align}\label{eq:Dc0}
\phi({\vct{g}},\vct{h}) \stackrel{{P}}{\longrightarrow} \max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
and \fx{using strict concave/convexity of $\mathcal{D}(u,\tau)$, we also have the parameter convergence}
\somm{This line follows from strict convex/concavity right?}
\begin{align}\label{eq:Dc}
\fx{(u_n,\tau_n) \stackrel{{P}}{\longrightarrow} (u_*,\tau_*):=\arg\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).}
\end{align}
In the proof of statement (iii) below, we show that the saddle point of \eqref{eq:Dc0} is $(u_*,\tau_*)$. In particular, $\tau_*$ is strictly in the interior of $\mathcal{T}$, which combined with convexity implies that
$$
\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau) = \max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau) =: \bar\phi.
$$
This, together with the first display above proves the second statement of the lemma.
\vspace{5pt}
\subsubsection{Proof of (iii).} Next, we compute the saddle point $(u_*,\tau_*)$ by studying the first-order optimality conditions of the strictly concave-convex $\mathcal{D}(u,\tau)$. Specifically, we consider the unconstrained minimization over $\tau$ and we will show that the minimum is achieved in the strict interior of $\mathcal{T}$. Direct differentiation of $\mathcal{D}(u,\tau)$ gives
\begin{subequations}
\begin{align}
{\tau} + \frac{\sigma^2}{\tau} - 2u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] + \frac{u^2}{\tau}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] + \frac{1}{\tau}\operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0, \label{eq:fo1}\\
{u} - \frac{u\sigma^2}{\tau^2} - \frac{u^3}{\tau^2}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} + \frac{u}{\tau}\right)^2}\right] - \frac{u}{\tau^2} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0,\label{eq:fo2}
\end{align}
\end{subequations}
Multiplying \eqref{eq:fo2} with $\frac{\tau}{u}$ and adding to \eqref{eq:fo1} results in the following equation
\begin{align}
\tau = u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] ~\Leftrightarrow ~
\operatorname{\mathbb{E}}\left[ \frac{1}{(\frac{u}{\tau}\Lambda)^{-1}+1} \right] = \frac{1}{\kappa} \label{eq:tauu}\,.
\end{align}
Thus, we have found that the ratio $\frac{u_*}{\tau_*}$ is the unique solution to the equation in \eqref{eq:tauu}. Note that this coincides with the Equation \eqref{eq:ksi} that defines the parameter $\xi$ in Definition \ref{def:Xi}. The fact that \eqref{eq:tauu} has a unique solution for all $\kappa>1$ can be easily seen as $F(x)=\operatorname{\mathbb{E}}\left[ \frac{1}{(x\Lambda)^{-1}+1} \right], x\in\mathbb{R}_+$ has range $(0,1)$ and is strictly increasing (by differentiation).
Thus, we call $\xi=\frac{u_*}{\tau_*}$. Moreover, multiplying \eqref{eq:fo2} with $u$ leads to the following equation for $\tau_*$:
\begin{align}
u_*^2 = \sigma^2\xi^2 + u_*^2 \xi^2 \kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right] + \xi^2 \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right] ~\Rightarrow~ \tau_*^2 = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\xi^2\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right]} = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left((\xi\Lambda)^{-1} +1\right)^2}\right]}.
\end{align}
{Again, note that this coincides with Equation \eqref{eq:gamma} that determines the parameter $\gamma$ in Definition \ref{def:Xi}, i.e., $\tau_*^2 = \gamma.$ }
\vspace{5pt}
\subsubsection{Proof of (iv).} For convenience, define
$$F_n({\boldsymbol{\hat{\beta}}},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Recall from \eqref{eq:w_n} the explicit expression for $\widehat{\vct{w}}_n$, repeated here for convenience.
\begin{align}
\widehat{\vct{w}}_n = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).\notag
\end{align}
Also, recall that ${\boldsymbol{\hat{\beta}}}_n = \boldsymbol{\Sigma}^{-1/2}\widehat{\vct{w}}_n+\betab^\star$. Thus, (and using the fact that $\bar{\h}$ is distributed as $-\bar{\h}$),
\begin{align}
{\boldsymbol{\hat{\beta}}}_n &=\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}u_n\sqrt{\kappa}\bar{\h} + \left({\mtx{I}}-\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\right)\betab^\star\notag\\
\Longrightarrow{\boldsymbol{\hat{\beta}}}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\bar{\h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i.\label{eq:betan}
\end{align}
For $i\in[p]$, define
\begin{align}
\vct{v}_{n,i} = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\bar{\h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i
\end{align}
In the above, for convenience, we have denoted $\xi_n:=u_n/\tau_n$ and recall that $\xi_*:=u_*/\tau_*$.
The proof proceeds in two steps. In the first step, we use the fact that $\xi_n\stackrel{{P}}{\longrightarrow}\xi_*$ and $u_n\stackrel{{P}}{\longrightarrow} u_\star$ (see \eqref{eq:Dc}) to prove that for any $\varepsilon\in(0,\xi_*/2)$, there exists an absolute constant $C>0$ such that wpa 1:
\begin{align}\label{eq:ivstep1}
|F_n(\sqrt{p}{\boldsymbol{\hat{\beta}}}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.
\end{align}
In the second step, we use \fx{pseudo-Lipschitzness of $f$ and Assumption \ref{ass:mu} to prove that}
\somm{Doesn't Assump 3 come out of nowhere? So far it looks like we need A1 and A2. We need to ensure consistency.}
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:ivstep2}
\end{align}
The desired follows by combining \eqref{eq:ivstep1} and \eqref{eq:ivstep2}. Thus, in what follows, we prove \eqref{eq:ivstep1} and \eqref{eq:ivstep2}.
\vspace{2pt}
\noindent\underline{Proof of \eqref{eq:ivstep1}.}~~Fix some $\varepsilon\in(0,\xi_*/2)$. From \eqref{eq:Dc}, we know that w.p.a. 1 $|\xi_n-\xi_*|\leq \varepsilon$ and $|u_n-u_*|\leq \varepsilon$. Thus, $\widehat{\vct{w}}_n$ is close to $\vct{v}_n$. Specifically, in this event, for every $i\in[p]$, it holds that:
\somm{Fourth line: Constant missing here. Powers are wrongish}
\begin{align}
\notag|{\boldsymbol{\hat{\beta}}}_{n,i} - \vct{v}_{n,i}| &\leq {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{\tau_n}{1+(\xi_n\bSi_{i,i})^{-1}}-\frac{\tau_*}{1+(\xi_*\bSi_{i,i})^{-1}}\right| \\
\notag&= {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{u_n}{\xi_n+\bSi_{i,i}^{-1}}-\frac{u_*}{\xi_*+\bSi_{i,i}^{-1}}\right| \\
\notag& \leq
{|\betab^\star_i|}\frac{|\bSi_{i,i}||\xi_n-\xi_*|}{|1+\xi_n\bSi_{i,i}||1+\xi_*\bSi_{i,i}|} + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}} \frac{u_*|\xi_n-\xi_*|}{(\xi_n+\bSi_{i,i}^{-1})(\xi_*+\bSi_{i,i}^{-1})} +
\sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}}\frac{|u_n-u_*|}{\xi_n+\bSi_{i,i}^{-1}}
\\
\notag& \leq
{|\betab^\star_i|}{\Sigma_{\max}\varepsilon} + \sqrt{\kappa}{|\bar{\h}_i|} u_*\Sigma_{\max}^{3/2} \varepsilon+
\sqrt{\kappa}|\bar{\h}_i|\Sigma_{\max}^{1/2}\varepsilon
\\
&\leq \varepsilon \cdot \max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\, \left(|\bar{\h}_i| + |\betab^\star_i|\right).\label{eq:betav}
\end{align
In the second line above, we recalled that $u_n=\tau_n\xi_n$ and $u_*=\tau_*\xi_*$. In the third line, we used the triangle inequality. In the fourth line, we used that $\xi_*>0$, $0<\bSi_{i,i}\leq\Sigma_{\max}$ and $\xi_n\geq \xi_*-\varepsilon \geq \xi_*/2 >0$.
Now, we will use this and Lipschitzness of $f$ to argue that there exists absolute constant $C>0$ such that wpa 1:
\begin{align}
|F_n(\sqrt{p}{\boldsymbol{\hat{\beta}}}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.\notag
\end{align}
Denote, $\vct{a}_i=(\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\vct{v}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$. Following the exact same argument as in \eqref{eq:dev2show} (just substitute ${\boldsymbol{\beta}}\leftrightarrow\vct{v}_n$ in the derivation), we have that for some absolute constant $C>0$ wpa 1:
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| &\leq C \|{\boldsymbol{\hat{\beta}}}_n-\vct{v}_n\|_2.
\end{align}
From this and \eqref{eq:betav}, we find that
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})|
&\leq
C\varepsilon\max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2} \right\} \left(\sum_{i=1}^p \left(|\bar{\h}_i|+|\betab^\star_i|\right)^2\right)^{1/2}\notag
\\
&\leq
C\varepsilon\sqrt{2}\max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\sqrt{\|\betab^\star\|_2^2 + \|\bar{\h}\|_2^2}.\label{eq:epsS}
\end{align}
But, \cts{recall that $\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^2<\infty$, as $p\rightarrow \infty$ by Assumption \ref{ass:mu}.}
Also, since $\bar{\h}_i\sim\mathcal{N}(0,1/p)$, it holds that $\|\bar{\h}\|_2^2\leq 2$, wpa 1 as $p\rightarrow\infty$. Therefore, from \eqref{eq:epsS}, wpa 1, there exists constant $C>0$ such that
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \leq C \cdot\varepsilon,
\end{align}
as desired.
\vspace{2pt}
\noindent\underline{Proof of \eqref{eq:ivstep2}.}~~Next, we will use \ref{ass:mu} to show that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:AO_conv}
\end{align}
Notice that $\vct{v}_n$ is a function of $\betab^\star,\boldsymbol{\Sigma},\bar{\h}$. Concretely, define ${\widetilde{g}}:\mathbb{R}^3\rightarrow\mathbb{R}$, such that
$$
\widetilde{g}(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi_* x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1,
$$
and notice that
$$
\sqrt{p}\vct{v}_{n,i} = \widetilde{g}\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right) = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\vct{h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\sqrt{p}\betab^\star_i.
$$
Thus,
$$
F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma}) = \frac{1}{p}\sum_{i=1}^p f\left(g\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right),\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) =: \frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right),
$$
where we have defined $h:\mathbb{R}^3\rightarrow\mathbb{R}$:
\begin{align}
h(x_1,x_2,x_3) := f\left(\widetilde{g}(x_2,x_3,x_1),x_2,x_3\right).\label{h func}
\end{align}
\cts{We will prove that $h\in\rm{PL}(4)$. Indeed, if that were the case, then Assumption \ref{ass:mu} gives}
\begin{align}
\frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[h(H,B,\Lambda)\right] &= \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[f\left(\widetilde{g}(B,\Lambda,H),B,\Lambda\right))\right]\\
& = \operatorname{\mathbb{E}}[f\left(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda\right)] = \alpha_*,
\end{align}
where the penultimate equality follows by recognizing that (cf. Eqn. \eqref{eq:X})
$$
\widetilde{g}(B,\Lambda,H) = (1-(1+\xi_*\Lambda)^{-1})B + \sqrt{\kappa}\frac{\tau_*\Lambda^{-1/2}}{1+(\xi_*\Lambda)^{-1}}H = X_{\kappa,\sigma^2}(\Lambda,B,H).
$$
It remains to show that $h\in\rm{PL}(4)$. Lemma \ref{lem:hPL} in Section \ref{SM useful fact} shows that if $f\in\rm{PL}(k)$, then $h\in\rm{PL}(k+1)$ for all integers $k\geq 2$. First, consider the case $f\in\rm{PL}(2)$. Then, $h\in\rm{PL}(3)$; thus, also $h\in\rm{PL}(4)$. Second, consider the case $f=f_{\cal{L}}$. Then, we prove in Lemma \ref{lem:fL} that $f_{\cal{L}}\in\rm{PL}(3)$. Thus, the desired holds in this case too.
\vspace{5pt}
\subsubsection{Proof of (v):} Let $\psi:\mathbb{R}\rightarrow\mathbb{R}$ be any bounded Lipschitz function. The function $f(a,b,c) = \psi(a)$ is trivially $\rm{PL}$ of order $2$. Thus, by directly applying statement (iv) of the lemma, we find that
$$
\frac{1}{p}\sum_{i=1}^p{\psi(\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i}({\vct{g}},\vct{h}))} \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\psi(X_{\kappa,\sigma^2})\right].
$$
Since this holds for any bounded Lipschitz function, we have shown that the empirical convergence of ${\boldsymbol{\hat{\beta}}}_n$ converges weakly to the distribution of $X_{\kappa,\sigma^2}$. It remains to prove boundedness of the $2$nd moment as advertised in \eqref{eq:k_AO}. Recall from
\eqref{eq:betan} that
\begin{align}
\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\vct{h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)(\sqrt{p}\betab^\star_i).\notag
\end{align}
Using this, boundedness of $\boldsymbol{\Sigma}_{i,i}$ from Assumption \ref{ass:mu}, and the fact that $\tau_n\stackrel{{P}}{\longrightarrow}\tau_\star, \xi_n\stackrel{{P}}{\longrightarrow}\xi_\star$, there exists constant $C=C(\Sigma_{\max},\Sigma_{\min},k,\tau_\star,\xi_\star)$ such that wpa 1,
\begin{align}
\frac{1}{p}\sum_{i=1}^p|\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i}|^{2} \leq C\left(\frac{1}{p}\sum_{i=1}^p|\vct{h}_{i}|^{2}+\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\betab^\star_{i}|^{2}\right).\notag
\end{align}
But the two summands in the expression above are finite in the limit of $p\rightarrow\infty$. Specifically, (i) from Assumption \ref{ass:mu}, $\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\betab^\star_{i}|^{2}\stackrel{{P}}{\longrightarrow}\operatorname{\mathbb{E}}[B^{2}]<\infty$; (ii) $\frac{1}{p}\sum_{i=1}^p|\vct{h}_{i}|^{2}\stackrel{{P}}{\longrightarrow}\operatorname{\mathbb{E}}[H^{2}]=1$, using the facts that $\vct{h}_i\stackrel{iid}{\sim}\mathcal{N}(0,1)$ and $H\sim\mathcal{N}(0,1)$. This proves \eqref{eq:k_AO}, as desired.
\input{risks_proofs}
\section{Useful results about pseudo-Lipschitz functions}\label{SM useful fact}
For $k\geq 1$ we say a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$ and denote it by $f\in \rm{PL}(k)$ if there exists a cosntant $L>0$ such that, for all $\vct{x},\vct{y}\in\mathbb{R}^m$:
\begin{align}
|f(\vct{x})-f(\vct{y})|\leq L\left(1+\|\vct{x}\|_2^{k-1}+\|\vct{y}\|_2^{k-1}\right)\|\vct{x}-\vct{y}\|_2.\label{PL func}
\end{align}
In particular, when $f\in\rm{PL}(k)$, the following properties hold:
\begin{enumerate}
\item There exists a constant $L'$ such that for all $\vct{x}\in\mathbb{R}^n$: $|f(\vct{x})|\leq L'(1+\|\vct{x}\|_2^k).$
\item $f$ is locally Lipschitz, that is for any $M>0$, there exists a constant $L_{M,m}<\infty$ such that for all $x,y\in[-M,M]^m$,
$
|f(\vct{x})-f(\vct{y})| \leq L_{M,m}\|\vct{x}-\vct{y}\|_2.
$
Further, $L_{M,m}\leq c(1+(M\sqrt{m})^{k-1})$ for some costant $c$.
\end{enumerate}
Using the above properties, we prove the following two technical lemmas used in the proof of Theorem \ref{thm:master_W2}
\begin{lemma}\label{lem:fL}
Let $g:\mathbb{R}\rightarrow\mathbb{R}$ be a Lipschitz function. Consider the function $f:\mathbb{R}^3\rightarrow\mathbb{R}$ defined as follows:
$$
f(\vct{x}) = x_1(x_2-g(x_3))^2.
$$
Then, $f\in\rm{PL}(3).$
\end{lemma}
\begin{proof}
Let $h:\mathbb{R}^2\rightarrow\mathbb{R}$ defined as $h({\vct{u}})=({\vct{u}}_1-g({\vct{u}}_2))^2$. The function $({\vct{u}}_1,{\vct{u}}_2)\mapsto{\vct{u}}_1-g({\vct{u}}_2)$ is clearly Lipschitz. Thus, $h\in\rm{PL}(2)$, i.e.,
\begin{align}
|h({\vct{u}})-h(\vct{v})| \leq C(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2\quad\text{and}\quad |h(\vct{v})|\leq C'(1+\|\vct{v}\|_2^2).\label{eq:h_pl}
\end{align}
Therefore, letting $\vct{x}=(x_1,{\vct{u}})\in\mathbb{R}^3$ and $\vct{y}=(y_1,\vct{v})\in\mathbb{R}^3$, we have that
\begin{align}
|f(\vct{x})-f(\vct{y})| &= |x_1h({\vct{u}}) - y_1h(\vct{v})| \leq |x_1||h({\vct{u}})-h(\vct{v})| + |h(\vct{v})| |x_1-y_1|\notag\\
&\leq C|x_1|(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(|x_1|^2+(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|\vct{x}-\vct{y}\|_2.
\end{align}
In the second line, we used \eqref{eq:h_pl}. In the third line, we used $2xy\leq x^2+y^2$. In the fourth line, we used Cauchy-Schwarz inequality. $C,C'>0$ are absolute constants that may change from line to line.
This completes the proof of the lemma.
\end{proof}
\fx{
\begin{lemma}[PL with Bounded Variables]\label{lem:PLbdd
Let $f:\mathbb{R}^{d_1}\rightarrow \mathbb{R}$ be a $PL(k)$ function. $M\subset \mathbb{R}^{d_2}$ be a compact set and $g$ be a continuously differentiable function over $M$. Then $h(\vct{x},\vct{y})=f(\vct{x})g(\vct{y})$ is $PL(k+1)$ over $\mathbb{R}^{d_1}\times M$.
\end{lemma}
\begin{proof} First observe that since $g$ has continuous derivatives and is continuous over a compact set. Thus $g$ and its gradient is bounded and $g$ is Lipschitz over $M$. Let $B=\sup_{\vct{x}\in M}\max |g(x)|,\tn{\nabla g(\vct{x})}$. To proceed, given pairs $(\vct{x},\vct{y})$ and $(\vct{x}',\vct{y}')$ over $\mathbb{R}\times M$, we have that
\begin{align}
|h(\vct{x},\vct{y})-h(\vct{x}',y')|&\leq |h(\vct{x},\vct{y})-h(\vct{x}',\vct{y})|+ |h(\vct{x}',\vct{y})-h(\vct{x}',\vct{y}')|\\
&\leq |f(\vct{x})-f(\vct{x}')||g(\vct{y})|+ |f(\vct{x}')||g(\vct{y})-g(\vct{y}')|\\
&\leq B|f(\vct{x})-f(\vct{x}')|+ B\tn{\vct{y}-\vct{y}'}|f(\vct{x}')|\\
&\leq B(1+\tn{\vct{x}'}^{k-1}+\tn{\vct{x}}^{k-1})\tn{\vct{x}-\vct{x}'}+ B\tn{\vct{y}-\vct{y}'}(1+\tn{\vct{x}'}^k)\\
&\lesssim (1+\tn{{\vct{z}}}^k+\tn{{\vct{z}}'}^{k})\tn{{\vct{z}}-{\vct{z}}'},
\end{align}
where ${\vct{z}}=[\vct{x}~\vct{y}]$. This shows the desired $\text{PL}(k+1)$ guarantee.
\end{proof}
The following lemma is in similar spirit to Lemma \ref{lem:PLbdd} and essentially follows from similar lines of arguments (i.e.~using Lipschitzness induced by boundedness).
}
\begin{lemma}\label{lem:hPL}
Let functions $f,g:\mathbb{R}^3\rightarrow\mathbb{R}$ such that $f\in\rm{PL}(k)$ and
$$
g(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi_* x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1.
$$
Here, $\xi_*,\tau_*,\kappa$ are positive constants.
Further define
\begin{align}
h(x,y,z) := f\left(g(y,z,x),y,z\right),
\end{align}
and assume that $y$ take values on a fixed bounded compact set $\mathcal{M}$.
Then, it holds that $h\in\rm{PL}(k+1)$.
\end{lemma}
\begin{proof}
Since $f$ is $\rm{PL}(k)$, for some $L>0$, \eqref{PL func} holds. Fix $x,x'\in\mathbb{R}$, $\vct{a}=[y,z]\in\mathbb{R}^{2}$ and $\vct{a}'=[y',z']\in\mathbb{R}^{2}$. Let $\vct{b}=[{\vct{g}},\vct{a}]=[{\vct{g}},y,z]\in\mathbb{R}^3$ where ${\vct{g}}=g(y,z,x)\in\mathbb{R}$ and define accordingly $\vct{b}'$ and ${\vct{g}}'$. We have that
\begin{align}
|h([x,\vct{a}])-h([x,\vct{a}'])|&= |f(\vct{b})-f(\vct{b}')|\notag\\
&\leq L\left(1+\|\vct{b}\|_2^{k-1}+\|\vct{b}'\|_2^{k-1}\right)\|\vct{b}-\vct{b}'\|_2\notag\\
&\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|{\vct{g}}|^{k-1}+|{\vct{g}}'|^{k-1}\right)(\|\vct{a}-\vct{a}'\|_2+|{\vct{g}}-{\vct{g}}'|),\label{eq:hlip}
\end{align}
for some constant $C>0$. In the last inequality we have repeatedly used the inequality
$
\left(\sum_{i=1}^m \|\vct{v}_i\|_2^2\right)^{\frac{d}{2}} \leq C(m)\cdot\sum_{i=1}^m\|\vct{v}_i\|_2^{d}.
$
Next, we need to bound the ${\vct{g}}$ term in terms of $\vct{a}$. This is accomplished as follows
\begin{align}
|{\vct{g}}|^{k-1} &= \left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x + (1-(1+\xi_*y)^{-1})z\right|^{k-1}\notag\\
&\leq C (|z|+|x|)^{k-1} \notag\\
&\leq C \left(|x|^{k-1}+|z|^{k-1}\right) \leq C (|x|^{k-1} + \|\vct{a}\|_2^{k-1}).\label{eq:gb}
\end{align}
Here, the value of the constant $C>0$ may change from line to line.
Secondly and similarly, we have the following perturbation bound on the ${\vct{g}}-{\vct{g}}'$. Recall that $(x,y)$ and $(x',y')$ are bounded. We will prove the following sequence of inequalities
\begin{align}
|{{\vct{g}}-{\vct{g}}'}|&= |g(y,z,x)-g(y',z',x')| \notag\\
&\leq \left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x - \frac{(y')^{-1/2}}{1+(\xi y')^{-1}}\sqrt{\kappa}\tau_* x'\right|
+ \left|(1-(1+\xi_*y)^{-1})z-(1-(1+\xi_*y')^{-1})z'\right|\notag\\
&\leq
\left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x - \frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x'\right| +\left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x'- \frac{(y')^{-1/2}}{1+(\xi y')^{-1}}\sqrt{\kappa}\tau_* x'\right|\notag\\
&\qquad+ \left|(1-(1+\xi_*y)^{-1})z-(1-(1+\xi_*y)^{-1})z'\right| + \left|(1-(1+\xi_*y)^{-1})z'-(1-(1+\xi_*y')^{-1})z'\right| \notag\\
&\leq C_1|x-x'| + C_2|x'||y-y'| + C_3|z-z'| + C_4|z'||y-y'|\notag\\
&\leq C(1+|x'| + |z'|)(|x-x'| + |z-z'| + |y-y'|)\\
&\leq C\sqrt{3}(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2
.\label{eq:gd}
\end{align}
In the fourth inequality above, we used the fact the assumption that $|y|$ is bounded. In the last line, we used Cauchy-Scwhartz.
Substituting \eqref{eq:gb} and \eqref{eq:gd} in \eqref{eq:hlip} gives:
\begin{align}
|h(x,y,z) - h(x',y',z')| &\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|x|^{k-1}+|x'|^{k-1}\right)(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2\notag \\
&\leq C\left(1+\|[\vct{a},x]\|_2^{k}+\|[\vct{a}',x']\|_2^{k}\right)\|[\vct{a},x]-[\vct{a}',x']\|_2.
\end{align}
Thus, $h\in\rm{PL}(k+1)$, as desired.
\end{proof}
\section{Proofs for overparameterized least-squares}\label{sec proof thm 1}
In this section, we assume the linear Gaussian problem (LGP) of Definition \ref{def LGP}, the overparameterized regime $k=p>n$ and the min-norm model $\hat\boldsymbol{\beta}$ of \eqref{eq:min_norm}. We prove Theorem \ref{thm:master_W2} that derives the asymptotic DC of $\hat\boldsymbol{\beta}$ and we show how this leads to sharp formulae for the risk of the Magnitude- and Hessian-pruned models.
\subsection{Notation and Assymptions}
For the reader's convenience, we recall some necessary notation and assumptions from Section \ref{sec main}. We say that a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$, denoted $f\in\rm{PL}(k)$, if there is a constant $L>0$ such that for all $\vct{x},\vct{y}\in\mathbb{R}^m$,
$
|f(\vct{x}) - f(\vct{y})|\leq L(1+\tn{\vct{x}}^{k-1}+\tn{\vct{y}}^{k-1})\|\vct{x}-\vct{y}\|_2
$
(See also Section \ref{SM useful fact}).
We say that a sequence of probability distributions $\nu_p$ on $\mathbb{R}^m$ {converges in $W_k$} to
$\nu$, written $\nu_p\stackrel{W_k}{\Longrightarrow} \nu$, if $W_k(\nu_p,\nu) \rightarrow 0$ as $p \rightarrow \infty$. An equivalent definition is that, for any $f\in\rm{PL}(k)$, $\lim_{p\rightarrow}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$ (e.g., \cite{montanari2017estimation}).
Finally, we use $C,C',c,c'$ to denote absolute constants (not depending on $n,p$) whose value might change from line to line.
We focus on a double asymptotic regime where:
$$n,p,s\rightarrow\infty \text{ at fixed overparameterization ratio } \kappa:=p/n>0 \text{ and sparsity level } \alpha:=s/p\in(0,1).$$
For a sequence of random variables $\mathcal{X}_{p}$ that converge in probability to some constant $c$ in the limit of Assumption \ref{ass:linear} below, we write $\mathcal{X}_{p}\stackrel{{P}}{\longrightarrow} c$. For a sequence of event $\mathcal{E}_p$ for which $\lim_{p\rightarrow}\mathbb{P}(\mathcal{E}_p) = 1$, we say that $\mathcal{E}_p$ occurs \emph{with probability approaching 1}. For this, we will often use the shorthand ``wpa. 1".
Next, we recall the set of assumption under which our analysis applies:
\textbf{(A1)[Diagonal covariance].}~~The covariance matrix $\boldsymbol{\Sigma}$ is diagonal.
\textbf{(A2)[Boundedness and empirical distribution].}~~There exist constants $\Sigma_{\min},\Sigma_{\max}\in(0,\infty)$ such that:
$
\Sigma_{\min}\leq{\boldsymbol{{\Sigma}}}_{i,i}\leq \Sigma_{\max}.
$
for all $i\in[p].$ There exists absolute constant $C>0$ such that $\sqrt{p}|\betab^\star_i|<C$, for all $i\in[p]$. Furthermore, the joint empirical distribution of $\{(\bSi_{i,i},\sqrt{p}\betab^\star_i)\}_{i\in[p]}$ converges in {Wasserstein-\cts{(k+1)}} distance to a probability distribution $\mu$ on $\mathbb{R}_{>0}\times\mathbb{R}$ {for some \cts{$k+1\geq 4$}}. That is
$
\frac{1}{p}\sum_{i\in[p]}\delta_{(\bSi_{i,i},\sqrt{p}\betab^\star_i)} \stackrel{W_{\cts{k+1}}}{\Longrightarrow} \mu.
$
\subsection{Proof of Theorem \ref{thm:master_W2}}
\vspace{5pt}
Let ${\mtx{X}}\in\mathbb{R}^{n\times p}$ have zero-mean and normally distributed rows with a diagonal covariance matrix ${\boldsymbol{{\Sigma}}}=\operatorname{\mathbb{E}}[\vct{x}\x^T]$. Given a ground-truth vector $\betab^\star$ and labels $\vct{y}={\mtx{X}}\betab^\star+\sigma {\vct{z}},~{\vct{z}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, we consider the least-squares problem subject to the minimum Euclidian norm constraint (as $\kappa=p/n>1$) given by
\begin{align}\label{eq:PO_beta}
\min_{{\boldsymbol{\beta}}}\frac{1}{2}\tn{{\boldsymbol{\beta}}}^2\quad\text{subject to}\quad \vct{y}={\mtx{X}}{\boldsymbol{\beta}}.
\end{align}
It is more convenient to work with the following change of variable: $\vct{w}:=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\beta}}-\betab^\star)$. With this, the optimization problem in \eqref{eq:min_norm} can be rewritten as
\begin{align}\label{eq:PO}
\Phi({\mtx{X}})=\min_{\vct{w}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
where we write ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$. First, using standard arguments, we show that the solution of \eqref{eq:PO} is bounded. Hence, we can constraint the optimization in a sufficiently large compact set without loss of generality.
\begin{lemma}[Boundedness of solution]\label{lem:bd_PO}
Let $\hat\vct{w}_n:=\hat\vct{w}_n({\mtx{X}},{\vct{z}})$ be the minimizer in \eqref{eq:PO}. Then, with probability approaching 1, it holds that $\hat\vct{w}_n\in\mathcal{B}$, where
$$\mathcal{B}:=\left\{\vct{w}\,|\,\|\vct{w}\|_2\leq B_{+} \right\},\qquad B_+:=4\sqrt{\Sigma_{\max}}\frac{2(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + \sigma}{\sqrt{\Sigma_{\min}}(\sqrt{\kappa}-1)} + \sqrt{\Sigma_{\max}}\sqrt{\operatorname{\mathbb{E}}\left[B^2\right]}.
$$
\end{lemma}
\begin{proof}
First, we show that the min-norm solution $\hat\boldsymbol{\beta}={\mtx{X}}^T({\mtx{X}}\X^T)^{-1}\vct{y}$ of \eqref{eq:PO_beta} is bounded. Here, we used the fact that $\kappa>1$, thus ${\mtx{X}}\X^T$ is invertible wpa 1. We
have,
\begin{align}
\tn{\hat\boldsymbol{\beta}_n}^2 = \vct{y}^T({\mtx{X}}\X^T)^{-1}\vct{y} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{X}}\X^T)} = \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\boldsymbol{\Sigma}{\mtx{\bar{X}}}^T)} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\Xb^T)\,\Sigma_{\min}} = \frac{\tn{\vct{y}}^2}{\sigma_{\min}^2({\mtx{\bar{X}}})\,\Sigma_{\min}}. \label{eq:Ubb}
\end{align}
But, wpa 1,
$
\sigma_{\min}({\mtx{\bar{X}}})/\sqrt{n} \geq \frac{1}{2}\left(\sqrt{\kappa}-1\right).
$
Furthermore,
$
\|\vct{y}\|_2 \leq \|{\mtx{\bar{X}}}\boldsymbol{\Sigma}^{1/2}\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2 \leq \sigma_{\max}({\mtx{\bar{X}}})\sqrt{\Sigma_{\max}} \|\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2.
$
Hence, wpa 1,
$$
\|\vct{y}\|_2/\sqrt{n} \leq 4(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma,
$$
where we used the facts that wpa 1: $\|z\|_2/\sqrt{n}\stackrel{{P}}{\longrightarrow} 1$, $\sigma_{\max}({\mtx{\bar{X}}})<2(\sqrt{\kappa}+1)$ and by Assumption \ref{ass:mu}:
$$
\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^{p}(\sqrt{p}\betab^\star_i)^2 \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[B^2\right].
$$
Put together in \eqref{eq:Ubb}, shows that
\begin{align}
\tn{\hat\boldsymbol{\beta}_n} < \frac{4(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma}{\sqrt{\Sigma_{\min}}(\sqrt{\kappa}-1)/2} =: \tilde{B}_+.\label{eq:bd_beta}
\end{align}
Recalling that $\hat\vct{w}_n= \sqrt{\boldsymbol{\Sigma}}\boldsymbol{\beta} - \sqrt{\boldsymbol{\Sigma}}\betab^\star$, we conclude, as desired, that wpa 1,
$
\tn{\hat\vct{w}_n} \leq \sqrt{\Sigma_{\max}}\tilde{B}_+ + \sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}[B^2]} =: B_+.
$
\end{proof}
Lemma \ref{lem:bd_PO} implies that nothing changes in \eqref{eq:PO} if we further constrain $\vct{w}\in\mathcal{B}$ in \eqref{eq:PO}. Henceforth, with some abuse of notation, we let
\begin{align}\label{eq:PO_bd}
\Phi({\mtx{X}}):=\min_{\vct{w}\in\mathcal{B}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
Next, in order to analyze the primary optimization (PO) problem in \eqref{eq:PO_bd} in apply the CGMT \cite{thrampoulidis2015lasso}. Specifically, we use the constrained formulation of the CGMT, Theorem \ref{thm closed}. Specifically, the auxiliary problem (AO) corresponding to \eqref{eq:PO_bd} takes the following form with ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$
\begin{align}
\phi({\vct{g}},\vct{h}) = \min_{\vct{w}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_con}
\end{align
We will prove the following result about the AO problem.
\begin{lemma}[Properties of the AO -- Overparameterized regime]\label{lem:AO}
Let $\phi_n=\phi({\vct{g}},\vct{h})$ be the optimal cost of the minimization in \eqref{eq:AO_con}. Define $\bar\phi$ as the optimal cost of the following deterministic min-max problem
\begin{align}\label{eq:AO_det}
\bar\phi:=\max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau):=\frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{N^2}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).
\end{align}
The following statements are true.
\noindent{(i).}~The AO minimization in \eqref{eq:AO_con} is $\frac{1}{\Sigma_{\max}}$-strongly convex and has a unique minimizer $\vct{w}_n:=\vct{w}_n({\vct{g}},\vct{h})$.
\noindent{(ii).}~In the limit of $n,p\rightarrow\infty, p/n=\kappa$, it holds that $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$, i.e., for any $\varepsilon>0$:
$$
\lim_{n\rightarrow\infty}\P\left(|\phi({\vct{g}},\vct{h})-\bar\phi|>\varepsilon\right) = 0.
$$
\noindent{(iii).} The max-min optimization in \eqref{eq:AO_det} has a unique saddle point $(u_*,\tau*)$ satisfying the following:
$$
u_*/\tau_* = \xi\quad\text{and}\quad\tau_* = \gamma,
$$
where $\xi, \gamma$ are defined in Definition \ref{def:Xi}.
\noindent{(iv).}~Let $f:\mathbb{R}\rightarrow\mathbb{R}$ be a $\rm{PL}({k})$ function. Let $\boldsymbol{\beta}_n=\boldsymbol{\Sigma}^{-1/2}\vct{w}_n + \betab^\star$. Then,
$$
\frac{1}{p}\sum_{i=1}^{p}f\left(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(B,\Lambda,H)\sim\mu\otimes \mathcal{N}(0,1)}\left[f\left(X_{\kappa,\sigma^2}(B,\Lambda,H),B,\Lambda\right) \right].
$$
\end{lemma}
We prove Lemma \ref{lem:AO} in Section \ref{sec:proofAO}. Here, we show how this leads to the proof of Theorem \ref{thm:master_W2} when combined with the CGMT.
For convenience, define
$$F_n(\hat\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}\hat\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Fix any $\varepsilon>0$. It suffices to prove that the solution $\hat\boldsymbol{\beta}_n$ of the PO in \eqref{eq:PO_beta} satisfies $\hat\boldsymbol{\beta}_n\not\in\mathcal{S}$ wpa. 1, where
\begin{align}\label{eq:S_set}
\mathcal{S} = \mathcal{S}(\betab^\star,\boldsymbol{\Sigma}) =\{\boldsymbol{\beta}{~\big |~} |F_n(\hat\boldsymbol{\beta}_n,\betab^\star,\boldsymbol{\Sigma})-\alpha_*|\geq 2\varepsilon\}.
\end{align}
In particular, define the ``perturbed" PO and AO problems (compare to \eqref{eq:PO} and \eqref{eq:AO_con}) as:
\begin{align}\label{eq:PO_S}
\Phi_S({\mtx{X}})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
and
\begin{align}
\phi_S({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_S}
\end{align
where recall that ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$, ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$ and we have used the change of variables $\vct{w}:=\sqrt{\boldsymbol{\Sigma}}(\boldsymbol{\beta}-\betab^\star)$ for convenience.
Using \cite[Theorem 6.1(iii)]{thrampoulidis2018precise} it suffices to find costants $\bar\phi, \bar\phi_S$ and $\eta>0$ such that the following hold:
\begin{enumerate}
\item $\bar\phi_S \geq \bar\phi + 3\eta$,
\item $\phi({\vct{g}},\vct{h}) \leq \bar\phi + \eta$, with probability approaching 1,
\item $\phi_S({\vct{g}},\vct{h}) \geq \bar\phi_S - \eta$, with probability approaching 1.
\end{enumerate}
In what follows, we explicitly find $\bar\phi, \bar\phi_S,\eta$ such that the three conditions above hold.
\vspace{5pt}
\noindent\underline{Satisfying Condition 2}: Recall the deterministic min-max optimization in \eqref{eq:AO_det}. Choose $\bar\phi=\mathcal{D}(u_*,\tau_*)$ be the optimal cost of this optimization. From Lemma \ref{lem:AO}(ii), $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$. Thus, for any $\eta>0$, with probability approaching 1:
\begin{align}\label{eq:phi_lim}
\bar\phi + \eta \geq \phi({\vct{g}},\vct{h}) \geq \bar\phi - \eta.
\end{align}
Clearly then, Condition 2 above holds for any $\eta>0$.
\vspace{5pt}
\noindent\underline{Satisfying Condition 3}: Next, we will show that the third condition holds for appropriate $\bar\phi$.
Let $\vct{w}_n=\vct{w}_n({\vct{g}},\vct{h})$ be the unique minimizer of \eqref{eq:AO_con} as per Lemma \ref{lem:AO}(i), i.e., $\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}_n+\boldsymbol{\beta}}^2 = \phi({\vct{g}},\vct{h})$. Again from Lemma \ref{lem:AO}, the minimization in \eqref{eq:AO_con} is $1/\Sigma_{\max}$-strongly convex in $\vct{w}$. Here, $\Sigma_{\max}$ is the upper bound on the eigenvalues of $\boldsymbol{\Sigma}$ as per Assumption \ref{ass:inv}. Thus, for any $\tilde\varepsilon>0$ and any feasible $\vct{w}$ the following holds (deterministically):
\begin{align}\label{eq:sc}
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\boldsymbol{\beta}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2C},~\text{provided that}~ \|\vct{w}-\vct{w}_n({\vct{g}},\vct{h})\|_2 \geq \tilde\epsilon.
\end{align}
Now, we argue that $\boldsymbol{\beta}\in\mathcal{S}$ implies that $\|\vct{w}-\vct{w}_n({\vct{g}},\vct{h})\|_2\geq \tilde\varepsilon$ wpa 1, for
appropriate value of $\tilde\varepsilon$ and $\vct{w}=\sqrt{\boldsymbol{\Sigma}}(\boldsymbol{\beta}-\betab^\star)$. Consider any $\boldsymbol{\beta}\in\mathcal{S}$. First, by definition in \eqref{eq:S_set},
$$
|F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \geq 2\varepsilon.
$$
Second, by Lemma \ref{lem:AO}(iv), with probability approaching 1,
$$
|F(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \leq \epsilon.
$$
Third, we will show that wpa 1, there exists universal constant $C>0$ such that
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| \leq C {\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}\label{eq:dev2show}.
\end{align}
Before proving \eqref{eq:dev2show}, let us argue how combining the above three displays shows the desired. Indeed, wpa. 1,
\begin{align*}
2\varepsilon &\leq |F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \leq |F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| + |F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \\
&\leq \epsilon + C \,\|\boldsymbol{\beta}-\boldsymbol{\beta}_n\|_2. \\
&\qquad\Longrightarrow \|\boldsymbol{\beta}-\boldsymbol{\beta}_n\|_2 \geq {\varepsilon}/{C}=:\hat\varepsilon\\
&\qquad\Longrightarrow \|\vct{w}-\vct{w}_n\|_2 \geq \hat\varepsilon\sqrt{\Sigma_{\min}}=:\tilde\varepsilon.
\end{align*}
In the last line above, we recalled that $\boldsymbol{\beta}=\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star$ and $\boldsymbol{\Sigma}_{i,i}\geq\Sigma_{\min},~i\in[p]$ by Assumption \ref{ass:inv}.
From this and \eqref{eq:sc}, we find that wpa 1,
$
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\boldsymbol{\beta}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2C},~\text{for all}~ \vct{w}\in\mathcal{S}.
$
Thus,
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}.\notag
\end{align}
When combined with \eqref{eq:phi_lim}, this shows that
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta.
\end{align}
Thus, choosing $\bar\phi_S = \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}$ proves the Condition 3 above.
To complete the proof, let us now show \eqref{eq:dev2show}. Henceforth, $C$ is used to denote a universal constant whose value can change from line to line.
To simplify notation, let $\vct{a}_i=(\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}),\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\boldsymbol{\beta}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$, for $i\in[p]$. Then, using $f\in\rm{PL}(k)$,
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| &\leq \frac{L}{p}\sum_{i=1}^p (1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\})\|\vct{a}_i-\vct{b}_i\|_2\notag \\
&= \frac{L}{p}\sum_{i=1}^p \left(1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\}\right)\cdot|\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}) - \sqrt{p}\boldsymbol{\beta}_i|\notag\\
&\leq L\left(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} \right)^{1/2}{\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}\label{eq:LipC}\,.
\end{align}
The last inequality above follows by applying Cauchy-Schwartz.
To reach \eqref{eq:dev2show} from the above, it suffices to show that
\begin{align}
\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} \leq C,\label{eq:leqC}
\end{align}
for some universal constant $C$ (that may depend on $k$, but not on $n,p$).
To prove \eqref{eq:leqC}, note that
using $a^{\ell_1}b^{\ell_2}c^{\ell_3}\leq a^{\ell}+b^{\ell}+c^\ell,~\ell=\ell_1+\ell_2+\ell_3$, for some constant $C=C(k)>0$ it holds:
\begin{align}
\frac{1}{p}\sum_{i=1}^p \|\vct{b}_i\|_2^{2k-2} {\leq} C\left(\frac{1}{p}\sum_{i=1}^p |\sqrt{p}\boldsymbol{\beta}_{i}|^{2k-2} + \frac{1}{p}\sum_{i=1}^p |\sqrt{p}\betab^\star_{i}|^{2k-2} + \frac{1}{p}\sum_{i=1}^p |\boldsymbol{\Sigma}_{i,i}|^{2k-2} \right)\label{eq:bdmom}\,.
\end{align}
\cts{The terms $\frac{1}{p}\sum_{i=1}^p |\sqrt{p}\betab^\star_{i}|^{2k-2} $ and $ \frac{1}{p}\sum_{i=1}^p |\boldsymbol{\Sigma}_{i,i}|^{2k-2}$ are bounded by Assumption (A2).}\ct{Need to add assumption that $ \frac{1}{p}\sum_{i=1}^p |\boldsymbol{\Sigma}_{i,i}|^{2k-2}<\infty$ and $\frac{1}{p}\sum_{i=1}^p (\sqrt{p}|\betab^\star_{i}|^{2k-2} $. OK by just assuming that $B$ is Definition \ref{def:Xi} is bounded.}. {\color{red}Furthermore we need to argue that ,
\begin{align}
\frac{1}{p}\sum_{i=1}^p |\sqrt{p}\boldsymbol{\beta}_{i}|^{2k-2} <\infty
\end{align}
}
These together with \eqref{eq:bdmom} show that $\frac{1}{p}\sum_{i=1}^p \|\vct{b}_i\|_2^{2k-2}<C$ wpa 1. Similarly, we can argue for $\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2}$, completing the proof of \eqref{eq:leqC}.
\vspace{5pt}
\noindent\underline{Satisfying Condition 1:} To prove Condition 1, we simply pick $\eta$ to satisfy the following
\begin{align}
\bar\phi_S > \bar\phi + 3 \eta ~\Leftarrow~ \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta \geq \bar\phi + 3 \eta ~\Leftarrow~ \eta \leq \frac{\tilde\epsilon^2}{8\Sigma_{\max}}.
\end{align}
This completes the proof of Theorem \ref{thm:master_W2}.
\subsection{Proof of Lemma \ref{lem:AO}}\label{sec:proofAO}
~~~~
~~~~
\vspace{3pt}
\noindent\underline{Proof of (i):}~Strong convexity of the objective function in \eqref{eq:PO} is easily verified by the second derivative test. Note here that we use Assumption \ref{ass:inv} that $\bSi_{i,i}\leq\Sigma_{\max},~i\in[p].$ Uniqueness of the solution follows directly from strong convexity. \ct{Strictly speaking we might need to also argue existence, i.e., feasibility of the AO. An indirect way is to show feasibility using the CGMT, but it seems unnecessarily complicated?}
\vspace{5pt}
\noindent\underline{Proof of (ii):}~Using Lagrangian formulation, the solution $\vct{w}_n$ to \eqref{eq:AO_con} is the same as the solution to the following:
\begin{align}
\left(\vct{w}_n,u_n\right) :=\arg\min_{\vct{w}\in\mathcal{B}}\max_{u\geq 0} ~\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + u \left( \sqrt{\tn{\vct{w}}^2+\sigma^2} \tn{\bar{\g}} - \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} + \frac{\sigma h}{\sqrt{n}} \right)\label{eq:AO_2}
\end{align}
where we have: (i) set $\bar{\g} := {\vct{g}}/\sqrt{n}$ and $\bar{\h}:= \vct{h}/\sqrt{p}$; (ii) recalled that $p/n=\kappa$; and, (iii) used $\left(\vct{w}_n,u_n\right)$ to denote the optimal solutions in \eqref{eq:AO_2}. The subscript $n$ emphasizes the dependence of $\left(\vct{w}_n,u_n\right)$ on the problem dimensions. Also note that (even though not explicit in the notation) $\left(\vct{w}_n,u_n\right)$ are random variables depending on the realizations of $\bar{\g},\bar{\h}$ and $h$.
Notice that the objective function above is convex in $\vct{w}$ and linear (thus, concave) in $u$. Thus, strong duality holds and we can flip the order of min-max. Moreover, in order to make the objective easy to optimize with respect to $\vct{w}$, we use the following variational expression for the square-root term $\sqrt{\tn{\vct{w}}^2+\sigma^2}$:
$$
\tn{\bar{\g}}\sqrt{\tn{\vct{w}}^2+\sigma^2} = \tn{\bar{\g}}\cdot\min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau}{2} + \frac{\tn{\vct{w}}^2+\sigma^2}{2\tau} \right\} = \min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau\tn{\bar{\g}}^2}{2} + \frac{\sigma^2}{2\tau} + \frac{\tn{\vct{w}}^2}{2\tau} \right\},
$$
where $B$ is defined in Lemma \ref{lem:bd_PO}. For convenience define the constraint set for the variable $\tau$ as $\mathcal{T}':=[\sigma,\sqrt{\sigma^2+B_+^2}]$. For reasons to be made clear later in the proof (see proof of statement (iii)), we consider the (possibly larger) set:
\[
\mathcal{T}:=[\sigma,\max\{\sqrt{\sigma^2+B_+^2},2\tau_*\}]\,
\]
where $\tau_*$ is as in the statement of the lemma.
The above lead to the following equivalent formulation of \eqref{eq:AO_2}:
\begin{align}
\left(\vct{w}_n,u_n,\tau_n\right) = \max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} + \min_{\vct{w}} \left\{ \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + \frac{u}{2\tau}\tn{\vct{w}}^2 - u \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} \right\} .\label{eq:AO_3}
\end{align}
\ct{To be fully rigorous, need to show here that the unconstrained min over $\vct{w}$ is the same as the constrained $\vct{w}\in\mathcal{B}$.}
The minimization over $\vct{w}$ is easy as it involves a strongly convex quadratic function. The optimal $\vct{w}':=\vct{w}'(\tau,u)$ (for fixed $(\tau,u)$) is given by
\begin{align}\label{eq:w'}
\vct{w}':=\vct{w}'(\tau,u) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right),
\end{align}
and \eqref{eq:AO_3} simplifies to
\begin{align}
\left(u_n,\tau_n\right)=\max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} - \frac{1}{2} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)\,=:\mathcal{R}(u,\tau) .\label{eq:AO_4}
\end{align}
It can be checked by direct differentiation and the second-derivative test that the objective function in \eqref{eq:AO_4} is strictly convex in $\tau$ and strictly concave in $u$ in the domain $\{(u,\tau)\in\mathbb{R}_+\times\mathbb{R}_+\}$. Thus, the saddle point $(u_n,\tau_n)$ is unique. Specifically, this implies that the optimal $\vct{w}_n$ in \eqref{eq:AO_3} is given by (cf. \eqref{eq:w'})
\begin{align}\label{eq:w_n}
\vct{w}_n=\vct{w}'(\tau_n,u_n) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).
\end{align}
In what follows, we characterize the high-dimensional limit of the optimal pair $(u_n,\tau_n)$ in the limit $n,p\rightarrow\infty,~p/n\rightarrow\kappa$.
We start by analyzing the (point-wise) convergence of $\mathcal{R}(u,\tau)$.
For the first three summands in \eqref{eq:AO_4}, we easily find that
$$
\left\{\frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}}\right\}~ \stackrel{{P}}{\longrightarrow}~\left\{
\frac{u\tau}{2} + \frac{u\sigma^2}{2\tau} \right\}.
$$
Next, we study the fourth summand. First, note that
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}(u\sqrt{\kappa}\bar{\h}) &= u^2\kappa\,\frac{1}{p}\vct{h}^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\vct{h} \notag\\
&= u^2\kappa\,\frac{1}{p}\sum_{i=1}^{p}\frac{\vct{h}_i^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}} \notag\\
&\stackrel{{P}}{\longrightarrow} u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right].
\end{align}
In the last line, $\Lambda$ is a random variable as in Definition \ref{def:Xi}. \cts{Also, we used Assumption (A2) together with the facts that $\vct{h}$ is independent of $\boldsymbol{\Sigma}$ and that the function $(x_1,x_2)\mapsto f(x_1,x_2)=x_1^2(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(2)$ assuming $x_2$ is bounded.} \ct{Question: Is it immediate that the empirical distribution of $(\beta,\boldsymbol{\Sigma},\vct{h})$ converges in $W_k$ to $\mu\otimes\mathcal{N}(0,1)$ given that $(\beta,\boldsymbol{\Sigma})$ converges to $\mu$ and $\vct{h}$ is independent???}
Second, we find that
\begin{align}
(\betab^\star)^T\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star &= \frac{1}{p}(\sqrt{p}\betab^\star)^T\left({\mtx{I}}+\frac{u}{\tau}\boldsymbol{\Sigma}\right)^{-1}(\sqrt{p}\betab^\star) \notag\\
&= \frac{1}{p}\sum_{i=1}^{p}\frac{\left(\sqrt{p}\betab^\star_i/\sqrt{\bSi_{i,i}}\right)^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}}\notag\\
&\stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right].
\end{align}
Here, $\Lambda,B$ are random variables as in Definition \ref{def:Xi} and \cts{we also used Assumption (A2) together with the fact that the function $(x_1,x_2)\mapsto f(x_1,x_2)=x_1^2x_2^{-1}(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(2)$ assuming $x_2$ is bounded.}
Third, \cts{by independence of $(\betab^\star, \boldsymbol{\Sigma})$ from $\vct{h}$}
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star = u\sqrt{\kappa} \cdot \frac{1}{p}\sum_{i=1}
^p \frac{\vct{h}_i\boldsymbol{\Sigma}_{i,i}^{-1/2}(\sqrt{p}\betab^\star_i)}{\boldsymbol{\Sigma}_{i,i}^{-1}+\frac{u}{\tau}{\mtx{I}}} ~\stackrel{{P}}{\longrightarrow}~ 0.
\end{align}
Putting these together, the objective $\mathcal{R}(u,\tau)$ in \eqref{eq:AO_4} converges point-wise in $u,\tau$ to
\begin{align}
\mathcal{R}(u,\tau)\stackrel{{P}}{\longrightarrow}\mathcal{D}(u,\tau) := \frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).\label{eq:conv_pt}
\end{align}
Note that $\mathcal{R}(u,\tau)$ (and thus, $\mathcal{D}(u,\tau)$) is convex in $\tau$ and concave in $u$. Thus, the convergence in \eqref{eq:conv_pt} is in fact uniform (e.g., \cite{AG1982}) and we can conclude that
\begin{align}\label{eq:Dc0}
\phi({\vct{g}},\vct{h}) \stackrel{{P}}{\longrightarrow} \max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
and
\begin{align}\label{eq:Dc}
(u_n,\tau_n) \stackrel{{P}}{\longrightarrow} (u_*,\tau_*):=\arg\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
In the proof of statement (iii) below, we show that the saddle point of \eqref{eq:Dc0} is $(u_*,\tau_*)$. In particular, $\tau_*$ is strictly in the interior of $\mathcal{T}$, which combined with convexity implies that
$$
\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau) = \max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau) =: \bar\phi.
$$
This, together with the first display above proves the second statement of the lemma.
\vspace{5pt}
\noindent\underline{Proof of (iii):}
Next, we compute the saddle point $(u_*,\tau_*)$ by studying the first-order optimality conditions of the srtictly concave-convex $\mathcal{D}(u,\tau)$. Specifically, we consider unconstrained minimization over $\tau$ and we will show that the minimum is achieved in the strict interior of $\mathcal{T}$. Direct differentiation gives
\begin{subequations}
\begin{align}
{\tau} + \frac{\sigma^2}{\tau} - 2u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] + \frac{u^2}{\tau}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] + \frac{1}{\tau}\operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0, \label{eq:fo1}\\
{u} - \frac{u\sigma^2}{\tau^2} - \frac{u^3}{\tau^2}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} + \frac{u}{\tau}\right)^2}\right] - \frac{u}{\tau^2} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0,\label{eq:fo2}
\end{align}
\end{subequations}
Multiplying \eqref{eq:fo2} with $\frac{\tau}{u}$ and adding to \eqref{eq:fo1} results in the following equation
\begin{align}
\tau = u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] ~\Leftrightarrow ~
\operatorname{\mathbb{E}}\left[ \frac{1}{(\frac{u}{\tau}\Lambda)^{-1}+1} \right] = \frac{1}{\kappa} \label{eq:tauu}\,.
\end{align}
Thus, we have found that the ratio $\frac{u_*}{\tau_*}$ is the unique solution to the equation in \eqref{eq:tauu}. Note that this coincides with the Equation \eqref{eq:ksi} that defines the parameter $\xi$ in Definition \ref{def:Xi}. The fact that \eqref{eq:tauu} has a unique solution for all $\kappa>1$ can be easily seen as $F(x)=\operatorname{\mathbb{E}}\left[ \frac{1}{(x\Lambda)^{-1}+1} \right], x\in\mathbb{R}_+$ has range $(0,1)$ and is strictly increasing (by differentiation).
Thus, we call $\xi=\frac{u_*}{\tau_*}$. Moreover, multiplying \label{eq:fo2} with $u$ leads to the following equation for $\tau_*$:
\begin{align}
u_*^2 = \sigma^2\xi^2 + u_*^2 \xi^2 \kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right] + \xi^2 \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right] ~\Rightarrow~ \tau_*^2 = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\xi^2\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right]} = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left((\xi\Lambda)^{-1} +1\right)^2}\right]}.
\end{align}
{Again, note that this coincides with Equation \eqref{eq:gamma} that determines the parameter $\gamma$ in Definition \ref{def:Xi}, i.e., $\tau_*^2 = \gamma.$ }
\vspace{3pt}\noindent\underline{Proof of (iv):}
For convenience, define
$$F_n(\hat\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}\hat\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Recall from \eqref{eq:w_n} the explicit expression for $\vct{w}_n$, repeated here for convenience.
\begin{align}
\vct{w}_n = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).\notag
\end{align}
Also, recall that $\boldsymbol{\beta}_n = \boldsymbol{\Sigma}^{-1/2}\vct{w}_n+\betab^\star$. Thus, (and using the fact that $\bar{\h}$ is distributed as $-\bar{\h}$),
\begin{align}
\boldsymbol{\beta}_n &=\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}u_n\sqrt{\kappa}\bar{\h} + \left({\mtx{I}}-\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\right)\betab^\star\notag\\
\Longrightarrow\boldsymbol{\beta}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\bar{\h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i.
\end{align}
For $i\in[p]$, define
\begin{align}
\vct{v}_{n,i} = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\bar{\h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i
\end{align}
In the above, for convenience, we have denoted $\xi_n:=u_n/\tau_n$ and recall that $\xi_*:=u_*/\tau_*$.
The proof proceeds in two steps. In the first step, we use the fact that $\xi_n\stackrel{{P}}{\longrightarrow}\xi_*$ and $u_n\stackrel{{P}}{\longrightarrow} u_\star$ (see \eqref{eq:Dc}) to prove that for any $\varepsilon\in(0,\xi_*/2)$, there exists absolute constant $C>0$ such that wpa 1:
\begin{align}\label{eq:ivstep1}
|F_n(\sqrt{p}\boldsymbol{\beta}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.
\end{align}
In the second step, we use Lipschitzness of $f$ and Assumption \ref{ass:mu} to prove that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:ivstep2}
\end{align}
The desired follows by combining \eqref{eq:ivstep1} and \eqref{eq:ivstep2}. Thus, in what follows, we prove \eqref{eq:ivstep1} and \eqref{eq:ivstep2}.
\vspace{2pt}
\noindent\emph{Proof of \eqref{eq:ivstep1}.}~~Fix some $\varepsilon\in(0,\xi_*/2)$. From \eqref{eq:Dc}, we know that w.p.a. 1 $|\xi_n-\xi_*|\leq \varepsilon$ and $|u_n-u_*|\leq \varepsilon$. Thus, $\vct{w}_n$ is close to $\vct{v}_n$. Specifically, in this event, for every $i\in[p]$, it holds that:
\begin{align}
\notag|\boldsymbol{\beta}_{n,i} - \vct{v}_{n,i}| &\leq {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{\tau_n}{1+(\xi_n\bSi_{i,i})^{-1}}-\frac{\tau_*}{1+(\xi_*\bSi_{i,i})^{-1}}\right| \\
\notag&= {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{u_n}{\xi_n+\bSi_{i,i}^{-1}}-\frac{u_*}{\xi_*+\bSi_{i,i}^{-1}}\right| \\
\notag& \leq
{|\betab^\star_i|}\frac{|\bSi_{i,i}||\xi_n-\xi_*|}{|1+\xi_n\bSi_{i,i}||1+\xi_*\bSi_{i,i}|} + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}} \frac{u_*|\xi_n-\xi_*|}{(\xi_n+\bSi_{i,i}^{-1})(\xi_*+\bSi_{i,i}^{-1})} +
\sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}}\frac{|u_n-u_*|}{\xi+\bSi_{i,i}^{-1}}
\\
\notag& \leq
{|\betab^\star_i|}{\Sigma_{\max}\varepsilon} + \sqrt{\kappa}{|\bar{\h}_i|} u_*\Sigma_{\max}^{3/2} \varepsilon+
\sqrt{\kappa}|\bar{\h}_i|\Sigma_{\max}^{1/2}\varepsilon
\\
&\leq \varepsilon \cdot \max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\, \left(|\bar{\h}_i| + |\betab^\star_i|\right).\label{eq:betav}
\end{align}\som{$\varepsilon$ missing}
In the second line above, we recalled that $u_n=\tau_n\xi_n$ and $u_*=\tau_*\xi_*$. In the third line, we used the triangle inequality. In the fourth line, we used that $\xi_*>0$, $0<\bSi_{i,i}\leq\Sigma_{\max}$ and $\xi_n\geq \xi_*-\varepsilon \geq \xi_*/2 >0$.
Now, we will use this and Lipschitzness of $f$ to argue that there exists absolute constant $C>0$ such that wpa 1:
\begin{align}
|F_n(\sqrt{p}\boldsymbol{\beta}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.\notag
\end{align}
Denote, $\vct{a}_i=(\sqrt{p}\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\vct{v}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$. Following the exact same argument as in \eqref{eq:LipC}, we have
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| &\leq \frac{L}{p}\sum_{i=1}^p (1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\})\|\vct{a}_i-\vct{b}_i\|_2\notag \\
&\leq \frac{L}{p}\sum_{i=1}^p \left(1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\}\right)\cdot|\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}) - \sqrt{p}\vct{v}_{n,i}|\notag\\
&\leq L\left(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} \right)^{1/2}{\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \vct{v}_n\|_2}\notag.
\end{align}
From this and \eqref{eq:betav}, we find
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| &\leq
L(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} )^{1/2}
\\
&\qquad\qquad\varepsilon\cdot\sqrt{2}\max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\sqrt{\|\betab^\star\|_2^2 + \|\bar{\h}\|_2^2}.\label{eq:epsS}
\end{align}
As in \eqref{eq:bdmom}, it can be shown that wpa 1: $\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2}<\infty$ and $\frac{1}{p}\sum_{i=1}^p \|\vct{b}_i\|_2^{2k-2}<\infty$, as $p\rightarrow\infty$. Similarly, $\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^2<\infty$, as $p\rightarrow \infty$ by Assumption (A2).
Finally, $\|\bar{\h}\|_2^2\leq 2$, wpa 1 as $p\rightarrow\infty$. Therefore, from \eqref{eq:epsS}, wpa 1, there exists constant $C>0$ such that
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \leq C \cdot\varepsilon,
\end{align}
as desired.
\vspace{2pt}
\noindent\emph{Proof of \eqref{eq:ivstep2}.}~~Next, we will use Assumption \ref{ass:mu} to show that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:AO_conv}
\end{align}
Notice that $\vct{v}_n$ is a function of $\betab^\star,\boldsymbol{\Sigma},\bar{\h}$. Concretely, define $g:\mathbb{R}^3\rightarrow\mathbb{R}$, such that
$$
g(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1,
$$
and notice that
$$
\sqrt{p}\vct{v}_{n,i} = g\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right) = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\vct{h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\sqrt{p}\betab^\star_i.
$$
Thus,
$$
F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma}) = \frac{1}{p}\sum_{i=1}^p f\left(g\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right),\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) =: \frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right),
$$
where we have defined $h:\mathbb{R}^3\rightarrow\mathbb{R}$:
\begin{align}
h(x_1,x_2,x_3) := f\left(g(x_2,x_3,x_1),x_2,x_3\right).\label{h func}
\end{align}
It will suffice to prove that $h\in\rm{PL}(k+1)$. Indeed, if that were the case, then Assumption (A2) gives
\begin{align}
\frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[h(H,B,\Lambda)\right] &= \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[f\left(g(B,\Lambda,H),B,\Lambda\right))\right]\\
& = \operatorname{\mathbb{E}}[f\left(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda\right)] = \alpha_*,
\end{align}
where the penultimate equality follows by recognizing that (cf. Eqn. \eqref{eq:X})
$$
g(B,\Lambda,H) = (1-(1+\xi_*\Lambda)^{-1})B + \sqrt{\kappa}\frac{\tau_*\Lambda^{-1/2}}{1+(\xi_*\Lambda)^{-1}}H = X_{\kappa,\sigma^2}(\Lambda,B,H).
$$
We prove that $h\in\rm{PL}(k)$ in Lemma \ref{lem:hPL} in Section \ref{SM useful fact}.
\begin{lemma} The function $h$ in \eqref{h func} is $\rm{PL}(k)$ as long as $f$ is $\rm{PL}(k)$.
{\color{red}The function $h$ is $\mathbb{R}^3\rightarrow\mathbb{R}$, not $\mathbb{R}^{3p}\rightarrow\mathbb{R}$! See Lemma 6 below.}
\end{lemma}
\begin{proof}
Since $f$ is $\rm{PL}(k)$, for some $L>0$, \eqref{PL func} holds. Fix $\vct{a}=[\vct{x},\vct{y},{\vct{z}}]\in\mathbb{R}^{3p},\vct{a}'=[\vct{x}',\vct{y}',{\vct{z}}']\in\mathbb{R}^{3p}$. Let $\vct{b}=[{\vct{g}},\vct{y},{\vct{z}}]$ where ${\vct{g}}=g(\vct{y},{\vct{z}},\vct{x})$. We have that
\begin{align}
|h(\vct{a})-h(\vct{a}')|&\leq |f(\vct{b})-f(\vct{b}')|\\
&\leq L\left(1+\|\vct{b}\|_2^{k-1}+\|\vct{b}'\|_2^{k-1}\right)\|\vct{b}-\vct{b}'\|_2\\
&\lesssim L\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+\|{\vct{g}}\|_2^{k-1}+\|{\vct{g}}'\|_2^{k-1}\right)(\|\vct{a}-\vct{a}'\|_2+\|{\vct{g}}-{\vct{g}}'\|_2).
\end{align}
Next, we need to bound the ${\vct{g}}$ term in terms of $\vct{a}$. This is accomplished as follows
\begin{align}
\|{\vct{g}}\|_2^{k-1}&\lesssim \frac{1}{p}\sum_{i=1}^p|g_i|^{k-1}\\
&\lesssim \frac{1}{p}\sum_{i=1}^p\left|\frac{y_i^{-1/2}}{1+(\xi y_i)^{-1}}\sqrt{\kappa}\tau_* z_i + (1-(1+\xi_*y_i)^{-1})x_i\right|\\
&\lesssim \frac{1}{p}\sum_{i=1}^p(|z_i|+|x_i|)^{k-1}\lesssim \|\vct{x}\|_2^{k-1}+\|{\vct{z}}\|_2^{k-1}\\
&\lesssim \|\vct{a}\|^{k-1}.
\end{align}
Secondly and similarly, we have the following perturbation bound on the ${\vct{g}}-{\vct{g}}'$. \so{Suppose the triples $(x_i,y_i,z_i)$ takes values in a fixed bounded compact set $\mathcal{M}$. We will prove the following sequence of inequalities}
\begin{align}
\tn{{\vct{g}}-{\vct{g}}'}^2&\leq \sum_{i=1}^p|g(x_i,y_i,z_i)-g(x'_i,y'_i,z'_i)|^2\\
&\leq C^2_{\mathcal{M}} \sum_{i=1}^p(|x_i-x'_i|^2+|y_i-y'_i|^2+|z_i-z'_i|^2)\label{sec line eq}\\
&\leq C^2_{\mathcal{M}} (\tn{\vct{x}-\vct{x}'}^2+\tn{\vct{y}-\vct{y}'}^2+\tn{{\vct{z}}-{\vct{z}}'}^2)\\
&\lesssim \tn{\vct{a}-\vct{a}'}^2.
\end{align}
In what follows, we prove the second line \eqref{sec line eq} i.e.~the fact that for any triples $a=(x,y,z),a'=(x',y',z')$ (with $a,a'\in\mathbb{R}^3$), we have that
\[
|g(a)-g(a')|\leq C_{\mathcal{M}}\tn{a-a'}.
\]
Set $C_\nabla=\sup_{a\in \mathcal{M}}\tn{\nabla g(a)}$. By definition of gradient, the inequality above holds with $C_\mathcal{M}=C_\nabla$. Thus, all that remains is proving that $C_\nabla$ is upper bounded by a constant. \so{However, this automatically holds because from the definition of $g(\dots)$ function, it is clear that $C_\nabla$ is defined everywhere and continuous thus it has a finite maximum over a compact set.}\cts{Only the problem is that $z_i$ (i.e., $\vct{h}_i$) is $\mathcal{N}(0,1)$ so it is not bounded.}
\end{proof}
But from Assumption \ref{ass:mu} and $f\in\rm{PL(2)}$ it follows immediately that with probability approaching 1,
\begin{align}
|\frac{1}{p}\sum_{i=1}^{p}f(\sqrt{p}\vct{v}_{n,i})-\operatorname{\mathbb{E}}_\mu\left[X_{\kappa,\sigma^2}(\Lambda,N,H) \right]| \leq \varepsilon.
\end{align}
Combining the above displays and using triangle inequality completes the proof of the statement.
\subsection{Asymptotic formulas on Magnitude- and Hessian- pruning}
Here, we use Theorem \ref{thm:master_W2} to characterize the risk of the magnitude- and Hessian- pruned solutions. This section supplements the discussion in Section \ref{sec:risk}. First, we formally state the result of Section \ref{sec:risk} as Corollary \ref{cor:mag} characterizing magnitude pruning. Next, we further discuss Hessian pruning.
\subsubsection{Magnitude-based pruning.}
We begin with the following necessary definitions. Define the hard-thresholding function with fixed threshold $t\in\mathbb{R}_+$ as follows:
\begin{align}\label{eq:threshold_app}
\mathcal{T}_t(x) = \begin{cases} x & \text{if } |x|>t \\ 0 & \text{otherwise} \end{cases}.
\end{align}
Further define threshold $t^\star$ as follows:
\begin{align}\label{eq:tstar_app}
t^\star:=\inf\left\{t\in\mathbb{R}\,:\, \Pr(|X_{\kappa,\sigma^2}|\geq t) \geq \alpha \right\}.
\end{align}
\begin{corollary}[Risk of Magnitude-pruning]\label{cor:mag}
Let the same assumptions and notation as in the statement of Theorem \ref{thm:master_W2} hold. Specifically, let $\hat\boldsymbol{\beta}$ be the min-norm solution in \eqref{eq:min_norm} and $\hat\beta_s^M:=\mathbb{T}_s({\boldsymbol{\beta}})$ the magnitude-pruned model at sparsity $s$. Recall the notations in \eqref{eq:threshold_app} and \eqref{eq:tstar_app}. The risk of the magnitude-pruned model satisfies the following in the limit of $n,p,s\rightarrow\infty$ at rates $\kappa:=p/n>1$ and $\alpha:=s/p\in(0,1)$ (cf. Assumption \ref{ass:linear}):
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^M_s) \stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}\left[ \Lambda\left(B-\mathcal{T}_{t^\star}(X_{\kappa,\sigma^2})\right) \right],\notag
\end{align}
where the expectation is over $(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1).$
\end{corollary}
The proof of the corollary above, is given in Section \ref{sec:risk}. Below, we extend the results to Hessian-based pruning.
\subsubsection{Hessian-based pruning.}
Let $\hat\beta$ be the min-norm solution in \eqref{eq:min_norm}. Recall that the Hessian-pruned model $\boldsymbol{\beta}_s^H$ at sparsity $s$ is given by
\begin{align}\label{eq:hp}
\boldsymbol{\beta}_s^H = \hat\boldsymbol{\Sigma}^{-1/2}\mathbb{T}_s(\hat\boldsymbol{\Sigma}^{1/2}\boldsymbol{\beta}),
\end{align}
where $\hat\boldsymbol{\Sigma}=\diag{{\mtx{X}}^T{\mtx{X}}}/n$ the diagonal of the empirical covariance matrix.
\begin{corollary}[Risk of Hessian-pruning]\label{cor:hes}
Let the same assumptions and notation as in the statement of Theorem \ref{thm:master_W2} hold. Specifically, let $\hat\boldsymbol{\beta}$ be the min-norm solution in \eqref{eq:min_norm} and $\hat\beta_s^H$ the Hessian-pruned model at sparsity $s$. Recall the notation in \eqref{eq:threshold_app} and define
\begin{align}\label{eq:tdiam_app}
t^\diamond:=\inf\left\{t\in\mathbb{R}\,:\, \Pr(|\Lambda^{1/2}X_{\kappa,\sigma^2}|\geq t) \geq \alpha \right\}.
\end{align}
The risk of the magnitude-pruned model satisfies the following in the limit of $n,p,s\rightarrow\infty$ at rates $\kappa:=p/n>1$ and $\alpha:=s/p\in(0,1)$ (cf. Assumption \ref{ass:linear}):
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^H_s) \stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}\left[ \Lambda\left(B-\Lambda^{-1/2}\mathcal{T}_{t^\diamond}(\Lambda^{1/2}X_{\kappa,\sigma^2})\right) \right],\notag
\end{align}
where the expectation is over $(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1).$
\end{corollary}
\begin{proof}
\end{proof}
\section{Facts about Lipschitz functions}\label{SM useful fact}
For $k\geq 1$ we say a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$ and denote it by $f\in \rm{PL}(k)$ if there exists a cosntant $L>0$ such that, for all $\vct{x},\vct{y}\in\mathbb{R}^m$:
\begin{align}
|f(\vct{x})-f(\vct{y})|\leq L\left(1+\|\vct{x}\|_2^{k-1}+\|\vct{y}\|_2^{k-1}\right)\|\vct{x}-\vct{y}\|_2.\label{PL func}
\end{align}
In particular, when $f\in\rm{PL}(k)$, the following properties hold:
\begin{enumerate}
\item There exists a constant $L'$ such that for all $\vct{x}\in\mathbb{R}^n$: $|f(\vct{x})|\leq L'(1+\|\vct{x}\|_2^k).$
\item $f$ is locally Lipschitz, that is for any $M>0$, there exists a constant $L_{M,m}<\infty$ such that for all $x,y\in[-M,M]^m$,
$
|f(\vct{x})-f(\vct{y})| \leq L_{M,m}\|\vct{x}-\vct{y}\|_2.
$
Further, $L_{M,m}\leq c(1+(M\sqrt{m})^{k-1})$ for some costant $c$.
\end{enumerate}
\begin{lemma}
Let $g:\mathbb{R}\rightarrow\mathbb{R}$ be a Lipschitz function. Consider the function $f:\mathbb{R}^3\rightarrow\mathbb{R}$ defined as follows:
$$
f(\vct{x}) = x_1(x_2-g(x_3))^2.
$$
Then, $f\in\rm{PL}(3).$
\end{lemma}
\begin{proof}
Let $h:\mathbb{R}^2\rightarrow\mathbb{R}$ defined as $h({\vct{u}})=({\vct{u}}_1-g({\vct{u}}_2))^2$. The function $({\vct{u}}_1,{\vct{u}}_2)\mapsto{\vct{u}}_1-g({\vct{u}}_2)$ is clearly Lipschitz. Thus, $h\in\rm{PL}(2)$, i.e.,
\begin{align}
|h({\vct{u}})-h(\vct{v})| \leq C(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2\quad\text{and}\quad |h(\vct{v})|\leq C'(1+\|\vct{v}\|_2^2).\label{eq:h_pl}
\end{align}
Therefore, letting $\vct{x}=(x_1,{\vct{u}})\in\mathbb{R}^3$ and $\vct{y}=(y_1,\vct{v})\in\mathbb{R}^3$, we have that
\begin{align}
|f(\vct{x})-f(\vct{y})| &= |x_1h({\vct{u}}) - y_1h(\vct{v})| \leq |x_1||h({\vct{u}})-h(\vct{v})| + |h(\vct{v})| |x_1-y_1|\notag\\
&\leq C|x_1|(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(|x_1|^2+(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|\vct{x}-\vct{y}\|_2.
\end{align}
In the second line, we used \eqref{eq:h_pl}. In the third line, we used $2xy\leq x^2+y^2$. In the fourth line, we used Cauchy-Schwarz inequality. $C,C'>0$ are absolute constants that may change from line to line.
This completes the proof of the lemma.
\end{proof}
\begin{lemma}\label{lem:hPL}
Let functions $f,g:\mathbb{R}^3\rightarrow\mathbb{R}$ such that $f\in\rm{PL}(k)$ and
$$
g(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi_* x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1.
$$
Here, $\xi_*,\tau_*,\kappa$ are positive constants.
Further define
\begin{align}
h(x,y,z) := f\left(g(y,z,x),y,z\right),
\end{align}
and assume that $y$ take values on a fixed bounded compact set $\mathcal{M}$.
Then, it holds that $h\in\rm{PL}(k+1)$.
\end{lemma}
\begin{proof}
Since $f$ is $\rm{PL}(k)$, for some $L>0$, \eqref{PL func} holds. Fix $x,x'\in\mathbb{R}$, $\vct{a}=[y,z]\in\mathbb{R}^{2}$ and $\vct{a}'=[y',z']\in\mathbb{R}^{2}$. Let $\vct{b}=[{\vct{g}},\vct{a}]=[{\vct{g}},y,z]\in\mathbb{R}^3$ where ${\vct{g}}=g(y,z,x)\in\mathbb{R}$ and define accordingly $\vct{b}'$ and ${\vct{g}}'$. We have that
\begin{align}
|h([x,\vct{a}])-h([x,\vct{a}'])|&= |f(\vct{b})-f(\vct{b}')|\notag\\
&\leq L\left(1+\|\vct{b}\|_2^{k-1}+\|\vct{b}'\|_2^{k-1}\right)\|\vct{b}-\vct{b}'\|_2\notag\\
&\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|{\vct{g}}|^{k-1}+|{\vct{g}}'|^{k-1}\right)(\|\vct{a}-\vct{a}'\|_2+|{\vct{g}}-{\vct{g}}'|),\label{eq:hlip}
\end{align}
for some constant $C>0$. In the last inequality we have repeatedly used the inequality
$
\left(\sum_{i=1}^m \|\vct{v}_i\|_2^2\right)^{\frac{d}{2}} \leq C(m)\cdot\sum_{i=1}^m\|\vct{v}_i\|_2^{d}.
$
Next, we need to bound the ${\vct{g}}$ term in terms of $\vct{a}$. This is accomplished as follows
\begin{align}
|{\vct{g}}|^{k-1} &= \left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x + (1-(1+\xi_*y)^{-1})z\right|^{k-1}\notag\\
&\leq C (|z|+|x|)^{k-1} \notag\\
&\leq C \left(|x|^{k-1}+|z|^{k-1}\right) \leq C (|x|^{k-1} + \|\vct{a}\|_2^{k-1}).\label{eq:gb}
\end{align}
Here, the value of the constant $C>0$ may change from line to line.
Secondly and similarly, we have the following perturbation bound on the ${\vct{g}}-{\vct{g}}'$. Recall that $(x,y)$ and $(x',y')$ are bounded. We will prove the following sequence of inequalities
\begin{align}
|{{\vct{g}}-{\vct{g}}'}|&= |g(y,z,x)-g(y',z',x')| \notag\\
&\leq \left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x - \frac{(y')^{-1/2}}{1+(\xi y')^{-1}}\sqrt{\kappa}\tau_* x'\right|
+ \left|(1-(1+\xi_*y)^{-1})z-(1-(1+\xi_*y')^{-1})z'\right|\notag\\
&\leq
\left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x - \frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x'\right| +\left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x'- \frac{(y')^{-1/2}}{1+(\xi y')^{-1}}\sqrt{\kappa}\tau_* x'\right|\notag\\
&\qquad+ \left|(1-(1+\xi_*y)^{-1})z-(1-(1+\xi_*y)^{-1})z'\right| + \left|(1-(1+\xi_*y)^{-1})z'-(1-(1+\xi_*y')^{-1})z'\right| \notag\\
&\leq C_1|x-x'| + C_2|x'||y-y'| + C_3|z-z'| + C_4|z'||y-y'|\notag\\
&\leq C(1+|x'| + |z'|)(|x-x'| + |z-z'| + |y-y'|)\\
&\leq C\sqrt{3}(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2
.\label{eq:gd}
\end{align}
In the fourth inequality above, we used the fact the assumption that $|y|$ is bounded. In the last line, we used Cauchy-Scwhartz.
Substituting \eqref{eq:gb} and \eqref{eq:gd} in \eqref{eq:hlip} gives:
\begin{align}
|h(x,y,z) - h(x',y',z')| &\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|x|^{k-1}+|x'|^{k-1}\right)(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2\notag \\
&\leq C\left(1+\|[\vct{a},x]\|_2^{k}+\|[\vct{a}',x']\|_2^{k}\right)\|[\vct{a},x]-[\vct{a}',x']\|_2.
\end{align}
Thus, $h\in\rm{PL}(k+1)$, as desired.
\end{proof}
\section{Proofs for overparameterized least-squares}\label{sec proof thm 1}
In this section, we assume the linear Gaussian problem (LGP) of Definition \ref{def LGP}, the overparameterized regime $k=p>n$ and the min-norm model $\hat\boldsymbol{\beta}$ of \eqref{eq:min_norm}. We prove Theorem \ref{thm:master_W2} that derives the asymptotic DC of $\hat\boldsymbol{\beta}$ and we show how this leads to sharp formulae for the risk of the Magnitude- and Hessian-pruned models.
\subsection{Notation and Assumptions}\label{sec:ass_app}
For the reader's convenience, we recall some necessary notation and assumptions from Section \ref{sec main}. We say that a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$, denoted $f\in\rm{PL}(k)$, if there is a constant $L>0$ such that for all $\vct{x},\vct{y}\in\mathbb{R}^m$,
$
|f(\vct{x}) - f(\vct{y})|\leq L(1+\tn{\vct{x}}^{k-1}+\tn{\vct{y}}^{k-1})\|\vct{x}-\vct{y}\|_2
$
(See also Section \ref{SM useful fact}).
We say that a sequence of probability distributions $\nu_p$ on $\mathbb{R}^m$ {converges in $W_k$} to
$\nu$, written $\nu_p\stackrel{W_k}{\Longrightarrow} \nu$, if $W_k(\nu_p,\nu) \rightarrow 0$ as $p \rightarrow \infty$. An equivalent definition is that, for any $f\in\rm{PL}(k)$, $\lim_{p\rightarrow\infty}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$ (e.g., \cite{montanari2017estimation}).
Finally, recall that a sequence of probability distributions $\nu_n$ on $\mathbb{R}^m$ \emph{converges weakly} to $\nu$, if for any bounded Lipschitz function $f$: $\lim_{p\rightarrow\infty}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$.
Throughout, we use $C,C',c,c'$ to denote absolute constants (not depending on $n,p$) whose value might change from line to line.
We focus on a double asymptotic regime where:
$$n,p,s\rightarrow\infty \text{ at fixed overparameterization ratio } \kappa:=p/n>0 \text{ and sparsity level } \alpha:=s/p\in(0,1).$$
For a sequence of random variables $\mathcal{X}_{p}$ that converge in probability to some constant $c$ in the limit of Assumption \ref{ass:linear} below, we write $\mathcal{X}_{p}\stackrel{{P}}{\longrightarrow} c$. For a sequence of event $\mathcal{E}_p$ for which $\lim_{p\rightarrow}\mathbb{P}(\mathcal{E}_p) = 1$, we say that $\mathcal{E}_p$ occurs \emph{with probability approaching 1}. For this, we will often use the shorthand ``wpa. 1".
Next, we recall the set of assumption under which our analysis applies:
\textbf{(A1)[Diagonal covariance].}~~The covariance matrix $\boldsymbol{\Sigma}$ is diagonal.
\textbf{(A2)[Boundedness and empirical distribution].}~~There exist constants $\Sigma_{\min},\Sigma_{\max}\in(0,\infty)$ such that:
$
\Sigma_{\min}\leq{\boldsymbol{{\Sigma}}}_{i,i}\leq \Sigma_{\max}.
$
for all $i\in[p].$
\cts{Furthermore, the joint empirical distribution of $\{(\bSi_{i,i},\sqrt{p}\betab^\star_i)\}_{i\in[p]}$ converges weakly to a probability distribution $\mu$ on $\mathbb{R}_{>0}\times\mathbb{R}$ with bounded $(2k-2)$-th moment, and assume that
$$
\frac{1}{p}\sum_{i\in[p]}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^{2k-2} \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(\Lambda,B)\sim\mu}\left[ B^{2k-2}\right]<\infty,
$$
as $p\rightarrow\infty$ for some $k\geq 3$.}
We remark that Assumption (A2) above implies (see \cite[Lem.~5]{bayati2011dynamics}) that for any pseudo-Lipschitz function $\psi:\mathbb{R}^2\rightarrow\mathbb{R}$ of order $2k-2$:
$$
\frac{1}{p}\sum_{i=1}^p\psi(\bSi_{i,i},\sqrt{p}\betab^\star_i) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(\Lambda,B)\sim\mu}\left[ \psi(\Lambda,B)\right].
$$
\subsection{Proof of Asymptotic DC}
In this section, we prove the following main result (which is a restatement of Theorem \ref{thm:master_W2}.
\begin{theorem}[Asymptotic DC for LGP]\label{thm:master_W2}
Fix $\kappa>1$. Let Assumptions (A1) and (A2) in Section \ref{sec:ass_app} hold for some $k\geq 3$. Consider $\hat{{\boldsymbol{\beta}}}$ as in \eqref{eq:min_norm} and $\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p}\sum_{i=1}^{p}\delta_{\sqrt{p}\hat{{\boldsymbol{\beta}}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}}$, the joint empirical distribution of $(\sqrt{p}\hat\boldsymbol{\beta},\sqrt{p}\betab^\star,\boldsymbol{\Sigma})$. Recall the definition of the measure $\Pi_{\kappa,\sigma^2}$ in Definition \ref{def:Xi}. Then, $\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma})$ converges in Wasserstein-k distance to $\Pi_{\kappa,\sigma^2}\otimes\mu$.
Specifically, for any function $f:\mathbb{R}^k\rightarrow\mathbb{R}$, $f\in\rm{PL}(k)$, it holds that
\begin{align}\label{eq:thm}
\frac{1}{p} \sum_{i=1}^{p} f(\sqrt{p}\hat\boldsymbol{\beta}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)}\left[f(X_{\kappa,\sigma^2},B,\Lambda) \right].
\end{align}
\end{theorem}
\vspace{5pt}
Let ${\mtx{X}}\in\mathbb{R}^{n\times p}$ have zero-mean and normally distributed rows with a diagonal covariance matrix ${\boldsymbol{{\Sigma}}}=\operatorname{\mathbb{E}}[\vct{x}\x^T]$. Given a ground-truth vector $\betab^\star$ and labels $\vct{y}={\mtx{X}}\betab^\star+\sigma {\vct{z}},~{\vct{z}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, we consider the least-squares problem subject to the minimum Euclidian norm constraint (as $\kappa=p/n>1$) given by
\begin{align}\label{eq:PO_beta}
\min_{{\boldsymbol{\beta}}}\frac{1}{2}\tn{{\boldsymbol{\beta}}}^2\quad\text{subject to}\quad \vct{y}={\mtx{X}}{\boldsymbol{\beta}}.
\end{align}
It is more convenient to work with the following change of variable: $\vct{w}:=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\beta}}-\betab^\star)$. With this, the optimization problem in \eqref{eq:min_norm} can be rewritten as
\begin{align}\label{eq:PO}
\Phi({\mtx{X}})=\min_{\vct{w}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
where we write ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$. First, using standard arguments, we show that the solution of \eqref{eq:PO} is bounded. Hence, we can constraint the optimization in a sufficiently large compact set without loss of generality.
\begin{lemma}[Boundedness of solution]\label{lem:bd_PO}
Let $\hat\vct{w}_n:=\hat\vct{w}_n({\mtx{X}},{\vct{z}})$ be the minimizer in \eqref{eq:PO}. Then, with probability approaching 1, it holds that $\hat\vct{w}_n\in\mathcal{B}$, where
$$\mathcal{B}:=\left\{\vct{w}\,|\,\|\vct{w}\|_2\leq B_{+} \right\},\qquad B_+:=4\sqrt{\Sigma_{\max}}\frac{2(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + \sigma}{\sqrt{\Sigma_{\min}}(\sqrt{\kappa}-1)} + \sqrt{\Sigma_{\max}}\sqrt{\operatorname{\mathbb{E}}\left[B^2\right]}.
$$
\end{lemma}
\begin{proof}
First, we show that the min-norm solution $\hat\boldsymbol{\beta}={\mtx{X}}^T({\mtx{X}}\X^T)^{-1}\vct{y}$ of \eqref{eq:PO_beta} is bounded. Here, we used the fact that $\kappa>1$, thus ${\mtx{X}}\X^T$ is invertible wpa 1. We
have,
\begin{align}
\tn{\hat\boldsymbol{\beta}_n}^2 = \vct{y}^T({\mtx{X}}\X^T)^{-1}\vct{y} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{X}}\X^T)} = \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\boldsymbol{\Sigma}{\mtx{\bar{X}}}^T)} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\Xb^T)\,\Sigma_{\min}} = \frac{\tn{\vct{y}}^2}{\sigma_{\min}^2({\mtx{\bar{X}}})\,\Sigma_{\min}}. \label{eq:Ubb}
\end{align}
But, wpa 1,
$
\sigma_{\min}({\mtx{\bar{X}}})/\sqrt{n} \geq \frac{1}{2}\left(\sqrt{\kappa}-1\right).
$
Furthermore,
$
\|\vct{y}\|_2 \leq \|{\mtx{\bar{X}}}\boldsymbol{\Sigma}^{1/2}\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2 \leq \sigma_{\max}({\mtx{\bar{X}}})\sqrt{\Sigma_{\max}} \|\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2.
$
Hence, wpa 1,
$$
\|\vct{y}\|_2/\sqrt{n} \leq 4(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma,
$$
where we used the facts that wpa 1: $\|z\|_2/\sqrt{n}\stackrel{{P}}{\longrightarrow} 1$, $\sigma_{\max}({\mtx{\bar{X}}})<2(\sqrt{\kappa}+1)$ and by Assumption \ref{ass:mu}:
$$
\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^{p}(\sqrt{p}\betab^\star_i)^2 \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[B^2\right].
$$
Put together in \eqref{eq:Ubb}, shows that
\begin{align}
\tn{\hat\boldsymbol{\beta}_n} < \frac{4(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma}{\sqrt{\Sigma_{\min}}(\sqrt{\kappa}-1)/2} =: \tilde{B}_+.\label{eq:bd_beta}
\end{align}
Recalling that $\hat\vct{w}_n= \sqrt{\boldsymbol{\Sigma}}\boldsymbol{\beta} - \sqrt{\boldsymbol{\Sigma}}\betab^\star$, we conclude, as desired, that wpa 1,
$
\tn{\hat\vct{w}_n} \leq \sqrt{\Sigma_{\max}}\tilde{B}_+ + \sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}[B^2]} =: B_+.
$
\end{proof}
\input{kth_bounded}
Lemma \ref{lem:bd_PO} implies that nothing changes in \eqref{eq:PO} if we further constrain $\vct{w}\in\mathcal{B}$ in \eqref{eq:PO}. Henceforth, with some abuse of notation, we let
\begin{align}\label{eq:PO_bd}
\Phi({\mtx{X}}):=\min_{\vct{w}\in\mathcal{B}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
Next, in order to analyze the primary optimization (PO) problem in \eqref{eq:PO_bd} in apply the CGMT \cite{thrampoulidis2015lasso}. Specifically, we use the constrained formulation of the CGMT, Theorem \ref{thm closed}. Specifically, the auxiliary problem (AO) corresponding to \eqref{eq:PO_bd} takes the following form with ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$
\begin{align}
\phi({\vct{g}},\vct{h}) = \min_{\vct{w}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_con}
\end{align
We will prove the following result about the AO problem.
\begin{lemma}[Properties of the AO -- Overparameterized regime]\label{lem:AO}
Let $\phi_n=\phi({\vct{g}},\vct{h})$ be the optimal cost of the minimization in \eqref{eq:AO_con}. Define $\bar\phi$ as the optimal cost of the following deterministic min-max problem
\begin{align}\label{eq:AO_det}
\bar\phi:=\max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau):=\frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{N^2}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).
\end{align}
The following statements are true.
\noindent{(i).}~The AO minimization in \eqref{eq:AO_con} is $\frac{1}{\Sigma_{\max}}$-strongly convex and has a unique minimizer $\vct{w}_n:=\vct{w}_n({\vct{g}},\vct{h})$.
\noindent{(ii).}~In the limit of $n,p\rightarrow\infty, p/n=\kappa$, it holds that $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$, i.e., for any $\varepsilon>0$:
$$
\lim_{n\rightarrow\infty}\P\left(|\phi({\vct{g}},\vct{h})-\bar\phi|>\varepsilon\right) = 0.
$$
\noindent{(iii).} The max-min optimization in \eqref{eq:AO_det} has a unique saddle point $(u_*,\tau*)$ satisfying the following:
$$
u_*/\tau_* = \xi\quad\text{and}\quad\tau_* = \gamma,
$$
where $\xi, \gamma$ are defined in Definition \ref{def:Xi}.
\noindent{(iv).}~Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be a $\rm{PL}({k})$ function. Let $\boldsymbol{\beta}_n=\boldsymbol{\Sigma}^{-1/2}\vct{w}_n + \betab^\star$. Then,
$$
\frac{1}{p}\sum_{i=1}^{p}f\left(\sqrt{p}\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\sqrt{p}\betab^\star,\boldsymbol{\Sigma}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(B,\Lambda,H)\sim\mu\otimes \mathcal{N}(0,1)}\left[f\left(X_{\kappa,\sigma^2}(B,\Lambda,H),B,\Lambda\right) \right].
$$
\noindent{(v).}~\cts{The empirical distribution of $\boldsymbol{\beta}_n$ converges weakly to the measure of $X_{\kappa,\sigma^2}$, and also,
\begin{align}\label{eq:k_AO}
\frac{1}{p}\sum_{i=1}^{p}|\sqrt{p}\boldsymbol{\beta}_n({\vct{g}},\vct{h})|^{2k-2} < \infty.
\end{align}
}
\end{lemma}
We prove Lemma \ref{lem:AO} in Section \ref{sec:proofAO}. Here, we show how this leads to the proof of Theorem \ref{thm:master_W2} when combined with the CGMT.
For convenience, define
$$F_n(\hat\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}\hat\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Fix any $\varepsilon>0$. It suffices to prove that the solution $\hat\boldsymbol{\beta}_n$ of the PO in \eqref{eq:PO_beta} satisfies $\hat\boldsymbol{\beta}_n\not\in\mathcal{S}$ wpa. 1, where
\begin{align}\label{eq:S_set}
\mathcal{S} = \mathcal{S}(\betab^\star,\boldsymbol{\Sigma}) =\{\boldsymbol{\beta}{~\big |~} |F_n(\hat\boldsymbol{\beta}_n,\betab^\star,\boldsymbol{\Sigma})-\alpha_*|\geq 2\varepsilon\}.
\end{align}
In particular, define the ``perturbed" PO and AO problems (compare to \eqref{eq:PO} and \eqref{eq:AO_con}) as:
\begin{align}\label{eq:PO_S}
\Phi_S({\mtx{X}})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
and
\begin{align}
\phi_S({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_S}
\end{align
where recall that ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$, ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$ and we have used the change of variables $\vct{w}:=\sqrt{\boldsymbol{\Sigma}}(\boldsymbol{\beta}-\betab^\star)$ for convenience.
Using \cite[Theorem 6.1(iii)]{thrampoulidis2018precise} it suffices to find costants $\bar\phi, \bar\phi_S$ and $\eta>0$ such that the following hold:
\begin{enumerate}
\item $\bar\phi_S \geq \bar\phi + 3\eta$,
\item $\phi({\vct{g}},\vct{h}) \leq \bar\phi + \eta$, with probability approaching 1,
\item $\phi_S({\vct{g}},\vct{h}) \geq \bar\phi_S - \eta$, with probability approaching 1.
\end{enumerate}
In what follows, we explicitly find $\bar\phi, \bar\phi_S,\eta$ such that the three conditions above hold.
\vspace{5pt}
\noindent\underline{Satisfying Condition 2}: Recall the deterministic min-max optimization in \eqref{eq:AO_det}. Choose $\bar\phi=\mathcal{D}(u_*,\tau_*)$ be the optimal cost of this optimization. From Lemma \ref{lem:AO}(ii), $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$. Thus, for any $\eta>0$, with probability approaching 1:
\begin{align}\label{eq:phi_lim}
\bar\phi + \eta \geq \phi({\vct{g}},\vct{h}) \geq \bar\phi - \eta.
\end{align}
Clearly then, Condition 2 above holds for any $\eta>0$.
\vspace{5pt}
\noindent\underline{Satisfying Condition 3}: Next, we will show that the third condition holds for appropriate $\bar\phi$.
Let $\vct{w}_n=\vct{w}_n({\vct{g}},\vct{h})$ be the unique minimizer of \eqref{eq:AO_con} as per Lemma \ref{lem:AO}(i), i.e., $\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}_n+\boldsymbol{\beta}}^2 = \phi({\vct{g}},\vct{h})$. Again from Lemma \ref{lem:AO}, the minimization in \eqref{eq:AO_con} is $1/\Sigma_{\max}$-strongly convex in $\vct{w}$. Here, $\Sigma_{\max}$ is the upper bound on the eigenvalues of $\boldsymbol{\Sigma}$ as per Assumption \ref{ass:inv}. Thus, for any $\tilde\varepsilon>0$ and any feasible $\vct{w}$ the following holds (deterministically):
\begin{align}\label{eq:sc}
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\boldsymbol{\beta}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2C},~\text{provided that}~ \|\vct{w}-\vct{w}_n({\vct{g}},\vct{h})\|_2 \geq \tilde\epsilon.
\end{align}
Now, we argue that $\boldsymbol{\beta}\in\mathcal{S}$ implies that $\|\vct{w}-\vct{w}_n({\vct{g}},\vct{h})\|_2\geq \tilde\varepsilon$ wpa 1, for
appropriate value of $\tilde\varepsilon$ and $\vct{w}=\sqrt{\boldsymbol{\Sigma}}(\boldsymbol{\beta}-\betab^\star)$. Consider any $\boldsymbol{\beta}\in\mathcal{S}$. First, by definition in \eqref{eq:S_set},
$$
|F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \geq 2\varepsilon.
$$
Second, by Lemma \ref{lem:AO}(iv), with probability approaching 1,
$$
|F(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \leq \epsilon.
$$
Third, we will show that wpa 1, there exists universal constant $C>0$ such that
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| \leq C {\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}\label{eq:dev2show}.
\end{align}
Before proving \eqref{eq:dev2show}, let us argue how combining the above three displays shows the desired. Indeed, wpa. 1,
\begin{align*}
2\varepsilon &\leq |F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \leq |F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| + |F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \\
&\leq \epsilon + C \,\|\boldsymbol{\beta}-\boldsymbol{\beta}_n\|_2. \\
&\qquad\Longrightarrow \|\boldsymbol{\beta}-\boldsymbol{\beta}_n\|_2 \geq {\varepsilon}/{C}=:\hat\varepsilon\\
&\qquad\Longrightarrow \|\vct{w}-\vct{w}_n\|_2 \geq \hat\varepsilon\sqrt{\Sigma_{\min}}=:\tilde\varepsilon.
\end{align*}
In the last line above, we recalled that $\boldsymbol{\beta}=\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star$ and $\boldsymbol{\Sigma}_{i,i}\geq\Sigma_{\min},~i\in[p]$ by Assumption \ref{ass:inv}.
From this and \eqref{eq:sc}, we find that wpa 1,
$
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\boldsymbol{\beta}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2C},~\text{for all}~ \vct{w}\in\mathcal{S}.
$
Thus,
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}.\notag
\end{align}
When combined with \eqref{eq:phi_lim}, this shows that
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta.
\end{align}
Thus, choosing $\bar\phi_S = \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}$ proves the Condition 3 above.
To complete the proof, let us now show \eqref{eq:dev2show}. Henceforth, $C$ is used to denote a universal constant whose value can change from line to line.
To simplify notation, let $\vct{a}_i=(\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}),\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\boldsymbol{\beta}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$, for $i\in[p]$. Then, using $f\in\rm{PL}(k)$,
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| &\leq \frac{L}{p}\sum_{i=1}^p (1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\})\|\vct{a}_i-\vct{b}_i\|_2\notag \\
&= \frac{L}{p}\sum_{i=1}^p \left(1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\}\right)\cdot|\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}) - \sqrt{p}\boldsymbol{\beta}_i|\notag\\
&\leq L\left(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} \right)^{1/2}{\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \boldsymbol{\beta}\|_2}\label{eq:LipC}\,.
\end{align}
The last inequality above follows by applying Cauchy-Schwartz.
To reach \eqref{eq:dev2show} from the above, it suffices to show that
\begin{align}
\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} \leq C,\label{eq:leqC}
\end{align}
for some universal constant $C$ (that may depend on $k$, but not on $n,p$).
To prove \eqref{eq:leqC}, note that
using $a^{\ell_1}b^{\ell_2}c^{\ell_3}\leq a^{\ell}+b^{\ell}+c^\ell,~\ell=\ell_1+\ell_2+\ell_3$, for some constant $C=C(k)>0$ it holds:
\begin{align}
\frac{1}{p}\sum_{i=1}^p \|\vct{b}_i\|_2^{2k-2} {\leq} C\left(\frac{1}{p}\sum_{i=1}^p |\sqrt{p}\boldsymbol{\beta}_{i}|^{2k-2} + \frac{1}{p}\sum_{i=1}^p |\sqrt{p}\betab^\star_{i}|^{2k-2} + \frac{1}{p}\sum_{i=1}^p |\boldsymbol{\Sigma}_{i,i}|^{2k-2} \right)\label{eq:bdmom}\,.
\end{align}
\cts{The terms $\frac{1}{p}\sum_{i=1}^p |\sqrt{p}\betab^\star_{i}|^{2k-2} $ and $ \frac{1}{p}\sum_{i=1}^p |\boldsymbol{\Sigma}_{i,i}|^{2k-2}$ are bounded by Assumption (A2).}
{\color{red}Furthermore we additionally know from Lemma \ref{subexp bth} that,
\begin{align}
\frac{1}{p}\sum_{i=1}^p |\sqrt{p}\boldsymbol{\beta}_{i}|^{2k-2} <\infty
\end{align}
}
These together with \eqref{eq:bdmom} show that $\frac{1}{p}\sum_{i=1}^p \|\vct{b}_i\|_2^{2k-2}<C$ wpa 1. Similarly, we can argue for $\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2}$, completing the proof of \eqref{eq:leqC}.
\cts{\textbf{Remark}:Note that for the specific $f(x,y,z) = z(y-g(x))^2$ that we care about for evaluating the risk (see Lemma 5) we can write \eqref{eq:LipC} simpler as follows:
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma})| &=
\frac{1}{p}\sum_{i=1}^p|\bSi_{i,i}|\left| (\sqrt{p}\betab^\star_i-g(\sqrt{p}\boldsymbol{\beta}_{n,i}))^2-(\sqrt{p}\betab^\star_i-g(\sqrt{p}\boldsymbol{\beta}_{i}))^2 \right|\notag\\
&\leq
\Sigma_{\max} \frac{1}{p}\sum_{i=1}^p \left| (\sqrt{p}\betab^\star_i-g(\sqrt{p}\boldsymbol{\beta}_{n,i}))^2-(\sqrt{p}\betab^\star_i-g(\sqrt{p}\boldsymbol{\beta}_{i}))^2 \right|\notag\\
&\leq
\Sigma_{\max} L \frac{1}{p}\sum_{i=1}^p (1+ \|\sqrt{p}[\betab^\star_i,\boldsymbol{\beta}_{n,i}]\|_2 + \|\sqrt{p}[\betab^\star_i,\boldsymbol{\beta}_{i}]\|_2) \sqrt{p}|\boldsymbol{\beta}_{n,i}-\boldsymbol{\beta}_{i}|\notag\\
&\leq
\Sigma_{\max} L \left(1+ \frac{1}{p}\sum_{i=1}^{p}\|\sqrt{p}[\betab^\star_i,\boldsymbol{\beta}_{n,i}]\|_2^2 + \frac{1}{p}\sum_{i=1}^p\|\sqrt{p}[\betab^\star_i,\boldsymbol{\beta}_{i}]\|_2^2\right) \|\boldsymbol{\beta}_{n}-\boldsymbol{\beta}\|_2\notag\\
&\leq
C \left(1+ \|\betab^\star\|_2^2+ \|\boldsymbol{\beta}_{n}\|_2^2 + \|\boldsymbol{\beta}\|_2^2\right) \|\boldsymbol{\beta}_{n}-\boldsymbol{\beta}\|_2.
\end{align}
And everything is easy now since $\|\betab^\star\|_2^2$ is bounded as long as second moments of empirical distribution as bounded, $\|\boldsymbol{\beta}_n\|_2^2$ is bounded if $\|\betab^\star\|_2^2$ is bounded, and, $\|\boldsymbol{\beta}\|_2^2$ is bounded, since $\boldsymbol{\beta}\in\mathcal{B}$ by Lemma 1.
}
\vspace{5pt}
\noindent\underline{Satisfying Condition 1:} To prove Condition 1, we simply pick $\eta$ to satisfy the following
\begin{align}
\bar\phi_S > \bar\phi + 3 \eta ~\Leftarrow~ \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta \geq \bar\phi + 3 \eta ~\Leftarrow~ \eta \leq \frac{\tilde\epsilon^2}{8\Sigma_{\max}}.
\end{align}
This completes the proof of Theorem \ref{thm:master_W2}.
\subsection{Proof of Lemma \ref{lem:AO}}\label{sec:proofAO}
~~~~
~~~~
\vspace{3pt}
\noindent\underline{Proof of (i):}~Strong convexity of the objective function in \eqref{eq:PO} is easily verified by the second derivative test. Note here that we use Assumption \ref{ass:inv} that $\bSi_{i,i}\leq\Sigma_{\max},~i\in[p].$ Uniqueness of the solution follows directly from strong convexity. \ct{Strictly speaking we might need to also argue existence, i.e., feasibility of the AO. An indirect way is to show feasibility using the CGMT, but it seems unnecessarily complicated?}
\vspace{5pt}
\noindent\underline{Proof of (ii):}~Using Lagrangian formulation, the solution $\vct{w}_n$ to \eqref{eq:AO_con} is the same as the solution to the following:
\begin{align}
\left(\vct{w}_n,u_n\right) :=\arg\min_{\vct{w}\in\mathcal{B}}\max_{u\geq 0} ~\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + u \left( \sqrt{\tn{\vct{w}}^2+\sigma^2} \tn{\bar{\g}} - \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} + \frac{\sigma h}{\sqrt{n}} \right)\label{eq:AO_2}
\end{align}
where we have: (i) set $\bar{\g} := {\vct{g}}/\sqrt{n}$ and $\bar{\h}:= \vct{h}/\sqrt{p}$; (ii) recalled that $p/n=\kappa$; and, (iii) used $\left(\vct{w}_n,u_n\right)$ to denote the optimal solutions in \eqref{eq:AO_2}. The subscript $n$ emphasizes the dependence of $\left(\vct{w}_n,u_n\right)$ on the problem dimensions. Also note that (even though not explicit in the notation) $\left(\vct{w}_n,u_n\right)$ are random variables depending on the realizations of $\bar{\g},\bar{\h}$ and $h$.
Notice that the objective function above is convex in $\vct{w}$ and linear (thus, concave) in $u$. Thus, strong duality holds and we can flip the order of min-max. Moreover, in order to make the objective easy to optimize with respect to $\vct{w}$, we use the following variational expression for the square-root term $\sqrt{\tn{\vct{w}}^2+\sigma^2}$:
$$
\tn{\bar{\g}}\sqrt{\tn{\vct{w}}^2+\sigma^2} = \tn{\bar{\g}}\cdot\min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau}{2} + \frac{\tn{\vct{w}}^2+\sigma^2}{2\tau} \right\} = \min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau\tn{\bar{\g}}^2}{2} + \frac{\sigma^2}{2\tau} + \frac{\tn{\vct{w}}^2}{2\tau} \right\},
$$
where $B$ is defined in Lemma \ref{lem:bd_PO}. For convenience define the constraint set for the variable $\tau$ as $\mathcal{T}':=[\sigma,\sqrt{\sigma^2+B_+^2}]$. For reasons to be made clear later in the proof (see proof of statement (iii)), we consider the (possibly larger) set:
\[
\mathcal{T}:=[\sigma,\max\{\sqrt{\sigma^2+B_+^2},2\tau_*\}]\,
\]
where $\tau_*$ is as in the statement of the lemma.
The above lead to the following equivalent formulation of \eqref{eq:AO_2}:
\begin{align}
\left(\vct{w}_n,u_n,\tau_n\right) = \max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} + \min_{\vct{w}} \left\{ \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + \frac{u}{2\tau}\tn{\vct{w}}^2 - u \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} \right\} .\label{eq:AO_3}
\end{align}
\ct{To be fully rigorous, need to show here that the unconstrained min over $\vct{w}$ is the same as the constrained $\vct{w}\in\mathcal{B}$.}
The minimization over $\vct{w}$ is easy as it involves a strongly convex quadratic function. The optimal $\vct{w}':=\vct{w}'(\tau,u)$ (for fixed $(\tau,u)$) is given by
\begin{align}\label{eq:w'}
\vct{w}':=\vct{w}'(\tau,u) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right),
\end{align}
and \eqref{eq:AO_3} simplifies to
\begin{align}
\left(u_n,\tau_n\right)=\max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} - \frac{1}{2} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)\,=:\mathcal{R}(u,\tau) .\label{eq:AO_4}
\end{align}
It can be checked by direct differentiation and the second-derivative test that the objective function in \eqref{eq:AO_4} is strictly convex in $\tau$ and strictly concave in $u$ in the domain $\{(u,\tau)\in\mathbb{R}_+\times\mathbb{R}_+\}$. Thus, the saddle point $(u_n,\tau_n)$ is unique. Specifically, this implies that the optimal $\vct{w}_n$ in \eqref{eq:AO_3} is given by (cf. \eqref{eq:w'})
\begin{align}\label{eq:w_n}
\vct{w}_n=\vct{w}'(\tau_n,u_n) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).
\end{align}
In what follows, we characterize the high-dimensional limit of the optimal pair $(u_n,\tau_n)$ in the limit $n,p\rightarrow\infty,~p/n\rightarrow\kappa$.
We start by analyzing the (point-wise) convergence of $\mathcal{R}(u,\tau)$.
For the first three summands in \eqref{eq:AO_4}, we easily find that
$$
\left\{\frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}}\right\}~ \stackrel{{P}}{\longrightarrow}~\left\{
\frac{u\tau}{2} + \frac{u\sigma^2}{2\tau} \right\}.
$$
Next, we study the fourth summand. First, note that
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}(u\sqrt{\kappa}\bar{\h}) &= u^2\kappa\,\frac{1}{p}\vct{h}^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\vct{h} \notag\\
&= u^2\kappa\,\frac{1}{p}\sum_{i=1}^{p}\frac{\vct{h}_i^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}} \notag\\
&\stackrel{{P}}{\longrightarrow} u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right].
\end{align}
In the last line, $\Lambda$ is a random variable as in Definition \ref{def:Xi}. \cts{Also, we used Assumption (A2) together with the facts that $\vct{h}$ is independent of $\boldsymbol{\Sigma}$ and that the function $(x_1,x_2)\mapsto f(x_1,x_2)=x_1^2(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(2)$ assuming $x_2$ is bounded.} \ct{Question: Is it immediate that the empirical distribution of $(\beta,\boldsymbol{\Sigma},\vct{h})$ converges in $W_k$ to $\mu\otimes\mathcal{N}(0,1)$ given that $(\beta,\boldsymbol{\Sigma})$ converges to $\mu$ and $\vct{h}$ is independent???}
Second, we find that
\begin{align}
(\betab^\star)^T\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star &= \frac{1}{p}(\sqrt{p}\betab^\star)^T\left({\mtx{I}}+\frac{u}{\tau}\boldsymbol{\Sigma}\right)^{-1}(\sqrt{p}\betab^\star) \notag\\
&= \frac{1}{p}\sum_{i=1}^{p}\frac{\left(\sqrt{p}\betab^\star_i/\sqrt{\bSi_{i,i}}\right)^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}}\notag\\
&\stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right].
\end{align}
Here, $\Lambda,B$ are random variables as in Definition \ref{def:Xi} and \cts{we also used Assumption (A2) together with the fact that the function $(x_1,x_2)\mapsto f(x_1,x_2)=x_1^2x_2^{-1}(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(2)$ assuming $x_2$ is bounded.}
Third, \cts{by independence of $(\betab^\star, \boldsymbol{\Sigma})$ from $\vct{h}$}
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star = u\sqrt{\kappa} \cdot \frac{1}{p}\sum_{i=1}
^p \frac{\vct{h}_i\boldsymbol{\Sigma}_{i,i}^{-1/2}(\sqrt{p}\betab^\star_i)}{\boldsymbol{\Sigma}_{i,i}^{-1}+\frac{u}{\tau}{\mtx{I}}} ~\stackrel{{P}}{\longrightarrow}~ 0.
\end{align}
Putting these together, the objective $\mathcal{R}(u,\tau)$ in \eqref{eq:AO_4} converges point-wise in $u,\tau$ to
\begin{align}
\mathcal{R}(u,\tau)\stackrel{{P}}{\longrightarrow}\mathcal{D}(u,\tau) := \frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).\label{eq:conv_pt}
\end{align}
Note that $\mathcal{R}(u,\tau)$ (and thus, $\mathcal{D}(u,\tau)$) is convex in $\tau$ and concave in $u$. Thus, the convergence in \eqref{eq:conv_pt} is in fact uniform (e.g., \cite{AG1982}) and we can conclude that
\begin{align}\label{eq:Dc0}
\phi({\vct{g}},\vct{h}) \stackrel{{P}}{\longrightarrow} \max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
and
\begin{align}\label{eq:Dc}
(u_n,\tau_n) \stackrel{{P}}{\longrightarrow} (u_*,\tau_*):=\arg\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
In the proof of statement (iii) below, we show that the saddle point of \eqref{eq:Dc0} is $(u_*,\tau_*)$. In particular, $\tau_*$ is strictly in the interior of $\mathcal{T}$, which combined with convexity implies that
$$
\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau) = \max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau) =: \bar\phi.
$$
This, together with the first display above proves the second statement of the lemma.
\vspace{5pt}
\noindent\underline{Proof of (iii):}
Next, we compute the saddle point $(u_*,\tau_*)$ by studying the first-order optimality conditions of the srtictly concave-convex $\mathcal{D}(u,\tau)$. Specifically, we consider unconstrained minimization over $\tau$ and we will show that the minimum is achieved in the strict interior of $\mathcal{T}$. Direct differentiation gives
\begin{subequations}
\begin{align}
{\tau} + \frac{\sigma^2}{\tau} - 2u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] + \frac{u^2}{\tau}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] + \frac{1}{\tau}\operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0, \label{eq:fo1}\\
{u} - \frac{u\sigma^2}{\tau^2} - \frac{u^3}{\tau^2}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} + \frac{u}{\tau}\right)^2}\right] - \frac{u}{\tau^2} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0,\label{eq:fo2}
\end{align}
\end{subequations}
Multiplying \eqref{eq:fo2} with $\frac{\tau}{u}$ and adding to \eqref{eq:fo1} results in the following equation
\begin{align}
\tau = u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] ~\Leftrightarrow ~
\operatorname{\mathbb{E}}\left[ \frac{1}{(\frac{u}{\tau}\Lambda)^{-1}+1} \right] = \frac{1}{\kappa} \label{eq:tauu}\,.
\end{align}
Thus, we have found that the ratio $\frac{u_*}{\tau_*}$ is the unique solution to the equation in \eqref{eq:tauu}. Note that this coincides with the Equation \eqref{eq:ksi} that defines the parameter $\xi$ in Definition \ref{def:Xi}. The fact that \eqref{eq:tauu} has a unique solution for all $\kappa>1$ can be easily seen as $F(x)=\operatorname{\mathbb{E}}\left[ \frac{1}{(x\Lambda)^{-1}+1} \right], x\in\mathbb{R}_+$ has range $(0,1)$ and is strictly increasing (by differentiation).
Thus, we call $\xi=\frac{u_*}{\tau_*}$. Moreover, multiplying \label{eq:fo2} with $u$ leads to the following equation for $\tau_*$:
\begin{align}
u_*^2 = \sigma^2\xi^2 + u_*^2 \xi^2 \kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right] + \xi^2 \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right] ~\Rightarrow~ \tau_*^2 = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\xi^2\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right]} = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left((\xi\Lambda)^{-1} +1\right)^2}\right]}.
\end{align}
{Again, note that this coincides with Equation \eqref{eq:gamma} that determines the parameter $\gamma$ in Definition \ref{def:Xi}, i.e., $\tau_*^2 = \gamma.$ }
\vspace{3pt}\noindent\underline{Proof of (iv):}
For convenience, define
$$F_n(\hat\boldsymbol{\beta},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}\hat\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Recall from \eqref{eq:w_n} the explicit expression for $\vct{w}_n$, repeated here for convenience.
\begin{align}
\vct{w}_n = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).\notag
\end{align}
Also, recall that $\boldsymbol{\beta}_n = \boldsymbol{\Sigma}^{-1/2}\vct{w}_n+\betab^\star$. Thus, (and using the fact that $\bar{\h}$ is distributed as $-\bar{\h}$),
\begin{align}
\boldsymbol{\beta}_n &=\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}u_n\sqrt{\kappa}\bar{\h} + \left({\mtx{I}}-\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\right)\betab^\star\notag\\
\Longrightarrow\boldsymbol{\beta}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\bar{\h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i.\label{eq:betan}
\end{align}
For $i\in[p]$, define
\begin{align}
\vct{v}_{n,i} = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\bar{\h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i
\end{align}
In the above, for convenience, we have denoted $\xi_n:=u_n/\tau_n$ and recall that $\xi_*:=u_*/\tau_*$.
The proof proceeds in two steps. In the first step, we use the fact that $\xi_n\stackrel{{P}}{\longrightarrow}\xi_*$ and $u_n\stackrel{{P}}{\longrightarrow} u_\star$ (see \eqref{eq:Dc}) to prove that for any $\varepsilon\in(0,\xi_*/2)$, there exists absolute constant $C>0$ such that wpa 1:
\begin{align}\label{eq:ivstep1}
|F_n(\sqrt{p}\boldsymbol{\beta}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.
\end{align}
In the second step, we use Lipschitzness of $f$ and Assumption \ref{ass:mu} to prove that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:ivstep2}
\end{align}
The desired follows by combining \eqref{eq:ivstep1} and \eqref{eq:ivstep2}. Thus, in what follows, we prove \eqref{eq:ivstep1} and \eqref{eq:ivstep2}.
\vspace{2pt}
\noindent\emph{Proof of \eqref{eq:ivstep1}.}~~Fix some $\varepsilon\in(0,\xi_*/2)$. From \eqref{eq:Dc}, we know that w.p.a. 1 $|\xi_n-\xi_*|\leq \varepsilon$ and $|u_n-u_*|\leq \varepsilon$. Thus, $\vct{w}_n$ is close to $\vct{v}_n$. Specifically, in this event, for every $i\in[p]$, it holds that:
\begin{align}
\notag|\boldsymbol{\beta}_{n,i} - \vct{v}_{n,i}| &\leq {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{\tau_n}{1+(\xi_n\bSi_{i,i})^{-1}}-\frac{\tau_*}{1+(\xi_*\bSi_{i,i})^{-1}}\right| \\
\notag&= {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{u_n}{\xi_n+\bSi_{i,i}^{-1}}-\frac{u_*}{\xi_*+\bSi_{i,i}^{-1}}\right| \\
\notag& \leq
{|\betab^\star_i|}\frac{|\bSi_{i,i}||\xi_n-\xi_*|}{|1+\xi_n\bSi_{i,i}||1+\xi_*\bSi_{i,i}|} + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}} \frac{u_*|\xi_n-\xi_*|}{(\xi_n+\bSi_{i,i}^{-1})(\xi_*+\bSi_{i,i}^{-1})} +
\sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}}\frac{|u_n-u_*|}{\xi+\bSi_{i,i}^{-1}}
\\
\notag& \leq
{|\betab^\star_i|}{\Sigma_{\max}\varepsilon} + \sqrt{\kappa}{|\bar{\h}_i|} u_*\Sigma_{\max}^{3/2} \varepsilon+
\sqrt{\kappa}|\bar{\h}_i|\Sigma_{\max}^{1/2}\varepsilon
\\
&\leq \varepsilon \cdot \max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\, \left(|\bar{\h}_i| + |\betab^\star_i|\right).\label{eq:betav}
\end{align}\som{$\varepsilon$ missing}
In the second line above, we recalled that $u_n=\tau_n\xi_n$ and $u_*=\tau_*\xi_*$. In the third line, we used the triangle inequality. In the fourth line, we used that $\xi_*>0$, $0<\bSi_{i,i}\leq\Sigma_{\max}$ and $\xi_n\geq \xi_*-\varepsilon \geq \xi_*/2 >0$.
Now, we will use this and Lipschitzness of $f$ to argue that there exists absolute constant $C>0$ such that wpa 1:
\begin{align}
|F_n(\sqrt{p}\boldsymbol{\beta}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.\notag
\end{align}
Denote, $\vct{a}_i=(\sqrt{p}\boldsymbol{\beta}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\vct{v}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$. Following the exact same argument as in \eqref{eq:LipC}, we have
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| &\leq \frac{L}{p}\sum_{i=1}^p (1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\})\|\vct{a}_i-\vct{b}_i\|_2\notag \\
&\leq \frac{L}{p}\sum_{i=1}^p \left(1+\max\{\|\vct{a}_i\|_2^{k-1},\|\vct{b}_i\|_2^{k-1}\}\right)\cdot|\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}) - \sqrt{p}\vct{v}_{n,i}|\notag\\
&\leq L\left(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} \right)^{1/2}{\|\boldsymbol{\beta}_{n}({\vct{g}},\vct{h}) - \vct{v}_n\|_2}\notag.
\end{align}
From this and \eqref{eq:betav}, we find
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| &\leq
L(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2k-2}\} )^{1/2}
\\
&\qquad\qquad\varepsilon\cdot\sqrt{2}\max\left\{\Sigma_{\max}^{3/2},\Sigma_{\max}^{1/2}\right\}\sqrt{\|\betab^\star\|_2^2 + \|\bar{\h}\|_2^2}.\label{eq:epsS}
\end{align}
As in \eqref{eq:bdmom}, it can be shown that wpa 1: $\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2k-2}<\infty$ and $\frac{1}{p}\sum_{i=1}^p \|\vct{b}_i\|_2^{2k-2}<\infty$, as $p\rightarrow\infty$. Similarly, $\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^2<\infty$, as $p\rightarrow \infty$ by Assumption (A2).
Finally, $\|\bar{\h}\|_2^2\leq 2$, wpa 1 as $p\rightarrow\infty$. Therefore, from \eqref{eq:epsS}, wpa 1, there exists constant $C>0$ such that
\begin{align}
|F_n(\boldsymbol{\beta}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \leq C \cdot\varepsilon,
\end{align}
as desired.
\vspace{2pt}
\noindent\emph{Proof of \eqref{eq:ivstep2}.}~~Next, we will use Assumption \ref{ass:mu} to show that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:AO_conv}
\end{align}
Notice that $\vct{v}_n$ is a function of $\betab^\star,\boldsymbol{\Sigma},\bar{\h}$. Concretely, define $g:\mathbb{R}^3\rightarrow\mathbb{R}$, such that
$$
g(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1,
$$
and notice that
$$
\sqrt{p}\vct{v}_{n,i} = g\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right) = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\vct{h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\sqrt{p}\betab^\star_i.
$$
Thus,
$$
F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma}) = \frac{1}{p}\sum_{i=1}^p f\left(g\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right),\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) =: \frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right),
$$
where we have defined $h:\mathbb{R}^3\rightarrow\mathbb{R}$:
\begin{align}
h(x_1,x_2,x_3) := f\left(g(x_2,x_3,x_1),x_2,x_3\right).\label{h func}
\end{align}
\cts{It will suffice to prove that $h\in\rm{PL}(k+1)$. Indeed, if that were the case, then Assumption (A2) (note that $2k-2\geq k+1$ for $k\geq3$) gives}
\begin{align}
\frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[h(H,B,\Lambda)\right] &= \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[f\left(g(B,\Lambda,H),B,\Lambda\right))\right]\\
& = \operatorname{\mathbb{E}}[f\left(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda\right)] = \alpha_*,
\end{align}
where the penultimate equality follows by recognizing that (cf. Eqn. \eqref{eq:X})
$$
g(B,\Lambda,H) = (1-(1+\xi_*\Lambda)^{-1})B + \sqrt{\kappa}\frac{\tau_*\Lambda^{-1/2}}{1+(\xi_*\Lambda)^{-1}}H = X_{\kappa,\sigma^2}(\Lambda,B,H).
$$
We prove that $h\in\rm{PL}(k+1)$ in Lemma \ref{lem:hPL} in Section \ref{SM useful fact}.
\vspace{5pt}
\vspace{3pt}\noindent\underline{Proof of (v):} Let $\psi:\mathbb{R}\rightarrow\mathbb{R}$ be any bounded Lipschitz function. The function $f(a,b,c) = \psi(a)$ is $\rm{PL}$ of order $k$. Thus, by directly applying statement (iv) of the lemma, we find that
$$
\frac{1}{p}\sum_{i=1}^p{\psi(\sqrt{p}\boldsymbol{\beta}_{n,i}({\vct{g}},\vct{h}))} \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\psi(X_{\kappa,\sigma^2})\right].
$$
Since this holds for any bounded Lipschitz function, we have shown that the empirical convergence of $\boldsymbol{\beta}_n$ converges to the distribution of $X_{\kappa,\sigma^2}$. It remains to prove boundedness of the $k$-th moment as advertised in \eqref{eq:k_AO}. Recall from
\eqref{eq:betan} that
\begin{align}
\sqrt{p}\boldsymbol{\beta}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\vct{h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)(\sqrt{p}\betab^\star_i).\notag
\end{align}
Using this, boundedness of $\boldsymbol{\Sigma}_{i,i}$ from Assumption (A2), and the fact that $\tau_n\stackrel{{P}}{\longrightarrow}\tau_\star, \xi_n\stackrel{{P}}{\longrightarrow}\xi_\star$, there exists constant $C=C(\Sigma_{\max},\Sigma_{\min},k,\tau_\star,\xi_\star)$ such that wpa. 1,
\begin{align}
\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\boldsymbol{\beta}_{n,i}|^{2k-2} \leq C\left(\frac{1}{p}\sum_{i=1}^p|\vct{h}_{i}|^{2k-2}+\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\betab^\star_{i}|^{2k-2}\right).\notag
\end{align}
But the two summands in the expression above are finite in the limit of $p\rightarrow\infty$. Specifically, (i) from Assumption (A2), $\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\betab^\star_{i}|^{2k-2}\stackrel{{P}}{\longrightarrow}\operatorname{\mathbb{E}}[B^{2k-2}]<\infty$; (ii) $\frac{1}{p}\sum_{i=1}^p|\vct{h}_{i}|^{2k-2}\stackrel{{P}}{\longrightarrow}\operatorname{\mathbb{E}}[H^{2k-2}]<\infty$, using the facts that $\vct{h}_i\stackrel{iid}{\sim}\mathcal{N}(0,1)$ and $H\sim\mathcal{N}(0,1)$. This proves \eqref{eq:k_AO}, as desired.
\begin{lemma} The function $h$ in \eqref{h func} is $\rm{PL}(k)$ as long as $f$ is $\rm{PL}(k)$.
{\color{red}The function $h$ is $\mathbb{R}^3\rightarrow\mathbb{R}$, not $\mathbb{R}^{3p}\rightarrow\mathbb{R}$! See Lemma 6 below.}
\end{lemma}
\begin{proof}
Since $f$ is $\rm{PL}(k)$, for some $L>0$, \eqref{PL func} holds. Fix $\vct{a}=[\vct{x},\vct{y},{\vct{z}}]\in\mathbb{R}^{3p},\vct{a}'=[\vct{x}',\vct{y}',{\vct{z}}']\in\mathbb{R}^{3p}$. Let $\vct{b}=[{\vct{g}},\vct{y},{\vct{z}}]$ where ${\vct{g}}=g(\vct{y},{\vct{z}},\vct{x})$. We have that
\begin{align}
|h(\vct{a})-h(\vct{a}')|&\leq |f(\vct{b})-f(\vct{b}')|\\
&\leq L\left(1+\|\vct{b}\|_2^{k-1}+\|\vct{b}'\|_2^{k-1}\right)\|\vct{b}-\vct{b}'\|_2\\
&\lesssim L\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+\|{\vct{g}}\|_2^{k-1}+\|{\vct{g}}'\|_2^{k-1}\right)(\|\vct{a}-\vct{a}'\|_2+\|{\vct{g}}-{\vct{g}}'\|_2).
\end{align}
Next, we need to bound the ${\vct{g}}$ term in terms of $\vct{a}$. This is accomplished as follows
\begin{align}
\|{\vct{g}}\|_2^{k-1}&\lesssim \frac{1}{p}\sum_{i=1}^p|g_i|^{k-1}\\
&\lesssim \frac{1}{p}\sum_{i=1}^p\left|\frac{y_i^{-1/2}}{1+(\xi y_i)^{-1}}\sqrt{\kappa}\tau_* z_i + (1-(1+\xi_*y_i)^{-1})x_i\right|\\
&\lesssim \frac{1}{p}\sum_{i=1}^p(|z_i|+|x_i|)^{k-1}\lesssim \|\vct{x}\|_2^{k-1}+\|{\vct{z}}\|_2^{k-1}\\
&\lesssim \|\vct{a}\|^{k-1}.
\end{align}
Secondly and similarly, we have the following perturbation bound on the ${\vct{g}}-{\vct{g}}'$. \so{Suppose the triples $(x_i,y_i,z_i)$ takes values in a fixed bounded compact set $\mathcal{M}$. We will prove the following sequence of inequalities}
\begin{align}
\tn{{\vct{g}}-{\vct{g}}'}^2&\leq \sum_{i=1}^p|g(x_i,y_i,z_i)-g(x'_i,y'_i,z'_i)|^2\\
&\leq C^2_{\mathcal{M}} \sum_{i=1}^p(|x_i-x'_i|^2+|y_i-y'_i|^2+|z_i-z'_i|^2)\label{sec line eq}\\
&\leq C^2_{\mathcal{M}} (\tn{\vct{x}-\vct{x}'}^2+\tn{\vct{y}-\vct{y}'}^2+\tn{{\vct{z}}-{\vct{z}}'}^2)\\
&\lesssim \tn{\vct{a}-\vct{a}'}^2.
\end{align}
In what follows, we prove the second line \eqref{sec line eq} i.e.~the fact that for any triples $a=(x,y,z),a'=(x',y',z')$ (with $a,a'\in\mathbb{R}^3$), we have that
\[
|g(a)-g(a')|\leq C_{\mathcal{M}}\tn{a-a'}.
\]
Set $C_\nabla=\sup_{a\in \mathcal{M}}\tn{\nabla g(a)}$. By definition of gradient, the inequality above holds with $C_\mathcal{M}=C_\nabla$. Thus, all that remains is proving that $C_\nabla$ is upper bounded by a constant. \so{However, this automatically holds because from the definition of $g(\dots)$ function, it is clear that $C_\nabla$ is defined everywhere and continuous thus it has a finite maximum over a compact set.} \cts{Only the problem is that $z_i$ (i.e., $\vct{h}_i$) is $\mathcal{N}(0,1)$ so it is not bounded.}
\end{proof}
But from Assumption \ref{ass:mu} and $f\in\rm{PL(2)}$ it follows immediately that with probability approaching 1,
\begin{align}
|\frac{1}{p}\sum_{i=1}^{p}f(\sqrt{p}\vct{v}_{n,i})-\operatorname{\mathbb{E}}_\mu\left[X_{\kappa,\sigma^2}(\Lambda,N,H) \right]| \leq \varepsilon.
\end{align}
Combining the above displays and using triangle inequality completes the proof of the statement.
\subsection{Asymptotic formulas on Magnitude- and Hessian- pruning}
Here, we use Theorem \ref{thm:master_W2} to characterize the risk of the magnitude- and Hessian- pruned solutions. This section supplements the discussion in Section \ref{sec:risk}. First, we formally state the result of Section \ref{sec:risk} as Corollary \ref{cor:mag} characterizing magnitude pruning. Next, we further discuss Hessian pruning.
\subsubsection{Magnitude-based pruning.}
We begin with the following necessary definitions. Define the hard-thresholding function with fixed threshold $t\in\mathbb{R}_+$ as follows:
\begin{align}\label{eq:threshold_app}
\mathcal{T}_t(x) = \begin{cases} x & \text{if } |x|>t \\ 0 & \text{otherwise} \end{cases}.
\end{align}
Further define threshold $t^\star$ as follows:
\begin{align}\label{eq:tstar_app}
t^\star:=\inf\left\{t\in\mathbb{R}\,:\, \Pr(|X_{\kappa,\sigma^2}|\geq t) \geq \alpha \right\}.
\end{align}
\begin{corollary}[Risk of Magnitude-pruning]\label{cor:mag}
Let the same assumptions and notation as in the statement of Theorem \ref{thm:master_W2} hold. Specifically, let $\hat\boldsymbol{\beta}$ be the min-norm solution in \eqref{eq:min_norm} and $\hat\beta_s^M:=\mathbb{T}_s({\boldsymbol{\beta}})$ the magnitude-pruned model at sparsity $s$. Recall the notations in \eqref{eq:threshold_app} and \eqref{eq:tstar_app}. The risk of the magnitude-pruned model satisfies the following in the limit of $n,p,s\rightarrow\infty$ at rates $\kappa:=p/n>1$ and $\alpha:=s/p\in(0,1)$ (cf. Assumption \ref{ass:linear}):
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^M_s) \stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}\left[ \Lambda\left(B-\mathcal{T}_{t^\star}(X_{\kappa,\sigma^2})\right) \right],\notag
\end{align}
where the expectation is over $(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1).$
\end{corollary}
The proof of the corollary above, is given in Section \ref{sec:risk}. Below, we extend the results to Hessian-based pruning.
\subsubsection{Hessian-based pruning.}
Let $\hat\beta$ be the min-norm solution in \eqref{eq:min_norm}. Recall that the Hessian-pruned model $\boldsymbol{\beta}_s^H$ at sparsity $s$ is given by
\begin{align}\label{eq:hp}
\boldsymbol{\beta}_s^H = \hat\boldsymbol{\Sigma}^{-1/2}\mathbb{T}_s(\hat\boldsymbol{\Sigma}^{1/2}\boldsymbol{\beta}),
\end{align}
where $\hat\boldsymbol{\Sigma}=\diag{{\mtx{X}}^T{\mtx{X}}}/n$ the diagonal of the empirical covariance matrix.
\begin{corollary}[Risk of Hessian-pruning]\label{cor:hes}
Let the same assumptions and notation as in the statement of Theorem \ref{thm:master_W2} hold. Specifically, let $\hat\boldsymbol{\beta}$ be the min-norm solution in \eqref{eq:min_norm} and $\hat\beta_s^H$ the Hessian-pruned model at sparsity $s$. Recall the notation in \eqref{eq:threshold_app} and define
\begin{align}\label{eq:tdiam_app}
t^\diamond:=\inf\left\{t\in\mathbb{R}\,:\, \Pr(|\Lambda^{1/2}X_{\kappa,\sigma^2}|\geq t) \geq \alpha \right\}.
\end{align}
The risk of the magnitude-pruned model satisfies the following in the limit of $n,p,s\rightarrow\infty$ at rates $\kappa:=p/n>1$ and $\alpha:=s/p\in(0,1)$ (cf. Assumption \ref{ass:linear}):
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^H_s) \stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}\left[ \Lambda\left(B-\Lambda^{-1/2}\mathcal{T}_{t^\diamond}(\Lambda^{1/2}X_{\kappa,\sigma^2})\right) \right],\notag
\end{align}
where the expectation is over $(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1).$
\end{corollary}
\begin{proof}
Recall the definition of the hard-thresholding operator $\mathcal{T}_t(x)$. Similar to Section \ref{sec:risk}, we consider a threshold-based pruning vector $${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H}_{t} := \hat\boldsymbol{\Sigma}^{-1/2}\mathcal{T}_{t/\sqrt{p}}(\hat\boldsymbol{\Sigma}^{1/2}\hat{\boldsymbol{\beta}}),$$ where $\mathcal{T}_t$ acts component-wise. Further define
$${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_{t} := \boldsymbol{\Sigma}^{-1/2}\mathcal{T}_{t/\sqrt{p}}(\boldsymbol{\Sigma}^{1/2}\hat{\boldsymbol{\beta}}).$$
Note that ${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_{t}$ uses the true (diagonal) covariance matrix $\boldsymbol{\Sigma}$ instead of its sample estimate $\hat\boldsymbol{\Sigma}$. For later reference, note here that $\hat\boldsymbol{\Sigma}$ concentrates (entry-wise) to $\boldsymbol{\Sigma}$. Specifically, using boundedness of $\boldsymbol{\Sigma}$ and standard concentration of sub-exponential random variables
First, we compute the limiting risk of ${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_{t}$. Then, we will use the fact that $\hat\boldsymbol{\Sigma}$ concentrates (entry-wise) to $\boldsymbol{\Sigma}$, to show that the risks of ${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_{t}$ and ${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H}_{t}$ are arbitrarily close as $p\rightarrow\infty$.
We start by computing the risk of ${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_{t}$. Similar to \eqref{eq:loss_w2},
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_t) &= \sigma^2 + ({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_t)^T\boldsymbol{\Sigma}({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_t) \notag \\
&= \sigma^2 + \frac{1}{p}\sum_{i=1}^p\boldsymbol{\Sigma}_{i,i}\big(\sqrt{p}\betab^\star_i-\bSi_{i,i}^{-1/2}\mathcal{T}_{t}(\bSi_{i,i}^{1/2}\sqrt{p}\hat\boldsymbol{\beta}_i)\big)^2.
\end{align}
where in the last line we used that $\sqrt{p}\mathcal{T}_{t/\sqrt{p}}(x)=\mathcal{T}_{t}(\sqrt{p}x)$.
\end{proof}
\section{Facts about Lipschitz functions}\label{SM useful fact}
For $k\geq 1$ we say a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$ and denote it by $f\in \rm{PL}(k)$ if there exists a cosntant $L>0$ such that, for all $\vct{x},\vct{y}\in\mathbb{R}^m$:
\begin{align}
|f(\vct{x})-f(\vct{y})|\leq L\left(1+\|\vct{x}\|_2^{k-1}+\|\vct{y}\|_2^{k-1}\right)\|\vct{x}-\vct{y}\|_2.\label{PL func}
\end{align}
In particular, when $f\in\rm{PL}(k)$, the following properties hold:
\begin{enumerate}
\item There exists a constant $L'$ such that for all $\vct{x}\in\mathbb{R}^n$: $|f(\vct{x})|\leq L'(1+\|\vct{x}\|_2^k).$
\item $f$ is locally Lipschitz, that is for any $M>0$, there exists a constant $L_{M,m}<\infty$ such that for all $x,y\in[-M,M]^m$,
$
|f(\vct{x})-f(\vct{y})| \leq L_{M,m}\|\vct{x}-\vct{y}\|_2.
$
Further, $L_{M,m}\leq c(1+(M\sqrt{m})^{k-1})$ for some costant $c$.
\end{enumerate}
\begin{lemma}
Let $g:\mathbb{R}\rightarrow\mathbb{R}$ be a Lipschitz function. Consider the function $f:\mathbb{R}^3\rightarrow\mathbb{R}$ defined as follows:
$$
f(\vct{x}) = x_1(x_2-g(x_3))^2.
$$
Then, $f\in\rm{PL}(3).$
\end{lemma}
\begin{proof}
Let $h:\mathbb{R}^2\rightarrow\mathbb{R}$ defined as $h({\vct{u}})=({\vct{u}}_1-g({\vct{u}}_2))^2$. The function $({\vct{u}}_1,{\vct{u}}_2)\mapsto{\vct{u}}_1-g({\vct{u}}_2)$ is clearly Lipschitz. Thus, $h\in\rm{PL}(2)$, i.e.,
\begin{align}
|h({\vct{u}})-h(\vct{v})| \leq C(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2\quad\text{and}\quad |h(\vct{v})|\leq C'(1+\|\vct{v}\|_2^2).\label{eq:h_pl}
\end{align}
Therefore, letting $\vct{x}=(x_1,{\vct{u}})\in\mathbb{R}^3$ and $\vct{y}=(y_1,\vct{v})\in\mathbb{R}^3$, we have that
\begin{align}
|f(\vct{x})-f(\vct{y})| &= |x_1h({\vct{u}}) - y_1h(\vct{v})| \leq |x_1||h({\vct{u}})-h(\vct{v})| + |h(\vct{v})| |x_1-y_1|\notag\\
&\leq C|x_1|(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(|x_1|^2+(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)|x_1-y_1| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|\vct{x}-\vct{y}\|_2.
\end{align}
In the second line, we used \eqref{eq:h_pl}. In the third line, we used $2xy\leq x^2+y^2$. In the fourth line, we used Cauchy-Schwarz inequality. $C,C'>0$ are absolute constants that may change from line to line.
This completes the proof of the lemma.
\end{proof}
\begin{lemma}\label{lem:hPL}
Let functions $f,g:\mathbb{R}^3\rightarrow\mathbb{R}$ such that $f\in\rm{PL}(k)$ and
$$
g(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi_* x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1.
$$
Here, $\xi_*,\tau_*,\kappa$ are positive constants.
Further define
\begin{align}
h(x,y,z) := f\left(g(y,z,x),y,z\right),
\end{align}
and assume that $y$ take values on a fixed bounded compact set $\mathcal{M}$.
Then, it holds that $h\in\rm{PL}(k+1)$.
\end{lemma}
\begin{proof}
Since $f$ is $\rm{PL}(k)$, for some $L>0$, \eqref{PL func} holds. Fix $x,x'\in\mathbb{R}$, $\vct{a}=[y,z]\in\mathbb{R}^{2}$ and $\vct{a}'=[y',z']\in\mathbb{R}^{2}$. Let $\vct{b}=[{\vct{g}},\vct{a}]=[{\vct{g}},y,z]\in\mathbb{R}^3$ where ${\vct{g}}=g(y,z,x)\in\mathbb{R}$ and define accordingly $\vct{b}'$ and ${\vct{g}}'$. We have that
\begin{align}
|h([x,\vct{a}])-h([x,\vct{a}'])|&= |f(\vct{b})-f(\vct{b}')|\notag\\
&\leq L\left(1+\|\vct{b}\|_2^{k-1}+\|\vct{b}'\|_2^{k-1}\right)\|\vct{b}-\vct{b}'\|_2\notag\\
&\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|{\vct{g}}|^{k-1}+|{\vct{g}}'|^{k-1}\right)(\|\vct{a}-\vct{a}'\|_2+|{\vct{g}}-{\vct{g}}'|),\label{eq:hlip}
\end{align}
for some constant $C>0$. In the last inequality we have repeatedly used the inequality
$
\left(\sum_{i=1}^m \|\vct{v}_i\|_2^2\right)^{\frac{d}{2}} \leq C(m)\cdot\sum_{i=1}^m\|\vct{v}_i\|_2^{d}.
$
Next, we need to bound the ${\vct{g}}$ term in terms of $\vct{a}$. This is accomplished as follows
\begin{align}
|{\vct{g}}|^{k-1} &= \left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x + (1-(1+\xi_*y)^{-1})z\right|^{k-1}\notag\\
&\leq C (|z|+|x|)^{k-1} \notag\\
&\leq C \left(|x|^{k-1}+|z|^{k-1}\right) \leq C (|x|^{k-1} + \|\vct{a}\|_2^{k-1}).\label{eq:gb}
\end{align}
Here, the value of the constant $C>0$ may change from line to line.
Secondly and similarly, we have the following perturbation bound on the ${\vct{g}}-{\vct{g}}'$. Recall that $(x,y)$ and $(x',y')$ are bounded. We will prove the following sequence of inequalities
\begin{align}
|{{\vct{g}}-{\vct{g}}'}|&= |g(y,z,x)-g(y',z',x')| \notag\\
&\leq \left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x - \frac{(y')^{-1/2}}{1+(\xi y')^{-1}}\sqrt{\kappa}\tau_* x'\right|
+ \left|(1-(1+\xi_*y)^{-1})z-(1-(1+\xi_*y')^{-1})z'\right|\notag\\
&\leq
\left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x - \frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x'\right| +\left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* x'- \frac{(y')^{-1/2}}{1+(\xi y')^{-1}}\sqrt{\kappa}\tau_* x'\right|\notag\\
&\qquad+ \left|(1-(1+\xi_*y)^{-1})z-(1-(1+\xi_*y)^{-1})z'\right| + \left|(1-(1+\xi_*y)^{-1})z'-(1-(1+\xi_*y')^{-1})z'\right| \notag\\
&\leq C_1|x-x'| + C_2|x'||y-y'| + C_3|z-z'| + C_4|z'||y-y'|\notag\\
&\leq C(1+|x'| + |z'|)(|x-x'| + |z-z'| + |y-y'|)\\
&\leq C\sqrt{3}(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2
.\label{eq:gd}
\end{align}
In the fourth inequality above, we used the fact the assumption that $|y|$ is bounded. In the last line, we used Cauchy-Scwhartz.
Substituting \eqref{eq:gb} and \eqref{eq:gd} in \eqref{eq:hlip} gives:
\begin{align}
|h(x,y,z) - h(x',y',z')| &\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|x|^{k-1}+|x'|^{k-1}\right)(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2\notag \\
&\leq C\left(1+\|[\vct{a},x]\|_2^{k}+\|[\vct{a}',x']\|_2^{k}\right)\|[\vct{a},x]-[\vct{a}',x']\|_2.
\end{align}
Thus, $h\in\rm{PL}(k+1)$, as desired.
\end{proof}
\section{Proofs for overparameterized least-squares}\label{sec proof thm 1}
In this section, we assume the linear Gaussian problem (LGP) of Definition \ref{def LGP}, the overparameterized regime $k=p>n$ and the min-norm model ${\boldsymbol{\hat{\beta}}}$ of \eqref{eq:min_norm}. We prove Theorem \ref{thm:master_W2} that derives the asymptotic DC of ${\boldsymbol{\hat{\beta}}}$ and we show how this leads to sharp formulae for the risk of the Magnitude- and Hessian-pruned models.
\subsection{Notation and Assumptions}\label{sec:ass_app}
For the reader's convenience, we recall some necessary notation and assumptions from Section \ref{sec main}. We say that a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$, denoted $f\in\rm{PL}(k)$, if there is a constant $L>0$ such that for all $\vct{x},\vct{y}\in\mathbb{R}^m$,
$
|f(\vct{x}) - f(\vct{y})|\leq L(1+\tn{\vct{x}}^{k-1}+\tn{\vct{y}}^{k-1})\|\vct{x}-\vct{y}\|_2
$
(See also Section \ref{SM useful fact}).
We say that a sequence of probability distributions $\nu_p$ on $\mathbb{R}^m$ {converges in $W_k$} to
$\nu$, written $\nu_p\stackrel{W_k}{\Longrightarrow} \nu$, if $W_k(\nu_p,\nu) \rightarrow 0$ as $p \rightarrow \infty$. An equivalent definition is that, for any $f\in\rm{PL}(k)$, $\lim_{p\rightarrow\infty}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$ (e.g., \cite{montanari2017estimation}).
Finally, recall that a sequence of probability distributions $\nu_n$ on $\mathbb{R}^m$ \emph{converges weakly} to $\nu$, if for any bounded Lipschitz function $f$: $\lim_{p\rightarrow\infty}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$.
Throughout, we use $C,C',c,c'$ to denote absolute constants (not depending on $n,p$) whose value might change from line to line.
We focus on a double asymptotic regime where:
$$n,p,s\rightarrow\infty \text{ at fixed overparameterization ratio } \kappa:=p/n>1 \text{ and sparsity level } \alpha:=s/p\in(0,1).$$
For a sequence of random variables $\mathcal{X}_{p}$ that converge in probability to some constant $c$ in the limit of Assumption \ref{ass:linear} below, we write $\mathcal{X}_{p}\stackrel{{P}}{\longrightarrow} c$. For a sequence of event $\mathcal{E}_p$ for which $\lim_{p\rightarrow}\mathbb{P}(\mathcal{E}_p) = 1$, we say that $\mathcal{E}_p$ occurs \emph{with probability approaching 1}. For this, we will often use the shorthand ``wpa 1".
\vspace{5pt}
Next, we recall the set of assumption under which our analysis applies:
\asstwo*
\assthree*
\noindent We remark that Assumption \ref{ass:mu} above implies (see \cite[Lem.~4]{bayati2011dynamics} and \cite[Lem.~A3]{javanmard2013state}) that for any pseudo-Lipschitz function $\psi:\mathbb{R}^2\rightarrow\mathbb{R}$ of order $4$, i.e., $\psi\in\rm{PL}(4)$:
$$
\frac{1}{p}\sum_{i=1}^p\psi(\bSi_{i,i},\sqrt{p}\betab^\star_i) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(\Lambda,B)\sim\mu}\left[ \psi(\Lambda,B)\right].
$$
\subsection{Asymptotic distribution and risk characterizations}
{In this section, we prove our main result Theorem \ref{thm:master_W2}. Recall that ${\boldsymbol{\hat{\beta}}}$ is the min-norm solution. Since the distribution of ${\boldsymbol{\hat{\beta}}}$ depends on the problem dimensions (as it is a function of ${\mtx{X}},\vct{y}$), when necessary, we will use ${\boldsymbol{\hat{\beta}}}_n$ notation to make its dimension dependence explicit.} Let ${\boldsymbol{\hat{\beta}}}^P={\cal{P}}({\boldsymbol{\hat{\beta}}})$ be a pruned version of the min-norm solution ${\boldsymbol{\hat{\beta}}}$. Recall from Section \ref{sec:risk}, that the first crucial step in characterizing the risk ${\cal{L}}({\boldsymbol{\hat{\beta}}}^P)$ is studying the risk ${\cal{L}}({\boldsymbol{\hat{\beta}}}^\mathcal{T}_t)$ of a threshold-based pruned vector.
To keep things slightly more general, consider ${\boldsymbol{\hat{\beta}}}^{g}$ defined such that $\sqrt{p}{\boldsymbol{\hat{\beta}}}^{g}=g(\sqrt{p}{\boldsymbol{\hat{\beta}}})$, where $g$ is a Lipschitz function acting entry-wise on ${\boldsymbol{\hat{\beta}}}$ (for example, $g$ can be the {(arbitrarily close Lipschitz approximation of the)} thresholding operator $\mathcal{T}_t$ of Section \ref{sec:risk}). Then, the risk of ${\boldsymbol{\hat{\beta}}}^g$ can be written as
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^g) &= \operatorname{\mathbb{E}}_{\mathcal{D}}[(\vct{x}^T({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^g) + \sigma z)^2] = \sigma^2 + ({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^g)^T\boldsymbol{\Sigma}({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^g) \notag \\
&= \sigma^2 + \frac{1}{p}\sum_{i=1}^p\boldsymbol{\Sigma}_{i,i}\big(\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\hat{\beta}}}_i)\big)^2\notag
\\
&=: \sigma^2 + \frac{1}{p}\sum_{i=1}^p f\big(\sqrt{p}{\boldsymbol{\hat{\beta}}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}\big),\label{eq:risk_app_f}
\end{align}
where in the last line, we defined the function $f$ as $f=f_{\cal{L}}\in {\cal{F}}_{\cal{L}}\subset{\cal{F}}$ given by
\[
f_{\cal{L}}(x,y,z) := z(y-g(x))^2\quad\text{where}~g~\text{is Lipschitz.}
\]
Here, recall the definition of the families ${\cal{F}}_{\cal{L}}$ in \eqref{eq:fdef} and ${\cal{F}}$ in \eqref{eq:pdef}.
The following theorem establishes the asymptotic limit of \eqref{eq:risk_app_f}. For the reader's convenience, we repeat the notation introduced in Definition \ref{def:Xi}.
Let random variables $(\Lambda,B)\sim \mu$ (where $\mu$ is defined in Assumption \ref{ass:mu}) and fix $\kappa>1$. Define parameter $\xi$ as the unique positive solution to the following equation
$$
\operatorname{\mathbb{E}}_{\mu}\Big[ \big({1+(\xi\cdot\Lambda)^{-1}}\big)^{-1} \Big] = {\kappa^{-1}}\,.
$$
Further define the positive parameter $\gamma$ as follows:
$$
\hspace{-0.1in}\gamma :=
\Big({\sigma^2 + \operatorname{\mathbb{E}}_{\mu}\Big[\frac{B^2\Lambda}{(1+\xi\Lambda)^2}\Big]}\Big)\Big/\Big({1-\kappa\operatorname{\mathbb{E}}_{\mu}\Big[\frac{1}{\left(1+(\xi\Lambda)^{-1}\right)^2}\Big]}\Big).
$$
With these and $H\sim\mathcal{N}(0,1)$, define the random variable
$$
X_{\kappa,\sigma^2}:=X_{\kappa,\sigma^2}(\Lambda,B,H) := \Big(1-\frac{1}{1+ \xi\Lambda}\Big) B + \sqrt{\kappa}\frac{\sqrt{\gamma}\,\Lambda^{-1/2}}{1+(\xi\Lambda)^{-1}} H,
$$
and let $\Pi_{\kappa,\sigma^2}$ be its distribution.
\mainthm*
Before we prove the theorem, let us show how it immediately leads to a sharp prediction of the risk behavior. Indeed, a direct application of \eqref{eq:thm} for $f=f_{\cal{L}}$ to \eqref{eq:risk_app_f} shows that
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^g)\stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}_{(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)}\left[f_{\cal{L}}(X_{\kappa,\sigma^2},B,\Lambda) \right] = \sigma^2 + \operatorname{\mathbb{E}}_{(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)}\left[\Sigma\left(B-g(X_{\kappa,\sigma^2})\right)^2 \right].\label{eq:risk_app_f2}
\end{align}
We further remark on the following two consequences of Theorem \ref{thm:master_W2}.
First, since \eqref{eq:thm} holds for any $\rm{PL}(2)$ function, we have essentially shown that $\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma})$ converges in Wasserstein-2 distance to $\Pi_{\kappa,\sigma^2}\otimes\mu$, where recall that $\Pi_{\kappa,\sigma^2}$ is the distribution of the random variable $X_{\kappa,\sigma^2}$.
Second, the theorem implies that the empirical distribution of $\sqrt{p}{\boldsymbol{\hat{\beta}}}_n$ converges weakly to $\Pi_{\kappa,\sigma^2}$. To see this, apply \eqref{eq:thm} for the $\rm{PL}(2)$ function $f(x,y,z) = \psi(x)$ where $\psi:\mathbb{R}\rightarrow\mathbb{R}$ is a bounded Lipschitz test function.
\subsection{Proof of Theorem \ref{thm:master_W2}}
\vspace{5pt}
Let ${\mtx{X}}\in\mathbb{R}^{n\times p}$ have zero-mean and normally distributed rows with a diagonal covariance matrix ${\boldsymbol{{\Sigma}}}=\operatorname{\mathbb{E}}[\vct{x}\x^T]$. Given a ground-truth vector $\betab^\star$ and labels $\vct{y}={\mtx{X}}\betab^\star+\sigma {\vct{z}},~{\vct{z}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, we consider the least-squares problem subject to the minimum Euclidian norm constraint (as $\kappa=p/n>1$) given by
\begin{align}\label{eq:PO_beta}
\min_{{\boldsymbol{\beta}}}\frac{1}{2}\tn{{\boldsymbol{\beta}}}^2\quad\text{subject to}\quad \vct{y}={\mtx{X}}{\boldsymbol{\beta}}.
\end{align}
It is more convenient to work with the following change of variable:
\begin{align}\label{eq:w}
\vct{w}:=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\beta}}-\betab^\star).
\end{align}
With this, the optimization problem in \eqref{eq:min_norm} can be rewritten as
\begin{align}\label{eq:PO}
\Phi({\mtx{X}})=\min_{\vct{w}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
where we set ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$. First, using standard arguments, we show that the solution of \eqref{eq:PO} is bounded. Hence, we can constraint the optimization in a sufficiently large compact set without loss of generality.
\begin{lemma}[Boundedness of the solution]\label{lem:bd_PO}
Let $\widehat{\vct{w}}_n:=\widehat{\vct{w}}_n({\mtx{X}},{\vct{z}})$ be the minimizer in \eqref{eq:PO}. Then, with probability approaching 1, it holds that $\widehat{\vct{w}}_n\in\mathcal{B}$, where
$$\mathcal{B}:=\left\{\vct{w}\,|\,\|\vct{w}\|_2\leq B_{+} \right\},\qquad B_+:=5\sqrt{\frac{\Sigma_{\max}}{\Sigma_{\min}}}\frac{\sqrt{\kappa}+1}{\sqrt{\kappa}-1}(\sqrt{\Sigma_{\max}\operatorname{\mathbb{E}}\left[B^2\right]} + \sigma).
$$
\end{lemma}
\begin{proof}
First, we show that the min-norm solution ${\boldsymbol{\hat{\beta}}}={\mtx{X}}^T({\mtx{X}}\X^T)^{-1}\vct{y}$ of \eqref{eq:PO_beta} is bounded. Note that $\kappa>1$, thus ${\mtx{X}}\X^T$ is invertible wpa 1. We
have,
\begin{align}
\tn{{\boldsymbol{\hat{\beta}}}_n}^2 = \vct{y}^T({\mtx{X}}\X^T)^{-1}\vct{y} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{X}}\X^T)} = \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\boldsymbol{\Sigma}{\mtx{\bar{X}}}^T)} \leq \frac{\tn{\vct{y}}^2}{\lambda_{\min}({\mtx{\bar{X}}}\Xb^T)\,\Sigma_{\min}} = \frac{\tn{\vct{y}}^2}{\sigma_{\min}^2({\mtx{\bar{X}}})\,\Sigma_{\min}}. \label{eq:Ubb}
\end{align}
But, wpa 1,
$
\sigma_{\min}({\mtx{\bar{X}}})/\sqrt{n} \geq \frac{1}{2}\left(\sqrt{\kappa}-1\right).
$
Furthermore,
$
\|\vct{y}\|_2 \leq \|{\mtx{\bar{X}}}\boldsymbol{\Sigma}^{1/2}\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2 \leq \sigma_{\max}({\mtx{\bar{X}}})\sqrt{\Sigma_{\max}} \|\betab^\star\|_2 + \sigma\|{\vct{z}}\|_2.
$
Hence, wpa 1,
$$
\|\vct{y}\|_2/\sqrt{n} \leq 2(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma,
$$
where we used the facts that wpa 1: $\|z\|_2/\sqrt{n}\stackrel{{P}}{\longrightarrow} 1$, $\sigma_{\max}({\mtx{\bar{X}}})<\sqrt{2n}(\sqrt{\kappa}+1)$ and \cts{by Assumption \ref{ass:mu}}:
$$
\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^{p}(\sqrt{p}\betab^\star_i)^2 \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[B^2\right].
$$
Put together in \eqref{eq:Ubb}, shows that
\begin{align}
\tn{{\boldsymbol{\hat{\beta}}}_n} < \frac{2(\sqrt{\kappa}+1)\sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}\left[B^2\right]} + 2\sigma}{\sqrt{\Sigma_{\min}}(\sqrt{\kappa}-1)/2} =: \tilde{B}_+.\label{eq:bd_beta}
\end{align}
Recalling that $\widehat{\vct{w}}_n= \sqrt{\boldsymbol{\Sigma}}{\boldsymbol{\hat{\beta}}}_n - \sqrt{\boldsymbol{\Sigma}}\betab^\star$, we conclude, as desired, that wpa 1,
$
\tn{\widehat{\vct{w}}_n} \leq \sqrt{\Sigma_{\max}}\tilde{B}_+ + \sqrt{\Sigma_{\max}} \sqrt{\operatorname{\mathbb{E}}[B^2]} \leq B_+.
$
\end{proof}
Lemma \ref{lem:bd_PO} implies that nothing changes in \eqref{eq:PO} if we further constrain $\vct{w}\in\mathcal{B}$ in \eqref{eq:PO}. Henceforth, with some abuse of notation, we let
\begin{align}\label{eq:PO_bd}
\Phi({\mtx{X}}):=\min_{\vct{w}\in\mathcal{B}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
Next, in order to analyze the primary optimization (PO) problem in \eqref{eq:PO_bd} in apply the CGMT \cite{thrampoulidis2015regularized,thrampoulidis2018precise}. Specifically, we use the constrained formulation of the CGMT given by Theorem \ref{thm closed}. Specifically, the auxiliary problem (AO) corresponding to \eqref{eq:PO_bd} takes the following form with ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$
\begin{align}
\phi({\vct{g}},\vct{h}) = \min_{\vct{w}{\in\mathcal{B}}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_con}
\end{align
We will prove the following techincal result about the AO problem.
\begin{lemma}[Properties of the AO -- Overparameterized regime]\label{lem:AO}
{Let the assumptions of Theorem \ref{thm:master_W2} hold.} Let $\phi_n=\phi({\vct{g}},\vct{h})$ be the optimal cost of the minimization in \eqref{eq:AO_con}. Define $\bar\phi$ as the optimal cost of the following deterministic min-max problem
\begin{align}\label{eq:AO_det}
\bar\phi:=\max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau):=\frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - {\operatorname{\mathbb{E}}\left[\frac{B^2}{1+\frac{u}{\tau}\Lambda}\right]} \right).
\end{align}
The following statements are true.
\noindent{(i).}~The AO minimization in \eqref{eq:AO_con} is $\frac{1}{\Sigma_{\max}}$-strongly convex and has a unique minimizer $\hat{\vct{w}}^{\rm{AO}}_n:=\hat{\vct{w}}^{\rm{AO}}_n({\vct{g}},\vct{h})$.
\noindent{(ii).}~In the limit of $n,p\rightarrow\infty, p/n=\kappa$, it holds that $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$, i.e., for any $\varepsilon>0$:
$$
\lim_{n\rightarrow\infty}\P\left(|\phi({\vct{g}},\vct{h})-\bar\phi|>\varepsilon\right) = 0.
$$
\noindent{(iii).} The max-min optimization in \eqref{eq:AO_det} has a unique saddle point $(u_*,\tau_*)$ satisfying the following:
$$
u_*/\tau_* = \xi\quad\text{and}\quad\tau_* = \gamma,
$$
where $\xi, \gamma$ are defined in Definition \ref{def:Xi}.
\noindent{(iv).}~Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be a function in \fx{$\rm{PL}(3)$}. Let ${\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n={\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n({\vct{g}},\vct{h})=\boldsymbol{\Sigma}^{-1/2}\hat{\vct{w}}^{\rm{AO}}_n + \betab^\star$. Then,
$$
\frac{1}{p}\sum_{i=1}^{p}f\left(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(B,\Lambda,H)\sim\mu\otimes \mathcal{N}(0,1)}\left[f\left(X_{\kappa,\sigma^2}(B,\Lambda,H),B,\Lambda\right) \right].
$$
\fx{In particular, this holds for all functions $f\in {\cal{F}}$ defined in \eqref{eq:pdef}.
\noindent{(v).}~\cts{The empirical distribution of ${\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n$ converges weakly to the measure of $X_{\kappa,\sigma^2}$, and also, for some absolute constant $C>0$:
\begin{align}\label{eq:k_AO}
\tn{{\boldsymbol{\hat{\beta}}^{\rm{AO}}}}^2 < C\quad\text{wpa 1.}
\end{align}
}
\end{lemma}
We prove Lemma \ref{lem:AO} in Section \ref{sec:proofAO}. \fx{We remark that Assumption \ref{ass:mu} on $W_4$-convergence of the joint empirical distribution of $\{(\boldsymbol{\Sigma}_{i,i},\sqrt{p}\boldsymbol{\beta}^\star_i)\}_{i\in[p]}$ is required in the proof of the statement (iv) above. More generally if $W_k$-convergence is known for some integer $k$, then statement (iv) above holds for test functions $f\in\rm{PL}(k-1)$. This is the first place in the proof of Theorem \ref{thm:master_W2}, where we use the assumption $f\in{\cal{F}}$; indeed, we show in Lemma \ref{lem:fL} that ${\cal{F}}_{\cal{L}}\subset{\cal{F}}\subset\rm{PL}(3)$. The second part is in proving the perturbation result in \eqref{eq:dev2show} below. Unlike the former, when proving the perturbation result, the requirement $f\in{\cal{F}}$ cannot be relaxed (e.g.~to $f\in\rm{PL}(k-1)$) by simply increasing the order of $W_k$-convergence in Assumption \ref{ass:mu}.}
\subsubsection{Finalizing the proof of Theorem \ref{thm:master_W2}:}
Here, we show how Lemma \ref{lem:AO} leads to the proof of Theorem \ref{thm:master_W2} when combined with the CGMT framework \cite{thrampoulidis2015regularized,thrampoulidis2018precise}.
\cts{Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be a function in ${\cal{F}}$, where ${\cal{F}}$ was defined in \eqref{eq:pdef}.} For convenience,
define
$$F_n({\boldsymbol{\hat{\beta}}}_n,\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}{\boldsymbol{\hat{\beta}}}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Fix any $\varepsilon>0$ and define the set
\begin{align}\label{eq:S_set}
\mathcal{S} = \mathcal{S}(\betab^\star,\boldsymbol{\Sigma})
=\big\{\vct{w}=\sqrt{\boldsymbol{\Sigma}}(\boldsymbol{\beta}-\boldsymbol{\beta}^\star) \in\mathcal{B} {~\big |~} |F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})-\alpha_*|\geq 2\varepsilon\big\}.
\end{align}
\fy{With this definition, observe that, it suffices to prove that the solution $\hat\vct{w}_n=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\hat{\beta}}}_n-\boldsymbol{\beta}^\star)$ of the PO in \eqref{eq:PO_beta} satisfies $\hat\vct{w}_n\not\in\mathcal{S}$ wpa 1.}
To prove the desired, we need to consider the ``perturbed" PO and AO problems (compare to \eqref{eq:PO} and \eqref{eq:AO_con}) as:
\begin{align}\label{eq:PO_S}
\Phi_S({\mtx{X}})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad {\mtx{\bar{X}}}\vct{w}=\sigma {\vct{z}},
\end{align}
and
\begin{align}
\phi_S({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S}} \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2\quad\text{subject to}\quad \tn{{\vct{g}}}\tn{\vct{w}~{\sigma}}\leq \vct{h}^T\vct{w}+\sigma h.\label{eq:AO_S}
\end{align
Recall here, that ${\mtx{\bar{X}}}={\mtx{X}}\boldsymbol{\Sigma}^{-1/2}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$, ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$, $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$, $h\sim \mathcal{N}(0,1)$ and we have used the change of variables $\vct{w}:=\sqrt{\boldsymbol{\Sigma}}({\boldsymbol{\beta}}-\betab^\star)$ for convenience.
Using \cite[Theorem 6.1(iii)]{thrampoulidis2018precise} it suffices to find costants $\bar\phi, \bar\phi_S$ and $\eta>0$ such that the following three conditions hold:
\begin{enumerate}
\item $\bar\phi_S \geq \bar\phi + 3\eta$,
\item $\phi({\vct{g}},\vct{h}) \leq \bar\phi + \eta$, with probability approaching 1,
\item $\phi_S({\vct{g}},\vct{h}) \geq \bar\phi_S - \eta$, with probability approaching 1.
\end{enumerate}
In what follows, we explicitly find $\bar\phi, \bar\phi_S,\eta$ such that the three conditions above hold.
\vspace{5pt}
\noindent\underline{Satisfying Condition 2}: Recall the deterministic min-max optimization in \eqref{eq:AO_det}. Choose $\bar\phi=\mathcal{D}(u_*,\tau_*)$ be the optimal cost of this optimization. From Lemma \ref{lem:AO}(ii), $\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow}\bar\phi$. Thus, for any $\eta>0$, with probability approaching 1:
\begin{align}\label{eq:phi_lim}
\bar\phi + \eta \geq \phi({\vct{g}},\vct{h}) \geq \bar\phi - \eta.
\end{align}
Clearly then, Condition 2 above holds for any $\eta>0$.
\vspace{5pt}
\noindent\underline{Satisfying Condition 3}: Next, we will show that the third condition holds for appropriate $\bar\phi$.
Let $\hat{\vct{w}}^{\rm{AO}}_n=\hat{\vct{w}}^{\rm{AO}}_n({\vct{g}},\vct{h})$ be the unique minimizer of \eqref{eq:AO_con} as per Lemma \ref{lem:AO}(i), i.e., $\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\hat{\vct{w}}^{\rm{AO}}_n+{\boldsymbol{\beta}}}^2 = \phi({\vct{g}},\vct{h})$. Again from Lemma \ref{lem:AO}, the minimization in \eqref{eq:AO_con} is $1/\Sigma_{\max}$-strongly convex in $\vct{w}$. Here, $\Sigma_{\max}$ is the upper bound on the eigenvalues of $\boldsymbol{\Sigma}$ as per Assumption \ref{ass:mu}. Thus, for any $\tilde\varepsilon>0$ and any feasible $\vct{w}$ the following holds (deterministically):
\begin{align}\label{eq:sc}
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+{\boldsymbol{\beta}}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2{\Sigma_{\max}}},~\text{provided that}~ \|\vct{w}-\hat{\vct{w}}^{\rm{AO}}_n\|_2 \geq \tilde\epsilon.
\end{align}
Now, we argue that {wpa 1,}
\begin{align}\label{eq:dev_arg}
\text{for all}~\vct{w}\in \mathcal{S}~\text{it holds that}~\|\vct{w}-\hat{\vct{w}}^{\rm{AO}}_n\|_2\geq \tilde\varepsilon,
\end{align}
for an appropriate value of a constant $\tilde\varepsilon>0$.
Consider any $\vct{w}\in\mathcal{S}$.
First, by definition in \eqref{eq:S_set}, for $\boldsymbol{\beta}=\boldsymbol{\Sigma}^{-1/2}\vct{w}+\boldsymbol{\beta}^\star$ we have that
$$
|F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \geq 2\varepsilon.
$$
Second, by Lemma \ref{lem:AO}(iv), with probability approaching 1,
$$
|F({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n,\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \leq \epsilon.
$$
Third, we will show that wpa 1, there exists universal constant $C>0$ such that
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n,\betab^\star,\boldsymbol{\Sigma}) - F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})| \leq C {\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n - {\boldsymbol{\beta}}\|_2}\label{eq:dev2show}.
\end{align}
Before proving \eqref{eq:dev2show}, let us argue how combining the above three displays shows the desired. Indeed, in that case, wpa 1,
\begin{align*}
2\varepsilon &\leq |F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})-\alpha_*| \leq |F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n,\betab^\star,\boldsymbol{\Sigma}) - F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})| + |F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n,\betab^\star,\boldsymbol{\Sigma}) - \alpha_*| \\
&\leq \epsilon + C \,\|{\boldsymbol{\beta}}-{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n\|_2. \\
&\qquad\Longrightarrow \|{\boldsymbol{\beta}}-{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n\|_2 \geq {\varepsilon}/{C}=:\hat\varepsilon\\
&\qquad\Longrightarrow \|\vct{w}-\hat{\vct{w}}^{\rm{AO}}_n\|_2 \geq \hat\varepsilon\sqrt{\Sigma_{\min}}=:\tilde\varepsilon.
\end{align*}
In the last line above, we recalled that ${\boldsymbol{\beta}}=\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star$ and $\boldsymbol{\Sigma}_{i,i}\geq\Sigma_{\min},~i\in[p]$ by Assumption \ref{ass:mu}. This proves \eqref{eq:dev_arg}.
Next, combining \eqref{eq:dev_arg} and \eqref{eq:sc}, we find that wpa 1,
$
\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+{\boldsymbol{\beta}}}^2 \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2\Sigma_{\max}},~\text{for all}~ \vct{w}\in\mathcal{S}.
$
Thus,
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \phi({\vct{g}},\vct{h}) + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}.\notag
\end{align}
When combined with \eqref{eq:phi_lim}, this shows that
\begin{align}
\phi_S({\vct{g}},\vct{h}) \geq \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta.
\end{align}
Thus, choosing $\bar\phi_S = \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}}$ proves the Condition 3 above.
\vspace{3pt}
\noindent{\textbf{Perturbation analysis via Pseudo-Lipschitzness (Proof of \eqref{eq:dev2show}).}} To complete the proof, let us now show \eqref{eq:dev2show}. Henceforth, $C$ is used to denote a universal constant whose value can change from line to line. \fy{Recall that $f\in {\cal{F}}$ where ${\cal{F}}:\mathbb{R}^2\times {\cal{Z}}\rightarrow\mathbb{R}$ is the set of $\rm{PL}(3)$ functions such that $f(\cdot,\cdot,z)$ is $\rm{PL}(2)$ for all $z\in{\cal{Z}}$. Suppose that the $\rm{PL}(2)$ constant of $f(\cdot,\cdot,z)$ is upper bounded over $z\in {\cal{Z}}$ by some $C>0$. We also let $C$ change from line to line for notational simplicity. Then, we have the following chain of inequalities:
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})|
&=\frac{1}{p}\sum_{i=1}^p|f(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})-f(\sqrt{p}{\boldsymbol{\beta}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})|
\notag\\
&\leq
\frac{C}{p}\sum_{i=1}^p (1+ \|\sqrt{p}[\betab^\star_i,{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}]\|_2 + \|\sqrt{p}[\betab^\star_i,{\boldsymbol{\beta}}_{i}]\|_2) \sqrt{p}|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}-{\boldsymbol{\beta}}_{i}|\notag\\
&\leq
C \Big(1+ \frac{1}{\sqrt{p}}\big(\sum_{i=1}^{p}\|\sqrt{p}[\betab^\star_i,{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}]\|_2^2\big)^{1/2} + \frac{1}{\sqrt{p}}\big(\sum_{i=1}^p\|\sqrt{p}[\betab^\star_i,{\boldsymbol{\beta}}_{i}]\|_2^2\big)^{1/2}\Big) \|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n-{\boldsymbol{\beta}}\|_2\notag\\
&\leq
C \left(1+ \max\{\|\betab^\star\|_2^2,\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n\|_2^2,\|{\boldsymbol{\beta}}\|_2^2\}^{1/2} \right) \|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n-{\boldsymbol{\beta}}\|_2.\label{eq:fcase_main}
\end{align}
In the second line above, we used the fact that $f(\cdot,\cdot,z)$ is $\rm{PL}(2)$. The third line follows by Cauchy-Schwartz inequality. Finally, in the last line, we used the elementary fact that $a+b+c\leq 3\max\{a,b,c\}$ for $a=2\sum_{i=1}^p(\betab^\star_i)^2$ and $b=\sum_{i=1}^p({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i})^2$ and $c=\sum_{i=1}^p{\boldsymbol{\beta}}_{i}^2$.}
\begin{comment}
First, consider the case $f=f_{\cal{L}}\in{\cal{F}}_{\cal{L}}$, i.e., $f(x,y,z) = z(y-g(x))^2$ for Lipschitz $g$. We have the following chain of inequalities:
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})| &=
\frac{1}{p}\sum_{i=1}^p|\bSi_{i,i}|\left| (\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}))^2-(\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\beta}}_{i}))^2 \right|\notag\\
&\leq
\Sigma_{\max} \frac{1}{p}\sum_{i=1}^p \left| (\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}))^2-(\sqrt{p}\betab^\star_i-g(\sqrt{p}{\boldsymbol{\beta}}_{i}))^2 \right|\notag\\
&\leq
{C} \Sigma_{\max} \frac{1}{p}\sum_{i=1}^p (1+ \|\sqrt{p}[\betab^\star_i,{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}]\|_2 + \|\sqrt{p}[\betab^\star_i,{\boldsymbol{\beta}}_{i}]\|_2) \sqrt{p}|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}-{\boldsymbol{\beta}}_{i}|\notag\\
&\leq
{C} \Sigma_{\max} \Big(1+ \frac{1}{\sqrt{p}}\big(\sum_{i=1}^{p}\|\sqrt{p}[\betab^\star_i,{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}]\|_2^2\big)^{1/2} + \frac{1}{\sqrt{p}}\big(\sum_{i=1}^p\|\sqrt{p}[\betab^\star_i,{\boldsymbol{\beta}}_{i}]\|_2^2\big)^{1/2}\Big) \|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n-{\boldsymbol{\beta}}\|_2\notag\\
&\leq
C \left(1+ \max\{\|\betab^\star\|_2^2,\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n\|_2^2,\|{\boldsymbol{\beta}}\|_2^2\}^{1/2} \right) \|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n-{\boldsymbol{\beta}}\|_2.\label{eq:fcase1}
\end{align}
In the second line above, we used boundedness of $\bSi_{i,i}$ as per Assumption \ref{ass:mu}. {In the third line, we used the fact that the function $\psi(a,b) = (a-g(b))^2$ is $\rm{PL}(2)$ as the quadratic of a Lipschitz function.} The fourth line follows by Cauchy-Schwartz inequality. Finally, in the last line, we used the elementary fact that $a+b+c\leq 3\max\{a,b,c\}$ for $a=2\sum_{i=1}^p(\betab^\star_i)^2$ and $b=\sum_{i=1}^p({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i})^2$ and $c=\sum_{i=1}^p{\boldsymbol{\beta}}_{i}^2$.
Second, consider the case $f\in\rm{PL}(2)$. Let $\vct{a}_i=(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}{\boldsymbol{\beta}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$, for $i\in[p]$. A similar chain of inequality holds:
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n,\betab^\star,\boldsymbol{\Sigma}) - F_n({\boldsymbol{\beta}},\betab^\star,\boldsymbol{\Sigma})| &\leq \frac{{C}}{p}\sum_{i=1}^p (1+\max\{\|\vct{a}_i\|_2,\|\vct{b}_i\|_2\})\|\vct{a}_i-\vct{b}_i\|_2\notag \\
&= \frac{{C}}{p}\sum_{i=1}^p \left(1+\max\{\|\vct{a}_i\|_2,\|\vct{b}_i\|_2\}\right)\cdot|\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i} - \sqrt{p}{\boldsymbol{\beta}}_i|\notag\\
&\leq {C}\left(1+\max\{\frac{1}{p}\sum_{i=1}^p \|\vct{a}_i\|_2^{2},\frac{1}{p}\sum_{i=1}^p\|\vct{b}_i\|_2^{2}\} \right)^{1/2}{\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n - {\boldsymbol{\beta}}\|_2}\notag\\
&\leq C\left(1+\max\{\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n\|_2^2,\|\betab^\star\|_2^2,\|{\boldsymbol{\beta}}\|_2^2,\frac{1}{p}\sum_{i=1}^p\bSi_{i,i}^2\} \right)^{1/2}{\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n - {\boldsymbol{\beta}}\|_2}\notag\\
&\leq C\Sigma_{\max}^2\left(1+\max\{\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n\|_2^2,\|\betab^\star\|_2^2,\|{\boldsymbol{\beta}}\|_2^2\} \right)^{1/2}{\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n - {\boldsymbol{\beta}}\|_2}.\label{eq:fcase2}
\end{align}
The first inequality uses the fact that $f\in\rm{PL}(2)$. The second inequality in the third line follows by Cauchy-Schwartz. The last inequality used Assumption \ref{ass:mu} on boundedness of $\bSi_{i,i}$.
\end{comment}
Hence, it follows from \eqref{eq:fcase_main} that in order to prove \eqref{eq:dev2show}, we need to show boundedness of the following terms: $\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n\|_2$, $\|\betab^\star\|_2$ and $\|{\boldsymbol{\beta}}\|_2$. {By feasibility of ${\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n$ and ${\boldsymbol{\beta}}$, we know that ${\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n,{\boldsymbol{\beta}}\in\mathcal{B}$. Thus, the desired $\|{\boldsymbol{\beta}}\|_2<\infty$ and $\|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n\|_2<\infty$ follow directly by Lemma \ref{lem:bd_PO} (Alternatively, for ${\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n$ we conclude the desired by directly applying Lemma \ref{lem:AO}(v)).} Finally, to prove $\|\betab^\star\|_2<\infty$, note that
$$
\|\betab^\star\|_2^2 =\frac{1}{p}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^2,
$$
\cts{which is bounded wpa 1 by Assumption \ref{ass:mu}, which implies bounded second moments of $\sqrt{p}\betab^\star$. } This completes the proof of \eqref{eq:dev2show}, as desired.
\vspace{5pt}
\noindent\underline{Satisfying Condition 1:} To prove Condition 1, we simply pick $\eta$ to satisfy the following
\begin{align}
\bar\phi_S > \bar\phi + 3 \eta ~\Leftarrow~ \bar\phi + \frac{\tilde\epsilon^2}{2\Sigma_{\max}} - \eta \geq \bar\phi + 3 \eta ~\Leftarrow~ \eta \leq \frac{\tilde\epsilon^2}{8\Sigma_{\max}}.\notag
\end{align}
This completes the proof of Theorem \ref{thm:master_W2}.
~~~~
~~~~
\section{Supporting Results on CGMT}\label{SM cgmt res}
The following theorem replaces the compactness constraint with closedness in the CGMT and is borrowed from \cite{li2020exploring}. For related statements see also \cite[App.~A]{deng2019model}.
\begin{theorem} [CGMT with Closedness Constrains]\label{thm closed} Let $\psi$ be a convex function obeying $\lim_{\tn{\vct{w}}\rightarrow\infty}\psi(\vct{w})=\infty$. Given a closed set $\mathcal{S}$, define
\begin{align}
\Phi_\lambda({\mtx{X}})&=\min_{\vct{w}\in\mathcal{S}}\lambda\tn{{\mtx{X}}\vct{w}}+\psi(\vct{w})\\
\phi_\lambda({\vct{g}},\vct{h})&=\min_{\vct{w}\in\mathcal{S}}\lambda(\tn{\vct{w}}\tn{{\vct{g}}}-\vct{h}^T\vct{w})_++\psi(\vct{w}),
\end{align}
and
\begin{align}
&\Phi_\infty({\mtx{X}})=\min_{\vct{w}\in\mathcal{S},{\mtx{X}}\vct{w}=0}\psi(\vct{w})\\
&\phi_\infty({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S},\tn{\vct{w}}\tn{{\vct{g}}}\leq \vct{h}^T\vct{w}}\psi(\vct{w}).
\end{align}
For all $\lambda\in[0,\infty)\cup\{\infty\}$, we have that
\begin{itemize}
\item $\mathbb{P}(\Phi_\lambda({\mtx{X}})<t)\leq2\mathbb{P}(\phi_\lambda({\mtx{X}})\leq t)$.
\item If $\mathcal{S}$ is additionally convex, we additionally have that $\mathbb{P}(\Phi_\lambda({\mtx{X}})>t)\leq2\mathbb{P}(\phi_\lambda({\mtx{X}})\geq t)$. Combining with the first statement, this implies that for any $\mu,t>0$
\[
\mathbb{P}(|\Phi_\lambda({\mtx{X}})-\mu|>t)\leq2\mathbb{P}(|\phi_\lambda({\mtx{X}})-\mu|\geq t)
\]
\end{itemize}
\end{theorem}
\cmt{
\section{Proof of Constrained CGMT}
\subsection{Proof for the convex case}
\begin{lemma} \label{lem convex}Given a convex and compact $\mathcal{S}$, define the PO and AO problem
\begin{align}
&\Phi_\infty({\mtx{X}})=\min_{\vct{w}\in\mathcal{S},{\mtx{X}}\vct{w}=0}\psi(\vct{w})\\
&\phi_\infty({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S},\tn{\vct{w}}\tn{{\vct{g}}}\leq \vct{h}^T\vct{w}}\psi(\vct{w}).
\end{align}
Suppose ${\mtx{X}},{\vct{g}},\vct{h}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$. Then, we have that
\begin{align}
\mathbb{P}(\Phi_\infty({\mtx{X}})> t)\leq2\mathbb{P}(\phi_\infty({\vct{g}},\vct{h})\geq t).
\end{align}
\end{lemma}
\subsection{Proof for the general case}
\begin{lemma} \label{lem general constraint}Given a compact set $\mathcal{S}$, define the PO and AO problems as in Lemma \ref{lem convex}. We have that
\begin{align}
\mathbb{P}(\Phi_\infty({\mtx{X}})<t)\leq2\mathbb{P}(\phi_\infty({\vct{g}},\vct{h})< t).
\end{align}
\end{lemma}
\begin{proof} The proof is similar to that of Lemma \ref{lem convex}. For a general compact set $\mathcal{S}$, application of Gordon's theorem yields the one-sided bound
\begin{align}
\mathbb{P}(\Phi_\lambda({\mtx{X}})<t)\leq2\mathbb{P}(\phi_\lambda({\vct{g}},\vct{h})\leq t).
\end{align}
To move from finite $\lambda$ to infinite, we make use of Lemma \ref{lem continuous limit}. Define the indicator function $E_\lambda=1_{\Phi_\lambda({\mtx{X}})\leq t}$. Using Lemma \ref{lem continuous limit}, for any choice of ${\mtx{X}}$, $\lim_{\lambda\rightarrow\infty}E_\lambda=\lim_{\lambda\rightarrow\infty}1_{\Phi_\lambda({\mtx{X}})< t}=1_{\Phi_\infty({\mtx{X}})< t}$. Note again that, if the problem is infeasible, then $\lim_{\lambda\rightarrow\infty}E_\lambda=E_\infty=0$. To proceed, we are in a position to apply Dominated Convergence Theorem to find
\begin{align}
&\lim_{\lambda \rightarrow\infty}\operatorname{\mathbb{E}}[E_\lambda]=\operatorname{\mathbb{E}}[E_\infty]\iff\mathbb{P}(\Phi_\infty({\mtx{X}})< t)=\lim_{\lambda\rightarrow\infty}\mathbb{P}(\Phi_\lambda({\mtx{X}})< t).
\end{align}
Applying the identical argument on $\phi_{{\vct{g}},\vct{h}}$ to find $\mathbb{P}(\phi_\infty({\vct{g}},\vct{h})\leq t)=\lim_{\lambda\rightarrow\infty}\mathbb{P}(\phi_\lambda({\vct{g}},\vct{h})\leq t)$, we obtain the desired relation
\begin{align}
\mathbb{P}(\Phi_\infty({\mtx{X}})< t)&=\lim_{\lambda\rightarrow\infty}\mathbb{P}(\Phi_\lambda({\mtx{X}})< t)\\
&\leq 2\lim_{\lambda\rightarrow\infty}\mathbb{P}(\phi_\lambda({\vct{g}},\vct{h})\leq t)\\
&= 2\mathbb{P}(\phi_\infty({\vct{g}},\vct{h})\leq t).
\end{align}
\end{proof}
\begin{lemma}\label{lem continuous limit} Let $\mathcal{S}$ be a compact set and $\psi(\cdot)$ be a continuous function and $f(\vct{w})$ be a non-negative continuous function. Then
\[
\lim_{\lambda\rightarrow\infty}\min_{\vct{w}\in\mathcal{S}}\lambda f(\vct{w})+\psi(\vct{w})=\min_{\vct{w}\in\mathcal{S},f(\vct{w})=0}\psi(\vct{w})
\]
Thus, setting $f(\vct{w})=\tn{{\mtx{X}}\vct{w}}$ and $f(\vct{w})=\tn{\vct{w}}\tn{{\vct{g}}}-\vct{h}^T\vct{w}$, we have that
\begin{align*}
&\lim_{\lambda\rightarrow\infty}\Phi_{\lambda}({\mtx{X}})=\Phi_{\infty}({\mtx{X}})\\
&\lim_{\lambda\rightarrow\infty}\phi_{\lambda}({\vct{g}},\vct{h})=\phi_{\infty}({\vct{g}},\vct{h}).
\end{align*}
\end{lemma}}
\section{Conclusions and Future Directions}\label{sec discuss}
This paper sheds light on under-explored phenomena in pruning practices for neural network model compression. On a theoretical level, we prove an accurate distributional characterization (DC) for the solution of overparameterized least-squares for linear models with correlated Gaussian features. Our DC allows to precisely characterize the pruning performance of popular pruning methods, such as magnitude pruning. Importantly, our DC combined with a linear Gaussian equivalence, leads to precise analytic formulas for the pruning performance of nonlinear random feature models. On the experimental side, we provide a thorough study of overparameterization and pruning with experiments on linear models, random features and neural nets with growing complexity. Our experiments reveal striking phenomena such as a novel double descent behavior for model pruning and the power of overparameterization. They also shed light on common practices such as retraining after pruning.
Going forward, there are several exciting directions to pursue. First, it would be insightful to study whether same phenomena occur for other loss functions in particular for cross-entropy. Second, this work focuses on unregularized regression tasks and it is important to identify optimal regularization schemes for pruning purposes. For instance, should we use classical $\ell_1/\ell_2$ regularization or can we refine them by injecting problem priors such as covariance information? Finally, going beyond pruning, using DC, one can investigate other compression techniques that process the output of the initial overparameterized learning problem, such as model quantization and distillation.
\section{Motivating Examples}\label{sec mot}
\begin{figure}[t!]
\begin{subfigure}{1.5in}\vspace{-5pt}
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[scale=0.2]{figs/FLAT_SIM_s10_n30_p100_SCL25}};
\node at (-2.1,0) [rotate=90,scale=.9]{Test Risk};
\node at (0,-1.7) [scale=.9]{Problem Size ($k/p$)}
\end{tikzpicture}\caption{\small{Identity covariance, spiked latent weights.}}\label{fig1a}\vspace{-0.2cm}
\end{subfigure}~~~~~\begin{subfigure}{1.5in}\vspace{-5pt}
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[scale=0.2]{figs/SPIKE_SIM_s10_n30_p100_SCL25}};
\node at (0,-1.7) [scale=.9]{Problem Size ($k/p$)}
\end{tikzpicture}\caption{\small{Spiked covariance, identical latent weights.}}\label{fig1b}\vspace{-0.2cm}
\end{subfigure}\caption{\small{Our theoretical predictions for various pruning strategies in linear models with $s/p=0.1$ and $n/p=0.3$. We solve ERM using the first $k$ features and then prune to obtain an $s$-sparse model. The vertical dashed line shows the $k=s$ point. The horizontal dashed line highlights the minimum risk among all underparameterized solutions ($k\leq n$) and all solutions obtained by a final retraining. \fy{Retraining curves are omitted here, but they can be found in Fig.~7 of SM.}}}\label{fig1}\vspace{-15pt}
\end{figure}
\subsection{Linear Gaussian Problems}
We begin our study with linear Gaussian problems (LGP), which we formally define as follows.
\begin{definition}[Linear Gaussian Problem (LGP)] \label{def LGP}Given latent vector ${\boldsymbol{\beta}}^\star\in\mathbb{R}^d$, covariance ${\boldsymbol{{\Sigma}}}$ and noise level $\sigma$, assume that each example in $\mathcal{S}$ is generated independently as $y_i=\vct{x}_i^T{\boldsymbol{\beta}}^\star+ \sigma z_i$ where $z_i\sim\mathcal{N}(0,1)$ and $\vct{x}_i\sim\mathcal{N}(0,{\boldsymbol{{\Sigma}}})$. Additionally, the map $\phi(\cdot)$ is identity and $p=d$.
\end{definition}
Albeit simple, LGPs are of fundamental importance for the following reasons: (1) We show in Sec. \ref{sec main} that our theoretical framework rigorously characterizes pruning strategies for LGPs; (2) Through a ``linear Gaussian equivalence", we will use our results for linear models to obtain analytic predictions for nonlinear random features; (3) Our theoretical predictions and numerical experiments discussed next demonstrate that LGPs already capture key phenomena observed in more complex models (e.g., Fig. \ref{figNN}).
In Fig.~\ref{fig1}, we consider LGPs with diagonal covariance ${\boldsymbol{{\Sigma}}}$. We set the sparsity level $s/p=0.1$ and the relative dataset size $n/p=0.3$. To parameterize the covariance and ${\boldsymbol{\beta}}^\star$, we use a {\em{spiked}} vector $\boldsymbol{\lambda}$, the first $s$ entries of which are set equal to $C=25\gg 1$ and the remaining entries equal to $1$. $\boldsymbol{\lambda}$ corresponds to the latent saliency score (cf. ~\eqref{saliency eq}) of the indices. To understand the role of overparameterization, we vary the number of features used in the optimization. Specifically, we solve \eqref{eq:ERM} with $\Delta=[k]$ and vary the number of features $k$ from $0$ to $p$. Here we consider the {\em{train$\rightarrow$prune}} algorithm, where we first solve for ${\boldsymbol{\hat{\beta}}}([k])$ and obtain our pruned model ${\boldsymbol{\hat{\beta}}}^P_s([k])$ by applying magnitude, Hessian or Oracle pruning (cf. ~$P\in \{M,H,O\}$). Since $\boldsymbol{\lambda}$ is non-increasing, the indices are sorted by saliency score; thus, Oracle pruning always picks the first $s$ indices. Solid lines represent analytic predictions, while markers are empirical results.
The vertical dashed line is the sparsity level $s/p$.
\cmt{confusing about why setups of Fig.~\ref{fig1a}, \ref{fig1b}. Are we trying to keep saliency score the same?}
In Fig.~\ref{fig1a}, we set ${\boldsymbol{{\Sigma}}}={\mtx{I}}_p$ and ${\boldsymbol{\beta}}^\star=\sqrt{\boldsymbol{\lambda}}$. Note, that the analytic curves correctly predict the test risk and the double descent behavior. Observe that the Hessian and Magnitude pruning coincide here, since the diagonal of the empirical covariance is essentially identity. In contrast, Fig.~\ref{fig1b} emphasizes the role of the feature covariance by setting ${\boldsymbol{{\Sigma}}}=\text{diag}(\boldsymbol{\lambda})$ and ${\boldsymbol{\beta}}^\star$ to be the all ones vector. In this scenario, we observe that Hessian pruning performs better compared to Fig.~\ref{fig1a} and also outperforms Magnitude pruning. This is because the empirical covariance helps distinguish the salient indices. Importantly, for Hessian and Oracle pruning, the optimal sparse model is achieved in the highly overparameterized regime $k=p$. Notably, the achieved performance at $k=p$ is strictly better than the horizontal dashed line, which highlights the optimal risk among all underparameterized solutions $k\leq n$ and all retraining solutions (see also SM Sec. A). This has two striking consequences. First, {\em{retraining can in fact hurt the performance}}; because the {\em{train$\rightarrow$prune}} performance at $k=p$ is strictly better than {\em{train$\rightarrow$prune$\rightarrow$retrain}} for all $k$. Second, {\em{overparameterized pruning can be provably better than solving the sparse model with the knowledge of the most salient features}} as $k=p$ is also strictly better than $k=s$.
\begin{figure}[t!]
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[scale=0.28]{figs2/RF_4s_80_n200_p1500}};
\node at (-3.1,0) [rotate=90,scale=1.]{Test Risk};
\node at (0,-2.2) [scale=.9]{Overparameterization ($p/n$)}
\end{tikzpicture}
\vspace{-10pt}
\caption{\small{Illustration of the mismatch between pruning with retraining (red markers) and pruning with fresh samples (cyan markers/line). The setting here is exactly the same as in Fig.~\ref{figRF}, but we only show the case of sparsity $4s$ for which the mismatch is observed. Observe that our analytical predictions accurately capture the risk of retraining with fresh samples. However, we observe a discrepancy with the true risk of retraining (without fresh samples) around the interpolation threshold. Also shown the risk of the original ERM solution before pruning (in blue) and of the magnitude-pruned model (before any retraining).
}
}
\label{figRF2}\vspace{-15pt}
\end{figure}
\subsection{Random Features Regression} We relate an ERM problem \eqref{eq:ERM} with nonlinear map $\phi$ to an equivalent LGP. This will allow us to use our theoretical results about the latter to characterize the properties of the original nonlinear map. We ensure the equivalence by properly setting up the LGP to exhibit similar second order statistics as the original problem.
\begin{definition}[Equivalent Linear Problem] Given distribution $(\vct{x},y)$ $\sim\mathcal{D}$, the equivalent LGP(${\boldsymbol{\beta}},{\boldsymbol{{\Sigma}}},\sigma$) with $n$ samples is given with parameters ${\boldsymbol{{\Sigma}}}=\operatorname{\mathbb{E}}[\vct{x}\x^T]$, ${\boldsymbol{\beta}}^\star={\boldsymbol{{\Sigma}}}^{-1}\operatorname{\mathbb{E}}[y\vct{x}]$ and $\sigma=\operatorname{\mathbb{E}}[(y-\vct{x}^T{\boldsymbol{\beta}}^\star)^2]^{1/2}$.
\end{definition}
In Section \ref{sec main}, we formalize the DC of LGPs, which enables us to characterize pruning/retraining dynamics. Then, we empirically verify that DC and pruning dynamics of equivalent LGPs can successfully predict the original problem \eqref{eq:ERM} with non-linear features. The idea of setting up and studying equivalent LGPs as a proxy to nonlinear models, has been recently used in the emerging literature of high-dimensional learning, for predicting the performance of the original ERM task \cite{montanari2019generalization,goldt2020gaussian,abbasi2019universality,derezinski2019exact}. This work goes beyond prior art, which focuses on ERM, by demonstrating that we can also successfully predict the pruning/retraining dynamics. Formalizing the performance equivalence between LGP and equivalent problem is an important future research avenue and it can presumably be accomplished by building on the recent high-dimensional universality results such as \cite{oymak2018universality,hu2020universality,abbasi2019universality,goldt2020gaussian}.
In Fig. \ref{figRF}, we study random feature regression to approximate a synthetic nonlinear distribution. Specifically, data has the following distribution: Given input $\vct{a}\sim\mathcal{N}(0,{\mtx{I}}_d)$, we generate random unit norm ${\boldsymbol{\beta}}^1\in\mathbb{R}^d,{\boldsymbol{\beta}}^2\in\mathbb{R}^d$ and set the label to be a quadratic function given by $y=\vct{a}^T{\boldsymbol{\beta}}^1+(\vct{a}^T{\boldsymbol{\beta}}^2)^2$. Then, we fix ${\mtx{R}}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$ and we generate ReLU features $\vct{x}=\text{ReLU}({\mtx{R}}\vct{a})$, where ${\mtx{R}}$ corresponds to the input layer of a two-layer network. The markers in Fig. \ref{figRF} are obtained by solving RFR and pruning and retraining with varying sparsity targets ($s,2s,4s$ with $s/n=10\%$). Here, $d=10, n=200$. For each marker, the results are averages of 50 ${\mtx{R}}\in\mathbb{R}^{p\times d}$ realizations and 10 iterations for each choice of ${\mtx{R}}$. The lines are obtained via our DC of the equivalent LGP (by using Defs.~\ref{aux_def} and \ref{RT_def}) where the latent parameter ${\boldsymbol{\beta}}^\star$, noise $\sigma$ and the covariance ${\boldsymbol{{\Sigma}}}$ of the RFR problem are calculated for fixed realization of the input layer ${\mtx{R}}$ (similarly averaged over 50 random ${\mtx{R}}$). Our theory and empirical curves exhibit a good match. The results demonstrate the importance of overparameterization for RF pruning, which corresponds to picking {\em{random features smartly}}. Here, the coefficients of least-squares act like a scoring function for the saliency of random features and capture how well they are aligned with the target function. The fact that the risk of the pruned models is minimized in the overparameterized regime implies that least-squares regression succeeds in properly selecting salient random features from a larger candidate set. In the context of deep learning, our discussion can be interpreted as {\em{pruning hidden nodes of the network}}.
\noindent\textbf{Predicting retraining performance.} As discussed in Sec.~\ref{sec main} and Def.~\ref{RT_def}, for the retraining stage, our DC is accomplished by assuming that retraining phase uses $n$ fresh training examples (i.e.~a new dataset $\mathcal{S}_{\text{fresh}}$). Let us denote the resulting model by ${\boldsymbol{\hat{\beta}}}^{RT}_{\text{fresh}}$. Perhaps surprisingly, Fig.~\ref{figRF} shows that this DC correctly captures the performance of ${\boldsymbol{\hat{\beta}}}^{RT}$ with the exception of the red curve ($4s$). Fig.~\ref{figRF2} focuses on this instance and shows that our DC indeed perfectly predicts the fresh retraining performance and verifies the slight empirical mismatch between ${\boldsymbol{\hat{\beta}}}^{RT}$ and ${\boldsymbol{\hat{\beta}}}^{RT}_{\text{fresh}}$.
\vspace{-6pt}
\subsection{Neural Network Experiments}\label{ssec:NNE}
Finally, we study pruning deep neural networks.
Inspired by \cite{nakkiran2019deep}, we train ResNet-20 with changeable filters over CIFAR-10. Here, the filter number $k$ is equivalent to the width/channel of the model. As the width of ResNet-20 changes, the fitting performance of the dataset varies. Here, we apply {\em{train$\rightarrow$prune$\rightarrow$retrain}}.
Select $s$ as the sparsity target and $s$-filter ResNet-20 model as the base model with $N_s$ parameters. First, we train a dense model with $k$ filters and $N_k$ parameters, $N_k\gg N_s$, and prune it by only keeping the largest $N_s$ entries in absolute value non-zero. $N_k$ grows approximately quadratically in $k$. Now, the sparse model shares the same number of parameters amenable to training as the base model. Finally, we retain the pruned network and record its performance on the same dataset and same configuration.
In Fig.~\ref{figNN}, we plot the training and test errors of dense and sparse models. All the neural experiments are trained with Adam optimization and $0.001$ learning rate for $1000$ epochs, with data augmentation. Green, yellow and red lines correspond to $5$, $8$ or $10$ sparsity targets, with around 28,000, 70,000 and 109,000 trainable parameters, respectively. As the width $k$ grows, the training and test errors decrease for all $5$-,$8$-,$10$-filter base models, except for the shaded double descent range. These experiments again verify the main insight revealed to us by studying simpler linear and random-feature models, that is, training a larger model, followed by appropriate pruning, can preform better than training a small model from the beginning.
Another worth-mentioning observation is that with appropriate sparsity level (here, $10$) the pruned model has prediction performance comparable to the dense model. Finally and interestingly, the test error dynamics of the pruned model exhibit a double descent that resembles that of the dense model (previously observed in \cite{nakkiran2019deep}).
\section{Introduction}
Large model size and overparameterization in deep learning are known to improve generalization performance \cite{neyshabur2017geometry}, and, state-of-the-art deep neural networks (DNNs) can be outrageously large. However, such large models are not suitable for certain important application domains, such as mobile computing \cite{tan2019mnasnet,sandler2018mobilenetv2}. Pruning algorithms aim to address the challenge of building lightweight DNNs for such domains. While there are several pruning methods, their common goal is to compress large DNN models by removing weak connections/weights with minimal decline in accuracy. Here, a key empirical phenomenon is that {\em{it is often better to train and prune a large model rather than training a small model from scratch}}. Unfortunately, the mechanisms behind this phenomenon are poorly understood especially for practical gradient-based algorithms. This paper sheds light on this by answering: {\em{What are the optimization and generalization dynamics of pruning overparameterized models? Does gradient descent naturally select the good weights?}}
\vspace{3pt}
\noindent{\bf{Contributions:}}
We analytically study the performance of popular pruning strategies. First, we analyze linear models, and then, generalize the results to nonlinear feature maps. Through extensive simulations, we show that our analytical findings predict similar behaviors in more complex settings.
\noindent\textbf{(a)} {\bf{Distributional characterization (DC):}} The key innovation facilitating our results is a theoretical characterization of the distribution of the solution of overparameterized least-squares. This DC enables us to accurately answer {\em{``what happens to the accuracy if X\% of the weights are pruned?''}}.
\noindent\textbf{(b)} {\bf{Benefits of overparameterization:}} Using DC, we
{obtain rigorous precise characterizations of the pruning performance in linear problems. Furthermore, we use, so called ``linear gaussian equivalences", to obtain sharp analytic predictions for nonlinear maps, which we verify via extensive numerical simulations.}
By training models of growing size and compressing them to fixed sparsity, we identify a novel double descent behavior, where the risk of the pruned model is
{consistently} minimized in the overparameterized regime. Using our theory, we uncover rather surprising scenarios where pruning an overparameterized model is provably better than training a small model with the exact information of optimal nonzero locations.
\noindent\textbf{(c)} {\bf{Benefits of retraining:}} An important aspect of pruning is retraining the model using the favorable nonzero locations identified during the initial training. We show that retraining can actually hurt the performance when features are uncorrelated. However, it becomes critical as correlations increase. Importantly, we devise the DC of the {\em{train$\rightarrow$prune$\rightarrow$retrain process}} (see Figs.~\ref{figRF} and \ref{figRF2} and the discussion around Def.~\ref{RT_def} for details), and, we demonstrate that it correctly captures the pruning performance of random features that are known to be good proxies for understanding DNN behavior \cite{jacot2018neural}.
{We anticipate that our techniques towards establishing the DC of the overparameterized problems might be useful, beyond the context of pruning, in other statistical inference tasks that require careful distributional studies.}
\begin{figure}[t!]
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[scale=0.33]{figs/CIFAR10_New_DD}};
\node at (-3.5,0) [rotate=90,scale=1.]{Training / Test Error};
\node at (0,-2.6) [scale=.9]{Width Parameter (\# of filters $k$)}
\end{tikzpicture}
\vspace{-10pt}
\caption{\small{We train sparse ResNet-20 models on the CIFAR-10 dataset with varying width (i.e. number of filters) and sparsity targets. The width parameter controls the overall model size. The solid (resp. dashed) lines are test (resp. training) errors. The blue line corresponds to training of a dense model with width-$k$. The other three curves correspond to sparsity targets $s\in\{5,8,10\}$, for which a dense model of width-$k$ is first pruned to achieve the exact same number of nonzeros as a dense model of width-$s$ and then retrained over the identified nonzero pattern. Surprisingly, all curves interpolate (achieve zero training error) around the same width parameter despite varying sparsity. The best test error is always achieved in the overparameterized regime (large width). Test error curves have two local minima which uncovers a novel double descent phenomena for pruning. {The shaded region highlights the transition to zero training error, where the test error peaks.}}}
\label{figNN}\vspace{-15pt}
\end{figure}
\subsection{Prior Art}
This work relates to the literature on model compression and overparameterization in deep learning.
\vspace{-2pt}\noindent {\bf{Neural network pruning:}} Large model sizes in deep learning have led to a substantial interest in model pruning/quantization \cite{han2015deep,hassibi1993second,lecun1990optimal}. DNN pruning has a diverse literature with various architectural, algorithmic, and hardware considerations \cite{sze2017efficient,han2015learning}. The pruning algorithms can be applied before, during, or after training a dense model \cite{lee2018snip,wang2020picking,jin2016training,oymak2018learning} and in this work we focus on after training. Related to over-parameterizarion, \cite{frankle2018lottery} shows that a large DNN contains a small subset of favorable weights (for pruning), which can achieve similar performance to the original network when trained with the same initialization. \cite{zhou2019deconstructing,malach2020proving,pensia2020optimal} demonstrate that there are subsets with good test performance even without any training and provide theoretical guarantees. However, these works do not answer why practical gradient-based algorithms lead to good pruning outcomes. {Closer to us, \cite{li2020exploring} derives formulas for predicting the pruning performance of over-parameterized least-squares without proofs. In contrast, we provide provable guarantees, and, also obtain DC for more complex problems with general design matrices and nonlinearities.}
\begin{figure}[t!]
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[scale=0.33]{figs2/RF_COMBINED_80_n200_p1500}};
\node at (-3.5,0) [rotate=90,scale=1.]{Test Risk};
\node at (0,-2.6) [scale=.9]{Overparameterization ($p/n$)}
\end{tikzpicture}
\vspace{-10pt}
\caption{\small{Random feature regression (RFR) with ReLU feature-map $\phi(\vct{a})=\text{ReLU}({\mtx{R}}\vct{a})$. Here ${\mtx{R}}$ has i.i.d.~standard normal entries corresponding to the input layer of a shallow neural net and we regress the output layer. Solid lines follow from our distributional characterization and the markers are obtained by solving random feature regression, which exhibit a good match. The blue line is the performance of usual RFR with growing number of features $p$. The other lines are obtained by solving RFR with $p$ features and pruning and retraining the solution to fixed sparsity levels ($s,2s,4s$) with $s/n=0.1$. Importantly, the risks of the retrained models exhibit double descent and are minimized when $p\gg n$ despite fixed model size / sparsity. The slight mismatch of the red curve/markers is explained in Fig.~\ref{figRF2}.}
\cmt{We emphasize that RFR is a well-recognized proxy for the DNN regression due to the equivalence between kernels and wide DNNs \cite{jacot2018neural}.}
\label{figRF}\vspace{-15pt}
\end{figure}
\cmt{The green line prunes the resulting coefficients to find a sparse model with fixed sparsity $s/n=0.1$. The red line resolves RFR using the features identified after pruning. All lines exhibit double descent at $n=p$.}
\noindent {\bf{Benefits of overparameterization:}} Studies on the optimization and generalization properties of DNNs demonstrate that overparameterization acts as a catalyst for learning. \cite{arora2018optimization,neyshabur2014search,gunasekar2017implicit,ji2018risk} argue that gradient-based algorithms are implicitly biased towards certain favorable solutions (even without explicit regularization) to explain benign overfitting \cite{bartlett2020benign,oymak2020towards,du2019gradient,chizat2019lazy,belkin2018understand,belkin2019does,tsigler2020benign,liang2018just,mei2019generalization,ju2020overfitting}. More recently, these studies have led to interesting connections between kernels and DNNs and a flurry of theoretical developments. Closest to us, \cite{nakkiran2019deep,belkin2019two,belkin2019reconciling} uncover a double-descent phenomenon: the test risk has two minima as a function of model size. One minimum occurs in the classical underparameterized regime whereas the other minimum occurs when the model is overparameterized and the latter risk can in fact be better than former. Closer to our theory, \cite{derezinski2019exact,hastie2019surprises,montanari2019generalization,deng2019model,kini2020analytic,liang2020precise,salehi2020performance,ju2020overfitting} characterize the asymptotic performance of overparameterized learning problems. However these works are limited to characterizing the test error of regular (dense) training. In contrast, we use distributional characterization (DC) to capture the performance of more challenging pruning strategies and we uncover novel double descent phenomena (see Fig.~\ref{figNN}).
\cmt{For linear models, implicit bias phenomena have been studied for various loss functions and algorithms (e.g.~logistic loss converging to max-margin solution on separable data) \cite{ji2018risk,soudry2018implicit,nacson2019stochastic}. Follow-up works show that such results continue to hold for nonlinear problems \cite{gunasekar2017implicit,oymak2019overparameterized,azizan2018stochastic}. More recently, this line of works has further motivated the study of generalization/optimization guarantees for deep networks and their connections to random features \cite{du2019gradient,allen2019convergence,chizat2019lazy,belkin2018understand,belkin2019does,liang2018just,mei2019generalization}.}
\subsection{Proof of Lemma \ref{lem:AO}}\label{sec:proofAO}
\vspace{3pt}
\subsubsection{Proof of (i).} Strong convexity of the objective function in \eqref{eq:AO_con} is easily verified by the second derivative test and use of Assumption \ref{ass:mu} that $\bSi_{i,i}\leq\Sigma_{\max},~i\in[p].$ Uniqueness of the solution follows directly from strong convexity. \ct{Strictly speaking we might need to also argue existence, i.e., feasibility of the AO. An indirect way is to show feasibility using the CGMT, but it seems unnecessarily complicated?}
\vspace{5pt}
\subsubsection{Proof of (ii).}Using Lagrangian formulation, the solution $\hat{\vct{w}}^{\rm{AO}}_n$ to \eqref{eq:AO_con} is the same as the solution to the following:
\begin{align}
\left(\hat{\vct{w}}^{\rm{AO}}_n,u_n\right) :=\arg\min_{\vct{w}\in\mathcal{B}}\max_{u\geq 0} ~\frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + u \left( \sqrt{\tn{\vct{w}}^2+\sigma^2} \tn{\bar{\g}} - \sqrt{\kappa}\,\bar{\h}^T\vct{w} + \frac{\sigma h}{\sqrt{n}} \right)\label{eq:AO_2}
\end{align}
where we have: (i) set $\bar{\g} := {\vct{g}}/\sqrt{n}$ and $\bar{\h}:= \vct{h}/\sqrt{p}$; (ii) recalled that $p/n=\kappa$; and, (iii) used $\left(\hat{\vct{w}}^{\rm{AO}}_n,u_n\right)$ to denote the optimal solutions in \eqref{eq:AO_2}. The subscript $n$ emphasizes the dependence of $\left(\hat{\vct{w}}^{\rm{AO}}_n,u_n\right)$ on the problem dimensions. Also note that (even though not explicit in the notation) $\left(\hat{\vct{w}}^{\rm{AO}}_n,u_n\right)$ are random variables depending on the realizations of $\bar{\g},\bar{\h}$ and $h$.
Notice that the objective function above is convex in $\vct{w}$ and linear (thus, concave) in $u$. Also, $\mathcal{B}$ is compact. Thus, strong duality holds and we can flip the order of min-max \cite{fan1953minimax}. Moreover, in order to make the objective easy to optimize with respect to $\vct{w}$, we use the following variational expression for the square-root term $\sqrt{\tn{\vct{w}}^2+\sigma^2}$:
$$
\tn{\bar{\g}}\sqrt{\tn{\vct{w}}^2+\sigma^2} = \tn{\bar{\g}}\cdot\min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau}{2} + \frac{\tn{\vct{w}}^2+\sigma^2}{2\tau} \right\} = \min_{\tau\in[\sigma,\sqrt{\sigma^2+B_+^2}]} \left\{ \frac{\tau\tn{\bar{\g}}^2}{2} + \frac{\sigma^2}{2\tau} + \frac{\tn{\vct{w}}^2}{2\tau} \right\},
$$
where $B_+$ is defined in Lemma \ref{lem:bd_PO}. For convenience define the constraint set for the variable $\tau$ as $\mathcal{T}':=[\sigma,\sqrt{\sigma^2+B_+^2}]$. For reasons to be made clear later in the proof (see proof of statement (iii)), we consider the (possibly larger) set:
\[
\mathcal{T}:=[\sigma,\max\{\sqrt{\sigma^2+B_+^2},2\tau_*\}]\,
\]
where $\tau_*$ is as in the statement of the lemma.
The above lead to the following equivalent formulation of \eqref{eq:AO_2},
\begin{align}
\left(\hat{\vct{w}}^{\rm{AO}}_n,u_n,\tau_n\right) = \max_{u\geq 0}\min_{\vct{w}\in\mathcal{B},\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} + \min_{\vct{w}\in\mathcal{B}} \left\{ \frac{1}{2}\tn{\boldsymbol{\Sigma}^{-1/2}\vct{w}+\betab^\star}^2 + \frac{u}{2\tau}\tn{\vct{w}}^2 - u \sqrt{\kappa}\,{\bar{\h}}^T\vct{w} \right\} .\label{eq:AO_3}
\end{align}
The minimization over $\vct{w}$ is easy as it involves a strongly convex quadratic function. First, note that the unconstrained optimal $\vct{w}':=\vct{w}'(\tau,u)$ (for fixed $(\tau,u)$) is given by
\begin{align}\label{eq:w'}
\vct{w}':=\vct{w}'(\tau,u) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right),
\end{align}
and \eqref{eq:AO_3} simplifies to
\begin{align}
\left(u_n,\tau_n\right)=\max_{u\geq 0}\min_{\tau\in\mathcal{T}} ~ \frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}} - \frac{1}{2} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1} \left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u\sqrt{\kappa}\bar{\h}\right)\,=:\mathcal{R}(u,\tau) .\label{eq:AO_4}
\end{align}
It can be checked by direct differentiation and the second-derivative test that the objective function in \eqref{eq:AO_4} is strictly convex in $\tau$ and strictly concave in $u$ over the domain $\{(u,\tau)\in\mathbb{R}_+\times\mathbb{R}_+\}$ \footnote{{To analyze the matrix-vector product term in \eqref{eq:AO_4} for $(\tau,u)$ one can use the fact that ${\boldsymbol{{\Sigma}}}$ is diagonal. This way, as a function of $u$ and $\tau$ the analysis reduces to the properties of relatively simple functions. For instance, for $\tau$, this function is in the form $f(\tau)=-(a+b/\tau)^{-1}$ for $a,b>0$, which is strictly convex.}}. Thus, the saddle point $(u_n,\tau_n)$ is unique. Specifically, this implies that the optimal $\hat{\vct{w}}^{\rm{AO}}_n$ in \eqref{eq:AO_3} is given by (cf. \eqref{eq:w'})
\begin{align}\label{eq:w_n}
\hat{\vct{w}}^{\rm{AO}}_n=\vct{w}'(\tau_n,u_n) = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).
\end{align}
In Lemma \ref{lem:AO}(v) we will prove that wpa 1, in the limit of $p\rightarrow\infty$, $\|\hat{\vct{w}}^{\rm{AO}}_n\|_2\leq C$ for sufficiently large absolute constant $C>0$. Thus, by choosing the upper bound in the definition of $\mathcal{B}$ in Lemma \ref{lem:bd_PO} strictly larger than C, guarantees that the unconstrained $\hat{\vct{w}}^{\rm{AO}}_n$ in \eqref{eq:w_n} is feasible in \eqref{eq:AO_3}.
\noindent{\textbf{Asymptotic limit of the key quantities $\tau_n,u_n$:}} In what follows, we characterize the high-dimensional limit of the optimal pair $(u_n,\tau_n)$ in the limit $n,p\rightarrow\infty,~p/n\rightarrow\kappa$.
We start by analyzing the (point-wise) convergence of $\mathcal{R}(u,\tau)$.
For the first three summands in \eqref{eq:AO_4}, we easily find that
$$
\left\{\frac{u\tau\tn{\bar{\g}}^2}{2} + \frac{u\sigma^2}{2\tau} + \frac{u\sigma h}{\sqrt{n}}\right\}~ \stackrel{{P}}{\longrightarrow}~\left\{
\frac{u\tau}{2} + \frac{u\sigma^2}{2\tau} \right\}.
$$
Next, we study the fourth summand. First, note that
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}(u\sqrt{\kappa}\bar{\h}) &= u^2\kappa\,\frac{1}{p}\vct{h}^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\vct{h} \notag\\
&= u^2\kappa\,\frac{1}{p}\sum_{i=1}^{p}\frac{\vct{h}_i^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}} \notag\\
&\stackrel{{P}}{\longrightarrow} u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right].
\end{align}
In the last line, $\Lambda$ is a random variable as in Definition \ref{def:Xi}. {Also, we used Assumption \ref{ass:mu} together with the facts that $\vct{h}$ is independent of $\boldsymbol{\Sigma}$ and that the function $(x_1,x_2)\mapsto x_1^2(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(3)$ assuming $x_2$ is bounded (see Lemma \ref{lem:PLbdd} for proof).}
\ct{Question: Is it immediate that the empirical distribution of $(\beta,\boldsymbol{\Sigma},\vct{h})$ converges in $W_k$ to $\mu\otimes\mathcal{N}(0,1)$ given that $(\beta,\boldsymbol{\Sigma})$ converges to $\mu$ and $\vct{h}$ is independent???}
Second, we find that
\begin{align}
(\betab^\star)^T\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star &= \frac{1}{p}(\sqrt{p}\betab^\star)^T\left({\mtx{I}}+\frac{u}{\tau}\boldsymbol{\Sigma}\right)^{-1}(\sqrt{p}\betab^\star) \notag\\
&= \frac{1}{p}\sum_{i=1}^{p}\frac{\left(\sqrt{p}\betab^\star_i/\sqrt{\bSi_{i,i}}\right)^2}{\bSi_{i,i}^{-1}+\frac{u}{\tau}}\notag\\
&\stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right].
\end{align}
Here, $\Lambda,B$ are random variables as in Definition \ref{def:Xi} and {we also used Assumption \ref{ass:mu} together with the fact that the function $(x_1,x_2)\mapsto x_1^2x_2^{-1}(x_2^{-1}+u/\tau)^{-1}$ is $\rm{PL}(3)$ assuming $x_2$ is bounded (see Lemma \ref{lem:PLbdd} for proof).}
Third, \cts{by independence of $(\betab^\star, \boldsymbol{\Sigma})$ from $\vct{h}$}
\begin{align}
(u\sqrt{\kappa}\bar{\h})^T\left(\boldsymbol{\Sigma}^{-1}+\frac{u}{\tau}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\betab^\star = u\sqrt{\kappa} \cdot \frac{1}{p}\sum_{i=1}
^p \frac{\vct{h}_i\boldsymbol{\Sigma}_{i,i}^{-1/2}(\sqrt{p}\betab^\star_i)}{\boldsymbol{\Sigma}_{i,i}^{-1}+\frac{u}{\tau}{\mtx{I}}} ~\stackrel{{P}}{\longrightarrow}~ 0.
\end{align}
Putting these together, the objective $\mathcal{R}(u,\tau)$ in \eqref{eq:AO_4} converges point-wise in $u,\tau$ to
\begin{align}
\mathcal{R}(u,\tau)\stackrel{{P}}{\longrightarrow}\mathcal{D}(u,\tau) := \frac{1}{2}\left({u\tau} + \frac{u\sigma^2}{\tau} - u^2\kappa\,\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] - \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\Lambda^{-1}+\frac{u}{\tau}}\right] \right).\label{eq:conv_pt}
\end{align}
Note that $\mathcal{R}(u,\tau)$ (and thus, $\mathcal{D}(u,\tau)$) is convex in $\tau$ and concave in $u$. Thus, the convergence in \eqref{eq:conv_pt} is in fact uniform (e.g., \cite{AG1982}) and we can conclude that
\begin{align}\label{eq:Dc0}
\phi({\vct{g}},\vct{h}) \stackrel{{P}}{\longrightarrow} \max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).
\end{align}
and {using strict concave/convexity of $\mathcal{D}(u,\tau)$, we also have the parameter convergence \cite[Lem. 7.75]{NF36}}
\begin{align}\label{eq:Dc}
{(u_n,\tau_n) \stackrel{{P}}{\longrightarrow} (u_*,\tau_*):=\arg\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau).}
\end{align}
In the proof of statement (iii) below, we show that the saddle point of \eqref{eq:Dc0} is $(u_*,\tau_*)$. In particular, $\tau_*$ is strictly in the interior of $\mathcal{T}$, which combined with convexity implies that
$$
\max_{u\geq 0}\min_{\tau\in\mathcal{T}}~ \mathcal{D}(u,\tau) = \max_{u\geq 0}\min_{\tau>0}~ \mathcal{D}(u,\tau) =: \bar\phi.
$$
This, together with the first display above proves the second statement of the lemma.
\vspace{5pt}
\subsubsection{Proof of (iii).} Next, we compute the saddle point $(u_*,\tau_*)$ by studying the first-order optimality conditions of the strictly concave-convex $\mathcal{D}(u,\tau)$. Specifically, we consider the unconstrained minimization over $\tau$ and we will show that the minimum is achieved in the strict interior of $\mathcal{T}$. Direct differentiation of $\mathcal{D}(u,\tau)$ gives
\begin{subequations}
\begin{align}
{\tau} + \frac{\sigma^2}{\tau} - 2u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] + \frac{u^2}{\tau}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] + \frac{1}{\tau}\operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0, \label{eq:fo1}\\
{u} - \frac{u\sigma^2}{\tau^2} - \frac{u^3}{\tau^2}\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} + \frac{u}{\tau}\right)^2}\right] - \frac{u}{\tau^2} \operatorname{\mathbb{E}}\left[\frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1}+\frac{u}{\tau}\right)^2}\right] &= 0,\label{eq:fo2}
\end{align}
\end{subequations}
Multiplying \eqref{eq:fo2} with $\frac{\tau}{u}$ and adding to \eqref{eq:fo1} results in the following equation
\begin{align}
\tau = u\kappa\operatorname{\mathbb{E}}\left[ \frac{1}{\Lambda^{-1}+\frac{u}{\tau}} \right] ~\Leftrightarrow ~
\operatorname{\mathbb{E}}\left[ \frac{1}{(\frac{u}{\tau}\Lambda)^{-1}+1} \right] = \frac{1}{\kappa} \label{eq:tauu}\,.
\end{align}
Thus, we have found that the ratio $\frac{u_*}{\tau_*}$ is the unique solution to the equation in \eqref{eq:tauu}. Note that this coincides with the Equation \eqref{eq:ksi} that defines the parameter $\xi$ in Definition \ref{def:Xi}. The fact that \eqref{eq:tauu} has a unique solution for all $\kappa>1$ can be easily seen as $F(x)=\operatorname{\mathbb{E}}\left[ \frac{1}{(x\Lambda)^{-1}+1} \right], x\in\mathbb{R}_+$ has range $(0,1)$ and is strictly increasing (by differentiation).
Thus, we call $\xi=\frac{u_*}{\tau_*}$. Moreover, multiplying \eqref{eq:fo2} with $u$ leads to the following equation for $\tau_*$:
\begin{align}
u_*^2 = \sigma^2\xi^2 + u_*^2 \xi^2 \kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right] + \xi^2 \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right] ~\Rightarrow~ \tau_*^2 = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\xi^2\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left(\Lambda^{-1} +\xi\right)^2}\right]} = \frac{\sigma^2 + \operatorname{\mathbb{E}}\left[ \frac{B^2\Lambda^{-1}}{\left(\Lambda^{-1} + \xi\right)^2}\right]}{1-\kappa \operatorname{\mathbb{E}}\left[ \frac{1}{\left((\xi\Lambda)^{-1} +1\right)^2}\right]}.
\end{align}
{Again, note that this coincides with Equation \eqref{eq:gamma} that determines the parameter $\gamma$ in Definition \ref{def:Xi}, i.e., $\tau_*^2 = \gamma.$ By definition of $\mathcal{T}$ and of $\tau_*$, it is clear that $\tau_\star$ is in the strict interior of $\mathcal{T}$.}
\vspace{5pt}
\subsubsection{Proof of (iv).} For convenience, define
$$F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p} \sum_{i=1}^{p} f\left(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i},\sqrt{p}\betab^\star_{i},\boldsymbol{\Sigma}_{ii}\right)\quad\text{and}\quad\alpha_*:=\operatorname{\mathbb{E}}_\mu\left[f(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda)\right].$$
Recall from \eqref{eq:w_n} the explicit expression for $\hat{\vct{w}}^{\rm{AO}}_n$, repeated here for convenience.
\begin{align}
\hat{\vct{w}}^{\rm{AO}}_n = -\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\left(\boldsymbol{\Sigma}^{-1/2}\betab^\star-u_n\sqrt{\kappa}\bar{\h}\right).\notag
\end{align}
Also, recall that ${\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n = \boldsymbol{\Sigma}^{-1/2}\hat{\vct{w}}^{\rm{AO}}_n+\betab^\star$. Thus, (and using the fact that $\bar{\h}$ is distributed as $-\bar{\h}$),
\begin{align}
{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n &=\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}u_n\sqrt{\kappa}\bar{\h} + \left({\mtx{I}}-\boldsymbol{\Sigma}^{-1/2}\left(\boldsymbol{\Sigma}^{-1}+\frac{u_n}{\tau_n}{\mtx{I}}\right)^{-1}\boldsymbol{\Sigma}^{-1/2}\right)\betab^\star\notag\\
\Longrightarrow{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\bar{\h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i.\label{eq:betan}
\end{align}
For $i\in[p]$, define
\begin{align}
\vct{v}_{n,i} = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\bar{\h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\betab^\star_i
\end{align}
In the above, for convenience, we have denoted $\xi_n:=u_n/\tau_n$ and recall that $\xi_*:=u_*/\tau_*$.
The proof proceeds in two steps. In the first step, we use the fact that $\xi_n\stackrel{{P}}{\longrightarrow}\xi_*$ and $u_n\stackrel{{P}}{\longrightarrow} u_\star$ (see \eqref{eq:Dc}) to prove that for any $\varepsilon\in(0,\xi_*/2)$, there exists an absolute constant $C>0$ such that wpa 1:
\begin{align}\label{eq:ivstep1}
|F_n(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.
\end{align}
In the second step, we use pseudo-Lipschitzness of $f$ and Assumption \ref{ass:mu} to prove that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:ivstep2}
\end{align}
The desired follows by combining \eqref{eq:ivstep1} and \eqref{eq:ivstep2}. Thus, in what follows, we prove \eqref{eq:ivstep1} and \eqref{eq:ivstep2}.
\vspace{2pt}
\noindent\underline{Proof of \eqref{eq:ivstep1}.}~~Fix some $\varepsilon\in(0,\xi_*/2)$. From \eqref{eq:Dc}, we know that w.p.a. 1 $|\xi_n-\xi_*|\leq \varepsilon$ and $|u_n-u_*|\leq \varepsilon$. Thus, $\hat{\vct{w}}^{\rm{AO}}_n$ is close to $\vct{v}_n$. Specifically, in this event, for every $i\in[p]$, it holds that:
\begin{align}
\notag|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i} - \vct{v}_{n,i}| &\leq {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{\tau_n}{1+(\xi_n\bSi_{i,i})^{-1}}-\frac{\tau_*}{1+(\xi_*\bSi_{i,i})^{-1}}\right| \\
\notag&= {|\betab^\star_i|}\left|\frac{1}{1+\xi_n\bSi_{i,i}}-\frac{1}{1+\xi_*\bSi_{i,i}}\right| + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\boldsymbol{\Sigma}_{i,i}}}\left|\frac{u_n}{\xi_n+\bSi_{i,i}^{-1}}-\frac{u_*}{\xi_*+\bSi_{i,i}^{-1}}\right| \\
\notag& \leq
{|\betab^\star_i|}\frac{|\bSi_{i,i}||\xi_n-\xi_*|}{|1+\xi_n\bSi_{i,i}||1+\xi_*\bSi_{i,i}|} + \sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}} \frac{u_*|\xi_n-\xi_*|}{(\xi_n+\bSi_{i,i}^{-1})(\xi_*+\bSi_{i,i}^{-1})} +
\sqrt{\kappa}\frac{|\bar{\h}_i|}{\sqrt{\bSi_{i,i}}}\frac{|u_n-u_*|}{\xi_n+\bSi_{i,i}^{-1}}
\\
\notag& \leq
{|\betab^\star_i|}{\Sigma_{\max}\varepsilon} + \sqrt{\kappa}{|\bar{\h}_i|} u_*\Sigma_{\max}^{3/2} \varepsilon+
\sqrt{\kappa}|\bar{\h}_i|\Sigma_{\max}^{1/2}\varepsilon
\\
&\leq C\varepsilon \left(|\bar{\h}_i| + |\betab^\star_i|\right).\label{eq:betav}
\end{align
where $C=C(\Sigma_{\max},\kappa,u_*)$ is an absolute constant. In the second line above, we recalled that $u_n=\tau_n\xi_n$ and $u_*=\tau_*\xi_*$. In the third line, we used the triangle inequality. In the fourth line, we used that $\xi_*>0$, $0<\bSi_{i,i}\leq\Sigma_{\max}$ and $\xi_n\geq \xi_*-\varepsilon \geq \xi_*/2 >0$.
Now, we will use this and Lipschitzness of $f$ to argue that there exists absolute constant $C>0$ such that wpa 1,
\begin{align}
|F_n(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) - F_n(\sqrt{p}\vct{v}_n,\sqrt{p}\betab^\star,\boldsymbol{\Sigma}) | \leq C\varepsilon.\notag
\end{align}
Denote, $\vct{a}_i=(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$ and $\vct{b}_i=(\sqrt{p}\vct{v}_{n,i},\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})$. Following the exact same argument as in \eqref{eq:dev2show} (just substitute ${\boldsymbol{\beta}}\leftrightarrow\vct{v}_n$ in the derivation), we have that for some absolute constant $C>0$ wpa 1:
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| &\leq C \|{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n-\vct{v}_n\|_2.
\end{align}
From this and \eqref{eq:betav}, we find that
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})|
&\leq
C\varepsilon \left(\sum_{i=1}^p \left(|\bar{\h}_i|+|\betab^\star_i|\right)^2\right)^{1/2}\notag
\\
&\leq
C\varepsilon\sqrt{2}\sqrt{\|\betab^\star\|_2^2 + \|\bar{\h}\|_2^2}.\label{eq:epsS}
\end{align}
But, \cts{recall that $\|\betab^\star\|_2^2 = \frac{1}{p}\sum_{i=1}^p(\sqrt{p}\betab^\star_i)^2<\infty$, as $p\rightarrow \infty$ by Assumption \ref{ass:mu}.}
Also, since $\bar{\h}_i\sim\mathcal{N}(0,1/p)$, it holds that $\|\bar{\h}\|_2^2\leq 2$, wpa 1 as $p\rightarrow\infty$. Therefore, from \eqref{eq:epsS}, wpa 1, there exists constant $C>0$ such that
\begin{align}
|F_n({\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n({\vct{g}},\vct{h}),\betab^\star,\boldsymbol{\Sigma}) - F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \leq C \cdot\varepsilon,\notag
\end{align}
as desired.
\vspace{2pt}
\noindent\underline{Proof of \eqref{eq:ivstep2}.}~~Next, we will use Assumption \ref{ass:mu} to show that
\begin{align}
|F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma})| \stackrel{{P}}{\longrightarrow} \alpha_*.\label{eq:AO_conv}
\end{align}
Notice that $\vct{v}_n$ is a function of $\betab^\star,\boldsymbol{\Sigma},\bar{\h}$. Concretely, define ${\widetilde{g}}:\mathbb{R}^3\rightarrow\mathbb{R}$, such that
$$
\widetilde{g}(x_1,x_2,x_3) := \frac{x_2^{-1/2}}{1+(\xi_* x_2)^{-1}}\sqrt{\kappa}\tau_* x_3 + (1-(1+\xi_*x_2)^{-1})x_1,
$$
and notice that
$$
\sqrt{p}\vct{v}_{n,i} = \widetilde{g}\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right) = \frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_*\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_*\vct{h}_i +\left(1-\frac{1}{1+\xi_*\boldsymbol{\Sigma}_{i,i}}\right)\sqrt{p}\betab^\star_i.
$$
Thus,
$$
F_n(\vct{v}_n,\betab^\star,\boldsymbol{\Sigma}) = \frac{1}{p}\sum_{i=1}^p f\left(g\left(\sqrt{p}\betab^\star_i,\bSi_{i,i},\vct{h}_i\right),\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) =: \frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right),
$$
where we have defined $h:\mathbb{R}^3\rightarrow\mathbb{R}$:
\begin{align}
h(x_1,x_2,x_3) := f\left(\widetilde{g}(x_2,x_3,x_1),x_2,x_3\right).\label{h func}
\end{align}
\cts{We will prove that $h\in\rm{PL}(4)$. Indeed, if that were the case, then Assumption \ref{ass:mu} gives}
\begin{align}
\frac{1}{p}\sum_{i=1}^p h\left(\vct{h}_i,\sqrt{p}\betab^\star_i,\bSi_{i,i}\right) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[h(H,B,\Lambda)\right] &= \operatorname{\mathbb{E}}_{\mathcal{N}(0,1)\otimes \mu}\left[f\left(\widetilde{g}(B,\Lambda,H),B,\Lambda\right))\right]\\
& = \operatorname{\mathbb{E}}[f\left(X_{\kappa,\sigma^2}(\Lambda,B,H),B,\Lambda\right)] = \alpha_*,
\end{align}
where the penultimate equality follows by recognizing that (cf. Eqn. \eqref{eq:X})
$$
\widetilde{g}(B,\Lambda,H) = (1-(1+\xi_*\Lambda)^{-1})B + \sqrt{\kappa}\frac{\tau_*\Lambda^{-1/2}}{1+(\xi_*\Lambda)^{-1}}H = X_{\kappa,\sigma^2}(\Lambda,B,H).
$$
It remains to show that $h\in\rm{PL}(4)$. Lemma \ref{lem:hPL} in Section \ref{SM useful fact} shows that if $f\in\rm{PL}(k)$, then $h\in\rm{PL}(k+1)$ for all integers $k\geq 2$. \fy{Using this and the fact that ${\cal{F}}\subset\rm{PL}(3)$, for any $f\in{\cal{F}}$, we find that $h\in \rm{PL}(4)$. This completes the proof of \eqref{eq:ivstep2}.}
\vspace{5pt}
\subsubsection{Proof of (v):} Let $\psi:\mathbb{R}\rightarrow\mathbb{R}$ be any bounded Lipschitz function. The function $f(a,b,c) = \psi(a)$ is trivially $\rm{PL}$(2). Thus, by directly applying statement (iv) of the lemma, we find that
$$
\frac{1}{p}\sum_{i=1}^p{\psi(\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}({\vct{g}},\vct{h}))} \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[\psi(X_{\kappa,\sigma^2})\right].
$$
Since this holds for any bounded Lipschitz function, we have shown that the empirical distribution of ${\boldsymbol{\hat{\beta}}^{\rm{AO}}}_n$ converges weakly to the distribution of $X_{\kappa,\sigma^2}$. It remains to prove boundedness of the $2$nd moment as advertised in \eqref{eq:k_AO}. Recall from
\eqref{eq:betan} that
\begin{align}
\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}&=\frac{{\boldsymbol{\Sigma}_{i,i}^{-1/2}}}{1+(\xi_n\boldsymbol{\Sigma}_{i,i})^{-1}}\sqrt{\kappa}\tau_n\vct{h}_i +\left(1-\frac{1}{1+\xi_n\boldsymbol{\Sigma}_{i,i}}\right)(\sqrt{p}\betab^\star_i).\notag
\end{align}
Using this, boundedness of $\boldsymbol{\Sigma}_{i,i}$ from Assumption \ref{ass:mu}, and the fact that $\tau_n\stackrel{{P}}{\longrightarrow}\tau_\star, \xi_n\stackrel{{P}}{\longrightarrow}\xi_\star$, there exists constant $C=C(\Sigma_{\max},\Sigma_{\min},k,\tau_\star,\xi_\star)$ such that wpa 1,
\begin{align}
\frac{1}{p}\sum_{i=1}^p|\sqrt{p}{\boldsymbol{\hat{\beta}}^{\rm{AO}}}_{n,i}|^{2} \leq C\left(\frac{1}{p}\sum_{i=1}^p|\vct{h}_{i}|^{2}+\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\betab^\star_{i}|^{2}\right).\notag
\end{align}
But the two summands in the expression above are finite in the limit of $p\rightarrow\infty$. Specifically, (i) from Assumption \ref{ass:mu}, $\frac{1}{p}\sum_{i=1}^p|\sqrt{p}\betab^\star_{i}|^{2}\stackrel{{P}}{\longrightarrow}\operatorname{\mathbb{E}}[B^{2}]<\infty$; (ii) $\frac{1}{p}\sum_{i=1}^p|\vct{h}_{i}|^{2}\stackrel{{P}}{\longrightarrow}\operatorname{\mathbb{E}}[H^{2}]=1$, using the facts that $\vct{h}_i\stackrel{iid}{\sim}\mathcal{N}(0,1)$ and $H\sim\mathcal{N}(0,1)$. This proves \eqref{eq:k_AO}, as desired.
\section{Proof of Lemma \ref{lem rank one}}\label{SM lem proof}
\begin{lemma} [Lemma \ref{lem rank one} restated]Suppose $\mathcal{S}$ is drawn from an LGP$(\sigma,{\boldsymbol{{\Sigma}}},{\boldsymbol{\beta}}_\star)$ as in Def.~\ref{def LGP} where $\text{rank}({\boldsymbol{{\Sigma}}})=1$ with ${\boldsymbol{{\Sigma}}}=\boldsymbol{\lambda}\bla^T$ for $\boldsymbol{\lambda}\in\mathbb{R}^p$. Define $\zeta=\mathbb{T}_s(\boldsymbol{\lambda})^2/\tn{\boldsymbol{\lambda}}^2$. For magnitude and Hessian pruning ($P\in\{M,H\}$) and the associated retraining, we have the following excess risks with respect to ${\boldsymbol{\beta}}^\star
\begin{align
&\operatorname{\mathbb{E}}_{\mathcal{S}}[{\cal{L}}({\boldsymbol{\hat{\beta}}}_s^P)]-{\cal{L}}({\boldsymbol{\beta}}^\star)={\frac{\zeta^2\sigma^2}{n-2}}+\underbrace{(1-\zeta)^2(\boldsymbol{\lambda}^T{\boldsymbol{\beta}}^\star)^2}_{\text{Error due to bias}}\\
&\operatorname{\mathbb{E}}_{\mathcal{S}}[{\cal{L}}({\boldsymbol{\hat{\beta}}}_s^{RT})]-{\cal{L}}({\boldsymbol{\beta}}^\star)={\sigma^2}/({n-2}).
\end{align}
\end{lemma}
\begin{proof}
\noindent \textbf{Retraining analysis:} We claim that for any feature set $\Delta$ with $\boldsymbol{\lambda}_{\Delta}\neq 0$, the test risk of ${\boldsymbol{\hat{\beta}}}(\Delta)$ is exactly identical. Secondly, pruning is guaranteed to pick a nonzero support satisfying $\boldsymbol{\lambda}_{\Delta}\neq 0$ \footnote{This is because ${\boldsymbol{\hat{\beta}}}=c\boldsymbol{\lambda}$ for some scalar $c\neq 0$ as ${\boldsymbol{\hat{\beta}}}$ lies in the row space of ${\mtx{X}}$. Then, Hessian/Magnitude-pruning would pick a nonzero support of ${\boldsymbol{\hat{\beta}}}$ which corresponds to the nonzero support of $\boldsymbol{\lambda}$.}. Thus, as described next, retraining always achieves a fixed risk. Set $c^\star=\boldsymbol{\lambda}^T{\boldsymbol{\beta}}^\star$. By definition, each input example $\vct{x}_i$ has the form $\vct{x}_i=g_i\boldsymbol{\lambda}$ and $y_i=g_ic^\star+\sigma z_i$. Set ${\vct{g}}=[g_1~\dots~g_n]^T$ and $\bar{{\vct{g}}}={\vct{g}}/\tn{{\vct{g}}}$. Thus, we have ${\mtx{X}}={\vct{g}}\boldsymbol{\lambda}^T$ and $\vct{y}={\vct{g}}\boldsymbol{\lambda}^T{\boldsymbol{\beta}}^\star+\sigma{\vct{z}}$. Decompose ${\vct{z}}=\bar{{\vct{z}}}+\bar{{\vct{g}}}^T{\vct{z}}\bar{{\vct{g}}}$ where $\bar{{\vct{z}}}$ is orthogonal to ${\vct{g}}$. When solving the regression of $\Delta$, we have that
\[
{\mtx{X}}_\Delta={\vct{g}}\boldsymbol{\lambda}_\Delta^T,~\vct{y}=c^\star{\vct{g}}+\sigma(\bar{{\vct{z}}}+\bar{{\vct{g}}}^T{\vct{z}}\bar{{\vct{g}}})
\]
The least-squares solution has the form ${\boldsymbol{\hat{\beta}}}={\boldsymbol{\hat{\beta}}}(\Delta)=\hat{c}\boldsymbol{\lambda}_\Delta/\tn{\boldsymbol{\lambda}_\Delta}^2$ where
\begin{align}
&\hat{c}=\arg\min_{c}\tn{(c^\star-c){\vct{g}}+\sigma{\vct{z}}}\implies\hat{c}=c^\star+\sigma\gamma\label{beta}.
\end{align}
where $\gamma=\frac{\bar{{\vct{g}}}^T{\vct{z}}}{\tn{{\vct{g}}}}$. Observe that $\sqrt{p}\gamma$ has Student's t-distribution with $p$ degrees of freedom. Set $h$, $\epsilon\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,1)$. Now, observe that a fresh test sample with $y=\vct{x}^T{\boldsymbol{\beta}}^\star+\sigma \epsilon$ with $\vct{x}=g\boldsymbol{\lambda}$, the test error obeys
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}(\Delta))&=\operatorname{\mathbb{E}}(y-\vct{x}_{\Delta}^T{\boldsymbol{\hat{\beta}}}(\Delta))^2\\
&=\mathbb{E}[((c^\star-\hat{c})g+\sigma\epsilon)^2]\\
&=(\gamma^2+1)\sigma^2
\end{align}
Now, observe that the minimum risk is obviously ${\cal{L}}({\boldsymbol{\beta}}^\star)=\sigma^2$. Thus, the excess retraining risk becomes
\[
{\cal{L}}({\boldsymbol{\hat{\beta}}}(\Delta))-{\cal{L}}({\boldsymbol{\beta}}^\star)=\gamma^2\sigma^2.
\]
regardless of choice of $\Delta$. Finally, averaging this risk over $\mathcal{S}$ returns $\operatorname{\mathbb{E}}[\sigma^2\gamma^2]=\sigma^2/(n-2)$. Thus retraining risk has the fixed excess risk same as the one advertised in Lemma \ref{lem rank one}.
\noindent \textbf{Pruning analysis:} For pruning setting $\Delta=[p]$ above, we have that ${\boldsymbol{\hat{\beta}}}=\hat{c}\boldsymbol{\lambda}/\tn{\boldsymbol{\lambda}}^2$. This means that, for both Magnitude and Hessian pruning\footnote{They yield the same result since diagonal covariance is proportional to $\boldsymbol{\lambda}$ in magnitude.}, pruned vector takes the form ${\boldsymbol{\hat{\beta}}}_s=\hat{c}\mathbb{T}_s(\boldsymbol{\lambda})/\tn{\boldsymbol{\lambda}}^2$. Using the fact that $\boldsymbol{\lambda}^T\mathbb{T}_s(\boldsymbol{\lambda})=\tn{\mathbb{T}_s(\boldsymbol{\lambda})}^2=\zeta\tn{\boldsymbol{\lambda}}^2$, we find
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}_s)-{\cal{L}}({\boldsymbol{\beta}}^\star)&=\mathbb{E}[(c^\star- \hat{c}\frac{\boldsymbol{\lambda}^T\mathbb{T}_s(\boldsymbol{\lambda})}{\tn{\boldsymbol{\lambda}}^2})g+\sigma\epsilon)^2]-\sigma^2\\
&=\mathbb{E}[(c^\star- \hat{c}\zeta)g+\sigma\epsilon)^2]-\sigma^2\\
&=\operatorname{\mathbb{E}}[(((1-\zeta)c^\star-\zeta\sigma\gamma)g+\sigma\epsilon)^2]-\sigma^2\\
&=((1-\zeta)c^\star-\zeta\sigma\gamma)^2.
\end{align}
Finally, using zero-mean $\gamma$, we find
\[
\operatorname{\mathbb{E}}_{\mathcal{S}}[{\cal{L}}({\boldsymbol{\hat{\beta}}}_s)]-{\cal{L}}({\boldsymbol{\beta}}^\star)=\operatorname{\mathbb{E}}_{\mathcal{S}}[((1-\zeta)c^\star-\zeta\sigma\gamma)^2]=(1-\zeta)^2{c^\star}^2+\frac{\zeta^2\sigma^2}{n-2},
\]
which concludes the proof after observing ${c^\star}^2=(\boldsymbol{\lambda}^T{\boldsymbol{\beta}}^\star)^2$. Here, we call $(1-\zeta)^2{c^\star}^2$ ``the error due to bias''. The reason is that the predictable signal in the data is the noiseless component $\vct{x}^T{\boldsymbol{\beta}}^\star$. Pruning leads to an error in this predictable component by resulting in a biased estimate of the label (when conditioned on the random variable $g$ which controls the signal).
\cmt{
Now let $|\lambda_{k_1}|\geq|\lambda_{k_2}|\geq...\geq|\lambda_{k_p}|$. Assume we solved ERM to get $\hat{c}$ and ${\boldsymbol{\hat{\beta}}}=\hat{c}\boldsymbol{\lambda}$. Prune the trained model ${\boldsymbol{\hat{\beta}}}$ to $s$-sparse by keeping the largest entries. Set ${\vct{u}}=\mathbb{T}_s(\boldsymbol{\lambda})$ and $\zeta=\tn{{\vct{u}}}^2/\tn{\boldsymbol{\lambda}}^2$. We get ${\boldsymbol{\hat{\beta}}}_s=\hat{c}{\vct{u}}$.
\begin{align}
{\cal{L}}(\boldsymbol{\hat{\theta}})&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-{\vct{u}}^T\boldsymbol{\hat{\theta}})h+\sigma\epsilon)^2]\\
&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\sum_{i=1}^s\lambda_{k_i}^2\hat{c})h+\sigma\epsilon)^2]\\
&=[(1-\zeta)\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\zeta\sigma\gamma]^2+\sigma^2
\end{align}
\textbf{Case 1: }
Now, assume that we have already known the optimal entries and do pruning before training. Then the pruned model can be seen as pruning the inputs and then putting it through a non-sparse $s$-feature model. Let $\lambda_{t_1}{\boldsymbol{\beta}}_{t_1}\geq\lambda_{t_2}{\boldsymbol{\beta}}_{t_2}\geq...\geq\lambda_{t_s}{\boldsymbol{\beta}}_{t_s}$, $\vct{w}=[\lambda_{t_1}~\lambda_{t_2}~...~\lambda_{t_s}]$ and $\vct{t}=[{\boldsymbol{\beta}}_{t_1}~...~{\boldsymbol{\beta}}_{t_s}]$. Set data sample $\vct{x}_i'\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,{\boldsymbol{{\Sigma}}}')\in\mathbb{R}^s$ where are ${\boldsymbol{{\Sigma}}}'=\vct{w}\w^T$. Following what done in \ref{beta}, similarly
\[
\boldsymbol{\hat{\theta}}'=\theta\vct{w}\quad\text{where}\quad\theta=\frac{1}{\sum_{i=1}^s\lambda_{t_i}^2}(\vct{w}^T\vct{t}+\frac{\sigma}{\tn{{\vct{g}}}^2}{\vct{g}}^T{\vct{z}})
\]
\begin{align}
{\cal{L}}(\boldsymbol{\hat{\theta}}')&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\sum_{i=1}^s\lambda_{t_i}^2\theta)h+\sigma\epsilon)^2]\\
&=[(\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\vct{w}^T\vct{t})-\sigma\gamma]^2+\sigma^2
\end{align}
\textbf{Case 2: }
Now retrain with the pruned entries $\lambda_{k_!}$, ..., $\lambda_{k_s}$. Then the pruned model can be seen as pruning the inputs and then putting it through a non-sparse $s$-feature model. Let $\vct{t}=[{\boldsymbol{\beta}}_{k_1}~...~{\boldsymbol{\beta}}_{k_s}]$. Set data sample $\vct{x}_i'\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,{\boldsymbol{{\Sigma}}}')\in\mathbb{R}^s$ where are ${\boldsymbol{{\Sigma}}}'={\vct{u}}\ub^T$. Following what done in \ref{beta}, similarly
\[
\boldsymbol{\hat{\theta}}'=\theta\vct u\quad\text{where}\quad\theta=\frac{1}{\sum_{i=1}^s\lambda_{k_i}^2}(\vct u^T\vct{t}+\frac{\sigma}{\tn{{\vct{g}}}^2}{\vct{g}}^T{\vct{z}})
\]
\begin{align}
{\cal{L}}(\boldsymbol{\hat{\theta}}')&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\sum_{i=1}^s\lambda_{k_i}^2\theta)h+\sigma\epsilon)^2]\\
&=[(\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\vct u^T\vct{t})-\sigma\gamma]^2+\sigma^2
\end{align}
\textbf{It is obvious to know that loss in case 1 is always smaller than it in case 2. Then for both case, }
Set ${\boldsymbol{\beta}}^T{\boldsymbol{{\Sigma}}}{\boldsymbol{\beta}}=\alpha^2$ and $\vct{t}^T{\boldsymbol{{\Sigma}}}'\vct{t}=\rho^2$. We can write ${\cal{L}}(\boldsymbol{\hat{\theta}})$ and ${\cal{L}}(\boldsymbol{\hat{\theta}}')$ as
\[
{\cal{L}}(\boldsymbol{\hat{\theta}})=((1-\zeta)\alpha-\zeta\sigma\gamma)^2+\sigma^2\iff{\cal{L}}(\boldsymbol{\hat{\theta}}')=(\alpha-\rho-\sigma\gamma)^2+\sigma^2
\]
Now consider a special case which satisfies ${\cal{L}}(\boldsymbol{\hat{\theta}})<{\cal{L}}(\boldsymbol{\hat{\theta}}')$.
\[
0\leq(1-\zeta)\alpha-\zeta\sigma\gamma<\alpha-\rho-\sigma\gamma
\]
\[
\frac{\rho+\sigma\gamma}{\alpha+\sigma\gamma}<\zeta\leq\frac{\alpha}{\alpha+\sigma\gamma}
\]}
\end{proof}
\section{Main Results}\label{sec main}
\cmt{\begin{definition}[Linear Gaussian Problem (LGP)] \label{def LGP}Given latent vector ${\boldsymbol{\beta}}^\star\in\mathbb{R}^d$, covariance ${\boldsymbol{{\Sigma}}}$ and noise level $\sigma$, suppose each example in $\mathcal{S}$ is generated independently as $y_i=\vct{x}_i^T{\boldsymbol{\beta}}^\star+ z_i$ where $z_i\sim\mathcal{N}(0,\sigma^2)$ and $\vct{x}_i\sim\mathcal{N}(0,{\boldsymbol{{\Sigma}}})$. Additionally, the map $\phi(\cdot)$ is identity and $p=d$.
\end{definition}}
\cmt{Given an ERM problem \eqref{erm} with nonlinear map $\phi$, we will relate it to an equivalent LGP to characterize its properties. The equivalence is ensured by setting up LGP to exhibit similar second order characteristics as the original problem.
\begin{definition}[Equivalent Linear Problem] Given distribution $(\vct{x},y)$ $\sim\mathcal{D}$, the equivalent LGP(${\boldsymbol{\beta}},{\boldsymbol{{\Sigma}}},\sigma$) with $n$ samples is given with parameters ${\boldsymbol{{\Sigma}}}=\operatorname{\mathbb{E}}[\vct{x}\x^T]$, ${\boldsymbol{\beta}}^\star={\boldsymbol{{\Sigma}}}^{-1}\operatorname{\mathbb{E}}[y\vct{x}]$ and $\sigma=\operatorname{\mathbb{E}}[(y-\vct{x}^T{\boldsymbol{\beta}}^\star)^2]^{1/2}$.
\end{definition}
Our key theoretical contribution is formalizing the DC of the LGPs which also enables us to characterize pruning/retraining dynamics. We then empirically verify that DC and pruning dynamics of equivalent LGPs can successfully predict the original problem which can be random feature regression. In the context of high-dimensional learning literature, equivalent LGP is known to be a useful proxy for predicting the performance of the original ERM task \cite{}. This work goes beyond the existing works by demonstrating it can also successfully predict pruning/retraining dynamics. Formalizing this connection is an important future research avenue and it can presumably be accomplished by building on the recent high-dimensional universality results such as \cite{}.}
\cmt{
\noindent\textbf{Covariance diagonalization:} Let ${\boldsymbol{{\Sigma}}}$ have eigenvalue decomposition ${\mtx{U}}\bar{{\boldsymbol{{\Sigma}}}}{\mtx{U}}^T$. Let ${\boldsymbol{\hat{\beta}}}$ be the minimizer of empirical risk. Observe that ${\mtx{X}}=\bar{{\mtx{X}}}{\mtx{U}}^T$ where $\bar{{\mtx{X}}}$ has features with diagonal covariance $\bar{{\boldsymbol{{\Sigma}}}}$. Let ${\hat{\bar{\boldsymbol{\beta}}}}$ be the solution of ERM for $(\vct{y},\bar{{\mtx{X}}})$ pair. Then ${\boldsymbol{\hat{\beta}}}={\mtx{U}}{\hat{\bar{\boldsymbol{\beta}}}}$.
\begin{definition}[Linear model] \label{d model}Fix a positive semidefinite covariance matrix ${\boldsymbol{{\Sigma}}}\in\mathbb{R}^{p\times p}$ and latent parameter ${\boldsymbol{\beta}}\in\mathbb{R}^p$ satisfying ${\boldsymbol{\beta}}^T{\boldsymbol{{\Sigma}}}{\boldsymbol{\beta}}=\alpha$. Suppose $(\vct{x}_i)_{i=1}^n\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,{\boldsymbol{{\Sigma}}})$. The samples $(\vct{x}_i,y_i)_{i=1}^n$ are generated as
\[
y_i=\vct{x}_i^T{\boldsymbol{\beta}}+\sigma z_i,
\]
where $\sigma^2$ is the noise variance and $(z_i)_{i=1}^n$ is the noise sequence.
\end{definition}
Declare the first $s$ entries of ${\boldsymbol{\beta}}$ to be $\vct{t}=\mathbb{F}_s({\boldsymbol{\beta}})$. We estimate $\vct{t}$ via
\begin{align}
\boldsymbol{\hat{\theta}}=\mathbb{F}_s({\boldsymbol{\hat{\beta}}})\quad\text{where}\quad {\boldsymbol{\hat{\beta}}}=\arg\min_{{\boldsymbol{\beta}}'} \tn{\vct{y}-{\mtx{X}}{\boldsymbol{\beta}}'}.\label{optim me}
\end{align}
If the problem is over-parameterized ($p>n$), there are infinitely many feasible solutions ${\boldsymbol{\hat{\beta}}}$ achieving zero empirical loss. In this case, we investigate the minimum $\ell_2$ norm solution which is pseudo-inverse given by
\begin{align}
\boldsymbol{\hat{\theta}}=\mathbb{F}_s({\boldsymbol{\hat{\beta}}})\quad\text{where}\quad {\boldsymbol{\hat{\beta}}}=\arg\min_{{\boldsymbol{\beta}}'} \tn{{\boldsymbol{\beta}}'}\quad\text{subject to}\quad \vct{y}={\mtx{X}}{\boldsymbol{\beta}}'.\label{optim me2}
\end{align}
For the following discussion we set ${\mtx{X}}={\mtx{\bar{X}}}\sqrt{\bSi}$ so that ${\mtx{\bar{X}}}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$.}
Here, we present our main theoretical result: a sharp asymptotic characterization of the distribution of the solution to overparameterized least-squares for correlated designs. We further show how this leads to a sharp prediction of the risk of magnitude-based pruning.
Concretely, for the rest of this section, we assume the linear Gaussian problem (LGP) of Definition \ref{def LGP}, the overparameterized regime $k=p>n$ and the min-norm model
\begin{align}
\hat{\boldsymbol{\beta}} = \arg\min_{\boldsymbol{\beta}} \tn{\boldsymbol{\beta}} ~\text{s.t.}~\vct{y}={\mtx{X}}{\boldsymbol{\beta}}.\label{eq:min_norm}
\end{align}
As mentioned in Sec.~\ref{sec:setup}, $\hat{\boldsymbol{\beta}}$ is actually given in closed-form as ${\boldsymbol{\hat{\beta}}}={\mtx{X}}^\dagger \vct{y}$. Interestingly, our analysis of the distribution of ${\boldsymbol{\hat{\beta}}}$ does not rely on the closed-form expression, but rather follows by viewing ${\boldsymbol{\hat{\beta}}}$ as the solution to the convex linearly-constrained quadratic program in \eqref{eq:min_norm}. Specifically, our analysis uses the framework of the convex Gaussian min-max Theorem (CGMT) \cite{thrampoulidis2015regularized}, which allows to study rather general inference optimization problems such as the one in \eqref{eq:min_norm}, by relating them with an auxiliary optimization that is simpler to analyze \cite{StoLASSO,oymak2013squared,thrampoulidis2015regularized,thrampoulidis2018precise,salehi2019impact,taheri2020fundamental}.
Due to space considerations, we focus here on the more challenging overparameterized regime and defer the analysis of the underparameterized regime to the SM.
\subsection{Distributional Characterization of the Overparameterized Linear Gaussian Models}
\noindent\emph{Notation:} We first introduce additional notation necessary to state our theoretical results. $\odot$ denotes the entrywise product of two vectors and ${\mathbf{1}}_p$ is the all ones vector in $\mathbb{R}^p$.
The \emph{empirical distribution} of a vector $\vct{x}\in\mathbb{R}^p$ is given by
$ \frac{1}{p}\sum_{i=1}^p \delta_{x_i}$, where $\delta_{x_i}$ denotes a Dirac delta mass on $x_i$. Similarly, the empirical joint distribution of vectors $\vct{x}, \vct{x}'\in \mathbb{R}^p$ is $\frac{1}{p} \sum_{i=1}^p \delta_{(x_i, x'_i)}$. The \emph{Wasserstein-$k$} ($W_k$) distance between two measures $\mu$ and $\nu$ is defined as
$W_k(\mu,\nu) \equiv \left(\inf_{\rho}\operatorname{\mathbb{E}}_{(X,Y)\sim \rho}|X-Y|^k\right)^{1/k},$
where the infimum is over all the couplings of $\mu$ and $\nu$, i.e. all random variables $(X,Y)$ such that $X\sim\mu$ and $Y\sim\nu$ marginally.
A sequence of probability distributions $\nu_p$ on $\mathbb{R}^m$ {converges in $W_k$} to
$\nu$, written $\nu_p\stackrel{W_k}{\Longrightarrow} \nu$, if $W_k(\nu_p,\nu) \rightarrow 0$ as $p \rightarrow \infty$. Finally, we say that a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$, denoted $f\in\rm{PL}(k)$, if there is a constant $L>0$ such that for all $\vct{x},\vct{y}\in\mathbb{R}^m$,
$
|f(\vct{x}) - f(\vct{y})|\leq L(1+\tn{\vct{x}}^{k-1}+\tn{\vct{y}}^{k-1})\|\vct{x}-\vct{y}\|_2.
$
\fy{We call $L$ the $\rm{PL}(k)$ constant of $f$.}
An equivalent definition of $W_k$ convergence is that, for any $f\in\rm{PL}(k)$, $\lim_{p\rightarrow\infty}\operatorname{\mathbb{E}} f(X_p)=\operatorname{\mathbb{E}} f(X)$, where expectation is with respect to $X_p\sim\nu_p$ and $X\sim\nu$. For a sequence of random variables $\mathcal{X}_{p}$ that converge in probability to some constant $c$ in the limit of Assumption \ref{ass:linear} below, we write $\mathcal{X}_{p}\stackrel{{P}}{\longrightarrow} c$.
\vspace{3pt}
Next, we formalize the set of assumption under which our analysis applies. Our asymptotic results hold in the linear asymptotic regime specified below.
\begin{assumption}\label{ass:linear} We focus on a double asymptotic regime where $n,p,s\rightarrow\infty$ at fixed overparameterization ratio $\kappa:=p/n>0$ and sparsity level $\alpha:=s/p\in(0,1)$.
\end{assumption}
Additionally, we require certain mild assumptions on the behavior of the covariance matrix $\boldsymbol{\Sigma}$ and of the true latent vector ${\boldsymbol{\beta}}^\star$. For simplicity, we assume here that $\boldsymbol{\Sigma} = \diag{[\boldsymbol{\Sigma}_{1,1},\ldots,\boldsymbol{\Sigma}_{p,p}]}$
\begin{restatable}{assumption}{asstwo}\label{ass:inv} {The covariance matrix $\boldsymbol{\Sigma}$ is diagonal} and there exist constants $\Sigma_{\min},\Sigma_{\max}\in(0,\infty)$ such that:
$
\Sigma_{\min}\leq{\boldsymbol{{\Sigma}}}_{i,i}\leq \Sigma_{\max},
$
for all $i\in[p].$
\end{restatable}
\begin{restatable}{assumption}{assthree}\label{ass:mu}
The joint empirical distribution of $\{(\bSi_{i,i},\sqrt{p}\betab^\star_i)\}_{i\in[p]}$ converges in {Wasserstein-k} distance to a probability distribution $\mu$ on $\mathbb{R}_{>0}\times\mathbb{R}$ {for some {$k\geq 4$}}. That is
$
\frac{1}{p}\sum_{i\in[p]}\delta_{(\bSi_{i,i},\sqrt{p}\betab^\star_i)} \stackrel{W_k}{\Longrightarrow} \mu.
$
\end{restatable}
With these, we are ready to define, what will turn out to be, the asymptotic DC in the overparameterized regime.
\begin{definition}[Asymptotic DC -- Overparameterized regime]\label{def:Xi}
Let random variables $(\Lambda,B)\sim \mu$ (where $\mu$ is defined in Assumption \ref{ass:mu}) and fix $\kappa>1$. Define parameter $\xi$ as the unique positive solution to the following equation
\begin{align}\label{eq:ksi}
\operatorname{\mathbb{E}}_{\mu}\Big[ \big({1+(\xi\cdot\Lambda)^{-1}}\big)^{-1} \Big] = {\kappa^{-1}}\,.
\end{align}
Further define the positive parameter $\gamma$ as follows:
\begin{align}
\label{eq:gamma}
\hspace{-0.1in}\gamma :=
\Big({\sigma^2 + \operatorname{\mathbb{E}}_{\mu}\Big[\frac{B^2\Lambda}{(1+\xi\Lambda)^2}\Big]}\Big)\Big/\Big({1-\kappa\operatorname{\mathbb{E}}_{\mu}\Big[\frac{1}{\left(1+(\xi\Lambda)^{-1}\right)^2}\Big]}\Big).
\end{align}
With these and $H\sim\mathcal{N}(0,1)$, define the random variable
\begin{align}
X_{\kappa,\sigma^2}(\Lambda,B,H) := \Big(1-\frac{1}{1+ \xi\Lambda}\Big) B + \sqrt{\kappa}\frac{\sqrt{\gamma}\,\Lambda^{-1/2}}{1+(\xi\Lambda)^{-1}} H, \label{eq:X}
\end{align}
and let $\Pi_{\kappa,\sigma^2}$ be its distribution.
\end{definition}
\fy{Our main result establishes asymptotic convergence of the empirical distribution of $(\sqrt{p}{\boldsymbol{\hat{\beta}}},\sqrt{p}\betab^\star,\boldsymbol{\Sigma})$ for a rich class of test functions. These are the functions within $\rm{PL}(3)$ that become $\rm{PL}(2)$ when restricted to the first two indices. Formally, we define this class of functions as follows
\begin{align}\label{eq:pdef}
{\cal{F}} :=\{&f:\mathbb{R}^2\times {\cal{Z}}\rightarrow\mathbb{R},~f\in \rm{PL}(3)~\text{and}~\\
&\sup_{z\in{\cal{Z}}}\text{``}\rm{PL}(2)~\text{constant~of}~f(\cdot,\cdot,z)\text{''}<\infty\}.\notag
\end{align}
For pruning analysis, we set ${\cal{Z}}=[\Sigma_{\min},\Sigma_{\max}]$ and define
\begin{align}\label{eq:fdef}
{\cal{F}}_{\cal{L}}:= \{ f:\mathbb{R}^2\times {\cal{Z}}\rightarrow\mathbb{R}~\big|~&f(x,y,z) = z(y-g(x))^2\notag\\
&~\text{where}~g(\cdot)~\text{is Lipschitz}\}.
\end{align}
As discussed below, ${\cal{F}}_{\cal{L}}$ is important for predicting the risk of the (pruned) model. In the SM, we prove that ${\cal{F}}_{\cal{L}}\subset {\cal{F}}$. We are now ready to state our main theoretical result.
\cts{
\begin{restatable}[Asymptotic DC -- Overparameterized LGP]{theorem}{mainthm}\label{thm:master_W2}
Fix $\kappa>1$ and suppose Assumptions \ref{ass:inv} and \ref{ass:mu} hold. Recall the solution ${\boldsymbol{\hat{\beta}}}$ from \eqref{eq:min_norm} and let
\[
\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p}\sum_{i=1}^{p}\delta_{(\sqrt{p}{\boldsymbol{\hat{\beta}}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i})}
\] be the joint empirical distribution of $(\sqrt{p}{\boldsymbol{\hat{\beta}}},\sqrt{p}\betab^\star,\boldsymbol{\Sigma})$. Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be a function in ${\cal{F}}$ defined in \eqref{eq:pdef}. We have that
\begin{align}\label{eq:thm}
\hspace{-0.1in}\frac{1}{p} \sum_{i=1}^{p} f(\sqrt{p}\hat\boldsymbol{\beta}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[f(X_{\kappa,\sigma^2},B,\Lambda) \right].
\end{align}
\end{restatable}
}
As advertised, Theorem \ref{thm:master_W2} fully characterizes the joint empirical distribution of the min-norm solution, the latent vector and the covariance spectrum. The asymptotic DC allows us to precisely characterize several quantities of interest, such as estimation error, generalization error etc.. For example, a direct application of \eqref{eq:thm} to the function $f(x,y,z)=z(y-x)^2\in {\cal{F}}_{\cal{L}}\subset{\cal{F}}$ directly yields the risk prediction of the min-norm solution recovering \cite[Thm.~3]{hastie2019surprises} as a special case. Later in this section, we show how to use Theorem \ref{thm:master_W2} towards the more challenging task of precisely characterizing the risk of magnitude-based pruning.
Before that, let us quickly remark on the technical novelty of the theorem. Prior work has mostly applied the CGMT to isotropic features. Out of these, only very few obtain DC, \cite{thrampoulidis2018symbol,miolane2018distribution}, while the majority focuses on simpler metrics, such as squared-error. Instead, Theorem \ref{thm:master_W2} considers correlated designs and the overparameterized regime. The most closely related work in that respect is \cite{montanari2019generalization}, which very recently obtained the DC of the max-margin classifier. Similar to us, they use the CGMT, but their analysis of the auxiliary optimization is technically different to ours. Our approach is similar to \cite{thrampoulidis2018symbol}, but extra technical effort is needed to account for correlated designs and the overparameterized regime.
\subsection{From DC to Risk Characterization}\label{sec:risk}
First, we consider a simpler ``threshold-based" pruning method that applies a fixed threshold at every entry of $\hat\boldsymbol{\beta}$. Next, we relate this to magnitude-based pruning and obtain a characterization for the performance of the latter.
In order to define the threshold-based pruning vector, let \[
\mathcal{T}_t(x) = \begin{cases} x & \text{if } |x|>t \\ 0 & \text{otherwise} \end{cases},
\] be the hard-thresholding function with fixed threshold $t\in\mathbb{R}_+$. Define ${\boldsymbol{\hat{\beta}}}^\mathcal{T}_{t} := \mathcal{T}_{t/\sqrt{p}}(\hat{\boldsymbol{\beta}})$, where $\mathcal{T}_t$ acts component-wise. Then, the population risk of ${\boldsymbol{\hat{\beta}}}^\mathcal{T}_t$ becomes
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^\mathcal{T}_t) &= \operatorname{\mathbb{E}}_{\mathcal{D}}[(\vct{x}^T({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^\mathcal{T}_t) + \sigma z)^2] \notag\\
\cmt{&= \sigma^2 + ({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^\mathcal{T}_t)^T\boldsymbol{\Sigma}({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^\mathcal{T}_t) \notag \\}
&= \sigma^2 + \frac{1}{p}\sum_{i=1}^p\boldsymbol{\Sigma}_{i,i}\big(\sqrt{p}\betab^\star_i-\mathcal{T}_{t}(\sqrt{p}\hat\boldsymbol{\beta}_i)\big)^2\notag\\
&\stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}\left[ \Lambda\left(B-\mathcal{T}_t(X_{\kappa,\sigma^2})\right)^2 \right]\,.
\label{eq:loss_w2}
\end{align}
In the second line above, we note that $\sqrt{p}\mathcal{T}_{t'}(x)=\mathcal{T}_{\sqrt{p}t'}(\sqrt{p}x)$. In the last line, we apply \eqref{eq:thm}, after recognizing that the function $(x,y,z)\mapsto z(y-\mathcal{T}_t(x))^2$
is a member of the ${\cal{F}}_{\cal{L}}$ family defined in \eqref{eq:fdef}.
As in \eqref{eq:thm}, the expectation here is with respect to $(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)$.
Now, we show how to use \eqref{eq:loss_w2} and Theorem \ref{thm:master_W2} to characterize the risk of the magnitude-based pruned vector $\boldsymbol{\beta}^M_s:=\mathbb{T}_s({\boldsymbol{\hat{\beta}}})$. Recall, here from Assumption \ref{ass:linear} that $s=\alpha p$. To relate $\hat{\boldsymbol{\beta}}^M_s$ to ${\boldsymbol{\hat{\beta}}}^\mathcal{T}_t$, consider the set $\mathcal{S}_t:=\{i\in[p]\,:\,\sqrt{p}|\hat\boldsymbol{\beta}_i| \geq t \}$ for some constant $t\in\mathbb{R}_+$ (not scaling with $n,p,s$). Note that the ratio ${|\mathcal{S}_t|}/{p}$ is equal to
\begin{align}
p^{-1}\sum_{i=1}^p\mathbb{1}_{[\sqrt{p}|\hat\boldsymbol{\beta}_i| \geq t]} \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}[\mathbb{1}_{[|X_{\kappa,\sigma^2}|\geq t]}] = \P\left(|X_{\kappa,\sigma^2}|\geq t\right).\notag
\end{align}
Here, $\mathbb{1}$ denotes the indicator function and the convergence follows from Theorem \ref{thm:master_W2} when applied to a sequence of bounded Lipschitz functions approximating the indicator. Thus, by choosing
\begin{align}\label{eq:tstar}
t^\star:=\sup\left\{t\in\mathbb{R}\,:\, \P(|X_{\kappa,\sigma^2}|\geq t) \geq \alpha \right\},
\end{align}
it holds that ${|\mathcal{S}_t|}/{p}\stackrel{{P}}{\longrightarrow}\alpha$. In words, and {observing that $X_{\kappa,\sigma^2}$ admits a continuous density (due to the Gaussian variable $H$)}: for any $\varepsilon>0$, in the limit of $n,p,s\rightarrow\infty$, the vector ${\boldsymbol{\hat{\beta}}}^\mathcal{T}_{t^\star}$ has $(1\pm\varepsilon) \alpha p= (1\pm\varepsilon) s$ non-zero entries, which correspond to the largest magnitude entries of ${\boldsymbol{\hat{\beta}}}$, with probability approaching1. Since this holds for arbitrarily small $\varepsilon>0$, recalling $t^\star$ as in \eqref{eq:tstar}, we can conclude from \eqref{eq:loss_w2} that the risk of the magnitude-pruned model converges as follows.
{\begin{restatable}[Risk of Magnitude-pruning]{corollary}{cormag}\label{cor:mag}
Let the same assumptions and notation as in the statement of Theorem \ref{thm:master_W2} hold. Specifically, let $\hat\boldsymbol{\beta}$ be the min-norm solution in \eqref{eq:min_norm} and ${\boldsymbol{\hat{\beta}}}_s^M:=\mathbb{T}_s({\boldsymbol{\beta}})$ the magnitude-pruned model at sparsity $s$. Recall the threshold $t^\star$ from \eqref{eq:tstar}. The risk of ${\boldsymbol{\hat{\beta}}}^M_s$ satisfies the following in the limit of $n,p,s\rightarrow\infty$ at rates $\kappa:=p/n>1$ and $\alpha:=s/p\in(0,1)$ (cf. Assumption \ref{ass:linear}):
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^M_s) \stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}\left[ \Lambda\left(B-\mathcal{T}_{t^\star}(X_{\kappa,\sigma^2})\right)^2 \right],\notag
\end{align}
where the expectation is over $(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1).$
\end{restatable}}
\cmt{LET US NOT FORGET TO ADD OR COMMENT ABOUT UNDERPARAM!}
\subsection{Non-asymptotic DC and Retraining Formula}
While Theorem \ref{thm:master_W2} is stated in the asymptotic regime, during analysis, the DC arises in a non-asymptotic fashion. The following definition is the non-asymptotic counterpart of Def.~\ref{def:Xi}. We remark that this definition applies to arbitrary covariance (not necessarily diagonal) by applying a simple eigen-rotation before and after the DC formula associated with the diagonalized covariance.
\cmt{
\begin{definition}[Non-asymptotic DC] \label{aux_def}Fix $p>n\geq 1$ and set $\kappa=p/n>1$. Given $\sigma>0$, covariance ${\boldsymbol{{\Sigma}}}={\mtx{U}}\text{diag}(\boldsymbol{\lambda}){\mtx{U}}^T$ and latent vector ${\boldsymbol{\beta}}$, set $\bar{{\boldsymbol{\beta}}}={\mtx{U}}^T{\boldsymbol{\beta}}$ and define the unique non-negative terms $\Xi,\Gamma,\boldsymbol{\zeta}\in\mathbb{R}^p$ and $\boldsymbol{\phi}\in\mathbb{R}^p$ as follows\vspace{-0pt}
\begin{align}
&\Xi>0\quad\text{is the solution of}\quad 1=\frac{\kappa}{p}\sum_{i=1}^p\frac{1}{1+(\Xi\boldsymbol{\lambda}_i)^{-1}},\notag\\
&\Gamma=\frac{\sigma^2+\sum_{i=1}^p\boldsymbol{\lambda}_i\boldsymbol{\zeta}_i^2\bar{{\boldsymbol{\beta}}}_i^2}{\kappa(1-\frac{\kappa}{p}\sum_{i=1}^p{(1+(\Xi\boldsymbol{\lambda}_i)^{-1})^{-2}})},\notag\\
&\boldsymbol{\zeta}_i=\frac{1}{1+\Xi\boldsymbol{\lambda}_i}\quad\text{and}\quad \boldsymbol{\phi}_i=\frac{\kappa\sqrt{\Gamma}}{1+(\Xi\boldsymbol{\lambda}_i)^{-1}}\quad\text{for}\quad 1\leq i\leq p.\notag
\end{align}
The non-asymptotic OP-LS distribution is defined as the following ${\mtx{U}}$-rotated normal distribution
\[
\mathcal{D}_{\sigma,{\boldsymbol{{\Sigma}}},{\boldsymbol{\beta}}}={\mtx{U}}\mathcal{N}(({\mathbf{1}}_p-\boldsymbol{\zeta})\odot\bar{{\boldsymbol{\beta}}},p^{-1}\text{diag}(\boldsymbol{\lambda}^{-1}\odot\boldsymbol{\phi}^2)).
\]
\end{definition}}
\begin{definition}[Non-asymptotic DC] \label{aux_def}Fix $p>n\geq 1$ and set $\kappa=p/n>1$. Given $\sigma>0$, covariance ${\boldsymbol{{\Sigma}}}={\mtx{U}}\text{diag}(\boldsymbol{\lambda}){\mtx{U}}^T$ and latent vector ${\boldsymbol{\beta}}$, set $\bar{{\boldsymbol{\beta}}}={\mtx{U}}^T{\boldsymbol{\beta}}$ and define the unique non-negative terms $\xi,\gamma,\boldsymbol{\zeta}\in\mathbb{R}^p$ and $\boldsymbol{\phi}\in\mathbb{R}^p$ as follows:\vspace{-0pt}
\begin{align}
&\xi>0\quad\text{is the solution of}\quad \kappa^{-1}={p^{-1}}\sum_{i=1}^p\big({1+(\xi\boldsymbol{\lambda}_i)^{-1}}\big)^{-1},\notag\\
&\gamma=\frac{\sigma^2+\sum_{i=1}^p\boldsymbol{\lambda}_i\boldsymbol{\zeta}_i^2\bar{{\boldsymbol{\beta}}}_i^2}{1-\frac{\kappa}{p}\sum_{i=1}^p{(1+(\xi\boldsymbol{\lambda}_i)^{-1})^{-2}}},\notag\\
&\boldsymbol{\zeta}_i=({1+\xi\boldsymbol{\lambda}_i})^{-1}\quad\text{,}\quad \boldsymbol{\phi}_i={\kappa\gamma}(1+(\xi\boldsymbol{\lambda}_i)^{-1})^{-2},~ 1\leq i\leq p.\notag
\end{align}
The non-asymptotic distributional prediction is given by the following ${\mtx{U}}$-rotated normal distribution
\[
\mathcal{D}_{\sigma,{\boldsymbol{{\Sigma}}},{\boldsymbol{\beta}}}={\mtx{U}}\mathcal{N}(({\mathbf{1}}_p-\boldsymbol{\zeta})\odot\bar{{\boldsymbol{\beta}}},p^{-1}\text{diag}(\boldsymbol{\lambda}^{-1}\odot\boldsymbol{\phi})).
\]
\end{definition}
{We remark that this definition is similar in spirit to the concurrent/recent work \cite{li2020exploring}. However, unlike this work, here we prove the asymptotic correctness of the DC, we use it to rigorously predict the pruning performance and also extend this to retraining DC as discussed next.}
\noindent \textbf{Retraining DC.} As the next step, we would like to characterize the DC of the solution after retraining, i.e.,~${\boldsymbol{\hat{\beta}}}^{RT}$. We carry out the retraining derivation (for magnitude pruning) as follows. Let ${\mathcal{I}}\subset[p]$ be the nonzero support of the pruned vector ${\boldsymbol{\hat{\beta}}}^M_s$. Re-solving \eqref{eq:ERM} restricted to the features over ${\mathcal{I}}$ corresponds to a linear problem with effective feature covariance ${\boldsymbol{{\Sigma}}}_{\mathcal{I}}$ with support of non-zeros restricted to ${\mathcal{I}}\times {\mathcal{I}}$. For this feature covariance, we can also calculate the effective noise level and global minima of the population risk ${\boldsymbol{\beta}}^\star_{\mathcal{I}}$. The latter has the closed-form solution ${\boldsymbol{\beta}}^\star_{\mathcal{I}}={\boldsymbol{{\Sigma}}}_{\mathcal{I}}^\dagger{\boldsymbol{{\Sigma}}}{\boldsymbol{\beta}}^\star$. The effective noise is given by accounting for the risk change due to the missing features via $\sigma_{\mathcal{I}}=(\sigma^2+{{\boldsymbol{\beta}}^\star}^T{\boldsymbol{{\Sigma}}}{\boldsymbol{\beta}}^\star-{{\boldsymbol{\beta}}^\star_{\mathcal{I}}}^T{\boldsymbol{{\Sigma}}}_{\mathcal{I}}{\boldsymbol{\beta}}^\star_{\mathcal{I}})^{1/2}$. With these terms in place, fixing ${\mathcal{I}}$ and using Def.~\ref{aux_def}, the retraining prediction becomes $\mathcal{D}_{\sigma_{{\mathcal{I}}},{\boldsymbol{{\Sigma}}}_{{\mathcal{I}}},\boldsymbol{\beta}^\star_{\mathcal{I}}}$. This process is summarized below.
\begin{definition}[Retraining DC] \label{RT_def}Consider the setting of Def.~\ref{aux_def} with $\sigma,{\boldsymbol{{\Sigma}}},{\boldsymbol{\beta}}^\star$ and sparsity target $s$. The sample ${\boldsymbol{\hat{\beta}}}^{RT}$ from the retraining distribution $\mathcal{D}^{\text{RT,s}}_{\sigma,{\boldsymbol{{\Sigma}}},{\boldsymbol{\beta}}^\star}$ is constructed as follows. Sample ${\boldsymbol{\hat{\beta}}}\sim \mathcal{D}_{\sigma,{\boldsymbol{{\Sigma}}},{\boldsymbol{\beta}}^\star}$ and compute the set of the top-$s$ indices ${\mathcal{I}}={\mathcal{I}}(\mathbb{T}_s({\boldsymbol{\hat{\beta}}}))$. Given ${\mathcal{I}}$, obtain the effective covariance ${\boldsymbol{{\Sigma}}}_{\mathcal{I}}\in\mathbb{R}^{p\times p}$, population minima ${\boldsymbol{\beta}}^\star_{\mathcal{I}}\in\mathbb{R}^p$, and the noise level $\sigma_{{\mathcal{I}}}>0$ as described above. Draw ${\boldsymbol{\hat{\beta}}}^{RT}\sim\mathcal{D}_{\sigma_{{\mathcal{I}}},{\boldsymbol{{\Sigma}}}_{{\mathcal{I}}},\boldsymbol{\beta}^\star_{\mathcal{I}}}$.
\cmt{Obtain ${\boldsymbol{{\Sigma}}}_{\mathcal{I}}\in\mathbb{R}^{p\times p}$ by restricting the nonzero support of ${\boldsymbol{{\Sigma}}}$ to ${\mathcal{I}}\times {\mathcal{I}}$. Set ${\mathcal{I}}$ restricted population minima ${\boldsymbol{\beta}}^\star_{\mathcal{I}}={\boldsymbol{{\Sigma}}}_{\mathcal{I}}^\dagger{\boldsymbol{{\Sigma}}}{\boldsymbol{\beta}}^\star$ and set the new noise level $\sigma_{{\mathcal{I}}}=(\sigma^2+{{\boldsymbol{\beta}}^\star}^T{\boldsymbol{{\Sigma}}}{\boldsymbol{\beta}}^\star-{{\boldsymbol{\beta}}^\star_{\mathcal{I}}}^T{\boldsymbol{{\Sigma}}}_{\mathcal{I}}{\boldsymbol{\beta}}^\star_{\mathcal{I}})^{1/2}$. Draw ${\boldsymbol{\hat{\beta}}}^{RT}\sim\mathcal{D}_{\sigma_{{\mathcal{I}}},{\boldsymbol{{\Sigma}}}_{{\mathcal{I}}},{\boldsymbol{\beta}}_{\mathcal{I}}}$. }
\end{definition}
Observe that, the support ${\mathcal{I}}$ depends on the samples $\mathcal{S}$ via ${\boldsymbol{\hat{\beta}}}$. Thus, our retraining DC is actually derived for the scenario when the retraining phase uses a fresh set of $n$ samples to break the dependence between ${\mathcal{I}},\mathcal{S}$ (which obtains ${\boldsymbol{\hat{\beta}}}^{RT}_{\text{fresh}}$). Despite this, we empirically observe that the retraining DC predicts the regular retraining (reusing $\mathcal{S}$) performance remarkably well and perfectly predicts ${\boldsymbol{\hat{\beta}}}^{RT}_{\text{fresh}}$ as discussed in Figs.~\ref{figRF} and \ref{figRF2}. Finally, we defer the formalization of the retraining analysis to a future work. This includes proving that ${\boldsymbol{\hat{\beta}}}^{RT}_{\text{fresh}}$ obeys Def.~\ref{RT_def} asymptotically as well as directly studying ${\boldsymbol{\hat{\beta}}}^{RT}$ by capturing the impact of the ${\mathcal{I}},\mathcal{S}$ dependency.
\subsection{Overdetermined Analysis}
\begin{theorem}[Overdetermined estimation] Consider a linear model described in Definition \ref{d model}. Assume $\kappa=p/n<1$. Let $\vct{h}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,{\mtx{I}}_p)$ and $\bar{\h}=\vct{h}/\tn{\vct{h}}$. Define
\[
{\boldsymbol{\hat{\beta}}}_{AO}={\boldsymbol{\beta}}+w_*{\boldsymbol{{\Sigma}}}^{-1/2}\bar{\h}\qquad\text{and}\qquad w_*=\sqrt{\frac{\kappa}{1-\kappa}}\sigma
\]
Then, for any $1$-Lipschitz function $F$, with probability $1-C\mathrm{e}^{-c\varepsilon^2p}$, we have that
\[
|\tn{F({\boldsymbol{\hat{\beta}}})}-\operatorname{\mathbb{E}}[\tn{F({\boldsymbol{\hat{\beta}}}_{AO})}]|\leq \varepsilon
\]
\end{theorem}\som{I think we can do a change of variable to fully get rid of the role of ${\boldsymbol{{\Sigma}}}$}
\begin{proof}
Fix $h\in\mathbb{R},{\vct{g}}\in\mathbb{R}^n,\vct{h}\in\mathbb{R}^p,{\mtx{\bar{X}}}\in\mathbb{R}^{n\times p}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$. We are interested in ordinary least-square solution ${\boldsymbol{\hat{\beta}}}=\arg\min_{{\boldsymbol{\beta}}}{\tn{\vct{y}-{\mtx{\bar{X}}}\sqrt{{\boldsymbol{{\Sigma}}}}{\boldsymbol{\beta}}}}$. Set $\vct{w}={\boldsymbol{\beta}}-{\boldsymbol{\hat{\beta}}}$, ${\mtx{X}}'=[{\mtx{\bar{X}}}~{\vct{z}}]$ and $\boldsymbol{\omega}=\sqrt{{\boldsymbol{{\Sigma}}}}\vct{w}$. Hence consider CGMT with $\lambda=1$ and $\psi(\cdot)=0$, we have
\begin{align}
\Phi_\mathcal{S}({\mtx{X}})&=\min_{\boldsymbol{\omega}\in\mathcal{S}}\tn{{\mtx{X}}'\left[\begin{matrix}
\boldsymbol{\omega} \\
\sigma
\end{matrix}\right]}\label{overder Phi}\\
\phi_\mathcal{S}({\vct{g}},\vct{h}, h)&=\min_{\boldsymbol{\omega}\in\mathcal{S}}(\tn{\left[\begin{matrix}
\boldsymbol{\omega} \\
\sigma
\end{matrix}\right]}\tn{{\vct{g}}}-\vct{h}^T\boldsymbol{\omega}+h\sigma)_+\label{overder phi}
\end{align}
We will argue that the solution of \ref{overder phi} has similar statistical properties to
\[
\vct{w}_{AO}=w_*{\boldsymbol{{\Sigma}}}^{-1/2}\bar{\h}\qquad\text{where}\qquad w_*=\sqrt{\frac{\kappa}{1-\kappa}}\sigma
\]
We see that the loss function is almost in the format of \eqref{overder Phi} except for a set restriction $\boldsymbol{\omega}\in\mathcal{S}$. To proceed, following Lemma \ref{dist cont lem} with a proper choice set $\mathcal{S}$, the solution of unconstrained problem lies in $\mathcal{S}^c$ with high probability. Specifically, we are interested in showing similar behavior to the $1$-Lipschitz function $F'(\boldsymbol{\omega}_{AO})=F(\sqrt{{\boldsymbol{{\Sigma}}}}{\boldsymbol{\beta}}-\boldsymbol{\omega}_{AO})$ where $\boldsymbol{\omega}_{AO}=\sqrt{{\boldsymbol{{\Sigma}}}}\vct{w}_{AO}$. Hence, in the following discussion, we choose
\[
\mathcal{S}=\{\boldsymbol{\omega}{~\big |~} |\tn{F'(\boldsymbol{\omega})}-{\operatorname{\mathbb{E}}[\tn{F'(\boldsymbol{\omega}_{AO})}]}|\geq 2\varepsilon\}.
\]
and will argue that with high probability, $\boldsymbol{\omega}_{PO}\not\in\mathcal{S}$. From Gaussian concentration of Lipschitz functions, we have that
\begin{align}
\mathbb{P}(|\tn{F'(\boldsymbol{\omega}_{AO})}-\operatorname{\mathbb{E}}[\tn{F'(\boldsymbol{\omega}_{AO})}]|\geq \varepsilon)\leq C\mathrm{e}^{-c\varepsilon^2p}.\label{con gauss}
\end{align}
Define the set
\[
\mathcal{B}=\{\boldsymbol{\omega}{~\big |~}\tn{\boldsymbol{\omega}-\boldsymbol{\omega}_{AO}}\geq \varepsilon\}.
\]
Let $E$ be the event that $\mathcal{B}\supseteq\mathcal{S}$. Then $\mathbb{P}(E)\geq 1-C\mathrm{e}^{-c\varepsilon^2p}$, as whenever \eqref{con gauss} holds, for any $\boldsymbol{\omega}\not\in\mathcal{B}$, we find $\boldsymbol{\omega}\not\in\mathcal{S}$ as follows
\begin{align}
|\tn{F'(\boldsymbol{\omega})}-\operatorname{\mathbb{E}}[\tn{F'(\boldsymbol{\omega}_{AO})}]|&\leq|\tn{F'(\boldsymbol{\omega})}-\tn{F'(\boldsymbol{\omega}_{AO})}|+|\operatorname{\mathbb{E}}[\tn{F'(\boldsymbol{\omega}_{AO})}]-\tn{F'(\boldsymbol{\omega}_{AO})}|\\
&\leq\tn{F'(\boldsymbol{\omega})-F'(\boldsymbol{\omega}_{AO})}+\varepsilon\\
&< 2\varepsilon.
\end{align}
Thus, in light of Lemma \ref{dist cont lem}, observe that
\[
\mathbb{P}(\phi_\mathcal{S}({\vct{g}},\vct{h})\leq t)\leq \mathbb{P}(E^c)+\mathbb{P}(\{\phi_{\mathcal{S}}({\vct{g}},\vct{h})\leq t\}\cap E)\leq C\mathrm{e}^{-c\varepsilon^2p}+\mathbb{P}(\phi_{\mathcal{B}}({\vct{g}},\vct{h})\leq t).
\]
In the subsequent discussion, we will bound $\mathbb{P}(\phi_{\mathcal{B}}({\vct{g}},\vct{h})\leq t)$ and $\mathbb{P}(\phi_{\mathbb{R}^p}({\vct{g}},\vct{h})\geq t)$ for a proper choice of $t$.
\noindent{\bf{Case 1: Upper bound analysis over $\mathbb{R}^p$}}
\som{These are minor but you are supposed to show that $l_g>l_h$ and the objective $\phi_{\mathbb{R}^p}>0$}
Let $\tn{{\vct{g}}}=l_g$, $\tn{\vct{h}}=l_h$, $h'=h/l_g$, $\rho=l_h^2/l_g^2<1$ and $\omega=\tn{\boldsymbol{\omega}}$. With at least probability $1-C\mathrm{e}^{-n(1-\sqrt{\kappa})^2/8}$, $l_g-l_h+h'>0$ and $\phi_{\mathbb{R}^p}>0$. We consider the auxiliary problem
\begin{align}
\phi_{\mathbb{R}^p}({\vct{g}},\vct{h},h)&=\min_{\boldsymbol{\omega}\in\mathbb{R}^p}(\tn{\left[\begin{matrix}
\boldsymbol{\omega} \\
\sigma
\end{matrix}\right]}\tn{{\vct{g}}}-\vct{h}^T\boldsymbol{\omega}+h\sigma)_+\\
&=\min_{\omega\in\mathbb{R}}(\sqrt{\omega^2+\sigma^2}l_g-\omega l_h+h\sigma)_+\\
&=l_g\min_{\omega\in\mathbb{R}}\sqrt{\omega^2+\sigma^2}-\omega \sqrt{\rho}+h'\sigma\\
&=(\sqrt{1-\rho}+h')\sigma l_g
\end{align}
We can derive the global solution
\begin{align}
\boldsymbol{\omega}^*=\sqrt{\frac{\rho}{1-\rho}}\sigma\bar{\h}\label{wo over R}
\end{align}
\noindent{\bf{Case 2: Lower bound analysis over $\mathcal{B}$}}
\som{Assuming ${\boldsymbol{{\Sigma}}}$ is full rank, you can do change of variable $\vct{v}\leftrightarrow{\boldsymbol{{\Sigma}}}^{-1/2}\vct{v}$ to simplify notation. You can get rid of ${\boldsymbol{{\Sigma}}}$ and $\gamma$!}
\som{How do you show: ``This implies that'' explain further please}
Let $\tn{{\vct{g}}}=l_g$, $\tn{\vct{h}}=l_h$, $h'=h/l_g$ and $\rho=l_h^2/l_g^2<1$. Set $\boldsymbol{\omega}\in\mathcal{B}$ as $\boldsymbol{\omega}=\boldsymbol{\omega}_{AO}+\vct{v}$ where $\tn{\vct{v}}\geq\varepsilon$. From \ref{wo over R} we can see that $\vct{v}*=(\sqrt{\frac{\rho}{1-\rho}}-\sqrt{\frac{\kappa}{1-\kappa}})\sigma\bar{\h}$ over $\phi_{\mathbb{R}^p}$. Hence for any $\rho$ making $\varepsilon\leq|\sqrt{\frac{\rho}{1-\rho}}-\sqrt{\frac{\kappa}{1-\kappa}}|\sigma$, $\phi_{\mathcal{B}}=\phi_{\mathbb{R}^p}$. However for sufficiently large $n$ and $p$ and any $\varepsilon>0$, with high probability, $\varepsilon>|\sqrt{\frac{\rho}{1-\rho}}-\sqrt{\frac{\kappa}{1-\kappa}}|\sigma$ and $\phi_{\mathcal{B}}>\phi_{\mathbb{R}^p}$. There exists sufficiently small $\gamma>0$, the auxiliary problem can be presented as
\begin{align}
\phi_{\mathcal{B}}({\vct{g}},\vct{h},h)&=\min_{\tn{\vct{v}}\geq\varepsilon}\tn{\left[\begin{matrix}
w_*\bar{\h}+\vct{v} \\
\sigma
\end{matrix}\right]}\tn{{\vct{g}}}-(w_*\tn{\vct{h}}+\vct{h}^T\vct{v})+h\sigma\\
&\geq
\end{align}
Write $\phi_{\mathcal{B}}$ as $\phi_{\mathcal{B}}=l_g\min_{\vct{v}}f(\vct{v})$ where $f(\vct{v})=\sqrt{(w_*\bar{\h}+\vct{v})^2+\sigma^2}-(w_*+\bar{\h}\vct{v})\sqrt{\rho}+h'\sigma$. Clearly $f(\vct{v})$ is convex over $\vct{v}$. Do derivation directly to find optima.
\begin{align}
\frac{w_*\bar{\h}+\vct{v}}{\sqrt{(w_*\bar{\h}+\vct{v})^2+\sigma^2}}-\bar{\h}\sqrt{\rho}=0\implies \vct{v}=(\sqrt{\frac{\rho}{1-\rho}}\sigma-w_*)\bar{\h}
\end{align}
This implies that $\vct{v}=v\frac{{\boldsymbol{{\Sigma}}}^{-1/2}\bar{\h}}{\tn{{\boldsymbol{{\Sigma}}}^{-1/2}\bar{\h}}}=v/\gamma{\boldsymbol{{\Sigma}}}^{-1/2}\bar{\h}$ where $v=\tn{\vct{v}}\geq\varepsilon$. Then for any $\varepsilon>(\sqrt{\frac{\rho}{1-\rho}}\sigma-w_*)\gamma$, $v^*=\varepsilon$
\[
\vct{v}*=\varepsilon/\gamma{\boldsymbol{{\Sigma}}}^{-1/2}\bar{\h}\implies\vct{w}^*=(w_*+\varepsilon/\gamma){\boldsymbol{{\Sigma}}}^{-1/2}\bar{\h}
\]
{\color{red}
$f(\vct{v})=\tn{w_*\bar{\h}+\sqrt{{\boldsymbol{{\Sigma}}}}\vct{v}~\sigma}\tn{{\vct{g}}}-(w_*\tn{\vct{h}}+\vct{h}^T\sqrt{{\boldsymbol{{\Sigma}}}}\vct{v})+h\sigma$ is convex over $\vct{v}$. Do derivation directly to find optima.
}
\som{You want to simplify your bound on $\phi_{\mathcal{B}}$ and argue it is larger than $\phi_{\mathbb{R}^p}$}
\begin{align}
\phi_{\mathcal{B}}({\vct{g}},\vct{h},h)&=\min_{v\geq\varepsilon}l_g\sqrt{(v/\gamma+w_*)^2+\sigma^2}-l_h(v/\gamma+w_*)+h\sigma\\
&=l_g\sqrt{(\varepsilon/\gamma+w_*)^2+\sigma^2}-l_h(\varepsilon/\gamma+w_*)+h\sigma
\end{align}
\som{I don't follow this condition on $\varepsilon$}Or if $\varepsilon\leq(\sqrt{\frac{\rho}{1-\rho}}\sigma-w_*)\gamma$,
\[
v*=(\sqrt{\frac{\rho}{1-\rho}}\sigma-w_*)\gamma\implies\phi_{\mathcal{B}}({\vct{g}},\vct{h},h)=\phi_{\mathbb{R}^p}({\vct{g}},\vct{h},h
\]
{\color{red}
For any ${\vct{g}},\vct{h},h$, we have $\phi_{\mathcal{B}}\geq\phi_{\mathbb{R}}$. Only when $\varepsilon>(\sqrt{\frac{\rho}{1-\rho}}\sigma-w_*)\gamma$, we have $\phi_{\mathcal{B}}>\phi_{\mathbb{R}}$
}
Set $\kappa=p/n$. For any $\varepsilon>(\sqrt{\frac{\rho}{1-\rho}}-\sqrt{\frac{\kappa}{1-\kappa}})\sigma\gamma$, we have proper choice $t$\som{You need a $t$ that achieves high probability. Inequality is not enough}
{\color{red}\\\\
1. With high probability, $(\sqrt{\frac{\rho}{1-\rho}}-\sqrt{\frac{\kappa}{1-\kappa}})\sigma\gamma<\epsilon$ where $\epsilon$ is small and positive\\
2. Then, there exists $\varepsilon<\epsilon$ that inequalities below hold\\
3. For any $t\in(\sqrt{n}(\sqrt{1-\kappa}\sigma+\varepsilon_1),\sqrt{n}(\sqrt{(\varepsilon/\gamma+w_*)^2+\sigma^2}-\sqrt{\kappa}(\varepsilon/\gamma+w_*)-\varepsilon_1))$ where $\sqrt{(\varepsilon/\gamma+w_*)^2+\sigma^2}-\sqrt{\kappa}(\varepsilon/\gamma+w_*)-\sqrt{1-\kappa}\sigma>2\varepsilon_1$ with high probability $Ce^{-cn\varepsilon_1^2/\sigma^2}$, we have $\phi_{\mathcal{B}}>\phi_{\mathbb{R}}$.
}
\begin{align}
\sqrt{n-p}\sigma+h\sigma<&t<\sqrt{n}\sqrt{(\varepsilon/\gamma+w_*)^2+\sigma^2}-\sqrt{p}(\varepsilon/\gamma+w_*)+h\sigma\\
\sqrt{1-\kappa}\sigma<&\frac{t-h\sigma}{\sqrt{n}}<\sqrt{(\varepsilon/\gamma+w_*)^2+\sigma^2}-\sqrt{\kappa}(\varepsilon/\gamma+w_*)
\end{align}
and
\som{This is not a rigorous conclusion. You are supposed to show >0}
\[
\sqrt{(\varepsilon/\gamma+w_*)^2+\sigma^2}-\sqrt{\kappa}(\varepsilon/\gamma+w_*)-\sqrt{1-\kappa}\sigma\nrightarrow0
\]
\end{proof}
\section{Underparameterized analysis}\label{SM overdet}
This section provides our results for the asymptotic DC in the underparameterized regime. This results establish direct counterparts of the overparameterized results Definition \ref{def:Xi} and Theorem \ref{thm:master_W2}. However, underparameterized DC is substantially less involved compared to overparameterized. A key reason is that underparameterized least-squares returns an unbiased estimate of the ground-truth parameter. Similar to Section \ref{sec proof thm 1}, for simplicity, we assume diagonal covariance however results can be translated to arbitrary covariance via eigen-rotation trick (e.g.~recall Def.~\ref{aux_def}). Throughout, we solve the following proble
\begin{align}
{\boldsymbol{\hat{\beta}}}={\mtx{X}}^\dagger\vct{y}=\arg\min_{{\boldsymbol{\beta}}} \tn{\vct{y}-{\mtx{X}}{\boldsymbol{\beta}}}^2\label{bth up}
\end{align}
where $\vct{y}={\mtx{X}}\boldsymbol{\beta}^\star+\sigma{\vct{z}}$ and ${\mtx{X}}={\mtx{\bar{X}}}\sqrt{{\boldsymbol{{\Sigma}}}}$. Now, set $\boldsymbol{\omega}=\sqrt{{\boldsymbol{{\Sigma}}}}({\boldsymbol{\beta}}-\boldsymbol{\beta}^\star)$ as previously. We can rewrite
\begin{align}
{\boldsymbol{\hat{\beta}}}={\mtx{X}}^\dagger\vct{y}=\boldsymbol{\beta}^\star+{\boldsymbol{{\Sigma}}}^{-1/2}\vct{w}^\star\quad\text{where}\quad \vct{w}^\star=\arg\min_{\boldsymbol{\omega}} \tn{\sigma{\vct{z}}-{\mtx{\bar{X}}}\boldsymbol{\omega}}^2.\label{wst up}
\end{align}
We will prove the following DC for the underparameterized problem with $n<p$ and $p/n=\kappa<1$.
\begin{definition}[Asymptotic DC -- Underparameterized regime]\label{def:Xi_under}
Let random variables $(B,\Lambda)\sim \mu$ (where $\mu$ is defined in Assumption \ref{ass:mu}) and fix $\kappa<1$. Let $H\sim\mathcal{N}(0,1)$ and define the random variable
\begin{align}
X_{\kappa,\sigma^2}(B,H) := B + \sigma\frac{\Lambda^{-1/2}H}{\sqrt{\kappa^{-1}-1}}, \label{eq:X
\end{align}
and let $\Pi_{\kappa,\sigma^2}$ be its distribution.
\end{definition}
We are now ready to state our main theoretical result.
\cmt{
\begin{theorem}[Asymptotic DC -- Underparameterized LGP]\label{thm:master_W2_under}
Let Assumption \ref{ass:linear} hold with $\kappa<1$ and further let Assumptions \ref{ass:inv} and \ref{ass:mu} hold. Consider $\hat{{\boldsymbol{\beta}}}$ as in \eqref{bth up} and $\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p}\sum_{i=1}^{p}\delta_{\sqrt{p}\hat{{\boldsymbol{\beta}}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}}$, the joint empirical distribution of $(\sqrt{p}\hat\boldsymbol{\beta},\sqrt{p}\betab^\star,\boldsymbol{\Sigma})$. Recall the definition of the measure $\Pi_{\kappa,\sigma^2}$ in Def.~\ref{def:Xi_under}. Then, $\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma})$ converges in Wasserstein-k distance to $\Pi_{\kappa,\sigma^2}\otimes\mu$.
Specifically, for any function $f:\mathbb{R}^3\rightarrow\mathbb{R}$, $f\in\rm{PL}(k)$ with $k\geq 3$, it holds that
\begin{align}\label{eq:thm_up}
\hspace{-0.1in}p^{-1} \sum_{i=1}^{p} f(\sqrt{p}\hat\boldsymbol{\beta}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}\left[f(X_{\kappa,\sigma^2},B,\Lambda) \right],
\end{align}
where the expectation is over $(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)$. Specifically, the asymptotic test risk is given by $\frac{\sigma^2}{1-\kappa}$.
\end{theorem}}
\begin{theorem}[Asymptotic DC -- Underparameterized LGP]\label{thm:master_W2_under}
Fix $\kappa<1$. Let Assumptions \ref{ass:inv} and \ref{ass:mu} hold. Consider $\hat{{\boldsymbol{\beta}}}$ as in \eqref{bth up} and $\hat\Pi_n(\vct{y},{\mtx{X}},\betab^\star,\boldsymbol{\Sigma}):=\frac{1}{p}\sum_{i=1}^{p}\delta_{\sqrt{p}\hat{{\boldsymbol{\beta}}}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}}$, the joint empirical distribution of $(\sqrt{p}\hat\boldsymbol{\beta},\sqrt{p}\betab^\star,\boldsymbol{\Sigma})$. Recall the definition of the measure $\Pi_{\kappa,\sigma^2}$ in Def.~\ref{def:Xi_under}. {Let $f:\mathbb{R}^3\rightarrow\mathbb{R}$ be a function in ${\cal{F}}$ where ${\cal{F}}$ is defined in \eqref{eq:pdef}.} We have that
\begin{align}\label{eq:thm_up}
\frac{1}{p} \sum_{i=1}^{p} f(\sqrt{p}\hat\boldsymbol{\beta}_i,\sqrt{p}\betab^\star_i,\boldsymbol{\Sigma}_{i,i}) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}_{(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)}\left[f(X_{\kappa,\sigma^2},B,\Lambda) \right].
\end{align}
Specifically, the asymptotic test risk of ${\boldsymbol{\hat{\beta}}}$ is given by $\frac{\sigma^2}{1-\kappa}$.
\end{theorem}
\begin{proof} \so{To avoid repetition, we will not provide the full proof as the technical details of the proofs for over/under-parameterized overlap to a significant extent. Instead, we will provide the part of the proof that deviates from the overparameterized.}
Since $\kappa<1$, the problem has a unique solution. Set $\boldsymbol{\omega}=\sqrt{{\boldsymbol{{\Sigma}}}}({\boldsymbol{\beta}}-\boldsymbol{\beta}^\star)$. Define ${\mtx{X}}_{\vct{z}}=[{\mtx{\bar{X}}}~{\vct{z}}]$ and $\boldsymbol{\omega}_\sigma=[\boldsymbol{\omega}~\sigma]$. This leads to the optimization problem
\[
\hat{\boldsymbol{\omega}}=\arg\min_{\boldsymbol{\omega}} \tn{{\mtx{\bar{X}}}\boldsymbol{\omega}+\sigma {\vct{z}}}=\arg\min_{\boldsymbol{\omega}} \tn{{\mtx{X}}_{\vct{z}}\boldsymbol{\omega}_\sigma}.
\]
Fix ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_p),\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_n),g\sim\mathcal{N}(0,1)$. Applying CGMT leads to the following Auxiliary Optimizatio
\begin{align}
\phi({\vct{g}},\vct{h})&=\min_{\boldsymbol{\omega}} \max_{\tn{\vct{a}}\leq 1} \vct{h}^T\vct{a} \tn{\boldsymbol{\omega}_{\sigma}}-\tn{\vct{a}}{\vct{g}}^T\boldsymbol{\omega}+g\sigma\\
&=\min_{\boldsymbol{\omega}} \tn{\vct{h}} \tn{\boldsymbol{\omega}_{\sigma}}-{\vct{g}}^T\boldsymbol{\omega}+g\sigma.
\end{align}
Solving for optimal $\boldsymbol{\omega}$ leads to the solution
\[
\boldsymbol{\omega}^{\text{AO}}=\arg\min_{\boldsymbol{\omega}}\tn{\vct{h}} \tn{\boldsymbol{\omega}_{\sigma}}-{\vct{g}}^T\boldsymbol{\omega}\implies \tn{\vct{h}}\frac{\boldsymbol{\omega}^{\text{AO}}}{\sqrt{\tn{\boldsymbol{\omega}^{\text{AO}}}^2+\sigma^2}}-{\vct{g}}\implies \boldsymbol{\omega}^{\text{AO}}=\frac{\sigma{\vct{g}}}{\sqrt{\tn{\vct{h}}^2-\tn{{\vct{g}}}^2}}.
\]
Observing $\tn{\boldsymbol{\omega}^{\text{AO}}}^2+\sigma^2=\frac{\sigma^2\tn{\vct{h}}^2}{\tn{\vct{h}}^2-\tn{{\vct{g}}}^2}$ and plugging $\boldsymbol{\omega}^{\text{AO}}$ in, we find
\[
\sigma^{-1}\phi({\vct{g}},\vct{h})=\frac{\tn{\vct{h}}^2-\tn{{\vct{g}}}^2}{\sqrt{\tn{\vct{h}}^2-\tn{{\vct{g}}}^2}}+g=\sqrt{\tn{\vct{h}}^2-\tn{{\vct{g}}}^2}+g.
\]
Thus, in the asymptotic regime $\phi({\vct{g}},\vct{h})$ converges to the objectiv
\[
\phi({\vct{g}},\vct{h})\stackrel{{P}}{\longrightarrow} \bar{\phi}=\sigma\sqrt{n-p}.
\]
The remaining arguments are same as in Lemma \ref{lem:AO}. First, the problem is strongly convex with $\sigma^2_{\min}({\mtx{X}})$, which satisfies $\sigma^2_{\min}({\mtx{X}})/p\gtrsim 1$ wpa.~1. Thus, the solution $\vct{w}^\star$ of the primary problem \eqref{wst up} will not deviate from $\boldsymbol{\omega}^{\text{AO}}$. Secondly, the empirical distribution of
\[
\sqrt{p}\boldsymbol{\omega}^\star=\frac{\sigma\sqrt{p}{\vct{g}}}{\sqrt{\tn{\vct{h}}^2-\tn{{\vct{g}}}^2}}\stackrel{{P}}{\longrightarrow} \frac{\sigma\sqrt{p}{\vct{g}}}{\sqrt{n-p}}= \frac{\sigma{\vct{g}}}{\sqrt{n/p-1}}=\frac{\sigma{\vct{g}}}{\sqrt{\kappa^{-1}-1}},
\]
converges to $\sigma H/\sqrt{\kappa^{-1}-1}$. By Assumption \ref{ass:mu}, the empirical distribution of $\sqrt{p}{\boldsymbol{\hat{\beta}}}=\sqrt{p}({\boldsymbol{\beta}}^\star+{\boldsymbol{{\Sigma}}}^{-1/2}\boldsymbol{\omega}^\star)$ converges to $B+\sigma \Lambda^{-1/2}H/\sqrt{\kappa^{-1}-1}\sim \Pi_{\kappa,\sigma^2}$. Finally, again by Assumption \ref{ass:mu}, for any $f\in {\cal{F}}$, we obtain the advertised result \eqref{eq:thm_up}. The asymptotic test risk is given by
\[
{\cal{L}}({\boldsymbol{\hat{\beta}}})=\operatorname{\mathbb{E}}[\tn{{\vct{g}}^T\sqrt{{\boldsymbol{{\Sigma}}}}({\boldsymbol{\hat{\beta}}}-\boldsymbol{\beta}^\star)+\sigma z}^2]=\sigma^2+\sum_{i=1}^p ({\boldsymbol{\hat{\beta}}}_i-\boldsymbol{\beta}^\star)^2{\boldsymbol{{\Sigma}}}_{i,i}\stackrel{{P}}{\longrightarrow}\sigma^2+\frac{\sigma^2}{\kappa^{-1}-1}=\frac{\sigma^2}{1-\kappa}.
\]
\end{proof}
In the main body of the paper, we claim that the optimal $s$ features to use in the underparameterized regime is given by the features with the maximum saliency score. This is proven below.
\begin{lemma}[Optimal $s$ features to use]\label{lem best s} Fix a sequence of sets $\Delta_p\subset[p]$ of size $s$ such that $\sum_{i\in \Delta_p} {\boldsymbol{\beta}^\star_i}^2{\boldsymbol{{\Sigma}}}_{i,i}\stackrel{{P}}{\longrightarrow} B(\Delta)$. Set $\kappa=s/n$. Under same assumptions as in Thm~\ref{thm:master_W2_under}, the asymptotic test risk of ${\boldsymbol{\hat{\beta}}}(\Delta)$ is given by
\[
{\cal{L}}({\boldsymbol{\hat{\beta}}}(\Delta))\stackrel{{P}}{\longrightarrow} \frac{B-B(\Delta)+\sigma^2}{1-\kappa}.
\]
Thus, the optimal feature set $\Delta$ chooses the indices with maximum Saliency Score \eqref{saliency eq} which maximizes $B(\Delta)$.
\end{lemma}
\begin{proof} The key idea is the fact that we can treat the missing features as uncorrelated noise. First, due to diagonal covariance, observe that, over the feature set $\Delta$, the optimal population model (i.e.~infinite sample) is $\boldsymbol{\beta}^\star_{\Delta}$. Thus, the $s$ feature problem minimized by ${\boldsymbol{\hat{\beta}}}(\Delta)$ can be written as the dataset model
\[
y=\vct{x}_{\Delta}^T\boldsymbol{\beta}^\star_{\Delta}+\sigma_{\Delta}^2,
\]
where the noise level is given by
\[
\operatorname{\mathbb{E}}[(y-\vct{x}_{\Delta}\boldsymbol{\beta}^\star_{\Delta})^2]=\sigma^2+\operatorname{\mathbb{E}}[(\vct{x}_{\bar{\Delta}}\boldsymbol{\beta}^\star_{\bar{\Delta}})^2]=\sigma^2+\sum_{i\not\in\Delta}{\boldsymbol{{\Sigma}}}_{i,i}{\boldsymbol{\beta}^\star_{i}}^2.
\]
The latter quantity converges to $B-B(\Delta)$ wpa.~1. Thus, applying Theorem \ref{thm:master_W2_under}, ${\boldsymbol{\hat{\beta}}}(\Delta)$ achieves the advertised asymptotic risk.
\end{proof}
\cmt{
where $\mathcal{D}(u,\tau)=\dots$. Specifically, we have that
\begin{align}\\
&=\min_{\vct{w}} \sqrt{n} \tn{\boldsymbol{\omega}_{\sigma}}-{\vct{g}}^T\boldsymbol{\omega}_{\sigma}\\
&=\min_{\vct{w}} \sqrt{n} \sqrt{w^2+\sigma^2}-w\tn{{\vct{g}}}\\
&\propto\min_{\vct{w}} \sqrt{w^2+\sigma^2}-w\sqrt{{\bar{p}}}.
\end{align}
This implies $w=\sqrt{\frac{{\bar{p}}}{1-{\bar{p}}}}\sigma$ via
\[
\sqrt{{\bar{p}}}=\frac{w}{\sqrt{w^2+\sigma^2}}\iff {\bar{p}} (w^2+\sigma^2)=w^2\iff w=\sqrt{\frac{{\bar{p}}}{1-{\bar{p}}}}\sigma
\]
Now, recall that ${\boldsymbol{\hat{\beta}}}={\boldsymbol{\beta}}+{\boldsymbol{{\Sigma}}}^{-1/2}\boldsymbol{\omega}$. Since the entries of $\boldsymbol{\omega}$ are normally distributed, we have that
\begin{align}
\sqrt{{\boldsymbol{{\Sigma}}}}\boldsymbol{\hat{\theta}}&=\sqrt{{\boldsymbol{{\Sigma}}}}\mathbb{F}_s({\boldsymbol{\hat{\beta}}})=\sqrt{{\boldsymbol{{\Sigma}}}}(\mathbb{F}_s({\boldsymbol{\beta}})+\mathbb{F}_s({\boldsymbol{{\Sigma}}}^{-1/2}\boldsymbol{\omega}))\\
&=\sqrt{{\boldsymbol{{\Sigma}}}}\mathbb{F}_s({\boldsymbol{\beta}})+\mathbb{F}_s(\boldsymbol{\omega})\\
&=\vct{t}+\mathbb{F}_s(\boldsymbol{\omega}).
\end{align}
Note that $\tn{\mathbb{F}_s(\boldsymbol{\omega})}=\sqrt{\frac{s}{p}}\sqrt{\frac{{\bar{p}}}{1-{\bar{p}}}}\sigma=\sqrt{\frac{{\bar{s}}}{1-{\bar{p}}}}\sigma$. Thus the test error \eqref{test formula} of ${\boldsymbol{\hat{\beta}}}$ and $\boldsymbol{\hat{\theta}}$ are
\[
{\cal{L}}({\boldsymbol{\hat{\beta}}})=\sigma^2+w^2=\frac{\sigma^2}{1-{\bar{p}}}\quad\text{and}\quad {\cal{L}}(\boldsymbol{\hat{\theta}})=\frac{{\bar{s}}}{1-{\bar{p}}}\sigma^2+\alpha-\theta+\sigma^2.
\]}
\cmt{
\begin{lemma}[To be proven rigorously] Set ${\bar{p}}=p/n$ and ${\bar{s}}=s/n$. Suppose ${\bar{p}}<1$ and suppose $\vct{t}^T{\boldsymbol{{\Sigma}}}\vct{t}=\theta$. Then
\[
{\cal{L}}({\boldsymbol{\hat{\beta}}})=\sigma^2+w^2=\frac{\sigma^2}{1-{\bar{p}}}\quad\text{and}\quad {\cal{L}}(\boldsymbol{\hat{\theta}})=\frac{1-{\bar{p}}+{\bar{s}}}{1-{\bar{p}}}\sigma^2+\alpha-\theta.
\]
\end{lemma}
\subsection{Analysis}
Since $\kappa<1$, the problem has a unique solution. Set $\boldsymbol{\omega}=\sqrt{{\boldsymbol{{\Sigma}}}}\vct{w}$ where $\vct{w}={\boldsymbol{\beta}}-{\boldsymbol{\beta}}'$. Define ${\mtx{X}}_{\vct{z}}=[{\mtx{\bar{X}}}~{\vct{z}}]$ and $\boldsymbol{\omega}_\sigma=[\boldsymbol{\omega}^T~\sigma]^T$. This leads to the optimization problem
\[
\hat{\boldsymbol{\omega}}=\arg\min_{\boldsymbol{\omega}} \tn{{\mtx{\bar{X}}}\boldsymbol{\omega}+\sigma {\vct{z}}}=\arg\min_{\boldsymbol{\omega}} \tn{{\mtx{X}}_{\vct{z}}\boldsymbol{\omega}_\sigma}.
\]
Fix ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_p),\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_n)$. Note that that $s/p={\bar{s}}/{\bar{p}}$. Applying Gordon's Lemma (non-rigorously) and noticing optimal $\boldsymbol{\omega}$ has form $\boldsymbol{\omega}=w\bar{{\vct{g}}}$ where $\bar{{\vct{g}}}={\vct{g}}/\tn{{\vct{g}}}$, we find
\begin{align}
\min_{\vct{w}} \max_{\tn{\vct{a}}\leq 1} \vct{h}^T\vct{a} \tn{\boldsymbol{\omega}_{\sigma}}-\tn{\vct{a}}{\vct{g}}^T\boldsymbol{\omega}_{\sigma}&=\min_{\vct{w}} \sqrt{n} \tn{\boldsymbol{\omega}_{\sigma}}-{\vct{g}}^T\boldsymbol{\omega}_{\sigma}\\
&=\min_{\vct{w}} \sqrt{n} \sqrt{w^2+\sigma^2}-w\tn{{\vct{g}}}\\
&\propto\min_{\vct{w}} \sqrt{w^2+\sigma^2}-w\sqrt{{\bar{p}}}.
\end{align}
This implies $w=\sqrt{\frac{{\bar{p}}}{1-{\bar{p}}}}\sigma$ via
\[
\sqrt{{\bar{p}}}=\frac{w}{\sqrt{w^2+\sigma^2}}\iff {\bar{p}} (w^2+\sigma^2)=w^2\iff w=\sqrt{\frac{{\bar{p}}}{1-{\bar{p}}}}\sigma
\]
Now, recall that ${\boldsymbol{\hat{\beta}}}={\boldsymbol{\beta}}+{\boldsymbol{{\Sigma}}}^{-1/2}\boldsymbol{\omega}$. Since the entries of $\boldsymbol{\omega}$ are normally distributed, we have that
\begin{align}
\sqrt{{\boldsymbol{{\Sigma}}}}\boldsymbol{\hat{\theta}}&=\sqrt{{\boldsymbol{{\Sigma}}}}\mathbb{F}_s({\boldsymbol{\hat{\beta}}})=\sqrt{{\boldsymbol{{\Sigma}}}}(\mathbb{F}_s({\boldsymbol{\beta}})+\mathbb{F}_s({\boldsymbol{{\Sigma}}}^{-1/2}\boldsymbol{\omega}))\\
&=\sqrt{{\boldsymbol{{\Sigma}}}}\mathbb{F}_s({\boldsymbol{\beta}})+\mathbb{F}_s(\boldsymbol{\omega})\\
&=\vct{t}+\mathbb{F}_s(\boldsymbol{\omega}).
\end{align}
Note that $\tn{\mathbb{F}_s(\boldsymbol{\omega})}=\sqrt{\frac{s}{p}}\sqrt{\frac{{\bar{p}}}{1-{\bar{p}}}}\sigma=\sqrt{\frac{{\bar{s}}}{1-{\bar{p}}}}\sigma$. Thus the test error \eqref{test formula} of ${\boldsymbol{\hat{\beta}}}$ and $\boldsymbol{\hat{\theta}}$ are
\[
{\cal{L}}({\boldsymbol{\hat{\beta}}})=\sigma^2+w^2=\frac{\sigma^2}{1-{\bar{p}}}\quad\text{and}\quad {\cal{L}}(\boldsymbol{\hat{\theta}})=\frac{{\bar{s}}}{1-{\bar{p}}}\sigma^2+\alpha-\theta+\sigma^2.
\]
}
\subsection{Overparameterized Analysis}
\begin{lemma}[Asymptotic Strong Convexity]\label{asym str} Fix $\sigma>0$ and $\rho>1$. Consider the loss function over nonnegative variables $v,w,\gamma$
\[
f(\gamma,w,v)=(1-\gamma)^2+w^2+v^2+\max_{\lambda\geq 0}\lambda (v^2+\gamma^2+\sigma^2-(\rho-1)w^2),
\]
and the optimal value
\[
f^*=\min_{w^2+v^2+\gamma^2+\sigma^2\leq w^2\rho}(1-\gamma)^2+w^2+v^2.
\]
Then, $f$ has a unique global minima $(\gamma^*,w^*,v^*)$ given by
\begin{align}
\gamma^*=1-\frac{1}{\rho},~w^*=\sqrt{\frac{\sigma^2}{\rho-1}+\frac{\rho-1}{\rho^2}},~v^*=0.\label{opt choices}
\end{align}
Additionally, for any $(\gamma,w,v)$, we have $f(\gamma,w,v)-f^*\geq (\gamma-\gamma^*)^2+v^2+(w-w^*)^2$.
\end{lemma}
\begin{proof} First observe that, we can restrict our attention to feasible values satisfying $v^2+\gamma^2+\sigma^2\leq (\rho-1)w^2$. Suppose $v,w,\gamma$ be a feasible triple. Then $0,w_\gamma=\sqrt{\frac{\gamma^2+\sigma^2}{\rho-1}},\gamma$ is also a feasible triple and obeys
\[
f(\gamma,w,v)-f(\gamma,_\gamma,0)\geq v^2+w^2-w_\gamma^2.
\]
Plugging in $w_\gamma$ we get
\[
f(\gamma,w_\gamma,0)=\frac{\gamma^2+\sigma^2}{\rho-1}+(1-\gamma)^2.
\]
Differentiating we find $1=\gamma^*+\frac{\gamma^*}{\rho-1}\implies \gamma^*=1-1/\rho$ and
\[
w^*=w_{\gamma^*}=\sqrt{\frac{\sigma^2}{\rho-1}+\frac{\rho-1}{\rho^2}}.
\].
Observe that $f$ is $\frac{2\rho}{\rho-1}$-strongly convex function of $\gamma$ thus
\[
f(\gamma,w_\gamma,0)\geq f^*+\frac{\rho}{\rho-1}(\gamma-\gamma^*)^2\geq f^*+(\gamma-\gamma^*)^2+(w_\gamma-w^*)^2.
\]
In summary, we find
\[
f(\gamma,w,v)-f^*\geq (\gamma-\gamma^*)^2+[w^2-w_\gamma^2+(w_\gamma-w^*)^2]+v^2.
\]
To conclude, note that
\[
w^2-w_\gamma^2+(w_\gamma-w^*)^2=w^2-2w_\gamma w^*+{w^*}^2=(w-w^*)^2+2(w-w_\gamma) w^*\geq (w-w^*)^2.
\]
\end{proof}
\begin{theorem}[Overparameterized estimation, Identity ${\boldsymbol{{\Sigma}}}$] Consider a dataset according to linear model with $\tn{{\boldsymbol{\beta}}}=1$ and assume $\kappa=p/n>1$. Let $\vct{h}\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$ and $\bar{\h}=\vct{h}/\tn{\vct{h}}$. Define
\[
{\boldsymbol{\hat{\beta}}}_{AO}=(1-\gamma_*){\boldsymbol{\beta}}+w_*\bar{\h}\quad\text{and}\quad w_*=\sqrt{\frac{\sigma^2}{\kappa-1}+\frac{\kappa-1}{\kappa^2}},~\gamma_*=1-\frac{1}{\kappa}.
\]
Then, for any $1$-Lipschitz function $F$, with probability $1-C\mathrm{e}^{-c\varepsilon^2p}$, we have that
\[
|\tn{F({\boldsymbol{\hat{\beta}}})}-\operatorname{\mathbb{E}}[\tn{F({\boldsymbol{\hat{\beta}}}_{AO})}]|\leq \varepsilon
\]
{\bf{Asymptotic:}} Then, for any $1$-Lipschitz function $F$, we have that
\[
\lim_{p\rightarrow\infty}\tn{F({\boldsymbol{\hat{\beta}}})}= \lim_{p\rightarrow\infty}\tn{F({\boldsymbol{\hat{\beta}}}_{AO})}.
\]\som{``for convex functions pointwise convergence in probability implies uniform convergence in compact subsets''}
\end{theorem}
\begin{proof} Fix $h,{\vct{g}}\in\mathbb{R}^n,\vct{h}\in\mathbb{R}^p,{\mtx{X}}\in\mathbb{R}^{n\times p}\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,1)$. Set $\vct{w}={\boldsymbol{\beta}}-{\boldsymbol{\hat{\beta}}}$. We are interested in the minimum norm solution to ${\mtx{X}}{\boldsymbol{\beta}}+\sigma{\vct{z}}={\mtx{X}}{\boldsymbol{\hat{\beta}}}$ i.e.~${\mtx{X}}'\vct{w}_\sigma=0$ where $\vct{w}_\sigma=[\vct{w}^T~\sigma]^T$ and ${\mtx{X}}'=[{\mtx{X}}~{\vct{z}}]$. Thus, with a change of variable, following Theorem \ref{lem cgmt constrained}, we consider
\begin{align}
&\Phi_\mathcal{S}({\mtx{X}}')=\min_{\vct{w}\in\mathcal{S},{\mtx{X}}'\vct{w}_\sigma=0}\tn{{\boldsymbol{\beta}}-\vct{w}}\\
&\phi_\mathcal{S}({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S},\tn{\vct{w}_\sigma}\tn{{\vct{g}}}\leq \vct{h}^T\vct{w}+\sigma h}\tn{{\boldsymbol{\beta}}-\vct{w}}.\label{overparm phi}
\end{align}
To proceed, we will apply Lemma \ref{dist cont lem} with a proper choice of set $\mathcal{S}$ to conclude that the solution $\vct{w}_{PO}$ of the unconstrained least-squares problem lies inside $\mathcal{S}^c$. Set the normalized vector $\bar{\h}=\vct{h}/\tn{\vct{h}}$. Using CGMT and following Lemma \ref{asym str}, we will argue that the solution exhibits similar statistical properties to
\[
\vct{w}_{AO}=\gamma_*{\boldsymbol{\beta}}+w_*\bar{\h}\quad\text{where}\quad w_*=\sqrt{\frac{\sigma^2}{\kappa-1}+\frac{\kappa-1}{\kappa^2}},~\gamma_*=1-\frac{1}{\kappa}.
\]
Specifically, we are interested in showing similar behavior to the $1$-Lipschitz function $F'(\vct{w}_{AO})=F({\boldsymbol{\beta}}-\vct{w}_{AO})$. Hence, in the following discussion, we choose
\[
\mathcal{S}=\{\vct{w}{~\big |~} |\tn{F'(\vct{w})}-{\operatorname{\mathbb{E}}[\tn{F'(\vct{w}_{AO})}]}|\geq 2\varepsilon\}.
\]
and will argue that with high probability, $\vct{w}_{PO}\not\in\mathcal{S}$. From Gaussian concentration of Lipschitz functions, we have that
\begin{align}
\mathbb{P}(|\tn{F'(\vct{w}_{AO})}-\operatorname{\mathbb{E}}[\tn{F'(\vct{w}_{AO})}]|\geq \varepsilon)\leq C\mathrm{e}^{-c\varepsilon^2p}.\label{cond gauss}
\end{align}
Define the set
\[
\mathcal{B}=\{\vct{w}{~\big |~}\tn{\vct{w}-\vct{w}_{AO}}\geq \varepsilon\}.
\]
Let $E$ be the event that $\mathcal{B}\supseteq\mathcal{S}$. Then $\mathbb{P}(E)\geq 1-C\mathrm{e}^{-c\varepsilon^2p}$, as whenever \eqref{cond gauss} holds, for any $\vct{w}\not\in\mathcal{B}$, we find $\vct{w}\not\in\mathcal{S}$ as follows
\begin{align}
|\tn{F'(\vct{w})}-\operatorname{\mathbb{E}}[\tn{F'(\vct{w}_{AO})}]|&\leq|\tn{F'(\vct{w})}-\tn{F'(\vct{w}_{AO})}|+|\operatorname{\mathbb{E}}[\tn{F'(\vct{w}_{AO})}]-\tn{F'(\vct{w}_{AO})}|\\
&\leq\tn{F'(\vct{w})-F'(\vct{w}_{AO})}+\varepsilon\\
&< 2\varepsilon.
\end{align}
Thus, in light of Lemma \ref{dist cont lem}, observe that
\[
\mathbb{P}(\phi_\mathcal{S}({\vct{g}},\vct{h})\leq t)\leq \mathbb{P}(E^c)+\mathbb{P}(\{\phi_{\mathcal{S}}({\vct{g}},\vct{h})\leq t\}\cap E)\leq C\mathrm{e}^{-c\varepsilon^2p}+\mathbb{P}(\phi_{\mathcal{B}}({\vct{g}},\vct{h})\leq t).
\]
In the subsequent discussion, we will bound $\mathbb{P}(\phi_{\mathcal{B}}({\vct{g}},\vct{h})\leq t)$ and $\mathbb{P}(\phi_{\mathbb{R}^p}({\vct{g}},\vct{h})\geq t)$ for a proper choice of $t$.
\noindent{\bf{Case 1: Upper bound analysis}}
\noindent{\bf{Case 2: Lower bound analysis over $\mathcal{B}$}}
Let $\tn{{\vct{g}}}=l_g,\tn{\vct{h}}=l_h$, $\rho=l_h^2/l_g^2$. Set $b=\bar{\h}^T{\boldsymbol{\beta}}$ and $\widetilde\vct{b}=\frac{{\boldsymbol{\beta}}-\bar{\h}\hb^T{\boldsymbol{\beta}}}{\tn{{\boldsymbol{\beta}}-\bar{\h}\hb^T{\boldsymbol{\beta}}}}=\frac{{\boldsymbol{\beta}}-\bar{\h} b}{\tn{{\boldsymbol{\beta}}-\bar{\h} b}}$. Observe that
\[
\tn{{\boldsymbol{\beta}}-\bar{\h} b}^2=1-2b^2+b^2=1-b^2\quad\text{and}\quad {\boldsymbol{\beta}}^T\widetilde\vct{b}=\sqrt{1-b^2}.
\]
Let us represent our variable $\vct{w}$ as
\[
\vct{w}=w\bar{\h}+\gamma\widetilde\vct{b}+\vct{v}
\]
where $v=\tn{\vct{v}}$. Observe that $\bar{\h},\widetilde\vct{b}$ are orthonormal vectors. We consider the objective
\begin{align}
\phi_\mathcal{S}^2({\vct{g}},\vct{h})&=\min_{\vct{w}\in\mathcal{S}}\tn{{\boldsymbol{\beta}}-\vct{w}}^2\quad\text{s.t.}\quad \tn{\vct{w}_\sigma}\tn{{\vct{g}}}\leq \vct{h}^T\vct{w}+\sigma h.
\end{align}
Let us rewrite both sides carefully as follows. Objective can be written as
\begin{align}
\tn{{\boldsymbol{\beta}}-\vct{w}}^2&=\tn{{\boldsymbol{\beta}}-\gamma\widetilde\vct{b}-w\vct{h}+\vct{v}}^2=\tn{({\boldsymbol{\beta}}-\widetilde\vct{b})+(1-\gamma)\widetilde\vct{b}-w\vct{h}-\vct{v}}^2\\
&=w^2+v^2+(1-\gamma)^2+\tn{{\boldsymbol{\beta}}-\widetilde\vct{b}}^2+2\left< {\boldsymbol{\beta}}-\widetilde\vct{b},(1-\gamma)\widetilde\vct{b}-w\vct{h}\right>\\
&=w^2+v^2+(1-\gamma)^2+2(1-\sqrt{1-b^2})+2(1-\gamma)(\sqrt{1-b^2}-1)-2wb\\
&=w^2+v^2+(1-\gamma)^2+2(\gamma (1-\sqrt{1-b^2})-wb).
\end{align}
Here, we will argue that the $b$ term is small and ignorable when $p$ is sufficiently large. Set $h'=\sigma h/l_g$. Next, the constraint can be written as
\begin{align}
\tn{\vct{w}_\sigma}\tn{{\vct{g}}}\leq \vct{h}^T\vct{w}+\sigma h&\iff \sqrt{w^2+\gamma^2+v^2+\sigma^2}l_g\leq wl_h+\sigma h\\
&\iff \sqrt{w^2+\gamma^2+v^2+\sigma^2}\leq w\sqrt{\rho}+h'\\
&\iff \gamma^2+v^2+\sigma^2\leq (\rho-1)w^2+2\sqrt{\rho}h'w+h'^2.
\end{align}
Together these two yield the parameterized loss function (which depends on ${\vct{g}},\vct{h},h$)
\[
L(w,\gamma,v)=w^2+v^2+(1-\gamma)^2+2(\gamma (1-\sqrt{1-b^2})-wb)~\text{s.t.}~\gamma^2+v^2+\sigma^2\leq (\rho-1)w^2+2\sqrt{\rho}h'w+h'^2.
\]
Also define the proxy loss function which is identical to the setup of Lemma \ref{asym str}
\[
L'(w,\gamma,v)=w^2+v^2+(1-\gamma)^2~\text{s.t.}~\gamma^2+v^2+\sigma^2\leq (\rho-1)w^2.
\]
We have the following lemma relating these two loss functions.
\begin{lemma} \label{lem nearby} Fix $\delta>0$. With probability $1-C\mathrm{e}^{-c\delta^2p}$, for any feasible solution $(w,\gamma,v)$ of the loss $L$, there exists a feasible solution $(w',\gamma,v)$ such that $|w-w'|\leq \delta$ and $|L(w,\gamma,v)-(w',\gamma,v)|\leq \delta$ and vice versa (i.e.~similar statement holds for mapping the solutions from $L'$ to $L$).
\end{lemma}
\begin{proof} Fix $(w,\gamma,v)$. The proof of the converse statement is similar.
\end{proof}
Using this lemma, let us first bound $\mathbb{P}(\phi_{\mathbb{R}^p}({\vct{g}},\vct{h})\geq t)$. Via Lemma \ref{asym str}, observe that
\[
\phi_*=\min_{w,\gamma,v}L'(w,\gamma,v)=L'(w_*,\gamma_*,0)
\]
where $w_*,\gamma_*$ is as in \eqref{opt choices} and $\rho=l_h^2/l_g^2$. Via Lemma \ref{lem nearby}, there exists $w'$ such that $|w'-w| \leq\delta$ and $L(w',\gamma_*,0)\leq\phi_*+\delta$.
\end{proof}
\newpage
\section{Overdetermined analysis}
\begin{lemma}[To be proven rigorously] Set ${\bar{p}}=p/n$ and ${\bar{s}}=s/n$. Suppose ${\bar{p}}<1$ and suppose $\vct{t}^T{\boldsymbol{{\Sigma}}}\vct{t}=\theta$. Then
\[
{\cal{L}}({\boldsymbol{\hat{\beta}}})=\sigma^2+w^2=\frac{\sigma^2}{1-{\bar{p}}}\quad\text{and}\quad {\cal{L}}(\boldsymbol{\hat{\theta}})=\frac{1-{\bar{p}}+{\bar{s}}}{1-{\bar{p}}}\sigma^2+\alpha-\theta.
\]
\end{lemma}
\subsection{Analysis}
Since ${\bar{p}}<1$, \eqref{optim me} has a unique solution. Set $\boldsymbol{\omega}=\sqrt{{\boldsymbol{{\Sigma}}}}\vct{w}$ where $\vct{w}={\boldsymbol{\beta}}-{\boldsymbol{\beta}}'$. Define ${\mtx{X}}_{\vct{z}}=[{\mtx{\bar{X}}}~{\vct{z}}]$ and $\boldsymbol{\omega}_\sigma=[\boldsymbol{\omega}^T~\sigma]^T$. This leads to the optimization problem
\[
\hat{\boldsymbol{\omega}}=\arg\min_{\boldsymbol{\omega}} \tn{{\mtx{\bar{X}}}\boldsymbol{\omega}+\sigma {\vct{z}}}=\arg\min_{\boldsymbol{\omega}} \tn{{\mtx{X}}_{\vct{z}}\boldsymbol{\omega}_\sigma}.
\]
Fix ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_p),\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_n)$. Note that that $s/p={\bar{s}}/{\bar{p}}$. Applying Gordon's Lemma (non-rigorously) and noticing optimal $\boldsymbol{\omega}$ has form $\boldsymbol{\omega}=w\bar{{\vct{g}}}$ where $\bar{{\vct{g}}}={\vct{g}}/\tn{{\vct{g}}}$, we find
\begin{align}
\min_{\vct{w}} \max_{\tn{\vct{a}}\leq 1} \vct{h}^T\vct{a} \tn{\boldsymbol{\omega}_{\sigma}}-\tn{\vct{a}}{\vct{g}}^T\boldsymbol{\omega}_{\sigma}&=\min_{\vct{w}} \sqrt{n} \tn{\boldsymbol{\omega}_{\sigma}}-{\vct{g}}^T\boldsymbol{\omega}_{\sigma}\\
&=\min_{\vct{w}} \sqrt{n} \sqrt{w^2+\sigma^2}-w\tn{{\vct{g}}}\\
&\propto\min_{\vct{w}} \sqrt{w^2+\sigma^2}-w\sqrt{{\bar{p}}}.
\end{align}
This implies $w=\sqrt{\frac{{\bar{p}}}{1-{\bar{p}}}}\sigma$ via
\[
\sqrt{{\bar{p}}}=\frac{w}{\sqrt{w^2+\sigma^2}}\iff {\bar{p}} (w^2+\sigma^2)=w^2\iff w=\sqrt{\frac{{\bar{p}}}{1-{\bar{p}}}}\sigma
\]
Now, recall that ${\boldsymbol{\hat{\beta}}}={\boldsymbol{\beta}}+{\boldsymbol{{\Sigma}}}^{-1/2}\boldsymbol{\omega}$. Since the entries of $\boldsymbol{\omega}$ are normally distributed, we have that
\begin{align}
\sqrt{{\boldsymbol{{\Sigma}}}}\boldsymbol{\hat{\theta}}&=\sqrt{{\boldsymbol{{\Sigma}}}}\mathbb{F}_s({\boldsymbol{\hat{\beta}}})=\sqrt{{\boldsymbol{{\Sigma}}}}(\mathbb{F}_s({\boldsymbol{\beta}})+\mathbb{F}_s({\boldsymbol{{\Sigma}}}^{-1/2}\boldsymbol{\omega}))\\
&=\sqrt{{\boldsymbol{{\Sigma}}}}\mathbb{F}_s({\boldsymbol{\beta}})+\mathbb{F}_s(\boldsymbol{\omega})\\
&=\vct{t}+\mathbb{F}_s(\boldsymbol{\omega}).
\end{align}
Note that $\tn{\mathbb{F}_s(\boldsymbol{\omega})}=\sqrt{\frac{s}{p}}\sqrt{\frac{{\bar{p}}}{1-{\bar{p}}}}\sigma=\sqrt{\frac{{\bar{s}}}{1-{\bar{p}}}}\sigma$. Thus the test error \eqref{test formula} of ${\boldsymbol{\hat{\beta}}}$ and $\boldsymbol{\hat{\theta}}$ are
\[
{\cal{L}}({\boldsymbol{\hat{\beta}}})=\sigma^2+w^2=\frac{\sigma^2}{1-{\bar{p}}}\quad\text{and}\quad {\cal{L}}(\boldsymbol{\hat{\theta}})=\frac{{\bar{s}}}{1-{\bar{p}}}\sigma^2+\alpha-\theta+\sigma^2.
\]
\section{Overparameterized analysis}
Let $\lambda_i$ be the $i$th diagonal entry of the covariance matrix ${\boldsymbol{{\Sigma}}}$.
\begin{lemma}[To be proven rigorously] Set ${\bar{p}}=p/n$ and ${\bar{s}}=s/n$. Then, define the quantities
\begin{align}
&\Lambda\quad\text{is solution of}\quad n=\sum_{i=1}^p\frac{1}{1+(\Lambda\lambda_i)^{-1}},\\
&\zeta_i=\frac{1}{1+\Lambda \lambda_i},\\
&\gamma_i=\frac{\Gamma}{n(1+(\Lambda \lambda_i)^{-1})}\quad\text{where}\quad n(\sigma^2+\sum_{i=1}^n \gamma_i^2+\lambda_i\zeta_i^2{\boldsymbol{\beta}}_i^2)=\Gamma^2.
\end{align}
$\bar{\gamma}=\sqrt{p}\gamma,\bar{\Gamma}=\Gamma/\sqrt{p}$
\begin{align}
&\bar{\gamma}_i=\frac{\sqrt{p}\sqrt{p}\bar{\Gamma}}{n(1+(\Lambda \lambda_i)^{-1})}\quad\text{where}\quad n(\sigma^2+\sum_{i=1}^n \frac{\bar{\gamma}_i^2}{p}+\lambda_i\zeta_i^2{\boldsymbol{\beta}}_i^2)=p\bar{\Gamma}^2.
\end{align}
Then, the distribution of $\hat{{\boldsymbol{\beta}}}$ can be approximated as
\[
\hat{{\boldsymbol{\beta}}}\sim (1-\zeta)\odot{\boldsymbol{\beta}}-{\boldsymbol{{\Sigma}}}^{-1/2}\gamma\odot{\vct{g}}.
\]
Then, the test errors are given by
\begin{align}
&{\cal{L}}({\boldsymbol{\hat{\beta}}})=\sigma^2+\sum_{i=1}^p\gamma_i^2+\lambda_i\zeta_i^2{\boldsymbol{\beta}}_i^2\\
&{\cal{L}}(\boldsymbol{\hat{\theta}})=\sigma^2+\sum_{i=1}^s(\gamma_i^2+\lambda_i\zeta_i^2{\boldsymbol{\beta}}_i^2)+\sum_{i=s+1}^p\lambda_i{\boldsymbol{\beta}}_i^2.
\end{align}
\end{lemma}
\subsection{Derivation}
Since ${\bar{p}}>1$, we solve the least-squares problem \eqref{optim me2} subject to the min norm constraint. Using change of variable, this yields
\begin{align}
\min_{\vct{w}} \tn{{\boldsymbol{\beta}}-\vct{w}}\quad\text{subject to}\quad \tn{{\mtx{X}}\vct{w}+\sigma {\vct{z}}}=0.
\end{align}
Noticing $\tn{{\mtx{X}}\vct{w}+\sigma {\vct{z}}}=\tn{{\mtx{\bar{X}}}\sqrt{{\boldsymbol{{\Sigma}}}}\vct{w}+\sigma {\vct{z}}}$, Gordon's Lemma formulation takes the form (non-rigorously)
\begin{align}
\min_{\vct{w}} \tn{{\boldsymbol{\beta}}-\vct{w}}\quad\text{subject to}\quad \sqrt{n}\tn{\sqrt{{\boldsymbol{{\Sigma}}}}\vct{w}~{\sigma}}={\vct{g}}^T\sqrt{{\boldsymbol{{\Sigma}}}}\vct{w}.
\end{align}
Setting $w_i=\frac{\gamma_i}{\sqrt{\lambda_i}}g_i+\zeta_i{\boldsymbol{\beta}}_i$, we have $\operatorname{\mathbb{E}}[{\vct{g}}^T\sqrt{{\boldsymbol{{\Sigma}}}}\vct{w}]=\operatorname{\mathbb{E}}[\sum_{i=1}^p\gamma_ig_i^2]=\sum_{i=1}^p\gamma_i:=\Gamma$. Taking expectations, we can write
\begin{align}
\min_{\gamma_i} \sum_{i=1}^p (1-\zeta_i)^2{\boldsymbol{\beta}}_i^2+\frac{\gamma_i^2}{\lambda_i} \quad\text{subject to}\quad n(\sigma^2+\sum_{i=1}^n \gamma_i^2+\lambda_i\zeta_i^2{\boldsymbol{\beta}}_i^2)=\Gamma^2.
\end{align}
In Lagrangian form, we obtain
\begin{align}
\min_{\gamma_i} \sum_{i=1}^p (1-\zeta_i)^2{\boldsymbol{\beta}}_i^2+\frac{\gamma_i^2}{\lambda_i} +\Lambda [(\sigma^2+\sum_{i=1}^n \gamma_i^2+\lambda_i\zeta_i^2{\boldsymbol{\beta}}_i^2)-\frac{\Gamma^2}{n}].
\end{align}
Differentiating with respect to $\gamma_i,\zeta_i$, we obtain the equations
\begin{align}
\frac{\gamma_i}{\lambda_i}+\Lambda [\gamma_i-\Gamma/n]=0&\iff \gamma_i=\frac{\Gamma}{n(1+(\Lambda\lambda_i)^{-1})}\\
(\zeta_i-1){\boldsymbol{\beta}}_i^2+\Lambda \lambda_i{\boldsymbol{\beta}}_i^2\zeta_i=0&\iff \zeta_i=\frac{1}{1+\Lambda \lambda_i}.
\end{align}
$\Lambda$ has to satisfy the inequality
\begin{align}
n=\sum_{i=1}^p\frac{1}{1+(\Lambda\lambda_i)^{-1}}\iff \sum_{i=1}^p\gamma_i=\Gamma\sum_{i=1}^p\frac{1}{n(1+(\Lambda\lambda_i)^{-1})}.
\end{align}
Finally, we pick $\Gamma>0$ and $\gamma_i=\frac{\Gamma}{n(1+(\Lambda \lambda_i)^{-1})}$ such that the following equality is satisfied
\[
n(\sigma^2+\sum_{i=1}^n \gamma_i^2+\lambda_i\zeta_i^2{\boldsymbol{\beta}}_i^2)=\Gamma^2.
\]
Finally, fixing a fresh example $\vct{x}=\sqrt{{\boldsymbol{{\Sigma}}}}\bar{\vct{x}}$, the test error is given as follows
\[
{\cal{L}}({\boldsymbol{\hat{\beta}}})=\sigma^2+\tn{\sum_{i=1}^p \bar{x}_i\sqrt{\lambda_i}(\frac{\gamma_i}{\sqrt{\lambda_i}}g_i+\zeta_i{\boldsymbol{\beta}}_i)}^2=\sigma^2+\sum_{i=1}^p\gamma_i^2+\lambda_i\zeta_i^2{\boldsymbol{\beta}}_i^2.
\]
To proceed, the error of the sparse model can be calculated as follows.
\[
{\cal{L}}(\boldsymbol{\hat{\theta}})=\sigma^2+\tn{\sum_{i=1}^s\bar{x}_i \sqrt{\lambda_i}(\frac{\gamma_i}{\sqrt{\lambda_i}}g_i+\zeta_i{\boldsymbol{\beta}}_i)+\sum_{i=s+1}^p\bar{x}_i\sqrt{\lambda_i}{\boldsymbol{\beta}}_i}^2=\sigma^2+\sum_{i=1}^s(\gamma_i^2+\lambda_i\zeta_i^2{\boldsymbol{\beta}}_i^2)+\sum_{i=s+1}^p\lambda_i{\boldsymbol{\beta}}_i^2.
\]
\section{Linear Random Features}
\subsection{When $k>p$}
Let ${\mtx{F}}\in\mathbb{R}^{k\times p}$ be the feature matrix. In this case, we solve
\[
\min_{\vct{t}}\tn{\vct{y}-{\mtx{X}}{\mtx{F}}^T\vct{t}}=\min_{\vct{t}}\tn{{\mtx{X}}{\boldsymbol{\beta}}+\sigma{\vct{z}}-{\mtx{X}}{\mtx{F}}^T\vct{t}}
\]
Let ${\boldsymbol{\hat{\beta}}}$ be the min-norm solution of $\min_{{\boldsymbol{\beta}}'}\tn{\vct{y}-{\mtx{X}}{\boldsymbol{\beta}}'}$. Since ${\mtx{F}}$ is unitary, we have that
\[
\boldsymbol{\hat{\theta}}={\mtx{F}}{\boldsymbol{\hat{\beta}}}.
\]
Thus random features has the effect of diluting the target vector before thresholding. The norm of ${\boldsymbol{\hat{\beta}}}$ is preserved but each coordinate has less say in the decision. Specifically, the sparsified $\boldsymbol{\hat{\theta}}_{\text{sparse}}=\mathbb{T}_s(\boldsymbol{\hat{\theta}})$ leads to the error
\[
{\cal{L}}(\boldsymbol{\hat{\theta}}_{\text{sparse}})=\tn{{\boldsymbol{\beta}}-{\mtx{F}}^T\mathbb{T}_s({\mtx{F}}{\boldsymbol{\hat{\beta}}})}^2
\]
Let $\vct{w}={\boldsymbol{\beta}}-{\boldsymbol{\beta}}'$. Draw $\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p),h\sim\mathcal{N}(0,1)$ and ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n)$. Consider the primary optimization (PO)
\begin{align}
\Phi({\mtx{X}})&=\min_{\vct{w}}\tn{{\mtx{X}}\vct{w}+\sigma {\vct{z}}}\\
&=\min_{\vct{w}}\max_{\tn{\vct{a}}\leq 1}\vct{a}^T{\mtx{X}}_{{\vct{z}}}\vct{w}_{\sigma},
\end{align}
\begin{align}
\Phi({\mtx{X}})&=\min_{\vct{w}}0.5\lambda\tn{{\mtx{X}}_{\vct{z}}\vct{w}_\sigma}^2+0.5\tn{{\boldsymbol{\beta}}-\vct{w}}^2\\
&=\min_{\vct{w}}\max_{\vct{a}}\lambda\vct{a}^T{\mtx{X}}_{{\vct{z}}}\vct{w}_{\sigma}-0.5\lambda\tn{\vct{a}}^2+0.5\tn{{\boldsymbol{\beta}}-\vct{w}}^2
\end{align}
Differentiating the first line, we obtain
\[
\vct{w}-{\boldsymbol{\beta}}+\lambda {\mtx{X}}^T({\mtx{X}}\vct{w}+\sigma {\vct{z}})=0\implies \vct{w}=(\lambda {\mtx{X}}^T{\mtx{X}}+{\mtx{I}})^{-1}({\boldsymbol{\beta}}-\sigma\lambda{\mtx{X}}^T{\vct{z}})
\]
{\bf{Opt 1:}} and the corresponding auxiliary optimization (AO)
\begin{align}
\phi({\vct{g}},\vct{h})&=\min_{\vct{w}}\max_{\vct{a}} \lambda(\tn{\vct{w}_{\sigma}}{\vct{g}}^T\vct{a}+\tn{\vct{a}}(\vct{w}^T\vct{h}+\sigma h)-0.5\tn{\vct{a}}^2)+0.5\tn{{\boldsymbol{\beta}}-\vct{w}}^2\\
&=\min_{\vct{w}}\max_{a} \lambda(a\tn{\vct{w}_{\sigma}}\tn{{\vct{g}}}+a(\vct{w}^T\vct{h}+\sigma h)-0.5a^2)+0.5\tn{{\boldsymbol{\beta}}-\vct{w}}^2\\
&=\min_{\vct{w}} 0.5\lambda(\tn{\vct{w}_{\sigma}}\tn{{\vct{g}}}+(\vct{w}^T\vct{h}+\sigma h))^2+0.5\tn{{\boldsymbol{\beta}}-\vct{w}}^2.
\end{align}
\[
\vct{w}-{\boldsymbol{\beta}}+\lambda(\vct{h}+\frac{\tn{{\vct{g}}}}{\sqrt{\tn{\vct{w}}^2+\sigma^2}}\vct{w})(\tn{\vct{w}_{\sigma}}\tn{{\vct{g}}}+(\vct{w}^T\vct{h}+\sigma h))=0.
\]
{\bf{Opt 2:}} and the corresponding auxiliary optimization (AO)
\begin{align}
\phi({\vct{g}},\vct{h})&=\min_{\vct{w}}\max_{\vct{a}} \tn{\vct{w}_{\sigma}}{\vct{g}}^T\vct{a}+\tn{\vct{a}}(\vct{w}^T\vct{h}+\sigma h)\\\
&=\min_{\vct{w}}(\tn{\vct{w}_{\sigma}}\tn{{\vct{g}}}+\vct{w}^T\vct{h}+\sigma h)_+\\
&=(\min_{\vct{w}}\sqrt{\tn{\vct{w}}^2+\sigma^2}\tn{{\vct{g}}}+\vct{w}^T\vct{h}+\sigma h)_+
\end{align}
Regularized version becomes
\begin{align}
\phi({\vct{g}},\vct{h})&=\lambda (\min_{\vct{w}}\sqrt{\tn{\vct{w}}^2+\sigma^2}\tn{{\vct{g}}}+\vct{w}^T\vct{h}+\sigma h)_++\tn{{\boldsymbol{\beta}}-\vct{w}}^2.
\end{align}
Differentiating, we have two options
\begin{align}
\vct{w}-{\boldsymbol{\beta}}+\lambda(\vct{h}+\frac{\tn{{\vct{g}}}}{\sqrt{\sigma^2+\tn{\vct{w}}^2}}\vct{w})=0.
\end{align}
\[
(1+\lambda \frac{l_g}{\sqrt{\sigma^2+w^2}})\vct{w}={\boldsymbol{\beta}}-\lambda \vct{h}.
\]
Decompose $\vct{w}=w\omega$ where $\omega=\frac{{\boldsymbol{\beta}}-\lambda \vct{h}}{\tn{{\boldsymbol{\beta}}-\lambda \vct{h}}}$. We find
\[
(1+\lambda \frac{l_g}{\sqrt{\sigma^2+w^2}})w=\tn{{\boldsymbol{\beta}}-\lambda \vct{h}}.
\]
Primary problem is
\[
\min_{{\boldsymbol{\beta}}}\lambda \tn{\vct{y}-{\mtx{X}}{\boldsymbol{\beta}}}+0.5\tn{{\boldsymbol{\beta}}}^2\implies {\boldsymbol{\beta}}+\lambda \frac{{\mtx{X}}^T({\mtx{X}}{\boldsymbol{\beta}}-\vct{y})}{\tn{{\mtx{X}}{\boldsymbol{\beta}}-\vct{y}}}=0
\]
Setting $\lambda'=\lambda/\tn{{\mtx{X}}{\boldsymbol{\beta}}-\vct{y}}$
\[
{\boldsymbol{\beta}}=(\frac{{\mtx{I}}}{\lambda'}+{\mtx{X}}^T{\mtx{X}})^{-1}{\mtx{X}}^T\vct{y}
\]
\newpage
and the corresponding auxiliary optimization (AO)
\begin{align}
\phi({\vct{g}},\vct{h})&=\min_{\vct{w}}\max_{\vct{a}} \tn{\vct{w}_{\sigma}}{\vct{g}}^T\vct{a}+\tn{\vct{a}}(\vct{w}^T\vct{h}+\sigma h)\\\
&=\min_{\vct{w}}(\tn{\vct{w}_{\sigma}}\tn{{\vct{g}}}+\vct{w}^T\vct{h}+\sigma h)_+\\
&=(\min_{\vct{w}}\sqrt{\tn{\vct{w}}^2+\sigma^2}\tn{{\vct{g}}}+\vct{w}^T\vct{h}+\sigma h)_+
\end{align}
Additionally denoting the pseudo-inverse by ${\mtx{X}}^+$, define the minimum norm solution of PO via
\begin{align}
\vct{w}_\Phi({\mtx{X}})&={\boldsymbol{\beta}}-{\mtx{X}}^+({\mtx{X}}{\boldsymbol{\beta}}+\sigma{\vct{z}})\\
&=\begin{cases}\arg\min_{\vct{w}}\tn{{\mtx{X}}\vct{w}+\sigma {\vct{z}}}\quad \text{if}\quad n\geq p\\\arg\min_{\vct{w}}\tn{{\boldsymbol{\beta}}-\vct{w}}\quad\text{subject to}\quad{\mtx{X}}\vct{w}+\sigma {\vct{z}}=0\quad\text{if}\quad n<p\end{cases}
\end{align}
\subsection{Overparameterized analysis}
\subsubsection{General idea - overparameterized line of attack}
Suppose $\Phi({\mtx{X}})=0$ and we are interested in
\begin{align}
{\cal{L}}_{\text{PO}}'(\vct{w})&=\min_{\vct{w}} \|\vct{w}\|\quad\text{subject to}\quad {\cal{L}}_{\text{PO}}(\vct{w},{\mtx{X}})=0\\
&=\min_{\vct{w}}\max_{\lambda} \|\vct{w}\|+\lambda{\cal{L}}_{\text{PO}}(\vct{w},{\mtx{X}}).
\end{align}
\subsubsection{Proof stuff}
Fix a scalar $\theta>0$ and define the set $\Theta=\{x{~\big |~} \theta(1-\varepsilon)\leq x\leq \theta(1+\varepsilon)\}$. Consider the optimization
\[
\min_{\tn{{\boldsymbol{\beta}}}\not\in \Theta}\tn{{\mtx{X}}\vct{w}}
\]
We would like to assess the properties of the solution ${\boldsymbol{\hat{\beta}}}$.
In this case, we solve least squares subject to min norm constraint. Gordon takes the form
\begin{align}
\min_{{\boldsymbol{\beta}}'} \tn{{\boldsymbol{\beta}}'}\quad\text{subject to}\quad \sqrt{n}\tn{\vct{w}_{\sigma}}={\vct{g}}^T\vct{w}_{\sigma}.
\end{align}
This time $\vct{w}$ will choose the form
\[
\vct{w}=w\bar{{\vct{g}}}-\gamma {\boldsymbol{\beta}}.
\]
which yields the scalar problem
\[
\min_{w,\gamma} \alpha^2(1-\gamma)^2+w^2\quad\text{subject to}\quad (w^2+\alpha^2\gamma^2+\sigma^2)=w^2\kappa.
\]
This means
\[
w^2=\frac{\alpha^2\gamma^2+\sigma^2}{\kappa-1}\implies \min_{\gamma} \alpha^2(1-\gamma)^2+\frac{\alpha^2\gamma^2+\sigma^2}{\kappa-1}
\]
This implies
\[
1-\gamma=\frac{\gamma}{\kappa-1}\implies 1=\gamma(1+\frac{1}{\kappa-1})\implies \gamma=\frac{\kappa-1}{\kappa}.
\]
In terms of $w,\gamma$,
\begin{align}
&{\boldsymbol{\hat{\beta}}}=(1-\gamma){\boldsymbol{\beta}}+w\bar{{\vct{g}}}=\frac{{\boldsymbol{\beta}}}{\kappa}+w\bar{{\vct{g}}}\\
&\boldsymbol{\hat{\theta}}=(1-\gamma)\vct{t}+w\bar{{\vct{g}}}[1:s]
\end{align}
Thus,
\begin{align}
&{\boldsymbol{\beta}}-{\boldsymbol{\hat{\beta}}}={\boldsymbol{\beta}}(1-\frac{1}{\kappa})+{\boldsymbol{\beta}}[k:]+w\bar{{\vct{g}}}\\
&{\boldsymbol{\beta}}-\boldsymbol{\hat{\theta}}={\boldsymbol{\beta}}[1:s](1-\frac{1}{\kappa})+{\boldsymbol{\beta}}[s:]+w\bar{{\vct{g}}}[1:s]
\end{align}
Secondly, note that
\[
w^2=\frac{\alpha^2\gamma^2+\sigma^2}{\kappa-1}=\frac{\sigma^2}{\kappa-1}+\frac{\alpha^2(\kappa-1)}{\kappa^2}
\]
Thus the test error of ${\boldsymbol{\hat{\beta}}}$ is
\[
err({\boldsymbol{\hat{\beta}}})=\sigma^2+w^2+\alpha^2(1-\frac{1}{\kappa})^2=\frac{\kappa \sigma^2}{\kappa-1}+\alpha^2\frac{\kappa-1}{\kappa}
\]
Similarly, the error will be proportionally distributed over $\boldsymbol{\hat{\theta}}$ thus test error of $\boldsymbol{\hat{\theta}}$ is
\[
err(\boldsymbol{\hat{\theta}})=1-\theta^2+\bar{\sigma}^2+\frac{\xi}{\kappa}w^2+\theta^2(1-\frac{1}{\kappa})^2=1-\theta^2+\bar{\sigma}^2+\frac{\xi (1-\alpha^2+\bar{\sigma}^2)}{\kappa(\kappa-1)}+\frac{\kappa-1}{\kappa^2}(\theta^2(\kappa-1)+\frac{\xi}{\kappa}\alpha^2)
\]
\section{Intuitions for Understanding the Benefits of Overparameterization and Retraining}\label{sec intu}
So far we have a couple of interesting conclusions. First, solving overparameterized problem and then pruning can be better than pruning with optimal sparse features. Secondly, retraining may hurt the performance in theory however it helps the performance for neural nets and random features. This section aims to provide further insights into these.
\noindent\textbf{Denoising effect of overparameterization:} Consider a simple linear model with noise level $\sigma=0$, $n\geq p\gg s$ and identity covariance. Pick index set $\Delta\subset[p]$ with $|\Delta|=s$. Suppose we wish to estimate the coefficients ${\boldsymbol{\beta}}^\star_{\Delta}$. For pruning, we can pick $\Delta$ to be the most salient/largest entries. If we solve the smaller problem over $\Delta$, ${\boldsymbol{\hat{\beta}}}[\Delta]$ will only provide a noisy estimate of ${\boldsymbol{\beta}}^\star_{\Delta}$. The reason is that, the signal energy of missing features $[p]-\Delta$ act as noise uncorrelated with features over $\Delta$. Conversely, if we solve ERM with all features (the larger problem), we perfectly recover ${\boldsymbol{\beta}}^\star$ due to lack of noise and $n>p$ from which we perfectly estimate ${\boldsymbol{\beta}}^\star_{\Delta}$. This simple argument, which is in similar spirit to \cite{}, shows that solving the larger problem with more parameters can have a denoising like effect and perform better than the small problem. It also explains why retraining again with few features can hurt the performance. Our contribution obviously goes well beyond this discussion and theoretically characterizes exact asymptotics and also the important overparameterized regime $n\ll p$.
\som{also see SM for DNN experiments}
\noindent\textbf{Why \& When Retraining Helps:} In practice, as seen in the RFR experiments of Fig.~\ref{figRF2}, retraining improves the performance. Here, we argue that the benefit of retraining is connected to the correlations among input features. Indeed, covariance/Hessian associated with RF and DNN regressions are not diagonal matrices. Imagine only a single feature is sufficient to explain the label. If there are multiple other features that can similarly explain the label, the prediction of ${\boldsymbol{\hat{\beta}}}$ will be shared across these features. Then, pruning will lead to a biased estimate which can be mitigated by retraining. The following lemma formalizes this and demonstrates improvement thanks to retraining. This lemma considers a simple setup where the features are perfectly correlated via a rank-1 covariance
\cmt{
train$\rightarrow$prune With growing correlations, the This is in contrast to the diagonal covariance with orthogonal features where retraining may hurt. wheredemonstrate a simple scenario which shows that retraining can actually
retraining is critical for training sparse neural networks \cite{frankle2019lottery}. In this section, we identify feature correlation as a critical factor that necessitates retraining. This is in contrast to our results in Section \ref{doubdec} which considers independent features.
In order to convey the point, let us focus on a simplistic scenario where covariance matrix is rank one and features are perfectly correlated with each other. In this case, the diagonal entries of the covariance matrix essentially dictates the form of the solution.
}
\begin{lemma} Suppose $\mathcal{S}$ is drawn from an LGP$(\sigma,{\boldsymbol{{\Sigma}}},{\boldsymbol{\beta}}_\star)$ as in Def.~\ref{def LGP} where $\text{rank}({\boldsymbol{{\Sigma}}})=1$ with ${\boldsymbol{{\Sigma}}}=\boldsymbol{\lambda}\bla^T$ for $\boldsymbol{\lambda}\in\mathbb{R}^p$. Define $\zeta=\mathbb{T}_s(\boldsymbol{\lambda})^2/\tn{\boldsymbol{\lambda}}^2$. For magnitude and Hessian pruning ($P\in\{M,H\}$) and the associated retraining, we have the following excess risks with respect to ${\boldsymbol{\beta}}^\star
\begin{align
&\operatorname{\mathbb{E}}_{\mathcal{S}}[{\cal{L}}({\boldsymbol{\hat{\beta}}}_s^P)]-{\cal{L}}({\boldsymbol{\beta}}^\star)={\frac{\zeta^2\sigma^2}{n-2}}+\underbrace{(1-\zeta)^2(\boldsymbol{\lambda}^T{\boldsymbol{\beta}}^\star)^2}_{\text{Error due to bias}}\\
&\operatorname{\mathbb{E}}_{\mathcal{S}}[{\cal{L}}({\boldsymbol{\hat{\beta}}}_s^{RT})]-{\cal{L}}({\boldsymbol{\beta}}^\star)=\frac{\sigma^2}{n-2}.
\end{align}
\end{lemma}
In essence, this lemma says that pruning the model leads to a biased estimator of the label where the bias coefficient $1-\zeta$ arises from the missing predictions of pruned features. These correspond to the small coefficients of $\boldsymbol{\lambda}$. In contrast, regardless of $s$, retraining always results in an unbiased estimator with the exact same risk as the dense model which quickly decays in sample size $n$. The reason is that retraining gives the remaining features to re-explain the label and account for the missing predictions and they can do it well since features are perfectly correlated. While this lemma provides insights for a simplistic model, we emphasize that our distributional predictions in the next section enables us to capture the retraining benefits for arbitrary covariance matrices
\cmt{
\begin{proof}
Set $c^\star=\boldsymbol{\lambda}^T{\boldsymbol{\beta}}^\star$. By definition, each input example $\vct{x}_i$ has the form $\vct{x}_i=g_i\boldsymbol{\lambda}$ and $y_i=g_ic^\star+\sigma z_i$. Set ${\vct{g}}=[g_1~\dots~g_n]^T$ and $\bar{{\vct{g}}}={\vct{g}}/\tn{{\vct{g}}}$. Thus, we have ${\mtx{X}}={\vct{g}}\boldsymbol{\lambda}^T$ and $\vct{y}={\vct{g}}\boldsymbol{\lambda}^T{\boldsymbol{\beta}}^\star+\sigma{\vct{z}}$. Decompose ${\vct{z}}=\bar{{\vct{z}}}+\bar{{\vct{g}}}^T{\vct{z}}\bar{{\vct{g}}}$. The least-squares solution has the form ${\boldsymbol{\hat{\beta}}}=\hat{c}\boldsymbol{\lambda}/\tn{\boldsymbol{\lambda}}^2$ where
\begin{align}
&\hat{c}=\arg\min_{c}\tn{(c^\star-c){\vct{g}}+\sigma{\vct{z}}}\implies\\
&\hat{c}=c^\star+\sigma\gamma\label{beta}.
\end{align}
where $\gamma=\frac{\bar{{\vct{g}}}^T{\vct{z}}}{\tn{{\vct{g}}}}$. Set $h$, $\epsilon\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,1)$. Note that
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}})=\mathbb{E}[((c^\star-\hat{c})h+\sigma\epsilon)^2]
=(\gamma^2+1)\sigma^2
\end{align}
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}})&=\mathbb{E}[(c^\star- \hat{c}\frac{\boldsymbol{\lambda}^T{\vct{u}}}{\tn{\boldsymbol{\lambda}}^2})h+\sigma\epsilon)^2]\\
&=\mathbb{E}[(c^\star- \hat{c}\zeta)h+\sigma\epsilon)^2]\\
&=((1-\zeta)c^\star-\zeta\sigma\gamma)^2
\end{align}
Now let $|\lambda_{k_1}|\geq|\lambda_{k_2}|\geq...\geq|\lambda_{k_p}|$. Assume we solved ERM to get $\hat{c}$ and ${\boldsymbol{\hat{\beta}}}=\hat{c}\boldsymbol{\lambda}$. Prune the trained model ${\boldsymbol{\hat{\beta}}}$ to $s$-sparse by keeping the largest entries. Set ${\vct{u}}=\mathbb{T}_s(\boldsymbol{\lambda})$ and $\zeta=\tn{{\vct{u}}}^2/\tn{\boldsymbol{\lambda}}^2$. We get ${\boldsymbol{\hat{\beta}}}_s=\hat{c}{\vct{u}}$.
\begin{align}
{\cal{L}}(\boldsymbol{\hat{\theta}})&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-{\vct{u}}^T\boldsymbol{\hat{\theta}})h+\sigma\epsilon)^2]\\
&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\sum_{i=1}^s\lambda_{k_i}^2\hat{c})h+\sigma\epsilon)^2]\\
&=[(1-\zeta)\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\zeta\sigma\gamma]^2+\sigma^2
\end{align}
\textbf{Case 1: }
Now, assume that we have already known the optimal entries and do pruning before training. Then the pruned model can be seen as pruning the inputs and then putting it through a non-sparse $s$-feature model. Let $\lambda_{t_1}{\boldsymbol{\beta}}_{t_1}\geq\lambda_{t_2}{\boldsymbol{\beta}}_{t_2}\geq...\geq\lambda_{t_s}{\boldsymbol{\beta}}_{t_s}$, $\vct{w}=[\lambda_{t_1}~\lambda_{t_2}~...~\lambda_{t_s}]$ and $\vct{t}=[{\boldsymbol{\beta}}_{t_1}~...~{\boldsymbol{\beta}}_{t_s}]$. Set data sample $\vct{x}_i'\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,{\boldsymbol{{\Sigma}}}')\in\mathbb{R}^s$ where are ${\boldsymbol{{\Sigma}}}'=\vct{w}\w^T$. Following what done in \ref{beta}, similarly
\[
\boldsymbol{\hat{\theta}}'=\theta\vct{w}\quad\text{where}\quad\theta=\frac{1}{\sum_{i=1}^s\lambda_{t_i}^2}(\vct{w}^T\vct{t}+\frac{\sigma}{\tn{{\vct{g}}}^2}{\vct{g}}^T{\vct{z}})
\]
\begin{align}
{\cal{L}}(\boldsymbol{\hat{\theta}}')&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\sum_{i=1}^s\lambda_{t_i}^2\theta)h+\sigma\epsilon)^2]\\
&=[(\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\vct{w}^T\vct{t})-\sigma\gamma]^2+\sigma^2
\end{align}
\textbf{Case 2: }
Now retrain with the pruned entries $\lambda_{k_!}$, ..., $\lambda_{k_s}$. Then the pruned model can be seen as pruning the inputs and then putting it through a non-sparse $s$-feature model. Let $\vct{t}=[{\boldsymbol{\beta}}_{k_1}~...~{\boldsymbol{\beta}}_{k_s}]$. Set data sample $\vct{x}_i'\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,{\boldsymbol{{\Sigma}}}')\in\mathbb{R}^s$ where are ${\boldsymbol{{\Sigma}}}'={\vct{u}}\ub^T$. Following what done in \ref{beta}, similarly
\[
\boldsymbol{\hat{\theta}}'=\theta\vct u\quad\text{where}\quad\theta=\frac{1}{\sum_{i=1}^s\lambda_{k_i}^2}(\vct u^T\vct{t}+\frac{\sigma}{\tn{{\vct{g}}}^2}{\vct{g}}^T{\vct{z}})
\]
\begin{align}
{\cal{L}}(\boldsymbol{\hat{\theta}}')&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\sum_{i=1}^s\lambda_{k_i}^2\theta)h+\sigma\epsilon)^2]\\
&=[(\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\vct u^T\vct{t})-\sigma\gamma]^2+\sigma^2
\end{align}
\textbf{It is obvious to know that loss in case 1 is always smaller than it in case 2. Then for both case, }
Set ${\boldsymbol{\beta}}^T{\boldsymbol{{\Sigma}}}{\boldsymbol{\beta}}=\alpha^2$ and $\vct{t}^T{\boldsymbol{{\Sigma}}}'\vct{t}=\rho^2$. We can write ${\cal{L}}(\boldsymbol{\hat{\theta}})$ and ${\cal{L}}(\boldsymbol{\hat{\theta}}')$ as
\[
{\cal{L}}(\boldsymbol{\hat{\theta}})=((1-\zeta)\alpha-\zeta\sigma\gamma)^2+\sigma^2\iff{\cal{L}}(\boldsymbol{\hat{\theta}}')=(\alpha-\rho-\sigma\gamma)^2+\sigma^2
\]
Now consider a special case which satisfies ${\cal{L}}(\boldsymbol{\hat{\theta}})<{\cal{L}}(\boldsymbol{\hat{\theta}}')$.
\[
0\leq(1-\zeta)\alpha-\zeta\sigma\gamma<\alpha-\rho-\sigma\gamma
\]
\[
\frac{\rho+\sigma\gamma}{\alpha+\sigma\gamma}<\zeta\leq\frac{\alpha}{\alpha+\sigma\gamma}
\]
\end{proof}}
\cmt{
\begin{proof} Let ${\boldsymbol{{\Sigma}}}=\boldsymbol{\lambda}\bla^T$. Each $\vct{x}_i$ has the form $\vct{x}_i=g_i\boldsymbol{\lambda}$ and $y_i=x_i{\boldsymbol{\beta}}^T\boldsymbol{\lambda}+\sigma z_i$. Thus, we have ${\mtx{X}}={\vct{g}}\boldsymbol{\lambda}$ and $\vct{y}={\vct{g}}\boldsymbol{\lambda}^T{\boldsymbol{\beta}}+\sigma{\vct{z}}$. Let ${\vct{z}}=\bar{{\vct{z}}}+\frac{{\vct{g}}^T{\vct{z}}}{\tn{{\vct{g}}}^2}{\vct{g}}$. The least-squares solution has the form ${\boldsymbol{\hat{\beta}}}=\hat{c}\boldsymbol{\lambda}$ where
\[
\hat{c}
\]
\end{proof}}
\subsection{Further Intuitions on The Denoising Effect of Overparameterization}
To provide further insights into the pruning benefits of overparameterization, consider a simple linear model (as in Def~\ref{def LGP}) with $n\geq p\geq s$, noise level $\sigma=0$ and identity covariance. Suppose our goal is estimating the coefficients ${\boldsymbol{\beta}}^\star_{\Delta}$ for some fixed index set $\Delta\subset[p]$ with $|\Delta|=s$. For pruning, we can pick $\Delta$ to be the most salient/largest entries. If we solve the smaller regression problem over $\Delta$, ${\boldsymbol{\hat{\beta}}}(\Delta)$ will only provide a noisy estimate of ${\boldsymbol{\beta}}^\star_{\Delta}$. The reason is that, the signal energy of the missing features $[p]-\Delta$ acts as a noise uncorrelated with the features in $\Delta$. Conversely, if we solve ERM with all features (the larger problem), we perfectly recover ${\boldsymbol{\beta}}^\star$ due to zero noise and invertibility ($n\geq p$). Then one can also perfectly estimate ${\boldsymbol{\beta}}^\star_{\Delta}$. This simple argument, which is partly inspired by the missing feature setup in \cite{hastie2019surprises}, shows that solving the larger problem with more parameters can have a ``denoising-like effect" and perform better than the small problem. Our contribution obviously goes well beyond this discussion and theoretically characterizes the exact asymptotics, handles the general covariance model and all $(n,p,s)$ regimes, and also highlights the importance of the overparameterized regime $n\ll p$.
\section{Understanding the Benefits of Retraining}\label{sec intu}
\cmt{\noindent\textbf{Denoising effect of overparameterization:} Consider a simple linear model with noise level $\sigma=0$, $n\geq p\gg s$ and identity covariance. Pick index set $\Delta\subset[p]$ with $|\Delta|=s$. Suppose we wish to estimate the coefficients ${\boldsymbol{\beta}}^\star_{\Delta}$. For pruning, we can pick $\Delta$ to be the most salient/largest entries. If we solve the smaller problem over $\Delta$, ${\boldsymbol{\hat{\beta}}}[\Delta]$ will only provide a noisy estimate of ${\boldsymbol{\beta}}^\star_{\Delta}$. The reason is that, the signal energy of the missing features $[p]-\Delta$ act as noise uncorrelated with features over $\Delta$. Conversely, if we solve ERM with all features (the larger problem), we perfectly recover ${\boldsymbol{\beta}}^\star$ due to lack of noise and $n>p$ from which we perfectly estimate ${\boldsymbol{\beta}}^\star_{\Delta}$. This simple argument, which is in similar spirit to \cite{}, shows that solving the larger problem with more parameters can have a denoising like effect and perform better than the small problem. It also explains why retraining again with few features can hurt the performance. Our contribution obviously goes well beyond this discussion and theoretically characterizes exact asymptotics and also the important overparameterized regime $n\ll p$.}
On the one hand, the study of LGPs in \fx{Fig.~\ref{fig1} and Fig.~7 of SM} suggest that retraining can actually hurt the performance. On the other hand, in practice and in the RFR experiments of Fig.~\ref{figRF2}, retraining is crucial; compare the green ${\boldsymbol{\hat{\beta}}}^M$ and red ${\boldsymbol{\hat{\beta}}}^{RT}$ curves and see SM Section A for further DNN experiments. Here, we argue that the benefit of retraining is connected to the correlations between input features. Indeed, the covariance/Hessian matrices associated with RF and DNN regression are not diagonal (as was the case in Fig.~\ref{fig1}). To build intuition, imagine that only a single feature suffices to explain the label. If there are multiple other features that can similarly explain the label, the model prediction will be shared across these features. Then, pruning will lead to a biased estimate, which can be mitigated by retraining. The following lemma formalizes this intuition under an instructive setup, where the features are perfectly correlated
\cmt{
train$\rightarrow$prune With growing correlations, the This is in contrast to the diagonal covariance with orthogonal features where retraining may hurt. wheredemonstrate a simple scenario which shows that retraining can actually
retraining is critical for training sparse neural networks \cite{frankle2019lottery}. In this section, we identify feature correlation as a critical factor that necessitates retraining. This is in contrast to our results in Section \ref{doubdec} which considers independent features.
In order to convey the point, let us focus on a simplistic scenario where covariance matrix is rank one and features are perfectly correlated with each other. In this case, the diagonal entries of the covariance matrix essentially dictates the form of the solution.
}
\begin{lemma} \label{lem rank one}Suppose $\mathcal{S}$ is drawn from an LGP$(\sigma,{\boldsymbol{{\Sigma}}},\boldsymbol{\beta}^\star)$ as in Def.~\ref{def LGP} where $\text{rank}({\boldsymbol{{\Sigma}}})=1$ with ${\boldsymbol{{\Sigma}}}=\boldsymbol{\lambda}\bla^T$ for $\boldsymbol{\lambda}\in\mathbb{R}^p$. Define $\zeta=\mathbb{T}_s(\boldsymbol{\lambda})^2/\tn{\boldsymbol{\lambda}}^2$. For magnitude and Hessian pruning ($P\in\{M,H\}$) and the associated retraining, we have the following excess risks with respect to ${\boldsymbol{\beta}}^\star
\begin{align
&\operatorname{\mathbb{E}}_{\mathcal{S}}[{\cal{L}}({\boldsymbol{\hat{\beta}}}_s^P)]-{\cal{L}}({\boldsymbol{\beta}}^\star)={\frac{\zeta^2\sigma^2}{n-2}}+\underbrace{(1-\zeta)^2(\boldsymbol{\lambda}^T\boldsymbol{\beta}^\star)^2}_{\text{Error due to bias}}\\
&\operatorname{\mathbb{E}}_{\mathcal{S}}[{\cal{L}}({\boldsymbol{\hat{\beta}}}_s^{RT})]-{\cal{L}}({\boldsymbol{\beta}}^\star)={\sigma^2}/({n-2}).
\end{align}
\end{lemma}
The lemma reveals that pruning the model leads to a biased estimator of the label. Specifically, the bias coefficient $1-\zeta$ arises from the missing predictions of the pruned features (which correspond to the small coefficients of $|\boldsymbol{\lambda}|$). In contrast, regardless of $s$, retraining always results in an unbiased estimator with the exact same risk as the dense model which quickly decays in sample size $n$. The reason is that retraining enables the remaining features to account for the missing predictions. Here, this is accomplished perfectly, due to the fully correlated nature of the problem. {In particular, this is in contrast to the diagonal covariance (Fig.~\ref{fig1}), where the missing features act like uncorrelated noise during retraining.} \cmt{Our distributional predictions in the next section enable us to capture the benefits of retraining for arbitrary covariances going beyond this instructive example.
\cmt{
\begin{proof}
Set $c^\star=\boldsymbol{\lambda}^T{\boldsymbol{\beta}}^\star$. By definition, each input example $\vct{x}_i$ has the form $\vct{x}_i=g_i\boldsymbol{\lambda}$ and $y_i=g_ic^\star+\sigma z_i$. Set ${\vct{g}}=[g_1~\dots~g_n]^T$ and $\bar{{\vct{g}}}={\vct{g}}/\tn{{\vct{g}}}$. Thus, we have ${\mtx{X}}={\vct{g}}\boldsymbol{\lambda}^T$ and $\vct{y}={\vct{g}}\boldsymbol{\lambda}^T{\boldsymbol{\beta}}^\star+\sigma{\vct{z}}$. Decompose ${\vct{z}}=\bar{{\vct{z}}}+\bar{{\vct{g}}}^T{\vct{z}}\bar{{\vct{g}}}$. The least-squares solution has the form ${\boldsymbol{\hat{\beta}}}=\hat{c}\boldsymbol{\lambda}/\tn{\boldsymbol{\lambda}}^2$ where
\begin{align}
&\hat{c}=\arg\min_{c}\tn{(c^\star-c){\vct{g}}+\sigma{\vct{z}}}\implies\\
&\hat{c}=c^\star+\sigma\gamma\label{beta}.
\end{align}
where $\gamma=\frac{\bar{{\vct{g}}}^T{\vct{z}}}{\tn{{\vct{g}}}}$. Set $h$, $\epsilon\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,1)$. Note that
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}})=\mathbb{E}[((c^\star-\hat{c})h+\sigma\epsilon)^2]
=(\gamma^2+1)\sigma^2
\end{align}
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}})&=\mathbb{E}[(c^\star- \hat{c}\frac{\boldsymbol{\lambda}^T{\vct{u}}}{\tn{\boldsymbol{\lambda}}^2})h+\sigma\epsilon)^2]\\
&=\mathbb{E}[(c^\star- \hat{c}\zeta)h+\sigma\epsilon)^2]\\
&=((1-\zeta)c^\star-\zeta\sigma\gamma)^2
\end{align}
Now let $|\lambda_{k_1}|\geq|\lambda_{k_2}|\geq...\geq|\lambda_{k_p}|$. Assume we solved ERM to get $\hat{c}$ and ${\boldsymbol{\hat{\beta}}}=\hat{c}\boldsymbol{\lambda}$. Prune the trained model ${\boldsymbol{\hat{\beta}}}$ to $s$-sparse by keeping the largest entries. Set ${\vct{u}}=\mathbb{T}_s(\boldsymbol{\lambda})$ and $\zeta=\tn{{\vct{u}}}^2/\tn{\boldsymbol{\lambda}}^2$. We get ${\boldsymbol{\hat{\beta}}}_s=\hat{c}{\vct{u}}$.
\begin{align}
{\cal{L}}(\boldsymbol{\hat{\theta}})&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-{\vct{u}}^T\boldsymbol{\hat{\theta}})h+\sigma\epsilon)^2]\\
&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\sum_{i=1}^s\lambda_{k_i}^2\hat{c})h+\sigma\epsilon)^2]\\
&=[(1-\zeta)\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\zeta\sigma\gamma]^2+\sigma^2
\end{align}
\textbf{Case 1: }
Now, assume that we have already known the optimal entries and do pruning before training. Then the pruned model can be seen as pruning the inputs and then putting it through a non-sparse $s$-feature model. Let $\lambda_{t_1}{\boldsymbol{\beta}}_{t_1}\geq\lambda_{t_2}{\boldsymbol{\beta}}_{t_2}\geq...\geq\lambda_{t_s}{\boldsymbol{\beta}}_{t_s}$, $\vct{w}=[\lambda_{t_1}~\lambda_{t_2}~...~\lambda_{t_s}]$ and $\vct{t}=[{\boldsymbol{\beta}}_{t_1}~...~{\boldsymbol{\beta}}_{t_s}]$. Set data sample $\vct{x}_i'\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,{\boldsymbol{{\Sigma}}}')\in\mathbb{R}^s$ where are ${\boldsymbol{{\Sigma}}}'=\vct{w}\w^T$. Following what done in \ref{beta}, similarly
\[
\boldsymbol{\hat{\theta}}'=\theta\vct{w}\quad\text{where}\quad\theta=\frac{1}{\sum_{i=1}^s\lambda_{t_i}^2}(\vct{w}^T\vct{t}+\frac{\sigma}{\tn{{\vct{g}}}^2}{\vct{g}}^T{\vct{z}})
\]
\begin{align}
{\cal{L}}(\boldsymbol{\hat{\theta}}')&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\sum_{i=1}^s\lambda_{t_i}^2\theta)h+\sigma\epsilon)^2]\\
&=[(\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\vct{w}^T\vct{t})-\sigma\gamma]^2+\sigma^2
\end{align}
\textbf{Case 2: }
Now retrain with the pruned entries $\lambda_{k_!}$, ..., $\lambda_{k_s}$. Then the pruned model can be seen as pruning the inputs and then putting it through a non-sparse $s$-feature model. Let $\vct{t}=[{\boldsymbol{\beta}}_{k_1}~...~{\boldsymbol{\beta}}_{k_s}]$. Set data sample $\vct{x}_i'\overset{\text{i.i.d.}}{\sim} \mathcal{N}(0,{\boldsymbol{{\Sigma}}}')\in\mathbb{R}^s$ where are ${\boldsymbol{{\Sigma}}}'={\vct{u}}\ub^T$. Following what done in \ref{beta}, similarly
\[
\boldsymbol{\hat{\theta}}'=\theta\vct u\quad\text{where}\quad\theta=\frac{1}{\sum_{i=1}^s\lambda_{k_i}^2}(\vct u^T\vct{t}+\frac{\sigma}{\tn{{\vct{g}}}^2}{\vct{g}}^T{\vct{z}})
\]
\begin{align}
{\cal{L}}(\boldsymbol{\hat{\theta}}')&=\mathbb{E}[((\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\sum_{i=1}^s\lambda_{k_i}^2\theta)h+\sigma\epsilon)^2]\\
&=[(\boldsymbol{\lambda}^T{\boldsymbol{\beta}}-\vct u^T\vct{t})-\sigma\gamma]^2+\sigma^2
\end{align}
\textbf{It is obvious to know that loss in case 1 is always smaller than it in case 2. Then for both case, }
Set ${\boldsymbol{\beta}}^T{\boldsymbol{{\Sigma}}}{\boldsymbol{\beta}}=\alpha^2$ and $\vct{t}^T{\boldsymbol{{\Sigma}}}'\vct{t}=\rho^2$. We can write ${\cal{L}}(\boldsymbol{\hat{\theta}})$ and ${\cal{L}}(\boldsymbol{\hat{\theta}}')$ as
\[
{\cal{L}}(\boldsymbol{\hat{\theta}})=((1-\zeta)\alpha-\zeta\sigma\gamma)^2+\sigma^2\iff{\cal{L}}(\boldsymbol{\hat{\theta}}')=(\alpha-\rho-\sigma\gamma)^2+\sigma^2
\]
Now consider a special case which satisfies ${\cal{L}}(\boldsymbol{\hat{\theta}})<{\cal{L}}(\boldsymbol{\hat{\theta}}')$.
\[
0\leq(1-\zeta)\alpha-\zeta\sigma\gamma<\alpha-\rho-\sigma\gamma
\]
\[
\frac{\rho+\sigma\gamma}{\alpha+\sigma\gamma}<\zeta\leq\frac{\alpha}{\alpha+\sigma\gamma}
\]
\end{proof}}
\cmt{
\begin{proof} Let ${\boldsymbol{{\Sigma}}}=\boldsymbol{\lambda}\bla^T$. Each $\vct{x}_i$ has the form $\vct{x}_i=g_i\boldsymbol{\lambda}$ and $y_i=x_i{\boldsymbol{\beta}}^T\boldsymbol{\lambda}+\sigma z_i$. Thus, we have ${\mtx{X}}={\vct{g}}\boldsymbol{\lambda}$ and $\vct{y}={\vct{g}}\boldsymbol{\lambda}^T{\boldsymbol{\beta}}+\sigma{\vct{z}}$. Let ${\vct{z}}=\bar{{\vct{z}}}+\frac{{\vct{g}}^T{\vct{z}}}{\tn{{\vct{g}}}^2}{\vct{g}}$. The least-squares solution has the form ${\boldsymbol{\hat{\beta}}}=\hat{c}\boldsymbol{\lambda}$ where
\[
\hat{c}
\]
\end{proof}}
\section{Ridge Regression Analysis}
We estimate $\vct{t}$ via the ridge-regression
\begin{align}
\boldsymbol{\hat{\theta}}=\mathbb{F}_s({\boldsymbol{\hat{\beta}}})\quad\text{where}\quad {\boldsymbol{\hat{\beta}}}=\arg\min_{{\boldsymbol{\beta}}'} 0.5\tn{\vct{y}-{\mtx{X}}{\boldsymbol{\beta}}'}^2+0.5\lambda n \tn{{\boldsymbol{\beta}}'}^2.\label{optim me3}
\end{align}
The residual formulation is given via
\[
\hat{\vct{w}}=\arg\min_{\vct{w}} 0.5\tn{{\mtx{X}}_{\vct{z}}\vct{w}_\sigma}^2+0.5\lambda n \tn{{\boldsymbol{\beta}}-\vct{w}}^2
\]
Observing $0.5\tn{\vct{v}}^2=\max_{\vct{a}} \vct{a}^T\vct{v}-0.5\tn{\vct{a}}^2$, we find
\[
\hat{\vct{w}}=\arg\min_{\vct{w}_{\sigma}} \max_{\vct{a}} \vct{a}^T{\mtx{X}}_{\vct{z}}\vct{w}_\sigma-0.5\tn{\vct{a}}^2+0.5\lambda n \tn{{\boldsymbol{\beta}}-\vct{w}}^2.
\]
Fix ${\vct{g}}\sim\mathcal{N}(0,{\mtx{I}}_n),\vct{h}\sim\mathcal{N}(0,{\mtx{I}}_p)$. The associated Gordon's form is given by
\[
\hat{\vct{w}}=\arg\min_{\vct{w}_{\sigma}} \max_{\vct{a}}\tn{\vct{w}_\sigma}{\vct{g}}^T\vct{a}- \tn{\vct{a}}\vct{w}_\sigma^T\vct{h}-0.5\tn{\vct{a}}^2+0.5\lambda n \tn{{\boldsymbol{\beta}}-\vct{w}}^2.
\]
Setting $\vct{a}=a\bar{{\vct{g}}}$, this yields
\[
\min_{\vct{w}_{\sigma}} \max_{a\geq 0}a(\tn{\vct{w}_\sigma}\tn{{\vct{g}}}- \vct{w}_\sigma^T\vct{h})-0.5a^2+0.5\lambda n \tn{{\boldsymbol{\beta}}-\vct{w}}^2.
\]
Differentiating and setting $a=\tn{\vct{w}_\sigma}\tn{{\vct{g}}}- \vct{w}_\sigma^T\vct{h}$, we obtain \textcolor{red}{Here: need to be more careful. This holds only if $\tn{\vct{w}_\sigma}\tn{{\vct{g}}}- \vct{w}_\sigma^T\vct{h}>0$}
\[
\min_{\vct{w}_{\sigma}} \lambda n \tn{{\boldsymbol{\beta}}-\vct{w}}^2+(\tn{\vct{w}_\sigma}\tn{{\vct{g}}}- \vct{w}_\sigma^T\vct{h})^2.
\]
Now, set $\vct{w}=h\bar{\vct{h}}+\gamma {\boldsymbol{\beta}}$. We obtain
\[
\tn{\vct{w}_\sigma}\tn{{\vct{g}}}- \vct{w}_\sigma^T\vct{h}\approx \sqrt{n}\sqrt{h^2+\alpha\gamma^2+\sigma^2}-h\sqrt{p}.
\]
Secondly, note that
\[
\tn{{\boldsymbol{\beta}}-\vct{w}}^2=\alpha(1-\gamma)^2+h^2.
\]
Normalizing by $n$ and setting ${\bar{p}}=p/n$, we obtain
\begin{align}\label{eq:min_hgamma_reg}
\min_{h,\gamma} \lambda (\alpha(1-\gamma)^2+h^2)+(\sqrt{h^2+\alpha\gamma^2+\sigma^2}-h\sqrt{{\bar{p}}})^2.
\end{align}
Set $T=\sqrt{h^2+\alpha\gamma^2+\sigma^2}$. Differentiating, we find
\begin{align}
&\lambda \alpha (\gamma-1)+(T-h\sqrt{{\bar{p}}})\frac{\alpha\gamma}{T}=0\implies \gamma=\frac{1}{1+\lambda^{-1}(1-\frac{h\sqrt{{\bar{p}}}}{T})}\label{eq:reg1}\\
&\lambda h+(T-h\sqrt{{\bar{p}}})(\frac{h}{T}-\sqrt{{\bar{p}}})=0.\label{eq:reg2}
\end{align}
From \eqref{eq:reg1}, we find that
\begin{align}
T =\Big( \frac{\gamma\sqrt{{\bar{p}}}}{\lambda(\gamma-1)+\gamma}\Big) h
\end{align}
Substitute this in \eqref{eq:reg2} and use the fact that $T>0\rightarrow h\neq 0$, yields:
\begin{align}
\gamma = \frac{1}{2{\bar{p}}}\,\Big( -1 - \lambda + {\bar{p}} + \sqrt{(1+\lambda-{\bar{p}})^2+4\lambda {\bar{p}}} \,\Big) = \frac{1}{2}\,\Big( 1 - \labar -{\bar{p}}^{-1} + \sqrt{\Big(1 - \labar-{\bar{p}}^{-1} \Big)^2+4\labar} \,\Big)\label{eq:reg_gamma},
\end{align}
where we have also used that optimal $\gamma$ must be non-negative, which is easily seen from \eqref{eq:min_hgamma_reg}. We have also defined,
$$
\labar = \lambda/{\bar{p}}.
$$
Note that as $\lambda\rightarrow0$:
$$
\gamma \stackrel{\lambda\rightarrow0}{\longrightarrow} \frac{1}{2{\bar{p}}}\Big( {\bar{p}} - 1 + \sqrt{(1-{\bar{p}})^2} \,\Big) = \begin{cases}
0 &, \text{ if } {\bar{p}}<1,\\
1-\frac{1}{{\bar{p}}} &, \text{ if } {\bar{p}}>1.
\end{cases}
$$
Next, we can solve \eqref{eq:reg1} to find $h^2$:
\begin{align}\label{eq:h2_reg}
h^2 &= \frac{\alpha\gamma^2+\sigma^2}{\frac{\gamma^2 {\bar{p}}}{(\gamma\lambda+\gamma-\lambda)^2}-1} = \frac{\alpha\gamma^2+\sigma^2}{\frac{{\bar{p}}^{-1}}{\left(\labar+{\bar{p}}^{-1}-\frac{\labar}{\gamma}\right)^2}-1}\\
&= \frac{\frac{\alpha}{4}\left( 1 - \labar -{\bar{p}}^{-1} + \sqrt{\big(1 - \labar-{\bar{p}}^{-1} \Big)^2+4\labar} \,\right)^2 + \sigma^2}{\frac{4{\bar{p}}^{-1}}{\left(1+{\bar{p}}^{-1} + \labar - \sqrt{(1+{\bar{p}}^{-1} + \labar)^2-4{\bar{p}}^{-1}} \right)^2}-1}\notag
\end{align}
where to arrive in the last expression we used \eqref{eq:reg_gamma} and the fact that
$$
\labar+{\bar{p}}^{-1}-\frac{\labar}{\gamma} = \frac{1}{2}\left(1+{\bar{p}}^{-1} + \labar - \sqrt{(1+{\bar{p}}^{-1} + \labar)^2-4{\bar{p}}^{-1}} \right).
$$
With this and using the fact that $a-\sqrt{a^2-b}<\sqrt{b}$ it can be verified that the denominator in \eqref{eq:h2_reg} is positive.
From \eqref{eq:h2_reg} note that as $\lambda\rightarrow 0$
$$
h^2 \stackrel{\lambda\rightarrow0}{\longrightarrow} \begin{cases}
\sigma^2\frac{{\bar{p}}}{1-{\bar{p}}} &, \text{ if } {\bar{p}}<1,\\
\sigma^2\frac{1}{{\bar{p}}-1} \textcolor{red}{+ \alpha {\bar{p}}^{-1}({\bar{p}}^{-1}-1)} &, \text{ if } {\bar{p}}>1
\end{cases}
$$
\textcolor{red}{something is wrong in the ${\bar{p}}>1$ case above}
{\bf{Possibly useful idea:}} Bias-variance decomposition (for Gordon setup).
\noindent {\bf{Step 1:}} Set $\sigma=0$. In this case, show that $h=0$\dots?
\noindent {\bf{Step 2:}} Set ${\boldsymbol{\beta}}=\alpha=0$. In this case, we find $\sqrt{h^2+\sigma^2}$ and
\[
\min_{h} \lambda h^2+(\sqrt{h^2+\sigma^2}-h\sqrt{{\bar{p}}})^2\dots
\]
\newpage
\section{Asymptotic formulas on Magnitude- and Hessian- pruning}\label{sec prune risk}
Here, we use Theorem \ref{thm:master_W2} to characterize the risk of the magnitude- and Hessian- pruned solutions. This section supplements the discussion in Section \ref{sec:risk}. For completeness, first, we recall the magnitude-pruning results of Section \ref{sec:risk} and restate Corollary \ref{cor:mag} below. This corollary characterizes the performance of magnitude pruning. Following this, we shall further discuss Hessian pruning.
\subsection{Magnitude-based pruning}
We begin with the following necessary definitions. Define the hard-thresholding function with fixed threshold $t\in\mathbb{R}_+$ as follows:
\begin{align}\label{eq:threshold_app}
\mathcal{T}_t(x) = \begin{cases} x & \text{if } |x|>t \\ 0 & \text{otherwise} \end{cases}.
\end{align}
Further, {given model sparsity target $1>\alpha>0$}, define the threshold $t^\star$ as follows:
\begin{align}\label{eq:tstar_app}
t^\star:=\sup\left\{t\in\mathbb{R}\,:\, \Pr(|X_{\kappa,\sigma^2}|\geq t) \geq \alpha \right\}.
\end{align}
\cormag*
The proof of the corollary above, is given in Section \ref{sec:risk}. Below, we extend the results to Hessian-based pruning.
\subsection{Hessian-based pruning}
Let ${\boldsymbol{\hat{\beta}}}$ be the min-norm solution in \eqref{eq:min_norm}. Recall that the Hessian-pruned model (via Optimal Brain Damage) $\boldsymbol{\beta}_s^H$ at sparsity $s$ is given by
\begin{align}\label{eq:hp}
\boldsymbol{\beta}_s^H = \hat\boldsymbol{\Sigma}^{-1/2}\mathbb{T}_s(\hat\boldsymbol{\Sigma}^{1/2}\boldsymbol{\beta}),
\end{align}
where $\hat\boldsymbol{\Sigma}=\diag{{\mtx{X}}^T{\mtx{X}}}/n$ the diagonal of the empirical covariance matrix.
We will argue that the following formula characterizes the asymptotic risk of the Hessian pruning solution.
Recall the notation in \eqref{eq:threshold_app} and define
\begin{align}\label{eq:tdiam_app}
t^\diamond:=\sup\left\{t\in\mathbb{R}\,:\, \Pr(|\Lambda^{1/2}X_{\kappa,\sigma^2}|\geq t) \geq \alpha \right\}.
\end{align}
The risk of the Hessian-pruned model satisfies the following in the limit of $n,p,s\rightarrow\infty$ at rates $\kappa:=p/n>1$ and $\alpha:=s/p\in(0,1)$ (cf. Assumption \ref{ass:linear}):
\begin{align}\label{eq:hessian_risk}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^H_s) \stackrel{{P}}{\longrightarrow} \sigma^2 + \operatorname{\mathbb{E}}\left[ \left(\Lambda^{1/2}B-\mathcal{T}_{t^\diamond}(\Lambda^{1/2}X_{\kappa,\sigma^2})\right)^2 \right],
\end{align}
where the expectation is over $(\Lambda,B,H)\sim\mu\otimes\mathcal{N}(0,1)$. In our LGP experiments, we used this formula \eqref{eq:hessian_risk} to accurately predict the Hessian-based pruning performance.
Recall the definition of the hard-thresholding operator $\mathcal{T}_t(x)$. Similar to Section \ref{sec:risk}, we consider a threshold-based pruning vector $${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H}_{t} := \hat\boldsymbol{\Sigma}^{-1/2}\mathcal{T}_{t/\sqrt{p}}(\hat\boldsymbol{\Sigma}^{1/2}\hat{\boldsymbol{\beta}}),$$ where $\mathcal{T}_t$ acts component-wise. Further define
$${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_{t} := \boldsymbol{\Sigma}^{-1/2}\mathcal{T}_{t/\sqrt{p}}(\boldsymbol{\Sigma}^{1/2}\hat{\boldsymbol{\beta}}).$$
Note that ${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_{t}$ uses the true (diagonal) covariance matrix $\boldsymbol{\Sigma}$ instead of its sample estimate $\hat\boldsymbol{\Sigma}$. For later reference, note here that $\hat\boldsymbol{\Sigma}$ concentrates (entry-wise) to $\boldsymbol{\Sigma}$. Using boundedness of $\boldsymbol{\Sigma}$ and standard concentration of sub-exponential random variables.
First, we compute the limiting risk of ${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_{t}$. Then, we will use the fact that $\hat\boldsymbol{\Sigma}$ concentrates (entry-wise) to $\boldsymbol{\Sigma}$, to show that the risks of ${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_{t}$ and ${\boldsymbol{\hat{\beta}}}^{\mathcal{T},H}_{t}$ are arbitrarily close as $p\rightarrow\infty$.
Similar to \eqref{eq:loss_w2},
\begin{align}
{\cal{L}}({\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_t) &= \sigma^2 + ({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_t)^T\boldsymbol{\Sigma}({\boldsymbol{\beta}}^\star-{\boldsymbol{\hat{\beta}}}^{\mathcal{T},H^\star}_t) \notag \\
&= \sigma^2 + \frac{1}{p}\sum_{i=1}^p\boldsymbol{\Sigma}_{i,i}\big(\sqrt{p}\betab^\star_i-\bSi_{i,i}^{-1/2}\mathcal{T}_{t}(\bSi_{i,i}^{1/2}\sqrt{p}\hat\boldsymbol{\beta}_i)\big)^2 \notag
= \sigma^2 + \frac{1}{p}\sum_{i=1}^p\big(\bSi_{i,i}^{1/2}\sqrt{p}\betab^\star_i-\mathcal{T}_{t}(\bSi_{i,i}^{1/2}\sqrt{p}\hat\boldsymbol{\beta}_i)\big)^2
\\
&= \sigma^2 + \frac{1}{p}\sum_{i=1}^p\big(\bSi_{i,i}^{1/2}\sqrt{p}\betab^\star_i-\mathcal{T}_{t}(\widehat{\vct{w}}_i + \bSi_{i,i}^{1/2}\sqrt{p}\betab^\star_i\big)^2 \notag \\
&=: \sigma^2 + \frac{1}{p}\sum_{i=1}^p\big(\sqrt{p}\bla^\st_i-\mathcal{T}_{t}(\widehat{\vct{w}}_i + \bla^\st_i)\big)^2.\label{eq:hesa}
\end{align}
In the second line above, we used that $\sqrt{p}\mathcal{T}_{t/\sqrt{p}}(x)=\mathcal{T}_{t}(\sqrt{p}x)$. In the third line, we recalled the change of variable in \eqref{eq:w}, i.e., $\widehat{\vct{w}}$ is the solution to \eqref{eq:PO}. Finally, in the last line we defined $\bla^\st:=\sqrt{\boldsymbol{\Sigma}}\betab^\star$ (note that this is related to the saliency score $\boldsymbol{\lambda}$ defined in \eqref{saliency eq}).
To evaluate the limit of the empirical average in \eqref{eq:hesa}, we proceed as follows. First, we claim that the empirical distribution of $\sqrt{p}\bla^\st$ converges weakly to the distribution of the random variable $B\sqrt{\Lambda}$, where $(\Lambda,B)\sim\mu$. \so{Note that this convergence is already implied by the proof of Theorem \ref{thm:master_W2} by setting the $g$ function to be zero in \eqref{eq:fdef}.}
For an explicit proof, take any bounded Lipschitz test function $\psi:\mathbb{R}\rightarrow\mathbb{R}$ and call $\psi'(x,y):=\psi({\sqrt{x}}y)$. Then,
\begin{align}
|\psi'(\bSi_{i,i},\sqrt{p}\betab^\star_i)-\psi'(\bSi_{i,i}',\sqrt{p}{\betab^\star_i}')| &= |\psi(\sqrt{\bSi_{i,i}}\sqrt{p}\betab^\star_i)-\psi(\sqrt{\bSi_{i,i}'}\sqrt{p}{\betab^\star_i}')| \leq C |\sqrt{\bSi_{i,i}}\sqrt{p}\betab^\star_i - \sqrt{\bSi_{i,i}'}\sqrt{p}{\betab^\star_i}'| \notag
\\
&\leq
C |\sqrt{p}\betab^\star_i - \sqrt{p}{\betab^\star_i}'| + C'|\sqrt{p}{\betab^\star_i}'| |\sqrt{\bSi_{i,i}}- \sqrt{\bSi_{i,i}'}|\notag \\
&\leq
C (|\sqrt{p}\betab^\star_i - \sqrt{p}{\betab^\star_i}'| + |\sqrt{p}{\betab^\star_i}'| |{\bSi_{i,i}}- {\bSi_{i,i}'}|)\notag \\
&\leq
C (1 + \|[\sqrt{p}\betab^\star_i,\bSi_{i,i}]\|_2 +\|[\sqrt{p}{\betab^\star_i}',\bSi_{i,i}']\|_2 ) \sqrt{|\sqrt{p}\betab^\star_i - \sqrt{p}{\betab^\star_i}'|^2 + |{\bSi_{i,i}}- {\bSi_{i,i}'}|^2}\notag.
\end{align}
Thus, $\psi'$ is $\rm{PL(2)}$. Hence, from Assumption \ref{ass:mu},
\begin{align}
\frac{1}{p}\sum_{i=1}^p \psi(\sqrt{p}\bla^\st_i) = \frac{1}{p}\sum_{i=1}^p \psi'(\bSi_{i,i},\sqrt{p}\betab^\star_i) \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}[\psi'(B,\sqrt{\Lambda})] = \operatorname{\mathbb{E}}[\psi(B\sqrt{\Lambda})]\label{eq:convvvv}
\end{align}
Besides, from Theorem \ref{thm:master_W2} applied for $f_{\cal{L}}(x,y,z)=zy^2$ (i.e., set $g$ the zero function in \eqref{eq:fdef}), we have that
$$
\frac{1}{p}\sum_{i=1}^p \sqrt{p}(\bla^\st_i)^2 \stackrel{{P}}{\longrightarrow} \operatorname{\mathbb{E}}[B^2\Lambda].
$$
Therefore, convergence in \eqref{eq:convvvv} actually holds for any $\psi\in\rm{PL}(2)$. Next, observe that, $\boldsymbol{\omega}$ of \eqref{eq:w_n} can be written in terms of $\bla^\st$ via \begin{align}\label{eq:w_n2}
\vct{w}_n=\vct{w}'(\tau_n,u_n)& = -\left({\mtx{I}}+\frac{u_n}{\tau_n}\boldsymbol{\Sigma}\right)^{-1}\left(\boldsymbol{\Sigma}^{1/2}\betab^\star-u_n\sqrt{\kappa}\boldsymbol{\Sigma}\bar{\h}\right)\\
&=-\left({\mtx{I}}+\frac{u_n}{\tau_n}\boldsymbol{\Sigma}\right)^{-1}\left(\bla^\st-u_n\sqrt{\kappa}\boldsymbol{\Sigma}\bar{\h}\right).
\end{align}
After this observation, the convergence proof can be finalized by using a modified version of Assumption \ref{ass:mu} as follows
{
\begin{assumption}[Empirical distribution for saliency] \label{ass:lambda} Set $\bla^\st_i=\sqrt{\bSi_{i,i}}\boldsymbol{\beta}^\star_i$. The joint empirical distribution of $\{(\bSi_{i,i},\bla^\st_i)\}_{i\in[p]}$ converges in {Wasserstein-k} distance to a probability distribution $\mu=\mu(\Lambda,S)$ on $\mathbb{R}_{>0}\times\mathbb{R}$ for some \fx{$k\geq 4$}. That is
$
\frac{1}{p}\sum_{i\in[p]}\delta_{(\bSi_{i,i},\sqrt{p}\bla^\st_i)} \stackrel{W_k}{\Longrightarrow} \mu.
$
\end{assumption}
}
Under this assumption, it can be verified that, the exact same proof strategy we used for magnitude-based pruning would apply to Hessian-based pruning by replacing $\boldsymbol{\beta}^\star$ with $\bla^\st$. The reason is that the Hessian pruning risk is a $\text{PL}(4)$ function of $\vct{h},\bla^\st,{\boldsymbol{{\Sigma}}}$ (e.g.~can be shown in a similar fashion to Lemma \ref{lem:hPL}). Observe that Assumption \ref{ass:lambda} is a reasonable assumption given the naturalness of the saliency score. If we only wish to use the earlier Assumption \ref{ass:mu} rather than Assumption \ref{ass:lambda}, one can obtain the equivalent result by modifying \ref{ass:mu} to enforce a slightly higher order bounded moment and convergence condition. Finally, one needs to address the perturbation due to the finite sample estimation of the covariance. Note that, even if the empirical covariance doesn't converge to the population, its diagonal weakly converges to the population (as we assumed the population is diagonal). The (asymptotically vanishing) deviation due to the finite sample affects can be addressed in an identical fashion to the deviation analysis of $\tau_n,u_n$ at \eqref{eq:ivstep1} and \eqref{eq:betav}. While these arguments are reasonably straightforward and our Hessian pruning formula accurately predicts the empirical performance, the fully formal proof of the Hessian-based pruning is rather lengthy to write and does not provide additional insights.
\cmt{
if one assumes Assumption \ref{ass:mu} with the
The overall function $f(\bla^\st_i,$ becomes
\[
h(x,y,z)=f(g(y,z,x),y,z)=y(x-g(z))^2
\]}
\section{Problem Setup}\label{sec:setup}
Let us fix the notation. Let $[p]=\{1,2,\dots,p\}$. Given ${\boldsymbol{\beta}}\in \mathbb{R}^p$, let $\mathbb{T}_s({\boldsymbol{\beta}})$ be the pruning operator that sets the smallest $p-s$ entries in absolute value of ${\boldsymbol{\beta}}$ to zero. Let ${\mathcal{I}}({\boldsymbol{\beta}})\subset[p]$ return the index of the nonzero entries of ${\boldsymbol{\beta}}$. ${\mtx{I}}_n$ denotes the $n\times n$ identity matrix and $\mathcal{N}({\boldsymbol{{\mu}}},{\boldsymbol{{\Sigma}}})$ denotes the normal distribution with mean ${\boldsymbol{{\mu}}}$ and covariance ${\boldsymbol{{\Sigma}}}$. ${\mtx{X}}^\dagger$ denotes the pseudoinverse of matrix ${\mtx{X}}$.
\vspace{3pt}
\noindent \textbf{Data:} Let $(\vct{a}_i,y_i)_{i=1}^n\subset \mathbb{R}^d\times \mathbb{R}$ with i.i.d.~input-label pairs. Let $\phi(\cdot):\mathbb{R}^d\rightarrow\mathbb{R}^p$ be a (nonlinear) feature map. We generate $\vct{x}_i=\phi(\vct{a}_i)$ and work with the dataset $\mathcal{S}=(\vct{x}_i,y_i)_{i=1}^n$ coming i.i.d. from some distribution $\mathcal{D}$. As an example, of special interest to the rest of the paper, consider random feature regression, where $\vct{x}_i=\psi({\mtx{R}}\vct{a}_i)$ for a nonlinear activation function $\psi$ that acts entry-wise and a random matrix ${\mtx{R}}\in \mathbb{R}^{p\times d}$ with i.i.d.~$\mathcal{N}(0,1)$ entries; see Fig.~\ref{figRF}. In matrix notation, we let $\vct{y}=[y_1~\dots~y_n]^T\in\mathbb{R}^n$ and ${\mtx{X}}=[\vct{x}_1~\dots~\vct{x}_n]^T\in\mathbb{R}^{n\times p}$ denote the vector of labels and the feature matrix, respectively.
Throughout, we focus on regression tasks, in which the training and the test risks of a model ${\boldsymbol{\beta}}$ are defined a
\begin{align}
&\text{Population risk:}~~~{\cal{L}}({\boldsymbol{\beta}})=\operatorname{\mathbb{E}}_{\mathcal{D}}[(y-\vct{x}^T{\boldsymbol{\beta}})^2].\label{test formula}\\
&\text{Empirical risk:}~~~~{\hat{\cal{L}}}({\boldsymbol{\beta}})=\frac{1}{n}\tn{\vct{y}-{\mtx{X}}{\boldsymbol{\beta}}}^2.\label{erm}
\end{align}
\cmt{Let $\phi(\cdot):\mathbb{R}^d\rightarrow \mathbb{R}^k$ be a nonlinear map for feature extraction. $\phi$ will be important for random feature analysis.}
During training, we will solve the empirical risk minimization (ERM) problem over a set of selected features $\Delta\subset[p]$, from which we obtain the least-squares solution
\begin{align}
\label{eq:ERM}
{\boldsymbol{\hat{\beta}}}(\Delta)=\arg\min_{{\boldsymbol{\beta}}\,:\,{\mathcal{I}}({\boldsymbol{\beta}})=\Delta} {\hat{\cal{L}}}({\boldsymbol{\beta}}).
\end{align}
For example, regular ERM corresponds to $\Delta=[p]$, and we simply use ${\boldsymbol{\hat{\beta}}}={\boldsymbol{\hat{\beta}}}([p])$ to denote its solution above. Let ${\boldsymbol{{\Sigma}}}=\operatorname{\mathbb{E}}[\vct{x}\x^T]$ be the covariance matrix and $\vct{b}=\operatorname{\mathbb{E}}[y\vct{x}]$ be the cross-covariance. The parameter minimizing the test error is given by ${\boldsymbol{\beta}}^\star={\boldsymbol{{\Sigma}}}^{-1}\vct{b}$. We are interested in training a model over the training set $\mathcal{S}$ that not only achieves small test error, but also, it is sparse. We do this as follows. First, we run stochastic gradient descent (SGD) to minimize the empirical risk (starting from zero initialization). It is common knowledge that SGD on least-squares converges to the minimum $\ell_2$ norm solution given by ${\boldsymbol{\hat{\beta}}}={\mtx{X}}^\dagger \vct{y}$. Next, we describe our pruning strategies to compress the model.
\vspace{3pt}
\noindent \textbf{Pruning strategies:} Given dataset $\mathcal{S}$ and target sparsity level $s$, a pruning function $P$ takes a model ${\boldsymbol{\beta}}$ as input and outputs an $s$-sparse model ${\boldsymbol{\beta}}^P_s$. Two popular pruning functions are magnitude-based (MP) and Hessian-based (HP) (a.k.a.~optimal brain damage) pruning \cite{lecun1990optimal}. The latter uses a diagonal approximation of the covariance via $\hat{{\boldsymbol{{\Sigma}}}}=\text{diag}({\mtx{X}}^T{\mtx{X}})/n$ to capture \emph{saliency} (see \eqref{saliency eq}). Formally, we have the following definitions:
\begin{itemize}
\item {\em{Magnitude-based pruning:}} ${\boldsymbol{\beta}}^M_s=\mathbb{T}_s({\boldsymbol{\beta}})$.
\item {\em{Hessian-based pruning:}} ${\boldsymbol{\beta}}^H_s=\hat{{\boldsymbol{{\Sigma}}}}^{-1/2}\mathbb{T}_s(\hat{{\boldsymbol{{\Sigma}}}}^{1/2}{\boldsymbol{\beta}})$.
\item {\em{Oracle pruning:}} Let $\Delta^\star\subset[p]$ be the optimal $s$ indices so that ${\boldsymbol{\hat{\beta}}}(\Delta^\star)$ achieves the minimum population risk (in expectation over $\mathcal{S}$) among all ${\boldsymbol{\hat{\beta}}}(\Delta)$ and any subset $\Delta$ in \eqref{eq:ERM}. When ${\boldsymbol{{\Sigma}}}$ is diagonal and $s<n$, using rather classical results, it can be shown that (see Lemma 7 in the Supplementary Material (SM)) these {\em{oracle indices}} are the ones with the top-$s$ saliency score given by
\begin{align}
\text{Saliency score}={\boldsymbol{{\Sigma}}}_{i,i}{{\boldsymbol{\beta}}^\star_i}^2.\label{saliency eq}
\end{align} Oracle pruning {employs these latent saliency scores and} returns ${\boldsymbol{\beta}}^O_s$ by pruning the weights of ${\boldsymbol{\beta}}$ outside of $\Delta^\star$.
\end{itemize}
We remark that our distributional characterization might allow us to study more complex pruning strategies, such as optimal brain surgeon \cite{hassibi1994optimal}. However, we restrict our attention to the three aforementioned core strategies to keep the discussion focused.
\noindent\textbf{Pruning algorithm:} To shed light on contemporary pruning practices, we will study the following three-stage {\em{train$\rightarrow$prune$\rightarrow$retrain}} algorithms.
\begin{enumerate}
\item Find the empirical risk minimizer ${\boldsymbol{\hat{\beta}}}={\mtx{X}}^\dagger \vct{y}$.
\item Prune ${\boldsymbol{\hat{\beta}}}$ with strategy $P$ to obtain ${\boldsymbol{\hat{\beta}}}^P_s$.
\item {\em{Retraining:}} Obtain ${\boldsymbol{\hat{\beta}}}^{RT}_s={\boldsymbol{\hat{\beta}}}({\mathcal{I}}({\boldsymbol{\hat{\beta}}}^P_s))$.
\end{enumerate}
The last step obtains a new $s$-sparse model by solving ERM in \eqref{eq:ERM} with the features $\Delta={\mathcal{I}}({\boldsymbol{\hat{\beta}}}^P_s)$ identified by pruning. Figures \ref{figNN} and \ref{figRF} illustrate the performance of this procedure for ResNet-20~on the CIFAR-10 dataset and for a random feature regression on a synthetic problem, respectively . Our analytic formulas for RF, as seen in Fig.~\ref{figRF}, very closely match the empirical observations (see Sec.~\ref{sec mot}~for further explanations). Interestingly, the arguably simpler RF model already captures key behaviors (double-descent, better performance in the overparameterized regime, performance of sparse model comparable to large model) in ResNet.
\cmt{I think it is nice to add a sentence here pointing out the similarities between Fig. 1 and 2. In particular noting that the arguably simpler RF model already captures key behavior (double-descent, better performance in deep overparameterized regime, performance of sparse model comparable to large model) in ResNet. And then say: we have analytic formulas for the latter which as seen in Fig. 1 very closely match the empirical observations.}
\cmt{We remark that this does not necessarily hold in general (see~supplementary material (SM)).}
\cmt{The next section provides our results on distributional characterization of ${\boldsymbol{\hat{\beta}}}$ and provable guarantees on ${\boldsymbol{\hat{\beta}}}^P_s$. We will then explore when and how retraining stage is critical for the pruning performance.}
Sections \ref{sec mot} and \ref{sec intu} present numerical experiments on pruning that verify our analytical predictions, as well as, our insights on the fundamental principles behind the roles of overparameterization and retraining. Sec \ref{sec main} establishes our theory on the DC of ${\boldsymbol{\hat{\beta}}}$ and provable guarantees on pruning. All proofs are deferred to the Supplementary Material (SM)
\section{Further Experiments and Intuitions}\label{SM mot}
\subsection{Further discussion and experiments on CIFAR-10} \label{ssec:SMCF}
First, we provide further discussion on Figure \ref{figNN}. Recall that, in this figure, we apply \emph{train$\rightarrow$prune$\rightarrow$retrain} to obtain the sparse neural networks. The test error and the training errors for the sparse models are evaluated at the end of the retraining phase. Thus, it is rather surprising that sparse models manage to achieve zero training error around the same width parameter $k$ as the dense model because while parameter count of the dense model increases in $k$, it is fixed for sparse models.
Secondly, we complement Figure \ref{figNN} with two additional experiments. The first experiment assesses the benefit of pruning compared to using a random nonzero pattern with the same sparsity. The second experiment assesses the benefit of retraining by comparing the curves in Fig \ref{figNN} with the test errors obtained without retraining. These two experiments are shown in Figure \ref{figCFa} and \ref{figCFb} and all of them are trained over same dataset and configured as given in Section \ref{ssec:NNE}. Instead of applying magnitude-based pruning, dotted red and green lines in Fig \ref{figCFa} are sparse models over $5$ or $10$ sparsity targets, pruned randomly to achieve same number of nonzeros as magnitude-based pruning strategy. Although the double descent phenomenon and downward trend are still present on the dotted lines, the performance is worse than magnitude-base pruning method. Fig. \ref{figCFb} shows how retraining benefits pruning ResNet-20 models. The results agree with our intuition that the non-retrained (dotted) lines achieve much bigger test error than retrained (solid) lines and overparameterization does not help in improving performance.
\subsection{MNIST Experiments with two layers}
In Figure \ref{figMNIST} we train the simplest neural model with only 2 fully-connected layers over MNIST with various number of nodes to explore properties of magnitude-based pruning, random pruning and non-retraining on simple neural networks. Here, the number of nodes is equivalent to the width of the model, which directly controls the model size. Same as in Section \ref{ssec:NNE}, we select an $s$-width model as base model and prune trained dense models to the same sparsity. All experiments are trained with Adam optimization, $0.001$ learning rate and $200$ epochs under MNIST. Solid red, green and blue lines in Figure \ref{figMNIST} correspond to test error of $4$-, $8$-sparsity targets and dense models. In Figure \ref{figMNISTa} dotted lines are training errors of dense and $s$-sparse models. As the width $k$ grows, the training and test error decrease for all dense and sparse models. The behavior is similar to Figure \ref{figNN} and training larger models benefit pruned-model accuracy however double descent is not really visible. We suspect that this may be because of the simpler nature of the MNIST dataset and LeNet architecture compared to the CIFAR10 dataset and ResNet-20 model. In Figure \ref{figMNISTb}, dotted lines apply randomly pruning. Different to magnitude pruning, where training bigger models and then pruning results in better performance, randomly pruning hurts when the sparsity level $s/k$ is relatively low. This is because under magnitude-based pruning, we can identity most of optimal entries of weights from trained dense model and retraining with these entries achieves lower errors. In constrast, random pruning learns nothing from the trained model and as the sparsity level decreases, the probability that random operator selects the limited optimal entries by chance also decreases, leading to worse performance. Dotted lines in Figure \ref{figMNISTc} show test errors of sparse models before retraining which educes the same conclusion in Section \ref{ssec:SMCF}, that is retraining is crucial to improve the performance in neural networks.
\subsection{Experiments on LGP}
In Figure \ref{figSM1}, we carry out the identical experiments as in Figure \ref{fig1}. The difference is that we display two more figures which are the retraining curves for Magnitude- and Hessian-based pruning strategies shown in purple and yellow lines respectively. Figure \ref{figSM1a} is the counterpart of Figure \ref{fig1a} and Figure \ref{figSM1b} is the counterpart of Figure \ref{fig1b}. The main message in these experiments is that \emph{retraining hurts the performance}. This performance degradation is more emphasized in the overparameterized regime. Specifically, both retrained versions of Magnitude and Hessian pruning ${\boldsymbol{\hat{\beta}}}^{RT,M}$ and ${\boldsymbol{\hat{\beta}}}^{RT,H}$ perform consistently worse compared to their pruning-only counterparts ${\boldsymbol{\hat{\beta}}}^{M}$ and ${\boldsymbol{\hat{\beta}}}^{H}$. Observe that, the only regime where retraining outperforms the pruning-only approach is at the peak of double descent. This is the region where pruning-only risk diverges to infinity whereas retraining attains finite risk. This is because the retraining stage solves a well-conditioned problem and avoids the ill-conditioning occurring at $n=k$. Recall that, in light of Lemma \ref{lem rank one}, unlike the rank-one covariance case, retraining hurts because covariance is diagonal; thus, features are uncorrelated and do not have overlapping predictions.
\begin{figure}[t!]
\centering
\begin{subfigure}{3.5in}\vspace{-5pt}
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[scale=0.33]{figs/FLAT_SIM_RT_s10_n30_p100_SCL25}};
\node at (-3.5,0) [rotate=90,scale=.9]{Test Risk};
\node at (0,-2.8) [scale=.9]{Problem Size ($k/p$)}
\end{tikzpicture}\caption{\small{Identity covariance, spiked latent weights.}}\label{figSM1a}\vspace{-0.2cm}
\end{subfigure}~~~~~\begin{subfigure}{3.5in}\vspace{-5pt}
\centering
\begin{tikzpicture}
\node at (0,0) {\includegraphics[scale=0.33]{figs/SPIKE_SIM_RT_s10_n30_p100_SCL25}};
\node at (-3.5,0) [rotate=90,scale=.9]{Test Risk};
\node at (0,-2.8) [scale=.9]{Problem Size ($k/p$)}
\end{tikzpicture}\caption{\small{Spiked covariance, identical latent weights.}}\label{figSM1b}\vspace{-0.2cm}
\end{subfigure}\caption{\small{Same as Figure \ref{fig1} however retraining curves are included. Our asymptotic prediction for various pruning strategies in linear gaussian models with $s/p=0.1$ and $n/p=0.3$. We solve ERM using the first $k$ features and then prune to obtain an $s$-sparse model. The vertical dashed line shows the $k=s$ point. The horizontal dashed line highlights the minimum risk among all underparameterized solutions including retraining. Retraining curves are not displayed.}}\label{figSM1}\vspace{-15pt}
\end{figure}
\section*{Acknowledgments}
S. Oymak is partially supported by the NSF award CNS-1932254. C. Thrampoulidis is partially supported by the NSF under Grant Numbers CCF-2009030.
\section*{Potential Ethical Impacts}
While deep learning is transformative in wide swath of applications, it comes at a cost: State-of-the-art deep learning models tend to be very large and consume significant energy during inference. The race for larger and better models and growing list of applications exacerbates this carbon footprint problem. Thus there is an urgent need for better and more principled model compression methods to help build environmentally friendly ML models. This work responds to this need by establishing the fundamental algorithmic principles and guarantees behind the contemporary model compression algorithms and by shedding light on the design of lightweight energy- and compute-efficient neural networks. We do not see an ethical concern associated with this work.
\cmt{\subsection{Random Features}}
\section{Main Results}\label{doubdec}
\input{overdetermined}
\input{overparam}
\subsection{Distributional Control via CGMT}
\begin{lemma} [AO solution to PO solution] \label{dist cont lem}Let ${\mtx{X}}\in\mathbb{R}^{n\times p},{\vct{g}}\in\mathbb{R}^n,\vct{h}\in\mathbb{R}^p\overset{\text{i.i.d.}}{\sim}\mathcal{N}(0,1)$. Suppose we have two loss functions ${\cal{L}}_{PO}(\vct{w};{\mtx{X}})$ and ${\cal{L}}_{AO}(\vct{w};{\vct{g}},\vct{h})$ as a function of $\vct{w}$\footnote{${\cal{L}}(\vct{w},\vct{a})$ can account for additional set constraints of type $\vct{w}\in\mathcal{C}$ by adding the indicator penalty $\max_{\lambda\geq 0}\lambda 1_{\vct{w}\not\in\mathcal{C}}$.}. Given a set $\mathcal{S}$, define the objectives
\[
\Phi_{\mathcal{S}}({\mtx{X}})=\min_{\vct{w}\in\mathcal{S}}{\cal{L}}_{PO}(\vct{w},\vct{a})\quad\text{and}\quad\phi_{\mathcal{S}}({\vct{g}},\vct{h})=\min_{\vct{w}\in\mathcal{S}}{\cal{L}}_{AO}(\vct{w},\vct{b}).
\]
Suppose $\Phi$ and $\phi$ satisfies the following conditions for any closed set $\mathcal{S}$ and $t$
\begin{itemize}
\item $\mathbb{P}(\Phi_{\mathcal{S}}({\mtx{X}}) < t)\leq2\mathbb{P}(\phi_{\mathcal{S}}({\vct{g}},\vct{h})\leq t)$.
\item Furthermore, if $\mathcal{S}$ is convex, $\mathbb{P}(\Phi_{\mathcal{S}}({\mtx{X}}) > t)\leq2\mathbb{P}(\phi_{\mathcal{S}}({\vct{g}},\vct{h})\geq t)$
\end{itemize}
Define the set of global minima $\mathcal{M}=\{\vct{w}{~\big |~} {\cal{L}}(\vct{w};{\mtx{X}})=\Phi({\mtx{X}})\}$. For any closed set $\mathcal{S}$, we have that
\begin{align}
\mathbb{P}(\mathcal{M}\in\mathcal{S}^c)\geq 1-2\min_{t}(\mathbb{P}(\phi_{\mathbb{R}^p}({\vct{g}},\vct{h})\geq t)+\mathbb{P}(\phi_{\mathcal{S}}({\vct{g}},\vct{h})\leq t)).\label{min inside}
\end{align}
\end{lemma}
\begin{proof} Let $\vct{w}^*\in\mathcal{M}$. Suppose the events $\Phi_{\mathbb{R}^p}({\vct{g}},\vct{h})\leq t$ and $\Phi_{\mathcal{S}}({\vct{g}},\vct{h})> t$ hold. These two imply that $\vct{w}^*\not\in\mathcal{S}$ hence $\mathcal{M}\subseteq\mathcal{S}^c$. To proceed, for any choice of $t$
\begin{align}
\mathbb{P}(\mathcal{M}\in\mathcal{S}^c)&\geq \mathbb{P}(\{\Phi_{\mathbb{R}^p}({\vct{g}},\vct{h})\leq t\}\cap \{\Phi_{\mathcal{S}}({\vct{g}},\vct{h})> t\})\\
&\geq 1-\mathbb{P}(\Phi_{\mathbb{R}^p}({\vct{g}},\vct{h})> t)-\mathbb{P}(\Phi_{\mathcal{S}}({\vct{g}},\vct{h})\leq t)\\
&\geq 1-\mathbb{P}(\Phi_{\mathbb{R}^p}({\vct{g}},\vct{h})> t)-\lim_{t'\rightarrow t^+}\mathbb{P}(\Phi_{\mathcal{S}}({\vct{g}},\vct{h})<t')\\
&\geq 1-2\mathbb{P}(\phi_{\mathbb{R}^p}({\vct{g}},\vct{h})\geq t)-2\lim_{t'\rightarrow t^+}\mathbb{P}(\phi_{\mathcal{S}}({\vct{g}},\vct{h})\leq t').
\end{align}
Since this holds for all $t$ and cumulative distribution function is continuous, we get the advertised bound \eqref{min inside}.
\end{proof}
\newpage
\section{Useful results about pseudo-Lipschitz functions and CGMT}\label{SM useful fact}
For $k\geq 1$ we say a function $f:\mathbb{R}^m\rightarrow\mathbb{R}$ is pseudo-Lipschitz of order $k$ and denote it by $f\in \rm{PL}(k)$ if there exists a constant $L>0$ such that, for all $\vct{x},\vct{y}\in\mathbb{R}^m$:
\begin{align}
|f(\vct{x})-f(\vct{y})|\leq L\left(1+\|\vct{x}\|_2^{k-1}+\|\vct{y}\|_2^{k-1}\right)\|\vct{x}-\vct{y}\|_2.\label{PL func}
\end{align}
In particular, when $f\in\rm{PL}(k)$, the following properties hold; see \cite{bayati2011dynamics}:
\begin{enumerate}
\item There exists a constant $L'$ such that for all $\vct{x}\in\mathbb{R}^n$: $|f(\vct{x})|\leq L'(1+\|\vct{x}\|_2^k).$
\item $f$ is locally Lipschitz, that is for any $M>0$, there exists a constant $L_{M,m}<\infty$ such that for all $x,y\in[-M,M]^m$,
$
|f(\vct{x})-f(\vct{y})| \leq L_{M,m}\|\vct{x}-\vct{y}\|_2.
$
Further, $L_{M,m}\leq c(1+(M\sqrt{m})^{k-1})$ for some costant $c$.
\end{enumerate}
Using the above properties, we prove the following two technical lemmas used in the proof of Theorem \ref{thm:master_W2}
\begin{lemma}\label{lem:fL}
Let $g:\mathbb{R}\rightarrow\mathbb{R}$ be a Lipschitz function. Consider the function $f:\mathbb{R}^3\rightarrow\mathbb{R}$ defined as follows:
$$
f(\vct{x}) = x_3(x_2-g(x_1))^2.
$$
Then, $f\in\rm{PL}(3)$. If additionally, $f:\mathbb{R}^2\times {\cal{Z}}\rightarrow\mathbb{R}$ for a bounded set ${\cal{Z}}\subset\mathbb{R}$, then $f\in{\cal{F}}\subset\rm{PL}(3)$. Specifically, setting ${\cal{Z}}=[\Sigma_{\min},\Sigma_{\max}]$ (as per Assumption \ref{ass:mu}), we find that ${\cal{F}}_{\cal{L}}\subset{\cal{F}}$, where ${\cal{F}}_{\cal{L}}$ is defined in \eqref{eq:fdef}.
\end{lemma}
\begin{proof}
We first prove that $f\in\rm{PL}(3)$.
Let $h:\mathbb{R}^2\rightarrow\mathbb{R}$ defined as $h({\vct{u}})=({\vct{u}}_2-g({\vct{u}}_1))^2$. The function $({\vct{u}}_1,{\vct{u}}_2)\mapsto{\vct{u}}_2-g({\vct{u}}_1)$ is clearly Lipschitz. Thus, $h\in\rm{PL}(2)$, i.e.,
\begin{align}
|h({\vct{u}})-h(\vct{v})| \leq C(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2\quad\text{and}\quad |h(\vct{v})|\leq C'(1+\|\vct{v}\|_2^2).\label{eq:h_pl}
\end{align}
Therefore, letting $\vct{x}=({\vct{u}},x_3)\in\mathbb{R}^3$ and $\vct{y}=(\vct{v},y_3)\in\mathbb{R}^3$, we have that
\begin{align}
|f(\vct{x})-f(\vct{y})| &= |x_3h({\vct{u}}) - y_3h(\vct{v})| \leq |x_3||h({\vct{u}})-h(\vct{v})| + |h(\vct{v})| |x_3-y_3|\notag\\
&\leq C|x_3|(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_3-y_3| \notag\\
&\leq C(|x_3|^2+(1+\|{\vct{u}}\|_2+\|\vct{v}\|_2)^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{v}\|_2^2)|x_3-y_3| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|{\vct{u}}-\vct{v}\|_2 + C'(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)|x_3-y_3| \notag\\
&\leq C(1+\|\vct{x}\|_2^2+\|\vct{y}\|_2^2)\|\vct{x}-\vct{y}\|_2.
\end{align}
In the second line, we used \eqref{eq:h_pl}. In the third line, we used $2xy\leq x^2+y^2$. In the fourth line, we used Cauchy-Schwarz inequality. $C,C'>0$ are absolute constants that may change from line to line.
\fy{Now, we shall prove that $f\in{\cal{F}}$. To accomplish this, we simply need to show that for all $z\in {\cal{Z}}$, $f(\cdot,\cdot,z)$ is $\rm{PL}(2)$ (for some uniform PL constant). This can be shown as follows. Let $C=\sup_{z\in{\cal{Z}}}|z|$. Then
\begin{align}
|f(x,y,z) - f(x',y',z)| &=|z(y-g(x))^2-z(y'-g(x'))^2|\\
&\leq C|(y-g(x))^2-(y'-g(x'))^2|\\
&\leq C(1+\tn{{\vct{u}}}+\tn{\vct{v}})\tn{{\vct{u}}-\vct{v}},
\end{align}
where ${\vct{u}}=[x,y]$ and $\vct{v}=[x',y']$. In the second line above, we used boundedness of ${\cal{Z}}$. In the third line, we used the fact that the function $\psi(x,y) = (y-g(x))^2$ is $\rm{PL}(2)$ as it is a quadratic of a Lipschitz function.
}
This completes the proof of the lemma.
\end{proof}
{
\begin{lemma}[PL with Bounded Variables]\label{lem:PLbdd
Let $f:\mathbb{R}^{d_1}\rightarrow \mathbb{R}$ be a $PL(k)$ function, $\mathcal{M}\subset \mathbb{R}^{d_2}$ be a compact set and $g$ be a continuously differentiable function over $\mathcal{M}$. Then $h(\vct{x},\vct{y})=f(\vct{x})g(\vct{y})$ is $PL(k+1)$ over $\mathbb{R}^{d_1}\times \mathcal{M}$.
\end{lemma}
\begin{proof} First observe that $g$ has continuous derivatives and is continuous over a compact set. Thus $g$ and its gradient are bounded and $g$ is Lipschitz over $\mathcal{M}$. Let $B=\sup_{\vct{x}\in \mathcal{M}}\max |g(x)|,\tn{\nabla g(\vct{x})}$. To proceed, given pairs $(\vct{x},\vct{y})$ and $(\vct{x}',\vct{y}')$ over $\mathbb{R}\times \mathcal{M}$, we have that
\begin{align}
|h(\vct{x},\vct{y})-h(\vct{x}',y')|&\leq |h(\vct{x},\vct{y})-h(\vct{x}',\vct{y})|+ |h(\vct{x}',\vct{y})-h(\vct{x}',\vct{y}')|\\
&\leq |f(\vct{x})-f(\vct{x}')||g(\vct{y})|+ |f(\vct{x}')||g(\vct{y})-g(\vct{y}')|\\
&\leq B|f(\vct{x})-f(\vct{x}')|+ B\tn{\vct{y}-\vct{y}'}|f(\vct{x}')|\\
&\leq B(1+\tn{\vct{x}'}^{k-1}+\tn{\vct{x}}^{k-1})\tn{\vct{x}-\vct{x}'}+ B\tn{\vct{y}-\vct{y}'}(1+\tn{\vct{x}'}^k)\\
&\leq C (1+\tn{{\vct{z}}}^k+\tn{{\vct{z}}'}^{k})\tn{{\vct{z}}-{\vct{z}}'},
\end{align}
where ${\vct{z}}=[\vct{x}~\vct{y}]^T$ and $C$ an absolute constant. This shows the desired $\text{PL}(k+1)$ guarantee.
\end{proof}
The following lemma is in similar spirit to Lemma \ref{lem:PLbdd} and essentially follows from similar lines of arguments (i.e.~using Lipschitzness induced by boundedness).
}
\begin{lemma}\label{lem:hPL}
Let functions $f,g:\mathbb{R}^3\rightarrow\mathbb{R}$ such that $f\in\rm{PL}(k)$ and
$$
g(x,y,z) := \frac{y^{-1/2}}{1+(\xi_* y)^{-1}}\sqrt{\kappa}\tau_* z + (1-(1+\xi_*y)^{-1})x.
$$
Here, $\xi_*,\tau_*,\kappa$ are positive constants.
Further define
\begin{align}
h(x,y,z) := f\left(g(y,z,x),y,z\right),
\end{align}
and assume that $y$ take values on a fixed bounded compact set $\mathcal{M}{\subset\mathbb{R}^+}$.
Then, it holds that $h\in\rm{PL}(k+1)$.
\end{lemma}
\begin{proof}
Since $f$ is $\rm{PL}(k)$, for some $L>0$, \eqref{PL func} holds. Fix $x,x'\in\mathbb{R}$, $\vct{a}=[y,z]\in\mathbb{R}^{2}$ and $\vct{a}'=[y',z']\in\mathbb{R}^{2}$. Let $\vct{b}=[{\vct{g}},\vct{a}]=[{\vct{g}},y,z]\in\mathbb{R}^3$ where ${\vct{g}}=g(y,z,x)\in\mathbb{R}$ and define accordingly $\vct{b}'=[{\vct{g}}',\vct{a}']$ and ${\vct{g}}'=g(y',z',x')$. We have that
\begin{align}
|h([x,\vct{a}])-h([x',\vct{a}'])|&= |f(\vct{b})-f(\vct{b}')|\notag\\
&\leq L\left(1+\|\vct{b}\|_2^{k-1}+\|\vct{b}'\|_2^{k-1}\right)\|\vct{b}-\vct{b}'\|_2\notag\\
&\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|{\vct{g}}|^{k-1}+|{\vct{g}}'|^{k-1}\right)(\|\vct{a}-\vct{a}'\|_2+|{\vct{g}}-{\vct{g}}'|),\label{eq:hlip}
\end{align}
for some constant $C>0$. In the last inequality we have repeatedly used the inequality
$
\left(\sum_{i=1}^m \|\vct{v}_i\|_2^2\right)^{\frac{d}{2}} \leq C(m)\cdot\sum_{i=1}^m\|\vct{v}_i\|_2^{d}.
$
Next, we need to bound the ${\vct{g}}$ term in terms of $(x,\vct{a})$. This is accomplished as follows
\begin{align}
|{\vct{g}}|^{k-1} &= \left|\frac{y^{-1/2}}{1+(\xi y)^{-1}}\sqrt{\kappa}\tau_* z + (1-(1+\xi_*y)^{-1})x\right|^{k-1}\notag\\
&\leq C (|z|+|x|)^{k-1} \notag\\
&\leq C \left(|x|^{k-1}+|z|^{k-1}\right) \leq C (|x|^{k-1} + \|\vct{a}\|_2^{k-1}).\label{eq:gb}
\end{align}
Here, the value of the constant $C>0$ may change from line to line.
Secondly and similarly, we have the following perturbation bound on the ${\vct{g}}-{\vct{g}}'$. Recall that $y,y'\subset \mathcal{M}$ are bounded. Additionally, since $\mathcal{M}\subset\mathbb{R}^+$ and is compact, $\mathcal{M}$ is strictly bounded away from $0$. Let
\[
g_1(y)=\sqrt{\kappa}\tau_*\frac{y^{-1/2}}{1+(\xi_* y)^{-1}}\quad\text{and}\quad g_2(y)=1-(1+\xi_*y)^{-1}.
\]
It can be seen that $g_1,g_2$ are continuously differentiable functions over $\mathcal{M}$. Thus $g_1,g_2$ are bounded and have bounded derivatives over $\mathcal{M}$. We will prove the following sequence of inequalities
\begin{align}
|{{\vct{g}}-{\vct{g}}'}|&= |g(y,z,x)-g(y',z',x')| \notag\\
&\leq \left|g_1(y) x - g_1(y') x'\right|
+ \left|g_2(y)z-g_2(y')z'\right|\notag\\
&\leq
\left|g_1(y) x - g_1(y) x'\right| +\left|g_1(y) x'- g_1(y') x'\right|\notag\\
&\qquad+ \left|g_2(y)z-g_2(y)z'\right| + \left|g_2(y)z'-g_2(y')z'\right| \notag\\
&\leq C_1|x-x'| + C_2|x'||y-y'| + C_3|z-z'| + C_4|z'||y-y'|\notag\\
&\leq C(1+|x'| + |z'|)(|x-x'| + |z-z'| + |y-y'|)\\
&\leq C\sqrt{3}(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2
.\label{eq:gd}
\end{align}
In the fourth inequality above, we used the fact that $|g_i(y)|,|g_i'(y)|$ are bounded. In the last line, we used Cauchy-Scwhartz.
Substituting \eqref{eq:gb} and \eqref{eq:gd} in \eqref{eq:hlip} gives:
\begin{align}
|h(x,y,z) - h(x',y',z')| &\leq C\left(1+\|\vct{a}\|_2^{k-1}+\|\vct{a}'\|_2^{k-1}+|x|^{k-1}+|x'|^{k-1}\right)(1+|x'| + |z'|)\|[\vct{a},x]-[\vct{a}',x']\|_2\notag \\
&\leq C\left(1+\|[\vct{a},x]\|_2^{k}+\|[\vct{a}',x']\|_2^{k}\right)\|[\vct{a},x]-[\vct{a}',x']\|_2.
\end{align}
Thus, $h\in\rm{PL}(k+1)$, as desired.
\end{proof}
| 42b1d34e911c2b4eae4635cd5dc4bec9f98ee1db | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Central least square estimation of level set gradients}
\label{sec:central_least_square}
The second term in equation~(\ref{eq:reinit_new}) involves computation
of level set gradient terms.
\begin{equation}
\lvert \nabla \psi \rvert_i = \sqrt{\left(\frac{\partial \psi}{\partial
x} \right)_i +
\left(\frac{\partial \psi}{\partial y} \right)_i}
\end{equation}
These terms are evaluated here using central least square approach.
In order to construct cell center derivatives, a stencil consisting of
vertex based neighbours, as shown in
Figure~\ref{fig:vertex_based_stencils}, is considered. Using Taylor
series expansion, the neighbour cell values, $\psi_j$, of the level
set function are expressed in terms of the value at cell $i$, as,
\begin{equation}
\psi_j = \psi_i + (x_j - x_i) \left(\frac{\partial
\psi}{\partial x}\right)_i + (y_j - y_i) \left(\frac{\partial
\psi}{\partial y}\right)_i \dots
\label{eq:taylor-series}
\end{equation}
where, $(x_i, y_i)$ and $(x_j, y_j)$ are locations of the centroids of
cell $i$ and centroid of the neighbour cell $j$ respectively.
Upon truncating the higher order terms (after the third order term)
and re-arranging, equations~(\ref{eq:taylor-series}) can be written
as,
\begin{equation}
\Delta \psi = {\bf S}~\text{d}\psi
\label{eq:taylor-series_matrix}
\end{equation}
where,
\begin{center}
$\Delta \psi =\left\{\begin{array}{c}
\psi_1 - \psi_i\\
\psi_2 - \psi_i\\
\dots\\
\dots\\
\psi_l - \psi_i
\end{array}\right\}$;\hspace{0.5cm}${\bf S} =
\left[\begin{array}{cc}
x_1 - x_i & y_1 - y_i \\
x_2 - x_i & y_2 - y_i \\
\dots & \dots \\
\dots & \dots \\
x_l - x_i & y_l - y_i \\
\end{array}\right]$;\hspace{0.5cm}$\text{d}\psi
=\left\{\begin{array}{c}
\displaystyle \frac{\partial \psi}{\partial x}\\ \\
\displaystyle \frac{\partial \psi}{\partial y}\\
\end{array}\right\}$
\end{center}
The overdetermined system of equation~(\ref{eq:taylor-series_matrix})
can be solved as,
\begin{equation}
\text{d}\psi = \left({\bf S}^{\text{T}}{\bf S}\right)^{-1}
{\bf S}^{\text{T}} \Delta \psi
\label{eq:weighted_normal_method}
\end{equation}
Closed-form expressions for the derivatives can be obtained by
simplifying equation~(\ref{eq:weighted_normal_method}) as,
\begin{subequations}
\begin{align}
\left(\frac{\partial \psi}{\partial x}\right)_i
= \frac{\ell_{22} r_{1} - \ell_{21}
r_{2}}{G}\\
\left(\frac{\partial \psi}{\partial y}\right)_i
= \frac{\ell_{11} r_{2} - \ell_{12} r_{1}}{G}
\end{align}
\label{eq:derivatives}
\end{subequations}
where
$$
\ell_{11} = \sum_{j=1}^l (x_j - x_i)^2, \hspace{0.5cm} \ell_{22}
= \sum_{j=1}^l (y_j - y_i)^2, \hspace{0.5cm} \ell_{12} =
\ell_{21}
= \sum_{j=1}^l (x_j - x_i) (y_j - y_i)
$$
$$
r_{1} = \sum_{j=1}^l (x_j - x_i) (\psi_j - \psi_i),
\hspace{0.5cm} r_{2}
= \sum_{j=1}^l (y_j - y_i) (\psi_j - \psi_i), \hspace{0.5cm}
G
= \ell_{11} \ell_{22} - \ell_{12}^2
$$
\begin{figure}[H]
\centering
\includegraphics[scale=0.75]{./Images/finite_volume_vertex_based_stencil.pdf}
\caption{A schematic representation of a triangular shaped finite
volume cell $i$ and its vertex based neighbours, denoted as $j$.}
\label{fig:vertex_based_stencils}
\end{figure}
\section{Conservative Level set Method}
The fluid-fluid interface in conservative level set method is
represented in the form of an iso-contour of a hyperbolic tangent type
level set function, defined as,
\begin{equation}
\psi({\bf x}, t) = \frac{1}{1 + \exp \left(\frac{-\phi({\bf x}, t)}{\varepsilon}\right)} \equiv
\frac{1}{2} \left( \tanh \left(\frac{\phi({\bf x}, t)}{2 \varepsilon}
\right) + 1 \right)
\label{eq:level_set_function}
\end{equation}
where, $\phi({\bf x}, t)$ is the standard signed distance function
defined in terms of the minimum distance $d({\bf x}, t)$ from the
interface, as,
$$
\phi({\bf x}, t) =
\begin{cases}
-d({\bf x}, t), &\text{ inside the first fluid} \\
0, &\text{ at the fluid-fluid interface} \\
+d({\bf x}, t), &\text{ inside the second fluid}
\end{cases}
$$
The function $\psi$ takes a value $0$ at regions occupied by the
first fluid and $1$ at the second fluid. Within a thin transition
region between the two fluids, $\psi$ varies smoothly from $0$ to
$1$. Width of the transition region is dictated by the parameter
$\varepsilon$. The contour corresponds to $\psi({\bf x},0) = 0.5$
represents the actual fluid-fluid interface. The geometric parameters
associated with the interface, such as interface normal vector (${\bf
n}$) and interface curvature (${\kappa}$), are obtained from the
level set function as,
\begin{equation}
{\bf n} = \frac{\nabla \psi}{\lvert \nabla \psi \rvert}
\label{eq:normal}
\end{equation}
\begin{equation}
\kappa = - \nabla \cdot {\bf n}
\label{eq:curvature}
\end{equation}
Finally, movement of the fluid-fluid interface is achieved by
advecting the level set function according to the flow field, as,
\begin{equation}
\label{eq:ls-advec}
\frac{\partial \psi}{\partial t} + \nabla \cdot \left({\bf u}
\psi\right) = 0
\end{equation}
where, ${\bf u} = u \hat{\text{i}} + v \hat{\text{j}}$, is the
divergence-free velocity field.
\subsection{Artificial Compression based Reinitialization Procedure}
\label{sec:artificial_compression}
It is well known that the level set function suffers from excessive
dissipation due to numerical errors~\cite{Olsson2005}. This leads the
level set function to deviate from its original hyperbolic tangent
type profile. An artificial compression based reinitialization is
developed in reference~\cite{Olsson2007} in order to re-establish the
pre-specified thickness and the profile of the level set function. The
discretized level set advection equation together with the
reinitialization should satisfy the following three
requirements~\cite{Olsson2005}. Firstly, the method should ensure
discrete conservation of mass while advecting the level set
function. Secondly, the method should not introduce any spurious
oscillations. Finally, the initial properties of the level set
function should be maintained throughout the simulation.
The equation for reinitializing the level set function can be written
as per~\cite{Olsson2007} as,
\begin{equation}
\frac{\partial \psi}{\partial \tau_r} + \nabla \cdot \left(\psi (1 -
\psi) {\bf n}_0 \right) - \nabla \cdot \left(\varepsilon
\left(\nabla \psi \cdot {\bf n}_0 \right)~{\bf n}_0\right) = 0
\label{eq:reinit_htf}
\end{equation}
where, ${\bf n}_0 = \displaystyle \frac{\nabla \psi_0}{\lvert
\nabla \psi_0 \rvert}$ is the interface contour normal vector
defined before the reinitialization starts. The variable $\tau_r$ is a
time like variable and $\psi_0$ is the level set function defined at
$\tau_r = 0$. In equation~(\ref{eq:reinit_htf}), the second and third
terms are responsible for the interface compression and diffusion
respectively, which balances each other once the
equation~(\ref{eq:reinit_htf}) converges in time~$\tau_r$.
In practice, the reinitialization formulation suffers from various
deficiencies. As described in the introduction, the above
reinitialization procedure may involve error in the ${\bf n}_0$
computation. It may also be noted, the reinitialization equation moves
the interface contour based on the local curvature. In order to
identify terms that have potential to move the interface contour
during the reinitialization, equation~(\ref{eq:reinit_htf}) is
rewritten in non-conservative form by expanding the the compressive
and diffusive terms. After rewriting equation~(\ref{eq:reinit_htf}) in
non-conservative form, a curvature dependent velocity like term, ${\bf
v} = \varepsilon \kappa_0 \bigg({\bf n}_0 - \displaystyle
\frac{\nabla \phi}{\lvert \nabla \phi \rvert^2}\bigg)$, is isolated
which is found to be responsible for undesired movement of the interface
(details shown in reference \cite{Parameswaran2020}).
\subsection{Reformulation of the Reinitialization Equation}
\label{sec:reformulation}
As described above, the curvature dependent advection term moves the
interface; thus it is undesired in a reinitialization
procedure. Therefore, in the new formulation of the reinitialization
equation, we remove this curvature dependent advection term from the
reinitialization equation. It may also be noted, the reinitialization
equation is sensitive towards numerical errors arising from the
ill-conditioned behaviour of the contour normal vectors. In order to
overcome the difficulty to deal with the these terms, in the new
formulation we seek for terms that are easy to compute and are less
sensitive to numerical errors arising form the ill-conditioned contour
normal vectors. After some manipulations (details shown in reference
\cite{Parameswaran2020}), the final reformulated reinitialization
equation can be written as,
\begin{equation}
\frac{\partial \psi}{\partial \tau_n} = - \psi (1 - \psi) (1 - 2 \psi)
+ \varepsilon (1 - 2 \psi) \lvert \nabla \psi \rvert
\label{eq:reinit_new}
\end{equation}
It can be easily shown that the level set function, $\psi$, as given
in equation~(\ref{eq:level_set_function}), trivially satisfies the
steady state form of equation~(\ref{eq:reinit_new}). Furthermore, the
overall behaviour of equation~(\ref{eq:reinit_new}) can be described
by considering each terms in the RHS individually. In the absence of
the second term in RHS, equation~(\ref{eq:reinit_new}) behaves like an
ordinary differential equation with $\psi = 1$ and $\psi = 0$ as two
stable equilibrium points and $\psi=0.5$ as an unstable equilibrium
point. Figure~\ref{fig:phase_plot} shows the phase plot of
equation~(\ref{eq:reinit_new}) with only first term in the RHS. From
the phase plot, it is clear that the first term results in sharpening
the level set function profile. Moreover, the first term also helps in
stabilizing the overshoot ($\psi > 1.0$) and undershoot ($\psi < 0.0$)
issues arising in case of the use of non-TVD numerical schemes for the
original level set advection equation. Nature of the second term in
the RHS of equation~(\ref{eq:reinit_new}) is to balance the first
term. Since the $\varepsilon$ and $\lvert \nabla \psi \rvert$ are
always positive quantities, the sign of the second term depends on the
sign of $(1 - 2\psi)$. That is, the sign of the second term is
positive when $\psi < 0.5$ and negative when $\psi > 0.5$. At $\psi =
0.5$, the second term is zero. In other words, the second term drives
the level set function towards a flat profile with $\psi = 0.5$
everywhere, thus balancing the first term. With the above mentioned
sharpening and balancing actions, the equation~(\ref{eq:reinit_new})
reinitializes the level set function.
\begin{figure}[H]
\centering
\includegraphics[scale=1.0]{./Images/phase_plot_modified}
\caption{Phase plot for the new reinitialization equation. The arrows
indicate the magnitude and direction of change of $\psi$ with
respect to $\tau_n$. The blue lines show the trajectories of
different initial values of $\psi$ as $\tau_n$ progresses. The
initial and final values of $\psi$ are represented using circles and
squares respectively.}
\label{fig:phase_plot}
\end{figure}
\section{Conclusion}
A new approach to reinitialize the level set function for the CLS
method is formulated in this paper. The two major drawbacks of
the existing artificial compression based reinitialization procedure,
namely, the unwanted movement of the interface contour and strong
sensitivity towards numerical errors leading to formation of
unphysical fluid patches, are resolved with the new approach. Here,
the existing artificial compression based reinitialization equation is
first examined carefully in order to identify the term responsible for
the movement of the interface contour. After isolating and removing a
curvature dependent velocity term, that is responsible for moving the
interface contour, the reinitialization equation is revised. The
remaining terms in the reinitialization equation are then carefully
replaced with equivalent terms that do not involve contour normal
vectors. Unlike the compression and diffusion fluxes present in a
typical artificial compression approach, the newly reformulated
approach has a level set sharpening term, responsible for the
narrowing the level set profile, and a balancing term in order to
counteract the effect of sharpening. The combined effect of sharpening
and balancing restores the level set function without causing any
unwanted movement to the interface contour. Due to the absence of
terms involving contour normal vectors, the susceptibility towards
formation of unphysical fluid patches during reinitialization process
has been completely eliminated in the new reinitialization
procedure. As result of the new reinitialization approach, there has been significant improvement in the mass
conservation property. While
solving the new reinitialization equation, one can choose a larger
time step, approximately by a factor of $4/h$, in comparison with the
allowable time step of an artificial compression based
approach. Moreover, the simplified terms also help in significantly
reducing the numerical computation per reinitialization iteration,
aiding an overall reduction in the computational efforts.
In order to evaluate the performance of the new reinitialization
scheme, three types of numerical test cases are carried out. A
set of in-place reinitialization problems demonstrate that the new
reinitialization approach does not unnecessarily move the interface
contour even after large number of reinitialization iterations. The
area and shape errors of the new approach are quantified and compared
against other reinitialization schemes using a set of scalar advection
based test problems. In order to evaluate the performance on more
practical problems, a set of standard incompressible two-phase flow
problems, starting from an inviscid test case to complex test cases
involving viscous and surface tension forces, are solved. Finally, in
order to demonstrate the ability to deal with complex mesh types, an
incompressible two-phase flow problem is also solved on an
unstructured mesh consisting of finite volume cells having triangular
and quadrilateral shapes. The numerical results of the incompressible
two-phase flow problems show superior results as compared to the
existing approach and match very well with the reference solutions
reported in literature. With the enhanced accuracy and improved
ability to deal with complex mesh types, the proposed reinitialization
approach can be efficiently used in solving real life incompressible
two-phase flow problems.
\section*{Highlights}
\begin{itemize}
\item Developed new reinitialization procedure by reformulating the artificial compression approach.
\item Problems of interface movement and unphysical patch formation during reinitialization are completely resolved.
\item Applicable to a wide variety of meshes including unstructured hybrid meshes and computationally efficient.
\item Numerical results show superior mass conservation and good agreement with analytical and experimental results.
\end{itemize}
\section{Introduction}
\label{sec:intro}
Numerical simulation of incompressible two-phase flows poses great
challenges due to the presence of the fluid-fluid interface. Popular
contact capturing methods, such as, Volume of Fluid (VOF) method and
Level Set (LS) method, use an additional interface advection equation
along with the incompressible Navier-Stokes (NS) equations. Any
inaccuracy in solving the interface advection equation will
significantly affect the quality of the numerical solution. Issues of
the formation of jetsam/ floatsam in VOF method, violation of mass
conservation in LS method are a few examples associated with errors in
the computation of the interface advection equation. In order to
overcome the mass conservation error in the classical LS method, a
variant of the LS method, known as Conservative Level Set (CLS)
method, is proposed by Olsson and Kreiss~\cite{Olsson2005} and
Olsson~et~al~\cite{Olsson2007}. The enhanced mass conservation
property is achieved here by replacing the signed distance function
used in the classical level set method with a hyperbolic tangent type
level set function. This level set function is then advected using a
scalar conservation law. In order to recover from the excessive
numerical dissipation error, an artificial compression based
reinitialization procedure is also formulated for the level set
function. With the improved mass conservation property, the CLS method
shows promising capabilities and provides a good alternative to
the classical LS method in solving incompressible two-phase
flow problems. However, in practice, the reinitialization procedure
for level set function in CLS method often runs into various numerical
difficulties, leading to undesirable results.
Primarily, two major issues with the reinitialization procedure are
reported in the literature~\cite{Shukla2010, Mccaslin2014,
Desjardins2008}. The first one is the undesired movement of the
interface contour during reinitialization. The problem gets further
aggravated with the frequent use of reinitialization. This issue
arises particularly in two and three dimensions, where, the interface
curvature gets involved in the computations. It is demonstrated
in reference~\cite{Shukla2010} that the degree of movement of
interface contour depends upon the strength of the interface
curvature. Also, a set of numerical experiments presented in
reference~\cite{Mccaslin2014} verifies that the frequent use of
reinitialization results in substantial movement of the interface,
leading to inaccuracies in the numerical solution. Several attempts to
resolve this issue can be found in literature~\cite{Yohei2012,
Mccaslin2014}. The efforts are focused mainly on localizing the
reinitialization process only to a selected region of the level set
field. Excessive reinitialization at less dissipated regions is thus
avoided. In reference~\cite{Yohei2012}, the reinitialization is
localized by defining a local coefficient based on the degree of
sharpness of the level set function. Whereas, in
reference~\cite{Mccaslin2014}, a metric which depends on the local
flow conditions and the numerical diffusion errors is used. It is,
therefore, clear that these methods introduce additional complexity
and computational efforts in obtaining the local coefficients.
The second issue of the original CLS method is the formation of
unphysical fluid patches away from the interface due to the
ill-conditioned behaviour of the interface contour normal
vectors. The contour normal vectors decide the direction of the
compression and diffusion fluxes during the reinitialization
process. Far away from the fluid-fluid interface these contour normal
vectors are expected to be zero, resulting in no
reinitialization. However, due to their ill-conditioned nature, even
far away from the interface the contour normal vectors may not
necessarily be zero. This results in undesirable compression
at far away region, which leads to formation of unphysical fluid
patches there. Several cures for this problem can be found in
literature~\cite{Desjardins2008, Shukla2010, Zhao2014,
Waclawczyk2015, Chiodi2017, Tabar2018}. Most of the efforts are
targeted towards replacing the ill-conditioned contour normal vectors
with some alternatives. In the Accurate Conservative Level Set (ACLS)
method, proposed in reference~\cite{Desjardins2008}, the contour
normal vectors are computed from an auxiliary signed distance
function. The auxiliary signed distance function is constructed here
from the level set function using a fast marching method. In the
method proposed by Shukla~et~al.~\cite{Shukla2010}, a modified form of
the reinitialization equation is used. Here, the level set function is
replaced with a smooth function constructed from the level set
function itself by employing a mapping procedure. A technique, by
combining the reinitialization schemes of both the classical and
conservative level set method, named as Improved Conservative Level
Set (ICLS) Method, is reported in reference~\cite{Zhao2014}. In
reference~\cite{Waclawczyk2015} and its improved version in
reference~\cite{Chiodi2017}, reformulations of the original
reinitialization equation are presented which take care of the
spurious movement of interface contour as well as the ill-conditioned
behaviour of the contour normal vectors. Recently in
reference~\cite{Tabar2018}, the ill-conditioned unit contour normal
vectors are replaced with another normal vectors, such that, their
magnitudes start to diminish away from the interface. This ensures
that the reinitialization process is activated only near the interface
regions. Though improvements in the contour normal vectors partially
circumvent the issue of formation of unphysical fluid patches, they
involve evaluation of more complicated terms adding to the overall
computational cost.
In the present work, a much simpler technique to reinitialize the
level set function is presented. Here, the existing artificial
compression based reinitialization equation is revised by isolating
and removing terms that have potential to move the interface
contours. The remaining terms in the modified equation are then
reformulated considerably, such that, the usage of the contour normal
vectors is completely avoided. With the new reformulation, issues
such as, distortion of the interface contour and the unphysical patch
formation away from the interface are resolved. As a consequence, the area conservation property is improved significantly. In addition, absence of a
viscous dissipation like (second derivative) term in the new
reinitialization equation enables one to choose a much larger time
step during the reinitialization iteration. The simplified
terms also help in significantly reducing the numerical computations
per reinitialization time step, aiding an overall reduction in the
computational efforts. In order to demonstrate the efficacy of the
proposed reinitialization scheme, a set of standard two-dimensional
test problems involving reinitialization of stationary level set
functions (henceforth, we name it as in-place reinitialization
problems), advection of level set function under predefined velocity
fields and a few standard incompressible two-phase flow problems are
solved. Finally, in order to demonstrate the ability to deal with
complex mesh types, an incompressible two-phase flow problem is solved
on an unstructured mesh consisting of finite volume cells having
triangular and quadrilateral shapes.
Rest of the paper is organized as follows. The original CLS method and
its reinitialization scheme is briefly described in
section~2. The limitations of the existing artificial compression
based reinitialization approach and its new reformulation are also
discussed in the same section. In section~3, the mathematical
formulation of incompressible two-phase flows is briefly
described. The numerical discretization of the governing equations and
the new reinitialization equation are also described in
section~3. Several numerical test problems are solved in
section~4. Finally the conclusions are given in section~5.
\section{Mathematical Formulation of Incompressible Two-Phase Flows}
\subsection{Governing Equations}
A dual time-stepping based artificial compressibility approach is
followed here for modelling incompressible two-phase flows. The
governing system of equations describing the unsteady incompressible
viscous two-phase flow can be written as,
\begin{equation}
\frac{\partial \bf U}{\partial \tau} + I^t \frac{\partial \bf
U}{\partial t} + \left[\frac{\partial {\bf (F -
F_v)}}{\partial
x} + \frac{\partial {\bf (G - G_v)}}{\partial y}\right] =
{\bf F_g} + {\bf F_s}
\label{eq:gov_eq_two_phase}
\end{equation}
where,
\begin{center}
${\bf U} =\left\{\begin{array}{c}
p/\beta\\
\rho u\\
\rho v\\
\psi
\end{array}\right\}$; ${\bf F} = \left\{\begin{array}{c}
u\\
\rho u^{2}+p\\
\rho uv\\
u \psi
\end{array}\right\}$; ${\bf G} = \left\{\begin{array}{c}
v\\
\rho uv\\
\rho v^{2}+p\\
v \psi
\end{array}\right\}$; ${\bf F_v} = \left\{\begin{array}{c}
0 \\
2\mu \frac{\partial u}{\partial x} \\
\mu (\frac{\partial u}{\partial y} + \frac{\partial
v}{\partial x})\\
0
\end{array}\right\}$; ${\bf G_v} = \left\{\begin{array}{c}
0 \\
\mu (\frac{\partial u}{\partial y} + \frac{\partial
v}{\partial x})\\
2\mu \frac{\partial v}{\partial y}\\
0
\end{array}\right\}$; ${\bf F_g} = \left\{\begin{array}{c}
0 \\
-\rho g_x \\
-\rho g_y \\
0
\end{array}\right\}$; \hspace{0.5cm} ${\bf F_s} =
\left\{\begin{array}{c}
0 \\
-\sigma \kappa \frac{\partial \psi}{\partial x}\\ \\
-\sigma \kappa \frac{\partial \psi}{\partial y}\\
0
\end{array}\right\}$; \hspace{0.5cm} $I^t =
\left[\begin{array}{cccc}
0 & 0 & 0 & 0\\
0 & 1 & 0 & 0\\
0 & 0 & 1 & 0\\
0 & 0 & 0 & 1\\
\end{array}\right]$.
\end{center}
where, the vector ${\bf U}$ in equation~(\ref{eq:gov_eq_two_phase})
denotes the vector of conservative variables and the vectors $({\bf
F}, {\bf G})$ and $({\bf F_v}, {\bf G_v})$ denote the convective and
viscous flux vectors respectively. The vectors ${\bf F_g}$ and ${\bf
F_s}$ denote the source terms containing gravitational and surface
tension forces respectively. Here, the surface tension term is
modelled using a continuum surface forces (CSF) method, proposed by
Brackbill~et~al.~\cite{Brackbill1992}. The parameter $\sigma$ denotes
the surface tension coefficient per unit length of interface and $g_x$
and $g_y$ denote the $x$ and $y$ components of the acceleration due to
gravity. The variable $p$, $\rho$ and $\mu$ denote the pressure,
density and dynamic viscosity of the fluid respectively. The variables
$\tau$ and $t$ appearing in equation~(\ref{eq:gov_eq_two_phase})
denote the pseudo time and the real time respectively. The parameter
$\beta$ denotes the artificial compressibility parameter, which is
usually taken as constant for a given test problem. The artificial
compressibility term added in the continuity equation is similar to
the one introduced by Chorin in~\cite{Chorin1967}. Once
equation~(\ref{eq:gov_eq_two_phase}) converges to a pseudo-steady
state, it recovers the set of unsteady incompressible two-phase flow
equations. It can be noticed that the level set advection, described
by equation(\ref{eq:ls-advec}), is combined here with the system of
equations~(\ref{eq:gov_eq_two_phase}), and solved simultaneously along
with the Navier-Stokes equations. The density and viscosity used in
equation~(\ref{eq:gov_eq_two_phase}) are defined in terms of the level
set function, as,
\begin{eqnarray}
\label{eq:den_lsm}
\rho = \rho(\psi) = \rho_{\text{2}}\psi + (1 - \psi) \rho_{\text{1}}
\\
\mu = \mu(\psi) = \mu_{\text{2}}\psi + (1 - \psi) \mu_{\text{1}}
\label{eq:visc_lsm}
\end{eqnarray}
where the subscripts ``1'' and ``2'' indicate the properties
corresponds to the first and the second fluids respectively.
\subsection{Numerical Discretization of Governing Equations}
\label{sec:fv_formulation}
A finite volume approach is followed here for solving the governing
system of equations~(\ref{eq:gov_eq_two_phase}). In order to proceed
with finite volume discretization, the governing system of
equations~(\ref{eq:gov_eq_two_phase}) is first integrated over a
control volume. The computational domain is then discretized
into a finite number of non-overlapping finite volume cells. The final
space discretized form of equations~(\ref{eq:gov_eq_two_phase}) for
an $i^{\text{th}}$ finite volume cell can be written as,
\begin{equation}
\Omega_{i} \frac{\partial {\overline{ \bf U}}}{\partial \tau} + I^t
\Omega_{i} \frac{\partial {\overline{ \bf U}}}{\partial t} + R\left({
\overline{\bf U}}\right) = 0
\label{eq:fv_disc_eq}
\end{equation}
where,
\begin{equation*}
R\left({ \overline{\bf U}}\right) = \sum_{m=1}^{M} \left( {\bf F}~n^m_x + {\bf
G}~n^m_y \right)_m \Gamma_m - \sum_{m=1}^{M} \left( {\bf F_v}~n^m_x
+ {\bf G_v}~n^m_y \right)_m \Gamma_m - \Omega_{i} {\overline{\bf
F}_g} - \Omega_{i} {\overline{\bf F}_s},
\end{equation*}
$\Omega_i$ is the area, $\Gamma_m$ and ${\bf n}^m = (n^m_x,
n^m_y)$ are the length and edge normals of the $m^{\text{th}}$ edge
respectively and $M$ is the total number of edges of the finite volume
cell~$i$. The vectors ${\overline{\bf U}}$, ${\overline{\bf F}_g}$ and
${\overline{\bf F}_s}$ represent the cell averaged values of ${\bf
U}$, ${\bf F_g}$ and ${\bf F_s}$ respectively. The source
term vector, ${\bf \overline{F}_g}$, appearing in
equation~(\ref{eq:fv_disc_eq}) is computed by multiplying the cell
averaged value of density and the acceleration due to
gravity. Evaluation of the surface tension vector, ${\overline{\bf
F}_s}$, involves the computation
of gradient, contour normal and curvature of the level set
function. The contour normal and curvature are computed from the
gradient of the level set function as per equation~(\ref{eq:normal})
and equation~(\ref{eq:curvature}) respectively, where the gradient
vector is evaluated using central least square method as explained
in~\cite{Parameswaran2019-2}. The convective flux vector $\left( {\bf
F}~n^m_x + {\bf G}~n^m_y \right)$ and the viscous flux vector
$\left({\bf F_v}~n^m_x + {\bf G_v}~n^m_y \right)$ in
equation~(\ref{eq:fv_disc_eq}) are computed at the edges of each cell
using a Roe-type Riemann solver, developed in~\cite{Parameswaran2019},
and a Green-Gauss integral approach over a Coirier's diamond path,
described in~\cite{Parameswaran2019-2}, respectively. The real time
derivatives appearing in equation~(\ref{eq:fv_disc_eq}) are computed
using a three point implicit backward differencing procedure. Finally,
an explicit three stage Strong Stability Preserving Runge-Kutta
(SSP-RK) method, described in~\cite{Gottlieb2005}, is used for
iterating in pseudo-time. The time step required for the pseudo-time
iteration is computed by considering the convective, viscous and
surface tension effects. For faster convergence, a local time stepping
approach is adopted here, in which, each cell is updated using its own
$\Delta \tau_i$. The local time step $\Delta \tau_i$ is computed as,
\begin{equation}
\label{eq:pseudo_dt}
\Delta \tau_i = \text{min}(\Delta \tau^{\text{conv}}_i, \Delta \tau^{\text{visc}}_i, \Delta \tau^{\text{surf}}_i)
\end{equation}
where, $\Delta \tau^{\text{conv}}_i, \Delta \tau^{\text{visc}}_i$ and
$\Delta \tau^{\text{surf}}_i$ are the maximum allowed time steps due to
convective flux, viscous flux and surface tension force
respectively. These time steps are evaluated as,
\begin{equation*}
\Delta \tau^{\text{conv}}_i = \frac{\nu ~
\Omega_{i}}{\sum_{m=1}^{M}\left(\lvert u_n \rvert +
\sqrt{\left(u_n \right)^2 + \frac{\beta}{\rho_i}} \right)_m
\Gamma_m } \text{;}
\Delta \tau^{\text{visc}}_i = \frac{\nu ~
\Omega_{i}^2}{\left(\frac{8}{3}\right)\frac{\mu_i}{\rho_i}
\sum_{m=1}^{M}\left(\Gamma_m \right)^2} \text{ and }
\Delta \tau^{\text{surf}}_i = \nu ~ \sqrt{\frac{(\rho_1 + \rho_2)
h^3}{4 \pi \sigma}}
\end{equation*}
where, $\nu$ is the Courant number, $h$ is the average cell size and
$u_n$ is the velocity component along the edge normal direction. For
stability reasons, the Courant number, $\nu$, is always taken less
than unity. The $\Delta \tau_i$ computed using
equation~(\ref{eq:pseudo_dt}) is further restricted based on the real
time step~\cite{Gaitonde1998}, as, $\Delta \tau_i \leq
\frac{2}{3}\Delta t$. Detailed descriptions of the numerical methods
used for solving incompressible two-phase flows are excluded from here
due to brevity reasons. One can refer~\cite{Parameswaran2019} and
\cite{Parameswaran2019-2} for more details.
\subsection{Numerical Discretization of Reinitialization Equation}
The RHS of equation~(\ref{eq:reinit_new}) consists of two terms. In
order to evaluate the first term at a given cell, only the cell center
values of $\psi$ is sufficient. However, evaluation of the second term
involves computation of the gradient of $\psi$. Here, the gradient of
$\psi$ at the cell centers are evaluated using a central least square
approach discussed in~\ref{sec:central_least_square}. Finally,
the time integration of equation~(\ref{eq:reinit_new}) is carried out
using a three stage Strong Stability Preserving Runge-Kutta (SSP-RK-3)
method described in section~\ref{sec:ssprk-3}.
\subsubsection{Time integration for the reinitialization equation}
\label{sec:ssprk-3}
According to the SSP-RK-3 approach described in~\cite{Gottlieb2005},
the cell averaged value of the unknown function in
equation~(\ref{eq:reinit_new}) is updated as,
\begin{equation}
\begin{array}{r@{}l@{}l@{}l@{}l@{}l@{}l}
\overline{\psi_i}^{1} &{}=&{}&{}&{}\overline{\psi_i}^{n} &{}+&{} \displaystyle \frac{\Delta
\tau_n}{\Omega_i}~L\left({\overline{\psi_i}}^{n}\right)\\
\overline{\psi_i}^{2} &{}= \displaystyle
\frac{3}{4}&{}\overline{\psi_i}^{n} &{}+ \displaystyle
\frac{1}{4} &{}\overline{\psi_i}^{1} &{}+
\displaystyle \frac{1}{4} &{}\displaystyle \frac{\Delta
\tau_n}{\Omega_i}~L\left({\overline{\psi_i}}^{1}\right)\\
\overline{\psi_i}^{n+1} &{}= \displaystyle
\frac{1}{3}&{}\overline{\psi_i}^{n} &{}+ \displaystyle \frac{2}{3}&{}\overline{\psi_i}^{2} &{}+
\displaystyle \frac{2}{3} &{}\displaystyle \frac{\Delta \tau_n}{\Omega_i}~L\left({\overline{\psi_i}}^{2}\right)
\end{array}
\label{eq:ssprk}
\end{equation}
where, $\overline{\psi_i}^n$ and $\overline{\psi_i}^{n+1}$ are the cell
averaged level set function defined at $n^{\text{th}}$ and
$(n+1)^{\text{th}}$ time levels respectively, $\overline{\psi_i}^1$ and
$\overline{\psi_i}^2$ are the intermediate values of $\psi$ and
$L\left({\overline{\psi_i}}^{*}\right) = - \overline{\psi_i}^{*}(1
- \overline{\psi_i}^{*})(1 - 2\overline{\psi_i}^{*}) + \varepsilon
\lvert \nabla \overline{\psi_i}^{*} \rvert (1 -
2\overline{\psi_i}^{*})$. For the explicit time integration of
equation~(\ref{eq:reinit_new}), the time step is restricted based on
the nature of the reinitialization equation. In order to find out the
allowable time step, equation~(\ref{eq:reinit_new}) is rewritten in a
Hamilton-Jacobi form with a velocity like variable, ${\bf \mathcal{S}}
= \varepsilon(1 - 2\psi) \left( \frac{1 - \lvert \nabla \phi
\rvert}{\lvert \nabla \phi \rvert} \right){\bf n}$. Since the
solution variable $\psi$ is updated according to the sharpening
velocity vector ${\bf \mathcal{S}}$, a stable explicit time
integration scheme for equation~(\ref{eq:reinit_new}) is possible
only with a restricted time step,
\begin{equation}
\Delta \tau_n \leq 2h^d
\label{eq:reinit_time_step}
\end{equation}
It may be noticed that the time step, $\Delta \tau_n$, given in
equation~(\ref{eq:reinit_time_step}) is larger by a factor of
$4/h$ in comparison with the allowable time step for
the artificial compression based reinitialization
procedure~\cite{Olsson2005}. Presence of a viscous dissipation term in
the artificial compression based approach restricts the
reinitialization time step to a smaller value~\cite{Olsson2005}.
\section{Acknowledgement}
The present work is partially supported by Aeronautics Research \&
Development Board (AR\&DB), with the project Grant number
ARDB/01/1031930/M/I DT. 13.09.2019. We gratefully thank AR\&DB for the
support.
\newpage
\section{Numerical Experiments}
\label{sec:num_exp}
Performance of the new reinitialization procedure is evaluated using
three types of test problems. In order to illustrate the movement of
the interface contour during the reinitialization, a set of test
problems involving reinitialization of stationary level set function,
is carried out first in section~\ref{sec:stationary_pblms}. These
problems are named here as in-place reinitialization
problems. Followed to the in-place reinitialization problems, a set of
scalar advection based test problems are considered in
section~\ref{sec:scalar_aadv}, where, the area and shape errors during
the level set advection are quantified. In the
sections~\ref{sec:db}~-~\ref{sec:br}, several incompressible two-phase
flow test problems are presented. These problems are arranged
according to their increasing levels of complexities, starting from an
inviscid flow problem to problems involving viscous and surface
tension forces. All test problems are solved on simple Cartesian type
meshes. However, in order to demonstrate the the ability to deal with
complex mesh types, the last test problem is also solved on an
unstructured mesh consisting of finite volume cells having triangular
and quadrilateral shapes.
\subsection{In-Place Reinitialization Problems}
\label{sec:stationary_pblms}
As illustrated in section~2.1, the reinitialization scheme often
results in moving the interface contour according to the sign and
strength of the interface curvature. In order to demonstrate this, a
set of test problems involving reinitialization of stationary level
set function, similar to the one reported in~\cite{Chiodi2017}, is
carried out here. In order to perform in-place reinitialization tests,
a level set function, corresponds to some given geometry, is
constructed first. For the present study, three standard shapes,
namely, a circle with diameter of 4 units, an ellipse with 4 units and
2 units of major and minor axes respectively and a square with size 3
units, are chosen. The geometric shapes are placed at the center of a
computational domain of square shaped region bounded between $-5 \leq
x \leq 5$ and $-5 \leq y \leq 5$. The computational domain is
discretized using a $200 \times 200$ Cartesian mesh. The level set
function corresponds to the given shape is then taken as the initial
condition for the reinitialization equation and carried out large number
of pseudo-time iterations. Under ideal situations, the
reinitialization process should not result in movement of the
interface contour. However, due to errors in the reinitialization
scheme, this need not be satisfied always. Moreover, unlike other test
problems, numerical errors associated with the advection of level set
function are not present here. Therefore, these tests help in
isolating the errors associated with only the reinitialization
process.
In the present study, the original CLS reinitialization
algorithm, described in~\cite{Olsson2007}, and the newly proposed
reinitialization algorithm are compared. The deformation of the
interface contour in both the cases are monitored constantly during
the pseudo-time
iterations. Figure~\ref{fig:ipr_circle},~\ref{fig:ipr_ellipse}~and~\ref{fig:ipr_square}
show the interface contours during the in-place reinitialization
compared with the initial contours in case of the circle, ellipse and
square shapes respectively. The solid black curve denotes the
interface contour during reinitialization and the dashed black curve
denotes the initial interface contour. One can notice that, for all
the three shapes, up to 10 number of reinitialization iterations no
significant changes in the interface contours are visible. However, as
the number of reinitialization iterations increases, the original CLS
approach leads to interface contour deformations. Especially, more
deformations can be observed at regions having higher
curvature. Whereas, there are no visible deformations of the interface
contours even after 250 iterations in case of the new reinitialization approach. This
observation is in well agreement with the discussion given in
section~\ref{sec:artificial_compression}~and~\ref{sec:reformulation}.
\begin{figure}[H]
\centering
\subfloat[CLS-Olsson (no.iter = 0)]
{\includegraphics[scale=0.47]{Images/Satationary/Circle/CLS/0}}
\subfloat[CLS-Olsson (no.iter = 10)]
{\includegraphics[scale=0.47]{Images/Satationary/Circle/CLS/10}}
\subfloat[CLS-Olsson (no.iter = 100)]
{\includegraphics[scale=0.47]{Images/Satationary/Circle/CLS/100}}
\subfloat[CLS-Olsson (no.iter = 250)]
{\includegraphics[scale=0.47]{Images/Satationary/Circle/CLS/250}}
\subfloat[New (no.iter = 0)]
{\includegraphics[scale=0.47]{Images/Satationary/Circle/New/0}}
\subfloat[New (no.iter = 10)]
{\includegraphics[scale=0.47]{Images/Satationary/Circle/New/10}}
\subfloat[New (no.iter = 100)]
{\includegraphics[scale=0.47]{Images/Satationary/Circle/New/100}}
\subfloat[New (no.iter = 250)]
{\includegraphics[scale=0.47]{Images/Satationary/Circle/New/250}}
\caption{In-place reinitialization of a circular interface using the
original CLS approach (a) to (d) and the new approach (e) to
(h). The dashed black curve denotes the initial interface and the
solid black curves denote the interfaces at the respective
pseudo-time iterations.}
\label{fig:ipr_circle}
\end{figure}
\begin{figure}[H]
\centering
\subfloat[CLS-Olsson (no.iter = 0)]
{\includegraphics[scale=0.47]{Images/Satationary/Ellipse/CLS/0}}
\subfloat[CLS-Olsson (no.iter = 10)]
{\includegraphics[scale=0.47]{Images/Satationary/Ellipse/CLS/10}}
\subfloat[CLS-Olsson (no.iter = 100)]
{\includegraphics[scale=0.47]{Images/Satationary/Ellipse/CLS/100}}
\subfloat[CLS-Olsson (no.iter = 250)]
{\includegraphics[scale=0.47]{Images/Satationary/Ellipse/CLS/250}}
\subfloat[New (no.iter = 0)]
{\includegraphics[scale=0.47]{Images/Satationary/Ellipse/New/0}}
\subfloat[New (no.iter = 10)]
{\includegraphics[scale=0.47]{Images/Satationary/Ellipse/New/10}}
\subfloat[New (no.iter = 100)]
{\includegraphics[scale=0.47]{Images/Satationary/Ellipse/New/100}}
\subfloat[New (no.iter = 250)]
{\includegraphics[scale=0.47]{Images/Satationary/Ellipse/New/250}}
\caption{In-place reinitialization of an elliptical interface using
the original CLS approach (a) to (d) and the new approach (e)
to (h). The dashed black curve denotes the initial interface and the
solid black curves denote the interfaces at the respective
pseudo-time iterations.}
\label{fig:ipr_ellipse}
\end{figure}
\begin{figure}[H]
\centering
\subfloat[CLS-Olsson (no.iter = 0)]
{\includegraphics[scale=0.47]{Images/Satationary/Square/CLS/0}}
\subfloat[CLS-Olsson (no.iter = 10)]
{\includegraphics[scale=0.47]{Images/Satationary/Square/CLS/10}}
\subfloat[CLS-Olsson (no.iter = 100)]
{\includegraphics[scale=0.47]{Images/Satationary/Square/CLS/100}}
\subfloat[CLS-Olsson (no.iter = 250)]
{\includegraphics[scale=0.47]{Images/Satationary/Square/CLS/250}}
\subfloat[New (no.iter = 0)]
{\includegraphics[scale=0.47]{Images/Satationary/Square/New/0}}
\subfloat[New (no.iter = 10)]
{\includegraphics[scale=0.47]{Images/Satationary/Square/New/10}}
\subfloat[New (no.iter = 100)]
{\includegraphics[scale=0.47]{Images/Satationary/Square/New/100}}
\subfloat[New (no.iter = 250)]
{\includegraphics[scale=0.47]{Images/Satationary/Square/New/250}}
\caption{In-place reinitialization of a square shaped interface using
original CLS approach (a) to (d) and the the new approach (e)
to (h). The dashed black curve denotes the initial interface and the
solid black curves denote the interfaces at the respective
pseudo-time iterations.}
\label{fig:ipr_square}
\end{figure}
\subsection{Reinitialization of Scalar Advection Problems}
\label{sec:scalar_aadv}
In order to further study the performance of the new reinitialization
formulation, a set of standard two-dimensional scalar advection based
test problems are considered next. In the scalar advection problems,
the initial interface is placed at $(0.25,~0.5)$ on a unit square
domain and advected upon a predefined velocity field. Simulations are
carried out till the initial interface completes one full
rotation. The area confined by the interface and the $L^1$ and $L^2$
error norms are monitored during the simulation. The percentage area
error, at any given time $t$, is computed as,
\begin{equation}
\text{Area Error (\%)} = \Bigg(\frac{A^{t} - A^0}{A^0}\Bigg) \times 100
\label{eq:area_error}
\end{equation}
where, $A$ is the area enclosed by the 0.5 contour of the level set
function and the superscript $t$ and $0$ represent the data computed
at time $t$ and at the initial time, $t=0$, respectively. Similarly,
the $L^1$ and $L^2$ error norms are defined as,
\begin{equation}
L^1 = \left( \frac{1}{Nx \times Ny} \right)
\sum_{i=1}^{Nx}\sum_{j=1}^{Ny}\lvert \psi^{2\pi}_{ij} - \psi^0_{ij}\rvert
\end{equation}
and
\begin{equation}
L^2 = \left( \frac{1}{Nx \times Ny} \right)
\sqrt{\sum_{i=1}^{Nx}\sum_{j=1}^{Ny}\Big(\psi^{2\pi}_{ij} - \psi^0_{ij}\Big)^2}
\end{equation}
where, $\psi^{2\pi}_{ij}$ and $\psi^0_{ij}$ are the discretized level
set functions defined at time level $t = 2\pi$ and $t = 0$ respectively,
and $Nx$ and $Ny$ are the total number of cells along $x$ and $y$
directions respectively. In this section, the test problems are solved
using the reinitialization schemes reported in~\cite{Olsson2007,
Desjardins2008, Waclawczyk2015, Chiodi2017} along with the new
scheme. For better clarity, the schemes are denoted here as
CLS-Olsson, ACLS-Desjardins, CLS-Wac{\l}awczyk and CLS-Chiodi for
references \cite{Olsson2007}, \cite{Desjardins2008},
\cite{Waclawczyk2015} and \cite{Chiodi2017} respectively.
\subsubsection{Reinitialization of circular disc rotation problem}
Rotation of a circular disc, similar to the test reported in
\cite{Olsson2005}, is considered first, where, a circular disc of
radius 0.15 units is advected upon a velocity field $u = (y-0.5)$ and
$v = (0.5-x)$. Test problem is solved on four levels of Cartesian
meshes starting from $25 \times 25$ up to $200 \times 200$. The level
set contours correspond to $\psi = 0.05, \psi = 0.5$ and $\psi =
0.95$ for the $100 \times 100$ case at time levels $t = 0$,
$t = \pi/4$, $t = \pi/2$ and $t = 3\pi/4$ for all five
reinitialization schemes are plotted in
Figure~\ref{fig:OC_contours}. Further,
Table~\ref{tab:OC_area_error}~and~\ref{tab:OC_l1_l2} show the
percentage area error and error norms computed after the disc
completes one full rotation ($ie.,$ at $t = 2 \pi$). Moreover, the
error norms are plotted against the mesh size in
Figure~\ref{fig:OC_order_error} along with reference slopes for the
first and second order rate of convergence. Looking at the above
figures and tables, it can be noticed that the new reinitialization
shows the least error among all.
\begin{figure}[H]
\centering
\subfloat[CLS-Olsson]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Olsson-Circle/100/CLS-Olsson.png}}
\subfloat[ACLS-Desjardins]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Olsson-Circle/100/ACLS-Desjardins.png}}
\subfloat[CLS-Wac{\l}awczyk]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Olsson-Circle/100/CLS-Waclawszyk.png}}
\subfloat[CLS-Chiodi]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Olsson-Circle/100/CLS-Chiodi.png}}
\subfloat[New]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Olsson-Circle/100/New.png}}
\caption{Comparison of reinitialization schemes for the
reinitialization of circular disc rotation problem solved on $100
\times 100$ grid. Each subfigure shows the 0.05, 0.5 and 0.95 level
contours of the level set functions at time levels of 0, $\pi$/4,
$\pi$/2 and $3\pi$/4 (left, top, right and bottom respectively).}
\label{fig:OC_contours}
\end{figure}
\begin{table}[H]
\begin{center}
\caption{Percentage area errors computed after completion of one full rotation of different reinitialization schemes for the reinitialization of circular disc rotation problem.}
\label{tab:OC_area_error}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
{\color[HTML]{000000} } & \multicolumn{5}{c|}{{\color[HTML]{000000} \text{Reinitialization Schemes}}} \\ \cline{2-6}
\multirow{-2}{*}{{\color[HTML]{000000} \text{Mesh}}} & {\color[HTML]{000000} \text{CLS-Olsson}} & {\color[HTML]{000000} \text{ACLS-Desjardins}} & {\color[HTML]{000000} \text{CLS-Wac{\l}awszyk}} & {\color[HTML]{000000} \text{CLS-Chiodi}} & {\color[HTML]{000000} \text{New}} \\ \hline
{\color[HTML]{000000} $25 \times 25$} & {\color[HTML]{000000} -5.48500} & {\color[HTML]{000000} -5.04690} & {\color[HTML]{000000} -4.38460} & {\color[HTML]{000000} -4.41610} & {\color[HTML]{000000} -1.95050} \\ \hline
{\color[HTML]{000000} $50 \times 50$} & {\color[HTML]{000000} -2.21890} & {\color[HTML]{000000} -2.12190} & {\color[HTML]{000000} -2.15960} & {\color[HTML]{000000} -2.03820} & {\color[HTML]{000000} -0.15707} \\ \hline
{\color[HTML]{000000} $100 \times 100$} & {\color[HTML]{000000} -1.99710} & {\color[HTML]{000000} -1.94980} & {\color[HTML]{000000} -1.91780} & {\color[HTML]{000000} -1.74080} & {\color[HTML]{000000} -0.02158} \\ \hline
{\color[HTML]{000000} $200 \times 200$} & {\color[HTML]{000000} -1.71600} & {\color[HTML]{000000} -1.68580} & {\color[HTML]{000000} -1.71300} & {\color[HTML]{000000} -1.65980} & {\color[HTML]{000000} -0.00865} \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[H]
\begin{center}
\caption{The $L^1$ and $L^2$ error norms of different reinitialization schemes for the reinitialization of circular disc rotation problem.}
\label{tab:OC_l1_l2}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
{\color[HTML]{000000} } & {\color[HTML]{000000} } & \multicolumn{5}{c|}{{\color[HTML]{000000} \text{Reinitialization Schemes}}} \\ \cline{3-7}
\multirow{-2}{*}{{\color[HTML]{000000} \text{Errors}}} &
\multirow{-2}{*}{{\color[HTML]{000000} \text{Mesh}}} &
{\color[HTML]{000000} \text{CLS-Olsson}} &
{\color[HTML]{000000} \text{ACLS-Desjardins}} &
{\color[HTML]{000000} \text{CLS-Wac{\l}awszyk}} &
{\color[HTML]{000000} \text{CLS-Chiodi}} &
{\color[HTML]{000000} \text{New}} \\ \hline
\hline
{\color[HTML]{000000} } & {\color[HTML]{000000} $25 \times 25$} & {\color[HTML]{000000} 2.6488E-02} & {\color[HTML]{000000} 2.3154E-02} & {\color[HTML]{000000} 2.1835E-02} & {\color[HTML]{000000} 1.9378E-02} & {\color[HTML]{000000} 1.4665E-02} \\ \cline{2-7}
{\color[HTML]{000000} } & {\color[HTML]{000000} $50 \times 50$} & {\color[HTML]{000000} 5.2915E-03} & {\color[HTML]{000000} 5.4820E-03} & {\color[HTML]{000000} 5.2766E-03} & {\color[HTML]{000000} 5.5800E-03} & {\color[HTML]{000000} 5.5977E-03} \\ \cline{2-7}
{\color[HTML]{000000} } & {\color[HTML]{000000} $100 \times 100$} & {\color[HTML]{000000} 3.3020E-03} & {\color[HTML]{000000} 3.3816E-03} & {\color[HTML]{000000} 3.3176E-03} & {\color[HTML]{000000} 3.4732E-03} & {\color[HTML]{000000} 2.4589E-03} \\ \cline{2-7}
\multirow{-4}{*}{{\color[HTML]{000000} $L^1$ Error}} & {\color[HTML]{000000} $200 \times 200$} & {\color[HTML]{000000} 2.1747E-03} & {\color[HTML]{000000} 2.2122E-03} & {\color[HTML]{000000} 2.1747E-03} & {\color[HTML]{000000} 2.2454E-03} & {\color[HTML]{000000} 1.2175E-03} \\ \hline \hline
{\color[HTML]{000000} } & {\color[HTML]{000000} $25 \times 25$} & {\color[HTML]{000000} 3.5611E-03} & {\color[HTML]{000000} 2.8689E-03} & {\color[HTML]{000000} 2.4137E-03} & {\color[HTML]{000000} 2.4929E-03} & {\color[HTML]{000000} 2.4929E-03} \\ \cline{2-7}
{\color[HTML]{000000} } & {\color[HTML]{000000} $50 \times 50$} & {\color[HTML]{000000} 4.4320E-04} & {\color[HTML]{000000} 4.4905E-04} & {\color[HTML]{000000} 4.3691E-04} & {\color[HTML]{000000} 4.4555E-04} & {\color[HTML]{000000} 4.4555E-04} \\ \cline{2-7}
{\color[HTML]{000000} } & {\color[HTML]{000000} $100 \times 100$} & {\color[HTML]{000000} 1.6393E-04} & {\color[HTML]{000000} 1.6817E-04} & {\color[HTML]{000000} 1.6307E-04} & {\color[HTML]{000000} 1.6827E-04} & {\color[HTML]{000000} 1.6827E-04} \\ \cline{2-7}
\multirow{-4}{*}{{\color[HTML]{000000} $L^2$ Error}} & {\color[HTML]{000000} $200 \times 200$} & {\color[HTML]{000000} 7.3160E-05} & {\color[HTML]{000000} 6.9776E-05} & {\color[HTML]{000000} 7.4308E-05} & {\color[HTML]{000000} 6.9736E-05} & {\color[HTML]{000000} 6.9736E-05} \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[H]
\centering
\subfloat[$L^1$ Error]
{\includegraphics[scale=0.22]{Images/Scalar/Tables-and-Plots/OC_L1}} \hspace{0.5cm}
\subfloat[$L^2$ Error]
{\includegraphics[scale=0.22]{Images/Scalar/Tables-and-Plots/OC_L2}}
\caption{Convergence of the $L^1$ and $L^2$ error norms for the
reinitialization of circular disc rotation problem solved on four
($25 \times 25$, $50 \times 50$, $100 \times 100$ and $200 \times
200$) grid levels.}
\label{fig:OC_order_error}
\end{figure}
\subsubsection{Reinitialization of Zalesak's disc rotation problem}
The second problem considered is the advection of Zalesak's disc,
similar to the test reported in \cite{Desjardins2008}. The Zalesak's
disc is of radius 0.15 units, and notch length and width of 0.3 units
and 0.1 units respectively. Due to the presence of sharp corners, this
test problem shows more numerical errors compared to the previous
problem. Similar to the rotation of circular disc problem, this
problem also is solved on four levels of Cartesian meshes starting
from $25 \times 25$ up to $200 \times 200$. The level set contours
correspond to $\psi = 0.05, \psi = 0.5$ and $\psi = 0.95$ for the
$100 \times 100$ case at time levels $t = 0$, $t = \pi/4$, $t = \pi/2$
and $t = 3\pi/4$ for all five reinitialization schemes are plotted in
Figure~\ref{fig:ZD_contours}. Further,
Table~\ref{tab:ZD_area_error}~and~\ref{tab:ZD_l1_l2} show the
percentage area error and error norms computed after the disc
completes one full rotation ($ie.,$ at $t = 2 \pi$). Moreover, the
error norms are plotted against the mesh size in
Figure~\ref{fig:ZD_order_error} along with reference slopes for the
first and second order rate of convergence. Similar to the previous
problem, here also it can be seen that the new reinitialization scheme
is having the least error among all.
\begin{figure}[H]
\centering
\subfloat[CLS-Olsson]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Zalesak-Disc/100/CLS-Olsson.png}}
\subfloat[ACLS-Desjardins]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Zalesak-Disc/100/ACLS-Desjardins.png}}
\subfloat[CLS-Wac{\l}awczyk]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Zalesak-Disc/100/CLS-Waclawszyk.png}}
\subfloat[CLS-Chiodi]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Zalesak-Disc/100/CLS-Chiodi.png}}
\subfloat[New]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Zalesak-Disc/100/New.png}}
\caption{Comparison of reinitialization schemes for the
reinitialization of Zalesak's disc rotation problem solved on $100
\times 100$ grid. Each subfigure shows the 0.05, 0.5 and 0.95 level
contours of the level set functions at time levels of 0.0, $\pi$/4,
$\pi$/2 and $3\pi$/4 (left, top, right and bottom respectively).}
\label{fig:ZD_contours}
\end{figure}
\begin{table}[H]
\begin{center}
\caption{Percentage area errors after completion of one full rotation of different reinitialization schemes for the reinitialization of Zalesak's disc rotation problem.}
\label{tab:ZD_area_error}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
{\color[HTML]{000000} } & \multicolumn{5}{c|}{{\color[HTML]{000000} \text{Reinitialization Schemes}}} \\ \cline{2-6}
\multirow{-2}{*}{{\color[HTML]{000000} \text{Mesh}}} & {\color[HTML]{000000} \text{CLS-Olsson}} & {\color[HTML]{000000} \text{ACLS-Desjardins}} & {\color[HTML]{000000} \text{CLS-Wac{\l}awszyk}} & {\color[HTML]{000000} \text{CLS-Chiodi}} & {\color[HTML]{000000} \text{New}} \\ \hline
{\color[HTML]{000000} $25 \times 25$} & {\color[HTML]{000000} -9.49000} & {\color[HTML]{000000} -9.51580} & {\color[HTML]{000000} -8.99110} & {\color[HTML]{000000} -9.37870} & {\color[HTML]{000000} -7.10470} \\ \hline
{\color[HTML]{000000} $50 \times 50$} & {\color[HTML]{000000} -5.33530} & {\color[HTML]{000000} -5.22570} & {\color[HTML]{000000} -5.19310} & {\color[HTML]{000000} -4.68600} & {\color[HTML]{000000} -3.70040} \\ \hline
{\color[HTML]{000000} $100 \times 100$} & {\color[HTML]{000000} -3.09830} & {\color[HTML]{000000} -3.45670} & {\color[HTML]{000000} -3.29800} & {\color[HTML]{000000} -2.00520} & {\color[HTML]{000000} -1.28230} \\ \hline
{\color[HTML]{000000} $200 \times 200$} & {\color[HTML]{000000} -2.38510} & {\color[HTML]{000000} -2.33920} & {\color[HTML]{000000} -2.38360} & {\color[HTML]{000000} -2.33040} & {\color[HTML]{000000} -0.90177} \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[H]
\begin{center}
\caption{The $L^1$ and $L^2$ error norms of different reinitialization schemes for the reinitialization of Zalesak's disc rotation problem.}
\label{tab:ZD_l1_l2}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
{\color[HTML]{000000} } & {\color[HTML]{000000} } & \multicolumn{5}{c|}{{\color[HTML]{000000} \text{Reinitialization Schemes}}} \\ \cline{3-7}
\multirow{-2}{*}{{\color[HTML]{000000} \text{Errors}}} &
\multirow{-2}{*}{{\color[HTML]{000000} \text{Mesh}}} &
{\color[HTML]{000000} \text{CLS-Olsson}} &
{\color[HTML]{000000} \text{ACLS-Desjardins}} &
{\color[HTML]{000000} \text{CLS-Wac{\l}awszyk}} &
{\color[HTML]{000000} \text{CLS-Chiodi}} &
{\color[HTML]{000000} \text{New}} \\ \hline
\hline
{\color[HTML]{000000} } & {\color[HTML]{000000} $25 \times 25$} & {\color[HTML]{000000} 4.2972E-02} & {\color[HTML]{000000} 3.9579E-02} & {\color[HTML]{000000} 4.1547E-02} & {\color[HTML]{000000} 3.6041E-02} & {\color[HTML]{000000} 1.9959E-02} \\ \cline{2-7}
{\color[HTML]{000000} } & {\color[HTML]{000000} $50 \times 50$} & {\color[HTML]{000000} 1.3590E-02} & {\color[HTML]{000000} 1.3860E-02} & {\color[HTML]{000000} 1.4083E-02} & {\color[HTML]{000000} 1.4145E-02} & {\color[HTML]{000000} 1.2202E-02} \\ \cline{2-7}
{\color[HTML]{000000} } & {\color[HTML]{000000} $100 \times 100$} & {\color[HTML]{000000} 7.4338E-03} & {\color[HTML]{000000} 7.6362E-03} & {\color[HTML]{000000} 7.4889E-03} & {\color[HTML]{000000} 7.8407E-03} & {\color[HTML]{000000} 6.4325E-03} \\ \cline{2-7}
\multirow{-4}{*}{{\color[HTML]{000000} $L^1$ Error}} & {\color[HTML]{000000} $200 \times 200$} & {\color[HTML]{000000} 4.2656E-03} & {\color[HTML]{000000} 4.3188E-03} & {\color[HTML]{000000} 4.2485E-03} & {\color[HTML]{000000} 4.3666E-03} & {\color[HTML]{000000} 3.8105E-03} \\ \hline \hline
{\color[HTML]{000000} } & {\color[HTML]{000000} $25 \times 25$} & {\color[HTML]{000000} 5.5200E-03} & {\color[HTML]{000000} 4.8633E-03} & {\color[HTML]{000000} 4.6080E-03} & {\color[HTML]{000000} 4.6220E-03} & {\color[HTML]{000000} 2.4447E-03} \\ \cline{2-7}
{\color[HTML]{000000} } & {\color[HTML]{000000} $50 \times 50$} & {\color[HTML]{000000} 1.0000E-03} & {\color[HTML]{000000} 1.0103E-03} & {\color[HTML]{000000} 1.0334E-03} & {\color[HTML]{000000} 1.0402E-03} & {\color[HTML]{000000} 1.0036E-03} \\ \cline{2-7}
{\color[HTML]{000000} } & {\color[HTML]{000000} $100 \times 100$} & {\color[HTML]{000000} 3.7050E-04} & {\color[HTML]{000000} 3.8588E-04} & {\color[HTML]{000000} 3.7405E-04} & {\color[HTML]{000000} 3.9550E-04} & {\color[HTML]{000000} 3.8000E-04} \\ \cline{2-7}
\multirow{-4}{*}{{\color[HTML]{000000} $L^2$ Error}} & {\color[HTML]{000000} $200 \times 200$} & {\color[HTML]{000000} 1.0556E-04} & {\color[HTML]{000000} 1.0664E-04} & {\color[HTML]{000000} 1.0473E-04} & {\color[HTML]{000000} 1.0647E-04} & {\color[HTML]{000000} 1.2000E-04} \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[H]
\centering
\subfloat[$L^1$ Error]
{\includegraphics[scale=0.22]{Images/Scalar/Tables-and-Plots/ZD_L1}} \hspace{0.5cm}
\subfloat[$L^2$ Error]
{\includegraphics[scale=0.22]{Images/Scalar/Tables-and-Plots/ZD_L2}}
\caption{Convergence of the $L^1$ and $L^2$ error norms for the
reinitialization of Zalesak's disc rotation problem solved on four
($25 \times 25$, $50 \times 50$, $100 \times 100$ and $200 \times
200$) grid levels.}
\label{fig:ZD_order_error}
\end{figure}
\subsubsection{Reinitialization of circular disc deformation problem}
The last scalar advection test problem considered is the
advection of circular disc subjected to a shear velocity field, $u =
\sin^2(\pi x) \sin(2\pi y)$, $v = -\sin^2(\pi y) \sin(2\pi
x)$. Figure~\ref{fig:SW_contours} shows a qualitative comparison of the
interface contour for all five reinitialization scheme at $t = 4$~s
solved a $200 \times 200$ Cartesian mesh. From
Figure~\ref{fig:SW_contours}, one can clearly see that the tail of the
interface is fully resolved without breaking in case of the new
reinitialization scheme.
\begin{figure}[H]
\centering
\subfloat[CLS-Olsson]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Swirl/200/CLS-Olsson.png}}
\subfloat[ACLS-Desjardins]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Swirl/200/ACLS-Desjardins.png}}
\subfloat[CLS-Wac{\l}awczyk]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Swirl/200/CLS-Waclawszyk.png}}
\subfloat[CLS-Chiodi]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Swirl/200/CLS-Chiodi.png}}
\subfloat[New]
{\includegraphics[scale=1.8]{Images/Scalar/Contours/Swirl/200/New.png}}
\caption{Comparison of reinitialization schemes for the
reinitialization of circular disc deformation problem solved on $200
\times 200$ grid. Each subfigure shows the 0.5 level contours of
the level set function at time levels of 4~s.}
\label{fig:SW_contours}
\end{figure}
\subsection{Broken Dam Problem}
\label{sec:db}
In previous sections, in-place reinitialization problems and scalar advection
based test problems are solved. In order to compare its performance on
more realistic problems, an inviscid broken dam problem, reported in
\cite{Yang2014, Parameswaran2019} is considered next. Here, an
initial water column of height 0.9~m and width 0.45~m is kept at zero
velocity field and subjected to a hydrostatic pressure distribution
inside a computational domain bounded between $0 \leq x \leq 2$~m and
$0 \leq y \leq 1$~m. The density of water and air are taken as
998.2~kg/m$^3$ and 1.2040~kg/m$^3$ respectively. The computational
domain is discretized into $100 \times 50$ Cartesian mesh. The
artificial compressibility parameter, $\beta$, is taken as 10000. All
four boundaries are set to free-slip boundary condition. As time
progresses, due to the presence of gravitational force, the water
column collapses. In order to accurately capture the interface
movement, a small real-time step of 0.005~s is chosen. For stability
reasons, a smaller Courant number of 0.1 is chosen for the computation
of pseudo-time step.
Figure~\ref{fig:DB_Interfaces} shows the snapshots comparing the air-water
interface computed using the CLS-Olsson and the new reinitialization
procedure. One can see that the surge front, in case of the new
reinitialization scheme, touches the top wall, then reaches the left
wall and, finally, falls to the bottom pool of water. Since the
free-slip wall boundary conditions do not offer any frictional
resistance, such a behaviour is expected. However, in case of
CLS-Olsson, the thin surge front is spoiled due to inaccuracies
arising form the reinitialization scheme. One can see that, the surge
front does not even touch the top wall. In order to compare the
numerical results with experimental data reported
in~\cite{Martin1952}, the non-dimensional surge front, $s$, and
non-dimensional water column height, $h$, are plotted with respect to
the corresponding non-dimensional time scales $T_s = t\sqrt{2g/a}$ and
$T_h = t\sqrt{g/a}$ respectively in
Figure~\ref{fig:BD_exp_comparison}. The surge front and water column
heights are non-dimensionalized with respect to their initial
sizes. Looking at Figure~\ref{fig:BD_exp_comparison}, one can see an
excellent match of both the numerical results with the experimental
data reported by Martin and Moyce in~\cite{Martin1952}. Finally,
Figure~\ref{fig:DB_area_error} shows the percentage area errors,
computed using equation~(\ref{eq:area_error}), with respect to time
for both the cases. Clearly, the area loss is high for the CLS-Olsson
case, especially during $t = 1$~s, when the surge front becomes thin.
\begin{figure}[H]
\centering
\subfloat[CLS-Olsson ($t = 0s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/Olsson/0.pdf}} \hspace{2.5cm}
\subfloat[New ($t = 0s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/New/0.pdf}}
\subfloat[CLS-Olsson ($t = 0.5s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/Olsson/200.pdf}} \hspace{2.5cm}
\subfloat[New ($t = 0.5s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/New/200.pdf}}
\subfloat[CLS-Olsson ($t = 0.75s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/Olsson/300.pdf}} \hspace{2.5cm}
\subfloat[New ($t = 0.75s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/New/300.pdf}}
\subfloat[CLS-Olsson ($t = 1.0s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/Olsson/400.pdf}} \hspace{2.5cm}
\subfloat[New ($t = 1.0s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/New/400.pdf}}
\subfloat[CLS-Olsson ($t = 1.5s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/Olsson/600.pdf}} \hspace{2.5cm}
\subfloat[New ($t = 1.5s$)]
{\includegraphics[scale=0.8]{Images/NS-Equn/DB/DB-Inviscid/New/600.pdf}}
\caption{Interface profiles at different time levels, starting from $t
= 0.0$~s up to $t = 1.5$~s, for the broken dam problem. Subfigures
(a), (c), (e), (g) and (i) correspond to the CLS-Olsson scheme
and (b), (d), (f), (h) and (j) correspond to the new
reinitialization scheme.}
\label{fig:DB_Interfaces}
\end{figure}
\begin{figure}[H]
\centering
\subfloat[$s$ versus $T_s$]
{\includegraphics[scale=1.0]{Images/NS-Equn/DB/DB-Inviscid/Plots/s_T.pdf}} \hspace{2.5cm}
\subfloat[$h$ versus $T_h$]
{\includegraphics[scale=1.]{Images/NS-Equn/DB/DB-Inviscid/Plots/h_T.pdf}}
\caption{The non-dimensional surge front and water column height
plotted with respect to the corresponding non-dimensional time
scales for the broken dam problem.}
\label{fig:BD_exp_comparison}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=1.2]{Images/NS-Equn/DB/DB-Inviscid/Plots/area_error.pdf}
\caption{The percentage area error plotted with respect to time for
the broken dam problem.}
\label{fig:DB_area_error}
\end{figure}
\subsection{Rayleigh-Taylor Instability Problem}
\label{sec:rti}
The second test problem considered is a Rayleigh-Taylor Instability
problem similar to the one reported in~\cite{Puckett1997,
Parameswaran2019}. Unlike the previous problem, viscosity plays an
important role here. In this problem, a heavier fluid of density
1.225~kg/m$^3$ is placed on top of a lighter fluid of density
0.1694~kg/m$^3$ inside the computational domain bounded between $0
\leq x \leq 1$ and $0 \leq y \leq 4$. The dynamic viscosity for both
the fluids are taken to be the same, $\mu_1 = \mu_2 = 3.1304952 \times
10^{-3}$~kg/m~s. The two fluids are initially separated by an
interface, defined as, $y = 2.0 + 0.05\cos(2 \pi x)$. The problem is
solved on a Cartesian mesh of $32 \times 128$ finite volume cells. The
top and bottom boundaries are set to no-slip boundary condition,
whereas, the left and right boundaries are set as symmetric boundary
condition. The initial velocity field is set to be zero and the
pressure field is set based on gravity. The artificial compressibility
parameter, $\beta$, is taken as 1000 and the real-time step is taken
as 0.01~s. For stability reasons, the Courant number is chosen as 0.9
for the computation of the pseudo-time step.
As time progress, the top heavy fluid start to penetrate into the
bottom light fluid resulting formation of an inverted mushroom shaped
structure. Snapshots at different time levels during the evolution of
the fluid-fluid interface for both the existing and new
reinitialiation cases are shown in
Figure~\ref{fig:RTI_contours}. Similar to the previous problem, here
also one can see that the new reinitialization scheme is better in
capturing the thin fluid layer originating from the tips of the
inverted mushroom head. Figure~\ref{fig:RTI_area_error} compares the
percentage area errors computed using equation~(\ref{eq:area_error})
in both the cases. One can clearly see a higher area loss for the
CLS-Olsson case during the later stages of interface evolution.
\begin{figure}[H]
\centering
\subfloat[($t = 0.8s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/Olsson/8.pdf}} \hspace{0.8cm}
\subfloat[($t = 0.9s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/Olsson/9.pdf}} \hspace{0.8cm}
\subfloat[($t = 1.0s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/Olsson/10.pdf}} \hspace{0.8cm}
\subfloat[($t = 1.2s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/Olsson/12.pdf}} \hspace{0.8cm}
\subfloat[($t = 1.4s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/Olsson/14.pdf}}
\subfloat[($t = 0.8s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/New/8.pdf}} \hspace{0.8cm}
\subfloat[($t = 0.9s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/New/9.pdf}} \hspace{0.8cm}
\subfloat[($t = 1.0s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/New/10.pdf}} \hspace{0.8cm}
\subfloat[($t = 1.2s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/New/12.pdf}} \hspace{0.8cm}
\subfloat[($t = 1.4s$)]
{\includegraphics[scale=0.6]{Images/NS-Equn/RTI/New/14.pdf}}
\caption{Interface profiles at different time levels, starting from $t
= 0.8$~s up to $t = 1.4$~s, for the Rayleigh-Taylor instability
problem. Subfigures from (a) to (e) correspond to the
CLS-Olsson scheme and from (f) to (j) correspond to the new
reinitialization scheme.}
\label{fig:RTI_contours}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{Images/NS-Equn/RTI/Area_Error/area_error.pdf}
\caption{The percentage area error plotted with respect to time for
the Rayleigh-Taylor instability problem.}
\label{fig:RTI_area_error}
\end{figure}
\subsection{Rising Bubble Problem}
\label{sec:br}
The last problem considered in this section is a rising bubble problem
described in~\cite{Hysing2009}. Unlike previous problems, this one
is more challenging because of the presence of buoyancy, viscosity and
surface tension forces. The problem consists of a circular bubble of
diameter, $d_b = 0.5$~m, placed at the lower half of a rectangular domain
bounded between $0 \leq x \leq 1$~m and $0 \leq y \leq 2$~m, initially
filled with a quiescent liquid. Due to the presence of the buoyant force,
the initial circular bubble will start rising. With the interaction
of the surrounding liquid, the initially circular shape of the bubble
gets deformed. The degree of deformation of the circular bubble
depends upon the Reynolds number ($Re$) and the E\"{o}tv\"{o}s number
($Eo$). The $Re$ and $Eo$ are defined as,
\begin{equation}
Re = \frac{\rho_2 U_g L}{\mu_2}
\end{equation}
and
\begin{equation}
Eo = \frac{\rho_2 U_g^2 L}{\sigma}
\end{equation}
where, the characteristic length scale, $L = d_b$, and the
characteristic velocity scale, $U_g = \sqrt{g d_b}$. Based on the
level of difficulty, two versions of rising bubble problems are
reported in~\cite{Hysing2009}. The first one (denoted here as
``Case-1'') is relatively simple and the second one (denoted here as
``Case-2'') is more challenging. The physical parameters defining the
two dimensional rising bubble test cases are given in
Table~\ref{tab:phy_parameters}.
\begin{table}[H]
\centering
\caption{Physical parameters defining the two dimensional rising bubble test cases.}
\begin{tabular}{p{1.5cm} p{1.2cm} p{1.2cm} p{1.2cm} p{1.2cm} p{1.2cm} p{1.2cm} p{1.2cm} p{1.2cm} p{1.2cm} p{1.2cm}}
\hline
Test case & $\rho_2$ & $\rho_1$ & $\mu_2$ & $\mu_1$ & $g$ & $\sigma$ & $Re$ & $Eo$ &
$\rho_2/\rho_1$ & $\mu_2/\mu_1$ \\ \hline
Case-1 & 1000 & 100 & 10 & 1 & 0.98 & 24.5 & 35 & 10 & 10 & 10 \\
Case-2 & 1000 & 1 & 10 & 0.1 & 0.98 & 1.96 & 35 & 125 & 1000 & 100 \\ \hline
\end{tabular}
\label{tab:phy_parameters}
\end{table}
In both the cases, the initial velocity field is set to zero and the
initial pressure field is set based on gravity. The left and right
boundaries are set to free-slip boundary condition and the top and
bottom walls are set to no-slip boundary condition. The artificial
compressibility parameter, $\beta$, is taken as 10000. The real-time
step is taken as 0.05~s and the pseudo-time step is computed according
to the Courant number 0.9. Numerical simulations are carried out up to
a time level of 4~s. In order to make a quantitative comparison, three
parameters, namely, the rise velocity, the location of centroid and
the circularity of the bubble are reported in~\cite{Hysing2009}. These
parameters are computed as,
\begin{equation}
\text{Rise Velocity, }{v_c} = \frac{\int_{\Omega_b}{\bf u}\cdot{\bf e}_y~\text{d}\Omega_b}{\int_{\Omega_b}\text{d}\Omega_b},
\label{eq:rise_vel}
\end{equation}
\begin{equation}
\text{Centroidal Location, }y_c = \frac{\int_{\Omega_b}{\bf
x}_b\cdot{\bf e}_y~\text{d}\Omega_b}{\int_{\Omega_b}\text{d}\Omega_b}
\label{eq:centroid}
\end{equation}
and
\begin{equation}
\text{Circularity, }\zeta = \frac{P_a}{P_b} = \frac{\text{Perimeter of
area-equivalent bubble}}{\text{Perimeter of the bubble}} = \frac{\pi
d_a}{\int_{\Omega}\lVert \nabla \psi \rVert \text{d}\Omega}
\label{eq:circularity}
\end{equation}
where, $\Omega$ is the computational domain, $\Omega_b$ is the region
occupied by the bubble, ${\bf x}_b$ is the position vector inside the
bubble, ${\bf e}_y$ is the unit vector parallel to the $y-$axis and
$d_a$ is the diameter of a circle with area equal to that of the
bubble with circumference $P_b$.
\subsubsection{Case-1}
\label{sec:br1}
For the choice of physical parameters of Case-1, the bubble does not
undergo large deformation. The initial circular bubble first stretches
in the horizontal direction and, finally, settles down to an
ellipsoidal profile as it reaches its terminal speed. Numerical
simulations are carried out on a Cartesian mesh of $80 \times 160$
finite volume cells. The bubble profiles at different time levels for
both the CLS-Olsson and the new reinitialization cases are shown in
Figure~\ref{fig:BR_1_bubble_profiles}. One can see that the bubble
profiles for the CLS-Olsson and the new reinitialization cases are
quite similar and match very well with the results reported
in~\cite{Hysing2009}. In order to make a close comparison, the
terminal shape of the bubbles in both the cases are plotted in
Figure~\ref{fig:BR_1_bubble_profiles_enl} along with the reference
bubble profile of~\cite{Hysing2009}. One can see from
Figure~\ref{fig:BR_1_bubble_profiles_enl} that the bubble profiles of
both the CLS-Olsson and the new reinitialization schemes match very
well with the reference bubble profile. The rise velocity, centroid
location and the circularity of the rising bubble are plotted with
respect to time in Figure~\ref{fig:BR_1_parameters_plot}. Here also,
both the CLS-Olsson and the new reintialization results match
closely with the reference plots. Finally, the percentage area error,
computed using equation~(\ref{eq:area_error}), is plotted with respect
to time in Figure~\ref{fig:BR_1_area_error}. It can be noticed that
the area error is relatively less for the new reinitialization case
compared to that of CLS-Olsson.
\begin{figure}[H]
\centering
\subfloat[CLS-Olsson]
{\includegraphics[scale=0.65]{Images/NS-Equn/BR/Type-1/Bubble_Profiles/Olsson/bubble_profile.pdf}}
\subfloat[New]
{\includegraphics[scale=0.65]{Images/NS-Equn/BR/Type-1/Bubble_Profiles/New/bubble_profile.pdf}} \hspace{0.8cm}
\caption{Bubble profiles from $t = 0$~s up to $t = 4$~s for the rising bubble problem (Case-1).}
\label{fig:BR_1_bubble_profiles}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.6]{Images/NS-Equn/BR/Type-1/Bubble_Profiles/bubble_plot.pdf}
\caption{Enlarged bubble profiles at $t = 3$~s for the rising bubble problem (Case-1).}
\label{fig:BR_1_bubble_profiles_enl}
\end{figure}
\begin{figure}[H]
\centering
\subfloat[Rise Velocity Vs Time]
{\includegraphics[scale=0.6]{Images/NS-Equn/BR/Type-1/Plots/rise_vel.pdf}}
\subfloat[Centroid Vs Time]
{\includegraphics[scale=0.6]{Images/NS-Equn/BR/Type-1/Plots/centroid.pdf}}
\subfloat[Circularity Vs Time]
{\includegraphics[scale=0.6]{Images/NS-Equn/BR/Type-1/Plots/circularity.pdf}}
\caption{The rise velocity, centroid location and circularity plotted
with respect to time for the rising bubble problem (Case-1).}
\label{fig:BR_1_parameters_plot}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.6]{Images/NS-Equn/BR/Type-1/Plots/area_error.pdf}
\caption{The percentage area error plotted with respect to time for
the rising bubble problem (Case-1).}
\label{fig:BR_1_area_error}
\end{figure}
\subsubsection{Case-2}
\label{sec:br2}
Unlike the previous case, a large density ratio in this case results
the bubble to deform more and acquire a dimple cap profile with thin
elongated filament like structures originating from both sides. Due to
the complex shape, it is relatively difficult to capture the bubble
profile in Case-2 as compared to the Case-1. Numerical simulations are
carried out on a Cartesian mesh of $80 \times 160$ finite volume
cells. Moreover, in order to demonstrate the ability of the new
reinitialization scheme to deal with complex meshes, the problem is
also solved on an unstructured mesh consisting of 23331 finite volume
cells of triangular and quadrilateral shapes. Due to the clustering of
cells in the bubble path, one may expect improved results in case of
the unstructured mesh case. Figure~\ref{fig:BR_2_meshes} shows the
Cartesian mesh and the unstructured meshes considered in this
problem. The snapshots of bubble profiles at different time levels for
all the three cases are shown in
Figure~\ref{fig:BR_2_bubble_profiles}. One can clearly see from
Figure~\ref{fig:BR_2_bubble_profiles} that the elongated filament
structure is not captured very well in case of the CLS-Olsson
case. Whereas, a better profile of the elongated filament structure is
captured in case of the new reinitialization method solved on the
Cartesian mesh. The bubble profiles captured using the new
reinitialization approach solved on unstructured mesh, however, show a
close resemblance with the fine mesh results reported
in~\cite{Hysing2009}. The rise velocity, centroid location and
circularity of the bubble are plotted with respect to time in
Figure~\ref{fig:BR_2_parameter_plots}. One can see that the results
for the new reinitialization scheme on the unstructured mesh show very
good match with the reference results. Whereas, the results in case of
CLS-Olsson, especially the circularity profile, are far away from the
reference solution. Finally, the percentage area errors, computed
using equation~(\ref{eq:area_error}), are plotted in
Figure~\ref{fig:BR_2_area_error}. One can easily see that, the area
error is highest for the CLS-Olsson case. The new reinitialization
scheme solved on Cartesian mesh shows much less percentage area
error. Whereas, the new reinitialization method solved on unstructured
mesh shows the least area error.
\begin{figure}[H]
\centering
\subfloat[Cartesian Mesh]
{\includegraphics[scale=0.65]{Images/NS-Equn/BR/Type-2/Contours/New/0_mesh.pdf}
\subfloat[Unstructured Mesh]
{\includegraphics[scale=0.65]{Images/NS-Equn/BR/Type-2/Contours/New_Unstructured/0_mesh.pdf}
\caption{The different meshes used for solving the rising bubble
problem (Case-2). Subfigures (a) is the $80 \times 160$ Cartesian
mesh and (b) is the unstructured mesh consisting of 23331 finite
volume cells of triangular and quadrilateral shape.}
\label{fig:BR_2_meshes}
\end{figure}
\begin{figure}[H]
\centering
\subfloat[($t = 0.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/Olsson/0.pdf}
\subfloat[($t = 1.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/Olsson/1.pdf}
\subfloat[($t = 2.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/Olsson/2.pdf}
\subfloat[($t = 3.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/Olsson/3.pdf}
\subfloat[($t = 4.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/Olsson/4.pdf}}
\subfloat[($t = 0.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New/0.pdf}
\subfloat[($t = 1.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New/1.pdf}
\subfloat[($t = 2.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New/2.pdf}
\subfloat[($t = 3.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New/3.pdf}
\subfloat[($t = 4.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New/4.pdf}}
\subfloat[($t = 0.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New_Unstructured/0.pdf}
\subfloat[($t = 1.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New_Unstructured/1.pdf}
\subfloat[($t = 2.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New_Unstructured/2.pdf}
\subfloat[($t = 3.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New_Unstructured/3.pdf}
\subfloat[($t = 4.0s$)]
{\includegraphics[scale=0.47]{Images/NS-Equn/BR/Type-2/Contours/New_Unstructured/4.pdf}}
\caption{The bubble profiles at different time levels starting from $t
= 0$~s up to $t = 4$~s for the rising bubble problem (Case-2). The
subfigures from (a) to (e) correspond to the CLS-Olsson case
computed on $80 \times 160$ Cartesian mesh,
from (f) to (j) correspond to the new reinitialization computed on
$80 \times 160$ Cartesian mesh and from (k) to (o) correspond to
the new reinitialization computed on the unstructured mesh.}
\label{fig:BR_2_bubble_profiles}
\end{figure}
\begin{figure}[H]
\centering
\subfloat[Rise Velocity Vs Time]
{\includegraphics[scale=0.6]{Images/NS-Equn/BR/Type-2/Plots/rise_vel.pdf}}
\subfloat[Centroid Vs Time]
{\includegraphics[scale=0.6]{Images/NS-Equn/BR/Type-2/Plots/centroid.pdf}}
\subfloat[Circularity Vs Time]
{\includegraphics[scale=0.6]{Images/NS-Equn/BR/Type-2/Plots/circularity.pdf}}
\caption{The rise velocity, centroid location and circularity plotted
with respect to time for the rising bubble problem (Case-2).}
\label{fig:BR_2_parameter_plots}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.6]{Images/NS-Equn/BR/Type-2/Plots/area_error.pdf}
\caption{The percentage area error plotted with respect to time for
the rising bubble problem (Case-2).}
\label{fig:BR_2_area_error}
\end{figure}
| e44f1c0ce3fad23c492dadf9488553659a180bd9 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\vspace{-8pt}
Over the recent eight years, deep generative models have demonstrated stunning performance in generating photorealistic images, considerably boosted by the revolutionary technique of generative adversarial networks (GANs)~\citep{gan14nips,radford2015dcgan,gulrajani2017wgan,miyato2018spectral,brock2018BigGAN,karras2017ProGAN,karras2019StyleGAN,karras2019StyleGAN2}.
However, with the closing margins between real and fake, a flood of strong concerns arise~\citep{harris2018deepfakes,chesney2019deepfakes,brundage2018malicious}: how if these models are misused to spoof sensors, generate deep fakes, and enable misinformation at scale? Not only human beings have difficulties in distinguishing deep fakes, but dedicated research efforts on deep fake detection~\citep{durall2019unmasking,zhang2019detecting,frank2020dct2d_detect,zhang2020not} and attribution~\citep{marra2019gans,ning2019iccv_gan_detection,wang2020cnn} are also unable to sustain longer against the evolution of generative models. For example, researchers delve into details on how deep fake detection works, and learn to improve generation that better fits the detection criteria~\citep{zhang2020not,margret2020upconvolution}. In principle, any successful detector can play an auxiliary role in augmenting the discriminator in the next iteration of GAN techniques, and consequently results in an even stronger generator.
The dark side of deep generative models delays its industrialization process. For example, when commercializing the GPT-2~\citep{radford2019language} and GPT-3~\citep{brown2020language} models, OpenAI leans conservative to open-source their models but rather only release the black-box APIs\footnote{\url{https://openai.com/blog/openai-api/}}.
They involve expensive human labor in the loop to review user downloads and monitor the usage of their APIs. Yet still, it is a challenging and industry-wide task on how to trace the responsibility of the downstream use cases in an open end.
\begin{wrapfigure}[13]{R}{0.6\textwidth}
\centering
\includegraphics[width=\linewidth]{teaser.pdf}
\vspace{-8pt}
\caption{The diagram of our fingerprinting mechanism for generators. See main text for descriptions.
\label{fig:teaser}
\vspace{-16pt}
\end{wrapfigure}
To pioneer in this task, we propose a model fingerprinting mechanism that enables responsible release and regulation of generative models. In particular, we allow responsible model inventors to fingerprint their generators and disclose their responsibilities. As a result, the generated samples contain fingerprints that can be accurately detected and attributed to their sources. This is achieved by an efficient and scalable ad-hoc generation of a large population of generator instances with distinct fingerprints. See Figure~\ref{fig:teaser} Middle.
Similar in the spirit of the dynamic filter networks~\citep{jia2016dynamic} and style-based generator architectures~\citep{karras2019StyleGAN,karras2019StyleGAN2} where their network filters are not freely learned but conditioned on an input, we learn to parameterize a unique fingerprint into the filters of each generator instance. The core gist is to incorporate a fingerprint auto-encoder into a GAN framework while preserving the original generation performance. See Figure~\ref{fig:teaser} Left. In particular, given a GAN backbone, we use the fingerprint embedding from the encoder to modulate each convolutional filter of the generator (Figure~\ref{fig:modulation}), and try to decode this fingerprint from the generated images. We jointly train the fingerprint auto-encoder and GAN with our fingerprint related losses and the original adversarial loss. See Figure~\ref{fig:pipeline} for the diagram, and~\ref{sec:loss_design} for details.
After training, the responsible model inventor is capable of efficiently fingerprinting and releasing different generator instances to different user downloads, which are equipped with the same generation performance but with different fingerprints. \ning{Each user download corresponds to a unique fingerprint, which is maintained by the inventor's database.} As a result, when misuse of a model happens, the model inventor can use the decoder to detect the fingerprint from the generated images, \ning{match it in the database,} and then trace the responsibility of the user. See Figure~\ref{fig:teaser} Right. Based on this form of responsible disclosure, responsible model inventors, like OpenAI, have a way to mitigate adverse side effects on society when releasing their powerful models, while at the same time should have an automatic way to attribute misuses.
There are several key properties of our mechanism. The \textbf{efficiency} to instantiate a generator is inherently satisfied because, after training, the fingerprint encoding and filter modulation run with little overhead. We evaluate the \textbf{effectiveness} of our fingerprinting and obtain almost perfect fingerprint detection accuracy. We also justify the \textbf{fidelity} with a negligible side effect on the original generation quality. See Section~\ref{sec:effectiveness_fidelity}. Our recommended operation point uses a 128-bit fingerprint (Section~\ref{sec:capacity}) which in principle results in a \textbf{capacity} of more than $10^{38}$ identifiable generator instances. The \textbf{scalability} benefits from the fact that fingerprints are randomly sampled on the fly during training so that fingerprint detection generalizes well for the entire fingerprint space. See Section~\ref{sec:scalability} for validation. In addition, we validate in Section~\ref{sec:secrecy} the \textbf{secrecy} of presence and value of our fingerprints against shadow model attacks. We validate in Section~\ref{sec:robustness_immunizability} the \textbf{robustness} and \textbf{immunizability} against perturbation on generated images.
To target the initial motivation, we move the deep fake detection and attribution solutions from \textit{passive} detectors to \textit{proactive} fingerprinting. We show in Section~\ref{sec:detection_attribution} saturated performance and advantages over two state-of-the-art discriminative methods~\citep{ning2019iccv_gan_detection,wang2020cnn} especially in the open world. This is because, conditioned on user-specific fingerprint inputs, the presence of such fingerprints in generated images guarantees the margin between real and fake, and facilitates the attribution and responsibility tracing of deep fakes to their sources.
Our \textbf{contributions} are in four thrusts: (1) We \revise{enhance} the concept of fingerprinting for generative models that enables a responsible disclosure of state-of-the-art GAN models. (2) We pioneer in the novel direction for efficient and scalable GAN fingerprinting mechanism, i.e., only one generic GAN model is trained while more than $10^{38}$ fingerprinted generator instances can be obtained with little overhead during deployment. (3) We also justify several key properties of our fingerprinting, including effectiveness, fidelity, large capacity, scalability, secrecy, robustness, and immunizability. (4) Finally, for the deep fake detection and attribution tasks, we move the solution from \textit{passive} classifiers to \textit{proactive} fingerprinting, and validate its saturated performance and advantages. It makes our responsible disclosure independent of the evolution of GAN techniques.
\vspace{-8pt}
\section{Related work}
\vspace{-8pt}
\textbf{Deep fake detection and attribution.} These tasks come along with the increasing concerns on deep fake misuse~\citep{harris2018deepfakes,chesney2019deepfakes,brundage2018malicious}. Deep fake detection is a binary classification problem to distinguish fake samples from real ones, while attribution further traces their sources. The findings of visually imperceptible but machine-distinguishable patterns in GAN-generated images make these tasks viable by noise pattern matching~\citep{marra2019gans}, deep classifiers~\citep{afchar2018mesonet,hsu2018learning,ning2019iccv_gan_detection}, or deep Recurrent Neural Networks~\citep{guera2018deepfake}. \citep{zhang2019detecting,durall2019unmasking,margret2020upconvolution,liu2020texture_fake} observe that mismatches between real and fake in frequency domain or in texture representation can facilitate deep fake detection. \revise{\citep{wang2020cnn,girish2021towards} follow up with generalization across different GAN techniques towards open world. Beyond attribution, \citep{albright2019source,asnani2021reverse} even reverse the engineering to predict in the hyper-parameter space of potential generator sources.}
However, \ning{these \textit{passive} detection methods heavily rely on the inherent clues in deep fakes. Therefore, they can barely sustain a long time against the adversarial iterations of GAN techniques.} For example, \citep{margret2020upconvolution} improves generation realism by closing the gap in generated high-frequency components. To handle this situation, artificial fingerprinting is proposed in~\citep{yu2021artificial} to \textit{proactively} embed clues into generative models by rooting fingerprints into training data. This makes deep fake detection independent of GAN evolution. Yet, \ning{as \textit{indirect} fingerprinting}, \citep{yu2021artificial} cannot scale up to a large number of fingerprints because they have to pre-process training data for each individual fingerprint and re-train a generator with each fingerprint. Our method is similar in spirit to~\citep{yu2021artificial}, but possesses fundamental advantages \ning{by \textit{directly} and \textit{efficiently} fingerprinting generative models}: after training one generic fingerprinting model, we can instantiate a large number of generators ad-hoc with different fingerprints.
\textbf{Image steganography and watermarking.} \revise{Steganography targets to manipulate carrier images in a hidden manner such that the communication through the images can only be understood by the sender and the intended recipient~\citep{fridrich2009steganography}. Traditional methods rely on Fourier transform~\citep{cox2002digital,cayre2005watermarking}, JPEG compression\footnote{\url{http://www.outguess.org/}}\footnote{\url{http://steghide.sourceforge.net}}, or least significant bits modification~\citep{pevny2010using,holub2014universal}. Recent works utilize deep neural encoder and decoder to hide information~\citep{baluja2017hiding,tancik2019stegastamp,luo2020distortion}. Watermarking targets to embed ownership information into carrier images such that the owner's identity and authenticity can be verified. It belongs to a form of steganography that sometimes interacts with physical images~\citep{tancik2019stegastamp}. Existing methods rely on log-polar frequency domain~\citep{pereira2000robust,kang2010efficient}, printer-camera transform~\citep{solanki2006print,pramila2018increasing}, or display-camera transform~\citep{yuan2013spatially,fang2018screen}. Recent works also use deep neural networks to detect when an image has been re-imaged~\citep{fan2018rebroadcast,tancik2019stegastamp}. The concept and function of our fingerprinting solution is similar in spirit of watermarking, but differs fundamentally.} In particular, we did not retouch individual images. Rather, our solution is the first to retouch generator parameters so as to encode information into the model.
\textbf{Network watermarking.} Network watermarking techniques~\citep{uchida2017embedding,adi2018turning,zhang2018protecting,chen2019deepmarks,rouhani2019deepsigns,ong2021protecting,yu2021artificial} embed watermarks into network parameters rather than pixels while not deteriorating the original utility. Our solution shares motivations with them but substantially differs in terms of concepts, motivations, and techniques. For concepts, most existing works are applicable to only image classification models, \revise{only~\citep{ong2021protecting,yu2021artificial} work for generative models but suffer from poor efficiency and scalability.} For motivations, the existing works target to fingerprint a single model, while we are motivated by the limitation of~\citep{ong2021protecting,yu2021artificial} to scale up the fingerprinting to as many as $10^{38}$ various generator instances within one-time training. For techniques, most existing works embed fingerprints in the input-output behaviors~\citep{adi2018turning,zhang2018protecting,ong2021protecting}, while \revise{our solution gets rid of such trigger input for improved scalability.}
\vspace{-8pt}
\section{GAN fingerprinting networks}
\vspace{-8pt}
\label{sec:loss_design}
Throughout the paper, we stand for the responsible model inventor’s perspective, which is regarded as the regulation hub of our experiments. None of the encoder, decoder, and training data should be accessible to the public. Only fingerprinted generator instances are released to the open end.
We list symbol notations at the beginning. We use latent code ${\bm{z}}\sim\mathcal{N}(\mathbf{0},\mathbf{I}^{d_z})$ to control generated contents. We set $d_z=512$. We represent fingerprint ${\bm{c}}\sim\text{Ber}(0.5)^{d_c}$ as a sequence of bits. It follows Bernoulli distribution with probability $0.5$. We non-trivially choose the fingerprint length $d_c$ in Section~\ref{sec:capacity}. We denote encoder $E$ mapping ${\bm{c}}$ to its embedding, generator $G$ mapping $({\bm{z}},E({\bm{c}}))$ to the image domain, discriminator $D$ mapping an image ${\bm{x}}\sim p_\text{data}$ to the real/fake classification probability, and decoder $F$ mapping an image to the decoded latent code and fingerprint $(\widehat{{\bm{z}}},\widehat{{\bm{c}}})$. In the following formulations, we denote $G({\bm{z}}, E({\bm{c}}))$ as $G ({\bm{z}},{\bm{c}})$ for brevity.
\vspace{-8pt}
\subsection{Pipeline}
\vspace{-8pt}
We consider three goals in our training. First, we preserve the original functionality of GANs to generate realistic images, as close to real distribution as possible. We use the unsaturated logistic loss as in~\citep{gan14nips,karras2019StyleGAN,karras2019StyleGAN2} for real/fake binary classification:
\begin{equation}
{\mathcal{L}}_\textit{adv} = \underset{{\bm{x}}\sim p_\text{data}}{\mathbb{E}} \log D({\bm{x}})+ \underset{\substack{{\bm{z}}\sim\mathcal{N}(\mathbf{0},\mathbf{I}^{d_z})\\{\bm{c}}\sim\text{Ber}(0.5)^{d_c}}}{\mathbb{E}} \log \Big(1-D\big(G({\bm{z}},{\bm{c}})\big)\Big)
\label{eq:Ladv}
\end{equation}
In addition, similar to~\citep{srivastava2017veegan}, we reconstruct the latent code through the decoder $F$ to augment generation diversity and mitigate the mode collapse issue of GANs~\citep{srivastava2017veegan,li2018implicit,yu2020inclusive}.
\begin{equation}
{\mathcal{L}}_\textit{z}= \underset{\substack{{\bm{z}}\sim\mathcal{N}(\mathbf{0},\mathbf{I}^{d_z})\\{\bm{c}}\sim\text{Ber}(0.5)^{d_c}}}{\mathbb{E}} \sum_{k=1}^{d_z} \Big(z_k - F\big( G({\bm{z}},{\bm{c}}) \big)_k \Big)^2
\label{eq:Lz}
\end{equation}
where we use the first $d_z$ output elements of $F$ that correspond to the decoded latent code.
\begin{figure}
\centering
\subfigure[Pipeline.]{
\includegraphics[width=0.45\columnwidth]{method.pdf}
\label{fig:pipeline}
}
\centering
\subfigure[Modulated convolutional layer.]{
\includegraphics[width=0.45\columnwidth]{modulate.pdf}
\label{fig:modulation}}
\vspace{-8pt}
\caption{The diagrams of our fingerprinting pipeline and the modulated convolutional layer.}
\vspace{-8pt}
\end{figure}
The second goal is to reconstruct the fingerprint so as to allow fingerprint detection.
\begin{equation}
{\mathcal{L}}_\textit{c} = \underset{\substack{{\bm{z}}\sim\mathcal{N}(\mathbf{0},\mathbf{I}^{d_z})\\{\bm{c}}\sim\text{Ber}(0.5)^{d_c}}}{\mathbb{E}} \sum_{k=1}^{d_c} \Big[c_k \log\sigma\Big(F\big(G({\bm{z}}, {\bm{c}})\big)_{d_z+k}\Big) +(1-c_k) \log\Big(1-\sigma\Big(F\big(G({\bm{z}}, {\bm{c}})\big)_{d_z+k}\Big)\Big)\Big]
\label{eq:Lc}
\end{equation}
where we use the last $d_c$ output elements of $F$ as the decoded fingerprint. $\sigma(\cdot)$ denotes the sigmoid function that differentiably clips the output to the range of $[0,1]$. The reconstruction is therefore a combination of cross-entropy binary classification for each bit.
It is worth noting that we use one decoder to decode both the latent code and fingerprint, which benefits explicit \textbf{disentanglement} between their representations as discussed below.
The third goal is to disentangle the representation between latent code and fingerprint. Desirably, latent code should have exclusive control over the generated content. This sticks to the original generation functionality. Therefore, two images with different fingerprints but with identical latent code should have a consistent appearance. We formulate the consistency loss as:
\begin{align}
\label{eq:Lconst}
{\mathcal{L}}_\textit{const} = \underset{\substack{{\bm{z}}\sim\mathcal{N}(\mathbf{0},\mathbf{I}^{d_z})\\{\bm{c}}_1,{\bm{c}}_2\sim\text{Ber}(0.5)^{d_c}}}{\mathbb{E}} \Vert G({\bm{z}},{\bm{c}}_1) -G({\bm{z}},{\bm{c}}_2)\Vert_2^2
\end{align}
The disentangled effect is demonstrated in Figure~\ref{fig:samples} and Appendix Figure~\ref{fig:more_samples}.
Our final training objective is as follows. We optimize it under the adversarial training framework w.r.t. $E$, $G$, $F$, and $D$.
\begin{align}
\min_{E,F,G} \max_D
\lambda_1 {\mathcal{L}}_\textit{adv} + \lambda_2 {\mathcal{L}}_z + \lambda_3 {\mathcal{L}}_c + \lambda_4 {\mathcal{L}}_\textit{const}
\label{eq:final_loss}
\end{align}
where $\lambda_1 = 1.0$, $\lambda_2 = 1.0$, $\lambda_3 = 2.0$, and $\lambda_4 = 2.0$ are hyper-parameters to balance the magnitude of each loss term, \revise{Each loss contributes to a property of our solution. The weight settings are empirically not sensitive within its magnitude level.} See Figure~\ref{fig:pipeline} for the diagram.
\vspace{-8pt}
\subsection{Fingerprint modulation}
\vspace{-8pt}
At the architectural level, it is non-trivial how to embed $E({\bm{c}})$ into $G$. The gist is to embed fingerprints into the generator parameters rather than generator input, so that after training a generic model we can instantiate a large population of generators with different fingerprints. This is critical to make our fingerprinting efficient and scalable, as validated in Section~\ref{sec:scalability}. We then deploy only the fingerprinted generator instances to user downloads, not including the encoder.
We achieve this by modulating convolutional filters in the generator backbone with our fingerprint embedding, similar in spirit of~\citep{karras2019StyleGAN2}. Given a convolutional kernel $\mathbf{W}\in\mathbb{R}^{3 \times 3 \times d_l}$ at layer $l$, we first project the fingerprint embedding $E({\bm{c}})$ through an affine transformation $\phi_l$ such that $\phi_l(E({\bm{c}}))\in\mathbb{R}^{d_l}$. The transformation is implemented as a fully-connect neural layer with learnable parameters. We then scale each channel of $\mathbf{W}$ with the corresponding value in $\phi_l$. In specific,
\begin{equation}
\widetilde{\mathbf{W}}_{i,j,k} = \phi_l\big(E({\bm{c}})\big)_k \cdot \mathbf{W}_{i,j,k},\;\;\forall i,j,k
\label{eq:modulation}
\end{equation}
See Figure~\ref{fig:modulation} for a diagram illustration. We compare to the other fingerprint embedding architectures in Section~\ref{sec:effectiveness_fidelity} and validate the advantages of this one. We conduct modulation for all the convolutional filters at layer $l$ with the same fingerprint embedding. And we investigate in Appendix Section~\ref{sec:ablation} at which layer to modulate we can achieve the optimal performance. A desirable trade-off is to modulate all convolutional layers.
Note that, during training, latent code ${\bm{z}}$ and fingerprint ${\bm{c}}$ are jointly sampled. Yet for deployment, the model inventor first samples a fingerprint ${\bm{c}}_0$, then modulates the generator $G$ with ${\bm{c}}_0$, and then deploys only the modulated generator $G(\cdot,{\bm{c}}_0)$ to a user. For that user there allows only one input, i.e. the latent code, to the modulated generator. Once a misuse happens, the inventor uses the decoder to decode the fingerprint and attribute it to the user, so as to achieve responsible disclosure.
\vspace{-8pt}
\section{Experiments}
\vspace{-8pt}
\textbf{Datasets}. We conduct experiments on CelebA face dataset~\citep{liu2015faceattributes}, LSUN Bedroom and Cat datasets~\citep{yu2015lsun}. \ning{LSUN Cat is the most challenging one reported in StyleGAN2~\citep{karras2019StyleGAN2}}. We train/evaluate on 30k/30k CelebA, 30k/30k LSUN Bedroom at the size of 128$\times$128, \ning{and 50k/50k LSUN Cat at the size of 256$\times$256}.
\textbf{GAN backbone}. We build upon the most recent state-of-the-art StyleGAN2~\citep{karras2019StyleGAN2} config E. \ning{This aligns to the settings in~\citep{yu2021artificial} and facilitates our direct comparisons.} See Appendix for the implementation details.
\vspace{-8pt}
\subsection{Effectiveness and fidelity}
\vspace{-8pt}
\label{sec:effectiveness_fidelity}
\textbf{Evaluation.} The effectiveness indicates that the input fingerprints consistently appear in the generated images and can be accurately detected by the decoder. This is measured by fingerprint detection bitwise accuracy over 30k random samples (with random latent codes and random fingerprint codes). We use 128 bits to represent a fingerprint. This is a non-trivial setting as analyzed in Section~\ref{sec:capacity}.
\ning{In addition, bit matching may happen by chance. Following~\citep{yu2021artificial}, we perform a null hypothesis test to evaluate the chance, the lower the more desirable. Given the number of matching bits $k$ between the decoded fingerprint and its encoded ground truth, the null hypothesis $H_0$ is getting this number of matching bits by chance. It is calculated as $Pr(X>k|H_0) = \sum_{i=k}^{d_c} \binom{d_c}{i} 0.5^{d_c}$, according to the binomial probability distribution with $d_c$ trials, where $d_c$ is the fingerprint bit length. $p$-value should be lower than 0.05 to reject the null hypothesis.}
The fidelity reflects how imperceptibly the original generation is affected by fingerprinting. It also helps avoid one's suspect of the presence of fingerprints which may attract adversarial fingerprint removal. We report Fr\'{e}chet Inception Distance~(FID)~\citep{heusel2017gans} between 30k generated images and 30k real testing images. A lower value indicates the generated images are more realistic.
\textbf{Baselines.} We compare seven baseline methods. The first baseline is the StyleGAN2~\citep{karras2019StyleGAN2} backbone. It provides the upper bound of fidelity while has no fingerprinting functionality.
The second baseline is~\citep{yu2021artificial} which is the other proactive but \textit{indirect} fingerprinting method for GANs. Another two baselines, outguess\footnote{\url{http://www.outguess.org/}} and steghide\footnote{\url{http://steghide.sourceforge.net}}, are similar to \citep{yu2021artificial}. They just replace the deep image fingerprinting auto-encoder in \citep{yu2021artificial} with traditional JPEG-compression-based image watermarking techniques, and still suffer from low efficiency/scalability.
We also compare our mechanism to three architectural variants. The motivation of these variants is to incorporate fingerprints in different manners. Variant I: modulating convolutional filters with only latent code embedding, while instead feeding the fingerprint code through the input of the generator. \ning{This is to test the necessity of fingerprint modulation.} Variant II: modulating filters twice, with latent code embedding and fingerprint code embedding separately. Variant III: modulating filters with the embedding from the concatenation of latent code and fingerprint code.
\begin{table}
\begin{center}
\small
\resizebox{\linewidth}{!}{
\begin{tabular}{l|ccc|ccc|ccc}\toprule
& \multicolumn{3}{c|}{CelebA} & \multicolumn{3}{c|}{LSUN Bedroom} & \multicolumn{3}{c}{LSUN Cat}\\
Method & Bit acc $\Uparrow$ & $p$-value $\Downarrow$ & FID $\Downarrow$ & Bit acc $\Uparrow$ & $p$-value $\Downarrow$ & FID $\Downarrow$ & Bit acc $\Uparrow$ & $p$-value $\Downarrow$ & FID $\Downarrow$\\
\midrule
StyleGAN2 & - & - & 9.37 & - & - & 19.24 & - & - & 31.01\\
outguess & 0.533 & 0.268 & 10.02 & 0.526 & 0.329 & 20.15 & 0.523 & 0.329 & 32.30\\
steghide & 0.535 & 0.268 & 9.48 & 0.530 & 0.268 & 19.77 & 0.541 & 0.213 & 31.67\\
\citep{yu2021artificial} & 0.989 & $<10^{-36}$ & 14.13 & 0.983 & $<10^{-34}$ & 21.31 & 0.990 & $<10^{-36}$ & 32.60\\
\midrule
Ours & 0.991 & $<10^{-36}$ & 11.50 & 0.993 & $<10^{-36}$ & 20.50 & 0.996 & $<10^{-36}$ & 33.94\\
Ours Variant I & 0.999 & $<10^{-38}$ & 12.98 & 0.999 & $<10^{-38}$ & 20.68 & 0.500 & 0.535 & 34.23\\
Ours Variant II & 0.987 & $<10^{-36}$ & 13.86 & 0.927 & $<10^{-25}$ & 21.70 & 0.869 & $<10^{-17}$ & 34.33\\
Ours Variant III & 0.990 & $<10^{-36}$ & 22.59 & 0.896 & $<10^{-21}$ & 64.91 & 0.901 & $<10^{-23}$ & 51.74\\
\bottomrule
\end{tabular}}
\end{center}
\vspace{-8pt}
\caption{Fingerprint detection in bitwise accuracy with $p$-value to accept the null hypothesis test, and generation fidelity in FID. $\Uparrow$/$\Downarrow$ indicates a higher/lower value is more desirable. \revise{The baseline results are directly copied from \citep{yu2021artificial}.}}
\label{tab:effectiveness_fidelity}
\vspace{-8pt}
\end{table}
\begin{wrapfigure}[20]{R}{0.3\textwidth}
\centering
\vspace{-8pt}
\includegraphics[width=0.3\columnwidth]{samples_celeba.png}
\vspace{-8pt}
\captionsetup{width=0.3\textwidth}
\captionof{figure}{Fidelity and disentangled control on CelebA: generated samples from five generator instances. For each row, we use a unique fingerprint to instantiate a generator. For each column, we feed in the same latent code to the generator instances. More samples are in Appendix Figure~\ref{fig:more_samples}.}
\vspace{-8pt}
\label{fig:samples}
\end{wrapfigure}
\textbf{Results.} From Table~\ref{tab:effectiveness_fidelity}, we find that:
(1) The two traditional image watermarking methods, outguess and steghide, fail to deliver fingerprints in generated images, indicated by the random guess ($\sim 0.5$) detection accuracy. We attribute this to the representation gap between deep generative models and shallow watermarking techniques.
(2) On CelebA, all the other methods achieve almost perfect fingerprint detection accuracy with $p$-value close to zero. \ning{This is because CelebA is a landmark-aligned dataset with limited diversity. Fingerprinting synergizes well with generation regardless of model configuration.}
(3) On LSUN Bedroom and Cat, only \citep{yu2021artificial} and our optimal model obtain saturated fingerprint detection accuracy. Ours Variant I, II, and III do not always achieve saturated performance. \ning{Especially Ours Variant I fails on LSUN Cat. We reason that filter modulation is a strong formulation for reconstruction. Modulating fingerprints is necessary for their detection while modulating latent code along with fingerprint code distracts fingerprint reconstruction.}
(4) Our method has comparable performance to~\citep{yu2021artificial}, plus substantial advantages in practice: \ning{during deployment, we can fingerprint a generator instance in \textbf{5 seconds}, in contrast to \citep{yu2021artificial} that has to retrain a generator instance in \textbf{3-5 days}. This is a $\mathbf{50000\times}$ gain of efficiency.}
(5) Our method results in negligible $\leq$2.93 FID degradation. This is a worthy trade-off \revise{to introduce the fingerprinting function}.
(6) We show in Figure~\ref{fig:samples} and Appendix Figure~\ref{fig:more_samples} uncurated generated samples from several generator instances. Image qualities are high. Fingerprints are imperceptible. \ning{Thanks to the consistency loss ${\mathcal{L}}_\textit{const}$ in Eq.~\ref{eq:Lconst}, different generator instances can generate identical images given the same latent code. Their fingerprints are clued only in the non-salient background and distinguishable by our decoder.}
\vspace{-8pt}
\subsection{Capacity}
\vspace{-8pt}
\label{sec:capacity}
The capacity indicates the number of unique fingerprints our mechanism can accommodate without crosstalk between two fingerprints. This is determined by $d_c$, fingerprint bit length, and by our detection accuracy (according to Section~\ref{sec:effectiveness_fidelity}). The choice of fingerprint bit length is however non-trivial. A longer length can accommodate more fingerprints but is more challenging to reconstruct/detect
\begin{wrapfigure}[12]{R}{0.3\textwidth}
\centering
\vspace{-8pt}
\includegraphics[width=0.3\textwidth]{collisions_0911.png}
\vspace{-20pt}
\captionsetup{width=0.3\textwidth}
\captionof{figure}{Capacity: fingerprint detection bitwise accuracy and its bottom line requirement w.r.t. fingerprint bit length on CelebA.
\vspace{-16pt}
\label{fig:capacity}
\end{wrapfigure}
To figure out the optimal fingerprint bit length, we conduct the following experiments. On one hand, given one length, we evaluate our detection accuracy. On the other hand, we estimate the bottom-line requirement for detection accuracy. This is simulated as the maximal percentage of bit overlap among a large bag (1 million) of fingerprint samples. The gap between the detection accuracy and bottom-line requirement should be the larger the better.
In Figure~\ref{fig:capacity}, we vary the fingerprint bit length in the options of $\{32, 64, 128, 256, 512\}$, and plot the bitwise detection accuracy in red and the bottom line requirement in blue. We find:
(1) The bottom line requirement is monotonically decreasing w.r.t. the bit length of fingerprint because a larger bit length leads to less heavy fingerprint overlaps.
(2) The testing accuracy is also monotonically decreasing w.r.t. the bit length of fingerprints. This is due to the challenge of fingerprint reconstruction/detection.
(3) The testing accuracy is empirically decreasing more slowly at the beginning and then faster than its bottom-line requirement. We, therefore, pick the bit length 128 as the optimal choice for the maximal gap. We stick to this for all our experiments.
(4) Considering our detection bitwise accuracy $\geq$0.991 and our fingerprint bit length as 128, we derive in principle our mechanism can hold a large capacity of $2^{128\times0.991}\approx10^{38}$ identifiable fingerprints.
\vspace{-8pt}
\subsection{Scalability}
\vspace{-8pt}
\label{sec:scalability}
\begin{wraptable}[11]{r}{0.45\textwidth}
\vspace{-8pt}
\small
\resizebox{\linewidth}{!}{
\begin{tabular}{lcc}\toprule
Fingerprint set size & Training acc $\Uparrow$ & Testing acc $\Uparrow$\\% & FID \\
\midrule
10 & 1.000 & 0.512\\% & $389.10$\\
100 & 1.000 & 0.537\\% & $13.01$\\
1k & 1.000 & 0.752\\% & $12.56$\\
10k & 0.990 & 0.988\\% & $14.20$\\
100k & 0.983 & 0.981\\% & $17.32$\\
Sampling on the fly & 0.991 & 0.991\\% 11.50\\
\bottomrule
\end{tabular}}
\vspace{-8pt}
\captionof{table}{Scalability: fingerprint detection in bitwise accuracy w.r.t. set size on CelebA. $\Uparrow$ indicates a higher value is more desirable.
\vspace{-16pt}
\label{tab:scalability}
\end{wraptable}
Scalability is one of the advantageous properties of our mechanism: during training, we can efficiently instantiate a large capacity of generators with arbitrary fingerprints on the fly, so that fingerprint detection generalizes well during testing. To validate this, we compare to the baselines where we intentionally downgrade our method with access to only a finite set of fingerprints. These baselines stand for the category of non-scalable fingerprinting methods that have to re-train a generator instance for each fingerprint, e.g.~\citep{yu2021artificial}. We cannot directly compare to~\citep{yu2021artificial} because it is impractical (time-consuming) to instantiate a large number of their generators for analysis. \revise{As a workaround, we trained our detector using $\leq$1k real samples to simulate the non-scalable nature of the baseline.}
From Table~\ref{tab:scalability} we show that fingerprint detection fails to generalize unless we can instantiate generators with 10k or more fingerprint samples. This indicates the necessity to equip GANs with an efficient and scalable fingerprinting mechanism, preferably the one on the fly
\vspace{-8pt}
\subsection{Secrecy}
\vspace{-8pt}
\label{sec:secrecy}
\ning{The presence and value of a fingerprint should not be easily spotted by a third party, otherwise it would be potentially removed. In fact, secrecy of our fingerprints is another advantageous property, because our fingerprint encoder, different from image steganography or watermarking, does not directly retouch generated images. As a result, traditional secrecy attack protocols, e.g. Artificial Training Sets (ATS) attack used in~\citep{lerch2016unsupervised,yu2021artificial}, is \textbf{not} applicable.}
\ning{Instead, we employ the shadow-model-based attack~\citep{salem2020updates} to try detecting the presence and value of a fingerprint from generated images. We assume the attacker can access the model inventor's training data, fingerprint space, and training mechanism. He re-trains his own shadow fingerprint auto-encoder. For the \textbf{fingerprint presence attack}, on CelebA dataset, the attacker trains a \revise{ResNet-18-based~\citep{he2016deep}} binary classifier to distinguish 10k non-fingerprinted images \revise{(5k real plus 5k generated)} against 10k generated images from his fingerprinted generators. \revise{We find near-saturated 0.981 training accuracy.} Then he applies the classifier to 1k inventor's generated images. As a result, we find only \textbf{0.505} \revise{testing} accuracy on the presence of fingerprints, close to random guess. For the \textbf{fingerprint value attack}, on CelebA dataset, the attacker applies his shadow decoder \revise{(0.991 training bitwise accuracy)} to 1k inventor's generated images. As a result, we find only \textbf{0.513} \revise{testing} bitwise accuracy, also close to random guess. We conclude that the mismatch between different versions of fingerprinting systems disables the attacks, which guarantees its secrecy.}
\vspace{-8pt}
\subsection{Robustness and immunizability}
\vspace{-8pt}
\label{sec:robustness_immunizability}
Deep fakes in the open end may undergo post-processing environments and result in quality deterioration. Therefore, robustness against image perturbations is equally important to our mechanism. When it does not hold for some perturbations, our immunizability property compensates for it.
Following the protocol in~\citep{ning2019iccv_gan_detection}, we evaluate the robustness against five types of image perturbation: cropping and resizing, blurring with Gaussian kernel, JPEG compression, additive Gaussian noise, and random combination of them. We consider two versions of our model: the original version and the immunized version. An immunized model indicates that during training we augment generated images with the corresponding perturbation in random strengths before feeding them to the fingerprint decoder.
It is worth noting that none of the encoder, decoder, and training data are accessible to the public. Therefore, the robustness against perturbation has to be experimented with the black-box assumption, as protocoled in~\citep{ning2019iccv_gan_detection}. In other words, white-box perturbations such as adversarial image modifications~\citep{goodfellow2014explaining} and fingerprint overwriting, which requires access to the encoder, decoder, and/or training data, are not applicable in our scenario.
\begin{figure*}[b!]
\center
\vspace{-8pt}
\subfigure[]{
\centering
\includegraphics[width=0.19\linewidth]{robustness_plot_crop_comparisons.png}
\subfigure[]{
\centering
\includegraphics[width=0.19\linewidth]{robustness_plot_blur_comparisons.png}
\subfigure[]{
\centering
\includegraphics[width=0.19\linewidth]{robustness_plot_jpeg_comparisons.png}
\subfigure[]{
\centering
\includegraphics[width=0.19\linewidth]{robustness_plot_noise_comparisons.png}
\subfigure[]{
\centering
\includegraphics[width=0.19\linewidth]{robustness_plot_combo_.png}
\vspace{-8pt}
\caption{Robustness and immunizability: red/blue plots show, on CelebA, the fingerprint detection of our original/immunized model in bitwise accuracy w.r.t. the strength of perturbations. \revise{Green plots show those of~\citep{yu2021artificial} as references. The data points are directly copied from their paper.}
\vspace{-8pt}
\label{fig:robutstness_immunizability}
\end{figure*}
We plot in Figure~\ref{fig:robutstness_immunizability} the \revise{comparisons of fingerprint detection accuracy among our original/immunized models and the models of~\citep{yu2021artificial}} w.r.t. the strength of each perturbation. We find:
(1) For all the perturbations, fingerprint detection accuracy drops monotonically as we increase the strength of perturbation. For some perturbations in red plots, i.e., blurring and JPEG compression, accuracy drops slowly in a reasonably large range. We consider accepting accuracy $\geq$75\%. As a result, the robust working range under blurring is: Gaussian blur kernel size $\sim[0, 7]$; under JPEG compression is: JPEG quality $\sim[80, 100]$. Usually, the images turn not functional with perturbations heavier than this range. We, therefore, validate the robustness against blurring and JPEG compression.
(2) For the other perturbations, although our original model is not robust enough, perturbed augmentation compensates significantly in blue dots. We consider accepting accuracy $\geq$75\%. As a result, the immunized working range under cropping is: cropping size $\sim[60, 128]$; under Gaussian noise is: noise standard deviation $\sim[0.0, 0.4]$; under combined perturbation is: the combination of the original or immunized working ranges aforementioned. We, therefore, validate the immunizability of our model against cropping, Gaussian noise, and the combined perturbation.
\revise{(3) Comparing between~\citep{yu2021artificial} and our models, for blurring, their model in green plot is less robust than our original/immunized models. For the other perturbations, theirs are more robust than our original models but are outperformed by our immunized models. This indicates the importance of immunizability of a fingerprinting solution, which is however lacking in~\citep{yu2021artificial}.}
\vspace{-8pt}
\subsection{Deep fake detection and attribution}
\vspace{-8pt}
\label{sec:detection_attribution}
The effectiveness, robustness, and immunizability in turn benefit our initial motivation: deep fake detection and attribution. The former task is a binary classification problem to distinguish between real and fake. The latter task is to further finely label the source of a generated image.
We move the solution from \textit{passive} classifiers to \textit{proactive} fingerprinting, and merge the two tasks into one with 1+$N$ classes: 1 real-world source and $N$ GAN sources, where $N$ can be extremely large, as large as our capacity $10^{38}$ in Section~\ref{sec:capacity}. Then the tasks are converted to verifying if one decoded fingerprint is in our database or not. This is achieved by comparing the decoded fingerprint to each fingerprint in the database given a threshold of bit overlap. According to our $\geq$0.991 fingerprint detection accuracy, it should be reliable to set the threshold at $128\times0.95\approx121$. Then the attribution is trivial because we can directly look up the generator instance according to the fingerprint. If the fingerprint is not in the database, it should be a random fingerprint decoded from a real image. We use our immunized model against the combined perturbations in Section~\ref{sec:robustness_immunizability}.
\textbf{Baselines.} We compare to two state-of-the-art deep fake classifiers~\citep{ning2019iccv_gan_detection,wang2020cnn} as learning-based baselines \textit{passively} relying on inherent visual clues. Because a learning-based method can only enumerate a finite set of training labels, we consider two scenarios for it: closed world and open world. The difference is whether the testing GAN sources are seen during training or not. This does not matter to our method because ours can work with any $N\leq10^{38}$. For the closed world, we train/evaluate a baseline classifier on 10k/1k images from each of the $N+1$ sources. For the open world, we train $N+1$ 1-vs-the-other binary classifiers, and predict as "the other" label if and only if all the classifiers predict negative results. We test on 1k images from each of the real sources or $N$ unseen GAN sources. \ning{We in addition refer to~\citep{yu2021artificial} in comparisons as the other \textit{proactive} but \textit{indirect} model fingerprinting baseline.}
\begin{wraptable}[9]{R}{0.6\textwidth}
\vspace{-20pt}
\begin{center}
\small
\resizebox{\linewidth}{!}{
\begin{tabular}{l|ccc|ccc}
\toprule
& \multicolumn{3}{c|}{Closed world \#GANs} & \multicolumn{3}{c}{Open world \#GANs} \\
Method & 1 & 10 & 100 & 1 & 10 & 100 \\
\midrule
\citep{ning2019iccv_gan_detection} & 0.997 & 0.998 & 0.955 & 0.893 & 0.102 & N/A \\
\citep{wang2020cnn} & 0.890 & N/A & N/A & 0.883 & N/A & N/A \\
\citep{yu2021artificial} & 1.000 & 1.000 & N/A & 1.000 & 1.000 & N/A \\
Ours & 1.000 & 1.000 & 1.000 & 1.000 & 1.000 & 1.000 \\
\bottomrule
\end{tabular}}
\end{center}
\vspace{-8pt}
\caption{Deep fake detection and attribution accuracy on CelebA. A higher value is more desirable. \revise{The baseline results are directly copied from \citep{yu2021artificial}.}
\label{tab:detection_attribution}
\vspace{-8pt}
\end{wraptable}
\textbf{Results.} From Table~\ref{tab:detection_attribution} we find:
(1) Deep fake detection and attribution based on our fingerprints perform equally perfectly ($\sim100\%$ accuracy) \ning{to most of the baselines in the closed world when the number of GAN sources is not too large. However, when $N=100$, \citep{yu2021artificial} is not applicable due to its limited efficiency and scalability. Neither is \citep{wang2020cnn} due to its binary classification nature.}
(2) Open world is also a trivial scenario to our method but challenges the baseline classifiers~\citep{ning2019iccv_gan_detection,wang2020cnn}. When the number of unseen GAN sources increases to 10, \citep{ning2019iccv_gan_detection} even degenerates close to random guess. This is a common generalization issue of the learning-based method. \ning{\citep{yu2021artificial} is still impractical when $N$ is large.}
(3) Since deep fake detection and attribution is a trivial task to our method, it makes our advantages independent of the evolution of GAN techniques. It benefits model tracking and \revise{pushes forward} the emerging direction of model inventors' responsible disclosure.
\vspace{-8pt}
\section{Conclusion}
\vspace{-8pt}
We achieve responsible disclosure of generative models by a novel fingerprinting mechanism. It allows scalable ad-hoc generation of a large population of models with distinct fingerprints. We further validate its saturated performance in the deep fake detection and attribution tasks. \revise{We appeal to the initiatives of our community to maintain responsible release and regulation of generative models. We hope responsible disclosure would serve as one major foundation for AI security.}
\vspace{-8pt}
\section*{Reproducibility statement}
\vspace{-8pt}
The authors strive to make this work reproducible. The appendix contains plenty of implementation details. The source code and well-trained models are available at \href{https://github.com/ningyu1991/ScalableGANFingerprints}{GitHub}.
\vspace{-8pt}
\section*{Ethics statement}
\vspace{-8pt}
This work does not involve any human subject, nor is our dataset related to any privacy concerns. The authors strictly comply with the ICLR Code of Ethics\footnote{\url{https://iclr.cc/public/CodeOfEthics}}.
\vspace{-8pt}
\section*{Acknowledgement}
\vspace{-8pt}
Ning Yu was partially supported by Twitch Research Fellowship. Vladislav Skripniuk was partially supported by IMPRS scholarship from Max Planck Institute. This work was also supported, in part, by the US Defense Advanced Research Projects Agency (DARPA) Media Forensics (MediFor) Program under FA87501620191 and Semantic Forensics (SemaFor) Program under HR001120C0124. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the DARPA. We acknowledge the Maryland Advanced Research Computing Center for providing computing resources. We thank David Jacobs, Matthias Zwicker, Abhinav Shrivastava, and Yaser Yacoob for constructive advice in general.
| df0e45636e28a908aa199909f79ded2772045b5b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Deep neural networks (DNNs) have achieved an outstanding performance for various learning tasks such as speech recognition, image classification, and visual object recognition etc.\cite{lecun2015deep}.
It is well known that DNNs can apprioximate almost any nonlinear functions and make end-to-end learning possible \cite{mallat2016understanding, RaghuPKGS17}.
Most recently, there has been a surge of interest in graph neural networks (GNNs) since they can capture the dependency of graphs by accounting for the message passing between nodes \cite{zhou2018graph, abs-1901-00596}.
This appealing feature has renewed interest in answering a variety of fundamental questions involving the interpretation, generalization, model selection, and convergence of GNNs (DNNs) \cite{NovakBAPS18}.
When developing innovative GNN techniques, it is imperative to explore the physical and mathematical principles that explain the observed phenomenon, which ultimately provides guidelines for creative designs.
There is a rich literature studying the effectiveness of GNNs from various aspects \cite{RaghuPKGS17, DongSZZ17}. \footnote{This field is called the interpretability or expressivity, and here we use the later terminology to encompass all efforts in this area.}
For example, Scarselli et al. \cite{Scarselli2009Computational} first showed that GNNs can approximate a large class of functions in probability.
Kawamoto et al. \cite{KawamotoTO18} provided a theoretical analysis of GNNs based on mean-field theory for graph partitioning tasks.
Lei et al. \cite{lei2017deriving} designed a recurrent neural architecture inspired by graph kernels and discussed its equivalence between Weisfeiler-Lehman kernels.
Moreover, Xu et al. \cite{xu2018powerful} proved that the expressivity of GNNs was as powerful as that of the Weisfeler-Lehman graph isomorphism test.
These efforts have deepened our understanding of the expressive power of GNNs, however, general guidelines are still largely needed for designing better neural architectures and overcoming issues in the training of neural networks, such as parametric choices (e.g., width and depth), vanishing, and exploding gradient problems.
Based on the previous research regarding the expressivity of DNNs \cite{RaghuPKGS17, zhang2019expressivity}, we understand that DNNs use the spatial space that offers an informational representation and evolve toward a critical state (i.e., critical points $\bm{0}$) as the depth increase, corresponding to increasing informational entropy. The main goal of this paper is to further study the optimal topology design on GNNs based on the criticality theorem. The main contributions of this paper can be summarized as follows:
\begin{itemize}
\item We propose a tree-pipeline training framework involving global skeleton determination (width determination) and local topological link rewiring.
\item The weak improvement of current pruning algorithms is examined, in which a rank-constrained or sparsity-constrained regularization imposed on non-convex optimization will prevent the tendency toward the critical state $\bm{0}$.
\item The modularity (clustering) phenomenon in network topology is utilized for erroneous weight rewiring in the weight matrix, which in turn verify the modularity phenomenon in GNNs.
\end{itemize}
\section{Graph Neural Network and Its Critical Representation}\label{sec:GraphNeuralNetworks}
The goal of this paper is to explore the representation capabilities of GNNs on graphs.
To this end, we consider a vanilla GNN with feedforward dynamics.
Suppose the input graphs are characterized by a vertex set of size $V$ and a $D$-dimensional feature vector with elements $X_{iu} (i\in V, u\in \{1,\dots, D\})$,
then the state matrix $\bm{X} =[X_{iu}]$ is given by
\begin{equation}\label{eq:GNNModel}
X_{i u}^{t+1}=\sum_{j v} \phi\left(A_{i j} X_{j v}^{t}W_{v u}^{t}\right) +b_{i u}^{t},
\end{equation}
where $\phi(\cdot)$ is a non-linear activation function, $\bm{A}=[A_{ij}]$ is the adjacency matrix of network topology, $\bm{W}^{t}=\left[W_{vu}^{t}\right]$ is a linear transformation of feature space, $\bm{b}^{t}=\left[b_{i u}^{t}\right]$ is a bias term, and the layer is indexed by $t \in \{1,\dots, T\} $.
The general idea behind GNNs is that nodes can be recursively aggregated and propagated to the next layer for complex calculations.
In this respect, the network structure of neural networks is typically described as graphs in which nodes act as neurons, and each edge links the output of one neuron to the input of another. Graph matching refers to a computational problem of establishing a one-to-one bijective correspondence between the vertex set of graphs.
Therefore, graph matching between a pair of graphs is analogous to representing graphs using GNN \cite{LiGDVK19}.
In the next, we discuss the dynamics aspects of the graph matching.
Based on the Banach Fixed Point Theorem in dynamics, we know that the unique solution of differential equations in (\ref{eq:GNNModel}) can be obtained through an iterative process
\begin{equation}\label{eq:RecursionDynamics}
{{\bf{X}}^{t + 1}} = \phi(\bm{A}\dots\phi(\bm{A}\phi\left( \bm{A}{{{\bm{X}}^1}{\bm{W}^1}} \right){\bm{W}^2})\dots{\bm{W}^t}).
\end{equation}
To prevent the system from being chaotic, the eigenvalue of hidden states should satisfy $|\lambda_i(\bm{X}^t)|)<1, i\in {1,\dots,\infty}$.
Assume that the weight matrix $\bm{W}^t$ is randomly distributed. Then both forward propagation and backpropagation are the information transfer powered by dynamics from $[\lambda_1(\bm{X}^t),\dots,\lambda_i(\bm{X}^t),\dots]$ toward the critical state, i,e., critical points $\bm{0}$, which has abundant expressivity \cite{zhang2019expressivity}.
The results can be generalized to local topological vector spaces via Schauder fixed point theorem \cite{Bonsall1962}.
The theorem illustrates that there always exists a fixed point if $X$ is a closed convex subset of local topological space $S$ and $f$ is a continuous self-mapping such that $f(X)$ is contained in a compact subset of $X$.
In this respect, training a GNN is to construct an inexact graph matching through convex-relaxation.
In the next sections, we will further demonstrate this point in more detail.
\section{The Training Issues and Global Skeleton in Graph Neural Networks}\label{sec:SkeletonDesign}
From the Schauder fixed point theorem, to reach the critical state, one should construct a convex network structure and an input convex topological space.
First, we examine the topological structure issue.
The current training methods in GNNs are mainly based on backpropagation, including those gradient-based methods.
However, ordinary gradient descent cannot guarantee convergence to the global minimum, since the cost function is always non-convex.
Another impediment to the convex optimization is the presence of saddle points in high dimensional representation.
The current network structure is pre-set before training, and usually over-parameterized, which may generate many saddle points \cite{ChoromanskaHMAL15}.
In addition, (\ref{eq:RecursionDynamics}) suggests that a global minimum in low dimension may attenuate to a saddle point $0$ in a high dimensional setting by layer-wise multiplying $\lambda_i$, the so-called proliferation of saddle points \cite{DauphinPGCGB14}.
Mathematically, to determine whether a solution is a local minimum, a global minimum or a saddle point, one needs to calculate the eigenvalues of its Hessian matrix at any given point. If all the eigenvalues have both positive and negative values, there will also be a zero value, corresponding to a saddle point.
If all the eigenvalues are positive at any point, there exists a global minimum. Although some recent work addressed this issue either by adopting noisy stochastic gradient descent (SGD) or second-order Hessian information (e.g., Adam), they only avoided the local minimums, and the saddle point issue remains unresolved.
In addition to the backpropagation, another popular method for solving non-convex optimization is the alternating direction method (e.g., PARAFAC for matrix/tensor decomposition) \cite{cichocki2016tensor, AghasiANR17}, which is an alternating matrix optimization algorithm that solves optimization problems by breaking down the convex optimization into smaller parts. Taylor et al. \cite{taylor2016training} pointed out that the alternating direction method of multipliers (ADMM) could overcome the gradient vanishing or explosion issue in backpropagation, and could be implemented in parallel and distributed computing environment.
However, the theoretical understanding of the convergence of ADMM remains challenging when the objective function is non-convex, and simulation examples showed that ADMM could achieve high precision very slowly \cite{BoydPCPE11}.
To avoid the saddle points caused by over-parameterization and high-dimensional representation, it is recommended that the network structure should have a pyramid-like shrinkage property \footnote{Some literature calls it as model compression, here we are prone to dimensional contraction to describe the relationship between successive layers.}.
The shrinkage characteristics refer to the situation that the width of the next layer needs to shrink down compared to the current layer.
Specifically, a network structure is typically characterized by the width (i.e., the number of nodes in each layer) and the depth (measured by the number of hidden layers) of GNNs. In theory, the depth relies on the time dependence and period of data itself. Hence there is no definitive way to determine the optimal value for depth given a specific dataset, and this is usually obtained by numerical trials. Therefore we here mainly focus on estimating the width.
In mathematics, given a complete input, network width can be determined by identifying the latent rank of the observable matrix. This field is called low-rank recovery (or low-rank matrix completion).
By imposing a rank constraint at each layer, the network width should show a pyramid-like structure.
We provide more details about low-rank recovery in Appendix, and examine the proposed hypothesis via simulation experiments.
\subsection{What is Wrong with Existing Pruning Algorithms}\label{sec:PruningAlgorithm}
This section discuss the criticality issue by examining the current pruning algorithms.
Initially, Denil et al. showed in Denil et al. \cite{DenilSDRF13} that there was a considerable redundant structure in existing networks.
To reduce the number of parameters and nodes, researchers have developed various network pruning algorithms to eliminate unnecessary connections or neurons without negatively affecting convergence. A typical pruning algorithm has a three-stage pipeline, i.e., 1) training a large, over-parameterized model;
2) pruning the trained over-parameterized model according to specific criteria; 3) fine-tuning the pruned model to regain the optimal accuracy.
The core pruning procedure is divided into three categories: weight pruning, structured pruning, and layer pruning.
Since the layer pruning depends on the matching between the model and actual data, this paper focuses on the first two pruning techniques.
Weight pruning also learns networks by adding sparsity or rank constraints on GNNs, i.e.,
\begin{equation}
\bm{W}=\mathop{\arg\min}_{\bm{W}} (\bm{X}^{t+1}- \phi(\bm{A}\bm{X}\bm{W}^t)+\lVert\bm{W}^t\lVert).
\end{equation}
From the critical analysis, this constraint by imposing regularization will reverse the tendency toward criticality when the network approaches the critical state $\bm{0}$. From a searching perspective, by mixing up the topology search with weight evolution in one model, the resulting algorithm cannot achieve representation with high precision. Liu et al. \cite{LiuSZHD19} also showed in an experimental analysis that current pruning algorithms only gave a comparable or worse performance than training models with randomly initialized weights. They also emphasized that the pruned architecture, rather than ``significant'' weights, was more important in improving convergence, which is consistent with our analysis.
\section{Robust Topological link Rewiring}
As mentioned earlier, the assumption in global network skeleton design is built on a complete observation of the input $\bm{X}$. In real-world scenes, however, graphs often suffer from the missing edge or missing node features, and the inputs are incomplete \cite{DavenportR16}.
Besides, specific-task based backpropagation learn quickly from current inputs and may ``forget" the previous learning experience.
As a result, the potential accumulated erroneous inputs may eventually form an erroneous topology structure \cite{nelwamondo2007missing}.
In such settings, we need to recover a complete and accurate network topology via a robust design.
Therefore, this section introduces a robust topological design for potential erroneous wights.
A classical approach for increasing network robustness is the use of local (geometric) topological structures. An intuitive understanding of the topological robustness is to provide path redundancy between vertices.
When one path fails, communication can continue through other alternative routes.
Besides, the experiments visualizing the hidden states during training also observed a growing modularity or clustering phenomenon \cite{KawamotoTO18,hou2020learning}.
This phenomenon generally appears in the real-world coupled systems consisting of dynamics and local topological structures \cite{li2010global}.
In all, this general phenomenon implies one can rewire the possible erroneous links by exploiting the local topological structures as an informational redundancy for self-checking.
The modularity presented in the network topology can be viewed as a cluster consensus that each cluster consists of multiple interacting intelligent agents,
and training the network topology is a process of building consensus among each cluster.
Most consensus problem would converge to the average (proof is given in the Appendix), that is, the current state of each agent is an average of local objective function
\begin{equation}
\min \frac{1}{n} \sum_{i=1}^{n} f_{i}(x_i)=\sum_{j=1}^n A_{ij}x_j(t), i=1,\dots,n, \quad \bm{x_i} \in \mathcal{X},\\
\end{equation}
where $f_i(\cdot)$ is the loss function corresponding to agent $i$, and $x \in \mathcal{X}$ is an unknown state to be optimized.
Since network topology in GNNs can be viewed as a graph, its convergence can be handled via graph theory.
For weights in GNNs, there are both positive and negative values. For an undirected graph with all positive weights, that belongs to a class of $Z$ matrix admitting many favorable properties, has been widely studied.
For example, the spectrum of the positive weighted graph Laplacian $\mathfrak{S}(\bm{L})$ has the form: $\mathfrak{S}(L)=\left\{0=\lambda_{1} \leq \lambda_{2} \leq \cdots \leq \lambda_{\infty}\right\}$.
The second smallest Laplacian eigenvalue $\lambda_2(\bm{L})$ is considered as a measure of algebraic connectivity on graphs.
For directed graphs, algebraic connectivity also holds (proof is given in the Appendix).
The consensus can be reached when all weights within a connected graph are positive.
In contrast, negative weights indicate an antagonistic or anticorrelated interaction between nodes.
The existence of both positive and negative weights in the neural architecture may lead to network modularity (clustering) \cite{zelazo2014definiteness}.
The consensus of a graph with negative weights relies on the specific algebraic connectivity measure.
On the other hand, one can make graph cuts or graph partitioning in which the link with positively weighted edges is within one module and the negative ones are between modules.
Given the modularity feature exhibited in the evolutionary dynamics, an intuitive idea is to exploit local connectivity as redundant information to fine-tune the local link during training.
Since the original weights are in general randomly generated, and the algebraic connectivity increases monotonously to form clustering,
one can impose the algebraic connectivity based regularization on the loss function after several epochs waiting for the cluster forming \cite{tam2020fiedler}
\begin{equation}\label{eq:coupledConstraint}
\min _{\bm{W}} \mathcal{L}\left(Y, \hat{\bm{f}}_{\bm{W}}(X)\right)+\delta \lambda_{2}(|\bm{L}|),
\end{equation}
where $\bm{L}$ is the Laplacian matrix converted from the weight matrix $\bm{W}$, $\lambda_{2}(|\cdot|)$ is the Fielder value of the graph of each cluster, and $\delta$ is a tuning parameter.
By imposing the regularization term in the loss function, (\ref{eq:coupledConstraint}) becomes less transparent to observe the specific erroneous links.
Meanwhile, the link should be pruned to exert a localized influence, i.e., the regularization imposed on the overall topology may offset the effects of local link rewiring.
To achieve a better interpretation of the results, we choose to use a greedy algorithm to verify our hypothesis,
rewire possible erroneous links and better understand the clustering phenomena in the training procedure.
We discuss the localized link rewiring in GNNs in the next section.
\subsection{Link's Weight Rewiring to Enhance Algebraic Connectivity}\label{subsec:Rewiringedges}
\begin{figure*}[htpb!]
\centering
\includegraphics[width=\textwidth]{SGNNFrameworkNew.pdf}
\caption{A graph neural network architecture design.
The tree-pipeline successively includes global width contraction, weight evolution, and link's weight rewiring.
In the initial dynamics, information transfer is from front to back.
In supervised learning, information transfer is in the opposite direction since it is subject to specific task constraints imposed by outputs,
accompanied by declining transfer capacity in backpropagation caused by numerous local minimums or saddle points.
To accelerate the informational transfer, the global architecture should have a pyramid-like shrinkage shape to prevent the saddle points caused by over-parameterized settings.
After the weight evolution forming the modularity in topological structure, one can use the topological structure as redundant information to rewire possible erroneous topological links.
}
\label{fig:Frameworks}
\end{figure*}
For a disconnected graph, its algebraic connectivity is $0$, and one can increase the algebraic connectivity by rewiring links.
Note that in addition to the adjacency matrix, incidence matrices can also be used to reprensent a gragh
\begin{equation}\label{eq:LaplacianIncidenceMatrix}
\mathbf{L}=\mathbf{H} \mathbf{H}^{T}=\sum_{l=1}^{m} \mathbf{h}_{l} \mathbf{h}_{l}^{T}.
\end{equation}
$\bm{H} = [\bm{h}_1, \dots, \bm{h}_m]\in \mathbb{R}^{n\times m}$ is the node-edge incidence matrix of graph $\mathcal{G}_{\text{sub}}$, and each edge vector $\bm{h}_l$ denotes vertex $V_i$ joining with vertex $V_j$ whose entries are $[h_l]_i = 1, [h_l]_j =-1$ and $0$ elsewhere.
Given an initial graph $\mathcal{G}_0$, the connectivity of weighted Laplacian matrix $\bm{L}_0$ can be increased by adding new edges
\begin{equation}\label{eq:addingEdges}
\bm{L}(x)=\bm{L}_{0}+\sum_{l=1}^{L} \beta_l w_{l} \bm{h}_{l} \bm{h}_{l}^{T},
\end{equation}
where $\beta_l \in \{0, 1\}$ is a boolean variable indicating whether the $l$th edge is selected, and $w_l$ is the weight being added to edge $l$.
If edge $l$ is added to graph $\mathbb{G}$,
the partial derivative of $\lambda_2(\bm{L}(\beta))$ with respect to $\beta_l$ gives the first order approximation of the increase of $\lambda_2(\bm{L}(\beta))$.
According to the algebraic connectivity of directed graphs in Appendix, we have
\begin{equation}\label{eq:derivative}
\frac{\partial}{\partial \beta_l \lambda_{2}(L(\beta))}=\bm{v}^{T} \frac{\partial L(\beta)}{\partial \beta_{l}} \bm{v}.
\end{equation}
Substitute (\ref{eq:addingEdges}) into (\ref{eq:derivative}), we obtain
\begin{equation}\label{eq:ChoosingEdges}
\begin{aligned}
& \frac{\partial}{\partial \beta_{l}} \lambda_{2}(\bm{L}(\beta)) \\
=& \bm{v}^{T} \frac{\partial\left(\bm{L}_{0}+\sum_{l=1}^L\beta_{l} w_{l} \bm{h}_{l} \bm{h}_{l}^{T}\right)}{\partial \beta_{l}} \bm{v} \\
=& \bm{v}^{T}\left(w_{l} \bm{h}_{l} \bm{h}_{l}^{T}\right) \bm{v}=w_{l}\left(\bm{v}^{T} \bm{h}_{l}\right)\left(\bm{h}_{l}^{T} \bm{v}\right) \\
=& w_{l}\left(v_{i}-v_{j}\right)^{2}.
\end{aligned}
\end{equation}
which indicates that the largest connected edge can be found by maximizing $w_l(v_i-v_j)^2$, where $v_i$ and $v_j$ are the $i$th and $j$th items of Fielder vector $\bm{v}$.
Since the algebraic connectivity of a weighted graph can be measured with respect to each edge,
we can first use the graph partitioning for GNN's node classification, then if the nodes in one classification change in the later training,
we can detect them based on the greedy algorithm of algebraic connectivity, and rewire them via a link prediction method.
In real-world scenes, the dynamics and local connectivity also exhibit coupling characteristics,
therefore, one can use a coupling coefficient to measure their relationship.
Based on the above analysis on global skeleton and local link rewiring, we now present the new GNN architecture design in Fig.\ref{fig:Frameworks}.
\section{Experiments}\label{sec:experiments}
This section provides some empirical evaluations for the proposed architecture design via node classification tasks (the datasets and parameters is outlined in the Appendix).
\subsection{Model Contraction}
we Fig. \ref{fig:ActivationCompare} show the model contraction properties, where the results are based on $20$ Monte Carlo experiments.
Subfigure(above) show the network width after automatic pruning.
Here we indeed observe a layer-by-layer shrinkage width, confirming our proposed shrinkage property when the depth increases. These contraction ratios, however, seem to be relatively small.
Subfigure(bottom) compare the learning rates on test datasets, we find that the convergence continues to decrease even adopting a large rate (i.e., 0.2, 0.5),
which illustrates that shrinkage structure can overcome the saddle point problem, and ultimately improves the convergence.
To enhance the interpretability of GNN, Fig. \ref{fig:EvolutionDynamics} demonstrates the evolutionary dynamics with respect to $5$ prominent eigenvalues of hidden states.
We observe the eigenvalues of each layer conversation from descending to ascending during the training procedure, confirming the proposed information transfer in dynamics.
\begin{figure*}
\centering
\begin{subfigure}[b]{0.9\textwidth}
\centering
\includegraphics[width=\textwidth]{NewNetworkSizeAdjustCompare-GCN.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.9\textwidth}
\centering
\includegraphics[width=\textwidth]{LRCompare-GCN.pdf}
\end{subfigure}
\caption[]
{\small Performance analysis of the proposed framework.
(Above) the observed shrinking property of network width after automatical pruning.
(Bottom) the impact of learning rate on convergence.}
\label{fig:ActivationCompare}
\end{figure*}
\begin{figure*}[htpb!]
\centering
\includegraphics[width=\textwidth]{DiagElement-GCN.pdf}\\
\caption{Evolutionary dynamics in training process.
Each figure contains the $5$ prominent eigenvalues of hidden states.
}
\label{fig:EvolutionDynamics}
\end{figure*}
\subsection{Multi-agent Consensus-based Link's Weight Rewiring}\label{subsec:LinkRewiring}
This section investigates the performance of multi-agents consensus-based link's weight rewiring.
Fig. \ref{fig:LinkRewiring} shows the effects of coupling coefficients on the convergence of test error.
We see the link rewiring on Citeseer, Pubmed and CoraFull are clear, while it is not obvious on Cora datasets, since the erroneous weights are not obvious.
Tab. \ref{Tab:AlgoComp} shows the test accuracy of different graph classification methods.
The results show our shrinkage-rewiring structure (SRGCN and SRChebNet) could greatly improve the node classification accuracy after automatic width pruning.
\begin{figure*}[htpb!]
\centering
\includegraphics[width=\textwidth]{WeightRewiringCoefficientCompare-GCN.pdf}\\
\caption{The effect of coupling coefficient on convergence.}
\label{fig:LinkRewiring}
\end{figure*}
\begin{table}
\vspace{0in}
\centering
\begin{tabular}{l||c|c|c|c}
\toprule \toprule
Method&\textbf{Cora} & \textbf{Citeseer} & \textbf{Pubmed} &\textbf{CoraFull}\\
\midrule
DeepWalk &67.2 & 43.2 &65.3 &80.3\\
GAT &56.8 &72.5 &79.0 &82.5\\
ChebNet & 62.1 &69.8 &74.4&81.4\\
GraphSAGE &64.2 &70.6 &70.5&82.2\\
Planetoid &75.7 &64.7 &77.2&80.5\\
GPNN & 68.1 &79.0 &73.6 &80.4\\
MPNN &72.0& 64.0&75.6&79.8\\
GCN &76.5 &81.5 &79.0 &86.6\\
SRGCN (ours)&80.45 &83.43 &80.16&87.13\\
SRChebNet (ours)&79.96 &82.93&80.98 &86.09\\
\hline
\end{tabular}
\centering
\caption{Performance comparison of different graph classification methods.}\label{Tab:AlgoComp}
\end{table}
\section{Related Works}\label{sec:RelatedWorks}
More recently, many innovative GNN frameworks have been developed.
Notable methods include gated GNN \cite{li2016detecting}, GraphSAGE \cite{hamilton2017inductive}, message-passing neural networks \cite{GilmerSRVD17}, and pruning networks \cite{0022KDSG17}.
In terms of architecture design on topological spaces, the most related work to ours is that Li et al. \cite{LiGDVK19}, where the authors established the equivalence between GNN and graph matching, and emphasized modeling in GNN was a convex optimization process.
The major difference is that their work does not provide a specific network skeleton design, while our method provides a shrinkage network skeleton.
Various regularization methods performed by randomly deleting hidden weights or activations are all for forming convex sets \cite{srivastava2014dropout, Rodriguez0CGR17}.
Zhang \cite{ZhangSS19}
showed that the success of several recently proposed architectures (e.g., ResNet, Wide ResNet, Xception, SqueezeNet, and Inception) was mainly related to the fact that multi-branch structures help reduce the non-convex property of network topology.i
Regarding the observed modularity (clustering) features in weight evolution,
several authors suggested defining convolutional neural network or recurrent neural network modules composed of topologically identical or similar blocks to simplify the topology design.
Results illustrated these methods could achieve a large compression ratio in terms of parameters with excellent performance guarantees \cite{zoph2018learning}.
Compared to their works, our method offers a theoretical explanation for the observed modularity phenomenon, and further employ it as an informational redundancy to guarantee local topological accuracy.
\section{Concluding Remarks}\label{sec:conclusions}
This paper presents a three-pipeline training framework based on global criticality and local topological connectivity.
From the critical Theorem on topological spaces, to reach the critical state, input and network structure should match to build a convex matching (optimization).
In specific training, to promote the information transfer under the over-parameterized setting, we propose a layer-wise shrinkage topological structure to prevent the proliferation of saddle points in high dimensional spaces.
In facing actual erroneous inputs, we give a robust topological link rewiring method based on the local connectivity required by cluster consensus, which is similar to the idea of self-supervised learning that applies structural information as redundant information for self-checking.
Our work contributes by shedding light on the success of GNNs from dynamics and topological spaces aspect.
Due to current topological structure constraints, this paper only involves the intra-layer erroneous weight rewiring, the inter-layer link imputation is still unresolved.
Further exploiting the modularity in more general topological architecture and more complex data (e.g., attacked data) is our next concern, which may provide guidelines to approach the critical expressivity.
\bibliographystyle{unsrt}
| 827bde4940d84456fce0a175489d35cce8bb7bb5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{TRNNGCN}
\begin{algorithm}[!h]
\SetAlgoLined
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{Temporal graph $(A_1,A_2,...,A_T)$,\\ Membership matrix of training data $\Theta_T^{train}$}
\Output{Membership matrix estimate $\hat{\Theta}_T$}
$\hat{A}=A_0$, $H_0=I_N$\;
\For{\text{iteration} $i=1,...,I$}{
\For{$ t=2,...,T$ }{
$ \hat{A}_t=(1-\hat{\Theta}_{i-1} \Lambda (\hat{\Theta}_{i-1})^T) \circ \hat{A}_{t-1}+\hat{\Theta}_{i-1} \Lambda (\hat{\Theta}_{i-1})^T \circ \hat{A}_t.$}
$H^{(1)}=\sigma_1(\hat{A}_TH^{(0)}W^{(1)})$\\
$H^{(2)}=\sigma_2(\hat{A}_TH^{(1)}W^{(2)})$\\
CrossEntropyLoss($H_i^{train}$, ${\Theta}_i^{train}$)\\
Backward()\\
$\hat{\Theta}_{i}=\text{Onehot}(\argmax_{1\leq j \leq n}H_{jk}^{(2)})$\\
}
$\hat{\Theta}_T=\text{Onehot}(\argmax_{1\leq j \leq n}H_{jk}^{(2)})$
\caption{TRNNGCN}
\label{algo:TRNNGCN}
\end{algorithm}
\section{Recovery Requirements}
The goal of clustering or community detection is to recover the membership $\Theta$ by observing the graph $G$, up to some level of accuracy. We next define the relative error and rate of recovery.
\begin{definition}
[Relative error of $\hat{\Theta}$]
The relative error of a clustering estimate $\hat{\Theta}$ is
\begin{equation}
E(\hat{\Theta},\Theta)=\frac{1}{n} \min_{\pi\in \mathcal{P}} \|\hat{\Theta} \pi-\Theta\|_0,
\end{equation}
where $n$ denotes the number of nodes in graph $G$, $\mathcal{P}$ is the set of all $K\times K$ permutation matrices and $\|.\|_0$ counts the number of non-zero elements of a matrix.
\end{definition}
\begin{definition}
[Agreement or accuracy of $\hat{\Theta}$]
\begin{equation}
A(\hat{\Theta},\Theta)=1-E(\hat{\Theta},\Theta).
\end{equation}
\end{definition}
\begin{definition}{(Recovery rate of $\hat{\Theta}$)}
A clustering estimate $\hat{\Theta}$ achieves
\begin{itemize}
\item Exact Recovery when
\begin{equation*}
\mathbb{P}\{A(\hat{\Theta},\Theta)=1\}=1-o(1),
\end{equation*}
\item Almost Exact Recovery when
\begin{equation*}
\mathbb{P}\{A(\hat{\Theta},\Theta)=1-o(1)\}=1-o(1),
\end{equation*}
\item Partial Recovery when
\begin{equation*}
\mathbb{P}\{A(\hat{\Theta},\Theta)\geq \alpha\}=1-o(1),\alpha\in(\frac{1}{k},1).
\end{equation*}
\end{itemize}
\end{definition}
\begin{lemma}
\label{lemma:almost}
If
\begin{equation}
\mathbb{E}(A(\hat{\Theta},\Theta))=1-o(1),
\end{equation}where $\mathbb{E}$ denotes the expectation, then $\hat{\Theta}$ achieves almost exact recovery.
\end{lemma}
\begin{lemma}
\label{lemma:partial}
If \begin{equation}
\mathbb{E}(A(\hat{\Theta},\Theta))\geq \alpha-o(1), \alpha\in(\frac{1}{k},1),
\end{equation} then $\hat{\Theta}$ achieves partial recovery.
\end{lemma}
\section{Proof of Proposition 1}
\begin{lemma}
In \cite{abbe2017community}, by genie-aided hypothesis test, if the link probability $\alpha=\mathcal{O}(\frac{\log n}{n})$, at time step $t$, by spectral estimator, we have
$$\mathbb{P}(\hat{\Theta}_{t,i}=\Theta_{t,i})=\left\{
\begin{aligned}
1-o(1) & , & \frac t 2\geq \max {C_i}\\
o(1) & , & \frac t 2 < \max {C_i},
\end{aligned}
\right.$$
where $\Theta^t_i$ denotes the i-th row (membership of node $i$) of membership matrix $\Theta$ at timestep $t$ and $C_i$ denotes the most recent change time of node $i$.
\end{lemma}
\begin{prop}[Partial Recovery of Spectral Clustering]\label{prop:recover_rate_appendix}
When nodes change their cluster membership over time with change probabilities $\mathcal{O}(\frac{\log n}{n})$, Spectral Clustering recovers the true clusters at time $T$ with relative error $\mathcal{O}(\frac{\log n}{n} T)$.
\end{prop}
\begin{proof}
Let $X^c_t$ denote the set of changed nodes from timestep $1$ to timestep $t$, and $X^u_t$ denote the set of the remaining unchanged nodes. $\Theta^c_t$ means the membership matrix of $X^c_t$ and $\Theta^u_t$ is the membership matrix of $X^u_t$ . We have
\begin{equation*}
\begin{aligned}
&\mathbb{E}(A(\hat{\Theta}_t,\Theta_t))\\=&\mathbb{E}(A(\hat{\Theta}^u_{t}\cup \hat{\Theta}^c_{t},\Theta^u_{t}\cup \Theta^c_{t})=1)\\
\leq&\sum_{i\in X^u_{t}} \mathbb{E}(\hat{\Theta}_{t,i}=\Theta_{t,i})+\sum_{j\in X^c_{t}} \mathbb{E}(\hat{\Theta}_{t,j}=\Theta_{t,j})\\
=&\sum_{i\in X^u_{t}} \mathbb{E}(\hat{\Theta}_{t,i}=\Theta_{t,i})+\sum_{s=1}^{t}\sum_{j\in X^c_{s}/X^c_{s-1}} \mathbb{E}(\hat{\Theta}_{t,j}=\Theta_{t,j})\\
=&(1-\mathcal{O}(\frac{\log n}{n}t))\mathbb{P}(\hat{\Theta}_{t,i}=\Theta_{t,i})_{i\in X^u_{t}}\\
&+\mathcal{O}(\frac{\log n}{2n}t)\mathbb{P}(\hat{\Theta}_{t,j},\Theta_{t,j})_{j\in X^c_{t}, \frac{t}{2}\geq \max C_j}\\
&+\mathcal{O}(\frac{\log n}{2n}t)\mathbb{P}(\hat{\Theta}_{t,j},\Theta_{t,j})_{j\in X^c_{t}, \frac{t}{2}< \max C_j}\\
=&(1-\mathcal{O}(\frac{\log n}{n}t))(1-o(1))\\
&+\mathcal{O}(\frac{\log n}{2n}t)(1-o(1))+\mathcal{O}(\frac{\log n}{2n}t)o(1)\\
=& 1-\mathcal{O}(\frac{\log n}{n}t)
\end{aligned}
\end{equation*}
Then the expectation of relative error at timestep $T$ is
\begin{equation}
\mathbb{E}(E(\hat{\Theta}_T,\Theta_T))= \mathcal{O}(\frac{\log n}{n} T).
\end{equation}
When $T=\mathcal{O}(\frac{n}{\log n})$, we have $\mathbb{E}(A(\hat{\Theta}_T,\Theta_T))=1-\mathcal{O}(1)$. By Lemma \ref{lemma:partial}(Technical Appendix), it only solves partial recovery.
\end{proof}
\section{Proof of Proposition 2}
\begin{prop}[Optimal Decay Rate]\label{prop:opt_appendix}
The concentration of each block $k$ is upper-bounded by
\begin{equation}
\label{equ:bound}
\left\|\hat{A}_{t}^{k}-P_{t}^{k}\right\|\lesssim E_1(\beta_{k})+E_2(\beta_{k}),
\end{equation}
where $\beta_k$ denotes the max decay rate of class $k$ and
\begin{equation}
E_1(\beta_{k}) = \sqrt{ n \alpha \beta_{k}},\;
E_2(\beta_{k}) = \alpha \sqrt{\frac{n^2 \varepsilon_k}{\beta_{k}}},
\end{equation}
which is minimized when $\beta_{k}=\sqrt{ n \alpha \varepsilon_k}$.
\end{prop}
\subsection{Decay rule}
In dynamic SBM, we use the following decay rule:
\begin{equation}
\hat{A}_t=(1-\Theta_t \Lambda (\Theta_t)^T)\hat{A}_{t-1}+\Theta_t \Lambda (\Theta_t)^T A_t.
\end{equation}
Let $\hat{A}_{t}^{k}$ denote the block matrix corresponding to cluster $k$, and similarly consider $K$ blocks $P_t^k$ of the connection probability matrix $P_t$. Let $\lambda_k$ denote the $k$-th element in the diagonal of $\Lambda$. We have
\begin{equation}
\hat{A}_t^k=(1-\lambda_k)\hat{A}_{t-1}^k+\lambda_k A_t^k,
\end{equation}
which can also be written as
\begin{equation}
\hat{A}_t^k=\sum_{s=0}^{t}\beta_s^k {A}_{t-s}^k,
\end{equation}
where $\beta_s^k=\lambda_k (1-\lambda_k)^s$ for $s<t$ and $\beta_{t}^k=(1-\lambda)^t$. Then we denote the maximum of $\beta_s^k$ as $\beta_k$ and we have $\beta_k=\lambda_k$.
Similarly, we define
\begin{equation}
\hat{P}_t^k=\sum_{s=0}^{t}\beta_s^k {P}_{t-s}^k.
\end{equation}
\subsection{Error Bound}
The error bound $\left\|\hat{A}_{t}^{k}-P_{t}^{k}\right\|$ can be divided into two terms:
\begin{equation}
\left\|\hat{A}_{t}^{k}-P_{t}^{k}\right\|\leq \left\|\hat{A}_{t}^{k}-\hat{P}_{t}^{k}\right\|+ \left\|\hat{P}_{t}^{k}-P_{t}^{k}\right\|
\end{equation}
\subsubsection{Preliminaries}
The result is valid for any estimator with weights $\beta_s^k\geq 0$ that satisfy the property that there are constants $C_\beta,C'_\beta>0$ such that:
\begin{equation}
\begin{aligned}
&\sum_{s=0}^t \beta_s^k =1,\;\; \beta_s^k \leq \beta_{k},\;\;\sum_{s=0}^t (\beta_s^k)^2\leq C_\beta \beta_{k} \\
&\sum_{s=0}^t \beta_s^k \min(1,\sqrt{s \varepsilon_k})\leq C'_\beta \sqrt{\frac{\varepsilon_k}{\beta_k}}
\end{aligned}
\end{equation}
Our defined decay rate naturally satisfy the first three preliminaries. The last preliminary is satisfied when $t\geq \frac{\min (\log(\varepsilon_k/\beta_k),\log \beta_k)}{2 \log (1-\beta_k)}$
\subsubsection{Bound the first term}
\begin{theorem}
\cite{keriven2020sparse} Let $A_1,...,A_t\in\{0,1\}^{n}$ be $t$ symmetric Bernoulli matrices whose elements $a^{s}_{ij}$ are independent random variables:
\begin{equation}
a^{s}_{ij} \backsim Ber(p^{s}_{ij}), a^{s}_{ji}=a^{s}_{ij}, a^{s}_{ii}=0
\end{equation}
Assume $p^{s}_{ij}\leq \alpha$. Consider non-negative weights $\beta_s$ and $A=\sum_{s=0}^t \beta_s A_{t-s}$ and $P=\mathbb{E}(A)$, there is a universal constant $C$ such that for all $c>0$, we have
\begin{equation}
\begin{aligned}
&\mathbb{P}\left(\left\|A-P\right\|\geq C(1+c)\sqrt{n \alpha \beta_{\max}}\right)\\
&\leq e^{-(\frac{c^2/2}{2C_\beta + 2c/3}-\log14)n}
+e^{-\frac{c^2/2}{2C_\beta + 2c/3}\frac{n \alpha}{\beta_{\max}}+\log n}+n^{-\frac{c}{4}+6}.
\end{aligned}
\end{equation}
\label{theorem:first_term}
\end{theorem}
By applying Theorem \ref{theorem:first_term}, for fixed block $\Theta_0^k,...,\Theta_t^k$, if ${\frac{n \alpha}{\beta_k}}\gtrsim \log n$, then for any $\nu>0$, there is a constant $C_\nu$ such that with probability at least $1-n^\nu$
\begin{equation}
\begin{aligned}
\left\|\hat{A}_{t}^{k}-\hat{P}_{t}^{k}\right\|
&\leq \left\|\hat{A}_{t}^{k}-\mathbb{E}(\hat{A}_{t}^{k})\right\|+\left\|diag(\hat{P}_t^k) \right\|\\
&\leq C_\nu \sqrt{n\alpha \beta_k} + \alpha
\end{aligned}
\end{equation}
In all considered case, $\beta_k \gg \frac{1}{n}$, $\alpha$ is negligible, we have
\begin{equation}
\left\| \hat{A}_t^k - \hat{P}_t^k\right\|\lesssim \sqrt{n \alpha \beta_k}
\end{equation}
\subsubsection{Bound the second term}
Since $\sum_{s=0}^t \beta_s^k=1$, we have
\begin{equation}
\left\| \hat{P}_t^k - P_t^k\right\|\leq \sum_{s=0}^t \beta_s^k \left\|P_{t-s}^k-P_t^k \right\|\leq \sum_{s=0}^t \beta_s^k \left\|P_{t-s}^k-P_t^k \right\|_F,
\end{equation}
where $\|.\|_F$ is the Frobenius norm.
\begin{lemma}
\label{lemma:bound_p}
Consider $P=\Theta B \Theta^T$ and $P'=\Theta' B (\Theta')^T$. Let $S$ denotes the set of nodes that have changed clusters between $\Theta$ and $\Theta'$. Then we have
\begin{equation}
\begin{aligned}
\left\| P-P'\right\|^2_F &=\sum_{i\in S}\sum_j(p_{ij}-p'_{ij})+(p_{ji}-p'_{ji}) \\
&\leq 4 \sum_{i\in S} \sum_j p_{ij}^2 + (p'_{ij})^2 \\
& \leq 8 \alpha^2 |S| n \max_k \sum_l (B)^2_{kl} \\
&\leq 8 \alpha^2 |S| n
\end{aligned}
\end{equation}
\end{lemma}
By using Lemma \ref{lemma:bound_p} and the Lemma \cite{keriven2020sparse}
\begin{equation}
\begin{aligned}
\mathbb{P}( \exists k,\; \left\|P_{t-k}-P_t\right\|^2_F \geq (8+&C) \alpha n^2 \min (1,k\varepsilon) )\\
&\leq e^{-2C^2\varepsilon^2n+\log \frac{1}{\varepsilon}}
\end{aligned}
\end{equation}
we have, with probability at least $1-n^{\nu}$,
\begin{equation}
\left\|P_{t-s}^k-P_t^k \right\|_F\leq 2 \alpha n^2 \min (1,k\varepsilon_k))
\end{equation}
By the preliminary of $\sum_{s=0}^t \beta_s^k \min(1,\sqrt{s \varepsilon_k})$, we obtain the desired bound.
\begin{equation}
\left\| \hat{P}_t^k - P_t^k\right\|\lesssim \alpha \sqrt{\frac{n^2 \varepsilon_k}{\beta_{k}}}
\end{equation}
\section{Proof of Proposition 3}
\begin{prop}[Almost Exact Recovery]\label{prop:bound_appendix}
Let $\lambda_{\max}$ denote the maximum element on the diagonal of $\Lambda$. With probability at least $1-n^{-\nu}$ for any $\nu > 0$, at any time $t$ we have
\begin{equation}
\left\|\hat{A}_t-P_t\right\|\lesssim \sqrt{ n \alpha \lambda_{\max}}
\end{equation}
When $K$ is constant, $\varepsilon_k=\mathcal{O}(\frac{\log n}{n})$ and $\alpha=\mathcal{O}(\frac{\log n}{n})$, the relative error is $\mathcal{O}\left(\frac{1}{n^{\frac{1}{4}}\log n}\right)$, which implies almost exact recovery at time $T$.
\end{prop}
\begin{proof}
By Equation \ref{equ:bound}(Technical Appendix), we have
\begin{equation}
\left\|\hat{A}_{t}^{k}-P_{t}^{k}\right\| \lesssim \sqrt{ n \alpha \lambda_{k}}.
\end{equation}
We define the decay rates as
\begin{equation}
\Lambda_{jk}=\left\{
\begin{aligned}
\min(1,\sqrt{n \alpha \varepsilon_k})& , &j=k\\
1& , &j\neq k.
\end{aligned}
\right.
\end{equation}
We can separate $\hat{A}_t$ into $K^2$ blocks based on their belonging of clusters. Let $\hat{A}_t^d$ denote the matrix include the diagonal block of $\hat{A}_t$ and $\hat{A}_t^n$ denote the matrix include the non-diagonal block of $\hat{A}_t$. The same goes in $\hat{A}_t^d$ and $\hat{P}_t^n$. We have
\begin{equation}
\left\|\hat{A}_t-P_t\right\|\leq \left\|\hat{A}_t^d-P_t^d\right\| + \left\|\hat{A}_t^n-P_t^n\right\|
\end{equation}
For $\left\|\hat{A}_t^n-P_t^n\right\|$, since $\tau \ll 1$, by setting $\Lambda_{jk}=1, (j\neq k)$, spectral norms of the matrix is negligible. Then we have
\begin{equation}
\left\|\hat{A}_t-P_t\right\|\lesssim \left\|\hat{A}_t^d-P_t^d\right\|
\end{equation}
Since $K$ is constant, by the property of spectral norm, we have
\begin{equation}
\left\|\hat{A}_t-P_t\right\|\lesssim \max_k{\left\|\hat{A}_{t}^{k}-P_{t}^{k}\right\|} =\sqrt{ n \alpha \lambda_{\max}}
\end{equation}
When $\varepsilon_k=\mathcal{O}(\frac{\log n}{n})$ and $\alpha=\mathcal{O}(\frac{\log n}{n})$,
\begin{equation}
\begin{aligned}
&E(\hat{\Theta},\Theta)\\
\lesssim &(1+\delta)\frac{n_{\max}' K}{n\alpha^2 n_{\min}^2 \tau^2}\|\hat{A}-P\|^2\\
\lesssim & (1+\delta)\frac{n K}{n\alpha^2 n^2 \tau^2} \sqrt{ n \alpha \lambda_{\max}}\\
= & (1+\delta)\frac{K}{\mathcal{O}(\frac{\log n}{n})^2 n^2 \tau^2} \sqrt{ n \mathcal{O}(\frac{\log n}{n}) \sqrt{n \mathcal{O}(\frac{\log n}{n}) \mathcal{O}(\frac{\log n}{n})}}\\
=&(1+\delta)\frac{K}{\mathcal{O}({\log n})^2 \tau^2} \sqrt{ \mathcal{O}({\log n})^2 \sqrt{ \mathcal{O}(\frac{1}{n}) }}\\
=&(1+\delta)\frac{K}{ \tau^2} \mathcal{O}\left(\frac{1}{n^{\frac{1}{4}} \log n}\right).
\end{aligned}
\end{equation}
The relative error is $\mathcal{O}\left(\frac{1}{n^{\frac{1}{4}}\log n}\right)$ By Lemma \ref{lemma:almost}(Technical Appendix), it implies almost exact recovery at time $T$.
\end{proof}
\section{GCN and Spectral Clustering}
Figure \ref{fig:spec_norm_appendix} shows that Spectral Clustering and GCN have qualitatively similar accuracy on simulated data as we vary the decay rate $\lambda$ (the same $\lambda$ is used for all nodes). As expected from Proposition~\ref{prop:opt_appendix}, the optimal decay rate $\lambda$ increases as we increase the link probability $\alpha$, as does the value of $\lambda$ that minimizes the spectral norm $\|\hat{A} - P\|$. The optimal decay rate for GCN matches the decay rate value that minimizes the spectral norm, while the optimal decay rate for spectral clustering is larger than the one minimizing the spectral norm.
\begin{figure}[ht]
\centering
\includegraphics[width=0.23\textwidth,trim={0.3cm 0 0.8cm 0.3cm}, clip]{figure/alpha_specnorm_acc.eps}
\includegraphics[width=0.23\textwidth,trim={0.3cm 0 0.8cm 0.3cm}, clip]{figure/alpha_specnorm_norm.eps}
\caption{Accuracy and Spectral Norm as we vary $\alpha$. The optimal decay rate $\lambda$ increases with $\alpha$, as in Proposition~\ref{prop:opt_appendix}.}
\label{fig:spec_norm_appendix}
\end{figure}
\section{Experiment result on simulated data}
Figure \ref{fig:bar_simu} shows the accuracy, AUC, and F1 score comparison of all baseline methods on simulated data, averaged over all 50 timesteps. Our TRNNGCN and RNNGCN methods show the best performance across all three metrics.
\begin{figure}[ht]
\centering
\includegraphics[width=0.47\textwidth]{figure/simulate_bar.eps}
\caption{Comparison of methods on simulated data with heterogeneous cluster transition probabilities.}
\label{fig:bar_simu}
\end{figure}
\section{Introduction}
Clustering nodes based on their connections to each other is a common goal of analyzing graphs, with applications ranging from social to biological to logistics networks. Most such clustering approaches assume that the connections (i.e., edges) between nodes, and thus the optimal clusters, do not change over time~\cite{lei2015consistency,qin2013regularized}. In practice, however, many graph structures will evolve over time. Users in social networks, for example, may migrate from one community to another as their interests or employment status changes, forming new connections with other users (i.e., new edges in the graph) and changing the cluster or community to which they belong. Thus, clustering algorithms on such evolving graphs should be able to track changes in cluster membership over time.
A major challenge in tracking cluster membership changes is to carefully handle historical information and assess its value in predicting the current cluster membership. Since clusters will often evolve relatively slowly, an extreme approach that does not consider edges formed in the past risks ignoring useful information about the majority of nodes whose memberships have not changed. On the other hand, making no distinction between historical and more recently formed edges may lead to slow detection of nodes' membership changes, as the historically formed edges would dominate until the nodes have enough time to make connections within their new clusters. Prior works have balanced these effects by introducing a \emph{decay rate}: the weight of each edge is reduced by a constant decay factor in each time step between the connection formation and the time at which cluster membership is estimated. The cluster membership can then be estimated at any given time, by taking the weighted connections as input to a static algorithm like the well-studied spectral clustering~\cite{abbe2017community}.
Accounting for historical node connections with a single decay rate parameter offers the advantage of interpretability: the decay rate quantifies the emphasis put on historically formed edges, which can be tuned for specific datasets. Yet while prior works have examined the optimal decay rate for stylized network models, they use a single decay rate for all edges~\cite{keriven2020sparse}. In practice, the optimal rate will likely vary, e.g., with
higher decay rates for clusters with higher membership turnover where historical information might reflect outdated cluster memberships, making it less useful. Introducing different decay rates for each cluster, on the other hand, raises a new challenge: since we do not know the true cluster memberships for each node, we may use the wrong decay rate if a node is erroneously labeled. Moreover, the optimal decay rates for different clusters will be correlated due to connections between nodes in different clusters that themselves must be optimally weighted, potentially making the decay rates difficult to optimize.
More sophisticated semi-supervised clustering methods combine LSTM (long-short-term memory) or RNN (recurrent neural network) structures with graph convolutional networks (GCNs), producing a neural network that classifies nodes based on their cluster membership labels. This network can be carefully trained to optimize the use of historical edge information, without explicitly specifying different node decay rates~\cite{pareja2020evolvegcn}. However, while such algorithms show impressive empirical performance on large graph datasets, they are generally not easily interpretable.
Our work seeks to connect the theoretical analysis of graph clustering algorithms with the graph neural networks commonly used in practice. Our key insight in doing so is that \emph{prefacing a GCN with a RNN layer} can be interpreted as imposing a decay rate on node connections that depends on each node's current cluster membership, and then approximating spectral clustering on the resulting weighted graph via the GCN. Following this insight, we propose two new \emph{transitional RNN-GCN neural network architectures} (RNNGCN and TRNNGCN) for semi-supervised clustering. We derive the theoretically optimal decay rates for nodes in each cluster under stylized graph models, and show that the weights learned for the RNN layer in TRNNGCN qualitatively match the theoretically optimal ones.
After reviewing related work on theoretical and empirical graph clustering, we make the following specific contributions:
\begin{itemize}
\item A \textbf{theoretical analysis of the optimal decay rates} for spectral clustering algorithms applied to the dynamic stochastic block model, a common model of graph clustering dynamics~\cite{keriven2020sparse}.
\item Two \textbf{new RNN-GCN neural network architectures} that use an interpretable RNN layer to capture the dynamics of evolving graphs and GCN layers to cluster the nodes.
\item Our algorithm can achieve \textbf{almost exact recovery} by including a RNN layer that decays historical edge information. Static methods can only partially recover the true clusters when nodes change their cluster memberships with probability $\mathcal{O}\left(\frac{\log n}{n}\right)$, $n$ being the number of nodes.
\item \textbf{Experimental results} on real and simulated datasets that show our proposed RNN-GCN architectures outperform state-of-the-art graph clustering algorithms.
\end{itemize}
\section{Spectral Clustering with Decay Rates}
We now introduce the Spectral Clustering algorithm and optimize the decay rates to minimize its relative error.
\subsection{Spectral Clustering Algorithm}
Spectral Clustering is a commonly used unsupervised method for graph clustering. The key idea is to apply $K$-means clustering to the $K$-leading left singular vectors of the adjacency matrix $A$~\cite{stella2003multiclass}; we denote the corresponding matrix of singular vectors as $E_K$. We then estimate the membership matrix $\bar{\Theta}$ by solving
\begin{equation}
(\bar{\Theta},\bar{C})\in \argmin_{\Theta\in \{0,1\}^{n\times K}, C\in\mathbb{R}^{K\times K}} \|\Theta C - E_K\|_F^2,
\label{eq:kmeans}
\end{equation}
where $\|.\|_F$ denotes the Frobenius norm. It is well known that finding a global minimizer of Eq.~\eqref{eq:kmeans} is NP-hard. However, efficient algorithms~\cite{kumar2004simple} can find a $(1+\delta)$-approximate solution $(\hat{\Theta},\hat{C})$, i.e., with
$\|\hat{\Theta}\hat{C}-E_K\|_F^2\leq (1+\delta) \|\bar{\Theta}\bar{C}-E_K\|_F^2$.
\subsection{Introducing Decay Rates}
{In the dynamic SBM, the adjacency matrix $A$ includes edges formed from the initial time step $1$ to the current time step $T$. Let $A_t$ denotes the adjacency matrix only including edges formed at time step $t$. We have $A = \sum_{t=1}^T A_t$}.
Spectral clustering performs poorly on the dynamic SBM:
\begin{prop}[Partial Recovery of Spectral Clustering]\label{prop:recover_rate}
When nodes change their cluster membership over time with probabilities $\varepsilon_j = \mathcal{O}(\frac{\log n}{n})$, {by using the adjacency matrix $A$}, Spectral Clustering recovers the true clusters at time $T$ with relative error $E(\hat{\Theta}_T,\Theta_T) = \mathcal{O}(\frac{\log n}{n} T)$.
\end{prop}
To improve the performance, {one can use an exponentially smoothed version $\hat{A}_t$ as input for clustering:}
\begin{equation}
\hat{A}_t=(1-\lambda)\hat{A}_{t-1}+\lambda A_t
\end{equation}
where $\hat{A}_1 = A_1$ and we call $\lambda\in[0,1]$ the \emph{decay rate}~\cite{chi2009evolutionary}. Intuitively, a larger value of $\lambda$ puts less weight on the past information, ``forgetting'' it faster. However, in the dynamic SBM, each cluster $j$ may have a different change probability $\varepsilon_j$, implying that they may benefit from using different decay rates $\lambda$.
We thus introduce a decay matrix $\Lambda\in[0,1]^{K\times K}$ that gives a different decay rate to connections between each pair of clusters:
\begin{equation}
\hat{A}_t=(1-\Theta_t \Lambda (\Theta_t)^T)\odot \hat{A}_{t-1}+\Theta_t \Lambda (\Theta_t)^T \odot A_t.
\end{equation}
\subsection{Bounding the Relative Error}
Our analysis uses \citet{lei2015consistency}'s result that the relative error rate of the Spectral Clustering on the dynamic SBM at each time $t$ is bounded by the concentration of the adjacency matrix around its expectation:
\begin{equation}
E(\hat{\Theta},\Theta)\lesssim (1+\delta)\frac{n_{\max}' K}{n\alpha^2 n_{\min}^2 \tau^2}\|\hat{A}-P\|^2,
\label{eq:errorbound}
\end{equation}
where $n_{\max}'$ and $n_{\min}$ are respectively the second largest and smallest cluster sizes, and $\|.\|$ denotes the spectral norm.
Thus, $E(\hat{\Theta},\Theta)$ is determined by the concentration $\|\hat{A}-P\|$, where {$\hat{A}=\hat{A}_t$} and $P=P_t=\Theta_t B (\Theta_t)^T$ as defined in the SBM model. To bound this concentration, we consider $K$ diagonal blocks of the adjacency matrix $\hat{A}_t$, with each block corresponding to edges between nodes in a single cluster, after re-indexing the nodes as necessary.
Let $\hat{A}_{t}^{k}$ denote the block matrix corresponding to cluster $k$, and similarly consider $K$ blocks $P^t_k$ of the connection probability matrix $P_t$. We can then upper-bound $\left\|\hat{A}_{t}^{k}-P_{t}^{k}\right\|$ in terms of the decay rate:
\begin{prop}[Optimal Decay Rate]\label{prop:opt}
The concentration of each block $k$ is upper-bounded by
\begin{equation}
\left\|\hat{A}_{t}^{k}-P_{t}^{k}\right\|\lesssim E_1(\beta_{k})+E_2(\beta_{k}),
\end{equation}
where $\beta_k$ denotes the maximum decay rate of class $k$ and
\begin{equation}
E_1(\beta_{k}) = \sqrt{ n \alpha \beta_{k}},\;
E_2(\beta_{k}) = \alpha \sqrt{\frac{n^2 \varepsilon_k}{\beta_{k}}},
\end{equation}
which is minimized when $\beta_{k}=\sqrt{ n \alpha \varepsilon_k}$.
\end{prop}
We formally prove this result in our supplementary material. The intuition is that if the change probability $\varepsilon_k$ is larger, we need a higher decay rate to remember less past information.
We thus define the decay rates as
\begin{equation}
\Lambda_{jk}=\left\{
\begin{aligned}
\min(1,\sqrt{n \alpha \varepsilon_k})& , &j=k\\
1& , &j\neq k.
\end{aligned}
\right.
\end{equation}
This decay rate yields almost exact recovery:
\begin{prop}[Almost Exact Recovery]\label{prop:bound}
Let $\lambda_{\max}$ denote the maximum element on the diagonal of $\Lambda$. With probability at least $1-n^{-\nu}$ for any $\nu > 0$, at any time $t$ we have
\begin{equation}
\left\|\hat{A}_t-P_t\right\|\lesssim \sqrt{ n \alpha \lambda_{\max}}
\end{equation}
When $K$ is constant, $\varepsilon_k=\mathcal{O}\left(\frac{\log n}{n}\right)$ and $\alpha=\mathcal{O}\left(\frac{\log n}{n}\right)$, the relative error is $\mathcal{O}\left(\frac{1}{n^{\frac{1}{4}}\log n}\right)$, which implies almost exact recovery at time $T$.
\end{prop}
\subsection{Connection between GCN and Spectral Clustering}
We empirically demonstrate that Proposition~\ref{prop:opt}'s decay rate is optimal by varying the decay rates used in both spectral clustering and the commonly used Graph Convolutional Network (GCN), which is a first-order approximation of spectral convolutions on graphs~\cite{kipf2016semi}. A multi-layer GCN has the layer-wise propagation rule:
\begin{equation}
H^{(l+1)}=\sigma(\widetilde{D}^{-\frac{1}{2}}\widetilde{A}\widetilde{D}^{-\frac{1}{2}}H^{(l)}W^{(l)}),
\end{equation}
where $\widetilde{A}=A+I_N$, $I_N$ is the identity matrix, $\widetilde{D}_{ii}=\sum_j \widetilde{A}_{ij}$ and $W^{(l)}$ is a layer-specific trainable weight matrix. The activation function is $\sigma$, typically ReLU (rectified linear units), with a softmax in the last layer for graph clustering. The node embedding matrix in the $l$-th layer is $H^{(l)}\in \mathbb{R}^{N\times D}$, which contains high-level representations of the graph nodes transformed from the initial features; $H^{(0)}=I_N$.
\begin{figure}[htp]
\centering
\includegraphics[width=0.23\textwidth,trim={0.25cm 0.3cm 1.2cm 0.55cm}, clip]{figure/specnorm_acc.eps}
\includegraphics[width=0.23\textwidth,trim={0.25cm 0.3cm 1.2cm 0.55cm}, clip]{figure/specnorm_norm.eps}
\caption{Accuracy and Spectral Norm as we vary $n$. The optimal decay rate $\lambda$ increases with $n$, as in Proposition~\ref{prop:opt}.}
\label{fig:spec_norm}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.23\textwidth,trim={0cm 0.1cm 0cm 0cm}, clip]{figure/heat_map_spec.eps}
\includegraphics[width=0.23\textwidth,trim={0cm 0.1cm 0cm 0cm}, clip]{figure/heat_map_gcn.eps}
\caption{Accuracy as we vary $\Lambda_{1,1}$ and $\Lambda_{2,2}$. Spectral Clustering(left) and GCN(right) have similar optimal decay matrix. The change probabilities are $\varepsilon_1=0.05, \varepsilon_2=0.1$.}
\label{fig:heat_map}
\end{figure}
Figure \ref{fig:spec_norm} shows that Spectral Clustering and GCN have qualitatively similar accuracy on simulated data as we vary the decay rate $\lambda$ (the same $\lambda$ is used for all nodes). As expected from Eq.~\eqref{eq:errorbound} and Proposition~\ref{prop:opt}, the optimal decay rate $\lambda$ increases as we increase the number of nodes $n$, as does the value of $\lambda$ that minimizes the spectral norm $\|\hat{A} - P\|$. Although the optimal decay rate is consistently larger than the one minimizing the spectral norm (which upper-bounds the relative error as in Eq.~\eqref{eq:errorbound}), GCN's accuracy is more correlated with the spectral norm, which is the first singular value of the smoothed adjacency matrix.
We then perform a grid search for the optimal decay matrix $\Lambda$ on simulated data with $n = 200$ nodes and change probabilities $\varepsilon_1 = 0.05$ and $\varepsilon_2 = 0.1$. As shown in Figure \ref{fig:heat_map}, GCN and Spectral Clustering achieve high accuracy. Cluster 2, which has a higher $\varepsilon_2$, has larger decay rate $\Lambda_{2,2}$, as expected from Proposition~\ref{prop:opt}, for both GCN and Spectral Clustering.
\section{Decay Rates as RNNs}
Although searching for the optimal decay matrix in Spectral Clustering can result in good performance, this method is expensive: the grid search for the optimal decay matrix can be time-consuming, and the time complexity of calculating the spectral norm is $\mathcal{O}(n^3)$.
In this section, we propose two neural network architectures, RNNGCN and TRNNGCN, that use a single decay rate $\lambda$ and decay matrix $\Lambda$, respectively, and then show they perform well on simulated data.
\subsubsection{RNNGCN}
\begin{algorithm}[htp]
\SetAlgoLined
\SetAlgoLined
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{Temporal graph $(A_1,A_2,...,A_T)$,\\ Membership matrix of training data $\Theta^{train}_T$}
\Output{Membership matrix estimate $\hat{\Theta}_T$}
$\hat{A}=A_0$, $H_0=I_N$\;
\For{\text{iteration} $i=1,...,I$}{
\For{$ t=2,...,T$ }{$\hat{A}_t=(1-\lambda)\hat{A}_{t-1}+\lambda {A}_t$}
$H^{(1)}=\sigma_1(\hat{A}_TH^{(0)}W^{(1)})$\\
$H^{(2)}=\sigma_2(\hat{A}_TH^{(1)}W^{(2)})$\\
CrossEntropyLoss($H^{train}$, ${\Theta}_T^{train}$)\\
Backward()
}
$\hat{\Theta}_T=\text{Onehot}(\argmax_{1\leq j \leq n}H_{jk}^{(2)})$
\caption{RNNGCN}
\label{algo:RNNGCN}
\end{algorithm}
The RNNGCN model uses a single decay rate $\lambda\in[0,1]$ as the RNN parameter. RNNGCN first uses a Recurrent Neural Network to learn the decay rate, then uses a two-layer GCN to cluster the weighted graphs.
The formal model is shown in Algorithm \ref{algo:RNNGCN}, where $\sigma_1$ denotes a ReLU layer and $\sigma_2$ is a Softmax layer.
\subsubsection{Transitional RNNGCN (TRNNGCN)}
The TRNNGCN network is similar to the RNNGCN, but uses a matrix $\Lambda\in[0,1]^{K\times K}$ to learn the decay rates for different pairs of classes. During the training process, the labels (cluster memberships) of the training nodes are known while the labels of other nodes remain unknown, so we use the cluster prediction $\hat{\Theta}_{i-1}$ from each iteration $i-1$ to determine the decay rates for each node in iteration $i$.
The TRNNGCN model replaces the decay method (line 4 of Algorithm 1) with
\begin{equation*}
\hat{A}_t=(1-\hat{\Theta}_{i-1} \Lambda (\hat{\Theta}_{i-1})^T) \odot \hat{A}_{t-1}+\hat{\Theta}_{i-1} \Lambda (\hat{\Theta}_{i-1})^T \odot \hat{A}_t,
\end{equation*}
where $\odot$ denotes element-wise multiplication. After each iteration, it calculates $\hat{\Theta}_i$ as the input of the next iteration.
\subsubsection{Empirical Validation}
We validate the performance of RNNGCN and TRNNGCN on data generated by the dynamic stochastic block model. Our graph has 200 nodes, 23190 edges, 50 time steps and 2 clusters. The probabilities of forming an edge between two nodes of the same or different clusters are $\alpha = 0.02$ and $\tau\alpha = 0.001$, respectively, and a node changes its cluster membership with probability $\varepsilon_1 = 0.05$ and $\varepsilon_2 = 0.1$ for clusters 1 and 2 respectively.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.27\textwidth,trim={0.3cm 0 0.8cm 0.3cm}, clip]{figure/simulate_ACC.eps}
\includegraphics[width=0.27\textwidth,trim={0.3cm 0 0.8cm 0.3cm}, clip]{figure/simulate_AUC.eps}
\includegraphics[width=0.27\textwidth,trim={0.3cm 0 0.8cm 0.3cm}, clip]{figure/simulate_F1.eps}
\caption{On simulated data with heterogeneous cluster transition probabilities, TRNNGCN and RNNGCN, which use optimized decay rates to account for historical information, outperform the static GCN and Spectral Clustering methods. TRNNGCN slightly outperforms RNNGCN.}
\label{fig:simulate}
\end{figure*}
Figure \ref{fig:simulate} compares the RNNGCN and TRNNGCN performance with the static Spectral Clustering and GCN methods. For better visualization, the value at each time step is averaged with the 2 timesteps immediately before and after. The performance of GCN and Spectral Clustering decreases over time, as in later timesteps they use accumulated historical information that may no longer be relevant. RNNGCN and TRNNGCN show consistently high performance over time, indicating that they optimally utilize historical information.
On average over time, TRNNGCN leads to 5\% accuracy and AUC (area under the ROC curve) improvement, and a 10\% higher F1-score, than RNNGCN, due to using a lower decay rate for the class with smaller change probability.
\section{Experiments}
In this section, we validate the performance of RNNGCN and TRNNGCN on real datasets, compared to state-of-the-art baselines. We first describe the datasets used and the baselines considered, and then present our results.
\subsection{Datasets}
We conducted experiments on five real datasets, as shown in Table \ref{tab:datasets}, which have the properties shown in Table \ref{tab:dataattr}. All datasets have edges that form at different times, although only nodes in DBLP-E change their class (cluster membership) over time.
We include four datasets with separate, time-varying features associated with each node (DBLP-3, DBLP-5, Brain and Reddit) to test RNNGCN's and TRNNGCN's ability to generalize to datasets with node features.
\begin{table}[h]
\centering
\begin{tabular}{ccccc}
\hline
Dataset& Nodes &Edges & Time Steps &Classes \\
\hline
DBLP-E & 6942 & 327392 & 14 &2 \\
DBLP-3 & 4257 & 23540 & 10 &3\\
DBLP-5 & 6606 & 42815 & 10 & 5\\
Brain & 5000 & 1955488 & 12 &10\\
Reddit & 8291 & 264050 & 10 &4\\
\hline
\end{tabular}
\caption{Real datasets used to evaluate our methods.}
\label{tab:datasets}
\end{table}
\begin{table}[h]
\centering
\begin{tabular}{ccccc}
\hline
Dataset& Dynamic Edge & Dynamic Class & Features\\
\hline
DBLP-E & $\surd$ & $\surd$ &$\times$\\
DBLP-3 & $\surd$ & $\times$ &100\\
DBLP-5 & $\surd$ & $\times$ & 100\\
Brain & $\surd$ & $\times$ &20\\
Reddit & $\surd$ & $\times$ &20\\
\hline
\end{tabular}
\caption{Properties of datasets in Table~\ref{tab:datasets}.}
\label{tab:dataattr}
\end{table}
\subsubsection{DBLP-E} dataset is extracted from the computer science bibliography website DBLP\footnote{https://dblp.org}, which provides open bibliographic information on major computer science journals and conferences. Nodes represent authors, and edges represent co-authorship from 2004 to 2018. Each year is equivalent to one timestep, and co-author edges are added in the year a coauthored paper is published. Labels represent the author research area (``computer networks'' or ``machine learning'') and may change as authors switch their research focus.
\subsubsection{DBLP-3 \& DBLP-5} use the same node and edge definitions as DBLP-E, but also include node features extracted by \texttt{word2vec}~\cite{mikolov2013efficient} from the authors' paper titles and abstracts. The authors in DBLP-3 and DBLP-5 are clustered into three and five classes (research areas) respectively that do not change over time.
\subsubsection{Reddit} dataset is generated from Reddit\footnote{https://www.reddit.com/}, a social news aggregation and discussion website. The nodes represent posts and two posts are connected if they share keywords. Node features are generated by \texttt{word2vec} on the post comments~\cite{hamilton2017inductive}.
\subsubsection{Brain} dataset is generated from functional magnetic resonance imaging (fMRI) data\footnote{https://tinyurl.com/y4hhw8ro}. Nodes represent cubes of brain tissue, and two nodes are connected if they show similar degrees of activation during the time period. Node features are generated by principal component analysis on the fMRI.
\subsection{Baselines and Metrics}
We compare our RNNGCN and TRNNGCN with multiple baselines. GCN, GAT~\cite{velivckovic2017graph} and GraphSage~\cite{hamilton2017inductive} are supervised methods that include node features, while Spectral Clustering is unsupervised without features; all of these methods ignore temporal information. DynAERNN~\cite{goyal2020dyngraph2vec} is an unsupervised method, and GCNLSTM~\cite{chen2018gc} and EGCN~\cite{pareja2020evolvegcn} are supervised methods, which all utilize temporal information of both graphs and features. We evaluate the performance of methods with the standard accuracy (ACC), area under the ROC curve (AUC) and F1-score classification metrics.
\subsection{Experiment Settings}
We divide each dataset into 70\% training/ 20\% validation/ 10\% test points. Each method uses two hidden Graph Neural Network layers (GCN, GAT, GraphSage, etc.) with the layer size equal to the number of classes in the dataset. We add a dropout layer between the two layers with dropout rate $0.5$. We use the Adam optimizer with learning rate $0.0025$. Each method is trained with $500$ iterations.
For static methods (GCN, GAT, GraphSage and Spectral Clustering) we first accumulate the adjacency matrices of graphs at each time step, then cluster on the normalized accumulated matrix. DynAERNN, GCNLSTM, and EGCN use the temporal graphs and temporal node features as input. For our RNNGCN and TRNNGCN, we use the temporal graphs and the node features at the last time step as input. The code of all methods and datasets are publicly available\footnote{https://github.com/InterpretableClustering/InterpretableClustering}.
\subsection{Experimental Results}
\begin{figure}[ht]
\centering
\includegraphics[width=0.44\textwidth]{figure/dblp_double.eps}
\caption{The two classes in DBLP-E exhibit different change probabilities, with computer network authors more likely to change their labels to machine learning. This trend accelerates after 2014.}
\label{fig:dblpe_num}
\end{figure}
\subsubsection{Node Classification with Temporal Labels}
We first compare the predictions of temporally changing labels in DBLP-E. Figure \ref{fig:dblpe_num} shows the number of authors in the Machine Learning and Computer Network fields in the years 2004-2018, as well as the probabilities that authors in each class change their labels. We observe that (i) the classes have different change probabilities (with users more likely to move from computer networks to machine learning) and (ii) the change probabilities evolve over time, with more users migrating to machine learning since 2013. This dataset thus allows us to test RNNGCN's and TRNNGCN's abilities to adapt the optimal decay rate for each class.
\begin{figure*}[ht]
\centering
\includegraphics[width=0.23\textwidth,trim={0.25cm 0cm 0.8cm 0.5cm}, clip]{figure/dblpe_ACC.eps}
\includegraphics[width=0.23\textwidth,trim={0.25cm 0cm 0.8cm 0.5cm}, clip]{figure/dblpe_AUC.eps}
\includegraphics[width=0.23\textwidth,trim={0.25cm 0cm 0.8cm 0.5cm}, clip]{figure/dblpe_F1.eps}
\includegraphics[width=0.23\textwidth,trim={0.25cm 0cm 1.2cm 0cm}, clip]{figure/dblpe_bar.eps}
\caption{On DBLP-E data, TRNNGCN and RNNGCN outperform the static GCN and spectral clustering methods and show better performance over time, indicating that they can optimize their decay rates to account for historical information. TRNNGCN slightly outperforms RNNGCN.}
\label{fig:dblpe}
\end{figure*}
\begin{table*}[ht]
\centering
\begin{tabular}{ccccccccccccc}
\hline
& \multicolumn{3}{c|}{DBLP-3} &\multicolumn{3}{c|}{DBLP-5} &\multicolumn{3}{c|}{Reddit} &\multicolumn{3}{c}{Brain}\\
\hline
& ACC&AUC&F1& ACC&AUC&F1&ACC&AUC&F1&ACC&AUC&F1 \\
\hline
GCN&71.6&62.2&35.8&64.9&51.0&\textbf{\textit{58.7}}&31.0&24.5&47.4&35.2&25.0&80.3\\
GAT&70.9&59.4&57.8&62.3&48.2&51.4&16.8&4.8&50.0&34.6&26.4&81.6\\
GraphSage &74.5&63.6&55.0&\textbf{\textit{66.5}}&53.9&55.1&29.2&20.7&42.5&\textbf{44.2}&\textbf{\textit{41.9}}&\textbf{86.7}\\
Spectral&45.7&51.6&51.2&43.8&45.6&51.3&30.1&24.1&51.7&42.7&41.7&68.1\\
DynAERNN &48.1&54.2&50.8&33.1&39.1&51.2&31.1&\textbf{31.7}&\textbf{54.1}&20.5&20.3&55.6\\
GCNLSTM&74.5&63.6&48.4&\textbf{\textit{66.5}}&53.2&54.6&31.9&25.5&46.1&38.8&32.9&\textbf{\textit{85.9}}\\
EGCN&72.3&60.7&48.1&63.2&50.6&53.2&28.3&12.5&50.0&28.6&26.1&73.7\\
RNNGCN&\textbf{\textit{75.9}}&\textbf{\textit{68.0}}&\textbf{\textit{66.7}}&65.7&\textbf{\textit{55.4}}&58.6&\textbf{33.6}&20.5&49.7&41.0&38.6&84.7\\
TRNNGCN&\textbf{78.0}&\textbf{72.1}&\textbf{73.8}&\textbf{67.4}&\textbf{57.9}&\textbf{63.5}&\textbf{33.6}&\textbf{\textit{25.6}}&\textbf{\textit{53.2}}&\textbf{\textit{43.8}}&\textbf{42.4}&85.7\\
\hline
\end{tabular}
\caption{TRNNGCN consistently achieves the best (bold) or second-best (bold italics) accuracy (ACC), area under the ROC curve (AUC), and F1 score compared to baseline algorithms on four temporal datasets.}
\label{tab:result_temporal}
\end{table*}
Figure \ref{fig:dblpe} shows the performance of RNNGCN and TRNNGCN in DBLP-E. Similar to Figure~\ref{fig:simulate}'s result on the simulated data, GCN's performance decreases over time as the accumulated effect of class change increases. Spectral Clustering consistently performs poorly since it cannot learn the high-dimensional patterns of the DBLP-E graph. RNNGCN and TRNNGCN maintain good performance and fully utilize the temporal information. As the two classes have different change probabilities, TRNNGCN learns a better decay rate and performs better. We further show the average accuracy, AUC, and F1-score for each baseline method at each timestep; RNNGCN and TRNNGCN consistently outperform the other baselines. GCNLSTM comes the closest to matching their performance. GCNLSTM uses a LSTM layer to account for historical information, which is similar to our methods but lacks interpretability as the LSTM operates on the output of the GCN layer (which is not readily interpretable) instead of the original graph adjacency information. We use a RNN instead of LSTM layer in our algorithms for computational efficiency.
\textbf{Node Classification with Temporal Features}
Although our analysis is based on dynamic networks without features, Table \ref{tab:result_temporal}'s performance results demonstrate the applicability of our RNNGCN and TRNNGCN algorithms to the DBLP-3, DBLP-5, Brain, and Reddit datasets with node features. TRNNGCN achieves the best or second-best performance consistently on all datasets, even outperforming EGCN and GCNLSTM, which unlike TRNNGCN fully utilize the historical information of node features. While GraphSAGE shows good accuracy, AUC, and F1-score on the Brain dataset, no other baseline method does well across all three metrics for any other dataset. RNNGCN performs second-best on DBLP-3 but worse on the other datasets, likely because those datasets have more than three classes, which would likely have different optimal decay rates. TRNNGCN can account for these differences, but RNNGCN cannot.
We further highlight the importance of taking into account historical information by noting that the static baselines (GCN, GAT, GraphSAGE, and spectral clustering) generally perform poorly compared to the dynamic baselines (DynAERNN, GCNLSTM, EGCN). DynAERNN can perform significantly worse than GCNLSTM and EGCN, likely because it is an unsupervised method that cannot take advantage of labeled training data. Thus, RNNGCN and TRNNGCN's good performance is likely due to their ability to optimally take advantage of historical graph information, even if they cannot use historical node feature information.
\section{Conclusion}
This work proposes RNNGCN and TRNNGCN, two new neural network architectures for clustering on dynamic graphs. These methods are inspired by the insight that RNNs progressively decrease the weight placed on their inputs over time according to a learned decay rate parameter.
This decay rate can in turn be interpreted as the importance of historical connection information associated with each community or cluster in the graph.
We show that decaying historical connection information can achieve almost exact recovery when used for spectral clustering on dynamic stochastic block models, and that the RNN decay rates on simulated data match the theoretically optimal decay rates for such stochastic block models. We finally validate the performance of RNNGCN and TRNNGCN on a range of real datasets, showing that TRNNGCN consistently outperforms static clustering methods as well as previously proposed dynamic clustering methods. This performance is particularly remarkable compared to dynamic clustering methods that account for historical information of both the connections between nodes and the node features; TRNNGCN ignores the historical feature information. We plan to investigate neural network architectures that reveal the importance of these dynamic node features in our future work. Much work also remains on better establishing
the models' interpretability.
\newpage
\section*{Acknowledgements}
This research was partially supported by NSF grant CNS-1909306. The authors would like to thank Jinhang Zuo, Mengqiu Teng and Xiao Zeng for their inputs to the work.
\section{Model}
We first introduce a dynamic version of the Stochastic Block Model (SBM) often used to study graph clustering~\cite{holland1983stochastic,abbe2017community}, which we will use for our theoretical analysis in the rest of the paper.
\subsection{Stochastic Block Model} For positive integers $K$ and $n$, a probability vector $p\in [0,1]^K$, and a symmetric connectivity matrix $B\in[0,1]^{K\times K}$, the SBM defines a random graph with $n$ nodes split into $K$ clusters. The goal of a prediction method for the SBM is to correctly divide nodes into their corresponding clusters, based on the graph structure. Each node is independently and randomly assigned a cluster in $\{1,...,K\}$ according to the distribution $p$; we can then say that a node is a ``member'' of this cluster. Undirected edges are independently created between any pair of nodes in clusters $i$ and $j$ with probability $B_{ij}$,
where the $(i,j)$ entry of $B$ is
\begin{equation}
B_{ij}=\left\{
\begin{aligned}
\alpha& ,\;i=j\\
\tau \alpha & ,\;i\neq j,
\end{aligned}
\right.
\end{equation}
for $\alpha\in(0,1)$ and $\tau\in(0,1)$, implying that the probability of an edge forming between nodes in the same cluster is $\alpha$ (which is the same for each cluster) and the edge formation probability between nodes in different clusters is $\tau \alpha$.
Let $\Theta \in {\{0,1\}}^{n\times K}$ denotes the matrix representing the nodes' cluster memberships, where $\Theta_{ik}=1$ indicates that node $i$ belongs to the $k$-th cluster, and is $0$ otherwise.
We use $A\in\{0,1\}^{n \times n}$ to denote the (symmetric) adjacency matrix of the graph, where $A_{ij}$ indicates whether there is a connection (edge) between node $i$ and node $j$. From our node connectivity model, we find that given $\Theta$, for $i<j$, we have
\begin{equation}
A_{ij}|\{\Theta_{ik}=1,\Theta_{jl}=1\} \backsim \text{Ber}(B_{kl}),
\end{equation}
where $\text{Ber}(p)$ indicates a Bernoulli random variable with parameter $p$. We define $A_{ii}=0$ (nodes are not connected directly to themselves) and since all edges are undirected, $A_{ij}=A_{ji}$. We further define the connection probability matrix $P=\Theta B \Theta^T \in [0,1]^{n\times n}$, where $P_{ij}$ is the connection probability of node $i$ and node $j$ and
$\mathbb{E}[A]=P-\text{diag}(P)$.
\subsection{Dynamic Stochastic Block Model} We now extend the SBM model to include how the graph evolves over time. We consider a set of discrete time steps $t = 1,2,\ldots,T$.
At each time step $t$, the Dynamic SBM generates new intra- and inter-cluster edges according to the probabilities $\alpha$ and $\tau\alpha$ as defined for the SBM above. All edges persist over time.
We assume a constant number of nodes $n$, number of clusters $K$, and connectivity matrix $B$, but the node membership matrix $\Theta_t$ depends on time $t$, i.e., nodes' cluster memberships may change over time. We similarly define the connectivity matrix $P_t = \Theta_t B (\Theta_t)^T$.
We model changes in nodes' cluster memberships as a Markov process with a constant transition probability matrix $H\in[0,1]^{K\times K}$. Let $\varepsilon_j\in(0,1)$ denotes the change probability of nodes in cluster $j$, i.e., the probability a node in cluster $j$ changes its membership. At each time step, node $v_i$ in cluster $j$ changes its membership to cluster $k$ with the following probability (independently from other nodes):
\begin{equation*}
H_{j,k}=\mathbb{P}\left[\Theta_{ik}^{t}=1|\Theta_{ij}^{t-1}=1\right]=\left\{
\begin{aligned}
1-\varepsilon_j & , & j=k\\
\frac{\varepsilon_{j}}{K-1} & , & j\neq k,
\end{aligned}
\right.
\end{equation*}
Note that $\varepsilon_j$ may be specific to cluster $j$, e.g., if some clusters experience less membership turnover. We give an example of such a graph in our experimental evaluation.
The goal of a clustering algorithm on a graph is to recover the membership matrix $\Theta$ up to column permutation. Static clustering algorithms give an estimate $\hat{\Theta}$ of the node membership; a dynamic clustering algorithm should produce such an estimate for each time $t$. We define two performance metrics for these estimates (in dynamic graphs, they may be evaluated for an estimate $\hat{\Theta} = \hat{\Theta}_t$ relative to $\Theta = \Theta_t$ at any time $t$):
\begin{definition}
[Relative error of $\hat{\Theta}$]
The relative error of a clustering estimate $\hat{\Theta}$ is
\begin{equation}
E(\hat{\Theta},\Theta)=\frac{1}{n} \min_{\pi\in \mathcal{P}} \|\hat{\Theta} \pi-\Theta\|_0,
\end{equation}
where $\mathcal{P}$ is the set of all $K\times K$ permutation matrices and $\|.\|_0$ counts the number of non-zero elements of a matrix.
\end{definition}
\begin{definition}
[Almost Exact Recovery]
A clustering estimate $\hat{\Theta}$ achieves almost exact recovery when
\begin{equation}
\mathbb{P}\left[1- \frac{1}{n} \min_{\pi\in \mathcal{P}} \|\hat{\Theta} \pi-\Theta\|_0=1-o(1)\right]=1-o(1).
\end{equation}
which also implies that the expectation of $E(\hat{\Theta},\Theta)$ is $o(1)$.
\end{definition}
Our goal is then to find an algorithm that produces an estimate $\hat{\Theta}$ minimizing $E(\hat{\Theta},\Theta)$. In the next section, we discuss the well-known (static) spectral clustering algorithm and analyze a simple decay-based method that allows a static algorithm to make dynamic membership estimates.
\section{Related Work}
Over the past few years, there has been significant research dedicated to graph clustering algorithms, motivated by applications such as community detection in social networks. While some works consider theoretical analysis of such clustering algorithms, more recently representation learning algorithms have been proposed that perform well in practice with few theoretical guarantees. We aim to \emph{connect} these approaches by using a theoretical analysis of decay-based dynamic clustering algorithms to design a new neural network-based approach that is easily interpretable.
\textbf{Theoretical analyses.}
Traditional spectral clustering algorithms use the spectrum of the graph adjacency matrix to generate a compact representation of the graph connectivity~\cite{lei2015consistency,qin2013regularized}. A line of work on static clustering algorithms uses the stochastic block model for graph connectivity~\cite{abbe2017community}, which more recent works have extended to a dynamic stochastic block model~\cite{keriven2020sparse,pensky2019spectral}. While these works do not distinguish between clusters with different transition probabilities, earlier models incorporate such heterogeneity~\cite{xu2015stochastic}. Other works use a Bayesian approach~\cite{yang2011detecting}, scoring metrics~\cite{agarwal2018dyperm,zhuang2019dynamo}, or multi-armed bandits~\cite{mandaglio2019dynamic} to detect communities and their evolution, while~\citet{xu2010evolutionary} use a decay rate similar to the one we propose.
As \textbf{representation learning} becomes popular, graph neural networks~\cite{zhang2018link,wu2020comprehensive,kipf2016semi} such as
GraphSage~\cite{hamilton2017inductive} have been used to cluster nodes in graphs based on (static) connections between nodes and node features. Graph Attention Networks (GAT)~\cite{velivckovic2017graph,xu2019self} use attention-based methods to construct a neural network that highlights the relative importance of each feature, while dynamic supervised~\cite{kumar2019predicting} and unsupervised~\cite{goyal2020dyngraph2vec} methods can track general network dynamics, or may be designed for clustering on graphs with dynamic edges and dynamic node features~\cite{chen2018gc,xu2019spatio,xu2020inductive}. EvolveGCN~\cite{pareja2020evolvegcn} usesa GCN to evolve the RNN weights, which is similar to our approach; however, we ensure interpretability of the RNN weights by placing the RNN before the GCN layers, which we show improves the clustering performance.
Finally, several works have considered the \textbf{interpretability of general GCN and RNN structures}~\cite{dehmamy2019understanding,liang2017interpretable,guo2019exploring}, such as GNNExplainer~\cite{ying2019gnnexplainer}. In the context of graph clustering, some works have used attention mechanisms to provide interpretable weights on node features~\cite{xu2019spatio}, but attention may not capture true feature importance~\cite{serrano2019attention}. Moreover, these works do not consider the importance of \emph{historical} information, as we consider in this work.
| 62b41a49355aecc6bc1a6f6f1571484428c4a5fc | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\subsection{A few auxiliary notions}
\textbf{Complemented subspaces in Banach spaces.}
Let $X$ be a real or complex Banach space.
By a subspace of $X$ we will mean a linear subset of $X$.
Let $M$ be a subspace of $X$.
$M$ is said to be complemented in $X$
if there exists a subspace $N$ (a complement) such that
$X$ is the topological direct sum of $M$ and $N$.
This means that the sum operator $S:M\times N\to X$
defined by $S(x,y)=x+y$, $x\in M, y\in N$ is an isomorphism
(of normed linear spaces).
Here $M\times N$ is the linear space of all pairs $(x,y)$
with $x\in M, y\in N$ endowed with the norm $\|(x,y)\|=\|x\|+\|y\|$.
One can easily check that $M$ is complemented in $X$ if and only if
there exists a continuous linear projection onto $M$, i.e.,
a continuous linear operator $P:X\to X$ such that
$Px\in M$ for all $x\in X$ and $Px=x$ for $x\in M$.
Each complemented subspace is closed
(this follows from the fact that $M=S(M\times\{0\})$).
Note that one can give the following (equivalent)
definition of complementability:
a subspace $M$ is said to be complemented in $X$ if $M$ is closed
and there exists a closed subspace $N$ such that
$M\cap N=\{0\}$ and $M+N=X$
(the equivalence of the definition to the original
follows from the fact that each complemented subspace is closed and
the Banach inverse mapping theorem).
If $X$ is a Hilbert space,
then each closed subspace $M$ of $X$ is complemented in $X$
(one can consider the orthogonal decomposition $X=M\oplus M^\bot$
or, equivalently, the orthogonal projection onto $M$).
\textbf{Sum of subspaces.}
Let $V$ be a vector space and $V_1,...,V_n$ be subspaces of $V$.
Define the sum of $V_1,...,V_n$ in the natural way, namely,
$$
V_1+...+V_n:=\{x_1+...+x_n\,|\,x_1\in V_1,...,x_n\in V_n\}.
$$
It is clear that $V_1+...+V_n$ is a subspace of $V$.
\textbf{Linear independence.}
Let $V$ be a vector space and $V_1,...,V_n$ be subspaces of $V$.
The system of subspaces $V_1,...,V_n$ is said to be linearly independent
if an equality $x_1+...+x_n=0$, where $x_1\in V_1,...,x_n\in V_n$,
implies that $x_1=...=x_n=0$.
\subsection{Notation}
Throughout the paper, $X$ is a real or complex Banach space
with norm $\|\cdot\|$.
When $X$ is a Hilbert space we denote by $\langle\cdot,\cdot\rangle$
the inner product in $X$.
The identity operator on $X$ is denoted by $I$
(throughout the paper it is clear which Banach space is being considered).
All operators in the paper are continuous linear operators.
In particular, by a projection we always mean a continuous linear projection.
The kernel and range of an operator $T$
will be denoted by $\ker(T)$ and $Ran(T)$, respectively.
For a continuous linear operator $T$ between two Hilbert spaces
we denote by $T^*$ its adjoint.
All vectors are vector-columns;
the superscript "t" means transpose.
\subsection{}
Starting point for this paper are the main results of our paper \cite{F20}
where the following questions are studied.
Let $X$ be a Banach space and $X_1,...,X_n$ be complemented subspaces of $X$.
\textbf{Question 1:}
\textit{Is $X_1+...+X_n$ complemented in $X$?}
If Question 1 has positive answer (for given $X_1,...,X_n$),
then the next natural question arises:
\textbf{Question 2:}
\textit{Suppose that we know some projections
$P_1,...,P_n$ onto $X_1,...,X_n$, respectively.
Is there a formula for a projection onto $X_1+...+X_n$
(in terms of $P_1,...,P_n$) (of course, under certain conditions)?}
\subsection{}
Let us present the main results of \cite{F20}.
First, we provided a sufficient condition for the sum
of complemented subspaces to be complemented.
\begin{A}(\cite[Theorem 2.1]{F20})
Let $X$ be a Banach space,
$X_1,...,X_n$ be complemented subspaces of $X$ and
$P_1,...,P_n$ be projections onto $X_1,...,X_n$, respectively.
Let $\varepsilon_{ij}$, $i\neq j$, $i,j\in\{1,...,n\}$,
be nonnegative numbers such that
\begin{equation*}
\|P_i x\|\leqslant\varepsilon_{ij}\|x\|,\quad x\in X_j,
\end{equation*}
for any distinct $i,j\in\{1,...,n\}$.
Define the $n\times n$ matrix $E=(e_{ij})$ by
$$e_{ij}=
\begin{cases}
0, &\text{if $i=j$,}\\
\varepsilon_{ij}, &\text{if $i\neq j$.}
\end{cases}
$$
Denote by $r(E)$ the spectral radius of $E$ and set $A:=P_1+...+P_n$.
If $r(E)<1$,
then the subspaces $X_1,...,X_n$ are linearly independent,
their sum is complemented in $X$, and
$\ker(P_1)\cap...\cap\ker(P_n)$ is a complement for $X_1+...+X_n$ in $X$.
Moreover, $I-(I-A)^N$ converges uniformly to the projection $P$
onto $X_1+...+X_n$ along $\ker(P_1)\cap...\cap\ker(P_n)$ as $N\to\infty$.
\end{A}
\subsection{}
The next result shows that the rate of convergence of
$I-(I-A)^N$ to $P$ can be estimated from above by $C\alpha^N$,
where $\alpha\in[0,1)$.
To formulate the result we need the following notation:
for two vectors $u,v\in\mathbb{R}^n$ we will write $u\leqslant v$
if $u\leqslant v$ coordinatewise.
To make the formulation of the result clearer
we first make the following important remark.
Since $E$ is a nonnegative matrix, the condition $r(E)<1$ is equivalent to
the existence of a vector $w\in\mathbb{R}^n$ with positive coordinates and
a number $\alpha\in[0,1)$ such that $Ew\leqslant\alpha w$.
More precisely, if such $w$ and $\alpha$ exist,
then $r(E)\leqslant\alpha<1$ (see \cite[Corollary 8.1.29]{HJ13}).
Conversely, suppose that $r(E)<1$.
If $E$ is irreducible,
then one can take $\alpha$ to be $r(E)$ and
$w$ a Perron-Frobenius vector of $E$.
If $E$ is not irreducible,
then consider the matrix $E'=(e_{ij}+\delta)$ for sufficiently small $\delta>0$,
and take $\alpha$ to be $r(E')$ and $w$ a Perron-Frobenius vector of $E'$.
Since $r(E^t)=r(E)$, we see that the condition $r(E)<1$
is also equivalent to the existence of a vector $w$ with positive coordinates
and a number $\alpha\in[0,1)$ such that $E^t w\leqslant \alpha w$.
\begin{B}(\cite[Theorem 2.2]{F20})
(1) Suppose $w=(w_1,...,w_n)^t$ with positive coordinates and
$\alpha\in[0,1)$ satisfy $Ew\leqslant\alpha w$.
Then for each $N\geqslant 1$,
$$
\|I-(I-A)^N-P\|\leqslant
(w_1+...+w_n)\max\{(1/w_1)\|P_1\|,...,(1/w_n)\|P_n\|\}\frac{\alpha^N}{1-\alpha}.
$$
(2) Suppose $w=(w_1,...,w_n)^t$ with positive coordinates and
$\alpha\in[0,1)$ satisfy $E^t w\leqslant \alpha w$.
Then for each $N\geqslant 1$,
$$
\|I-(I-A)^N-P\|\leqslant
(w_1\|P_1\|+...+w_n\|P_n\|)\max\{(1/w_1),...,(1/w_n)\}\frac{\alpha^N}{1-\alpha}.
$$
\end{B}
Using Theorem B, we can get concrete estimates for
the rate of convergence of $I-(I-A)^N$ to $P$.
Suppose $E$ is irreducible and $r(E)<1$.
Take $\alpha$ to be $r(E)$ and $w$ a Perron-Frobenius vector of $E$.
Then we get
$$
\|I-(I-A)^N-P\|\leqslant
(w_1+...+w_n)\max\{(1/w_1)\|P_1\|,...,(1/w_n)\|P_n\|\}\frac{(r(E))^N}{1-r(E)}.
$$
Similarly, we can take $\alpha$ to be $r(E)$ and $w$ a Perron-Frobenius vector of $E^t$.
Then we get
$$
\|I-(I-A)^N-P\|\leqslant
(w_1\|P_1\|+...+w_n\|P_n\|)\max\{(1/w_1),...,(1/w_n)\}\frac{(r(E))^N}{1-r(E)}.
$$
\subsection{}
The assumption $r(E)<1$ is a \textit{sharp} sufficient condition
for $X_1+...+X_n$ to be complemented in $X$.
More precisely, we have the following result.
\begin{C}(\cite[Theorem 2.3]{F20})
Let $E=(e_{ij})$ be an $n\times n$ matrix with
$e_{ii}=0$ for $i=1,...,n$ and
$e_{ij}\geqslant 0$ for any distinct $i,j\in\{1,...,n\}$.
If $r(E)=1$,
then there exist a Banach space $X$,
complemented subspaces $X_1,...,X_n$ of $X$ and
projections $P_1,...,P_n$ onto $X_1,...,X_n$, respectively, such that
\begin{enumerate}
\item
$\|P_i x\|=e_{ij}\|x\|$, $x\in X_j$, for any distinct $i,j\in\{1,...,n\}$;
\item
$X_1,...,X_n$ are linearly independent;
\item
$X_1+...+X_n$ is closed and not complemented in $X$.
\end{enumerate}
\end{C}
\begin{remark}
If $r(E)>1$, the theorem can be applied to the matrix $(1/r(E))E$.
\end{remark}
\subsection{}
The aim of this paper is to obtain analogues of Theorems A, B, C
for the case when $X$ is a Hilbert space,
$X_1,...,X_n$ are closed subspaces of $X$,
and $P_1,...,P_n$ are orthogonal projections onto $X_1,...,X_n$, respectively.
\section{Results}
Let $X$ be a real or complex Hilbert space,
$X_1,...,X_n$ be closed subspaces of $X$,
and $P_1,...,P_n$ be orthogonal projections onto $X_1,...,X_n$, respectively.
As in the Banach space setting, we assume that nonnegative numbers
$\varepsilon_{ij}$, $i\neq j$, $i,j\in\{1,...,n\}$ are such that
\begin{equation}\label{eq:H varepsilon_ij}
\|P_i x\|\leqslant\varepsilon_{ij}\|x\|,\quad x\in X_j,
\end{equation}
for any distinct $i,j\in\{1,...,n\}$.
Clearly, \eqref{eq:H varepsilon_ij} is equivalent to the
inequality $\|P_i|_{X_j}\|\leqslant\varepsilon_{ij}$.
Now observe that for arbitrary closed subspaces $M$ and $N$ of
the space $X$ $\|P_M|_N\|=\|P_N|_M\|$,
where $P_M$ and $P_N$ are the orthogonal projections onto $M$ and $N$,
respectively.
Indeed, if $M=\{0\}$ or $N=\{0\}$, then $\|P_M|_N\|=\|P_N|_M\|=0$.
Assume that $M$ and $N$ are nonzero.
One can easily check that $(P_M|_N:N\to M)^*=P_N|_M:M\to N$.
Therefore $\|P_M|_N\|=\|P_N|_M\|$.
Hence, we can and will assume that
$\varepsilon_{ij}=\varepsilon_{ji}$ for any distinct $i,j\in\{1,...,n\}$.
Define the $n\times n$ matrix $E=(e_{ij})$ by
$$e_{ij}=
\begin{cases}
0, &\text{if $i=j$,}\\
\varepsilon_{ij}, &\text{if $i\neq j$.}
\end{cases}
$$
It is clear that $E$ is symmetric and nonnegative.
It follows that $r(E)$, the spectral radius of $E$,
is the maximum eigenvalue of $E$.
Set $A:=P_1+...+P_n$.
Applying Theorem A and noting that
$$
\ker(P_1)\cap...\cap\ker(P_n)=
X_1^\bot\cap...\cap X_n^\bot=
(X_1+...+X_n)^\bot,
$$
we get the following result.
\begin{theorem}\label{Th:H}
If $r(E)<1$,
then the subspaces $X_1,...,X_n$ are linearly independent and
their sum is closed in $X$.
Moreover, the sequence of operators $I-(I-A)^N$
converges uniformly to the orthogonal projection $P$
onto $X_1+...+X_n$ as $N\to\infty$.
\end{theorem}
\begin{remark}
A part on the closedness of Theorem~\ref{Th:H} follows from
\cite[Theorem 1.2]{K11}.
\end{remark}
\begin{remark}
$r(E)<1$ $\Leftrightarrow$
the matrix $I-E$ is positive definite $\Leftrightarrow$
every leading principal minor of the matrix is positive.
For $n=2$ the inequality $r(E)<1$ is equivalent to
$\varepsilon_{12}^2<1$ $\Leftrightarrow$ $\varepsilon_{12}<1$.
For $n=3$ the inequality $r(E)<1$ is equivalent to
$$
\varepsilon_{12}^2+\varepsilon_{23}^2+\varepsilon_{31}^2+
2\varepsilon_{12}\varepsilon_{23}\varepsilon_{31}<1.
$$
\end{remark}
\begin{remark}
For the optimal choice $\varepsilon_{ij}=\|P_i|_{X_j}\|$, $i\neq j$,
the condition $r(E)<1$ can be formulated in terms of
the minimal angles between the subspaces.
Let us recall the definition of the minimal angle between two subspaces;
the notion was introduced by J.~Dixmier in \cite{D49}.
Let $M$ and $N$ be two closed subspaces of $X$.
Define the number $c_0(M,N)\in[0,1]$ by
$$
c_0(M,N)=\sup\{|\langle x,y\rangle|\,|\,
x\in M, \|x\|\leqslant 1, y\in N, \|y\|\leqslant 1\}.
$$
The minimal angle between $M$ and $N$ is the angle
$\varphi_0(M,N)\in[0,\pi/2]$
whose cosine is equal to $c_0(M,N)$.
It is well known, and one can easily prove, that
$c_0(M,N)=\|P_M|_N\|$, where $P_M$ is the orthogonal projection onto $M$.
Set $\varphi_{ij}:=\varphi_0(X_i,X_j)$, $i\neq j$, $i,j\in\{1,2,...,n\}$.
Then $\varepsilon_{ij}=\|P_i|_{X_j}\|=c_0(X_i,X_j)=\cos\varphi_{ij}$,
$i\neq j$, and therefore
$$
e_{ij}=
\begin{cases}
0, &\text{if $i=j$,}\\
\cos\varphi_{ij}, &\text{if $i\neq j$.}
\end{cases}
$$
$r(E)<1$ $\Leftrightarrow$
the matrix $I-E$ is positive definite $\Leftrightarrow$
every leading principal minor of the matrix is positive.
For $n=2$ the inequality $r(E)<1$ is equivalent to
$\cos^2\varphi_{12}<1$ $\Leftrightarrow$
$\cos\varphi_{12}<1$ $\Leftrightarrow$
$\varphi_{12}>0$.
Thus, if $\varphi_{12}>0$, then $X_1\cap X_2=\{0\}$ and
the subspace $X_1+X_2$ is closed in $X$.
It is worth mentioning that the converse is also true, i.e.,
if $X_1\cap X_2=\{0\}$ and the subspace $X_1+X_2$ is closed in $X$,
then $\varphi_{12}>0$ (see \cite[Theorem 12]{D95}).
For $n=3$ the inequality $r(E)<1$ is equivalent to
$$
\cos^2\varphi_{12}+\cos^2\varphi_{23}+\cos^2\varphi_{31}+
2\cos\varphi_{12}\cos\varphi_{23}\cos\varphi_{31}<1.
$$
One can easily check that the last inequality is equivalent to
$\varphi_{12}+\varphi_{23}+\varphi_{31}>\pi$.
\end{remark}
To make the paper self-contained we will prove
Theorem~\ref{Th:H} in Section~\ref{S:proofs}.
For the rate of convergence of $I-(I-A)^N$ to $P$
we have the following estimate
which is more precise than that given by Theorem B.
\begin{theorem}\label{Th:H rate}
If $r(E)<1$, then $\|I-(I-A)^N-P\|\leqslant (r(E))^N$
for each $N\geqslant 1$.
\end{theorem}
The assumption $r(E)<1$ is a \textit{sharp} sufficient condition
for $X_1+...+X_n$ to be closed in $X$.
More precisely, we have the following result.
\begin{theorem}\label{Th:H r(E)=1}
Let $E=(e_{ij})$ be a symmetric $n\times n$ matrix
with $e_{ii}=0$ for $i=1,...,n$ and
$e_{ij}\geqslant 0$ for any distinct $i,j\in\{1,...,n\}$.
If $r(E)=1$,
then there exist a separable infinite dimensional Hilbert space $X$ and
infinite dimensional closed subspaces $X_1,...,X_n$ of $X$ such that
\begin{enumerate}
\item
$\|P_i|_{X_j}\|=e_{ij}$ for any distinct $i,j\in\{1,...,n\}$;
here $P_i$ is the orthogonal projection onto $X_i$, $i=1,...,n$.
\item
$X_1,...,X_n$ are linearly independent;
\item
$X_1+...+X_n$ is not closed in $X$.
\end{enumerate}
\end{theorem}
\begin{remark}
If $r(E)>1$, the theorem can be applied to the matrix $(1/r(E))E$.
\end{remark}
\begin{remark}
For a similar result see \cite[Theorem 1.2, part ``Moreover,...'']{K11}.
However, in \cite{K11} there is no proof for this part of Theorem 1.2.
\end{remark}
\section{Proofs}\label{S:proofs}
\subsection{Proof of Theorems \ref{Th:H} and \ref{Th:H rate}}
We first formulate the following lemma.
\begin{lemma}\label{L:H}
Let $H,K$ be Hilbert spaces and $T:H\to K$ be a continuous linear operator.
Suppose that
$$
\alpha\|x\|\leqslant\|Tx\|\leqslant\beta\|x\|,\quad x\in H,
$$
where $\alpha$ and $\beta$ are positive numbers.
Then
$$
\alpha\|y\|\leqslant\|T^* y\|\leqslant\beta\|y\|,\quad y\in Ran(T).
$$
\end{lemma}
The proof is simple and is omitted.
Now we are ready to prove Theorems \ref{Th:H} and \ref{Th:H rate}.
For simplicity of notation, we set $r:=r(E)$.
Let $X_1\oplus...\oplus X_n$ be
the (orthogonal) direct sum of Hilbert spaces $X_1,...,X_n$.
Define the operator $S:X_1\oplus...\oplus X_n\to X$ by
$$
S(x_1,...,x_n)^t=x_1+...+x_n,\quad x_1\in X_1,...,x_n\in X_n.
$$
One can easily check that $S^*:X\to X_1\oplus...\oplus X_n$
acts as follows:
$$
S^*x=(P_1 x,...,P_n x)^t,\quad x\in X.
$$
Therefore $SS^*=P_1+...+P_n=A$.
For every $v=(v_1,...,v_n)^t\in X_1\oplus...\oplus X_n$ we have
\begin{align*}
&|\|Sv\|^2-\|v\|^2|=
|\|v_1+...+v_n\|^2-\|v_1\|^2-...-\|v_n\|^2|=\\
&=|\sum_{i\neq j}\langle v_i,v_j\rangle|\leqslant
\sum_{i\neq j}|\langle v_i,v_j\rangle|=
\sum_{i\neq j}|\langle P_i v_i,v_j\rangle|=\\
&=\sum_{i\neq j}|\langle v_i,P_i v_j\rangle|\leqslant
\sum_{i\neq j}\|v_i\|\|P_i v_j\|\leqslant
\sum_{i\neq j}\varepsilon_{ij}\|v_i\|\|v_j\|=\\
&=\langle E(\|v_1\|,...,\|v_n\|)^t,(\|v_1\|,...,\|v_n\|)^t\rangle\leqslant\\
&\leqslant r\|(\|v_1\|,...,\|v_n\|)^t\|^2=
r(\|v_1\|^2+...+\|v_n\|^2)=
r\|v\|^2.
\end{align*}
So $|\|Sv\|^2-\|v\|^2|\leqslant r\|v\|^2$, i.e.,
\begin{equation}\label{ineq:S}
(1-r)\|v\|^2\leqslant\|Sv\|^2\leqslant(1+r)\|v\|^2,\quad
v\in X_1\oplus...\oplus X_n.
\end{equation}
Since $\|Sv\|\geqslant\sqrt{1-r}\|v\|$, $v\in X_1\oplus...\oplus X_n$,
we conclude that $\ker(S)=\{0\}$ and $Ran(S)$ is closed in $X$
$\Rightarrow$
the subspaces $X_1,...,X_n$ are linearly independent and their sum
$X_1+...+X_n$ is closed in $X$.
Further, from \eqref{ineq:S} and Lemma~\ref{L:H} it follows that
$$
(1-r)\|x\|^2\leqslant\|S^* x\|^2\leqslant(1+r)\|x\|^2,\quad x\in Ran(S).
$$
We have
$$
\|S^* x\|^2=\langle S^*x,S^*x\rangle=\langle SS^*x,x\rangle=\langle Ax,x\rangle.
$$
The closed subspace $Ran(S)=X_1+...+X_n$ is invariant with respect to $A$.
Denote by $A'$ the restriction of $A$ to $X_1+...+X_n$.
Then we get
$$
(1-r)\|x\|^2\leqslant\langle A' x,x\rangle\leqslant(1+r)\|x\|^2,
\quad x\in X_1+...+X_n,
$$
and hence
$$
|\langle(A'-I)x,x\rangle|\leqslant r\|x\|^2,\quad x\in X_1+...+X_n.
$$
Since $A'-I$ is self-adjoint, we conclude that $\|A'-I\|\leqslant r$.
Let us estimate $\|I-(I-A)^N-P\|$,
where $P$ is the orthogonal projection onto $X_1+...+X_n$.
Consider the orthogonal decomposition
$$
X=(X_1+...+X_n)\oplus(X_1+...+X_n)^\bot=
(X_1+...+X_n)\oplus (X_1^\bot\cap...\cap X_n^\bot).
$$
With respect to the decomposition we have
$P=I\oplus 0$ and $A=A'\oplus 0$.
Thus
$$
I-(I-A)^N-P=-(I-A')^N\oplus 0
$$
and
$$
\|I-(I-A)^N-P\|=\|-(I-A')^N\|\leqslant\|A'-I\|^N\leqslant r^N\to 0
$$
as $N\to\infty$.
Theorems \ref{Th:H} and \ref{Th:H rate} are proved.
\subsection{Proof of Theorem \ref{Th:H r(E)=1}}
We will prove Theorem~\ref{Th:H r(E)=1} for the case when
the base field of scalars is $\mathbb{R}$;
the proof for $\mathbb{C}$ is similar.
For a number $\alpha\in(0,1)$ consider the matrix $I-\alpha E$.
Note that this matrix has the following properties:
\begin{enumerate}
\item
$I-\alpha E$ is a real symmetric matrix with diagonal elements equal to $1$;
\item
the least eigenvalue of this matrix is equal to $1-\alpha$.
Consequently, this matrix is positive definite.
\end{enumerate}
Therefore $I-\alpha E$ is the Gram matrix of some
linearly independent collection of unit vectors of $\mathbb{R}^n$, say
$v^{(i)}=v^{(i)}(\alpha)$, $i=1,...,n$.
Let $L_i=L_i(\alpha)$ be the one-dimensional subspace
spanned by $v^{(i)}$, $i=1,...,n$.
Denote by $P_i=P_i(\alpha)$ the orthogonal projection onto $L_i$, $i=1,...,n$.
Clearly, $P_i v=\langle v,v^{(i)}\rangle v^{(i)}$, $v\in\mathbb{R}^n$, $i=1,...,n$.
It follows that
$P_i v^{(j)}=\langle v^{(j)},v^{(i)}\rangle v^{(i)}=-\alpha e_{ij}v^{(i)}$
and consequently
$\|P_i|_{L_j}\|=\alpha e_{ij}$
for arbitrary $i\neq j$.
Further, since $1$ is an eigenvalue of $E$,
there exists a unit vector $c=(c_1,...,c_n)\in\mathbb{R}^n$ such that $Ec=c$.
Then $(I-\alpha E)c=(1-\alpha)c$ and consequently
$\langle (I-\alpha E)c,c\rangle=1-\alpha$.
We rewrite this equality as
$\sum_{i,j}\langle v^{(j)},v^{(i)}\rangle c_j c_i=1-\alpha,$
which is equivalent to
\begin{equation}\label{eq:norm of linear combination}
\|c_1 v^{(1)}+...+c_n v^{(n)}\|^2=1-\alpha.
\end{equation}
Now we are ready to construct a Hilbert space $X$ and
its closed subspaces $X_1,...,X_n$ with the needed properties.
Take an arbitrary sequence $\alpha_k\in(0,1)$, $k=1,2,...$,
which converges to $1$ as $k\to\infty$.
Set
$$
X=\bigoplus_{k=1}^\infty\mathbb{R}^n
$$
and
$$
X_i=\bigoplus_{k=1}^\infty L_i(\alpha_k),\quad i=1,...,n,
$$
where $\bigoplus$ is the (orthogonal) direct sum of Hilbert spaces.
First, let us show that $\|P_i|_{X_j}\|=e_{ij}$ for each pair $i\neq j$
(here $P_i$ is the orthogonal projection onto $X_i$).
It is clear that
$$
P_i=\bigoplus_{k=1}^\infty P_i(\alpha_k).
$$
Hence,
$$
\|P_i|_{X_j}\|=
\sup\{\|P_i(\alpha_k)|_{L_j(\alpha_k)}\|\,|\,k=1,2,...\}=
\sup\{\alpha_k e_{ij}\,|\,k=1,2,...\}=
e_{ij}.
$$
Further, since the vectors
$v^{(1)}(\alpha_k),...,v^{(n)}(\alpha_k)$ are linearly independent
for $k=1,2,...$, we conclude that the subspaces $X_1,...,X_n$
are linearly independent.
It remains to show that $X_1+...+X_n$ is not closed in $X$.
Suppose that $X_1+...+X_n$ is closed in $X$.
Let $X_1\oplus...\oplus X_n$ be
the (orthogonal) direct sum of Hilbert spaces $X_1,...,X_n$.
Define the operator $S:X_1\oplus...\oplus X_n\to X$ by
$$
S(x_1,...,x_n)^t=x_1+...+x_n,\quad x_1\in X_1,...,x_n\in X_n.
$$
Since $X_1,...,X_n$ are linearly independent, we conclude that
$\ker(S)=\{0\}$.
Moreover, $Ran(S)=X_1+...+X_n$ is closed in $X$.
These properties imply that $S$ is an isomorphic embedding.
Hence, there exists a number $\beta>0$ such that
$\|Su\|\geqslant\beta\|u\|, u\in X_1\oplus...\oplus X_n.$
We rewrite this inequality as
\begin{equation}\label{ineq:norm of sum}
\|x_1+...+x_n\|^2\geqslant
\beta^2(\|x_1\|^2+...+\|x_n\|^2),\quad x_1\in X_1,...,x_n\in X_n.
\end{equation}
Now we choose
$x_i=(0,...,0,c_i v^{(i)}(\alpha_k),0,0,...)^t, i=1,...,n.$
By~\eqref{eq:norm of linear combination}
and~\eqref{ineq:norm of sum}
we get $1-\alpha_k\geqslant\beta^2$.
But $\alpha_k\to 1$ as $k\to\infty$ and thus we get a contradiction.
Hence, $X_1+...+X_n$ is not closed in $X$.
Theorem~\ref{Th:H r(E)=1} is proved.
\subsection*{Acknowledgements}
The research was funded by
Institute of Mathematics of NAS of Ukraine.
This research was supported by
the Project 2017-3M from the Department of Targeted Training of
Taras Shevchenko National University of Kyiv at the NAS of Ukraine.
| 3c647818e5e30e453388ed119ed83f5f287affc6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $G$ be a finite group and ${{\operatorname{Irr}}}(G)$ the set of its ordinary irreducible
characters. Frobenius showed the classical formula
$$\sum_{\chi\in{{\operatorname{Irr}}}(G)}\chi(1)^2=|G|$$
for the dimensions of the irreducible complex representations of $G$.
No analogue of this equation is known in the modular setting, that is, for the
set ${{\operatorname{IBr}}}(G)$ of irreducible $p$-Brauer characters of $G$, when $p$ is a prime
dividing $|G|$. In 2005, Willems \cite{Wi05} put forward a conjecture giving a
lower bound in terms of the prime-to-$p$ part $|G|_{p'}$ of the group order:
\begin{conjA}[Willems (2005)] \label{conj:W}
Let $G$ be a finite group and $p$ be a prime. Then
$$|G|_{p'}\le\sum_{\vhi\in{{\operatorname{IBr}}}(G)}\vhi(1)^2.$$
\end{conjA}
Willems \cite{Wi05} points out that his conjecture holds for groups with cyclic
Sylow $p$-subgroups as well as for $p$-solvable groups, and he proves it for
groups of Lie type in defining characteristic~$p$. Here we show:
\begin{thmA} \label{thm:main}
Willems' Conjecture~\ref{conj:W} holds for all groups for the prime $p=2$.
\end{thmA}
Our proof relies on a reduction of (a strengthening of) the conjecture by
Tong-Viet \cite[Prop.~1.1]{TV19} to the case of quasi-simple groups. He shows
that Conjecture~\ref{conj:W} holds for all finite groups at the prime $p$ if the
following conjecture on $p$-Brauer characters is true for all quasi-simple
groups:
\begin{conjA}[Tong-Viet (2019)] \label{conj:TV}
Let $G$ be a finite quasi-simple group with centre $Z(G)$ of $p'$-order. Then
$$|G/Z(G)|_{p'}\le\sum_{\vhi\in{{\operatorname{IBr}}}(G|\theta)}\vhi(1)^2$$
for all faithful characters $\theta\in{{\operatorname{Irr}}}(Z(G))$.
\end{conjA}
Here, for $N\unlhd G$ and $\theta\in{{\operatorname{IBr}}}(N)$ we denote by ${{\operatorname{IBr}}}(G|\theta)$ the
set of irreducible Brauer characters of $G$ above $\theta$. Also, recall that
a finite group $G$ is \emph{quasi-simple} if $G$ is perfect and $G/Z(G)$ is
simple. \par
Tong-Viet \cite[Prop.~2.1]{TV19} has verified his conjecture for the
finitely many quasi-simple groups of Lie type with exceptional Schur multiplier,
as well as for all covering groups of sporadic simple groups and for alternating
groups of small degree. We invoke the classification of finite simple groups to
deal with the general case, obtaining a complete answer at least when $p=2$. It
will turn out that in many cases the inequality in Conjecture~\ref{conj:TV} is
already satisfied with one single suitable Brauer character, like for symmetric
groups in characteristic~2 or many groups of Lie type in defining
characteristic, but on the other hand there are groups for which the degrees of
a large number of characters have to be taken into consideration; there is no
absolute upper bound on the number of necessary characters even in the class of
quasi-simple groups. \par
To show that there are sufficiently many such characters, we are led to derive
lower bounds for the number of conjugacy classes of finite groups of Lie type
containing regular semisimple elements (Proposition~\ref{prop:reg elt}) which
may be of independent interest. As an application we obtain that any simple
group of Lie type in characteristic~$p$ has at least two conjugacy classes of
length divisible by the $p$-part of the group order
(Corollary~\ref{cor:sambale}); this was recently used by Sambale \cite{Sa20}.
\medskip
After collecting some basic observations, we study the finite simple groups
according to their classification, starting with the groups of Lie type. In
Section~\ref{sec:defchar} we suppose that $p$ is the defining prime; here our
results are only partial for certain types. In Section~\ref{sec:crosschar} we
show Conjecture~\ref{conj:TV} in the non-defining characteristic situation for
all cases. Finally, in Section~\ref{sec:alt} we verify Conjecture~\ref{conj:TV}
for alternating groups for the prime $p=2$.
\section{Preliminary results}
\begin{prop} \label{prop:normal}
Let $G$ be a finite group and $N\unlhd G$ with $G/N$ solvable. Assume
Conjecture~\ref{conj:W} holds for $G$ at the prime $p$. Then it also holds for
$N$ at $p$.
\end{prop}
\begin{proof}
Let $N\le M\le G$ be a maximal (normal) subgroup. Then $G/M$ is cyclic of
prime order. Using Clifford theory and an easy counting argument we conclude
that Conjecture~\ref{conj:W} holds for $M$. The general statement thus follows
by induction over a composition series of $G/N$.
\end{proof}
We will use the following consequence of a result of Kiyota and Wada:
\begin{lem} \label{lem:cyclic}
Conjecture~\ref{conj:TV} holds for quasi-simple groups with cyclic Sylow
$p$-subgroups.
\end{lem}
\begin{proof}
It was shown by Kiyota and Wada \cite[Prop.~4.7]{KW93} (see also
\cite[Prop.~3.1]{HW07}) that for every $p$-block $B$ of a finite group $G$ with
a cyclic defect group $D$ we have
$$\dim B\le |D|\sum_{\vhi\in{{\operatorname{IBr}}}(B)}\vhi(1)^2.$$
Now assume that $G$ is quasi-simple with a cyclic Sylow $p$-subgroup and centre
$Z(G)$ of $p'$-order. Let $\theta\in{{\operatorname{Irr}}}(Z(G))$ be faithful. By
\cite[Thm.~(9.2)]{Na98}, for example, there is a
union $\cB$ of $p$-blocks of $G$ with ${{\operatorname{Irr}}}(G|\theta)=\bigcup_{B\in\cB}{{\operatorname{Irr}}}(B)$
and ${{\operatorname{IBr}}}(G|\theta)=\bigcup_{B\in\cB}{{\operatorname{IBr}}}(B)$. Let $d$ be the maximal order of
a defect group of any $B\in\cB$. Then, using the above inequality we have
$$|G/Z(G)|_{p'}\le \frac{1}{d}|G/Z(G)|
=\frac{1}{d}\sum_{\chi\in {{\operatorname{Irr}}}(G|\theta)}\chi(1)^2
=\frac{1}{d}\sum_{B\in\cB}\dim B \le \sum_{\vhi\in{{\operatorname{IBr}}}(G|\theta)}\vhi(1)^2,$$
as claimed by Conjecture~\ref{conj:TV}.
\end{proof}
In many cases, we will make use of characters of \emph{$p$-defect zero}, that
is, irreducible characters $\chi$ of a finite group $G$ such that $\chi(1)$
contains the full $p$-part of the group order $|G|$. It is a basic result of
Brauer that these remain irreducible under $p$-modular reduction and thus
furnish irreducible $p$-Brauer characters of the same degree.
\section{Groups of Lie type in defining characteristic} \label{sec:defchar}
Throughout this section $\bG$ denotes a simple, simply connected linear
algebraic group over the algebraic closure of a finite field and $F:\bG\to\bG$ a
Steinberg endomorphism with finite group of fixed point $\bG^F$.
Any simple group of Lie type can then be obtained as $\bG^F/Z(\bG^F)$ for a
suitable such $\bG$, except for $\tw2F_4(2)'$, for which
Conjecture~\ref{conj:TV} already shown in \cite[Prop.~2.1]{TV19}.
In this section we consider the case where $p$ is the defining characteristic
of $\bG$. Here, Willems observed that the Steinberg character (of $p$-defect
zero) already has large enough degree for Conjecture~\ref{conj:W} to hold. Note,
however, that this result does not imply Conjecture~\ref{conj:TV} for these
groups, since generally there will be more than one block of positive defect.
Here, substantially more work is needed. In fact, it seems that not enough is
currently known about large degree irreducible modular characters in this
situation to derive a complete answer unless $p=2$. We will use the theory of
highest weight representations; a basic introduction can be found for example in
\cite[\S16]{MT}.
\begin{rem} \label{rem:St}
Let $\bG$ be simply connected and $F:\bG\to\bG$ a Frobenius endomorphism with
respect to an ${\mathbb{F}}_p$-structure. Recall that any irreducible representation
of $\bG^F$ over $\overline{{\mathbb{F}}_p}$ is the restriction of a highest weight
representation of $\bG$ with $p$-restricted highest weight. Let
$\{\psi_i\mid i\in I\}$ be a set of irreducible representations of $\bG$ (and
hence of $\bG^F$) with restricted highest weights, all lying over the same
central character $\theta$ of $\bG$ (and hence over the same central character
of $\bG^F$). Further assume that
$\sum_{i\in I}\psi_i(1)^2\ge |\bG^F:(\ker\theta)^F|_{p'}$. Let $\chi$ be the
Steinberg representation of $\bG$ and let $r\ge1$. Then by Steinberg's tensor
product theorem (see, e.g., \cite[Thm.~16.12]{MT}), all
$\psi_{i,r}:=\psi_i\otimes\chi^{(1)}\otimes\cdots\otimes\chi^{(r-1)}$ are
irreducible representations of $\bG^{F^r}$, where $\chi^{(j)}$ denotes the
$j$th Frobenius twist of $\chi$. Since we have $\chi(1)^2>|\bG^F|_{p'}$,
$$\sum_{i\in I}\psi_{i,r}(1)^2\ge |\bG^{F^r}:(\ker\theta)^{F^r}|_{p'}.$$
Thus, if Conjecture~\ref{conj:TV} holds for $\bG^F$, it holds for all
$\bG^{F^r}$, $r\ge1$.
In proving Conjecture~\ref{conj:TV} for $\bG^{F^r}$ in the defining
characteristic, we may therefore restrict attention to the groups defined over
the prime field whenever convenient.
\end{rem}
Let us first deal with some small rank cases.
\begin{prop} \label{prop:A2}
Conjecture~\ref{conj:TV} holds for all central quotients of the groups
$\operatorname{SL}_2(q)$, $\operatorname{SL}_3(q)$, $\operatorname{SU}_3(q)$, $\operatorname{SL}_4(q)$, $\operatorname{SU}_4(q)$ and $\operatorname{Sp}_4(q)$ in
defining characteristic.
\end{prop}
\begin{proof}
Let $G$ be as in the statement. By \cite[Exmp.~2.3(a)]{Wi05} we may restrict our
investigations to the case that $Z(G)\ne1$ and by Remark~\ref{rem:St} we may
assume that $q=p$. For $G=\operatorname{SL}_2(q)$, we may hence assume that $q=p$ is odd. Here
$G$ has a $p$-restricted faithful irreducible Brauer character of degree
$p-1$, and $(p-1)^2\ge(p^2-1)/2=|G/Z(G)|_{p'}$.
\par
Next assume $G=\operatorname{SL}_3(p)$ or $\operatorname{SU}_3(p)$ with $Z(G)$ non-trivial of order~3. For
$1\le i\le p-1$ let $\chi_i$ be the irreducible $p$-Brauer character of $G$ with
(restricted) highest weight $(i,i-1)$. We obtain a lower bound on $\chi_i(1)$ by
adding up the dimensions of certain weight spaces in the corresponding highest
weight module $L(i,i-1)$. According to the result of Premet \cite[Thm.~1]{Pr87},
all weights in the corresponding characteristic~0 irreducible highest weight
module also occur in $L(i,i-1)$. But these are exactly the weights subdominant
to $(i,i-1)$. It is now an easy exercise to determine all these weights and to
find the lengths of their orbits under the Weyl group. This shows that
$\chi_i(1)=\dim L(i,i-1)\ge 3i^2$. Hence
$\sum_{i=1}^{p-1}\chi_i(1)^2\ge (p^2-1)(p^3+1)/3$ by a straightforward
calculation. Thus we conclude, since all characters $\chi_i$ lie above a fixed
faithful character of $Z(\operatorname{SL}_3)$ (see e.g.\ \cite[App.~A.2]{Lue01}).
\par
Next, assume that $G=\operatorname{Sp}_4(p)$. Since $|Z(G)|=(2,p-1)$ we may assume that $p$
is odd. Note that a module with restricted highest weight $(i,j)$ is faithful on
$Z(G)$ if
and only if $j$ is odd. Arguing as in the previous case, one sees that a lower
bound for the dimension of the restricted module with highest weight $(i,2j+1)$
is $2i^2+(8j+6)i+4(j+1)^2$, and the sum of the squares of this over all
$i=0,\ldots,p-1$ and $j=0,\ldots,(p-3)/2$ is larger than $p^6$ and thus larger
than $|G/Z(G)|_{p'}=(p^2-1)(p^4-1)/2$. \par
Finally, for $G=\operatorname{SL}_4(p)$ or $\operatorname{SU}_4(p)$, lower bounds for the weight space
dimensions in restricted representations with highest weight $(m_1,m_2,m_3)$
with $m_i\le p-1$ and $m_1+2m_2+3m_3\equiv1\pmod4$ (for faithful representations
of $G$), respectively $m_1+2m_2+3m_3\equiv2\pmod4$ (for those with a central
subgroup of order~2 in the kernel), yield a sum of squares at least
$(p^2-1)(p^3\pm1)(p^4-1)/(4,p\pm1)$.
\end{proof}
\begin{prop} \label{prop:gen def}
Let $G$ be quasi-simple of Lie type in characteristic~$p$, but not a spin,
half spin or symplectic group, nor of type $\tw{(2)}A_{n-1}$ with $n\le 5$.
Then Conjecture~\ref{conj:TV} holds for $G$ at the prime~$p$.
\end{prop}
\begin{proof}
We may assume that $G$ is not an exceptional covering group of its simple
quotient, since these have centre of order divisible by the characteristic~$p$
(see e.g. \cite[Tab.~24.3]{MT}). Thus, $G$ is a central factor group of a simply
connected group of Lie type. For those, as pointed out by Willems
\cite[Exmp.~2.3(a)]{Wi05}, Conjecture~\ref{conj:TV} holds for groups with
trivial centre. It remains to discuss the groups with non-trivial centre. In
particular we need not worry about Suzuki- and Ree groups. Let $G=\bG^F$, with
$\bG$ simple of simply connected type and $F$ a Frobenius endomorphism of $\bG$
with respect to an ${\mathbb{F}}_q$-structure.
First assume that $G=E_6(q)$ or $\tw2E_6(q)$ with $|Z(G)|=3$. Then $G$ has a
subgroup $H=F_4(q)\times Z(G)$, where the first factor is the centraliser of a
graph automorphism of $G$. Let $\psi\in{{\operatorname{IBr}}}(H)$ be the Steinberg character
of $F_4(q)$ times a faithful character of $Z(G)$. Then $\psi(1)=q^{24}$.
Clearly, $G$ has to have a faithful irreducible $p$-Brauer character $\vhi$ of
at least that degree. But then $\vhi(1)^2\ge q^{48}>|G|_{p'}$. Next, assume
$G=E_7(q)$ with $|Z(G)|=2$. Here, consider the subgroup $H=E_6(q)\times Z(G)$,
and let $\psi\in{{\operatorname{IBr}}}(H)$ be the Steinberg character of $E_6(q)$ times the
faithful linear character of $Z(G)$. This shows that $G$ has an irreducible
$p$-Brauer character $\vhi$ with $\vhi(1)^2\ge\psi(1)^2=q^{72}>|G|_{p'}$.
Next, let $G=\operatorname{SL}_n(q)$. Then $G$ has a subgroup $H=\operatorname{SL}_{n-1}(q)\times Z(G)$.
The Steinberg character of $\operatorname{SL}_{n-1}(q)$ times a linear character of $Z(G)$
has degree $q^{(n-1)(n-2)/2}$, so $G$ has an irreducible $p$-Brauer character
of degree at least that large, while $|G|_{p'}\le q^{(n-1)(n+2)/2}$. Thus, we
are done when $n\ge6$.
For $G=\operatorname{SU}_n(q)$ with $n\ge6$, we again obtain an irreducible $p$-Brauer
character from a Steinberg character $\psi$ of a subgroup
$\operatorname{SU}_{n-1}(q)\times Z(G)$, with $\psi(1)^2=q^{(n-1)(n-2)}$. Now using that
$(q^k-1)(q^{k+1}+1)\le q^{2k+1}$ for all $k\ge2$ we find
$$|G|_{p'}=\prod_{k=2}^n(q^k-(-1)^k)\le q^{(n-1)(n+2)/2},$$
which again is smaller than $\psi(1)^2$ when $n\ge6$.
Since we exclude the spin, half spin and symplectic groups by assumption,
it only remains to discuss the groups $G=\operatorname{SO}_{2n}^\pm(q)'$, $q$ odd, $n\ge4$.
These contain a subgroup $H=\operatorname{SO}_{2n-1}(q)'Z(G)$, and the product of the
Steinberg character of $\operatorname{SO}_{2n-1}(q)'$ with the faithful character of $Z(G)$
yields an irreducible Brauer character of degree $q^{(n-1)^2}$. Thus $G$ itself
also has a faithful character of at least that degree, and its square is larger
than $|G|_{p'}$ when $n\ge4$.
\end{proof}
\begin{prop} \label{prop:class def}
Let $G$ be quasi-simple of classical Lie type in characteristic~$p=2$. Then
Conjecture~\ref{conj:TV} holds for $G$ at the prime $2$.
\end{prop}
\begin{proof}
By the results of Propositions~\ref{prop:A2} and~\ref{prop:gen def} we only
need to consider the groups $\operatorname{SL}_n(q)$, $\operatorname{SU}_n(q)$ for $n=5$ and $q=2^f$, as
the spin, half spin and symplectic groups have trivial centre in
characteristic~2.
For $\operatorname{SL}_5(q)$ the tensor product of the natural module with $f-1$ twists of the
1024-dimensional restricted Steinberg module yields an irreducible
representation of sufficiently large degree when $f\ge3$. For $f\le2$ the
centre of $\operatorname{SL}_5(2^f)$ is trivial. For the unitary groups, we can argue
similarly for $f\ge3$. For $f=2$ we take the tensor product of the Steinberg
character with the exterior square of the natural representation, and for
$\operatorname{SU}_5(2)$ the centre is trivial.
\end{proof}
We now state some further partial results.
\begin{lem} \label{lem:Spn}
Let $G=\operatorname{Sp}_{2n}(q)$ with $q=p^f$. Then Conjecture~\ref{conj:TV} holds for $G$
at $p$ if either $n\ge5$ or $f\ge2$.
\end{lem}
\begin{proof}
Recall that we only need to consider the faithful characters of $G$.
The symplectic group $\operatorname{Sp}_{2n+2}(q)$ has a subsystem subgroup
$\operatorname{Sp}_{2n}(q)\times\operatorname{Sp}_2(q)$, and its image in $H:=\operatorname{PSp}_{2n+2}(q)$ is a central
product $G_1:=\operatorname{Sp}_{2n}(q)\circ\operatorname{Sp}_2(q)$ containing a subgroup $\operatorname{Sp}_{2n}(q)$.
Now the Steinberg character of $H$ of degree $q^{(n+1)^2}$ has $p$-defect zero,
so yields an irreducible $p$-Brauer character of $H$ of that degree.
Hence the subgroup $G_1$ has to have a faithful irreducible Brauer character
of degree at least $q^{(n+1)^2}/|H:G_1|$. Since the faithful irreducible Brauer
characters of $\operatorname{Sp}_2(q)$ have degree at most $q-1$, $G$ must have a faithful
irreducible Brauer character of degree at least $q^{(n+1)^2}/((q-1)|H:G_1|)$. But
the square of this is at least $|G|_{p'}$ whenever $n\ge5$. When $n\le4$ but
$f\ge2$, we may take the tensor product of that character at $q=p$ with twists
of the Steinberg character.
\end{proof}
Thus, with Proposition~\ref{prop:A2} among symplectic groups $\operatorname{Sp}_{2n}(q)$ only
the faithful block for $n=3,4$ and $q=p$ odd remains open.
\begin{prop}
Let $G$ be a covering group of a simple orthogonal group of rank $n\ge3$
over ${\mathbb{F}}_q$, and assume that $q=p^f$ with $f\ge2$. Then
Conjecture~\ref{conj:TV} holds for all covering groups of $G$ at $p$ unless
possibly when $G=\operatorname{Spin}_8^-(q)$, where it holds when $f\ge3$.
\end{prop}
\begin{proof}
By Proposition~\ref{prop:class def} it remains to discuss the spin groups and
half spin groups for odd~$q$. Also, $\operatorname{Spin}_{4n}^+(q)$ has non-cyclic centre, so
we need not consider it. First assume that
$G=\operatorname{Spin}_{2n}^+(p)$ with $n\ge3$ odd. Let $H$ be the preimage in $G$ of the
stabiliser $\operatorname{GL}_n(p)$ of a totally isotropic subspace for $\operatorname{SO}_{2n}^+(p)$. As
$n$ is odd, $\operatorname{PSL}_n(p)$ has odd Schur multiplier, so the Steinberg character of
$\operatorname{SL}_n(p)$ extends to a faithful $p$-Brauer character $\psi$ of $H$, of degree
$p^{n(n-1)/2}$. Thus, $G$ has a faithful irreducible character of at least that
degree. Then, by Steinberg's tensor product theorem, $\operatorname{Spin}_{2n}^+(q)$ has a
faithful irreducible Brauer character of degree at least $(q/p)^{n^2}\psi(1)$,
which is larger than $|G|_{p'}$ when $f\ge2$. Now $\operatorname{Spin}_{2n+1}(p)$ contains a
subgroup $\operatorname{Spin}_{2n}^+(p)$ whose centre lies in the centre of $\operatorname{Spin}_{2n+1}(p)$.
As the Steinberg character of $\operatorname{Spin}_{2n+1}(p)$ has degree $p^{n^2}$, this
shows that $\operatorname{Spin}_{2n+1}(q)$ has a character as claimed (where still $n$ is odd).
Unless $n=3$ the same argument also applies to $\operatorname{HSpin}_{2n+2}^+(q)$,
$\operatorname{Spin}_{2n+2}^-(q)$, $\operatorname{Spin}_{2n+3}(q)$ and $\operatorname{Spin}_{2n+4}^-(q)$.
The group $\operatorname{HSpin}_8^+(q)$ is isomorphic to $\operatorname{SO}_8^+(q)$ by the triality
automorphism and thus by Proposition~\ref{prop:gen def} has a faithful
Brauer character of degree $q^9$. Using this, we also obtain our claim for
$\operatorname{Spin}_9(q)$ and $\operatorname{Spin}_{10}^-(q)$.
\end{proof}
\section{The non-defining characteristic case} \label{sec:crosschar}
We now turn to groups of Lie type $G$ in cross characteristic. That is,
we assume that $p$ is not the defining characteristic of $G$. Here, Willems
\cite[Thm.~3.1]{Wi05} obtained certain asymptotic results on
Conjecture~\ref{conj:W}, which were improved upon for several families of groups
in the thesis of Maslowski \cite{Ma05}. Nonetheless, both of these fall short of
proving Conjecture~\ref{conj:TV} in the case at hand.
We will again argue using suitable characters of $p$-defect zero, as considered
by Willems \cite{Wi88}, but in order to obtain complete results we need to show
that there exist sufficiently many of these.
Let $\bG$ be a simple algebraic group and $F:\bG\to\bG$ a Steinberg
endomorphism. The regular semisimple elements are known to be dense in $\bG$, so
we certainly expect a large proportion of regular semisimple classes in the
finite group $\bG^F$. Here, motivated by our application in character theory,
we quantify this expectation by giving a (rather weak) lower bound for their
number.
\begin{prop} \label{prop:reg elt}
Let $(\bG,F)$ be as above, with
$$\bG^F\ne \operatorname{SU}_4(2),\ \operatorname{Sp}_4(2),\ \operatorname{Sp}_6(2),\ \operatorname{O}_8^+(2),\ \operatorname{O}_8^-(2),\
\operatorname{SU}_3(3).$$
Then a lower bound for the number $n_{{\operatorname{reg}}}(T)$ of conjugacy classes of regular
semisimple conjugacy classes meeting certain maximal tori $T$ of $\bG^F$ is
bounded below as given in Table~$\ref{tab:bnd}$ for classical groups, and in
Table~$\ref{tab:tori exc}$ for exceptional type groups.
\end{prop}
\begin{proof}
The proof is independent of the isogeny type of $\bG$. Let $q$ be the absolute
value of the eigenvalues of $F$ on the character group of an $F$-stable maximal
torus of $\bG$. We will use the following
well-known result of Zsigmondy: for every integer $e\ge3$ and prime power $q$
there exists a prime $z_e(q)$ dividing $q^e-1$ but no $q^m-1$ for $1\le m<e$,
unless $(e,q)=(6,2)$. Since $q$ has multiplicative order $e$ modulo $z_e(q)$, we
see that $z_e(q)\ge e+1$.
\par
For each group $G=\bG^F$ of classical type we have given in Table~\ref{tab:bnd}
(the order of) two maximal tori $T\le G$. These two tori have been chosen such
that the greatest common divisor of their orders is exactly the order of the
centre of the simply connected group $G_{{\operatorname{sc}}}$ of that type, which is also the
order of the commutator factor group of the adjoint group of that type.
Furthermore, the image of $T$ in $G_{{\operatorname{sc}}}/Z(G_{{\operatorname{sc}}})$ is cyclic, except for the
first two tori listed for type $D_n$. For each torus the table also gives an
integer $e$ such that $|T|$ is divisible by a Zsigmondy prime $r:=z_e(q)$,
whenever that exists, that is, if $e\ge3$ and $(e,q)\ne(6,2)$. In all cases
listed in the table, the description of Sylow $p$-subgroups in
\cite[Thm.~25.14]{MT} together with the order formula for $G$ (see
\cite[Tab.~24.1]{MT}) shows that the Sylow $r$-subgroups of $G$ are cyclic.
Moreover, by the parametrisation of maximal tori via conjugacy classes of the
Weyl group (see \cite[Prop.~25.3]{MT}), the conjugates of $T$
are the only maximal tori of $G$ containing elements of order $r$, except for
the torus of order $(q^{n-1}-1)(q+1)$ in types $B_n,C_n$. In particular, apart
from that latter case, all elements of $T$ of order divisible by $r$ are
necessarily regular. Now $T$ certainly contains at least $|T|(r-1)/r$ such
elements.
Since $r\ge e+1$, we thus find at least $|T|e/(e+1)$ regular elements in~$T$.
In the case of the torus $T$ of order $(q^{n-1}-1)(q+1)$, all products of
elements of order $r$ by an element of order at least~3 in the factor of order
$q+1$ are regular, and thus there exist at least $(q^{n-1}-1)(q-1)e/(e+1)$
regular elements in $T$.
The fusion of elements in any maximal torus is controlled by its normaliser,
so the number $n_{{\operatorname{reg}}}(T)$ of regular conjugacy classes in $G$ with
representatives in $T$ is at least $|N_G(T):T|^{-1}$ times the number of
regular elements in $T$. These are exactly the entries in the last column of
Table~\ref{tab:bnd}.
\begin{table}[ht]
\caption{Lower bounds in classical groups} \label{tab:bnd}
$$\begin{array}{c|cccc}
G& |T|& e& |N_G(T):T|& n_{{\operatorname{reg}}}(T)\ge\cr
\hline\hline
A_1(q)& q+1& 2& 2& (q-1)/2\\
& q-1& 1& 2& (q-3)/2\\
\hline
A_{n-1}(q),\ n\ge3& (q^n-1)/(q-1)& n& n& |T|/(n+1)\\
& q^{n-1}-1& n-1& n-1& |T|/n\\
\hline
\tw2A_{n-1}(q),& (q^n+1)/(q+1)& 2n& n& 2|T|/(2n+1)\\
3\le n\equiv1\,(2)& q^{n-1}-1& n-1& n-1& |T|/n\\
\hline
\tw2A_{n-1}(q),& q^{n-1}+1& 2n-2& n-1& 2|T|/(2n-1)\\
4\le n\equiv0\,(2)& (q^n-1)/(q+1)& n& n& |T|/(n+1)\\
\hline
B_n(q),C_n(q),\ n\ge2& q^n+1& 2n& 2n& |T|/(2n+1)\\
n=2:& q^2-1& 2& 4& (q-1)(q-2)/4\\
4\le n\equiv0\,(2):& (q^{n-1}-1)(q+1)& n-1& 4n-4& (q^{n-1}-1)(q-1)/(4n)\\
n\equiv1\,(2):& q^n-1& n& 2n& |T|/(2n+2)\\
\hline
D_n(q),\ n\ge4& (q^{n-1}+1)(q+1)& 2n-2& 2n-2& |T|/(2n-1)\\
n\equiv0\,(2):& (q^{n-1}-1)(q-1)& n-1& 2n-2& |T|/(2n)\\
n\equiv1\,(2):& q^n-1& n& n& |T|/(n+1)\\
\hline
\tw2D_n(q),\ n\ge4& q^n+1& 2n& n& 2|T|/(2n+1)\\
& (q^{n-1}+1)(q-1)& 2n-2& 2n-2& |T|/(2n-1)\\
\end{array}$$
\end{table}
We now discuss the cases in Table~\ref{tab:bnd} when there do not always exist
Zsigmondy primes. For $G=A_1(q)$ a maximal torus of order $q\pm1$ contains at
least $q\pm1-\gcd(2,q-1)$ regular elements and hence representatives from
$(q\pm1-\gcd(2,q-1))/2$ regular conjugacy classes. For $G=A_2(q)$, it can be
seen by direct
calculation that a maximal torus of order $q^2-1$ intersects $(q^2-q)/2\ge |T|/3$
regular classes, as stated in the table. For $G=\tw2A_2(q)$, a maximal torus of
order $q^2-1$ meets $(q+1)(q-2)/2$ regular classes, which is smaller than
$|T|/3$ only when $q=2,3$. In the first case, $G$ is solvable, the second was
excluded. For $G=B_2(q)$ a maximal torus of order $q^2-1$ meets $(q-1)(q-2)/4$
regular semisimple classes, as stated in Table~\ref{tab:bnd}.
To complete the discussion of Zsigmondy exceptions, we finally consider the
cases when $(e,q)=(6,2)$ in Table~\ref{tab:bnd}. This concerns the following
groups:
$$\begin{array}{c|cccccccc}
G& \operatorname{SL}_6(2)& \operatorname{SL}_7(2)& \operatorname{SU}_4(2)& \operatorname{SU}_6(2)& \operatorname{SU}_7(2)& \operatorname{Sp}_6(2)& \operatorname{O}_8^+(2)& \operatorname{O}_8^-(2)\\
\hline
|T|& 63& 63& 9& 21& 63& 9& 27& 9\\
n_{{\operatorname{reg}}}(T)& 9& 9& 2& 3& 9& 1& 3& 1\\
\end{array}$$
The numbers $n_{{\operatorname{reg}}}(T)$ can be read off from the known character tables.
For $\operatorname{SL}_6(2)$, $\operatorname{SL}_7(2)$, $\operatorname{SU}_6(2)$ and $\operatorname{SU}_7(2)$ this agrees with the
numbers given in Table~\ref{tab:bnd}, while the other groups are listed as
exceptions.
We now turn to the groups of exceptional type, for which the arguments are
very similar. In Table~\ref{tab:tori exc} for each type we give a maximal torus
$T$, respectively two tori for $G$ of type $E_7$. In all cases, $T$ is cyclic
by \cite[Thm.~25.14]{MT} and for all groups different from $E_7(q)$, all
elements of $T$ of order not dividing $|Z(G_{{\operatorname{sc}}})|$ are regular by the order
formula for $G$ (see \cite[Tab.~24.1]{MT}). For $E_7(q)$, the maximal tori
are Coxeter tori of maximal rank subgroups of type $\tw2A_7(q)$, $A_7(q)$
respectively, and all of their elements not lying in the subgroup of order
$q\pm1$ are regular.
\begin{table}[ht]
\caption{Lower bounds in exceptional groups} \label{tab:tori exc}
$$\begin{array}{l|ccc}
G& |T|& |N_G(T):T|& n_{{\operatorname{reg}}}(T)\cr
\hline
\!\tw2B_2(q^2),\ q^2\ge8& \Phi_8''& \ 4& (|T|-1)/4\cr
\!\!^2G_2(q^2),\ q^2\ge27& \Phi_{12}''& \ 6& (|T|-1)/6\cr
G_2(q),\,q\equiv1\,(3)& \Phi_6& \ 6& (|T|-1)/6\cr
\hphantom{G_2(q),\ }q\not\equiv1\,(3)& \Phi_3& \ 6& (|T|-1)/6\cr
\!\tw3D_4(q)& \Phi_{12}& \ 4& (|T|-1)/4\cr
\!\tw2F_4(q^2),\ q^2\ge8& \Phi_{24}''& 12& (|T|-1)/12\cr
F_4(q)& \Phi_{12}& 12& (|T|-1)/12\cr
E_6(q)& \Phi_9& \ \,9& (|T|-(3,q-1))/9\cr
\!\tw2E_6(q)& \Phi_{18}& \ \,9& (|T|-(3,q+1))/9\cr
E_7(q)& \Phi_2\Phi_{14}& 14& (q^7-q)/14\cr
& \Phi_1\Phi_7& 14& (q^7-q)/14\cr
E_8(q)& \Phi_{24}& 24& (|T|-1)/24\cr
\end{array}$$
Here $\Phi_8''=q^2+\sqrt{2}q+1$, $\Phi_{12}''=q^2+\sqrt{3}q+1$,
$\Phi_{24}''=q^4+\sqrt{2}q^3+q^2+\sqrt{2}q+1$.
\end{table}
Each regular element in $T$ is conjugate to $|N_G(T):T|$ elements of $T$. Since
torus normalisers control fusion of their elements, we obtain the stated lower
bounds for the number $n_{{\operatorname{reg}}}(T)$ of $G$-conjugacy classes of regular semisimple
elements meeting $T$.
\end{proof}
Note that the total number of semisimple classes for $G$ of simply connected
type was shown by Steinberg to be equal to $q^l$, where $l$ is the rank of
$\bG$. With a lot more effort it would be possible to derive estimates for the
number of regular semisimple classes that are asymptotically much closer to
$q^l$.
We note the following easy consequence, which has been used in the proof of
a recent result by Sambale \cite[Thm.~10]{Sa20}:
\begin{cor} \label{cor:sambale}
Let $G$ be a simple group of Lie type in characteristic $p$. Then there exist
at least two conjugacy classes of elements of $G$ with centraliser order prime
to $p$.
\end{cor}
\begin{proof}
Let $\bG$ be simple of simply connected type with a Steinberg endomorphism $F$
such that $G=\bG^F/Z(\bG^F)$. This is possible unless $G=\tw2F_4(2)'$, for which
the claim is easily verified. Let $x\in\bG^F$ be regular semisimple. Then its
centraliser is a maximal torus, of order prime to $p$. Thus it suffices to
show that $\bG^F$ has at least two conjugacy classes of regular semisimple
elements which do not have the same image in $G$. For exceptional type groups
$G$ it is immediate that the numbers $n_{{\operatorname{reg}}}(T)$ in Table~\ref{tab:tori exc}
are all strictly bigger than $|Z(\bG^F)|$, and so we obtain at least two
regular semisimple $G$-classes. For the groups of classical type, we take
regular semisimple classes coming from the two different types of maximal tori
given in Table~\ref{tab:bnd}. For the exceptions in
Proposition~\ref{prop:reg elt} the claim is again readily verified.
\end{proof}
\begin{thm} \label{thm:class cross}
Conjecture~\ref{conj:TV} holds for quasi-simple groups of Lie type for
non-defining primes $p$.
\end{thm}
\begin{proof}
By \cite[Prop.~2.1]{TV19} the claim holds for the exceptional covering groups,
thus we may assume that $G$ is a central quotient $G=\bG^F/Z$ of a quasi-simple
group $\bG^F$ of simply connected Lie type. We will construct sufficiently
many irreducible Deligne--Lusztig characters of $G$ of $p$-defect zero using
Proposition~\ref{prop:reg elt}. First, the assertion is readily checked from
the known Brauer character tables for the six groups listed as exceptions in
that result. So we may assume that $G$ is not a covering group of one of
those. As discussed in the proof of Proposition~\ref{prop:reg elt}, for $G$ of
classical type, the two maximal tori of $\bG^F$ in Table~\ref{tab:bnd} are such
that their images in $\bG^F/Z(\bG^F)$ have coprime orders. Thus, we may choose
at least one of them, say $T$, with $|T|_p=|Z(G)|_p$ for our given prime~$p$.
The same applies to groups of type $E_7$. For the remaining groups of
exceptional type, all elements of $T\setminus Z(\bG^F)$ are regular and the
corresponding Sylow subgroups of $G$ are cyclic. Since by
Lemma~\ref{lem:cyclic} we need not consider primes $p$ for which Sylow
subgroups are cyclic, we again have that $|T|_p=|Z(G)|_p$ for the tori $T$ in
Table~\ref{tab:tori exc}.
\par
Let $\bT\le\bG$ be an $F$-stable maximal torus with $\bT^F=T$, and let
$(\bT^*,\bG^*,F)$ be dual to $\bT,\bG,F)$ (see \cite[Def.~1.5.17]{GM20}), so
$T^*:=\bT^{*F}$ has the same order as $T$. To each regular element $s\in T^*$,
up to $G^*$-conjugation, there exists an irreducible Deligne--Lusztig character
$\pm R_s$ of~$\bG^F$ of degree $|\bG^F:T|_{q'}$ (see \cite[Def.~2.5.17 and
Thm.~2.2.12]{GM20}). Hence, by the choice of $T$, $R_s$ is of central
$p$-defect, that is, the defect of $\chi$ equals the $p$-part of the centre
of $\bG^F$. So the $p$-modular reduction of $R_s$ is irreducible (see e.g.
\cite[Thm.~(9.13)]{Na98}). Now by the character formula
\cite[Prop.~2.2.18]{GM20}, $R_s$ is a faithful character of $G=\bG^F/Z$ with
$Z=\ker(\theta)\cap Z(\bG^F)$, where $s$ is in duality with $(\bT,\theta)$, for
$\theta\in{{\operatorname{Irr}}}(T)$. By construction the Zsigmondy prime $r$ for $T$ occurring in
the proof of Proposition~\ref{prop:reg elt} does not divide
$|\bG^{*F}:[\bG^{*F},\bG^{*F}]|$. So we may choose a system of representatives
$R$ for the cosets of $T^*\cap[\bG^{*F},\bG^{*F}]$ in $T^*$ consisting of
elements of order prime to $r$. Thus, if $s\in T^*$ has order divisible by~$r$
then so has $st$, for $t\in R$. It follows that the regular elements in $T^*$ of
order divisible by $r$ are distributed equally across the various cosets of
$T^*\cap[\bG^{*F},\bG^{*F}]$ in $T^*$. Thus, the irreducible Deligne--Lusztig
characters of central $p$-defect constructed above contribute
$$n_{{\operatorname{reg}}}(T)/|Z|\cdot|\bG^F:T|_{q'}^2$$
to the sum in Conjecture~\ref{conj:TV}. So, we need to see that
$$n_{{\operatorname{reg}}}\ge \frac{|G|_{q}\cdot |T|^2}{|G|_{q'}\cdot|P|},\eqno{(*)}$$
where $P$ denotes a Sylow $p$-subgroup of $G$.
For this, we estimate $|P|$. Since, as pointed out before, we only need to
consider
primes $p$ for which Sylow $p$-subgroups of $G$ are non-cyclic, $p$ divides
at least two (not necessarily distinct) cyclotomic factors occurring in the
order formula for $G$ (see \cite[Thm.~25.14]{MT}). Note that if $p>2$ and $q$
has order $d$ modulo~$p$, then $p\ge d+1$, while for $p=2$, at least one of
$q-1$, $q+1$ is divisible at least by~4. Again first assume that $\bG$ is of
classical type, say of rank~$n$. A Sylow $p$-subgroup $P$ of $G$ has order at
least~$p^2$. As the Weyl group $W$ of $\bG$ contains a symmetric group $\fS_n$,
this shows $|P|\ge(n+1)^2$ if $p$ does not divide the order of $W$. If $p\le n$,
then $|P|\ge p^{\lfloor n/p\rfloor+p}$, which is still at least $(n+1)^2$ unless
$p=2$. But for $p=2$ we have $|P|\ge 2^{\lfloor n/2\rfloor+4}$, and it follows
that $|P|\ge(n+1)^2$ in all cases.
For groups of exceptional type, from the order formula for $G$ (see
\cite[Tab.~24.1]{MT}) it is easy to check that the order of the smallest
non-cyclic Sylow $p$-subgroups is at least~25, and even at least~121 when
$G=E_8(q)$.
Now comparing $(*)$ with the right hand side of the formula in
Conjecture~\ref{conj:TV} we see that our conclusion follows except when $G$ is
one of $\operatorname{SL}_3(2),\operatorname{Sp}_4(2)$, for which the claim is easily checked directly,
or $G=\operatorname{Sp}_{2n}(q)$ or $\operatorname{Spin}_{2n+1}(q)$ with $n=4$ and $q\le5$, $n=6$ and
$q\le4$, $n=8,10$ with $q\le3$ or $(n,q)=(12,2)$, and in all cases, $p$ divides
the order $q^n+1$ of the first torus from Table~\ref{tab:bnd}. Among these, the
only cases in which Sylow $p$-subgroups of $G$ are non-cyclic are for
$$(n,q,p)\in\{(6,2,5),(6,3,5),(6,4,17),(10,2,5),(10,3,5),(12,2,17)\}.$$
In all these cases except for $(n,p,q)=(6,2,5)$, the desired inequality holds
when we use the exact order of $P$ (which is at least $p^3$ here). Finally, for
$\operatorname{Sp}_{12}(2)$, the ordinary character table allows us to see that the sum of
squares of 5-defect zero characters exceeds the bound.
\end{proof}
\section{Alternating groups} \label{sec:alt}
It remains to discuss the covering groups of simple alternating groups. At
present we only see how to treat the prime $p=2$.
\begin{thm} \label{thm:alt}
Let $G$ be a covering group of the alternating group $\fA_n$, $n\ge5$. Then
Conjecture~\ref{conj:TV} holds for $G$ at the prime~$2$.
\end{thm}
\begin{proof}
By Proposition~\ref{prop:normal} it is sufficient to consider the covering
groups of the symmetric groups. The exceptional covering groups of $\fA_6$ and
$\fA_7$ are easily checked, so we may assume that $n\ge8$ and so $Z(G)=1$, that
is $G=\fS_n$. We will show that for large enough $n$ there exists a single
irreducible 2-Brauer character whose degree satisfies the desired inequality.
In fact, as for the groups of Lie type, it is the 2-modular reduction of an
ordinary character, namely, a faithful character of the 2-fold cover $2.\fS_n$,
but not of 2-defect zero. \par
For $l\ge1$ consider the two partitions
$$\begin{aligned}
p_1&=(4l-3,4l-7,\ldots,1)\vdash n_{l,1}:=l(2l-1),\\
p_2&=(4l-1,4l-5,\ldots,3)\vdash n_{l,2}:=l(2l+1).
\end{aligned}$$
These label irreducible characters $\chi_l^i$ of the 2-fold cover
$2.\fS_{n_{l,i}}$ of $\fS_{n_{l,i}}$. Their degree is given by the analogue of
the hook length formula (see \cite[Thm.~2.8]{Fa18}): for a partition
$\la=(\la_1,\la_2,\ldots,\la_m)\vdash n$ with distinct parts $\la_i$, the degree
of the corresponding spin character of $2.\fS_n$ equals
$$2^{\lfloor(n-m)/2\rfloor}\frac{n!}{\prod_i\la_i!}\cdot
\prod_{i<j}\frac{\la_i-\la_j}{\la_i+\la_j}.$$
From this one easily computes that
$$\frac{\chi_{l+1}^1(1)}{\chi_l^1(1)}
=2^{4l-1}\binom{n_{l+1,1}}{4l+1}\binom{4l-1}{2l}^{-1}\quad\text{and}\quad
\frac{\chi_{l+1}^2(1)}{\chi_l^2(1)}
=2^{4l+1}\binom{n_{l+1,2}}{4l+3}\binom{4l+1}{2l+1}^{-1}.$$
We claim that for $l\ge8$ we have
$$\chi_l^1(1)^2 \ge (n_{l,2}-1)!_{2'}\qquad\text{and}\qquad
\chi_l^2(1)^2 \ge (n_{l+1,1}-1)!_{2'}\,.\eqno{(*)}
$$
Assume that the inequalities have already been shown up to $l$. Now
$$\binom{4l-1}{2l}=\frac{1}{2}\binom{4l}{2l}\le 2^{4l-1}/\sqrt{4l}\quad\text{and}\quad
\binom{4l+1}{2l+1}=\frac{1}{2}\binom{4l+2}{2l+1}\le 2^{4l+1}/\sqrt{4l+2},$$
by the standard estimate for the middle binomial coefficient, and so
$$\chi_{l+1}^1(1)^2\ge 4l\binom{n_{l+1,1}}{4l+1}^2(n_{l,2}-1)!_{2'},
\qquad\chi_{l+1}^2(1)^2\ge(4l+2)\binom{n_{l+1,2}}{4l+3}^2(n_{l+1,1}-1)!_{2'}$$
by our inductive assumption. Further, with $n:=n_{l+1,1}$ we have
$$\begin{aligned}
\binom{n}{4l+1}
&\ge c_1\Big(\frac{n}{4l+1}\Big)^{4l+1}\Big(\frac{n}{n-4l-1}\Big)^{n-4l-1}
\sqrt{\frac{n}{2\pi(4l+1)((n-4l-1)}}\\
&\ge c_1\sqrt\frac{1}{2\pi(4l+1)}\Big(1+\frac{4l+1}{n-4l-1}\Big)^{n-4l-1}
\Big(\frac{n}{4l+1}\Big)^{4l+1}\\
&\ge c_2\sqrt\frac{1}{2\pi(4l+1)}\ e^{4l+1}\Big(\frac{n}{4l+1}\Big)^{4l+1}
\end{aligned}$$
for some constants $c_1,c_2\ge0.5$ independent of $l$ by a well-known estimate
for binomial coefficients. On the other hand by Stirling's formula
$$\begin{aligned}
\frac{(n+2l+1)!}{(n-2l-2)!}
&\le c_3 \frac{\sqrt{2\pi(n+2l+1)}}{\sqrt{2\pi(n-2l-2)}}\cdot
\frac{(n+2l+1)^{n+2l+1}\,e^{n-2l-2}}{(n-2l-2)^{n-2l-2}\,e^{n+2l-1}}\\
&=c_4 e^{-4l-3}\Big(1+\frac{4l+3}{n-2l-2}\Big)^{n-2l-2}(n+2l+1)^{4l+3}\\
&\le c_4(n+2l+1)^{4l+3},
\end{aligned}$$
where again $c_3,c_4\le2$ are independent of $l$. Putting things together,
this shows that
$$\frac{\chi_{l+1}^1(1)^2}{(n_{l+1,2}-1)!_{2'}}
\ge \frac{c_2^2}{c_4} 4l\frac{e^{8l+2}}{2\pi(4l+1)}\Big(\frac{n}{4l+1}\Big)^{8l+2}
\Big(\frac{1}{n+2l+1}\Big)^{4l+3}\frac{(n_{l+1,2}-1)!_2}{(n_{l,2}-1)!_2}.
$$
Now, using that $n=n_{l+1,1}=(l+1)(2l+1)$ we find
$$\frac{n^2}{(4l+1)^2(n+2l+1)}\ge \frac{1}{8},$$
and since $(n_{l+1,2}-1)!_2/(n_{l,2}-1)!_2\ge 2^{4l+3}/(4l+3)$,
$$\frac{\chi_{l+1}^1(1)^2}{(n_{l+1,2}-1)!_{2'}}
\ge c_5 \frac{4l}{(4l+1)(4l+3)} \Big(\frac{1}{n+2l+1}\Big)^{2}
\Big(\frac{e^2}{4}\Big)^{4l-1}.$$
The last term on the right hand side clearly dominates when $l\to\infty$ and so
the left hand side is eventually bigger than~1. A more
precise estimate and checking the first 50 values by computer then completes
the proof of our claim for $\chi_l^1$. A very similar computation shows~(*)
for $\chi_l^2$.
\par
The desired assertion then follows: by a result of Fayers \cite[Thm.~3.3]{Fa18},
both $\chi_l^1$ and $\chi_l^2$ remain irreducible modulo~2. So for all
$n\ge n_{l+1,i}$, with $l\ge8$, there exists an irreducible 2-Brauer character
of $\fS_n$ of degree at least $\chi_{l+1}^i(1)$. By (*), $\chi_{l+1}^1(1)$
satisfies the inequality from Conjecture~\ref{conj:TV} for $\fS_n$ for all
$n_{l+1,1}\le n\le n_{l+1,1}+2l+1=n_{l+1,2}-1$, and similarly,
$\chi_{l+1}^2(1)$ satisfies our inequality for $\fS_n$ whenever
$n_{l+1,2}\le n\le n_{l+2,1}-1$. Finally, for $l<8$, that is, for $n<120$ it is
easily checked using the
criterion in \cite[Thm.~3.3]{Fa18} that there always is an irreducible
2-Brauer character with the desired property, labelled by a sum of one of
our partitions $\la_{l,1}$ or $\la_{l,2}$ and a suitable Carter partition.
\end{proof}
It should be noted that we need to use both characters $\chi_l^1$ and $\chi_l^2$
for our approach to work.
For odd primes, the combinatorics seem more daunting, and furthermore, there
are two quite different cases to consider, corresponding to faithful and
non-faithful characters of $2.\fS_n$. Note that by Lemma~\ref{lem:cyclic}, we may
assume that $p\le n/2$ for $\fS_n$ and $2.\fS_n$.
\medskip
Our Main Theorem~\ref{thm:main} now follows by combining
Proposition~\ref{prop:class def}, Theorem~\ref{thm:class cross} and
Theorem~\ref{thm:alt} with the result of Tong-Viet \cite[Prop.~2.1]{TV19} for
sporadic groups.
| a62c23bea158d3d2e90face2b2490b0adc06704b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Understanding the orbital magnetization in crystalline solids is among the
most important objectives of orbitronics \cite{Zhang2005,Go2017,Bhowal2020}.
Unlike its spin companion, the orbital magnetization of Bloch electrons in the
absence of external perturbations is already hard to access quantum
mechanically, as it does not correspond to a bounded operator. This
nonlocality is finally accounted for by a Berry phase formula
\cite{Xiao2005,Resta2005,Shi2007} that gives significantly distinct orbital
magnetization from the atom-centered approximation when combined with
\textit{ab-initio} calculations in various magnetic materials
\cite{Ceresoli2010,Lopez2012,Hanke2016}. In the presence of external driving
electric fields, the extrinsic responses of spin and orbital magnetization,
namely the spin and orbital Edelstein effects, are given similarly by the
magnetic moments averaged over current-carrying states
\cite{Edelstein1990,Murakami2015,Mak2017,Pesin2018,Salemi2019,Mertig2020}. In
contrast, the intrinsic responses of them, i.e., the spin and orbital
magnetoelectric effects, are completely different due to the nonlocal nature
of the magnetic dipole operator. Specifically, the spin magnetoelectricity
\cite{Garate2009,Garate2010} is dictated by a Berry curvature following the
ubiquitous character of intrinsic linear responses of a local operator
\cite{Dong2020}, while the orbital one consists of a Chern-Simons three form
and a perturbative term of the reciprocal-space Berry connection
\cite{Moore2010,Vanderbilt2010,Lee2011,Gao2014}. Besides, the magnetization
pumped by periodic adiabatic processes in band insulators has been studied
recently by a density matrix approach evaluating the time-averaged expectation
value of the spin and magnetic dipole operators \cite{Luka2019,Murakami2020}.
In this approach, the orbital magnetization can only be obtained in the
Wannier basis \cite{Luka2019}, in contrast to the spin one that can be
evaluated in the Bloch representation.
Up to date, the orbital magnetization induced by electric fields and by
periodic adiabatic processes are treated by different theories. Whether both
phenomena have a deep connection and if they can emerge in a unified
theoretical framework are still unknown. In this work, we develop a
semiclassical theory for the magnetization induced by adiabatic
evolutions of Bloch electronic states. In general, the adiabatically induced
orbital magnetization is gauge dependent due to the presence of the electric
current in the second Chern form of Berry curvatures. Noticeably, the orbital
magnetoelectric effect and the periodic-evolution pumped orbital magnetization
emerge as the only instances where the induced magnetization is
gauge invariant due to the elimination of its explicit time dependence. Our work
thus renders a unified theory of both phenomena in insulators with vanishing
Chern numbers. Besides, unlike the Chern-Simons contribution deduced from the
second Chern form current, the induced magnetization due to the perturbed
Berry connection is well defined irrespective of Chern invariants and of
insulators or metals. As a result, the orbital magnetoelectricity in
two-dimensional (2D) metals and Chern insulators, which had long been
hard to approach, is also attained in our theory.
We apply our theory to illustrate the orbital magnetization pumped by
microscopic atomic rotations, which correspond to phonon modes with angular
momentum \cite{Zhang2014,Garanin2015}, in toy models based on the honeycomb
lattice. The results are comparable to the pumped spin magnetization via a
strong Rashba spin orbit coupling \cite{Murakami2020}. We also show the vital
role played by the electric-field induced orbital magnetization in the
nonlinear intrinsic anomalous Ettingshausen effect in 2D metallic systems. In
particular, the Mott relation is validated in intrinsic nonlinear transport by
subtracting the magnetization component of the thermal current in the second
order of the electric field.
This paper is organized as follows. In Sec. II we lay out the semiclassical
theory of Bloch electrons, which is employed to study the adiabatically
induced orbital magnetization in metals and insulators, respectively, in Sec.
III and Sec. IV. A case study of the orbital magnetization pumped by local
rotations of atoms and the application of the orbital magnetoelectricity to
the intrinsic nonlinear anomalous Ettingshausen effect are shown in Sec. V,
followed by a summary in Sec. VI. Some technical details of the theory are
presented in Appendices A and B for the convenience of interested readers.
\section{Semiclassical theory}
In the semiclassical description, the Hamiltonian felt by a narrow wave packet
centered around position $\boldsymbol{r}_{c}$ is $\hat{H}=\hat{H}_{c}+\hat
{H}^{\prime}$ in the first order gradient expansion, where $\hat{H}_{c}%
=\hat{H}_{0}\left( \boldsymbol{\hat{r}},\boldsymbol{\hat{p}};w\left(
\boldsymbol{r}_{c}\right) ,R\left( t\right) \right) +\sum_{\alpha
}\boldsymbol{\hat{\theta}}^{\left( \alpha\right) }\cdot\boldsymbol{h}%
^{\left( \alpha\right) }\left( \boldsymbol{r}_{c},t\right) \ $is the local
Hamiltonian and $\hat{H}^{\prime}=\frac{1}{2}\{\left( \hat{r}-r_{c}\right)
_{i},\frac{\partial\hat{H}_{c}}{\partial r_{ci}}\}$ is the gradient correction
\cite{Sundaram1999}. The Einstein summation convention is implied for repeated
Cartesian indices $i$, $j$ and $s$ henceforth. $\hat{H}_{0}$ is the local
approximation of the genuine Hamiltonian. The most general $\hat{H}_{0}$
considered here includes $w\left( \boldsymbol{r}_{c}\right) $ that
represents possible nonuniform static mechanical fields varying slowly on the
scale of the wave packet as well as $R\left( t\right) $ serving as a
parameter whose time evolution is adiabatic. Besides, in order to implement
the variational approach to obtain the local density of a bounded observable,
we add the auxiliary term $\sum_{\alpha}\boldsymbol{\hat{\theta}}^{\left(
\alpha\right) }\cdot\boldsymbol{h}^{\left( \alpha\right) }\left(
\boldsymbol{r},t\right) $ into the Hamiltonian and expand it around
$\boldsymbol{r}_{c}$. Here $\boldsymbol{\hat{\theta}}^{\left( \alpha\right)
}$ ($\alpha=1,2,..$) is a set of bounded observable operators, each of which
is assumed to be a vector for simple notations, without losing generality.
$\boldsymbol{h}^{\left( \alpha\right) }\left( \boldsymbol{r},t\right) $
denotes the conjugate slowly varying external fields, and $\boldsymbol{h}%
^{\left( \alpha\right) }\left( \boldsymbol{r}_{c},t\right) $ changes
adiabatically in the parameter space $\left( t,\boldsymbol{r}_{c}\right) $.
At the end of the calculation the auxiliary term is set to zero, i.e.,
$\boldsymbol{h}^{\left( \alpha\right) }=0$.
In the semiclassical theory that is accurate to the first order of spatial
gradients and of time derivative \cite{Sundaram1999,Xiao2010}, the wave packet
$|W_{n}\left( \boldsymbol{k},\boldsymbol{r}_{c},t\right) \rangle
=\sum_{\boldsymbol{p}}C_{n\boldsymbol{p}}|\psi_{n\boldsymbol{p}}\rangle$ is
constructed by superposing the local Bloch states $|\psi_{n\boldsymbol{p}%
}\left( \boldsymbol{r}_{c},t\right) \rangle=e^{i\boldsymbol{p}%
\cdot\boldsymbol{\hat{r}}}|u_{n\boldsymbol{p}}\left( \boldsymbol{r}%
_{c},t\right) \rangle$ of $\hat{H}_{c}$. Here $n$ and $\hbar\boldsymbol{p}$
are the band index and crystal momentum, respectively, and the coefficient
$C_{n\boldsymbol{p}}$ is sharply distributed around the wave vector
$\boldsymbol{k}$ of the wave packet, obeying $\sum_{\boldsymbol{p}}\left\vert
C_{n\boldsymbol{p}}\right\vert ^{2}=1$. To simplify notations, all the band
quantities without explicit band index are considered for band $n$, unless
otherwise noted. Throughout this study we consider nondegenerate bands to
simplify the analysis and assume they are so separated that adiabatic
evolutions are feasible. The wave packet Lagrangian reads (set $\hbar=1$)
\begin{equation}
L=\langle W|i\frac{d}{dt}-\hat{H}|W\rangle=\boldsymbol{\dot{r}}_{c}%
\cdot\boldsymbol{k}+\boldsymbol{\dot{k}}\cdot\boldsymbol{\mathcal{A}}%
^{k}+\boldsymbol{\dot{r}}_{c}\cdot\boldsymbol{\mathcal{A}}^{r_{c}}%
+\mathcal{A}^{t}-\tilde{\varepsilon},\label{L}%
\end{equation}
where $\boldsymbol{\mathcal{A}}^{k/r_{c}}=\langle u_{n}|i\partial
_{\boldsymbol{k/r}_{c}}\mathbf{|}u_{n}\rangle$\ and $\mathcal{A}^{t}=\langle
u_{n}|i\partial_{t}\mathbf{|}u_{n}\rangle$ are the Berry connections derived
from the periodic part $|u_{n}\left( \boldsymbol{k},\boldsymbol{r}%
_{c},t\right) \rangle$ of the Bloch wave \cite{Sundaram1999}. The
noncanonical form of the Lagrangian due to Berry connections implies that
$\left( \boldsymbol{r}_{c},\boldsymbol{k}\right) $\ are not canonical
variables, thus the measure of the phase space spanned by $\left(
\boldsymbol{r}_{c},\boldsymbol{k}\right) $ should be modified, with the
result \cite{Xiao2005} $\mathcal{D}=1+\Omega_{ii}^{kr_{c}}$. Here $\Omega
_{ii}^{kr_{c}}$ is the trace of the Berry curvature $\Omega_{ij}^{kr_{c}%
}\equiv\partial_{k_{i}}\mathcal{A}_{j}^{r_{c}}-\partial_{r_{cj}}%
\mathcal{A}_{i}^{k}$, and other Berry curvatures are formed similarly. The
wave-packet energy is given by $\tilde{\varepsilon}=\varepsilon+\delta
\varepsilon$ up to first order gradients, where $\varepsilon_{n}\left(
\boldsymbol{k},\boldsymbol{r}_{c},t\right) $ is the local Bloch energy,
$\delta\varepsilon_{n}=\operatorname{Re}\sum_{n_{1}\neq n}%
\boldsymbol{\mathcal{A}}_{nn_{1}}^{k}\cdot(\partial_{\boldsymbol{r}_{c}}%
\hat{H}_{c})_{n_{1}n}$ and $\boldsymbol{\mathcal{A}}_{nn_{1}}^{k}=\langle
u_{n}|i\partial_{\boldsymbol{k}}\mathbf{|}u_{n_{1}}\rangle$ is the interband
Berry connection.
When going beyond the above first-order theory, the wave packet is no longer
dictated only by the Bloch states of local Hamiltonian $\hat{H}_{c}$ but is
modified by $\hat{H}^{\prime}$ up to the linear order of spatial gradients
\cite{Gao2014}. In the well established second-order theory \cite{Gao2019} the
inhomogeneity appears only in electromagnetic gauge potentials, whereas the
following results account for weak inhomogeneities of mechanical fields
conjugate to general bounded operators. This generalization is necessary to
obtain the adiabatically induced orbital magnetization carried by Bloch
electrons by calculating the magnetization current, which is manifested only
in nonuniform systems \cite{Cooper1997,Xiao2020EM}. For this purpose, it is
sufficient to retain results up to the order of the product spatial and time derivatives.
In the second-order theory the wave packet takes the form of \cite{Gao2019}
\begin{equation}
|W_{n}\left( \boldsymbol{k},\boldsymbol{r}_{c},t\right) \rangle
=\sum_{\boldsymbol{p}}e^{i\boldsymbol{p}\cdot\boldsymbol{\hat{r}}%
}[C_{n\boldsymbol{p}}|u_{n\boldsymbol{p}}\rangle+\sum_{n_{1}\neq n}%
C_{n_{1}\boldsymbol{p}}|u_{n_{1}\boldsymbol{p}}\rangle],
\end{equation}
where (derivations presented in Appendix A)
\begin{equation}
C_{n_{1}}=\frac{\Delta_{n_{1}n}C_{n}}{\varepsilon_{n\boldsymbol{p}%
}-\varepsilon_{n_{1}\boldsymbol{p}}}-i\boldsymbol{\mathcal{A}}_{n_{1}n}%
^{r_{c}}\cdot(i\partial_{\boldsymbol{p}}+\boldsymbol{\mathcal{A}}_{n}%
^{p}-\boldsymbol{r}_{c})C_{n}\label{C}%
\end{equation}
incorporates spatial-gradient induced corrections. Here
\begin{align}
\Delta_{n_{1}n} & \equiv\frac{i}{2}(\partial_{\boldsymbol{\hat{p}}}%
\cdot\partial_{\boldsymbol{r}_{c}}\hat{H}_{c})_{n_{1}n}+\sum_{n_{2}\neq
n}(\partial_{\boldsymbol{r}_{c}}\hat{H}_{c})_{n_{1}n_{2}}\cdot
\boldsymbol{\mathcal{A}}_{n_{2}n}^{p}\nonumber\\
& -\boldsymbol{\mathcal{A}}_{n_{1}n}^{r_{c}}\cdot\boldsymbol{v}%
_{n}\label{delta}%
\end{align}
has the dimension of energy, with $\boldsymbol{\mathcal{A}}_{nn_{1}}^{r_{c}%
}=\langle u_{n}|i\partial_{\boldsymbol{r}_{c}}\mathbf{|}u_{n_{1}}\rangle$. The
wave-packet center appearing in Eq. (\ref{C}) is determined by $\boldsymbol{r}%
_{c}=\langle W|\boldsymbol{\hat{r}}|W\rangle=\partial_{\boldsymbol{k}}%
\gamma(\boldsymbol{k})+\boldsymbol{\mathcal{A}}^{k}+\boldsymbol{a}^{k}$, where
$-\gamma(\boldsymbol{p})$ is the phase of $C_{n\boldsymbol{p}}$. While the
first two terms appear already in the first order theory \cite{Sundaram1999},
the third term $\boldsymbol{a}_{n}^{k}\equiv2\operatorname{Re}\sum
_{\boldsymbol{p}}\sum_{n_{1}\neq n}C_{n\boldsymbol{p}}^{\ast}C_{n_{1}%
\boldsymbol{p}}\boldsymbol{\mathcal{A}}_{nn_{1}}^{p}$ is the inhomogeneity
induced positional shift of the wave-packet center. Gathering the above
results we get a gauge invariant expression
\begin{equation}
\boldsymbol{a}_{n}^{k}=2\operatorname{Re}\sum_{n_{1}\neq n}\frac
{\boldsymbol{\mathcal{A}}_{nn_{1}}^{k}\Delta_{n_{1}n}}{\varepsilon
_{n}-\varepsilon_{n_{1}}}-\partial_{\boldsymbol{k}}\cdot\mathcal{G}_{n}%
^{r_{c}k},\label{ak}%
\end{equation}
with $\mathcal{G}_{n}^{r_{c}k}=\operatorname{Re}\sum_{n_{1}\neq n}%
\boldsymbol{\mathcal{A}}_{nn_{1}}^{r_{c}}\boldsymbol{\mathcal{A}}_{n_{1}n}%
^{k}$ being the quantum metric tensor in $\left( \boldsymbol{r}%
_{c},\boldsymbol{k}\right) $ space.
One can then find that the wave-packet Lagrangian takes the same form [Eq.
(\ref{L})] as in the first order theory, but the involved Berry connections
are modified by inhomogeneity, i.e., $\boldsymbol{\mathcal{\tilde{A}}}%
^{k}=\boldsymbol{\mathcal{A}}^{k}+\boldsymbol{a}^{k}$ and $\mathcal{\tilde{A}%
}^{t}=\mathcal{A}^{t}+a^{t}$. The corrected Berry connections in fact take
similar structures, e.g.,%
\begin{equation}
a_{n}^{t}=2\operatorname{Re}\sum_{n_{1}\neq n}\frac{\mathcal{A}_{nn_{1}}%
^{t}\Delta_{n_{1}n}}{\varepsilon_{n}-\varepsilon_{n_{1}}}-\partial
_{\boldsymbol{k}}\cdot\boldsymbol{\mathcal{G}}_{n}^{r_{c}t},\label{at}%
\end{equation}
where $\boldsymbol{\mathcal{G}}_{n}^{r_{c}t}=\operatorname{Re}\sum_{n_{1}\neq
n}\boldsymbol{\mathcal{A}}_{nn_{1}}^{r_{c}}\mathcal{A}_{n_{1}n}^{t}$, and are
gauge invariant. Note that the correction to $\boldsymbol{\mathcal{A}}^{r_{c}%
}$ is not needed, since $\boldsymbol{\mathcal{A}}^{r_{c}}$ is already in the
first order of spatial gradients. Meanwhile, the wave-packet energy does not
receive further corrections at the order of the product spatial and time derivatives.
Having identified the wave-packet Lagrangian and the concomitant action $S$,
one gets directly the semiclassical dynamics of Bloch electrons following from
the Euler-Lagrange equation in $\left( \boldsymbol{r}_{c},\boldsymbol{k}%
\right) $ space. Furthermore, one can consider the local density of a bounded
observable $\boldsymbol{\hat{\theta}}$, of which the conjugate external field
is marked by $\boldsymbol{h}$, contributed by a Bloch-electron ensemble with
the occupation function $f_{n}\left( \boldsymbol{r}_{c},\boldsymbol{k}%
,t\right) $. The general recipe for this has been given recently as
\cite{Dong2020}
\begin{equation}
\boldsymbol{\theta}\left( \boldsymbol{r},t\right) =-\int\left[
d\boldsymbol{k}\right] d\boldsymbol{r}_{c}\mathcal{D}f\frac{\delta S}%
{\delta\boldsymbol{h}\left( \boldsymbol{r},t\right) }|_{\boldsymbol{h}%
\rightarrow0},\label{Key}%
\end{equation}
where $\left[ d\boldsymbol{k}\right] \equiv\sum_{n}d\boldsymbol{k}%
/(2\pi)^{d}$ with $d$ as the spatial dimensionality. In what follows we
suppress the notation $\boldsymbol{h}\rightarrow0$ but the results for various
adiabatic responses are calculated in this limit. We take $f(\tilde
{\varepsilon})$ as the Fermi distribution in order to focus on adiabatic
intrinsic contributions determined solely by band structures.
The spin magnetoelectricity and spin pumping by periodic adiabatic processes
can be readily obtained from the above formula in the first order of time
derivative, as detailed in Appendix B, while the orbital counterparts can only
be acquired through calculating the electric current, which is much more
involved and is elaborated in the next two sections.
\section{Orbital magnetization in metals}
\subsection{Nonlinear electric current}
In order to address the orbital magnetization induced by adiabatic evolutions,
we need to formulate the local charge current density up to the order of the
product spatial and time derivatives. To achieve this Eq. (\ref{Key}) is
considered in the case of $\boldsymbol{\hat{\theta}}=e\boldsymbol{\hat{v}}$
being the charge current operator, hence $\boldsymbol{h}=-\boldsymbol{A}$ is
the electromagnetic vector potential with a minus sign, which enters through
the minimal coupling, resulting in the chain rule $\partial_{-\boldsymbol{A}%
}=e\partial_{\boldsymbol{k}}$. We still use $\boldsymbol{k}$ to denote the
gauge invariant crystal momentum. In the following the center label $c$ is
suppressed, unless otherwise noted. After some manipulations, as shown in
Appendix B, we arrive at (hereafter $\int$ without integral variable is
shorthand for $\int\left[ d\boldsymbol{k}\right] $, $f^{0}=f\left(
\varepsilon\right) $)
\begin{align}
\boldsymbol{j}\left( \boldsymbol{r},t\right) & =\boldsymbol{\nabla}%
\times(\boldsymbol{M}^{0}+\int f^{0}\partial_{\boldsymbol{B}}a^{t}%
)+\partial_{t}\int f^{0}e\boldsymbol{a}^{k}\nonumber\\
& -e\int f^{0}(\Omega^{\boldsymbol{k}t}+\Omega_{s}^{k\left[ kr\right]
t}\boldsymbol{\hat{e}}_{s})\nonumber\\
& -e\int\partial_{\varepsilon}f^{0}\delta\varepsilon\Omega^{\boldsymbol{k}%
t}+e\int(\partial_{\boldsymbol{k}}f^{0}a^{t}-\partial_{t}f^{0}\boldsymbol{a}%
^{k}),\label{general}%
\end{align}
where the second Chern form of the Berry curvature \cite{Xiao2009,Zhou2013} is
labeled as
\begin{equation}
\Omega_{s}^{k\left[ kr\right] t}\equiv\Omega_{si}^{kk}\Omega_{i}^{rt}%
+\Omega_{si}^{kr}\Omega_{i}^{tk}+\Omega_{s}^{kt}\Omega_{ii}^{kr},
\end{equation}
$\boldsymbol{\hat{e}}_{s}$ is the unit vector in the $s$ direction, and
\begin{equation}
\boldsymbol{M}^{0}=\int(f^{0}\boldsymbol{m}+eg^{0}\boldsymbol{\Omega
}).\label{M-eq}%
\end{equation}
Here $\boldsymbol{m}_{n}=\frac{e}{2}\sum_{n_{1}\neq n}\boldsymbol{\mathcal{A}%
}_{nn_{1}}^{k}\times\boldsymbol{v}_{n_{1}n}$ is the orbital moment of a Bloch
electron, $\boldsymbol{\Omega}$ is the vector form of Berry curvature
$\Omega_{ij}^{kk}$, and $g^{0}=-\int_{\varepsilon}^{\infty}f(\eta)d\eta$ is
the grand potential density contributed by a Bloch electron. With the
symmetric gauge for the uniform magnetic field, Eq. (\ref{at}) gives
\begin{equation}
\partial_{\boldsymbol{B}}a^{t}=2\operatorname{Re}\sum_{n_{1}\neq n}%
\frac{-\mathcal{A}_{nn_{1}}^{t}\boldsymbol{m}_{n_{1}n}}{\varepsilon
_{n}-\varepsilon_{n_{1}}}+\frac{e}{2}\partial_{\boldsymbol{k}}\times
\boldsymbol{\mathcal{G}}_{n}^{kt},\label{at-1}%
\end{equation}
where $\boldsymbol{\mathcal{G}}_{n}^{kt}=\operatorname{Re}\sum_{n_{1}\neq
n}\boldsymbol{\mathcal{A}}_{nn_{1}}^{k}\mathcal{A}_{n_{1}n}^{t}$ is the
quantum metric in $\left( \boldsymbol{k},t\right) $ space, and
$\boldsymbol{m}_{n_{1}n}=-\partial_{\boldsymbol{B}}\Delta_{n_{1}n}$, which
will be elaborated later in combination with a more specific physical context.
Equation (\ref{general}) is the pivotal result of this paper. The first line
is of total spatial and time derivatives, hence is certainly intimately
related to the orbital magnetization and electric polarization. Apparently,
$\boldsymbol{M}^{0}$ is the magnetization that relies solely on instantaneous
electronic states and corresponds to the equilibrium orbital magnetization in
the static case \cite{Shi2007}. On the other hand, the magnetization and
polarization may not be determined by the first line of Eq. (\ref{general})
alone, as the second line can be relevant as well. This line consists of first
and second Chern forms of Berry curvatures, which underline various electronic
topological responses of insulators \cite{Xiao2009,Xiao2010,Qi2008,Essin2009}.
Besides, the last line signifies intrinsic Fermi-surface contributions to the
charge current density in metals, which are beyond the conventional Boltzmann
transport picture of conductors \cite{Ziman} and are distinct from intrinsic
Fermi-sea contributions to linear response.
Now we are in a position to compare Eq. (\ref{general}) with existing results
at the same order. The second line of this equation has been formulated in
inhomogeneous insulators \cite{Xiao2009,Qi2015} and metals \cite{Hayata2017}.
The specific case where the inhomogeneity enters only through the magnetic
vector potential has been studied in insulators with degenerate bands
\cite{Moore2010,Qi2008} and in metals \cite{Moore2016}. Meanwhile, these
pioneering studies disregarded the magnetization current in the first line of
Eq. (\ref{general}), especially the orbital magnetization induced by the Berry
connection $a^{t}$ due to adiabatic time evolutions [Eq. (\ref{at-1})].
However, there is an important physical context: orbital magnetoelectricity in
2D metals, which is contributed entirely by this gauge invariant term and
hence is beyond the scope of the aforementioned theories. We discuss this
subject shortly.
\subsection{Orbital magnetoelectricity in 2D}
To address the orbital magnetoelectricity, we consider the case that the
adiabatic time dependence stems entirely from the vector potential, i.e.,
$\boldsymbol{E}=-\partial_{t}\boldsymbol{A}$, with $\boldsymbol{E}$ being a
weak constant electric field, then $\partial_{t}=e\boldsymbol{E}\cdot
\partial_{\boldsymbol{k}}$, $a^{t}=\boldsymbol{a}^{k}\cdot e\boldsymbol{E}$
and $\Omega_{s}^{k\left[ kr\right] t}=\Omega_{sj}^{k\left[ kr\right]
k}eE_{j}$. Thus the local charge current density [Eq. (\ref{general})] reduces
to%
\begin{align}
\boldsymbol{j}\left( \boldsymbol{r}\right) & =\boldsymbol{\nabla}%
\times\lbrack\boldsymbol{M}^{0}+\int f^{0}\partial_{\boldsymbol{B}%
}(e\boldsymbol{E\cdot a}^{k})]\nonumber\\
& -e^{2}\boldsymbol{E}\times\int f^{0}\boldsymbol{\Omega}-e^{2}E_{i}\int
f^{0}\Omega_{si}^{k\left[ kr\right] k}\boldsymbol{\hat{e}}_{s}\nonumber\\
& +e^{2}\boldsymbol{E}\times\int\partial_{\varepsilon}f^{0}(\boldsymbol{v}%
\times\boldsymbol{a}^{k}-\delta\varepsilon\boldsymbol{\Omega}%
).\label{j-static}%
\end{align}
To understand this current we first inspect the case when the spatial
dependence originates only from the vector potential, i.e., $\boldsymbol{B}%
=\boldsymbol{\nabla}\times\boldsymbol{A}$. Then it is apparent that%
\begin{equation}
\boldsymbol{j}=-e^{2}\boldsymbol{E}\times\int\{f^{0}\boldsymbol{\Omega
}-\partial_{\varepsilon}f^{0}[\boldsymbol{v}\times(\boldsymbol{a}%
^{k})^{\boldsymbol{B}}+\left( \boldsymbol{m}\cdot\boldsymbol{B}\right)
\boldsymbol{\Omega}]\},\label{NLH}%
\end{equation}
where $(\boldsymbol{a}^{k})^{\boldsymbol{B}}$\ is proportional to the magnetic
field. This result recovers the intrinsic magneto-nonlinear Hall current of
order $EB$ that was obtained previously by a different method \cite{Gao2014}.
On the other hand, to identify the orbital magnetization one could introduce
the spatial dependence from other inhomogeneous external mechanical fields. By
doing so one may expect that the local current density in bulk can be
decomposed into a transport and a magnetization component, namely
\cite{Cooper1997}%
\begin{equation}
\boldsymbol{j}=\boldsymbol{j}^{\text{tr}}+\boldsymbol{\nabla}\times
\boldsymbol{M},\label{decomposition}%
\end{equation}
where the transport current $\boldsymbol{j}^{\text{tr}}$ contributes to the
net flow through the sample. In 2D the second Chern form current is enforced
to vanish due to $\Omega_{si}^{k\left[ kr\right] k}=0$, hence Eq.
(\ref{decomposition}) is rescued with $\boldsymbol{j}^{\text{tr}}$ taking the
same form as the above magneto-nonlinear Hall current Eq. (\ref{NLH}) and%
\begin{equation}
\boldsymbol{M}=\boldsymbol{M}^{0}+\int f^{0}\partial_{\boldsymbol{B}%
}(e\boldsymbol{E\cdot a}^{k}).\label{E-induced M}%
\end{equation}
Recall that $\boldsymbol{a}^{k}$ is physically a positional shift of a
semiclassical Bloch electron, the electric work upon this shift implies
immediately an orbital magnetization. This electric-field induced
magnetization is in agreement with what is obtained recently by a different
method \cite{Xiao2020OM}, but the present derivation is much simpler, even
though starting from a more generic framework. In 2D, $\boldsymbol{M}$ is a
pseudoscalar and is well defined irrespective of metals or Chern insulators.
Owing to the gauge invariance of $\boldsymbol{a}^{k}$, it is legitimate to
define the orbital magnetoelectric susceptibility contributed by each Bloch
electron in 2D $\alpha_{ij}^{\text{o}}$ via
\begin{equation}
\partial M_{j}/\partial E_{i}=\int f^{0}\alpha_{ij}^{\text{o}}. \label{2D OM}%
\end{equation}
$\alpha_{ij}^{\text{o}}$ takes a gauge invariant form ($j=z$ in 2D)%
\begin{equation}
\alpha_{ij}^{\text{o}}=-2e\operatorname{Re}\sum_{n_{1}\neq n}\frac
{(\mathcal{A}_{i}^{k})_{nn_{1}}(m_{j})_{n_{1}n}}{\varepsilon_{n}%
-\varepsilon_{n_{1}}}+\frac{e^{2}}{2}\epsilon_{jls}\partial_{k_{l}%
}(\mathcal{G}_{si}^{kk})_{n}, \label{OMP}%
\end{equation}
with $\mathcal{G}_{n}^{kk}=\operatorname{Re}\sum_{n_{1}\neq n}%
\boldsymbol{\mathcal{A}}_{nn_{1}}^{k}\boldsymbol{\mathcal{A}}_{n_{1}n}^{k}$ as
the $\boldsymbol{k}$-space quantum metric \cite{QM2020}.
It is interesting to compare $\alpha_{ij}^{\text{o}}$ with the spin
magnetoelectric susceptibility contributed by each Bloch electron, which is
given by the first term of Eq. (\ref{OMP}) with $\boldsymbol{m}_{n_{1}n}$
replaced by the interband elements of spin magnetic moment. Since
$\boldsymbol{m}_{n_{1}n}=-\frac{e}{2}\sum_{n_{2}\neq n}(\boldsymbol{v}%
_{n_{1}n_{2}}+\delta_{n_{2}n_{1}}\boldsymbol{v}_{n})\times
\boldsymbol{\mathcal{A}}_{n_{2}n}^{k}$ reduces to the familiar orbital moment
$\boldsymbol{m}_{n}$ when $n_{1}=n$, it can be deemed as an interband orbital
magnetic moment. Despite this similarity between spin and orbital
magnetoelectric susceptibility, the distinction is apparent: the
$\boldsymbol{k}$-space dipole moment of the quantum metric $\partial_{k_{l}%
}\mathcal{G}_{si}^{kk}$ does not have a counterpart in spin
magnetoelectricity. Noticeably, for two-band metallic systems with
particle-hole symmetry, the first term of $\alpha_{ij}^{\text{o}}$ vanishes,
hence $\alpha_{ij}^{\text{o}}=\frac{e^{2}}{2}\epsilon_{jls}\partial_{k_{l}%
}\mathcal{G}_{si}^{kk}$ is given solely by the quantum metric dipole, which is
an intrinsic Fermi surface effect.
Before closing this section, we note that in 3D insulators with nonvanishing
$\boldsymbol{k}$-space\ Chern invariants or 3D metals, the second Chern form
current in Eq. (\ref{j-static}) obviously poses a difficulty in pursuing a
gauge invariant decomposition in the form of Eq. (\ref{decomposition}). This
difficulty raises the question as to whether the electric-field induced
orbital magnetization can be defined as a bulk quantity in such systems. At
the present stage this is still an open question
\cite{Chen2012,Bergman2011,Qi2011} and is left for future efforts.
\section{Orbital magnetization in non-Chern insulators}
Now we turn to the nonlinear electric current in insulators in the general
case of adiabatic time evolutions and spatial dependence, under the assumption
of vanishing Chern numbers in all the pertinent parameter spaces. Great
simplifications of Eq. (\ref{general}) occur in insulators. First, the
Fermi-surface terms vanish and the Fermi-sea ones are contributed by fully
occupied bands. Then, according to the antisymmetric decomposition of the
second Chern form%
\begin{equation}
\Omega_{s}^{k\left[ kr\right] t}=\partial_{k_{s}}CS_{ii}^{tkr}%
-\partial_{k_{i}}CS_{si}^{tkr}-\partial_{r_{i}}CS_{is}^{tkk}-\partial
_{t}CS_{sii}^{kkr}, \label{CS}%
\end{equation}
where the involved Chern-Simons\ three forms read, e.g., $CS_{si}^{tkr}%
=\frac{1}{2}(\mathcal{A}^{t}\Omega_{si}^{kr}+\mathcal{A}_{s}^{k}\Omega
_{i}^{rt}+\mathcal{A}_{i}^{r}\Omega_{s}^{tk})$, $CS_{is}^{tkk}=\frac{1}%
{2}(\mathcal{A}^{t}\Omega_{is}^{kk}+\mathcal{A}_{i}^{k}\Omega_{s}%
^{kt}+\mathcal{A}_{s}^{k}\Omega_{i}^{tk})$ and $CS_{sii}^{kkr}=\frac{1}%
{2}(\mathcal{A}_{s}^{k}\Omega_{ii}^{kr}+\mathcal{A}_{i}^{k}\Omega_{is}%
^{rk}+\mathcal{A}_{i}^{r}\Omega_{si}^{kk})$, the current density takes the
form of%
\begin{equation}
\boldsymbol{j}\left( \boldsymbol{r},t\right) =\boldsymbol{\nabla}%
\times\boldsymbol{M}\left( \boldsymbol{r},t\right) +\partial_{t}%
\boldsymbol{P}\left( \boldsymbol{r},t\right) . \label{current-insulator}%
\end{equation}
Here we have taken the $\boldsymbol{k}$-space periodic gauge for Bloch wave
functions, and
\begin{align}
\boldsymbol{M} & =\boldsymbol{M}^{0}+\int(\partial_{\boldsymbol{B}}%
a^{t}-e\frac{1}{2}\epsilon_{lis}CS_{is}^{tkk}\boldsymbol{\hat{e}}%
_{l}),\label{M}\\
\boldsymbol{P} & =e\int(\boldsymbol{\mathcal{A}}^{k}+\boldsymbol{a}%
^{k}+CS_{sii}^{kkr}\boldsymbol{\hat{e}}_{s}) \label{P}%
\end{align}
can be deemed as the orbital magnetization and polarization induced,
respectively, by the adiabatic time evolution and spatial inhomogeneity.
One can tell from Eq. (\ref{general}) that the perturbative contribution
$\partial_{\boldsymbol{B}}a^{t}$ to the orbital magnetization is well defined
regardless of Chern numbers in $\left( \boldsymbol{k},t\right) $ space and
is invariant under a gauge transformation of Bloch wave functions (a phase
transformation is compatible with the $\boldsymbol{k}$-space periodic gauge).
In contrast, the Chern-Simons orbital magnetization deduced from the second
Chern form current is only well defined in insulators with vanishing $\left(
\boldsymbol{k},t\right) $-space Chern numbers. It changes under the gauge
transformation. It can be readily shown that this gauge dependence is
permitted by the inherent degrees of freedom of $\boldsymbol{M}\left(
\boldsymbol{r},t\right) $ and $\boldsymbol{P}\left( \boldsymbol{r},t\right)
$ determined by the invariance of the local current density Eq.
(\ref{current-insulator}) \cite{Hirst1997} (e.g., in the 2D case the inherent
degrees of freedom of $\boldsymbol{M}$ and $\boldsymbol{P}$ are $M_{z}%
\boldsymbol{\hat{z}}\rightarrow M_{z}\boldsymbol{\hat{z}}-\partial_{t}%
\chi\boldsymbol{\hat{z}}\ $and$\ \boldsymbol{P}\rightarrow\boldsymbol{P}%
+\boldsymbol{\nabla}\times(\chi\boldsymbol{\hat{z}})$, with a scalar field
$\chi(\boldsymbol{r},t)$).
This gauge dependence also implies, on the other hand, the necessity of
removing the time dependence of the orbital magnetization and the spatial
dependence of the electric polarization if one would like to pursue gauge
invariant definitions of them. Therefore, there are generally two ways to have
a gauge invariant orbital magnetization: to either eliminate the explicit time
dependence of $\boldsymbol{M}$ or pursue the definition upon an average over
time. These two approaches correspond to two important physical contexts --
orbital magnetoelectric response and orbital magnetization pumping --\ that
are addressed separately in the following two subsections.
\subsection{Orbital magnetoelectric response}
When the time and spatial dependence concerns only the electromagnetic gauge
potentials, the explicit time dependence of $\boldsymbol{M}$ and spatial
dependence of $\boldsymbol{P}$ are removed due to the minimal coupling. This
is the case of the orbital magnetoelectric response in insulators, which
includes two dual effects: a constant electric (magnetic) field induces an
orbital magnetization (electric polarization). Most previous derivations are
designed for only one of the two dual effects
\cite{Moore2010,Vanderbilt2010,Gao2014,Lee2011}, while a theory capable of
both simultaneously is rare \cite{Sipe2020}. Here we show that they are
readily derived from the present theory.
On the one hand, when the time dependence appears solely as $\boldsymbol{E}%
=-\partial_{t}\boldsymbol{A}$, one has $\int CS_{is}^{tkk}=\frac{-\theta
e}{4\pi^{2}}\epsilon_{isj}E_{j}$, where $\theta=-\int\frac{d^{3}k}{4\pi
}\boldsymbol{\mathcal{A}}^{k}\cdot\boldsymbol{\Omega}$ is the abelian version
of the so called $\theta$-term \cite{Xiao2010,Essin2009}. Then $\boldsymbol{M}%
$ [Eq. (\ref{M})] becomes a time-independent orbital magnetization%
\begin{equation}
\boldsymbol{M}=\boldsymbol{M}^{0}+\partial_{\boldsymbol{B}}\int\boldsymbol{a}%
^{k}\cdot e\boldsymbol{E}+\frac{e^{2}}{4\pi^{2}}\theta\boldsymbol{E}.
\label{OME}%
\end{equation}
On the other hand, when the spatial dependence appears only as a magnetic
field, $\int CS_{sii}^{kkr}=\frac{\theta e}{4\pi^{2}}B_{s}$. Thus one can
identify $\boldsymbol{P}$ as a uniform polarization, which is in agreement
with the previous theory \cite{Gao2014}, and verify $\partial M_{i}/\partial
E_{j}=\partial P_{j}/\partial B_{i}$.
\subsection{Periodic-evolution pumped orbital magnetization}
It is also possible to define, based on $\boldsymbol{M}$ and $\boldsymbol{P}$,
the time averaged orbital magnetization in time periodic systems and spatially
averaged polarization in spatially periodic systems. Here we concentrate on
the magnetization, and the polarization can be discussed similarly. If the
adiabatic time dependence of the electronic Hamiltonian is periodic with
period $T$ and the Chern invariants in $\left( \boldsymbol{k},t\right) $
space are zero, then the time averaged $\boldsymbol{M}$ is gauge invariant and
can be perceived as the orbital magnetization pumped by the periodic
evolution, namely
\begin{equation}
\boldsymbol{\bar{M}}\left( \boldsymbol{r}\right) =\int_{0}^{T}\frac{dt}%
{T}\boldsymbol{M}\left( \boldsymbol{r},t\right) .
\end{equation}
We use the notation $\boldsymbol{\bar{M}}$ to distinguish from the
instantaneous magnetization $\boldsymbol{M}\left( \boldsymbol{r},t\right) $
in Eq. (\ref{current-insulator}). In 2D the $\boldsymbol{\nabla}\phi$ degree
of freedom of the magnetization is irrelevant and the so defined
$\boldsymbol{\bar{M}}\left( \boldsymbol{r}\right) $ gives the orbital
magnetization unambiguously. In both 2D and 3D this $\boldsymbol{\bar{M}}%
$\ coincides with the abelian version of the so-called geometric orbital
magnetization obtained by a density matrix approach evaluating the
time-averaged expectation value of the magnetic dipole operator in the Wannier
basis in homogeneous band insulators \cite{Luka2019}. On the other hand, the
present theory does not invoke the Wannier basis and accounts also for weak
inhomogeneous systems.
\section{Applications}
\subsection{Model illustration of orbital magnetization pumped by local
rotations of atoms}
To illustrate the above theory, we consider a minimal model for the orbital
magnetization due to the periodic adiabatic evolution of electronic states
induced by microscopic local rotations of atoms. Such a model is not required
to possess the spin-orbit coupling, in contrast to the spin magnetization
induced by local circulations of atoms that is only possible with the aid of
spin-orbit coupling \cite{Murakami2020}. The minimal spatial dimensionality
for rotational motions is two, and the model should have a gap. Moreover, the
second Chern form current can be nonzero only if the dimension of the
Hamiltonian is larger than two \cite{Xiao2009}. Therefore, we here consider a
two-band model hence focus exclusively on the contribution
\begin{equation}
\boldsymbol{\bar{M}}=\int_{0}^{T}\frac{dt}{T}\int\frac{d\boldsymbol{k}}%
{(2\pi)^{2}}\partial_{\boldsymbol{B}}a^{t}\label{OM pumping}%
\end{equation}
from the perturbed Berry connection. According to the expression for
$\partial_{\boldsymbol{B}}a^{t}$ (only the first term of Eq. (\ref{at-1})
matters in insulators), one can easily verify that it can be nonvanishing in a
two-band model\ only if the particle-hole symmetry is broken.
\begin{figure}[ptb]
\includegraphics[width=9 cm]{Murakami2band_PhononMag.pdf}\caption{Model
illustration of the orbital magnetization of Bloch electrons induced by the
microscopic local rotation of atoms. (a) Band structure of a spinless-graphene
toy model with first and second nearest neighbor hoppings. (b) Time dependence
of the adiabatically induced magnetic moment in units of $e\omega a^{2}$ per
unit cell. Here $\omega$ is the angular frequency of the atomic rotation, $a$
is the lattice constant, and we take the parameters as $\Delta=0.2t_{0}$,
$t_{1}=0.1t_{0}$ and $\delta t_{0}=0.1t_{0}$. The insert shows that the pumped
orbital magnetization in one period of the local rotations of atoms is
proportional to the next nearest neighbor hopping parameter. (c)
$\boldsymbol{k}$-space\ distribution of $\partial_{\boldsymbol{B}}a^{t}$ of
the valence band. (d) Gap dependence of the pumped orbital magnetization.}%
\label{MurakamiSpinless}%
\end{figure}
Such a minimal model can thus be chosen as a spinless graphene-type one based
on the honeycomb lattice taking into account the next nearest neighbor
hopping, which is described by the Hamiltonian
\begin{equation}
\hat{H}(\boldsymbol{k})=t_{0}(F_{\mathrm{R}}\sigma_{x}-F_{\mathrm{Im}}%
\sigma_{y})+\Delta\sigma_{z}+t_{1}F_{\mathrm{NN}}\sigma_{0},
\end{equation}
where $F_{\mathrm{R}}=2\cos x\cos y+\cos2y$, $F_{\mathrm{Im}}=2\cos x\sin
y-\sin2y$, and $F_{\mathrm{NN}}=2\cos(2x/\sqrt{3})+4\cos(x/\sqrt{3})\cos
\sqrt{3}y+3$ with $x=k_{x}a_{0}\sqrt{3}/2$, $y=k_{y}a_{0}/2$ and $a_{0}$ being
the inter-atomic distance. The first and second nearest neighbor hoppings are
$t_{0}$ and $t_{1}$, respectively, and a nonzero $t_{1}$ breaks the
particle-hole symmetry. The staggered sublattice potential strength is
$\Delta$.
Next, we add an adiabatic perturbation term due to the microscopic local
rotation of atoms, and mainly follow the treatment introduced in Ref.
\cite{Murakami2020}, where a right handed circularly polarized optical phonon
mode at $\Gamma$ point is considered, with frequency $\omega$ and displacement
vectors%
\begin{equation}
\boldsymbol{u}_{A}=u_{0}(\cos\omega t,\sin\omega t),\text{ \ }\boldsymbol{u}%
_{B}=-\boldsymbol{u}_{A}%
\end{equation}
of A and B atoms on the two sublattices. There is a phase difference $\pi$
between the circular rotations of atoms A and B (see also Fig
\ref{MurakamiSpinless}(a)), thus the nearest neighbor bond lengths change with
time by the microscopic local rotations, while the next nearest neighbor ones
do not. One can hence take these rotations as the modulation of the nearest
neighbor hopping. By writing down the tight-binding Hamiltonian and converting
it to a $\boldsymbol{k}$-space one, the resultant adiabatic perturbation to
$\hat{H}(\boldsymbol{k})$ reads
\begin{align}
\delta\hat{H}(\boldsymbol{k},t)= & -\delta t_{0}(\sigma_{x}\sin y+\sigma
_{y}\cos y)\sqrt{3}\sin x\cos\omega t\nonumber\\
+ & \delta t_{0}[(\cos x\cos y-\cos2y)\sigma_{x}\nonumber\\
& -(\cos x\sin y+\sin2y)\sigma_{y}]\sin\omega t,
\end{align}
where $\delta t_{0}\propto2u_{0}$ arises from the change of the first nearest
neighbor hopping energy due to the variation of the inter-atomic distance by
the local rotations \cite{Murakami2020}. In our calculation, we consider the
chemical potential inside the band gap and set $t_{0}$ as the energy unit,
$t_{1}=0.1t_{0}$, and $\delta t_{0}=0.1t_{0}$.
With $\Delta=0.2t_{0}$, we plot energy bands in the absence of phonons in
Fig.~\ref{MurakamiSpinless}(a) where a band gap opens and the energy bands do
not show particle-hole symmetry. In the presence of phonons, the adiabatic
evolution of the electronic states due to the local rotations of A and B atoms
leads to a time-dependent orbital magnetic moment, which is plotted in
Fig.~\ref{MurakamiSpinless}(b) in units of $\frac{et_{0}}{\hbar}a^{2}%
\frac{\hbar\omega}{t_{0}}$ (magnetic moment upon an area of $a^{2}$) in an
evolution period. It is apparent that the induced orbital magnetization is
proportional to the phonon frequency $\omega$. By using the parameters of
graphene with $t_{0}=3$ eV and the lattice constant $a=\sqrt{3}a_{0}=2.46$
\r{A}, $\frac{et_{0}}{\hbar}a^{2}$ is about 4.77$\mu_{B}$ with $\mu_{B}$ as
the Bohr magneton. One can find a weak oscillation with a nonzero net
contribution (about $-3.5\times10^{-3}$ $\mu_{B}\frac{\hbar\omega}{t_{0}}$)
upon one period of the local rotations of atoms, which implies a nonzero
pumping of orbital magnetization. The $\boldsymbol{k}$-resolved instantaneous
orbital magnetization $\partial_{\boldsymbol{B}}a^{t}$ is plotted in
Fig.~\ref{MurakamiSpinless}(c) where one can find that the main contribution
is from K and K' points with minimal interband spacings. As the band gap
decreases, the magnitude of the induced magnetization increases rapidly
(Fig.~\ref{MurakamiSpinless}(d)).
We verify that the induced orbital magnetization vanishes with the next
nearest neighbor hopping parameter $t_{1}$ in the insert in
Fig.~\ref{MurakamiSpinless}(b), which also shows that the pumped magnetization
is proportional to $t_{1}$ at least up to $t_{1}=0.1t_{0}$. The resultant
orbital magnetization pumping in the above toy model is of the same order as
the pumped spin magnetization in a honeycomb lattice with very strong Rashba
spin-orbit coupling (the Rashba coefficient equals to $0.4t_{0}$)
\cite{Murakami2020}. We also mention that, in this latter four-band model,
breaking the particle-hole symmetry is not required for supporting nonzero
orbital magnetization pumped by rotations of atoms, and the pumped orbital and
spin magnetization are also generally comparable (not shown here).
Furthermore, taking $\hbar\omega/t_{0}=0.1$ (the ratio of the typical energy
scales of phonons and electrons), the magnetic-moment pumping due to the
periodic adiabatic change of electronic states induced by microscopic
rotations of atoms is of the order of the nuclear magneton.
\subsection{Intrinsic nonlinear anomalous Ettingshausen effect in 2D metals}
Not only the electric current but also a thermal current carried by Bloch
electrons can be induced by an applied electric field. In the intrinsic linear
thermal current response to the electric field, the zero-field orbital
magnetization plays a vital role \cite{Cooper1997,Xiao2006,Xiao2020EM}. It is
therefore anticipated that the electric-field induced orbital magnetization is
indispensable in the second-order nonlinear intrinsic thermal current response
to the electric field, i.e., the nonlinear intrinsic anomalous Ettingshausen
effect. In this subsection we discuss in more detail the semiclassical picture
of the electric-field induced orbital magnetization in 2D metals, and point
out its key role in the intrinsic nonlinear anomalous Ettingshausen effect.
First, the electric-field modified orbital magnetization can be recast into an
instructive form%
\begin{equation}
\boldsymbol{M}=\int(f^{0}\boldsymbol{\tilde{m}}+eg^{0}\boldsymbol{\tilde
{\Omega}})\label{M guess}%
\end{equation}
in analogy to the magnetization (\ref{M-eq}) in the absence of electric
fields. Here $\boldsymbol{\tilde{\Omega}}=\partial_{\boldsymbol{k}}%
\times\lbrack\boldsymbol{\mathcal{A}}^{k}+(\boldsymbol{a}^{k})^{\boldsymbol{E}%
}]$ is the electric-field modified Berry curvature, and%
\begin{equation}
\boldsymbol{\tilde{m}}=\boldsymbol{m}+\partial_{\boldsymbol{B}}%
(e\boldsymbol{E}\cdot\boldsymbol{a}^{k})+e\boldsymbol{v}^{0}\times
(\boldsymbol{a}^{k})^{\boldsymbol{E}}%
\end{equation}
is the orbital moment $\boldsymbol{\tilde{m}}=\frac{e}{2}\langle W|\left(
\boldsymbol{\hat{r}}-\boldsymbol{r}_{c}\right) \times\boldsymbol{\hat{v}%
}|W\rangle$ up to the first order of the electric field. $\boldsymbol{m}$ is
the zero-field orbital moment, $\boldsymbol{v}^{0}$ is the band velocity, and
$(\boldsymbol{a}^{k})^{\boldsymbol{E}}$ is the positional shift linear in the
electric field \cite{Gao2014}.
Second, a coarse graining process based on the wave-packet description of
Bloch electrons \cite{Xiao2006} shows that, the electric-field induced local
energy current density up to the second order is given by%
\begin{equation}
\boldsymbol{j}^{\text{E}}=-e\boldsymbol{E}\times\int f^{0}\varepsilon
\boldsymbol{\tilde{\Omega}}-\boldsymbol{E}\times\int f^{0}\boldsymbol{\tilde
{m}}.
\end{equation}
A magnetization current $\boldsymbol{j}^{\text{E,mag}}$ should be discounted
to obtain the transport energy current density \cite{Cooper1997}
$\boldsymbol{j}^{\text{E,tr}}=\boldsymbol{j}^{\text{E}}-\boldsymbol{j}%
^{\text{E,mag}}$. In uniform crystals, the energy magnetization current at the
linear order of the electric field is given by the the material-dependent part
of the Poynting vector describing the energy flow \cite{Xiao2006}:
$\boldsymbol{j}^{\text{E,mag}}=-\boldsymbol{E}\times\boldsymbol{M}^{0}$. In
the present nonlinear response, one has
\begin{equation}
\boldsymbol{j}^{\text{E,mag}}=-\boldsymbol{E}\times\boldsymbol{M}.
\end{equation}
Consequently, the transport thermal current is given by
\begin{equation}
\boldsymbol{j}^{\text{h,tr}}=\boldsymbol{j}^{\text{E,tr}}-\frac{\mu}%
{e}\boldsymbol{j}=-e\boldsymbol{E}\times T\int s\left( \varepsilon\right)
\boldsymbol{\tilde{\Omega}},
\end{equation}
where $s\left( \varepsilon\right) =\left[ \left( \varepsilon-\mu\right)
f^{0}-g^{0}\right] /T$ is the entropy density contributed by a particular
Bloch state, $\mu$ is the chemical potential, $T$ is the temperature, and we
have made use of the result for the intrinsic nonlinear Hall electric current
\cite{Gao2014} $\boldsymbol{j}=-e^{2}\boldsymbol{E}\times\int f^{0}%
\boldsymbol{\tilde{\Omega}}$.
By integration by parts, the entropy density takes the form of $s\left(
\varepsilon\right) =\int d\eta\left( \eta-\mu\right) \partial_{\mu}f\left(
\eta\right) \theta\left( \eta-\varepsilon\right) /T$, which renders the
thermal transport current to be%
\begin{equation}
\frac{\boldsymbol{j}^{\text{h,tr}}}{T}=-\frac{1}{e}\int d\eta\frac{\eta-\mu
}{T}\frac{\partial f\left( \eta\right) }{\partial\eta}\boldsymbol{j}\left(
\eta\right) .
\end{equation}
Here $\boldsymbol{j}\left( \eta\right) =-e^{2}\boldsymbol{E}\times\int
\theta\left( \eta-\varepsilon\right) \boldsymbol{\tilde{\Omega}}$ is the
nonlinear Hall electric current at zero temperature with Fermi energy $\eta$.
This equation is completely parallel to the generalized Mott relation between
the transport thermal current and electric current in the linear order of
electric fields. At low temperatures much less than the distances between the
chemical potential and band edges, the Sommerfeld expansion is legitimate
\cite{Xiao2016}, hence the entropy density reduces to $s\left( \varepsilon
\right) =\frac{1}{3}\pi^{2}k_{B}^{2}T\delta\left( \mu-\varepsilon\right) $,
which is concentrated on the Fermi surface and decays dramatically away from
it. Then the standard form of the Mott relation follows%
\begin{equation}
\boldsymbol{j}^{\text{h}}/T=(\pi^{2}k_{B}^{2}T/3e)[\partial\boldsymbol{j}%
\left( \varepsilon\right) /\partial\varepsilon]|_{\varepsilon=\mu}.
\end{equation}
Therefore, we extend the regime of validity of the Mott relation to the
second-order intrinsic thermoelectric current responses to a constant electric field.
\section{Summary}
We have formulated a semiclassical theory for the orbital magnetization
induced by general adiabatic evolutions of Bloch electronic states. This
theory starts from formulating the electric current density in bulk, from
which the magnetization can be extracted. The induced orbital magnetization is
gauge dependent in general case but is gauge invariant only when the adiabatic
time dependence is implicit or averaged out. These two cases correspond to the
orbital magnetoelectric response and the periodic-evolution pumped orbital magnetization.
In the orbital magnetoelectric effect the adiabatic evolution is driven by a
constant electric field, and the time dependence is only implicit through the
evolution of mechanical crystal momentum. Thus the pertinent second Chern form
current vanishes in 2D, making the 2D orbital magnetoelectricity governed
completely by the perturbative term of the reciprocal-space Berry connection
(Eqs. (\ref{2D OM}) and (\ref{OMP})), irrespective of insulators or metals.
The role of the orbital magnetoelectricity in the nonlinear intrinsic
anomalous Ettingshausen effect, which is proposed here as a transverse thermal
current response in the second order of the driving electric field, has also
been revealed in 2D metals. On the other hand, the orbital magnetoelectricity
and the nonlinear intrinsic anomalous Ettingshausen effect in 3D metals are
beyond the scope of the present theory. They may not be determined solely by
bulk considerations and are left for future efforts.
In the context of the orbital magnetization pumped by periodic adiabatic
evolutions in non-Chern insulators, the Chern-Simons contribution deduced from
the second Chern form current can be present even in 2D. Meanwhile, as a
second Chern form can be nonzero only if the system has more than two bands
\cite{Xiao2009}, in a two-band minimal model the pumped magnetization is
dictated solely by the perturbative term of the time component of the Berry
connection [Eqs. (\ref{OM pumping}) and (\ref{at-1})]. We illustrated the
orbital magnetization pumping due to the periodic adiabatic change of
electronic states induced by microscopic rotations of atoms in toy models
based on the honeycomb lattice. The induced magnetization is of the same order
as the pumped spin magnetization via strong Rashba spin orbit coupling.
The presented formulation is based on the assumption of well separated
nondegenerate Bloch bands, whereas to explore the semiclassical theories in
the case of degenerate bands and of closely located bands with possible
non-adiabatic effects \cite{Tu2020,Woods2020} need separate studies.
\begin{acknowledgments}
We thank Luka Trifunovic and Liang Dong for enlightening discussions.
This work was supported by NSF (EFMA-1641101) and Welch Foundation (F-1255).
\end{acknowledgments}
| 24d93cd6e99e75c996b6bb8d872dd647bbffd428 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The \ac{iot} provides a number of benefits,
to consumers as well as business organisations.
Flexibility and ease of use,
coupled with low management overhead and cost make it a desirable concept.
Smart cities,
for example,
can benefit from \ac{iot} solutions~\cite{formisano2015advantages},
as well as cloud applications~\cite{nastic2014provisioning}.
For business organisations,
logistics and asset management can be aided by the \ac{iot}~\cite{ding2013study}.
The \ac{iot} is characterised by its heterogeneity,
as an abundance of different devices can constitute an \ac{iot}.
Communication as well as embedded computation capabilities mark the common denominator on which \ac{iot} devices and networks depend.
While these features are a given in the age of constant mobile connectivity and open \acp{wlan},
they also constitute the vulnerabilities of \ac{iot} networks.
Being connected to networks leads to an increased attack surface.
Furthermore,
\ac{iot} devices are often cheap and manufactured in large numbers for short periods of time,
until they become obsolete.
After that,
newer versions are produced,
often for small prices.
Consequently,
re-use of hard- and software as well as low effort in programming contribute to insecure operating conditions~\cite{Spring.2016}.
The \ac{iot} is but a part of the development towards increased connectivity that inherently carries higher risks of cyber attacks.
Botnets targeting \ac{iot} devices,
such as \textit{Mirai}~\cite{kolias2017ddos},
industrial environments falling prey to attackers,
such as the power grid in the Ukraine in December 2015~\cite{Cherepanov.2017},
and ransomware attacks on healthcare infrastructure~\cite{slayton2018ransomware} and consumers~\cite{richardson2017ransomware} alike show the need for increased automated security in this brave new digital world.
As the networking paradigms are shifting from classic home and office networks to heterogeneous ad hoc networks,
security solutions have to adapt as well~\cite{Plaga.2019}.
Offensive security measures become an administrators friend to discover vulnerabilities along the attack phases.
By preemptive security,
such as vulnerability checking,
threats can be mitigated before an attacker can exploit them.
These capabilities are more relevant in the heterogeneous environments presented today.
Furthermore,
an insight regarding methods as well as tools of cyber criminals is becoming crucial for \ac{it} security professionals.
Since criminals often rely on publicly available tools,
an understanding of those allows security professionals to gain insight about the threat potential and possible attack vectors.
Furthermore,
if \ac{it} security professionals adapt to the methodology of a cyber criminal,
they obtain a new understanding of attacking a system,
potentially allowing for a more suited defense against attacks.
Additionally,
any vulnerability found by methods and tools of attackers is a vulnerability that can be mitigated before a real attack occurs.
The contribution of this work is
\begin{itemize}
\item the identification and collection of the most well-known and used open source security assessment tools and
\item the mapping of these tools to well-established attack models and.
\item the analysis, comparison and discussion of the capabilities of these tools.
\end{itemize}
The remainder of this work is structured as follows.
The state of the art is presented in Section~\ref{sec:sota}.
The methodology underlying this paper is introduced in Section~\ref{sec:methodology}.
The tools are introduced and evaluated in Section~\ref{sec:offsec_tools}.
A conclusion is drawn in Section~\ref{sec:conc}.
\section{State of the Art}
\label{sec:sota}
There is an abundance in literature regarding tools for offensive security purposes,
in numerous blogs,
but also in specialist books.
However,
an objective indication why the tools were chosen to be presented is not provided.
Either they are used and recommended by the author,
who usually is a security professional that has built a tool-box for themselves.
Or the tools are contained in a suite,
such as Kali Linux~\cite{kalilinux}.
\textit{Velu} presents the usage of Kali Linux for penetration testing in his book,
discussing the tools he deems most relevant~\cite{Velu.2016}.
\textit{Oakley} introduces red teaming in his book,
where tools are introduced in the respective stages based on experience of the author~\cite{Oakley.2019}.
\textit{Kim} presents practical penetration testing with the tools chosen in a similar fashion~\cite{Kim.2018}.
\textit{Forshaw} reduces the focus to tools for attacking network protocols,
thus setting a scope~\cite{Forshaw.2017}.
However,
the tools chosen are derived from his long experience.
In general,
it is beneficial to have tools introduced by professionals with a long experience in the fields,
as they took a long time to chose the right tool-box and become acquainted with it.
In this work,
however,
the focus is on creating tangible,
objective criteria for rating offensive security tools.
\section{Methodology}
\label{sec:methodology}
This section presents the methodology on which this work is founded.
First,
definitions of the terms are presented,
after that,
the sources from which the tools are collected.
Furthermore,
the scope of tools and applications is discussed,
a metric for attack stages is presented as well as the feature criteria of the tools.
\subsection{Definitons}
\label{ssec:definitions}
This paragraph presents the definitions of terms used in this work that are underlying to the evaluation. \par
\textit{Offensive information security}: Often called red teaming or penetration testing.
It is a concept that describes using tools and methods of an attacker to detect security vulnerabilities which then can be fixed before an attacker can exploit them. \par
\textit{Tool}: Finding a definition of the term tool in the context of software is exceedingly difficult.
For this work,
a tool is defined as a software program that can be used as such without further software,
except for operating system and corresponding environment.
A tool can consist of related parts that could be used in a stand alone-fashion,
but are distributed and commonly used together as they follow purposes along the path of a security assessment. \par
\textit{Freeware}: In the context of this work,
this term describes tools which can be obtained by private persons and professionals alike free of charge.
The free usage is not limited regarding the time of usage,
so trial versions of commercial software are not considered.
They are not specific to organisations. \par
\textit{Enterprise networks}: Consisting of \ac{it} infrastructures,
such as computers and servers.
Specifically excluded are \ac{ot} environments as found in industrial environments.
\subsection{Data Sources}
As the collection of exhaustive,
consistent lists of security tools with their attribution to a specific attack phase is difficult,
several sources were considered when identifying tools to evaluate in this work.
First,
the literature presented in Section~\ref{sec:sota} was used to extract the tools the authors used.
Second,
comprehensive lists that can be found online were employed.
The nmap project~\cite{nmap} provides a list of security tools,
called sectools~\cite{sectools}.
Furthermore,
\textit{r0lan} provides an overview of tools that is attributed to the phases they are used in~\cite{awesome-red-teaming}.
From these sources,
the most relevant tools were extracted.
Third,
well-known security distributions such as Kali~\cite{kalilinux} and Parrot~\cite{parrotlinux} Linux contain the tools that are most established in the security community.
\subsection{Scope}
The scope of this work are enterprise networks as discussed in Section~\ref{ssec:definitions},
consisting of computers,
servers and auxiliary devices.
Furthermore,
the scope regarding the tools is limited on tools with a security focus.
Since there is a trend in security research as well as cyber criminals to use tools that are already installed on the target machine for exploitation purposes,
many tasks in security assessment can be performed without security-specific tools.
This technique is called living off the land.
An example is the use of Microsoft PowerShell for enumerating users and directories.
\subsection{Attack Metrics: MITRE ATT\&CK}
\label{sec:attack_metrics}
The MITRE ATT\&CK matrix~\cite{mitre} was developed based on the well-established Lockheed Martin Cyber Kill Chain~\cite{cyberkillchain}.
Both aim at splitting a cyber attack into distinct phases during which an attacker follows a certain goal.
This is used to aid in comprehending the objectives of an attack and ultimately mitigating it.
The structure of the MITRE ATT\&CK model is shown in Figure~\ref{fig:mitre}.
\begin{figure*}
\includegraphics[width=\textwidth,height=7cm]{crop.pdf}
\caption{MITRE ATT\&CK Model}
\label{fig:mitre}
\end{figure*}
Each of these phases requires a different set of tools,
so the aim of this work is mapping tools to these phases as discussed in Section~\ref{sec:offsec_tools}.
The phases used in the MITRE ATT\&CK Enterprise matrix are as follows,
with the description according to the MITRE-homepage~\cite{mitre}. \par
\textit{Reconnaissance}: The phase during which an adversary collects information about the target.
Generally,
reconnaissance techniques are categorised in active,
i.e. with the adversary interacting with the target system in an unexpected way,
and passive,
i.e. the adversary not directly interacting with the target system.\par
\textit{Resource Development}: The adversary is obtaining resources that can aid in attacking the system,
such as accounts,
systems,
and other capabilities.
The resources might be used in later phases,
such as \ac{cc}.\par
\textit{Initial Access}: The adversary attempts to gain an initial foothold on the target system.
This is the first phase with direct adversarial action on the target.\par
\textit{Execution}: The adversary executes malicious code on the target system.
This code execution usually follows an underlying goal.
Often,
on-board capabilities of the target system,
such as compilers and interpreters,
aid in the execution of malicious code.\par
\textit{Persistence}: The phase in which the adversary aims to secure the foothold.
Persistence allows re-entry and access to the system after the adversary logged out or the system rebooted.\par
\textit{Privilege Escalation}: In this phase,
the adversary aims at obtaining higher privileges.
Often,
certain users are restricted from performing security-critical tasks,
and the first foothold was performed with such restricted accounts.
Elevating privileges allows the adversary to perform a wider variety of actions.\par
\textit{Defense Evasion}: After gaining access and elevating the privileges,
the adversary actively evades detection by \ac{ids}.
Obfuscation of the tools as well as deactivation of security measures are performed in this phase.\par
\textit{Credential Access}: The adversary aims at stealing account credentials for further use and to aid in following phases.\par
\textit{Discovery}: In this phase,
the adversary is gaining information about the environment in which the target system is located.
This includes machines and services,
accounts and users.\par
\textit{Lateral Movement}: In this phase,
the adversary is moving through the target environment and infecting new systems.
Often,
the foothold with which entry to the network was gained does not contain the desired target,
so lateral movement is necessary to reach devices that are not directly reachable from the outside.\par
\textit{Collection}: In this phase,
the adversary gathers the desired information from the target machine.\par
\textit{Command and Control (C\&C)}: In this phase,
the adversary executes control over the targeted systems and communicates with them.\par
\textit{Exfiltration}: The phase in which the adversary attempts to steal and obtain data without the owner of the target system noticing.
The data has to be sent in a fashion that does not cause suspicion.\par
\textit{Impact}: In this phase,
the adversary maliciously impacts the target system by destroying or restricting its functionality.
This activity can easily be detected by the owner of the system.\par
\subsection{Tool Features}
In order to evaluate and rate the tools,
a metric needs to be defined.
This metric should contain tangible,
verifiable features.
The features used in this work are listed as follows:
\begin{itemize}
\item Actively maintained: This feature is evaluated according to the latest release and the average number of releases per year. Actively maintained tools provide bug fixes and the integration of new features and protocols as well as a more active support.
\item Usage: This feature discusses the licence a tool is published under, the support a user can expect and whether or not a paid version of the tool is available.
\item Technical: This feature describes the interface of the tool for a user as well as the programming language the tool is programmed in. This is important as the way of interaction can make a tool more difficult or easier to apply, while the programming language in open source tools describes whether or not a user could adapt and extend the tool.
\end{itemize}
These features allow an assessment of the tools according to several dimensions.
It can be derived if the tool is actively developed and likely to be adapted to new technologies.
Furthermore,
the capabilities for extending and embedding the tool into a toolchain can be obtained from these features.
\section{Analysis}
\label{sec:offsec_tools}
In this section the relevant tools are identified and the metric is applied for comparison.
Then, the results for the comparison are discussed.
\subsection{Identification}
For the evaluation,
well-established,
commonly used security tools tailored for each of the phases as presented in Section~\ref{sec:attack_metrics} were identified,
collected and compared.
As it is not trivial to find an exhaustive overview of security tools,
several sources were used to obtain information.
Apart from books presented in Section~\ref{sec:sota},
web resources were considered as well.
The nmap network scanner~\cite{nmap} hosts a list of security tools~\cite{sectools}.
Every tool that fits the scope of this work in the top 50 tools of the sectools-list is evaluated in this work.
Thus,
a good coverage of relevant tools is expected.
Furthermore,
\textit{r0lan} provides an overview of tools commonly used for the individual phases~\cite{awesome-red-teaming}.
Among links to tools,
\textit{r0lan} provides an abundance of sources that describe methods rather than tools,
and information how to employ tools without a first focus on security,
such as PowerShell,
to perform security-relevant tasks.
From these sources,
the most used tools were extracted.
They are listed in Table~\ref{tab:methodology_vulnerabilities_siemens_overview},
categorised according to the phases of the MITRE ATT\&CK metric.
\subsection{Comparison}
The tools identified in the previous subsection are listed in Table~\ref{tab:methodology_vulnerabilities_siemens_overview},
with the criteria alongside which they are evaluated.
\begin{table*}[ht]
\renewcommand{\arraystretch}{1.3}
\caption{Security Assessment Tools Categorised According to the Phases They Are Used During}
\label{tab:methodology_vulnerabilities_siemens_overview}
\centering
\scriptsize
\begin{tabular}{l c c c c c c c c c c c}
\toprule
\textbf{Tools} & \phantom{a} & \multicolumn{3}{c}{\textbf{Releases}} & \phantom{a} & \multicolumn{3}{c}{\textbf{Usage aspects}} & \phantom{a} & \multicolumn{2}{c}{\textbf{Technical aspects}} \\
& & First & p.a. & Latest & & Licence & Supp. & Paid & & Interf. & Lang. \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Reconnaissance} \\
nmap~\cite{nmap} & & 1997 & 10 & 2020 & & Nmap Publ. Src. & Forum & no & & GUI, CLI & Lua, C, C++, Python, Shell \\
OpenVAS~\cite{openvas} & & 2006 & 9 & 2020 & & GNU GPLv2 2.0 & Forum & no & & GUI, CLI & C, NASL, Yacc, Shell, C++ \\
Maltego~\cite{maltego} & & 2007 & 11 & 2020 & & Proprietary & Forum & yes & & GUI & Java \\
Shodan~\cite{shodan} & & 2009 & n/a & 2020 & & Proprietary & Forum & yes & & WUI, CLI & n/a \\
TheHarvester~\cite{theharvester} & & 2011 & 3 & 2020 & & GPLv2 & Forum & no & & CLI & Python \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Initial Access} \\
Aircrack-ng~\cite{aircrackng} & & 2006 & 2 & 2020 & & GPLv2 & Forum & no & & GUI, CLI & C, M4, C\#, Shell, Python, Roff \\
GoPhish~\cite{gophish} & & 2013 & 2 & 2020 & & MIT & GitHub & no & & WUI, CLI & Go, JavaScript \\
msf~\cite{msf} & & 2014 & 2 & 2020 & & Apache 2.0 & Forum & no & & GUI, CLI & C++ \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Persistence} \\
C99 webshell~\cite{c99webshell} & & 2005 & n/a & n/a & & n/a & n/a & no & & CLI & PHP \\
Reptile~\cite{reptile} & & 2018 & 1 & 2020 & & n/a & n/a & no & & CLI & C, C++, Yacc, Perl, Shell \\
SharPersist~\cite{sharpersist} & & 2019 & 1 & 2020 & & Apache 2.0 & Forum & no & & CLI & C\# \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Privilege Escalation} \\
searchsploit~\cite{searchsploit} & & 2014 & n/a & 2020 & & GPL-2.0 & Forum & no & & CLI & C, Python, Ruby, Perl, PHP \\
msf~\cite{msf} & & 2014 & 2 & 2020 & & Apache 2.0 & Forum & yes & & GUI, CLI & C++ \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Defense Evasion} \\
veil-evasion~\cite{veil-evasion} & & 2015 & 21 & 2016 & & GPL-3.0 & Forum & no & & CLI & Python, C, Shell, C++ \\
shellter~\cite{shellter} & & 2014 & 14.7 & 2017 & & special & Forum & yes & & CLI & n/a \\
msf~\cite{msf} & & 2014 & 2 & 2020 & & Apache 2.0 & Forum & yes & & GUI, CLI & C++ \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Credential Access} \\
LaZange~\cite{lazagne} & & 2015 & 3.6 & 2019 & & LGPL-3.0 & Forum & no & & CLI & Python \\
Responder~\cite{responder} & & 2014 & 7.5 & 2015\footnotemark & & GPL-3.0 & Forum & no & & CLI & Python \\
Mimikatz~\cite{mimikatz} & & 2007 & 0.5 & 2020 & & CC-BY 4.0 & Forum & no & & CLI & C \\
msf~\cite{msf} & & 2014 & 2 & 2020 & & Apache 2.0 & Forum & no & & GUI, CLI & C++ \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Discovery} \\
linuxprivchecker~\cite{linuxprivchecker} & & 2015 & n/a & 2020 & & - & Forum & no & & CLI & Python \\
windows-privesc-check~\cite{windows-privesc-check} & & 2010 & n/a & 2015 & & GPL-2.0 & Forum & no & & CLI & Python \\
Bloodhound~\cite{bloodhound} & & 2016 & 7 & 2020 & & GPL-3.0 & Forum & no & & GUI, CLI & C++ \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Lateral Movement} \\
Mimikatz~\cite{mimikatz} & & 2007 & 0.5 & 2020 & & CC-BY 4.0 & Forum & no & & CLI & C \\
Rubeus~\cite{rubeus} & & 2018 & n/a & 2020 & & BSD 3-clause & Forum & no & & CLI & C\# \\
msf~\cite{msf} & & 2014 & 2 & 2020 & & Apache 2.0 & Forum & yes & & GUI, CLI & C++ \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Collection} \\
Veil-Pillage~\cite{veil-pillage} & & 2014 & n/a & 2015 & & GPL-3.0 & Forum & no & & CLI & PowerShell, Python \\
msf~\cite{msf} & & 2014 & 2 & 2020 & & Apache 2.0 & Forum & yes & & GUI, CLI & C++ \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Command and Control} \\
empire~\cite{empire} & & 2015 & 4.33 & 2018\footnotemark[\value{footnote}] & & BSD 3-clause & Forum & no & & CLI & PowerShell, Python \\
msf~\cite{msf} & & 2014 & 2 & 2020 & & Apache 2.0 & Forum & yes & & GUI, CLI & C++ \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Exfiltration} \\
DET~\cite{det} & & 2016 & n/a & 2019 & & MIT & Forum & no & & CLI & Python \\
Cloakify-Factory~\cite{cloakifyfactory} & & 2018 & 4 & 2018 & & MIT & Forum & no & & GUI, CLI & Python \\
msf~\cite{msf} & & 2014 & 2 & 2020 & & Apache 2.0 & Forum & yes & & GUI, CLI & C++ \\
\cmidrule{1-1} \cmidrule{3-5} \cmidrule{7-9} \cmidrule{11-12}
\textbf{Impact} \\
Veil-Pillage~\cite{veil-pillage} & & 2014 & n/a & 2015 & & GPL-3.0 & Forum & no & & CLI & PowerShell, Python \\
msf~\cite{msf} & & 2014 & 2 & 2020 & & Apache 2.0 & Forum & yes & & GUI, CLI & C++ \\
\bottomrule
\end{tabular}
\end{table*}
\footnotetext{These tools are indicated to be no longer actively maintained.}
It can be seen that the first stage has the most extensive number of tools available.
Most tools are still actively maintained,
meaning new releases are provided at the time of this work.
However,
a few tools,
empire~\cite{empire} and Responder~\cite{responder},
are indicated to be deprecated.
Furthermore,
several tools,
Veil-Evasion~\cite{veil-evasion},
Veil-Pillage~\cite{veil-pillage}
shellter~\cite{shellter},
windows-privesc-check~\cite{windows-privesc-check},
and Cloakify-Facory~\cite{cloakifyfactory} have their latest releases older than a year,
which indicates limited maintenance of these tools.
Since most versions are free,
do not have a paid version and are developed by members of the community,
most support is provided in terms of forums or mailing lists.
Every tool provides a \ac{cli},
some tools additionally provide a \ac{gui} or \ac{wui}.
The \ac{cli}-capabilities allow for the tool to be integrated into toolchains,
with the output being piped into other applications.
Furthermore,
since a number of tools is written in Python,
integration of the source code into user tools is easily possible.
All tools,
except for Maltego~\cite{maltego} and shellter~\cite{shellter},
provide their code for a user to extend and adapt.
\subsection{Discussion}
The features based on which the tools are evaluated are intended to provide information for a user to pick a tool that fits her need.
Furthermore,
the quality and suitability of tools,
as well as their versatility is evaluated.
For example,
msf~\cite{msf} can be used in nine of 14 phases,
making it the most versatile and powerful tool in this comparision.
This is due to the toolbox-approach that allows msf to load different modules for specific tasks.
Furthermore,
research underlying this work shows that some phases have significantly more tools created for them than others.
For example,
the execution phase does not have a singular tool dedicated to it.
Instead,
board measures of the target system are used,
or tools that are not security-specific,
such as web servers or communication tools.
Furthermore,
the variety of tools,
especially in the reconnaissance-phase,
is an indicator of knowledge and experience an \ac{it} security professional should have.
\section{Conclusion}
\label{sec:conc}
This work highlights a few insights.
First,
there is a plethora of \ac{it} security assessment tools available that can be used by professionals as well as cyber criminals.
Being aware of these tools and gaining familiarity is therefore crucial for security experts.
The majority is freely available and actively maintained,
with examples and help readily available.
Second,
some phases have more tools dedicated to them than others.
Reconnaissance has an abundance of publicly available tools solely for the purpose of gathering information that can be used to exploit a target.
The initial access phase has several tools as well,
as attacks can be aimed at different types of targets.
For example,
there is a number of tools to attack websites,
a different set of tools for attacking databases and further tools for other attack vectors.
Understanding these as a professional is crucial for hardening any potential attack vector and preventing attacks from happening in the first place.
Other phases rely on tools already available on the target systems,
such as programming compilers and interpreters,
PowerShell,
netcat and other tools.
These can be misused to perform malicious activity,
a fact of which \ac{it} security professionals need to be aware of as well.
This living off the land can be discussed in a future work,
as there is an abundance of tools that can be misused under the right circumstances.
Furthermore,
tools for cracking passwords and for monitoring and exploiting wireless devices were not considered in this work,
as they would exceed the scope.
An exhaustive overview of such tools can be discussed in a future work as well.
This work shows that gaining insight about attacks is crucial for defense.
A method commonly used to obtain information about attackers,
their tools and aims is honeypots~\cite{Fraunholz.2017c,Fraunholz.2017h}.
Attributing attackers and attacks is similarly important in order to implement counter measures suitable for the attacks which are identified as most likely or having most impact~\cite{Fraunholz.2017f}.
In general,
detecting attacks is becoming increasingly difficult due to the changes in \ac{it} infrastructure.
Using context information~\cite{Duque_Anton.2017c},
employing novel machine learning approaches~\cite{anton2019anomaly} and creating and providing data to train \acp{ids}~\cite{Duque_Anton.2019a} on are required to secure \ac{it} network against current and future attacks.
| 3c609f1b067084bc092810948e771b3b6ef86868 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |